diff --git a/parse/train/BkglSTNFDB/BkglSTNFDB.md b/parse/train/BkglSTNFDB/BkglSTNFDB.md new file mode 100644 index 0000000000000000000000000000000000000000..4bb32d970971b2c34ca2f6c759dd6d3321d32bc3 --- /dev/null +++ b/parse/train/BkglSTNFDB/BkglSTNFDB.md @@ -0,0 +1,683 @@ +# Q-LEARNING WITH UCB EXPLORATION IS SAMPLE EFFICIENT FOR INFINITE-HORIZON MDP + +Kefan Dong\*, Yuanhao Wang∗ +Institute for Interdisciplinary Information Sciences, Tsinghua University +{dkf16,yuanhao-16}@mails.tsinghua.edu. +Xiaoyu Chen +Key Laboratory of Machine Perception, MOE, School of EECS, +Peking University +cxy30@pku.edu.cn +Liwei Wang +Key Laboratory of Machine Perception, MOE, School of EECS +Center for Data Science, Peking University +wanglw@cis.pku.edu.cn + +# ABSTRACT + +A fundamental question in reinforcement learning is whether model-free algorithms are sample efficient. Recently, Jin et al. (2018) proposed a Q-learning algorithm with UCB exploration policy, and proved it has nearly optimal regret bound for finite-horizon episodic MDP. In this paper, we adapt Q-learning with UCB-exploration bonus to infinite-horizon MDP with discounted rewards without accessing a generative model. We show that the sample complexity of exploration of our algorithm is bounded by $\begin{array} { r } { \tilde { O } \big ( \frac { S A } { \epsilon ^ { 2 } ( 1 - \gamma ) ^ { 7 } } \big ) } \end{array}$ . This improves the previously best known result of $\begin{array} { r } { \tilde { O } \big ( \frac { S A } { \epsilon ^ { 4 } ( 1 - \gamma ) ^ { 8 } } \big ) } \end{array}$ in this setting achieved by delayed Q-learning (Strehl et al., 2006), and matches the lower bound in terms of $\epsilon$ as well as and up to logarithmic factors. + +# 1 INTRODUCTION + +The goal of reinforcement learning (RL) is to construct efficient algorithms that learn and plan in sequential decision making tasks when the underlying system dynamics are unknown. A typical model in RL is Markov Decision Process (MDP). At each time step, the environment is in a state $s$ . The agent takes an action $a$ , obtain a reward $r$ , and then the environment transits to another state. In reinforcement learning, the transition probability distribution is unknown. The algorithm needs to learn the transition dynamics of MDP, while aiming to maximize the cumulative reward. This poses the exploration-exploitation dilemma: whether to act to gain new information (explore) or to act consistently with past experience to maximize reward (exploit). + +Theoretical analyses of reinforcement learning fall into two broad categories: those assuming a simulator (a.k.a. generative model), and those without a simulator. In the first category, the algorithm is allowed to query the outcome of any state action pair from an oracle. The emphasis is on the number of calls needed to estimate the $Q$ value or to output a near-optimal policy. There has been extensive research in literature following this line of research, the majority of which focuses on discounted infinite horizon MDPs (Azar et al., 2011; Even-Dar & Mansour, 2003; Sidford et al., 2018b). The current results have achieved near-optimal time and sample complexities (Sidford et al., 2018b;a). + +Without a simulator, there is a dichotomy between finite-horizon and infinite-horizon settings. In finite-horizon settings, there are straightforward definitions for both regret and sample complexity; the latter is defined as the number of samples needed before the policy becomes near optimal. In this setting, extensive research in the past decade (Jin et al., 2018; Azar et al., 2017; Jaksch et al., 2010; Dann et al., 2017) has achieved great progress, and established nearly-tight bounds for both regret and sample complexity. + +The infinite-horizon setting is a very different matter. First of all, the performance measure cannot be a straightforward extension of the sample complexity defined above (See Strehl & Littman (2008) for detailed discussion). Instead, the measure of sample efficiency we adopt is the so-called sample complexity of exploration (Kakade et al., 2003), which is also a widely-accepted definition. This measure counts the number of times that the algorithm “makes mistakes” along the whole trajectory. See also (Strehl & Littman, 2008) for further discussions regarding this issue. + +Several model based algorithms have been proposed for infinite horizon MDP, for example Rmax (Brafman & Tennenholtz, 2003), MoRmax (Szita & Szepesvári, 2010) and UCRL- $\gamma$ (Lattimore & Hutter, 2012). It is noteworthy that there still exists a considerable gap between the state-of-the-art algorithm and the theoretical lower bound (Lattimore & Hutter, 2012) regarding $1 / ( 1 - \gamma )$ factor. + +Though model-based algorithms have been proved to be sample efficient in various MDP settings, most state-of-the-art RL algorithms are developed in the model-free paradigm (Schulman et al., 2015; Mnih et al., 2013; 2016). Model-free algorithms are more flexible and require less space, which have achieved remarkable performance on benchmarks such as Atari games and simulated robot control problems. + +For infinite horizon MDPs without access to simulator, the best model-free algorithm has a sample complexity of exploration $\begin{array} { r } { \tilde { \mathcal { O } } \big ( \frac { S A } { \epsilon ^ { 4 } ( 1 - \gamma ) ^ { 8 } } \big ) } \end{array}$ , achieved by delayed Q-learning (Strehl et al., 2006). The authors provide a novel strategy of argument when proving the upper bound for the sample complexity of exploration, namely identifying a sufficient condition for optimality, and then bound the number of times that this condition is violated. + +However, the results of Delayed Q-learning still leave a quadratic gap in $1 / \epsilon$ from the best-known lower bound. This is partly because the updates in Q-value are made in an over-conservative way. In fact, the loose sample complexity bound is a result of delayed Q-learning algorithm itself, as well as the mathematical artifact in their analysis. To illustrate this, we construct a hard instance showing that Delayed Q-learning incurs $\Omega ( 1 / \epsilon ^ { 3 } )$ sample complexity. This observation, as well as the success of the Q-learning with UCB algorithm (Jin et al., 2018) in proving a regret bound in finite-horizon settings, motivates us to incorporate a UCB-like exploration term into our algorithm. + +In this work, we propose a Q-learning algorithm with UCB exploration policy. We show the sample complexity of exploration bound of our algorithm is $\begin{array} { r } { \tilde { \mathcal { O } } \big ( \frac { S A } { \epsilon ^ { 2 } ( 1 - \gamma ) ^ { 7 } } \big ) } \end{array}$ . This strictly improves the previous best known result due to Delayed Q-learning. It also matches the lower bound in the dependence on $\epsilon$ , $S$ and $A$ up to logarithmic factors. + +We point out here that the infinite-horizon setting cannot be solved by reducing to finite-horizon setting. There are key technical differences between these two settings: the definition of sample complexity of exploration, time-invariant policies and the error propagation structure in Q-learning. In particular, the analysis techniques developed in (Jin et al., 2018) do not directly apply here. We refer the readers to Section 3.2 for detailed explanations and a concrete example. + +The rest of the paper is organized as follows. After introducing the notation used in the paper in Section 2, we describe our infinite Q-learning with UCB algorithm in Section 3. We then state our main theoretical results, which are in the form of PAC sample complexity bounds. In Section 4 we present some interesting properties beyond sample complexity bound. Finally, we conclude the paper in Section 5. + +# 2 PRELIMINARY + +We consider a Markov Decision Process defined by a five tuple $\langle S , \mathcal { A } , p , r , \gamma \rangle$ , where $s$ is the state space, $\mathcal { A }$ is the action space, $p ( s ^ { \prime } | s , a )$ is the transition function, $r : S \times \mathcal { A } [ 0 , 1 ]$ is the deterministic reward function, and $0 \leq \gamma < 1$ is the discount factor for rewards. Let $S = | S |$ and $A = | { \mathcal { A } } |$ denote the number of states and the number of actions respectively. + +Starting from a state $s _ { 1 }$ , the agent interacts with the environment for infinite number of time steps. At each time step, the agent observes state $s _ { t } \in S$ , picks action $a _ { t } \in \mathcal A$ , and receives reward $r _ { t }$ ; the system then transits to next state $s _ { t + 1 }$ . + +Using the notations in Strehl et al. (2006), a policy $\pi _ { t }$ refers to the non-stationary control policy of the algorithm since step $t$ . We use $V ^ { \pi _ { t } } ( s )$ to denote the value function under policy $\pi _ { t }$ , which is defined as $\begin{array} { r } { V ^ { \pi _ { t } } ( s ) = \mathbb { E } \bigl [ \sum _ { i = 1 } ^ { \infty } \gamma ^ { i - 1 } r ( s _ { i } , \pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = s \bigr ] } \end{array}$ . We also use $V ^ { * } ( s ) = \operatorname* { s u p } _ { \pi } V ^ { \pi } ( s )$ to denote the value function of the optimal policy. Accordingly, we define $\begin{array} { r } { \mathbb { E } [ \sum _ { i = 2 } ^ { \infty } \gamma ^ { i - 1 } r ( s _ { i } , \pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = \stackrel { \cdot } { s } , a _ { 1 } = \stackrel { \cdot } { a } ] } \end{array}$ as the $\mathrm { Q }$ function under policy $Q ^ { \pi _ { t } } ( s , a ) = r ( s , a ) +$ $\pi _ { t }$ ; $Q ^ { * } ( s , a )$ is the Q function under optimal policy $\pi ^ { * }$ . + +We use the sample complexity of exploration defined in Kakade et al. (2003) to measure the learning efficiency of our algorithm. This sample complexity definition has been widely used in previous works Strehl et al. (2006); Lattimore & Hutter (2012); Strehl $\&$ Littman (2008). + +Definition 1. Sample complexity of Exploration of an algorithm $\mathcal { A L G }$ is defined as the number of time steps t such that the non-stationary policy $\pi _ { t }$ at time $t$ is not $\epsilon$ -optimal for current state $s _ { t }$ , i.e. $\dot { V } ^ { \pi _ { t } } \left( s _ { t } \right) \dot { < } V ^ { * } \left( s _ { t } \right) - \epsilon .$ . + +Roughly speaking, this measure counts the number of mistakes along the whole trajectory. We use the following definition of PAC-MDP Strehl et al. (2006). + +Definition 2. An algorithm $\mathcal { A L G }$ is said to be PAC-MDP (Probably Approximately Correct in Markov Decision Processes) $i f ,$ for any  and $\delta$ , the sample complexity of $\mathcal { A L G }$ is less than some polynomial in the relevant quantities $( S , A , 1 / \epsilon , 1 / \delta , 1 / ( 1 - \gamma ) )$ , with probability at least $1 - \delta$ . + +Finally, recall that Bellman equation is defined as the following: + +$$ +\left\{ \begin{array} { l l } { V ^ { \pi _ { t } } ( s ) = Q ^ { \pi _ { t } } \left( s , \pi _ { t } ( s ) \right) } \\ { Q ^ { \pi _ { t } } ( s , a ) : = \left( r _ { t } + \gamma \mathbb { P } V ^ { \pi _ { t + 1 } } \right) ( s , a ) , } \end{array} \right. \quad \left\{ \begin{array} { l l } { V ^ { * } ( s ) = Q ^ { * } \left( s , \pi ^ { * } ( s ) \right) } \\ { Q ^ { * } ( s , a ) : = \left( r _ { t } + \gamma \mathbb { P } V ^ { * } \right) ( s , a ) , } \end{array} \right. +$$ + +which is frequently used in our analysis. Here we denote $\left[ \mathbb { P } V ^ { \pi _ { t } } \right] ( s , a ) : = \mathbb { E } _ { s ^ { \prime } \sim p ( \cdot \vert s , a ) } V ^ { \pi _ { t + 1 } } ( s ^ { \prime } ) .$ + +# 3 MAIN RESULTS + +In this section, we present the UCB Q-learning algorithm and the sample complexity bound. + +# 3.1 ALGORITHM + +# Algorithm 1 Infinite Q-learning with UCB + +Parameters: , γ, δ Initializ $\begin{array} { r l } & { \textnormal { \texttt { e } } Q ( s , a ) , \hat { Q } ( s , a ) \gets \frac { 1 } { 1 - \gamma } , N ( s , a ) \gets 0 , \epsilon _ { 1 } \gets \frac { \epsilon } { 2 4 R M \ln \frac { 1 } { 1 - \gamma } } , H \gets \frac { \ln 1 / ( ( 1 - \gamma ) \epsilon _ { 1 } ) } { \ln 1 / \gamma } . } \\ & { \cdot ( k ) = \ln ( S A ( k + 1 ) ( k + 2 ) / \delta ) , \alpha _ { k } = \frac { H + 1 } { H + k } . } \end{array}$ Define for $t = 1 , 2 , \dots \mathbf { d o }$ +5: Take action $a _ { t } \gets \arg \operatorname* { m a x } _ { a ^ { \prime } } \hat { Q } ( s _ { t } , a ^ { \prime } )$ Receive reward rt and transit to st+1 $\begin{array} { r l r } & { N ( s _ { t } , a _ { t } ) \gets N ( s _ { t } , a _ { t } ) + 1 } & \\ & { N ( s _ { t } , a _ { t } ) \gets N ( s _ { t } , a _ { t } ) + 1 } & \\ & { k \gets N ( s _ { t } , a _ { t } ) , b _ { k } \gets \frac { c _ { 2 } } { 1 - \gamma } \sqrt { \frac { H ( k ) } { k } } } & \\ & { \hat { V } ( s _ { t + 1 } ) \gets \operatorname* { m a x } _ { a \in A } \hat { Q } ( s _ { t + 1 } , a ) } & \\ & { Q ( s _ { t } , a _ { t } ) \gets ( 1 - \alpha _ { k } ) Q ( s _ { t } , a _ { t } ) + \alpha _ { k } \left[ r ( s _ { t } , a _ { t } ) + b _ { k } + \gamma \hat { V } ( s _ { t + 1 } ) \right] } \\ & { \hat { Q } ( s _ { t } , a _ { t } ) \gets \operatorname* { m i n } ( \hat { Q } ( s _ { t } , a _ { t } ) , Q ( s _ { t } , a _ { t } ) ) } & \end{array}$ t and can be set to $4 { \sqrt { 2 } }$ +10: end for + +Here $c _ { 2 } = 4 \sqrt { 2 }$ is a constant. $\begin{array} { r } { R = \lceil \ln \frac { 3 } { \epsilon ( 1 - \gamma ) } / ( 1 - \gamma ) \rceil } \end{array}$ , while the choice of $M$ can be found in Section. 3.3. ${ \cal { \left( M = 0 \left( \ln 1 / ( 1 - \gamma ) \epsilon \right) \right) } }$ ). The learning rate is defined as $\alpha _ { k } = ( H + 1 ) / ( H + k )$ . $H$ is chosen as $\frac { \ln { 1 } / ( ( 1 - \gamma ) \epsilon _ { 1 } ) } { \ln { 1 } / \gamma }$ , which satisfies $H \leq \frac { \ln 1 / ( ( 1 - \gamma ) \epsilon _ { 1 } ) } { 1 - \gamma }$ . + +Our UCB Q-learning algorithm (Algorithm 1) maintains an optimistic estimation of action value function $Q ( s , a )$ and its historical minimum value $\hat { Q } ( s , a )$ . $N _ { t } ( s , a )$ denotes the number of times that $( s , a )$ is experienced before time step $t$ ; $\tau ( s , a , k )$ denotes the time step $t$ at which $( s _ { t } , a _ { t } ) = ( s , a )$ for the $k$ -th time; if this state-action pair is not visited that many times, $\tau ( s , a , k ) = \infty$ . $Q _ { t } ( s , a )$ and $\hat { Q } _ { t } ( s , a )$ denotes the $Q$ and $\hat { Q }$ value of $( s , a )$ that the algorithm maintains when arriving at $s _ { t }$ respectively. + +# 3.2 SAMPLE COMPLEXITY OF EXPLORATION + +Our main result is the following sample complexity of exploration bound. + +Theorem 1. For any $\epsilon > 0$ , $\delta > 0 , 1 / 2 < \gamma < 1$ , with probability $1 - \delta$ , the sample complexity of exploration (i.e., the number of time steps t such that $\pi _ { t }$ is not $\epsilon$ -optimal at $s _ { t . }$ ) of Algorithm 1 is at most + +$$ +\tilde { \mathcal { O } } \left( \frac { S A \ln { 1 / \delta } } { \epsilon ^ { 2 } \left( 1 - \gamma \right) ^ { 7 } } \right) , +$$ + +where $\tilde { \mathcal { O } }$ suppresses logarithmic factors of $1 / \epsilon , 1 / ( 1 - \gamma )$ and $S A$ . + +We first point out the obstacles for proving the theorem and reasons why the techniques in Jin et al. +(2018) do not directly apply here. We then give a high level description of the ideas of our approach. + +One important issue is caused by the difference in the definition of sample complexity for finite and infinite horizon MDP. In finite horizon settings, sample complexity (and regret) is determined in the first $T$ timesteps, and only measures the performance at the initial state $s _ { 1 }$ (i.e. $( V ^ { * } - V ^ { \pi } ) ( s _ { 1 } ) )$ . However, in the infinite horizon setting, the agent may enter under-explored regions at any time period, and sample complexity of exploration characterizes the performance at all states the agent enters. + +The following example clearly illustrates the key difference between infinite-horizon and finitehorizon. Consider an MDP with a starting state $s _ { 1 }$ where the probability of leaving $s _ { 1 }$ is $o ( T ^ { - 1 } )$ . In this case, with high probability, it would take more than $T$ timesteps to leave $s _ { 1 }$ . Hence, guarantees about the learning in the first $T$ timesteps or about the performance at $s _ { 1 }$ imply almost nothing about the number of mistakes the algorithm would make in the rest of the MDP (i.e. the sample complexity of exploration of the algorithm). As a result, the analysis for finite horizon MDPs cannot be directly applied to infinite horizon setting. + +This calls for techniques for counting mistakes along the entire trajectory, such as those employed by Strehl et al. (2006). In particular, we need to establish convenient sufficient conditions for being $\epsilon$ -optimal at timestep $t$ and state $s _ { t }$ , i.e. $V ^ { \ast } ( s _ { t } ) - V ^ { \pi _ { t } } ( s _ { t } ) \leq \epsilon$ . Then, bounding the number of violations of such conditions gives a bound on sample complexity. + +Another technical reason why the proof in Jin et al. (2018) cannot be directly applied to our problem is the following: In finite horizon settings, Jin et al. (2018) decomposed the learning error at episode $k$ and time $h$ as errors from a set of consecutive episodes before $k$ at time $h + 1$ using a clever design of learning rate. However, in the infinite horizon setting, this property does not hold. Suppose at time $t$ the agent is at state $s _ { t }$ and takes action $a _ { t }$ . Then the learning error at $t$ only depends on those previous time steps such that the agent encountered the same state as $s _ { t }$ and took the same action as $a _ { t }$ . Thus the learning error at time $t$ cannot be decomposed as errors from a set of consecutive time steps before $t$ , but errors from a set of non-consecutive time steps without any structure. Therefore, we have to control the sum of learning errors over an unstructured set of time steps. This makes the analysis more challenging. + +Now we give a brief road map of the proof of Theorem 1. Our first goal is to establish a sufficient condition so that $\pi _ { t }$ learned at step $t$ is $\epsilon$ -optimal for state $s _ { t }$ . As an intermediate step we show that a sufficient condition for $V ^ { * } ( s _ { t } ) - V ^ { \pi _ { t } } ( s _ { t } ) \leq \epsilon$ is that $V ^ { \ast } ( s _ { t ^ { \prime } } ) - Q ^ { \ast } ( s _ { t ^ { \prime } } , a _ { t ^ { \prime } } )$ is small for a few time steps $t ^ { \prime }$ within an interval $[ t , t + R ]$ for a carefully chosen $R$ (Condition 1). Then we show the desired sufficient condition (Condition 2) implies Condition 1. We then bound the total number of bad time steps on which $V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )$ is large for the whole MDP; this implies a bound on the number of violations of Condition 2. This in turn relies on a key technical lemma (Lemma 2). + +The remaining part of this section is organized as follows. We establish the sufficient condition for $\epsilon$ -optimality in Section 3.3. The key lemma is presented in Section 3.4. Finally we prove Theorem 1 in Section 3.5. + +# 3.3 SUFFICIENT CONDITION FOR $\epsilon$ -OPTIMALITY + +In this section, we establish a sufficient condition (Condition 2) for $\epsilon$ -optimality at time step $t$ . + +For a fixed $s _ { t }$ , let TRAJ $( R )$ be the set of length- $R$ trajectories starting from $s _ { t }$ . Our goal is to give a sufficient condition so that $\pi _ { t }$ , the policy learned at step $t$ , is $\epsilon$ -optimal. For any $\epsilon _ { 2 } > 0$ , define $\begin{array} { r } { R : = \lceil \ln \frac { 1 } { \epsilon _ { 2 } ( 1 - \gamma ) } / ( 1 - \gamma ) \rceil } \end{array}$ . Denote $V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )$ by $\Delta _ { t }$ . We have + +$$ +\begin{array} { r l } & { \quad V ^ { * } ( s _ { t } ) - V ^ { \pi } ( s _ { t } ) } \\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + Q ^ { * } ( s _ { t } , a _ { t } ) - V ^ { \pi } ( s _ { t } ) } \\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \gamma \mathbb { P } ( V ^ { * } - V ^ { \pi } ) ( s _ { t } , \pi _ { t } ( s _ { t } ) ) } \\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \gamma \sum _ { s + 1 } ^ { \infty } p ( s _ { t + 1 } | s _ { t } , \pi _ { t } ( s _ { t } ) ) \cdot [ V ^ { * } ( s _ { t + 1 } ) - Q ^ { * } ( s _ { t + 1 } , a _ { t + 1 } ) ] + } \\ & { \quad \Big . \Big . \Big . \Big . \Big . \Big . } \\ & { \quad \quad \Big . \gamma _ { s + 1 , s ^ { * } + 2 } \mathrm { ~ } p ( s _ { t + 2 } | s _ { t + 1 } , \pi _ { t + 1 } ( s _ { t + 1 } ) ) \cdot p ( s _ { t + 1 } | s _ { t } , \pi _ { t } ( s _ { t } ) ) \big [ V ^ { * } ( s _ { t + 2 } ) - Q ^ { * } ( s _ { t + 2 } , a _ { t + 2 } ) \Big ] } \\ & { \quad \quad \cdot \dots } \\ & { \le \epsilon _ { 2 } + \displaystyle \sum _ { s ^ { \prime } = s ^ { \prime } } p ( t r a _ { t } ) \cdot [ \frac { n - 1 } { \lambda _ { 2 } } \gamma ^ { 2 } \Delta _ { t + 2 } ] , } \\ & { \quad \quad \pi _ { s \lambda ^ { ( 2 ) } ( L ) } ^ { R , a } } \end{array} +$$ + +where the last inequality holds because $\frac { \gamma ^ { R } } { 1 - \gamma } \leq \epsilon _ { 2 }$ , which follows from the definition of $R$ + +For any fixed trajectory of length $R$ starting from $s _ { t }$ , consider the sequence $\left( \Delta _ { t ^ { \prime } } \right) _ { t \leq t ^ { \prime } < t + R }$ . Let $X _ { t } ^ { ( i ) }$ be the $i$ -th largest item of $\left( \Delta _ { t ^ { \prime } } \right) _ { t \leq t ^ { \prime } < t + R }$ . Rearranging Eq. (1), we obtain + +$$ +V ^ { \ast } ( s _ { t } ) - V ^ { \pi _ { t } } ( s _ { t } ) \leq \epsilon _ { 2 } + E _ { t r a j } \left[ \sum _ { i = 1 } ^ { R } \gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \right] . +$$ + +We first prove that Condition 1 implies $\epsilon$ -optimality at time step $t$ when $\epsilon _ { 2 } = \epsilon / 3$ + +Condition 1. Let $\begin{array} { r } { \xi _ { i } : = \frac { 1 } { 2 ^ { i + 2 } } \epsilon _ { 2 } \left( \ln \frac { 1 } { 1 - \gamma } \right) ^ { - 1 } } \end{array}$ . For all $0 \leq i \leq \lfloor \log _ { 2 } R \rfloor$ + +$$ +E [ X _ { t } ^ { ( 2 ^ { i } ) } ] \leq \xi _ { i } . +$$ + +Claim 1. If Condition $^ { l }$ is satisfied at time step $t$ , the policy $\pi _ { t }$ is $\epsilon$ -optimal at state $s _ { t }$ , i.e. $V ^ { \ast } ( s _ { t } ) -$ +$V ^ { \pi _ { t } } ( s _ { t } ) \leq \epsilon$ . + +Proof. Note that $E [ X _ { t } ^ { ( 2 ^ { \lfloor \log _ { 2 } i \rfloor } ) } ]$ t. Eq. (3) implies that for $X _ { t } ^ { ( i ) }$ is monotonically decreasing with respect to $1 / 2 < \gamma < 1$ , $i$ . Therefore, $E [ X _ { t } ^ { ( i ) } ] \ \leq$ + +$$ +\begin{array} { r l } & { E \left[ \displaystyle \sum _ { i = 1 } ^ { R } \gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \right] = \displaystyle \sum _ { i = 1 } ^ { R } \gamma ^ { i - 1 } E [ X _ { t } ^ { ( i ) } ] \leq \displaystyle \sum _ { i = 1 } ^ { R } \gamma ^ { i - 1 } E [ X _ { t } ^ { ( 2 ^ { \lfloor \log _ { 2 } i \rfloor } ) } ] } \\ & { \leq \displaystyle \sum _ { i = 1 } ^ { R } \gamma ^ { i - 1 } 2 ^ { - \lfloor \log _ { 2 } i \rfloor - 2 } \epsilon _ { 2 } \left( \ln \frac { 1 } { 1 - \gamma } \right) ^ { - 1 } \leq \displaystyle \sum _ { i = 1 } ^ { R } \frac { \gamma ^ { i - 1 } } { i } \epsilon _ { 2 } \left( \ln \frac { 1 } { 1 - \gamma } \right) ^ { - 1 } \leq 2 \epsilon _ { 2 } , } \end{array} +$$ + +where the last inequality follows from the fact that $\begin{array} { r } { \sum _ { i = 1 } ^ { \infty } \frac { \gamma ^ { i - 1 } } { i } = \frac { 1 } { \gamma } \ln \frac { 1 } { 1 - \gamma } } \end{array}$ and $\gamma > 1 / 2$ + +Combining with Eq. 2, we have, $\begin{array} { r } { V ^ { \ast } ( s _ { t } ) - V ^ { \pi _ { t } } ( s _ { t } ) \leq \epsilon _ { 2 } + E \left[ \sum _ { i = 1 } ^ { R } \gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \right] \leq 3 \epsilon _ { 2 } = \epsilon . } \end{array}$ + +Next we show that given $i , t$ , Condition 2 implies Eq. (3). + +Condition 2. Define $L = \lfloor \log _ { 2 } R \rfloor$ . Let $\begin{array} { r } { M = \operatorname* { m a x } \left\{ \lceil 2 \log _ { 2 } \frac { 1 } { \xi _ { L } ( 1 - \gamma ) } \rceil , 1 0 \right\} } \end{array}$ , and $\begin{array} { r } { \eta _ { j } = \frac { \xi _ { i } } { M } \cdot 2 ^ { j - 1 } } \end{array}$ . +For all $\begin{array} { r } { \because j \le M , \eta _ { j } \operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \eta _ { j - 1 } ] \le \frac { \xi _ { i } } { M } , } \end{array}$ . + +Claim 2. Given i, t, Eq. (3) holds if Condition 2 is satisfied. + +Proof. The reason behind the choice of $M$ is to ensure that $\eta _ { M } > 1 / ( 1 - \gamma )$ 1. It follows that, assuming Condition 2 holds, for $1 \leq j \leq M$ , + +$$ +E \left[ { X _ { t } ^ { ( 2 ^ { i } ) } } \right] = \int _ { 0 } ^ { 1 / ( 1 - \gamma ) } \operatorname* { P r } \left[ { X _ { t } ^ { ( 2 ^ { i } ) } } > x \right] d x \le \eta _ { 1 } + \sum _ { j = 2 } ^ { M } \eta _ { j } \operatorname* { P r } [ { X _ { t } ^ { ( 2 ^ { i } ) } } > \eta _ { j - 1 } ] \le \xi _ { i } . +$$ + +Therefore, if a time step $t$ is not $\epsilon _ { 2 }$ -optimal, there exists $0 \leq i < \lfloor \log _ { 2 } R \rfloor$ and $2 \leq j \leq M$ such that + +$$ +\eta _ { j } \mathrm { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \eta _ { j - 1 } ] > \frac { \xi _ { i } } { M } . +$$ + +Now, the sample complexity can be bounded by the number of $( t , i , j )$ pairs that Eq. (4) is violated. Following the approach of Strehl et al. (2006), for a fixed $( i , j )$ -pair, instead of directly counting the number of time steps $t$ such that $\begin{array} { r } { \operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \eta _ { j - 1 } ] > \frac { \xi _ { i } } { M \eta _ { j } } } \end{array}$ , we count the number of time steps that $X _ { t } ^ { ( 2 ^ { i } ) } > \eta _ { j - 1 }$ . Lemma 1 provides an upper bound of the number of such $t$ . + +# 3.4 KEY LEMMAS + +In this section, we present two key lemmas. Lemma 1 bounds the number of sub-optimal actions, which in turn, bounds the sample complexity of our algorithm. Lemma 2 bounds the weighted sum of learning error, i.e. $( \hat { Q } _ { t } - Q ^ { * } ) ( s , a )$ , with the sum and maximum of weights. Then, we show that Lemma 1 follows from Lemma 2. + +Lemma 1. For fixed $t$ and $\eta > 0$ , let ${ B } _ { \eta } ^ { ( t ) }$ be the event that $\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \frac { \eta } { 1 - \gamma } } \end{array}$ in step t. If $\eta > 2 \epsilon _ { 1 }$ , then with probability at least $1 - \delta / 2$ , + +$$ +\sum _ { t = 1 } ^ { t = \infty } I \left[ B _ { \eta } ^ { ( t ) } \right] \leq \frac { S A \ln S A \ln 1 / \delta } { \eta ^ { 2 } ( 1 - \gamma ) ^ { 3 } } \cdot p o l y l o g \left( \frac { 1 } { \epsilon _ { 1 } } , \frac { 1 } { 1 - \gamma } \right) , +$$ + +where $I [ \cdot ]$ is the indicator function. + +Before presenting Lemma 2, we define a class of sequence that occurs in the proof. + +Definition 3. A sequence $( w _ { t } ) _ { t \geq 1 }$ is said to be a $( C , w )$ -sequence for $C , w > 0$ , $i f 0 \le w _ { t } \le w$ for all $t \geq 1$ , and $\begin{array} { r } { \sum _ { t \geq 1 } w _ { t } \leq C } \end{array}$ . + +Lemma 2. For every $( C , w )$ -sequence $( w _ { t } ) _ { t \geq 1 }$ , with probability $1 - \delta / 2$ , the following holds: + +$$ +\sum _ { t \ge 1 } w _ { t } ( \hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \le \frac { C \epsilon _ { 1 } } { 1 - \gamma } + \mathcal { O } \left( \frac { \sqrt { w S A C \ell ( C ) } } { ( 1 - \gamma ) ^ { 2 . 5 } } + \frac { w S A \ln C } { ( 1 - \gamma ) ^ { 3 } } \ln \frac { 1 } { ( 1 - \gamma ) \epsilon _ { 1 } } \right) . +$$ + +$\begin{array} { r } { \ell ( C ) = \iota ( C ) \ln { \frac { 1 } { ( 1 - \gamma ) \epsilon _ { 1 } } } } \end{array}$ + +Proof of Lemma 2 is quite technical, and is therefore deferred to supplementary materials. + +Now, we briefly explain how to prove Lemma 1 with Lemma 2. (Full proof can be found in supplementary materials.) Note that since $\hat { Q } _ { t } \geq Q ^ { * }$ and $a _ { t } = \arg \operatorname* { m a x } _ { a } \hat { Q } _ { t } ( s _ { t } , a )$ , + +$$ +V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \leq \hat { Q } _ { t } ( s _ { t } , a _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) . +$$ + +We now consider a set $J = \{ t : \ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \eta ( 1 - \gamma ) ^ { - 1 } \}$ , and consider the $( | J | , 1 )$ - weight sequence defined by $w _ { t } ~ = ~ I \left[ t \in J \right]$ . We can now apply Lemma 2 to weighted sum $\begin{array} { r } { \sum _ { t \geq 1 } \dot { w } _ { t } \left[ \bar { V } ^ { \ast } ( s _ { t } ) - Q ^ { \ast } ( s _ { t } , \dot { a } _ { t } ) \right] , } \end{array}$ . On the one hand, this quantity is obviously at least $| J | \bar { \eta ( 1 - \gamma ) } ^ { - 1 }$ . On the other hand, by lemma 2, it is upper bounded by the weighted sum of $( \hat { Q } - Q ^ { * } ) ( s _ { t } , a _ { t } )$ . Thus we get + +$$ +| J | \eta ( 1 - \gamma ) ^ { - 1 } \leq \frac { C \epsilon _ { 1 } } { 1 - \gamma } + \mathcal { O } \left( \frac { \sqrt { S A | J | \ell ( | J | ) } } { ( 1 - \gamma ) ^ { 2 . 5 } } + \frac { w S A \ln | J | } { ( 1 - \gamma ) ^ { 3 } } \ln \frac { 1 } { ( 1 - \gamma ) \epsilon _ { 1 } } \right) . +$$ + +Now focus on the dependence on $| J |$ . The left-hand-side has linear dependence on $| J |$ , whereas the left-hand-side has a $\tilde { \mathcal { O } } \left( \sqrt { | J | } \right)$ dependence. This allows us to solve out an upper bound on $| J |$ with quadratic dependence on $1 / \eta$ . + +# 3.5 PROOF FOR THEOREM 1 + +We prove the theorem by stitching Lemma 1 and Condition 2. + +Proof. (Proof for Theorem 1) + +By lemma 1, for any $2 \leq j \leq M , \sum _ { t = 1 } ^ { \infty } I \left[ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \eta _ { j - 1 } \right] \leq C$ , where $C = \frac { S A \ln S A \ln 1 / \delta } { \eta _ { j - 1 } ^ { 2 } ( 1 - \gamma ) ^ { 5 } } \cdot \tilde { P } .$ + +Here $\tilde { P }$ is a shorthand for polylog $\begin{array} { r } { \left( \frac { 1 } { \epsilon _ { 1 } } , \frac { 1 } { 1 - \gamma } \right) } \end{array}$ + +Let $A _ { t } = I [ X _ { t } ^ { ( 2 ^ { i } ) } \geq \eta _ { j - 1 } ]$ be a Bernoulli random variable, and $\{ \mathcal { F } _ { t } \} _ { t \ge 1 }$ be the filtration generated by random variables $\bar { \{ ( s _ { \tau } , a _ { \tau } ) : 1 \leq \tau \leq t \} }$ . Since $A _ { t }$ is $\mathcal { F } _ { t + R }$ −measurable, for any $0 \leq k < R$ , $\{ \dot { A } _ { k + t R } - E [ A _ { k + t R } \ | \ \mathcal { F } _ { k + t R } ] \} _ { t \geq 0 }$ is a martingale difference sequence. For now, consider a fixed $0 \leq k < R$ . By Azuma-Hoeffiding inequality, after $\begin{array} { r } { T = \mathcal { O } \left( \frac { C } { 2 ^ { i } } \cdot \frac { M \eta _ { j } } { \xi _ { i } } \ln ( R M L ) \right) } \end{array}$ · M ηjξ ln(RM L) time steps (if it happens that many times) with + +$$ +\operatorname* { P r } \left[ X _ { k + t R } ^ { ( 2 ^ { i } ) } \geq \eta _ { j - 1 } \right] = \mathbb { E } [ A _ { k + t R } ] > \frac { \xi _ { i } } { M \eta _ { j } } , +$$ + +we have $\textstyle \sum _ { t } A _ { k + t R } \geq C / 2 ^ { i }$ with probability at least $1 - \delta / ( 2 M R L )$ . + +On the other hand, if $A _ { k + t R }$ happens, within $[ k + t R , k + t R + R - 1 ]$ , there must be at least $2 ^ { i }$ time steps at which $V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > ^ { \mathsf { ^ { * } } } \eta _ { j - 1 }$ . The latter event happens at most $C$ times, and the $[ k + t R , k + t R + R - 1 ]$ are disjoint. Tppens at most erefore, times f $\textstyle \sum _ { t = 0 } ^ { \infty } A _ { k + t R } \leq C / 2 ^ { i }$ . This suggests that a union bound on $T$ $i$ $j$ $0 \leq k < R$ , we can show that with probability $1 - \delta / ( 2 M L )$ , there are at most $R T$ time steps where $\mathrm { P r } \left[ X _ { t } ^ { ( 2 ^ { i } ) } \geq \eta _ { j - 1 } \right] > \xi _ { i } / ( M \eta _ { j } )$ . Thus, the number of sub-optimal steps is bounded by, + +$$ +\begin{array} { r l r } { { \sum _ { t = 1 } ^ { \ell - 1 } I [ V ^ { * } ( s _ { t } ) - V ^ { \pi _ { t } } ( s _ { t } ) > \epsilon ] } } \\ & { \leq \sum _ { t = 1 } ^ { \infty } \sum _ { i = 0 } ^ { L } \sum _ { j = 2 } ^ { M } I [ \eta _ { j } \operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \eta _ { j - 1 } ] > \frac { \xi _ { i } } { M } ] = \sum _ { i = 0 } ^ { L } \sum _ { j = 2 } ^ { M } \sum _ { t = 1 } ^ { \infty } I [ \operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \eta _ { j - 1 } ] > \frac { \xi _ { i } } { \eta _ { j } M } ] } \\ & { \leq \sum _ { i = 0 } ^ { L } \sum _ { j = 2 } ^ { M } \frac { S A M R \ln 1 / \delta \ln S A } { \eta _ { j } \xi _ { i } \cdot 2 ^ { i } ( 1 - \gamma ) ^ { 5 } } \tilde { P } \leq \sum _ { i = 0 } ^ { L } \frac { S A \cdot 2 ^ { i + 4 } \ln S A \ln 1 / \delta } { \epsilon _ { 2 } ^ { 2 } ( 1 - \gamma ) ^ { 6 } } \tilde { P } \mathrm { ( B y ~ d e f i n i t i o n ~ o f } \xi _ { i } \mathrm { ~ a n d } \eta _ { j } ) } \\ & { \leq \frac { S A R \ln S A \ln 1 / \delta } { \epsilon _ { 2 } ^ { 2 } ( 1 - \gamma ) ^ { 6 } } \tilde { P } \leq \frac { S A \ln S A \ln 1 / \delta } { \epsilon _ { 2 } ^ { 2 } ( 1 - \gamma ) ^ { 7 } } \tilde { P } . } & { \mathrm { ( B y ~ d e f i n i t i o n ~ o f } \hbar ) } \end{array} +$$ + +It should be stressed that throughout the lines, $\tilde { P }$ is a shorthand for an asymptotic expression, instead of an exact value. Our final choice of 2 and 1 are 2 = 3 , and 1 = 24RM ln . It is not hard 1−γ to see that $\ln 1 / \epsilon _ { 1 } = \mathrm { p o l y } ( \ln \textstyle \frac { 1 } { \epsilon } , \ln \frac { 1 } { 1 - \gamma } )$ . This immediately implies that with probability $1 - \delta$ , the number of time steps such that $\left( V ^ { * } - \dot { V } ^ { \pi } \right) \left( s _ { t } \right) > \epsilon$ is + +$$ +\tilde { \mathcal { O } } \left( \frac { S A \ln 1 / \delta } { \epsilon ^ { 2 } ( 1 - \gamma ) ^ { 7 } } \right) , +$$ + +where hidden factors are $\begin{array} { r } { \operatorname { p o l y } ( \ln { \frac { 1 } { \epsilon } } , \ln { \frac { 1 } { 1 - \gamma } } , \ln S A ) } \end{array}$ . + +# 4 DISCUSSION + +In this section, we discuss the implication of our results, and present some interesting properties of our algorithm beyond its sample complexity bound. + +# 4.1 COMPARISON WITH PREVIOUS RESULTS + +Lower bound To the best of our knowledge, the current best lower bound for worst-case sample complexity is  SA2(1−γ)3 ln 1/δ due to Lattimore & Hutter (2012). The gap between our results and this lower bound lies only in the dependence on $1 / ( 1 - \gamma )$ and logarithmic terms of $\mathrm { { } } ^ { \mathrm { { 3 } } A , 1 / ( 1 - \gamma ) }$ and $1 / \epsilon$ . + +Model-free algorithms Previously, the best sample complexity bound for a model-free algorithm is $\begin{array} { r } { \tilde { \mathcal O } \left( \frac { S A } { \epsilon ^ { 4 } ( 1 - \gamma ) ^ { 8 } } \right) } \end{array}$ (suppressing all logarithmic terms), achieved by Delayed Q-learning Strehl et al. (2006). Our results improve this upper bound by a factor of $\frac { 1 } { \epsilon ^ { 2 } ( 1 - \gamma ) }$ , and closes the quadratic gap in $1 / \epsilon$ between Delayed Q-learning’s result and the lower bound. In fact, the following theorem shows that UCB Q-learning can indeed outperform Delayed Q-learning. + +Theorem 2. There exists a family of MDPs with constant $S$ and $A$ , in which with probability $1 - \delta$ , Delayed $Q$ -learning incurs sample complexity of exploration of $\Omega \left( \frac { \epsilon ^ { - 3 } } { \ln ( 1 / \delta ) } \right)$ , assuming that $\ln ( 1 / \delta ) < \epsilon ^ { - 2 }$ . + +The construction of this hard MDP family is given in the supplementary material. + +Model-based algorithms For model-based algorithms, better sample complexity results in infinite horizon settings have been claimed Szita & Szepesvári (2010). To the best of our knowledge, the best published result without further restrictions on MDPs is $\begin{array} { r } { \tilde { \mathcal O } \left( \frac { S A } { \epsilon ^ { 2 } ( 1 - \gamma ) ^ { 6 } } \right) } \end{array}$ claimed by Szita & Szepesvári (2010), which is $( 1 - \gamma )$ smaller than our upper bound. From the space complexity point of view, our algorithm is much more memory-efficient. Our algorithm stores $O ( S A )$ values, whereas the algorithm in Szita & Szepesvári (2010) needs $\Omega ( S ^ { 2 } A )$ memory to store the transition model. + +# 4.2 EXTENSION TO OTHER SETTINGS + +Due to length limits, detailed discussion in this section is deferred to supplementary materials. + +Finite horizon MDP The sample complexity of exploration bounds of UCB Q-learning implies $\tilde { \mathcal { O } } \left( \epsilon ^ { - 2 } \right)$ PAC sample complexity and a $\bar { \mathcal { O } } \left( T ^ { 1 / 2 } \right)$ regret bound in finite horizon MDPs. That is, our algorithm implies a PAC algorithm for finite horizon MDPs. We are not aware of reductions of the opposite direction (from finite horizon sample complexity to infinite horizon sample complexity of exploration). + +Regret The reason why our results can imply an $\tilde { \mathcal { O } } ( { \sqrt { T } } )$ regret is that, after choosing $\epsilon _ { 1 }$ , it follows from the argument of Theorem 1 that with probability $1 - \delta$ , for all $\epsilon _ { 2 } > \tilde { \mathcal { O } } ( \epsilon _ { 1 } / ( 1 - \gamma ) )$ , the number of $\epsilon _ { 2 }$ -suboptimal steps is bounded by + +$$ +\mathcal { O } \left( \frac { S A \ln S A \ln 1 / \delta } { \epsilon _ { 2 } ^ { 2 } ( 1 - \gamma ) ^ { 7 } } \mathrm { p o l y l o g } \left( \frac { 1 } { \epsilon _ { 1 } } , \frac { 1 } { 1 - \gamma } \right) \right) . +$$ + +In contrast, Delayed Q-learning Strehl et al. (2006) can only give an upper bound on $\epsilon _ { 1 }$ -suboptimal steps after setting parameter $\epsilon _ { 1 }$ . + +# 5 CONCLUSION + +Infinite-horizon MDP with discounted reward is a setting that is arguably more difficult than other popular settings, such as finite-horizon MDP. Previously, the best samby model-free reinforcement learning algorithms in this setting is $\tilde { \cal O } \big ( { \textstyle \frac { S A } { \epsilon ^ { 4 } ( 1 - \gamma _ { _ { - } } ) ^ { 8 } } } \big )$ y bound achieved, due to Delayed Q-learning Strehl et al. (2006). In this paper, we propose a variant of Q-learning that incorporates upper confidence bound, and show that it has a sample complexity of $\begin{array} { r } { \tilde { \mathcal { O } } \big ( \frac { S A } { \epsilon ^ { 2 } ( 1 - \gamma ) ^ { 7 } } \big ) } \end{array}$ . This matches the best lower bound except in dependence on $1 / ( 1 - \gamma )$ and logarithmic factors. + +# 6 ACKNOWLEDGEMENTS + +The authors thank Chi Jin and Chongjie Zhang for helpful discussions. This work is supported by National Basic Research Program of China (973 Program) (grant no. 2015CB352502), NSFC (61573026), BJNSF (L172037) and Beijing Acedemy of Artificial Intelligence. + +# REFERENCES + +Mohammad Gheshlaghi Azar, Remi Munos, Mohammad Ghavamzadeh, and Hilbert Kappen. Speedy q-learning. In Advances in neural information processing systems, 2011. + +Mohammad Gheshlaghi Azar, Ian Osband, and Rémi Munos. Minimax regret bounds for reinforcement learning. arXiv preprint arXiv:1703.05449, 2017. + +Ronen I. Brafman and Moshe Tennenholtz. R-max - a general polynomial time algorithm for near-optimal reinforcement learning. J. Mach. Learn. Res., 3:213–231, March 2003. ISSN 1532-4435. doi: 10.1162/153244303765208377. URL https://doi.org/10.1162/ 153244303765208377. + +Christoph Dann, Tor Lattimore, and Emma Brunskill. Unifying pac and regret: Uniform pac bounds for episodic reinforcement learning. In Advances in Neural Information Processing Systems, pp. 5713–5723, 2017. + +Eyal Even-Dar and Yishay Mansour. Learning rates for q-learning. Journal of Machine Learning Research, 5(Dec):1–25, 2003. + +Thomas Jaksch, Ronald Ortner, and Peter Auer. Near-optimal regret bounds for reinforcement learning. Journal of Machine Learning Research, 11(Apr):1563–1600, 2010. + +Chi Jin, Zeyuan Allen-Zhu, Sebastien Bubeck, and Michael I Jordan. Is q-learning provably efficient? In Advances in Neural Information Processing Systems, pp. 4864–4874, 2018. + +Sham Machandranath Kakade et al. On the sample complexity of reinforcement learning. PhD thesis, University of London London, England, 2003. + +Tor Lattimore and Marcus Hutter. Pac bounds for discounted mdps. In International Conference on Algorithmic Learning Theory, pp. 320–334. Springer, 2012. + +Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. + +Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pp. 1928–1937, 2016. + +John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015. + +Aaron Sidford, Mengdi Wang, Xian Wu, Lin Yang, and Yinyu Ye. Near-optimal time and sample complexities for solving markov decision processes with a generative model. In Advances in Neural Information Processing Systems, pp. 5186–5196, 2018a. + +Aaron Sidford, Mengdi Wang, Xian Wu, and Yinyu Ye. Variance reduced value iteration and faster algorithms for solving markov decision processes. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 770–787. Society for Industrial and Applied Mathematics, 2018b. + +Alexander L Strehl and Michael L Littman. An analysis of model-based interval estimation for markov decision processes. Journal of Computer and System Sciences, 74(8):1309–1331, 2008. + +Alexander L Strehl, Lihong Li, Eric Wiewiora, John Langford, and Michael L Littman. Pac modelfree reinforcement learning. In Proceedings of the 23rd international conference on Machine learning, pp. 881–888. ACM, 2006. + +István Szita and Csaba Szepesvári. Model-based reinforcement learning with nearly tight exploration complexity bounds. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pp. 1031–1038, 2010. + +# A PROOF OF LEMMA 1 + +Lemma 1. For fixed t and $\eta > 0$ , let $B _ { \eta } ^ { ( t ) }$ be the event that $\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \frac { \eta } { 1 - \gamma } } \end{array}$ in step t. If $\eta > 2 \epsilon _ { 1 }$ , then with probability at least $1 - \delta / 2$ , + +$$ +\sum _ { t = 1 } ^ { t = \infty } I \left[ B _ { \eta } ^ { ( t ) } \right] \leq \frac { S A \ln S A \ln 1 / \delta } { \eta ^ { 2 } ( 1 - \gamma ) ^ { 3 } } \cdot p o l y l o g \left( \frac { 1 } { \epsilon _ { 1 } } , \frac { 1 } { 1 - \gamma } \right) , +$$ + +where $I [ \cdot ]$ is the indicator function. + +Proof. When $\eta > 1$ the lemma holds trivially. Now consider the case that $\eta \leq 1$ . + +Let $\begin{array} { r } { I = \{ t \colon V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \frac { \eta } { 1 - \gamma } \} } \end{array}$ . By lemma 2, with probability $1 - \delta$ + +$$ +\begin{array} { r l } & { \frac { \eta | I | } { 1 - \gamma } \leq \displaystyle \sum _ { t \in I } \left( V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \right) \leq \displaystyle \sum _ { t \in I } \left[ \left( \hat { Q } _ { t } - Q ^ { * } \right) ( s _ { t } , a _ { t } ) \right] } \\ & { \qquad \leq \displaystyle \frac { | I | \epsilon _ { 1 } } { 1 - \gamma } + \mathcal { O } \left( \frac { 1 } { ( 1 - \gamma ) ^ { 5 / 2 } } \sqrt { S A | I | \ell | ( I | ) } + \frac { S A } { ( 1 - \gamma ) ^ { 3 } } \ln | I | \ln \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \right) } \\ & { \qquad \leq \displaystyle \frac { | I | \epsilon _ { 1 } } { 1 - \gamma } + \mathcal { O } \left( \ln \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \cdot \left( \frac { \sqrt { S A | I | \ln \frac { S A | I | } { \delta } } } { ( 1 - \gamma ) ^ { 5 / 2 } } + \frac { S A \ln | I | } { ( 1 - \gamma ) ^ { 3 } } \right) \right) } \\ & { \qquad \leq \displaystyle \frac { | I | \epsilon _ { 1 } } { 1 - \gamma } + \mathcal { O } \left( \sqrt { \ln \frac { 1 } { \delta } } \ln \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \cdot \left( \frac { \sqrt { S A | I | \ln S A | I | } } { ( 1 - \gamma ) ^ { 5 / 2 } } + \frac { S A \ln | I | } { ( 1 - \gamma ) ^ { 3 } } \right) \right) } \end{array} +$$ + +Suppose that $\begin{array} { r } { | I | = \frac { S A k ^ { 2 } } { \eta ^ { 2 } ( 1 - \gamma ) ^ { 3 } } \ln S A } \end{array}$ , for some $k > 1$ . Then it follows that for some constant $C _ { 1 }$ , + +$$ +\begin{array} { r l } & { \frac { \eta | I | } { 1 - \gamma } = \cfrac { k ^ { 2 } S A \ln S A } { ( 1 - \gamma ) ^ { 4 } \eta } \leq 2 \cfrac { ( \eta - \epsilon _ { 1 } ) | I | } { 1 - \gamma } } \\ & { \qquad \leq C _ { 1 } \sqrt { \ln \frac { 1 } { \delta } } \ln \cfrac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \left( \cfrac { \sqrt { S A | I | \ln \left( S A | I | \right) } } { ( 1 - \gamma ) ^ { 5 / 2 } } + \cfrac { S A \ln | I | } { ( 1 - \gamma ) ^ { 3 } } \right) } \\ & { \qquad \leq C _ { 1 } \sqrt { \ln \frac { 1 } { \delta } } \ln \cfrac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \left( \cfrac { S A k } { \eta ( 1 - \gamma ) ^ { 4 } } \sqrt { \ln S A \cdot ( \ln S A + \ln | I | ) } + \cfrac { S A \ln | I | } { ( 1 - \gamma ) ^ { 3 } } \right) . } \end{array} +$$ + +Therefore + +$$ +\begin{array} { r l } & { k ^ { 2 } \ln ( S A ) \le C _ { 1 } \sqrt { \ln \displaystyle \frac { 1 } { \delta } } \ln \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \left( k \left( \ln S A + \ln | I | \right) + \eta ( 1 - \gamma ) \ln | I | \right) } \\ & { \qquad \le k C _ { 1 } \sqrt { \ln \displaystyle \frac { 1 } { \delta } } \ln \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \cdot \left( \ln S A + 2 \ln | I | \right) } \\ & { \qquad \le k C _ { 1 } \sqrt { \ln \displaystyle \frac { 1 } { \delta } } \ln \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \cdot \left( 3 \ln S A + 4 \ln k + 6 \ln \frac { 1 } { \eta ( 1 - \gamma ) } \right) } \\ & { \qquad \le 6 k C _ { 1 } \sqrt { \ln \displaystyle \frac { 1 } { \delta } } \ln ^ { 2 } \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \left( \ln S A + \ln \epsilon k \right) . } \end{array} +$$ + +Let $\begin{array} { r } { C ^ { \prime } = \operatorname* { m a x } \lbrace 2 , 6 C _ { 1 } \sqrt { \ln \frac { 1 } { \delta } } \ln ^ { 2 } \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } \rbrace } \end{array}$ . Then + +$$ +k \leq C ^ { \prime } ( 2 + \ln k ) . +$$ + +If $k \geq 1 0 C ^ { \prime } \ln C ^ { \prime }$ , then + +$$ +\begin{array} { c } { k - C ^ { \prime } \left( 2 + \ln k \right) \geq 8 C ^ { \prime } \ln C ^ { \prime } - ( 2 + \ln 1 0 ) C ^ { \prime } } \\ { \geq 4 C ^ { \prime } \left( 2 \ln C ^ { \prime } - 4 \right) \geq 0 , } \end{array} +$$ + +which means violation of (9). Therefore, since $C ^ { \prime } \geq 2$ + +$$ +k \leq 1 0 C ^ { \prime } \ln C ^ { \prime } \leq 3 6 0 C _ { 1 } ^ { 2 } \operatorname* { m a x } \{ \ln ^ { 4 } { \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } } , 2 0 \ln 2 \} . +$$ + +It immediately follows that + +$$ +\begin{array} { r l r } { { \vert I \vert = \frac { S A k ^ { 2 } } { \eta ^ { 2 } ( 1 - \gamma ) ^ { 3 } } \ln S A } } \\ & { \leq \frac { S A \ln S A } { \eta ^ { 2 } ( 1 - \gamma ) ^ { 5 } } \cdot \ln \frac { 1 } { \delta } \cdot \mathcal { O } ( \ln ^ { 8 } \frac { 1 } { \epsilon _ { 1 } ( 1 - \gamma ) } ) . } \end{array} +$$ + +# B PROOF OF LEMMA 2 + +Lemma 2. For every $( C , w )$ -sequence $( w _ { t } ) _ { t \geq 1 }$ , with probability $1 - \delta / 2$ , the following holds: + +$$ +\sum _ { t \ge 1 } w _ { t } ( \hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \le \frac { C \epsilon _ { 1 } } { 1 - \gamma } + \mathcal { O } \left( \frac { \sqrt { w S A C \ell ( C ) } } { ( 1 - \gamma ) ^ { 2 . 5 } } + \frac { w S A \ln C } { ( 1 - \gamma ) ^ { 3 } } \ln \frac { 1 } { ( 1 - \gamma ) \epsilon _ { 1 } } \right) . +$$ + +where $\begin{array} { r } { \ell ( C ) = \iota ( C ) \ln { \frac { 1 } { ( 1 - \gamma ) \epsilon _ { 1 } } } } \end{array}$ is a log-factor. + +Fact 1. (1) The following statement holds throughout the algorithm, + +$$ +\hat { Q } _ { p + 1 } ( s , a ) \leq Q _ { p + 1 } ( s , a ) . +$$ + +(2) For any $p$ , there exists $p ^ { \prime } \leq p$ such that + +$$ +{ \hat { Q } } _ { p + 1 } ( s , a ) \geq Q _ { p ^ { \prime } + 1 } ( s , a ) . +$$ + +Proof. Both properties are results of the update rule at line 11 of Algorithm 1. + +Before proving lemma 2, we will prove two auxiliary lemmas. + +Lemma 3. The following properties hold for $\alpha _ { t } ^ { i }$ : + +1. $\begin{array} { r } { \sqrt { \frac { 1 } { t } } \leq \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \sqrt { \frac { 1 } { i } } \leq 2 \sqrt { \frac { 1 } { t } } } \end{array}$ for every $t \geq 1 , c > 0$ . + +2. $\begin{array} { r } { \operatorname* { m a x } _ { i \in [ t ] } \alpha _ { t } ^ { i } \leq \frac { 2 H } { t } } \end{array}$ and $\begin{array} { r } { \sum _ { i = 1 } ^ { t } ( \alpha _ { t } ^ { i } ) ^ { 2 } \le \frac { 2 H } { t } } \end{array}$ for every $t \geq 1$ + +3. $\textstyle \sum _ { t = i } ^ { \infty } \alpha _ { t } ^ { i } = 1 + 1 / H$ , for every $i \geq 1$ + +Proof. Recall that + +$$ +\alpha _ { t } = \frac { H + 1 } { H + t } , \quad \alpha _ { t } ^ { 0 } = \prod _ { j = 1 } ^ { t } ( 1 - \alpha _ { j } ) , \quad \alpha _ { t } ^ { i } = \alpha _ { i } \prod _ { j = i + 1 } ^ { t } ( 1 - \alpha _ { j } ) . +$$ + +Properties 1-3 are proven by Jin et al. (2018). Now we prove the last property. + +On the one hand, + +$$ +\sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \sqrt { \frac { \iota ( i ) } { i } } \leq \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \sqrt { \frac { \iota ( t ) } { i } } \leq 2 \sqrt { \frac { \iota ( t ) } { t } } , +$$ + +where the last inequality follows from property 1. + +The left-hand side is proven by induction on $t$ . For the base case, when $t = 1 , \alpha _ { t } ^ { t } = 1$ . For $t \geq 2$ , we have $\alpha _ { t } ^ { i } = ( 1 - \alpha _ { t } ) \dot { \alpha _ { t - 1 } ^ { i } }$ for $1 \leq i \leq t - 1$ . It follows that + +$$ +\sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \sqrt { \frac { \iota ( i ) } { i } } = \alpha _ { t } \sqrt { \frac { \iota ( t ) } { t } } + ( 1 - \alpha _ { t } ) \sum _ { i = 1 } ^ { t - 1 } \alpha _ { t - 1 } ^ { i } \sqrt { \frac { \iota ( i ) } { i } } \geq \alpha _ { t } \sqrt { \frac { \iota ( t ) } { t } } + ( 1 - \alpha _ { t } ) \sqrt { \frac { \iota ( t - 1 ) } { t - 1 } } . +$$ + +Since function $f ( t ) = \iota ( t ) / t$ is monotonically decreasing for $t \geq 1 , c \geq 1$ , we have + +$$ +\alpha _ { t } \sqrt { \frac { \iota ( t ) } { t } } + ( 1 - \alpha _ { t } ) \sqrt { \frac { \iota ( t - 1 ) } { t - 1 } } \geq \alpha _ { t } \sqrt { \frac { \iota ( t ) } { t } } + ( 1 - \alpha _ { t } ) \sqrt { \frac { \iota ( t ) } { t } } \geq \sqrt { \frac { \iota ( t ) } { t } } . +$$ + +Lemma 4. With probability at least $1 - \delta / 2$ , for all $p \geq 0$ and $( s , a )$ -pair, + +$$ +\begin{array} { r l } & { 0 \le ( Q _ { p } - Q ^ { * } ) ( s , a ) \le \displaystyle \frac { \alpha _ { t } ^ { 0 } } { 1 - \gamma } + \displaystyle \sum _ { i = 1 } ^ { t } \gamma \alpha _ { t } ^ { i } ( { \hat { V } } _ { t _ { i } } - V ^ { * } ) ( s _ { t _ { i } + 1 } ) + \beta _ { t } , } \\ & { 0 \le ( { \hat { Q } } _ { p } - Q ^ { * } ) ( s , a ) , } \end{array} +$$ + +where $t = N _ { p } ( s , a ) , t _ { i } = \tau ( s , a , i )$ and $\beta _ { t } = c _ { 3 } \sqrt { H \iota ( t ) / ( ( 1 - \gamma ) ^ { 2 } t ) }$ . + +Proof. Recall that + +$$ +\alpha _ { t } ^ { 0 } = \prod _ { j = 1 } ^ { t } ( 1 - \alpha _ { j } ) , \quad \alpha _ { t } ^ { i } = \alpha _ { i } \prod _ { j = i + 1 } ^ { t } ( 1 - \alpha _ { j } ) . +$$ + +From the update rule, it can be seen that our algorithm maintains the following $Q ( s , a )$ : + +$$ +Q _ { p } ( s , a ) = \alpha _ { t } ^ { 0 } \frac { 1 } { 1 - \gamma } + \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \left[ r ( s , a ) + b _ { i } + \gamma \hat { V } _ { t _ { i } } ( s _ { t _ { i } + 1 } ) \right] . +$$ + +Bellman optimality equation gives: + +$$ +Q ^ { * } ( s , a ) = r ( s , a ) + \gamma \mathbb { P } V ^ { * } ( s , a ) = \alpha _ { t } ^ { 0 } Q ^ { * } ( s , a ) + \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \left[ r ( s , a ) + \gamma \mathbb { P } V ^ { * } ( s , a ) \right] . +$$ + +Subtracting the two equations gives + +$$ +Q _ { p } - Q ^ { * } ) ( s , a ) = \alpha _ { t } ^ { 0 } ( \frac { 1 } { 1 - \gamma } - Q ^ { * } ( s , a ) ) + \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \left[ b _ { i } + \gamma \left( V _ { t _ { i } } - V ^ { * } \right) \left( s _ { t _ { i } + 1 } \right) + \gamma \left( V ^ { * } \left( s _ { t _ { i } + 1 } \right) - \mathbb { P } V ^ { * } \right) \right] , +$$ + +The identity above holds for arbitrary $p , s$ and $a$ . Now fix $s \in S$ , $a \in A$ and $p \in \mathbb N$ . Let $t = N _ { p } ( s , a )$ , $t _ { i } = \tau ( s , a , i )$ . The $t = 0$ case is trivial; we assume $t \geq 1$ below. Now consider an arbitrary fixed $k$ . Define + +$$ +\Delta _ { i } = \left( \alpha _ { k } ^ { i } \cdot I [ t _ { i } < \infty ] \cdot \left( \mathbb { P } V ^ { * } - \mathbb { \hat { P } } _ { t _ { i } } V ^ { * } \right) ( s , a ) \right) +$$ + +Let $F _ { i }$ be the $\sigma$ -Field generated by random variables $( s _ { 1 } , a _ { 1 } , . . . , s _ { t _ { i } } , a _ { t _ { i } } )$ . It can be seen that $\mathbb { E } \left[ \Delta _ { i } | F _ { i } \right] = 0$ , while $\Delta _ { i }$ is measurable in $F _ { i + 1 }$ . Also, since $\textstyle 0 \leq V ^ { \ast } ( s , { \bar { a } } ) \leq { \frac { 1 } { 1 - \gamma } }$ , $\begin{array} { r } { | \Delta _ { i } | \le \frac { 2 } { 1 - \gamma } } \end{array}$ . Therefore, $\Delta _ { i }$ is a martingale difference sequence; by the Azuma-Hoeffding inequality, + +$$ +\mathrm { P r } \left[ \left| \sum _ { i = 1 } ^ { k } \Delta _ { i } \right| > \eta \right] \leq 2 \exp \left\{ - \frac { \eta ^ { 2 } } { 8 \left( 1 - \gamma \right) ^ { - 2 } \sum _ { i = 1 } ^ { k } ( \alpha _ { k } ^ { i } ) ^ { 2 } } \right\} . +$$ + +By choosing $\eta$ , we can show that with probability $1 - \delta / \left[ S A ( k + 1 ) ( k + 2 ) \right]$ , + +$$ +\left| \sum _ { i = 1 } ^ { k } \Delta _ { i } \right| \leq \frac { 2 \sqrt { 2 } } { 1 - \gamma } \cdot \sqrt { \sum _ { i = 1 } ^ { k } ( \alpha _ { k } ^ { i } ) ^ { 2 } \cdot \ln \frac { 2 ( k + 1 ) ( k + 2 ) S A } { \delta } } \leq \frac { c _ { 2 } } { 1 - \gamma } \sqrt { \frac { H \iota ( k ) } { k } } . +$$ + +Here $c _ { 2 } = 4 \sqrt { 2 }$ , $\begin{array} { r } { \iota ( k ) = \ln { \frac { ( k + 1 ) ( k + 2 ) S A } { \delta } } } \end{array}$ . By a union bound for all $k$ , this holds for arbitrary $k > 0$ arbitrary $s \in S , a \in A$ simultaneously with probability + +$$ +1 - \sum _ { s ^ { \prime } \in S , a ^ { \prime } \in A } \sum _ { k = 1 } ^ { \infty } \frac { \delta } { 2 S A ( k + 1 ) ( k + 2 ) } = 1 - \frac { \delta } { 2 } . +$$ + +Therefore, we conclude that (16) holds for the random variable $t = N _ { p } ( s , a )$ and for all $p$ , with probability $1 - \delta / 2$ as well. + +Proof of the right hand side of (13): We also know that $\begin{array} { r } { ( b _ { k } = \frac { c _ { 2 } } { 1 - \gamma } \sqrt { \frac { H \iota ( k ) } { k } } ) } \end{array}$ + +$$ +\frac { c _ { 2 } } { 1 - \gamma } \sqrt { \frac { H \iota ( k ) } { k } } \leq \sum _ { i = 1 } ^ { k } \alpha _ { k } ^ { i } b _ { i } \leq \frac { 2 c _ { 2 } } { 1 - \gamma } \sqrt { \frac { H \iota ( k ) } { k } } . +$$ + +It is implied by (16) that + +$$ +\begin{array} { r l } & { ( Q _ { p } - Q ^ { * } ) ( s , a ) \leq \displaystyle \frac { \alpha _ { t } ^ { 0 } } { 1 - \gamma } + \gamma \left| \displaystyle \sum _ { i = 1 } ^ { t } \Delta _ { i } \right| + \displaystyle \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \left[ \gamma ( \hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \right] } \\ & { \qquad \leq \displaystyle \frac { \alpha _ { t } ^ { 0 } } { 1 - \gamma } + \displaystyle \frac { 3 c _ { 2 } } { 1 - \gamma } \sqrt { \frac { H \iota ( t ) } { t } } + \displaystyle \sum _ { i = 1 } ^ { t } \gamma \alpha _ { t } ^ { i } ( \hat { V } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) } \end{array} +$$ + +$$ +\leq \frac { \alpha _ { t } ^ { 0 } } { 1 - \gamma } + \sum _ { i = 1 } ^ { t } \gamma \alpha _ { t } ^ { i } ( { \hat { V } } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + \beta _ { t } . +$$ + +Note that $\beta _ { t } = c _ { 3 } ( 1 - \gamma ) ^ { - 1 } \sqrt { H \iota ( t ) / t }$ ; $c _ { 3 } = 3 c _ { 2 } = 1 2 { \sqrt { 2 } } .$ . + +Proof of the left hand side of (13): Now, we assume that event that (16) holds. We assert that $Q _ { p } \geq Q ^ { * }$ for all $( s , a )$ and $p \leq p ^ { \prime }$ . This assertion is obviously true when $p ^ { \prime } = 0$ . Then + +$$ +\begin{array} { c l } { ( Q _ { p } - Q ^ { * } ) ( s , a ) \geq - \gamma \displaystyle \left| \displaystyle \sum _ { i = 1 } ^ { t } \Delta _ { i } \right| + \displaystyle \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } \left[ \gamma ( \hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \right] } \\ { \geq \displaystyle \sum _ { i = 1 } ^ { t } \alpha _ { t } ^ { i } b _ { i } - \gamma \displaystyle \left| \displaystyle \sum _ { i = 1 } ^ { t } \Delta _ { i } \right| \geq 0 . } \end{array} +$$ + +Therefore the assertion holds for $p ^ { \prime } + 1$ as well. By induction, it holds for all $p$ . + +We now see that (13) holds for probability $1 - \delta / 2$ for all $p , s , a$ . Since $\hat { Q } _ { p } ( s , a )$ is always greater than $Q _ { p ^ { \prime } } ( s , a )$ for some $p ^ { \prime } \leq p$ , we know that $\hat { Q } _ { p } ( s , a ) \geq Q _ { p ^ { \prime } } ( s , a ) \geq Q ^ { * } ( s , a )$ , thus proving (14). + +We now give a proof for lemma 2. Recall the definition for a $( C , w )$ -sequence. A sequence $( w _ { t } ) _ { t \geq 1 }$ is said to be a $( C , w )$ -sequence for $C , w > 0$ , if $0 \leq w _ { t } \leq w$ for all $t \geq 1$ , and $\textstyle \sum _ { t \geq 1 } w _ { t } \leq C$ . + +Proof. Let $n _ { t } = N _ { t } ( s _ { t } , a _ { t } )$ for simplicity; we have + +$$ +\begin{array} { r l } { { \sum _ { t \geq 1 } w _ { t } ( \hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } } \\ & { \leq \sum _ { t \geq 1 } w _ { t } ( Q _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\ & { \leq \sum _ { t \geq 1 } w _ { t } [ \cfrac { \alpha _ { n _ { t } } ^ { 0 } } { 1 - \gamma } + \beta _ { n _ { t } } + \gamma \sum _ { i = 1 } ^ { n _ { t } } \alpha _ { n _ { t } } ^ { i } ( \hat { V } _ { \tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } ) ( s _ { \tau ( s _ { t } , a _ { t } , i ) + 1 } ) ] } \end{array} +$$ + +The last inequality is due to lemma 4. Note that $\alpha _ { n _ { t } } ^ { 0 } = \mathbb { I } [ n _ { t } = 0 ]$ , the first term in the summation can be bounded by, + +$$ +\sum _ { t \geq 1 } w _ { t } \frac { \alpha _ { n _ { t } } ^ { 0 } } { 1 - \gamma } \leq \frac { S A w } { 1 - \gamma } . +$$ + +For the second term, define $u ( s , a ) = \operatorname* { s u p } _ { t } N _ { t } ( s , a )$ .2 It follows that, + +$$ +\begin{array} { l } { \displaystyle \sum _ { t \geq 1 } w _ { t } \beta _ { n _ { t } } = \sum _ { s , o } ^ { u ( s , o ) } w _ { \tau ( s , a , i ) } \beta _ { i } } \\ { \displaystyle \qquad \leq \sum _ { s , a } ( 1 - \gamma ) ^ { - 1 } c _ { 3 } \sum _ { i = 1 } ^ { C _ { s , a } / w } \sqrt { \frac { H \iota ( i ) } { i } } w } \\ { \displaystyle \qquad \leq 2 \sum _ { s , a } ( 1 - \gamma ) ^ { - 1 } c _ { 3 } \sqrt { \iota ( C ) H C _ { s , a } w } } \\ { \displaystyle \qquad \leq 2 c _ { 3 } ( 1 - \gamma ) ^ { - 1 } \sqrt { w S A H C \iota ( C ) } . } \end{array} +$$ + +Where $\begin{array} { r } { C _ { s , a } = \sum _ { t \geq 1 , ( s _ { t } , a _ { t } ) = ( s , a ) } w _ { t } } \end{array}$ . Inequality (19) follows from rearrangement inequality, since $\iota ( x ) / x$ is monotonically decreasing. Inequality (21) follows from Jensen’s inequality. + +For the third term of the summation, we have + +$$ +\begin{array} { c } { { \displaystyle \sum _ { t \ge 1 } w _ { t } \sum _ { i = 1 } ^ { n _ { t } } \alpha _ { n _ { t } } ^ { i } \left( \hat { V } _ { \tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } \right) \left( s _ { \tau ( s _ { t } , a _ { t } , i ) + 1 } \right) } } \\ { { \displaystyle \le \sum _ { t ^ { \prime } \ge 1 } \left( \hat { V } _ { t ^ { \prime } } - V ^ { * } \right) \left( s _ { t ^ { \prime } + 1 } \right) \left( \sum _ { t = t ^ { \prime } + 1 } ^ { \infty } \alpha _ { n _ { t } } ^ { n _ { t ^ { \prime } } } w _ { t } \right) . } } \end{array} +$$ + +Define + +$$ +w _ { t ^ { \prime } + 1 } ^ { \prime } = \left( \sum _ { \stackrel { t = t ^ { \prime } + 1 } { ( s _ { t } , a _ { t } ) = ( s _ { t } ^ { \prime } , a _ { t } ^ { \prime } ) } } ^ { \infty } \alpha _ { n _ { t } } ^ { n _ { t ^ { \prime } } } w _ { t } \right) . +$$ + +We claim that $w _ { t + 1 } ^ { \prime }$ is a $\textstyle ( C , ( 1 + { \frac { 1 } { H } } ) w )$ -sequence. We now prove this claim. By lemma 3, for any $t ^ { \prime } \geq 0$ , + +$$ +w _ { t ^ { \prime } + 1 } ^ { \prime } \leq w \sum _ { j = n _ { t ^ { \prime } } } ^ { \infty } \alpha _ { j } ^ { n _ { t ^ { \prime } } } = ( 1 + 1 / H ) w . +$$ + +By $\textstyle \sum _ { j = 0 } ^ { i } \alpha _ { i } ^ { j } = 1$ , we have $\begin{array} { r } { \sum _ { t ^ { \prime } \geq 1 } w _ { t ^ { \prime } + 1 } ^ { \prime } \leq \sum _ { t \geq 1 } w _ { t } \leq C . } \end{array}$ . This proves the assertion. It follows from (22) that + +$$ +\begin{array} { r l } & { \quad \underset { t \geq 1 } { \sum \sigma } u _ { t + 1 } ^ { \prime } \left( \hat { V } _ { t } - V ^ { * } \right) \left( s _ { t + 1 } \right) } \\ & { = \underset { t \geq 1 } { \sum \sigma } u _ { t + 1 } ^ { \prime } \left( \hat { V } _ { t + 1 } - V ^ { * } \right) \left( s _ { t + 1 } \right) + \underset { t \geq 1 } { \sum \sigma } u _ { t + 1 } ^ { \prime } \left( \hat { V } _ { t } - \hat { V } _ { t + 1 } \right) \left( s _ { t + 1 } \right) } \\ & { \quad \leq \underset { t \geq 1 } { \sum \sigma } u _ { t + 1 } ^ { \prime } \left( \hat { V } _ { t + 1 } - V ^ { * } \right) \left( s _ { t + 1 } \right) + \underset { t \geq 1 } { \sum \sigma } u _ { t + 1 } ^ { \prime } \left( 2 \alpha _ { n _ { t } + 1 } \frac { 1 } { 1 - \gamma } \right) } \\ & { \quad \leq \underset { t \geq 1 } { \sum \sigma } w _ { t + 1 } ^ { \prime } \left( \hat { V } _ { t + 1 } - V ^ { * } \right) \left( s _ { t + 1 } \right) + \mathcal { O } \left( \frac { w S A H } { 1 - \gamma } \ln { C } \right) } \\ & { \quad \leq \underset { t \geq 1 } { \sum \sigma } w _ { t + 1 } ^ { \prime } \left( \hat { Q } _ { t + 1 } - Q ^ { * } \right) \left( s _ { t + 1 } , a _ { t + 1 } \right) + \mathcal { O } \left( \frac { w S A H } { 1 - \gamma } \ln { C } \right) } \end{array} +$$ + +Inequality (25) comes from the update rule of our algorithm. Inequality (26) comes from the fact that $\dot { \alpha } _ { t } = ( H + 1 ) / ( H + t ) \overset { \cdot } { \leq } H / t$ and Jensen’s Inequality. More specifically, let $C _ { s , a } ^ { \prime } =$ $\scriptstyle \sum _ { t \geq 1 , ( s _ { t } , a _ { t } = s , a } w _ { t + 1 } ^ { \prime }$ , $w ^ { \prime } = w ( 1 + 1 / H )$ . Then + +$$ +\sum _ { t \geq 1 } w _ { t + 1 } ^ { \prime } \alpha _ { n _ { t } + 1 } \leq \sum _ { s , a } \sum _ { n = 1 } ^ { C _ { s , a } ^ { \prime } / w ^ { \prime } } w ^ { \prime } \frac { H } { n } \leq \sum _ { s , a } H w ^ { \prime } \ln ( C _ { s , a } ^ { \prime } / w ) \leq 2 S A H w \ln C . +$$ + +Putting (18), (21) and (27) together, we have, + +$$ +\begin{array} { r l } & { \displaystyle \sum _ { t \geq 1 } w _ { t } ( \hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\ & { \leq 2 c _ { 3 } \frac { \sqrt { w S A H C \iota ( C ) } } { 1 - \gamma } + \mathcal { O } \left( \frac { w S A H } { 1 - \gamma } \ln C \right) + \gamma \displaystyle \sum _ { t \geq 1 } w _ { t + 1 } ^ { \prime } \left( \hat { Q } _ { t + 1 } - Q ^ { * } \right) ( s _ { t + 1 } , a _ { t + 1 } ) . } \end{array} +$$ + +Observe that the third term is another weighted sum with the same form as (17). Therefore, we can unroll this term repetitively with changing weight sequences.Suppose that our original weight sequence is also denoted by $\{ \bar { w _ { t } ^ { ( 0 ) } } \} _ { t \ge 1 }$ , while $\{ w _ { t } ^ { ( k ) } \} _ { t \ge 1 }$ denotes the weight sequence after unrolling for $k$ times. Let $w ^ { ( k ) }$ be $w \cdot \left( 1 + 1 / H \right) ^ { k }$ . Then we can see that $\{ w _ { t } ^ { ( k ) } \} _ { t \ge 1 }$ is a $( C , w ^ { ( k ) } )$ -sequence. Suppose that we unroll for $H$ times. Then + +$$ +\begin{array} { r l } & { \displaystyle \sum _ { t \geq 1 } w _ { t } ( \hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\ & { \leq 2 c _ { 3 } \frac { \sqrt { w ^ { ( H ) } S A H C \iota ( C ) } } { ( 1 - \gamma ) ^ { 2 } } + \mathcal { O } \left( \frac { w ^ { ( H ) } S A H } { ( 1 - \gamma ) ^ { 2 } } \ln C \right) + \gamma ^ { H } \displaystyle \sum _ { t \geq 1 } w _ { t } ^ { ( H ) } \left( \hat { Q } _ { t } - Q ^ { * } \right) ( s _ { t } , a _ { t } ) } \\ & { \leq 2 c _ { 3 } \frac { \sqrt { w ^ { ( H ) } S A H C \iota ( C ) } } { ( 1 - \gamma ) ^ { 2 } } + \mathcal { O } \left( \frac { w ^ { ( H ) } S A H } { ( 1 - \gamma ) ^ { 2 } } \ln C \right) + \gamma ^ { H } \frac { C } { 1 - \gamma } . } \end{array} +$$ + +We set $\begin{array} { r } { H = \frac { \ln 1 / ( ( 1 - \gamma ) \epsilon _ { 1 } ) } { \ln 1 / \gamma } \leq \frac { \ln 1 / ( ( 1 - \gamma ) \epsilon _ { 1 } ) } { 1 - \gamma } , } \end{array}$ . It follows that $w ^ { ( H ) } = ( 1 + 1 / H ) ^ { H } w ^ { ( 0 ) } \leq e w ^ { ( 0 ) }$ , and that γH C1−γ $\gamma ^ { H } \frac { C } { 1 - \gamma } \leq C \epsilon _ { 1 }$ . Also, let $\ell ( C ) = \iota ( C ) \ln ( ( 1 - \gamma ) ^ { - 1 } \epsilon _ { 1 } ^ { - 1 } )$ . Therefore, + +$$ +\sum _ { t \geq 1 } w _ { t } ( \hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \leq \frac { C \epsilon _ { 1 } } { 1 - \gamma } + \mathcal { O } \left( \frac { \sqrt { w S A C \ell ( C ) } } { ( 1 - \gamma ) ^ { 2 . 5 } } + \frac { w S A } { ( 1 - \gamma ) ^ { 3 } } \ln C \ln \frac { 1 } { ( 1 - \gamma ) \epsilon _ { 1 } } \right) . +$$ + +# C EXTENSION TO OTHER SETTINGS + +First we define a mapping from a finite horizon MDP to an infinite horizon MDP so that our algorithm can be applied. For an arbitrary finite horizon MDP $\mathcal { M } = ( S , A , H , r _ { h } ( s _ { \perp } a ) , p _ { h } ( s ^ { \prime } \mid s , a ) )$ where $H$ is the length of episode, the corresponding infinite horizon MDP $\mathcal { M } = ( S , A , \gamma , \bar { r } ( \bar { s } , \bar { a } ) , \bar { p } ( \bar { s } ^ { \prime } \mid \bar { s } , \bar { a } ) )$ is defined as, + +$\begin{array} { l } { { \bullet \ { \bar { S } } = S \times H , { \bar { A } } = A ; } } \\ { { \bullet \ \gamma = ( 1 - 1 / H ) ; } } \end{array}$ • for a state $s$ at step $h$ , let $\bar { s } _ { s , h }$ be the corresponding state. For any action $a$ and next state $s ^ { \prime }$ , define $\bar { r } ( \bar { s } _ { s , h } , a ) = \gamma ^ { H - h + 1 } r _ { h } ( s , a )$ and $\bar { p } ( \bar { s } _ { s ^ { \prime } , h + 1 } \mid \bar { s } _ { s , h } , a ) = p _ { h } ( s ^ { \prime } \mid s , h )$ . And for $h = H$ , set $\bar { r } ( \bar { s } _ { s , h } , a ) = 0$ and $\bar { p } ( \bar { s } _ { s ^ { \prime } , 1 } \mid \bar { s } _ { s , h } , a ) = I [ s ^ { \prime } = s _ { 1 } ]$ for a fixed starting state $s _ { 1 }$ . + +Let $\bar { V } _ { t }$ be the value function in $\bar { \mathcal { M } }$ at time $t$ and $V _ { h } ^ { k }$ the value function in $\mathcal { M }$ at episode $k$ , step $h$ . It follows that $\begin{array} { r } { \bar { V } ^ { \ast } ( \bar { s } _ { s _ { 1 } , 1 } ) = \frac { \gamma ^ { H } } { 1 - \gamma ^ { H } } V _ { 1 } ^ { \ast } ( s _ { 1 } ) } \end{array}$ . And the policy mapping is defined as $\pi _ { h } ( s ) = \bar { \pi } ( \bar { s } _ { s , h } )$ for policy $\bar { \pi }$ in $\bar { \mathcal { M } }$ . Value functions in MDP $\mathcal { M }$ and $\bar { \mathcal { M } }$ are closely related in a sense that, any $\epsilon$ -optimal policy $\bar { \pi }$ of $\bar { \mathcal { M } }$ corresponding to an $( \epsilon / \gamma ^ { H } )$ -optimal policy $\pi$ in $\mathcal { M }$ (see section C.1 for proof). Note that here $\gamma ^ { H } = ( 1 - \dot { 1 } / H ) ^ { \tilde { H } } = \mathcal { O } ( \dot { 1 } )$ is a constant. + +For any visited a $\epsilon > 0$ g our algorithm on times, and at most $\bar { M }$ for of $\tilde { \mathcal O } \big ( \frac { 3 S A H ^ { 9 } } { \epsilon ^ { 2 } } \big )$ tt me steps, the starting state -optimal. If we select the p $s _ { 1 }$ iscy $\tilde { \mathcal O } \big ( { \textstyle { \frac { 3 S A H ^ { 8 } } { \epsilon ^ { 2 } } } } \big )$ $1 / 3$ $\epsilon$ uniformly randomly from the policy $\pi ^ { t H + 1 }$ for $0 \leq t < T / H$ , with probability at least $2 / 3$ we can get an $\epsilon$ -optimal policy. Therefore the PAC sample complexity is $\tilde { \mathcal { O } } \left( \epsilon ^ { - 2 } \right)$ after hiding $S , A , H$ terms. + +On the other hand, we want to show that for any $K$ episodes, + +$$ +\mathrm { R e g r e t } ( T ) = \sum _ { k = 1 } ^ { T / H } \left[ V ^ { * } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } ) \right] \propto T ^ { 1 / 2 } . +$$ + +The reason why our algorithm can have a better reduction from regret to PAC is that, after choosing $\epsilon _ { 1 }$ , it follows from the argument of theorem 1 that for all $\epsilon _ { 2 } > \tilde { \mathcal { O } } ( \epsilon _ { 1 } / ( 1 - \gamma ) )$ , the number of $\epsilon _ { 2 }$ -suboptimal steps is bounded by + +$$ +\mathcal { O } \left( \frac { S A \ln S A \ln 1 / \delta } { \epsilon _ { 2 } ^ { 2 } ( 1 - \gamma ) ^ { 7 } } \mathrm { p o l y l o g } \left( \frac { 1 } { \epsilon _ { 1 } } , \frac { 1 } { 1 - \gamma } \right) \right) +$$ + +with probability $1 - \delta$ . In contrast, delayed Q-learning can only give an upper bound on $\epsilon _ { 1 }$ -suboptimal steps after setting parameter $\epsilon _ { 1 }$ . + +Formally, let $X _ { k } = V ^ { \ast } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } )$ be the regret of $k$ -th episode. For any $T$ , set $\epsilon = \sqrt { S A / T }$ and $\epsilon _ { 2 } = \tilde { \mathcal { O } } ( \epsilon _ { 1 } / ( 1 - \gamma ) )$ . Let $\begin{array} { r } { M = \lceil \log _ { 2 } \frac { 1 } { \epsilon _ { 2 } ( 1 - \gamma ) } \rceil } \end{array}$ . It follows that, + +$$ +\begin{array} { r l r } { { \operatorname { R e g r e t } ( T ) \le T \epsilon _ { 2 } + \sum _ { i = 1 } ^ { M } ( | k : \{ X _ { k } \ge \epsilon _ { 2 } \cdot 2 ^ { i - 1 } \} | ) \epsilon _ { 2 } \cdot 2 ^ { i } } } \\ & { } & \\ & { } & { \le \tilde { \mathcal { O } } ( T \epsilon _ { 2 } + \sum _ { i = 1 } ^ { M } \frac { S A \ln 1 / \delta } { \epsilon _ { 2 } \cdot 2 ^ { i - 2 } } ) } \\ & { } & { \le \tilde { \mathcal { O } } ( \sqrt { S A T } \ln 1 / \delta ) } \end{array} +$$ + +with probability $1 - \delta$ . Note that the $\tilde { \mathcal { O } }$ notation hides the poly $( 1 / ( 1 - \gamma ) , \log { 1 / \epsilon _ { 1 } } )$ which is, by our reduction, poly $( H , \log T , \log S , \log A )$ . + +# C.1 CONNECTION BETWEEN VALUE FUNCTIONS + +Recall that our MDP mapping from $\begin{array} { r c l } { \mathcal { M } } & { = } & { ( S , A , H , r _ { h } ( s , a ) , p _ { h } ( s ^ { \prime } \mathrm { ~ ~ \chi ~ } | \mathrm { ~ ~ \chi ~ } s , a ) ) } \end{array}$ to $\begin{array} { r l } { \bar { \mathcal { M } } } & { { } = } \end{array}$ $( \bar { S } , \bar { A } , \gamma , \bar { r } ( \bar { s } , \bar { a } ) , \bar { p } ( \bar { s } ^ { \prime } \mid \bar { s } , \bar { a } ) )$ is defined as, + +• $\bar { S } = S \times H , \bar { A } = A$ ; +• $\gamma = ( 1 - 1 / H )$ ; +• for a state $s$ at step $h$ , let $\bar { s } _ { s , h }$ be the corresponding state. For any action $a$ and next state $s ^ { \prime }$ , define $\bar { r } ( \bar { s } _ { s , h } , a ) = \gamma ^ { H - h + 1 } r _ { h } ( s , a )$ and $\bar { p } ( \bar { s } _ { s ^ { \prime } , h + 1 } \mid \bar { s } _ { s , h } , a ) = p _ { h } ( s , h )$ . And for $h = H$ , set $\bar { r } ( \bar { s } _ { s , h } , a ) = 0$ and $\bar { p } ( \bar { s } _ { s ^ { \prime } , 1 } \mid \bar { s } _ { s , h } , a ) = I [ s ^ { \prime } = s _ { 1 } ]$ for a fixed starting state $s _ { 1 }$ . + +For a trajectory $\left\{ { \left( { \bar { s } } _ { s _ { 1 } , 1 } , { \bar { a } } _ { 1 } \right) } , { \left( { \bar { s } } _ { s _ { 2 } , 2 } , { \bar { a } } _ { 2 } \right) } , \cdot \cdot \cdot \right\}$ in $\bar { \mathcal { M } }$ , let $\{ ( s _ { 1 } , a _ { 1 } ) , ( s _ { 2 } , a _ { 2 } ) , \cdot \cdot \cdot \}$ be the corresponding trajectory in $\mathcal { M }$ . Note that $\mathcal { M }$ has a unique fixed starting state $s _ { 1 }$ , which means that $s _ { t H + 1 } = s _ { 1 }$ for all $t \geq 0$ . Denote the corresponding policy of $\bar { \pi } ^ { t }$ as $\pi ^ { t }$ (may be non-stationary), then we have + +$$ +\begin{array} { r l } & { \bar { \gamma } ^ { \pi ^ { t } } ( \bar { s } _ { s _ { 1 } , 1 } ) = \mathbb { E } \left[ \bar { r } ( \bar { s } _ { s _ { 1 } , 1 } , \bar { a } _ { 1 } ) + \gamma \bar { r } ( \bar { s } _ { s _ { 2 } , 2 } , \bar { a } _ { 2 } ) + \cdot \cdot \cdot + \gamma ^ { H - 1 } \bar { r } ( \bar { s } _ { s _ { H - 1 } , H - 1 } , \bar { a } _ { H - 1 } ) + \gamma ^ { H } \bar { V } ^ { \pi _ { t + H - 1 } } ( \bar { s } _ { s _ { H + 1 } , 1 } ) \right] } \\ & { \phantom { \frac { 1 } { 1 } } = \gamma ^ { H } \mathbb { E } \left[ r _ { 1 } ( s _ { 1 } , a _ { 1 } ) + r _ { 2 } ( s _ { 2 } , a _ { 2 } ) + \cdot \cdot \cdot + r _ { H - 1 } ( s _ { H - 1 } , a _ { H - 1 } ) + \bar { V } ^ { \pi _ { t + H } } ( \bar { s } _ { s _ { H + 1 } , 1 } ) \right] } \\ & { \phantom { \frac { 1 } { 1 } } = \gamma ^ { H } V ^ { \pi ^ { t } } ( s _ { 1 } ) + \gamma ^ { H } \bar { V } ^ { \pi + H } ( \bar { s } _ { s _ { 1 } , 1 } ) . } \end{array} +$$ + +Then for a stationary policy $\bar { \pi }$ , we can conclude $\begin{array} { r } { \bar { V } ^ { \bar { \pi } } ( \bar { s } _ { s _ { 1 } , 1 } ) = \frac { \gamma ^ { H } } { 1 - \gamma ^ { H } } V ^ { \pi } ( s _ { 1 } ) } \end{array}$ . Since the optimal policy $\bar { \pi } ^ { * }$ is stationary, we have $\begin{array} { r } { \bar { V } ^ { \ast } ( \bar { s } _ { s _ { 1 } , 1 } ) = \frac { \gamma ^ { H } } { 1 - \gamma ^ { H } } V ^ { \ast } ( s _ { 1 } ) } \end{array}$ . + +By definition, $\bar { \pi }$ is $\epsilon$ -optimal at time step $t$ means that + +$$ +\begin{array} { r } { \bar { V } ^ { \bar { \pi } ^ { t } } ( \bar { s } _ { s _ { 1 } , 1 } ) \geq \bar { V } ^ { * } ( \bar { s } _ { s _ { 1 } , 1 } ) - \epsilon . } \end{array} +$$ + +It follows that + +$$ +\gamma ^ { H } V ^ { \pi ^ { t } } ( s _ { 1 } ) + \gamma ^ { H } \bar { V } ^ { \pi _ { t + H } } ( \bar { s } _ { s _ { 1 } , 1 } ) = \bar { V } ^ { \bar { \pi } } ( \bar { s } _ { s _ { 1 } , 1 } ) \geq \bar { V } ^ { * } ( \bar { s } _ { s _ { 1 } , 1 } ) - \epsilon , +$$ + +hence + +$$ +^ { H } V ^ { \pi ^ { \varepsilon } } ( s _ { 1 } ) \geq ( 1 - \gamma ^ { H } ) \bar { V } ^ { * } ( \bar { s } _ { s _ { 1 } , 1 } ) + \gamma ^ { H } ( \bar { V } ^ { * } ( \bar { s } _ { s _ { 1 } , 1 } ) - \bar { V } ^ { \pi _ { t + H } } ( \bar { s } _ { s _ { 1 } , 1 } ) ) - \epsilon \geq ( 1 - \gamma ^ { H } ) \bar { V } ^ { * } ( \bar { s } _ { s _ { 1 } , 1 } ) - \epsilon . +$$ + +Therefore we have + +$$ +{ V ^ { \pi } } ^ { t } ( s _ { 1 } ) \ge \frac { 1 - \gamma ^ { H } } { \gamma ^ { H } } \bar { V } ^ { * } ( \bar { s } _ { s 1 , 1 } ) - \epsilon / \gamma ^ { H } = V ^ { * } ( s _ { 1 } ) - \epsilon / \gamma ^ { H } , +$$ + +which means that $\pi ^ { t }$ is an $( \epsilon / \gamma ^ { H } )$ -optimal policy. + +# D A HARD INSTANCE FOR DELAYED Q-LEARNING + +In this section, we prove Theorem 2 regarding the performance of Delayed Q-learning. + +Theorem 2. There exists a family of MDPs with constant $S$ and $A$ , in which with probability $1 - \delta$ , Delayed $Q$ -learning incurs sample complexity of exploration of $\Omega \left( \frac { \epsilon ^ { - 3 } } { \ln ( 1 / \delta ) } \right)$ , assuming that $\ln ( 1 / \delta ) < \epsilon ^ { - 2 }$ . + +![](images/d5ea7466679255528080a8e623d46ab5efe02c60983c67939e2c04d8974b3167.jpg) +Figure 1: The MDP family. Actions are denoted by arrows. Actions with red color have reward 1, and reward 0 otherwise. + +Proof. For each $0 < \epsilon < \frac { 1 } { 1 0 }$ , consider the following MDP (see also Fig. 1): state space is ${ \boldsymbol { s } } =$ $\{ a , b , c \}$ while action set is $\overset { \vartriangle } { \mathcal { A } } = \{ \ v { x } , \ v { y } \}$ ; transition probabilities are $P ( b | a , y ) = 1 - 1 0 \epsilon$ , $P ( c | a , y ) =$ 10, $P ( b | a , x ) = 1$ , $P ( a | b , \cdot ) = P ( a | c , \cdot ) = 1$ . Rewards are all 1, except $R ( c , \cdot ) = 0$ . + +Assume that Delayed Q-learning is called for this MDP starting from state $a$ , with discount $\begin{array} { r } { \gamma > \frac { 1 } { 2 } } \end{array}$ and precision set as $\epsilon$ . Denote the $Q$ value maintained by the algorithm by $\hat { Q }$ . Without loss of generality, assume that the initial tie-breaking favors action $y$ when comparing $\hat { Q } ( a , x )$ and $\hat { Q } ( a , y )$ . In that case, unless $\hat { Q } ( a , y )$ is updated, the agent will always choose $y$ in state $a$ . Since $Q ( a , x ) - Q ( a , y ) = 1 0 \epsilon \gamma > \epsilon$ for any policy, choosing $y$ at state $a$ implies that the timestep is not $\epsilon$ -optimal. In other words, sample complexity for exploration is at least the number of times the agent visits $a$ before the first update of $\hat { Q } ( a , y )$ . + +In the Delayed Q-learning algorithm, $\hat { Q } ( \cdot , \cdot )$ are initialized to $1 / ( 1 - \gamma )$ . Therefore, $\hat { Q } ( a , y )$ could only be updated if max $\hat { Q } ( c , \cdot )$ is updated (and becomes smaller than $1 / ( 1 - \gamma ) )$ . According to the algorithm, this can only happen if $c$ is visited $\begin{array} { r } { m = \Omega \left( \frac { 1 } { \epsilon ^ { 2 } } \right) } \end{array}$ times. + +However, each time the agent visits $a$ , there is less than $1 0 \epsilon$ probability of transiting to $c$ . Let $t _ { 0 } = m / ( 1 0 \epsilon C )$ , where $\begin{array} { r } { C ^ { ' } = 3 \ln \frac { 1 } { \delta } + 1 } \end{array}$ . $\delta$ is chosen such that $C \leq m$ . In the first $2 t _ { 0 }$ timesteps, $a$ will be visited $t _ { 0 }$ times. By Chernoff’s bound, with probability $1 - \delta$ , state $c$ will be visited less than $m$ times. In that case, $\hat { Q } ( a , y )$ will not be updated in the first $2 t _ { 0 }$ timesteps. Therefore, with probability $1 - \delta$ , sample complexity of exploration is at least + +$$ +t _ { 0 } = \Omega \left( { \frac { 1 } { \epsilon ^ { 3 } \left( \ln { 1 / \delta } \right) } } \right) . +$$ + +When $\ln ( 1 / \delta ) < \epsilon ^ { - 2 }$ , it can be seen that $\begin{array} { r } { C = 3 \ln \frac { 1 } { \delta } + 1 < \frac { 4 } { \epsilon ^ { 2 } } < m } \end{array}$ \ No newline at end of file diff --git a/parse/train/BkglSTNFDB/BkglSTNFDB_content_list.json b/parse/train/BkglSTNFDB/BkglSTNFDB_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..47248d499e0138decfd6a6face1b216abc5e7766 --- /dev/null +++ b/parse/train/BkglSTNFDB/BkglSTNFDB_content_list.json @@ -0,0 +1,3148 @@ +[ + { + "type": "text", + "text": "Q-LEARNING WITH UCB EXPLORATION IS SAMPLE EFFICIENT FOR INFINITE-HORIZON MDP ", + "text_level": 1, + "bbox": [ + 176, + 98, + 823, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kefan Dong\\*, Yuanhao Wang∗ \nInstitute for Interdisciplinary Information Sciences, Tsinghua University \n{dkf16,yuanhao-16}@mails.tsinghua.edu. \nXiaoyu Chen \nKey Laboratory of Machine Perception, MOE, School of EECS, \nPeking University \ncxy30@pku.edu.cn \nLiwei Wang \nKey Laboratory of Machine Perception, MOE, School of EECS \nCenter for Data Science, Peking University \nwanglw@cis.pku.edu.cn ", + "bbox": [ + 183, + 169, + 555, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 184, + 247, + 604, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 183, + 324, + 601, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 416, + 544, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A fundamental question in reinforcement learning is whether model-free algorithms are sample efficient. Recently, Jin et al. (2018) proposed a Q-learning algorithm with UCB exploration policy, and proved it has nearly optimal regret bound for finite-horizon episodic MDP. In this paper, we adapt Q-learning with UCB-exploration bonus to infinite-horizon MDP with discounted rewards without accessing a generative model. We show that the sample complexity of exploration of our algorithm is bounded by $\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}$ . This improves the previously best known result of $\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}$ in this setting achieved by delayed Q-learning (Strehl et al., 2006), and matches the lower bound in terms of $\\epsilon$ as well as and up to logarithmic factors. ", + "bbox": [ + 233, + 450, + 766, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 632, + 336, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of reinforcement learning (RL) is to construct efficient algorithms that learn and plan in sequential decision making tasks when the underlying system dynamics are unknown. A typical model in RL is Markov Decision Process (MDP). At each time step, the environment is in a state $s$ . The agent takes an action $a$ , obtain a reward $r$ , and then the environment transits to another state. In reinforcement learning, the transition probability distribution is unknown. The algorithm needs to learn the transition dynamics of MDP, while aiming to maximize the cumulative reward. This poses the exploration-exploitation dilemma: whether to act to gain new information (explore) or to act consistently with past experience to maximize reward (exploit). ", + "bbox": [ + 173, + 665, + 825, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Theoretical analyses of reinforcement learning fall into two broad categories: those assuming a simulator (a.k.a. generative model), and those without a simulator. In the first category, the algorithm is allowed to query the outcome of any state action pair from an oracle. The emphasis is on the number of calls needed to estimate the $Q$ value or to output a near-optimal policy. There has been extensive research in literature following this line of research, the majority of which focuses on discounted infinite horizon MDPs (Azar et al., 2011; Even-Dar & Mansour, 2003; Sidford et al., 2018b). The current results have achieved near-optimal time and sample complexities (Sidford et al., 2018b;a). ", + "bbox": [ + 174, + 784, + 825, + 895 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Without a simulator, there is a dichotomy between finite-horizon and infinite-horizon settings. In finite-horizon settings, there are straightforward definitions for both regret and sample complexity; the latter is defined as the number of samples needed before the policy becomes near optimal. In this setting, extensive research in the past decade (Jin et al., 2018; Azar et al., 2017; Jaksch et al., 2010; Dann et al., 2017) has achieved great progress, and established nearly-tight bounds for both regret and sample complexity. ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The infinite-horizon setting is a very different matter. First of all, the performance measure cannot be a straightforward extension of the sample complexity defined above (See Strehl & Littman (2008) for detailed discussion). Instead, the measure of sample efficiency we adopt is the so-called sample complexity of exploration (Kakade et al., 2003), which is also a widely-accepted definition. This measure counts the number of times that the algorithm “makes mistakes” along the whole trajectory. See also (Strehl & Littman, 2008) for further discussions regarding this issue. ", + "bbox": [ + 174, + 194, + 825, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several model based algorithms have been proposed for infinite horizon MDP, for example Rmax (Brafman & Tennenholtz, 2003), MoRmax (Szita & Szepesvári, 2010) and UCRL- $\\gamma$ (Lattimore & Hutter, 2012). It is noteworthy that there still exists a considerable gap between the state-of-the-art algorithm and the theoretical lower bound (Lattimore & Hutter, 2012) regarding $1 / ( 1 - \\gamma )$ factor. ", + "bbox": [ + 174, + 285, + 825, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though model-based algorithms have been proved to be sample efficient in various MDP settings, most state-of-the-art RL algorithms are developed in the model-free paradigm (Schulman et al., 2015; Mnih et al., 2013; 2016). Model-free algorithms are more flexible and require less space, which have achieved remarkable performance on benchmarks such as Atari games and simulated robot control problems. ", + "bbox": [ + 174, + 347, + 825, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For infinite horizon MDPs without access to simulator, the best model-free algorithm has a sample complexity of exploration $\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}$ , achieved by delayed Q-learning (Strehl et al., 2006). The authors provide a novel strategy of argument when proving the upper bound for the sample complexity of exploration, namely identifying a sufficient condition for optimality, and then bound the number of times that this condition is violated. ", + "bbox": [ + 174, + 424, + 825, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the results of Delayed Q-learning still leave a quadratic gap in $1 / \\epsilon$ from the best-known lower bound. This is partly because the updates in Q-value are made in an over-conservative way. In fact, the loose sample complexity bound is a result of delayed Q-learning algorithm itself, as well as the mathematical artifact in their analysis. To illustrate this, we construct a hard instance showing that Delayed Q-learning incurs $\\Omega ( 1 / \\epsilon ^ { 3 } )$ sample complexity. This observation, as well as the success of the Q-learning with UCB algorithm (Jin et al., 2018) in proving a regret bound in finite-horizon settings, motivates us to incorporate a UCB-like exploration term into our algorithm. ", + "bbox": [ + 174, + 505, + 825, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a Q-learning algorithm with UCB exploration policy. We show the sample complexity of exploration bound of our algorithm is $\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}$ . This strictly improves the previous best known result due to Delayed Q-learning. It also matches the lower bound in the dependence on $\\epsilon$ , $S$ and $A$ up to logarithmic factors. ", + "bbox": [ + 174, + 609, + 825, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We point out here that the infinite-horizon setting cannot be solved by reducing to finite-horizon setting. There are key technical differences between these two settings: the definition of sample complexity of exploration, time-invariant policies and the error propagation structure in Q-learning. In particular, the analysis techniques developed in (Jin et al., 2018) do not directly apply here. We refer the readers to Section 3.2 for detailed explanations and a concrete example. ", + "bbox": [ + 174, + 676, + 825, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rest of the paper is organized as follows. After introducing the notation used in the paper in Section 2, we describe our infinite Q-learning with UCB algorithm in Section 3. We then state our main theoretical results, which are in the form of PAC sample complexity bounds. In Section 4 we present some interesting properties beyond sample complexity bound. Finally, we conclude the paper in Section 5. ", + "bbox": [ + 174, + 753, + 825, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 PRELIMINARY ", + "text_level": 1, + "bbox": [ + 176, + 856, + 323, + 872 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider a Markov Decision Process defined by a five tuple $\\langle S , \\mathcal { A } , p , r , \\gamma \\rangle$ , where $s$ is the state space, $\\mathcal { A }$ is the action space, $p ( s ^ { \\prime } | s , a )$ is the transition function, $r : S \\times \\mathcal { A } [ 0 , 1 ]$ is the deterministic reward function, and $0 \\leq \\gamma < 1$ is the discount factor for rewards. Let $S = | S |$ and $A = | { \\mathcal { A } } |$ denote the number of states and the number of actions respectively. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Starting from a state $s _ { 1 }$ , the agent interacts with the environment for infinite number of time steps. At each time step, the agent observes state $s _ { t } \\in S$ , picks action $a _ { t } \\in \\mathcal A$ , and receives reward $r _ { t }$ ; the system then transits to next state $s _ { t + 1 }$ . ", + "bbox": [ + 174, + 138, + 823, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the notations in Strehl et al. (2006), a policy $\\pi _ { t }$ refers to the non-stationary control policy of the algorithm since step $t$ . We use $V ^ { \\pi _ { t } } ( s )$ to denote the value function under policy $\\pi _ { t }$ , which is defined as $\\begin{array} { r } { V ^ { \\pi _ { t } } ( s ) = \\mathbb { E } \\bigl [ \\sum _ { i = 1 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = s \\bigr ] } \\end{array}$ . We also use $V ^ { * } ( s ) = \\operatorname* { s u p } _ { \\pi } V ^ { \\pi } ( s )$ to denote the value function of the optimal policy. Accordingly, we define $\\begin{array} { r } { \\mathbb { E } [ \\sum _ { i = 2 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = \\stackrel { \\cdot } { s } , a _ { 1 } = \\stackrel { \\cdot } { a } ] } \\end{array}$ as the $\\mathrm { Q }$ function under policy $Q ^ { \\pi _ { t } } ( s , a ) = r ( s , a ) +$ $\\pi _ { t }$ ; $Q ^ { * } ( s , a )$ is the Q function under optimal policy $\\pi ^ { * }$ . ", + "bbox": [ + 173, + 186, + 825, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the sample complexity of exploration defined in Kakade et al. (2003) to measure the learning efficiency of our algorithm. This sample complexity definition has been widely used in previous works Strehl et al. (2006); Lattimore & Hutter (2012); Strehl $\\&$ Littman (2008). ", + "bbox": [ + 176, + 279, + 821, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. Sample complexity of Exploration of an algorithm $\\mathcal { A L G }$ is defined as the number of time steps t such that the non-stationary policy $\\pi _ { t }$ at time $t$ is not $\\epsilon$ -optimal for current state $s _ { t }$ , i.e. $\\dot { V } ^ { \\pi _ { t } } \\left( s _ { t } \\right) \\dot { < } V ^ { * } \\left( s _ { t } \\right) - \\epsilon .$ . ", + "bbox": [ + 173, + 323, + 826, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Roughly speaking, this measure counts the number of mistakes along the whole trajectory. We use the following definition of PAC-MDP Strehl et al. (2006). ", + "bbox": [ + 174, + 375, + 823, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2. An algorithm $\\mathcal { A L G }$ is said to be PAC-MDP (Probably Approximately Correct in Markov Decision Processes) $i f ,$ for any \u000f and $\\delta$ , the sample complexity of $\\mathcal { A L G }$ is less than some polynomial in the relevant quantities $( S , A , 1 / \\epsilon , 1 / \\delta , 1 / ( 1 - \\gamma ) )$ , with probability at least $1 - \\delta$ . ", + "bbox": [ + 173, + 405, + 825, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, recall that Bellman equation is defined as the following: ", + "bbox": [ + 174, + 457, + 593, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/59cd62d5ec6d571803efed659b258dfba8d0141783a9422cc7a1cf15db20a44e.jpg", + "text": "$$\n\\left\\{ \\begin{array} { l l } { V ^ { \\pi _ { t } } ( s ) = Q ^ { \\pi _ { t } } \\left( s , \\pi _ { t } ( s ) \\right) } \\\\ { Q ^ { \\pi _ { t } } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { \\pi _ { t + 1 } } \\right) ( s , a ) , } \\end{array} \\right. \\quad \\left\\{ \\begin{array} { l l } { V ^ { * } ( s ) = Q ^ { * } \\left( s , \\pi ^ { * } ( s ) \\right) } \\\\ { Q ^ { * } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { * } \\right) ( s , a ) , } \\end{array} \\right.\n$$", + "text_format": "latex", + "bbox": [ + 232, + 473, + 754, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is frequently used in our analysis. Here we denote $\\left[ \\mathbb { P } V ^ { \\pi _ { t } } \\right] ( s , a ) : = \\mathbb { E } _ { s ^ { \\prime } \\sim p ( \\cdot \\vert s , a ) } V ^ { \\pi _ { t + 1 } } ( s ^ { \\prime } ) .$ ", + "bbox": [ + 168, + 508, + 805, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 MAIN RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 542, + 336, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the UCB Q-learning algorithm and the sample complexity bound. ", + "bbox": [ + 166, + 573, + 776, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 ALGORITHM ", + "text_level": 1, + "bbox": [ + 174, + 603, + 303, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Infinite Q-learning with UCB ", + "text_level": 1, + "bbox": [ + 176, + 633, + 455, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameters: \u000f, γ, δ Initializ $\\begin{array} { r l } & { \\textnormal { \\texttt { e } } Q ( s , a ) , \\hat { Q } ( s , a ) \\gets \\frac { 1 } { 1 - \\gamma } , N ( s , a ) \\gets 0 , \\epsilon _ { 1 } \\gets \\frac { \\epsilon } { 2 4 R M \\ln \\frac { 1 } { 1 - \\gamma } } , H \\gets \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } . } \\\\ & { \\cdot ( k ) = \\ln ( S A ( k + 1 ) ( k + 2 ) / \\delta ) , \\alpha _ { k } = \\frac { H + 1 } { H + k } . } \\end{array}$ Define for $t = 1 , 2 , \\dots \\mathbf { d o }$ \n5: Take action $a _ { t } \\gets \\arg \\operatorname* { m a x } _ { a ^ { \\prime } } \\hat { Q } ( s _ { t } , a ^ { \\prime } )$ Receive reward rt and transit to st+1 $\\begin{array} { r l r } & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { k \\gets N ( s _ { t } , a _ { t } ) , b _ { k } \\gets \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H ( k ) } { k } } } & \\\\ & { \\hat { V } ( s _ { t + 1 } ) \\gets \\operatorname* { m a x } _ { a \\in A } \\hat { Q } ( s _ { t + 1 } , a ) } & \\\\ & { Q ( s _ { t } , a _ { t } ) \\gets ( 1 - \\alpha _ { k } ) Q ( s _ { t } , a _ { t } ) + \\alpha _ { k } \\left[ r ( s _ { t } , a _ { t } ) + b _ { k } + \\gamma \\hat { V } ( s _ { t + 1 } ) \\right] } \\\\ & { \\hat { Q } ( s _ { t } , a _ { t } ) \\gets \\operatorname* { m i n } ( \\hat { Q } ( s _ { t } , a _ { t } ) , Q ( s _ { t } , a _ { t } ) ) } & \\end{array}$ t and can be set to $4 { \\sqrt { 2 } }$ \n10: end for ", + "bbox": [ + 179, + 651, + 823, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $c _ { 2 } = 4 \\sqrt { 2 }$ is a constant. $\\begin{array} { r } { R = \\lceil \\ln \\frac { 3 } { \\epsilon ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}$ , while the choice of $M$ can be found in Section. 3.3. ${ \\cal { \\left( M = 0 \\left( \\ln 1 / ( 1 - \\gamma ) \\epsilon \\right) \\right) } }$ ). The learning rate is defined as $\\alpha _ { k } = ( H + 1 ) / ( H + k )$ . $H$ is chosen as $\\frac { \\ln { 1 } / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln { 1 } / \\gamma }$ , which satisfies $H \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma }$ . ", + "bbox": [ + 174, + 872, + 825, + 928 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our UCB Q-learning algorithm (Algorithm 1) maintains an optimistic estimation of action value function $Q ( s , a )$ and its historical minimum value $\\hat { Q } ( s , a )$ . $N _ { t } ( s , a )$ denotes the number of times that $( s , a )$ is experienced before time step $t$ ; $\\tau ( s , a , k )$ denotes the time step $t$ at which $( s _ { t } , a _ { t } ) = ( s , a )$ for the $k$ -th time; if this state-action pair is not visited that many times, $\\tau ( s , a , k ) = \\infty$ . $Q _ { t } ( s , a )$ and $\\hat { Q } _ { t } ( s , a )$ denotes the $Q$ and $\\hat { Q }$ value of $( s , a )$ that the algorithm maintains when arriving at $s _ { t }$ respectively. ", + "bbox": [ + 173, + 103, + 825, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 SAMPLE COMPLEXITY OF EXPLORATION ", + "text_level": 1, + "bbox": [ + 173, + 213, + 500, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our main result is the following sample complexity of exploration bound. ", + "bbox": [ + 174, + 241, + 656, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. For any $\\epsilon > 0$ , $\\delta > 0 , 1 / 2 < \\gamma < 1$ , with probability $1 - \\delta$ , the sample complexity of exploration (i.e., the number of time steps t such that $\\pi _ { t }$ is not $\\epsilon$ -optimal at $s _ { t . }$ ) of Algorithm 1 is at most ", + "bbox": [ + 174, + 261, + 828, + 304 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/cc2b5aa0f63178566f02139857adabb797f2f963feb35907fc51fd7501fb1693.jpg", + "text": "$$\n\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln { 1 / \\delta } } { \\epsilon ^ { 2 } \\left( 1 - \\gamma \\right) ^ { 7 } } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 433, + 305, + 563, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde { \\mathcal { O } }$ suppresses logarithmic factors of $1 / \\epsilon , 1 / ( 1 - \\gamma )$ and $S A$ . ", + "bbox": [ + 173, + 357, + 612, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first point out the obstacles for proving the theorem and reasons why the techniques in Jin et al. \n(2018) do not directly apply here. We then give a high level description of the ideas of our approach. ", + "bbox": [ + 173, + 387, + 825, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One important issue is caused by the difference in the definition of sample complexity for finite and infinite horizon MDP. In finite horizon settings, sample complexity (and regret) is determined in the first $T$ timesteps, and only measures the performance at the initial state $s _ { 1 }$ (i.e. $( V ^ { * } - V ^ { \\pi } ) ( s _ { 1 } ) )$ . However, in the infinite horizon setting, the agent may enter under-explored regions at any time period, and sample complexity of exploration characterizes the performance at all states the agent enters. ", + "bbox": [ + 173, + 421, + 826, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The following example clearly illustrates the key difference between infinite-horizon and finitehorizon. Consider an MDP with a starting state $s _ { 1 }$ where the probability of leaving $s _ { 1 }$ is $o ( T ^ { - 1 } )$ . In this case, with high probability, it would take more than $T$ timesteps to leave $s _ { 1 }$ . Hence, guarantees about the learning in the first $T$ timesteps or about the performance at $s _ { 1 }$ imply almost nothing about the number of mistakes the algorithm would make in the rest of the MDP (i.e. the sample complexity of exploration of the algorithm). As a result, the analysis for finite horizon MDPs cannot be directly applied to infinite horizon setting. ", + "bbox": [ + 174, + 512, + 825, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This calls for techniques for counting mistakes along the entire trajectory, such as those employed by Strehl et al. (2006). In particular, we need to establish convenient sufficient conditions for being $\\epsilon$ -optimal at timestep $t$ and state $s _ { t }$ , i.e. $V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon$ . Then, bounding the number of violations of such conditions gives a bound on sample complexity. ", + "bbox": [ + 173, + 617, + 825, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another technical reason why the proof in Jin et al. (2018) cannot be directly applied to our problem is the following: In finite horizon settings, Jin et al. (2018) decomposed the learning error at episode $k$ and time $h$ as errors from a set of consecutive episodes before $k$ at time $h + 1$ using a clever design of learning rate. However, in the infinite horizon setting, this property does not hold. Suppose at time $t$ the agent is at state $s _ { t }$ and takes action $a _ { t }$ . Then the learning error at $t$ only depends on those previous time steps such that the agent encountered the same state as $s _ { t }$ and took the same action as $a _ { t }$ . Thus the learning error at time $t$ cannot be decomposed as errors from a set of consecutive time steps before $t$ , but errors from a set of non-consecutive time steps without any structure. Therefore, we have to control the sum of learning errors over an unstructured set of time steps. This makes the analysis more challenging. ", + "bbox": [ + 173, + 680, + 825, + 820 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now we give a brief road map of the proof of Theorem 1. Our first goal is to establish a sufficient condition so that $\\pi _ { t }$ learned at step $t$ is $\\epsilon$ -optimal for state $s _ { t }$ . As an intermediate step we show that a sufficient condition for $V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon$ is that $V ^ { \\ast } ( s _ { t ^ { \\prime } } ) - Q ^ { \\ast } ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } )$ is small for a few time steps $t ^ { \\prime }$ within an interval $[ t , t + R ]$ for a carefully chosen $R$ (Condition 1). Then we show the desired sufficient condition (Condition 2) implies Condition 1. We then bound the total number of bad time steps on which $V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )$ is large for the whole MDP; this implies a bound on the number of violations of Condition 2. This in turn relies on a key technical lemma (Lemma 2). ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The remaining part of this section is organized as follows. We establish the sufficient condition for $\\epsilon$ -optimality in Section 3.3. The key lemma is presented in Section 3.4. Finally we prove Theorem 1 in Section 3.5. ", + "bbox": [ + 173, + 103, + 825, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 SUFFICIENT CONDITION FOR $\\epsilon$ -OPTIMALITY ", + "text_level": 1, + "bbox": [ + 173, + 162, + 519, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we establish a sufficient condition (Condition 2) for $\\epsilon$ -optimality at time step $t$ . ", + "bbox": [ + 173, + 188, + 789, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a fixed $s _ { t }$ , let TRAJ $( R )$ be the set of length- $R$ trajectories starting from $s _ { t }$ . Our goal is to give a sufficient condition so that $\\pi _ { t }$ , the policy learned at step $t$ , is $\\epsilon$ -optimal. For any $\\epsilon _ { 2 } > 0$ , define $\\begin{array} { r } { R : = \\lceil \\ln \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}$ . Denote $V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )$ by $\\Delta _ { t }$ . We have ", + "bbox": [ + 174, + 210, + 826, + 256 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0a9ed2938f048cb014418095c8bf22495e13c88841960e9512dce70e6eff25c2.jpg", + "text": "$$\n\\begin{array} { r l } & { \\quad V ^ { * } ( s _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + Q ^ { * } ( s _ { t } , a _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { P } ( V ^ { * } - V ^ { \\pi } ) ( s _ { t } , \\pi _ { t } ( s _ { t } ) ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\sum _ { s + 1 } ^ { \\infty } p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\cdot [ V ^ { * } ( s _ { t + 1 } ) - Q ^ { * } ( s _ { t + 1 } , a _ { t + 1 } ) ] + } \\\\ & { \\quad \\Big . \\Big . \\Big . \\Big . \\Big . \\Big . } \\\\ & { \\quad \\quad \\Big . \\gamma _ { s + 1 , s ^ { * } + 2 } \\mathrm { ~ } p ( s _ { t + 2 } | s _ { t + 1 } , \\pi _ { t + 1 } ( s _ { t + 1 } ) ) \\cdot p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\big [ V ^ { * } ( s _ { t + 2 } ) - Q ^ { * } ( s _ { t + 2 } , a _ { t + 2 } ) \\Big ] } \\\\ & { \\quad \\quad \\cdot \\dots } \\\\ & { \\le \\epsilon _ { 2 } + \\displaystyle \\sum _ { s ^ { \\prime } = s ^ { \\prime } } p ( t r a _ { t } ) \\cdot [ \\frac { n - 1 } { \\lambda _ { 2 } } \\gamma ^ { 2 } \\Delta _ { t + 2 } ] , } \\\\ & { \\quad \\quad \\pi _ { s \\lambda ^ { ( 2 ) } ( L ) } ^ { R , a } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 202, + 263, + 792, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the last inequality holds because $\\frac { \\gamma ^ { R } } { 1 - \\gamma } \\leq \\epsilon _ { 2 }$ , which follows from the definition of $R$ ", + "bbox": [ + 173, + 474, + 763, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For any fixed trajectory of length $R$ starting from $s _ { t }$ , consider the sequence $\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }$ . Let $X _ { t } ^ { ( i ) }$ be the $i$ -th largest item of $\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }$ . Rearranging Eq. (1), we obtain ", + "bbox": [ + 173, + 506, + 821, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/353a318648b4e83bfc1d70cd63fedbd78d74dcde5d09d9484e43d06da944118d.jpg", + "text": "$$\nV ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E _ { t r a j } \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 331, + 547, + 666, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first prove that Condition 1 implies $\\epsilon$ -optimality at time step $t$ when $\\epsilon _ { 2 } = \\epsilon / 3$ ", + "bbox": [ + 173, + 604, + 705, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Condition 1. Let $\\begin{array} { r } { \\xi _ { i } : = \\frac { 1 } { 2 ^ { i + 2 } } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } } \\end{array}$ . For all $0 \\leq i \\leq \\lfloor \\log _ { 2 } R \\rfloor$ ", + "bbox": [ + 173, + 626, + 638, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/4c5556442dd84c67ed2f8017702e406cf2d4d5acde09b8b3242d907ec73dda13.jpg", + "text": "$$\nE [ X _ { t } ^ { ( 2 ^ { i } ) } ] \\leq \\xi _ { i } .\n$$", + "text_format": "latex", + "bbox": [ + 447, + 659, + 549, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Claim 1. If Condition $^ { l }$ is satisfied at time step $t$ , the policy $\\pi _ { t }$ is $\\epsilon$ -optimal at state $s _ { t }$ , i.e. $V ^ { \\ast } ( s _ { t } ) -$ \n$V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon$ . ", + "bbox": [ + 173, + 688, + 825, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. Note that $E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ]$ t. Eq. (3) implies that for $X _ { t } ^ { ( i ) }$ is monotonically decreasing with respect to $1 / 2 < \\gamma < 1$ , $i$ . Therefore, $E [ X _ { t } ^ { ( i ) } ] \\ \\leq$ ", + "bbox": [ + 176, + 733, + 825, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/62ba3bea0f084d11903c964028d906643070d34916bde88924637c5e08ea73d6.jpg", + "text": "$$\n\\begin{array} { r l } & { E \\left[ \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] = \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( i ) } ] \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ] } \\\\ & { \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } 2 ^ { - \\lfloor \\log _ { 2 } i \\rfloor - 2 } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\frac { \\gamma ^ { i - 1 } } { i } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq 2 \\epsilon _ { 2 } , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 235, + 779, + 761, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the last inequality follows from the fact that $\\begin{array} { r } { \\sum _ { i = 1 } ^ { \\infty } \\frac { \\gamma ^ { i - 1 } } { i } = \\frac { 1 } { \\gamma } \\ln \\frac { 1 } { 1 - \\gamma } } \\end{array}$ and $\\gamma > 1 / 2$ ", + "bbox": [ + 169, + 873, + 754, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining with Eq. 2, we have, $\\begin{array} { r } { V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] \\leq 3 \\epsilon _ { 2 } = \\epsilon . } \\end{array}$ ", + "bbox": [ + 166, + 901, + 777, + 928 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next we show that given $i , t$ , Condition 2 implies Eq. (3). ", + "bbox": [ + 173, + 102, + 552, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Condition 2. Define $L = \\lfloor \\log _ { 2 } R \\rfloor$ . Let $\\begin{array} { r } { M = \\operatorname* { m a x } \\left\\{ \\lceil 2 \\log _ { 2 } \\frac { 1 } { \\xi _ { L } ( 1 - \\gamma ) } \\rceil , 1 0 \\right\\} } \\end{array}$ , and $\\begin{array} { r } { \\eta _ { j } = \\frac { \\xi _ { i } } { M } \\cdot 2 ^ { j - 1 } } \\end{array}$ . \nFor all $\\begin{array} { r } { \\because j \\le M , \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] \\le \\frac { \\xi _ { i } } { M } , } \\end{array}$ . ", + "bbox": [ + 173, + 122, + 826, + 169 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Claim 2. Given i, t, Eq. (3) holds if Condition 2 is satisfied. ", + "bbox": [ + 173, + 171, + 571, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. The reason behind the choice of $M$ is to ensure that $\\eta _ { M } > 1 / ( 1 - \\gamma )$ 1. It follows that, assuming Condition 2 holds, for $1 \\leq j \\leq M$ , ", + "bbox": [ + 169, + 203, + 826, + 234 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/b1e98198c486d877466252bbfdd5d0afb8e4019bc37fc19f80c274168b0bb6a3.jpg", + "text": "$$\nE \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } \\right] = \\int _ { 0 } ^ { 1 / ( 1 - \\gamma ) } \\operatorname* { P r } \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } > x \\right] d x \\le \\eta _ { 1 } + \\sum _ { j = 2 } ^ { M } \\eta _ { j } \\operatorname* { P r } [ { X _ { t } ^ { ( 2 ^ { i } ) } } > \\eta _ { j - 1 } ] \\le \\xi _ { i } .\n$$", + "text_format": "latex", + "bbox": [ + 227, + 241, + 769, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, if a time step $t$ is not $\\epsilon _ { 2 }$ -optimal, there exists $0 \\leq i < \\lfloor \\log _ { 2 } R \\rfloor$ and $2 \\leq j \\leq M$ such that ", + "bbox": [ + 171, + 327, + 825, + 343 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/7d918d021bc39256ac2de308ee455c6a54396bc36335cd0f685f840f0c5005d1.jpg", + "text": "$$\n\\eta _ { j } \\mathrm { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } .\n$$", + "text_format": "latex", + "bbox": [ + 405, + 349, + 593, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Now, the sample complexity can be bounded by the number of $( t , i , j )$ pairs that Eq. (4) is violated. Following the approach of Strehl et al. (2006), for a fixed $( i , j )$ -pair, instead of directly counting the number of time steps $t$ such that $\\begin{array} { r } { \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } } \\end{array}$ , we count the number of time steps that $X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 }$ . Lemma 1 provides an upper bound of the number of such $t$ . ", + "bbox": [ + 173, + 387, + 826, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4 KEY LEMMAS ", + "text_level": 1, + "bbox": [ + 174, + 477, + 313, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present two key lemmas. Lemma 1 bounds the number of sub-optimal actions, which in turn, bounds the sample complexity of our algorithm. Lemma 2 bounds the weighted sum of learning error, i.e. $( \\hat { Q } _ { t } - Q ^ { * } ) ( s , a )$ , with the sum and maximum of weights. Then, we show that Lemma 1 follows from Lemma 2. ", + "bbox": [ + 173, + 502, + 826, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 1. For fixed $t$ and $\\eta > 0$ , let ${ B } _ { \\eta } ^ { ( t ) }$ be the event that $\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}$ in step t. If $\\eta > 2 \\epsilon _ { 1 }$ , then with probability at least $1 - \\delta / 2$ , ", + "bbox": [ + 171, + 563, + 825, + 598 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/8e6137e01a7beea297bf95283d863b21a785d17463e6944e2efedb515bf7d808.jpg", + "text": "$$\n\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 303, + 604, + 691, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I [ \\cdot ]$ is the indicator function. ", + "bbox": [ + 174, + 655, + 405, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Before presenting Lemma 2, we define a class of sequence that occurs in the proof. ", + "bbox": [ + 173, + 683, + 718, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Definition 3. A sequence $( w _ { t } ) _ { t \\geq 1 }$ is said to be a $( C , w )$ -sequence for $C , w > 0$ , $i f 0 \\le w _ { t } \\le w$ for all $t \\geq 1$ , and $\\begin{array} { r } { \\sum _ { t \\geq 1 } w _ { t } \\leq C } \\end{array}$ . ", + "bbox": [ + 171, + 702, + 823, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lemma 2. For every $( C , w )$ -sequence $( w _ { t } ) _ { t \\geq 1 }$ , with probability $1 - \\delta / 2$ , the following holds: ", + "bbox": [ + 171, + 737, + 792, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/576704e0ba7bc113c9eb1c2218a2bd6ace18a8d09188a27f397409e00cdaad49.jpg", + "text": "$$\n\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 207, + 779, + 789, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}$ ", + "bbox": [ + 174, + 838, + 472, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof of Lemma 2 is quite technical, and is therefore deferred to supplementary materials. ", + "bbox": [ + 169, + 869, + 764, + 886 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Now, we briefly explain how to prove Lemma 1 with Lemma 2. (Full proof can be found in supplementary materials.) Note that since $\\hat { Q } _ { t } \\geq Q ^ { * }$ and $a _ { t } = \\arg \\operatorname* { m a x } _ { a } \\hat { Q } _ { t } ( s _ { t } , a )$ , ", + "bbox": [ + 173, + 103, + 825, + 135 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/e36c4bae9f64d8b170464eae13de7df006383d23275ffcf8d31e0711f7c4678a.jpg", + "text": "$$\nV ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\leq \\hat { Q } _ { t } ( s _ { t } , a _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) .\n$$", + "text_format": "latex", + "bbox": [ + 338, + 137, + 653, + 156 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now consider a set $J = \\{ t : \\ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta ( 1 - \\gamma ) ^ { - 1 } \\}$ , and consider the $( | J | , 1 )$ - weight sequence defined by $w _ { t } ~ = ~ I \\left[ t \\in J \\right]$ . We can now apply Lemma 2 to weighted sum $\\begin{array} { r } { \\sum _ { t \\geq 1 } \\dot { w } _ { t } \\left[ \\bar { V } ^ { \\ast } ( s _ { t } ) - Q ^ { \\ast } ( s _ { t } , \\dot { a } _ { t } ) \\right] , } \\end{array}$ . On the one hand, this quantity is obviously at least $| J | \\bar { \\eta ( 1 - \\gamma ) } ^ { - 1 }$ . On the other hand, by lemma 2, it is upper bounded by the weighted sum of $( \\hat { Q } - Q ^ { * } ) ( s _ { t } , a _ { t } )$ . Thus we get ", + "bbox": [ + 173, + 157, + 826, + 233 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/dc3a723cc9eab16cf16971abf7bbf69938ee49002e529fdd61110b29bb930e71.jpg", + "text": "$$\n| J | \\eta ( 1 - \\gamma ) ^ { - 1 } \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { S A | J | \\ell ( | J | ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln | J | } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 238, + 234, + 758, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Now focus on the dependence on $| J |$ . The left-hand-side has linear dependence on $| J |$ , whereas the left-hand-side has a $\\tilde { \\mathcal { O } } \\left( \\sqrt { | J | } \\right)$ dependence. This allows us to solve out an upper bound on $| J |$ with quadratic dependence on $1 / \\eta$ . ", + "bbox": [ + 173, + 279, + 825, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.5 PROOF FOR THEOREM 1 ", + "text_level": 1, + "bbox": [ + 174, + 347, + 382, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We prove the theorem by stitching Lemma 1 and Condition 2. ", + "bbox": [ + 173, + 372, + 578, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Proof. (Proof for Theorem 1) ", + "bbox": [ + 174, + 402, + 372, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By lemma 1, for any $2 \\leq j \\leq M , \\sum _ { t = 1 } ^ { \\infty } I \\left[ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta _ { j - 1 } \\right] \\leq C$ , where $C = \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta _ { j - 1 } ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\tilde { P } .$ ", + "bbox": [ + 174, + 422, + 733, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here $\\tilde { P }$ is a shorthand for polylog $\\begin{array} { r } { \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) } \\end{array}$ ", + "bbox": [ + 173, + 481, + 475, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let $A _ { t } = I [ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } ]$ be a Bernoulli random variable, and $\\{ \\mathcal { F } _ { t } \\} _ { t \\ge 1 }$ be the filtration generated by random variables $\\bar { \\{ ( s _ { \\tau } , a _ { \\tau } ) : 1 \\leq \\tau \\leq t \\} }$ . Since $A _ { t }$ is $\\mathcal { F } _ { t + R }$ −measurable, for any $0 \\leq k < R$ , $\\{ \\dot { A } _ { k + t R } - E [ A _ { k + t R } \\ | \\ \\mathcal { F } _ { k + t R } ] \\} _ { t \\geq 0 }$ is a martingale difference sequence. For now, consider a fixed $0 \\leq k < R$ . By Azuma-Hoeffiding inequality, after $\\begin{array} { r } { T = \\mathcal { O } \\left( \\frac { C } { 2 ^ { i } } \\cdot \\frac { M \\eta _ { j } } { \\xi _ { i } } \\ln ( R M L ) \\right) } \\end{array}$ · M ηjξ ln(RM L)\u0011 time steps (if it happens that many times) with ", + "bbox": [ + 171, + 512, + 826, + 597 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/b83b246b4683938705ff27511cfa875c502578f40906ac3c2c49408d5a5881f3.jpg", + "text": "$$\n\\operatorname* { P r } \\left[ X _ { k + t R } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] = \\mathbb { E } [ A _ { k + t R } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } ,\n$$", + "text_format": "latex", + "bbox": [ + 354, + 598, + 642, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we have $\\textstyle \\sum _ { t } A _ { k + t R } \\geq C / 2 ^ { i }$ with probability at least $1 - \\delta / ( 2 M R L )$ . ", + "bbox": [ + 174, + 633, + 632, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the other hand, if $A _ { k + t R }$ happens, within $[ k + t R , k + t R + R - 1 ]$ , there must be at least $2 ^ { i }$ time steps at which $V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > ^ { \\mathsf { ^ { * } } } \\eta _ { j - 1 }$ . The latter event happens at most $C$ times, and the $[ k + t R , k + t R + R - 1 ]$ are disjoint. Tppens at most erefore, times f $\\textstyle \\sum _ { t = 0 } ^ { \\infty } A _ { k + t R } \\leq C / 2 ^ { i }$ . This suggests that a union bound on $T$ $i$ $j$ $0 \\leq k < R$ , we can show that with probability $1 - \\delta / ( 2 M L )$ , there are at most $R T$ time steps where $\\mathrm { P r } \\left[ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] > \\xi _ { i } / ( M \\eta _ { j } )$ . Thus, the number of sub-optimal steps is bounded by, ", + "bbox": [ + 173, + 655, + 826, + 752 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/19060e0af63eb5746ff2fa8e81a511f198a258681b5fb178d09da3690eab476b.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\sum _ { t = 1 } ^ { \\ell - 1 } I [ V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) > \\epsilon ] } } \\\\ & { \\leq \\sum _ { t = 1 } ^ { \\infty } \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } I [ \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } ] = \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\sum _ { t = 1 } ^ { \\infty } I [ \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { \\eta _ { j } M } ] } \\\\ & { \\leq \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\frac { S A M R \\ln 1 / \\delta \\ln S A } { \\eta _ { j } \\xi _ { i } \\cdot 2 ^ { i } ( 1 - \\gamma ) ^ { 5 } } \\tilde { P } \\leq \\sum _ { i = 0 } ^ { L } \\frac { S A \\cdot 2 ^ { i + 4 } \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\xi _ { i } \\mathrm { ~ a n d } \\eta _ { j } ) } \\\\ & { \\leq \\frac { S A R \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\tilde { P } . } & { \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\hbar ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 184, + 762, + 836, + 926 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It should be stressed that throughout the lines, $\\tilde { P }$ is a shorthand for an asymptotic expression, instead of an exact value. Our final choice of \u000f2 and \u000f1 are \u000f2 = \u000f3 , and \u000f1 = \u000f24RM ln . It is not hard 1−γ to see that $\\ln 1 / \\epsilon _ { 1 } = \\mathrm { p o l y } ( \\ln \\textstyle \\frac { 1 } { \\epsilon } , \\ln \\frac { 1 } { 1 - \\gamma } )$ . This immediately implies that with probability $1 - \\delta$ , the number of time steps such that $\\left( V ^ { * } - \\dot { V } ^ { \\pi } \\right) \\left( s _ { t } \\right) > \\epsilon$ is ", + "bbox": [ + 173, + 102, + 825, + 171 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/112b2a0ff239fa280515807bb22278ab6c902bb4dd2bada09b6a5e780e0dcd78.jpg", + "text": "$$\n\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln 1 / \\delta } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 434, + 176, + 560, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where hidden factors are $\\begin{array} { r } { \\operatorname { p o l y } ( \\ln { \\frac { 1 } { \\epsilon } } , \\ln { \\frac { 1 } { 1 - \\gamma } } , \\ln S A ) } \\end{array}$ . ", + "bbox": [ + 173, + 218, + 511, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 DISCUSSION ", + "text_level": 1, + "bbox": [ + 174, + 253, + 312, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we discuss the implication of our results, and present some interesting properties of our algorithm beyond its sample complexity bound. ", + "bbox": [ + 174, + 286, + 825, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 COMPARISON WITH PREVIOUS RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 332, + 490, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lower bound To the best of our knowledge, the current best lower bound for worst-case sample complexity is \u0010 SA\u000f2(1−γ)3 ln 1/δ\u0011 due to Lattimore & Hutter (2012). The gap between our results and this lower bound lies only in the dependence on $1 / ( 1 - \\gamma )$ and logarithmic terms of $\\mathrm { { } } ^ { \\mathrm { { 3 } } A , 1 / ( 1 - \\gamma ) }$ and $1 / \\epsilon$ . ", + "bbox": [ + 173, + 357, + 826, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model-free algorithms Previously, the best sample complexity bound for a model-free algorithm is $\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\right) } \\end{array}$ (suppressing all logarithmic terms), achieved by Delayed Q-learning Strehl et al. (2006). Our results improve this upper bound by a factor of $\\frac { 1 } { \\epsilon ^ { 2 } ( 1 - \\gamma ) }$ , and closes the quadratic gap in $1 / \\epsilon$ between Delayed Q-learning’s result and the lower bound. In fact, the following theorem shows that UCB Q-learning can indeed outperform Delayed Q-learning. ", + "bbox": [ + 173, + 429, + 826, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Theorem 2. There exists a family of MDPs with constant $S$ and $A$ , in which with probability $1 - \\delta$ , Delayed $Q$ -learning incurs sample complexity of exploration of $\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)$ , assuming that $\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }$ . ", + "bbox": [ + 173, + 518, + 825, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The construction of this hard MDP family is given in the supplementary material. ", + "bbox": [ + 173, + 582, + 705, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model-based algorithms For model-based algorithms, better sample complexity results in infinite horizon settings have been claimed Szita & Szepesvári (2010). To the best of our knowledge, the best published result without further restrictions on MDPs is $\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\right) } \\end{array}$ claimed by Szita & Szepesvári (2010), which is $( 1 - \\gamma )$ smaller than our upper bound. From the space complexity point of view, our algorithm is much more memory-efficient. Our algorithm stores $O ( S A )$ values, whereas the algorithm in Szita & Szepesvári (2010) needs $\\Omega ( S ^ { 2 } A )$ memory to store the transition model. ", + "bbox": [ + 173, + 603, + 826, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 EXTENSION TO OTHER SETTINGS ", + "text_level": 1, + "bbox": [ + 176, + 715, + 442, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to length limits, detailed discussion in this section is deferred to supplementary materials. ", + "bbox": [ + 173, + 739, + 790, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finite horizon MDP The sample complexity of exploration bounds of UCB Q-learning implies $\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)$ PAC sample complexity and a $\\bar { \\mathcal { O } } \\left( T ^ { 1 / 2 } \\right)$ regret bound in finite horizon MDPs. That is, our algorithm implies a PAC algorithm for finite horizon MDPs. We are not aware of reductions of the opposite direction (from finite horizon sample complexity to infinite horizon sample complexity of exploration). ", + "bbox": [ + 173, + 761, + 825, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regret The reason why our results can imply an $\\tilde { \\mathcal { O } } ( { \\sqrt { T } } )$ regret is that, after choosing $\\epsilon _ { 1 }$ , it follows from the argument of Theorem 1 that with probability $1 - \\delta$ , for all $\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )$ , the number of $\\epsilon _ { 2 }$ -suboptimal steps is bounded by ", + "bbox": [ + 174, + 840, + 825, + 887 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/4312c45d17cd6e06d125f0e95014a5792379fa26cd13205bd4de52e4ac033fac.jpg", + "text": "$$\n\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 344, + 893, + 651, + 929 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, Delayed Q-learning Strehl et al. (2006) can only give an upper bound on $\\epsilon _ { 1 }$ -suboptimal steps after setting parameter $\\epsilon _ { 1 }$ . ", + "bbox": [ + 173, + 103, + 823, + 132 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 151, + 318, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Infinite-horizon MDP with discounted reward is a setting that is arguably more difficult than other popular settings, such as finite-horizon MDP. Previously, the best samby model-free reinforcement learning algorithms in this setting is $\\tilde { \\cal O } \\big ( { \\textstyle \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma _ { _ { - } } ) ^ { 8 } } } \\big )$ y bound achieved, due to Delayed Q-learning Strehl et al. (2006). In this paper, we propose a variant of Q-learning that incorporates upper confidence bound, and show that it has a sample complexity of $\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}$ . This matches the best lower bound except in dependence on $1 / ( 1 - \\gamma )$ and logarithmic factors. ", + "bbox": [ + 174, + 183, + 825, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 ACKNOWLEDGEMENTS ", + "text_level": 1, + "bbox": [ + 176, + 297, + 395, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors thank Chi Jin and Chongjie Zhang for helpful discussions. This work is supported by National Basic Research Program of China (973 Program) (grant no. 2015CB352502), NSFC (61573026), BJNSF (L172037) and Beijing Acedemy of Artificial Intelligence. ", + "bbox": [ + 174, + 328, + 823, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 391, + 287, + 406 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mohammad Gheshlaghi Azar, Remi Munos, Mohammad Ghavamzadeh, and Hilbert Kappen. Speedy q-learning. In Advances in neural information processing systems, 2011. ", + "bbox": [ + 174, + 415, + 823, + 443 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mohammad Gheshlaghi Azar, Ian Osband, and Rémi Munos. Minimax regret bounds for reinforcement learning. arXiv preprint arXiv:1703.05449, 2017. ", + "bbox": [ + 173, + 452, + 823, + 481 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ronen I. Brafman and Moshe Tennenholtz. R-max - a general polynomial time algorithm for near-optimal reinforcement learning. J. Mach. Learn. Res., 3:213–231, March 2003. ISSN 1532-4435. doi: 10.1162/153244303765208377. URL https://doi.org/10.1162/ 153244303765208377. ", + "bbox": [ + 174, + 488, + 826, + 545 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Christoph Dann, Tor Lattimore, and Emma Brunskill. Unifying pac and regret: Uniform pac bounds for episodic reinforcement learning. In Advances in Neural Information Processing Systems, pp. 5713–5723, 2017. ", + "bbox": [ + 173, + 554, + 823, + 597 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Eyal Even-Dar and Yishay Mansour. Learning rates for q-learning. Journal of Machine Learning Research, 5(Dec):1–25, 2003. ", + "bbox": [ + 173, + 604, + 823, + 633 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thomas Jaksch, Ronald Ortner, and Peter Auer. Near-optimal regret bounds for reinforcement learning. Journal of Machine Learning Research, 11(Apr):1563–1600, 2010. ", + "bbox": [ + 173, + 642, + 823, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Chi Jin, Zeyuan Allen-Zhu, Sebastien Bubeck, and Michael I Jordan. Is q-learning provably efficient? In Advances in Neural Information Processing Systems, pp. 4864–4874, 2018. ", + "bbox": [ + 173, + 680, + 823, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sham Machandranath Kakade et al. On the sample complexity of reinforcement learning. PhD thesis, University of London London, England, 2003. ", + "bbox": [ + 173, + 718, + 825, + 747 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tor Lattimore and Marcus Hutter. Pac bounds for discounted mdps. In International Conference on Algorithmic Learning Theory, pp. 320–334. Springer, 2012. ", + "bbox": [ + 171, + 755, + 823, + 785 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013. ", + "bbox": [ + 176, + 791, + 825, + 835 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pp. 1928–1937, 2016. ", + "bbox": [ + 173, + 843, + 823, + 887 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015. ", + "bbox": [ + 173, + 895, + 820, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Aaron Sidford, Mengdi Wang, Xian Wu, Lin Yang, and Yinyu Ye. Near-optimal time and sample complexities for solving markov decision processes with a generative model. In Advances in Neural Information Processing Systems, pp. 5186–5196, 2018a. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aaron Sidford, Mengdi Wang, Xian Wu, and Yinyu Ye. Variance reduced value iteration and faster algorithms for solving markov decision processes. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pp. 770–787. Society for Industrial and Applied Mathematics, 2018b. ", + "bbox": [ + 174, + 155, + 825, + 212 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alexander L Strehl and Michael L Littman. An analysis of model-based interval estimation for markov decision processes. Journal of Computer and System Sciences, 74(8):1309–1331, 2008. ", + "bbox": [ + 171, + 219, + 821, + 250 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alexander L Strehl, Lihong Li, Eric Wiewiora, John Langford, and Michael L Littman. Pac modelfree reinforcement learning. In Proceedings of the 23rd international conference on Machine learning, pp. 881–888. ACM, 2006. ", + "bbox": [ + 174, + 257, + 826, + 301 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "István Szita and Csaba Szepesvári. Model-based reinforcement learning with nearly tight exploration complexity bounds. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pp. 1031–1038, 2010. ", + "bbox": [ + 174, + 309, + 825, + 353 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A PROOF OF LEMMA 1 ", + "text_level": 1, + "bbox": [ + 176, + 378, + 379, + 395 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lemma 1. For fixed t and $\\eta > 0$ , let $B _ { \\eta } ^ { ( t ) }$ be the event that $\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}$ in step t. If $\\eta > 2 \\epsilon _ { 1 }$ , then with probability at least $1 - \\delta / 2$ , ", + "bbox": [ + 171, + 407, + 825, + 443 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/0b9f09a826319782774c08f51f81a7e3a19a570518937c623163e273091ef961.jpg", + "text": "$$\n\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 303, + 449, + 691, + 492 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "where $I [ \\cdot ]$ is the indicator function. ", + "bbox": [ + 174, + 498, + 405, + 513 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Proof. When $\\eta > 1$ the lemma holds trivially. Now consider the case that $\\eta \\leq 1$ . ", + "bbox": [ + 173, + 527, + 704, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Let $\\begin{array} { r } { I = \\{ t \\colon V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } \\} } \\end{array}$ . By lemma 2, with probability $1 - \\delta$ ", + "bbox": [ + 173, + 549, + 702, + 568 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/c3600e24303e29ac9ff643f0f4892a82d147d38d65c5111e6d8ac103703617e6.jpg", + "text": "$$\n\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } \\leq \\displaystyle \\sum _ { t \\in I } \\left( V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\right) \\leq \\displaystyle \\sum _ { t \\in I } \\left[ \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { 1 } { ( 1 - \\gamma ) ^ { 5 / 2 } } \\sqrt { S A | I | \\ell | ( I | ) } + \\frac { S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln | I | \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln \\frac { S A | I | } { \\delta } } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln S A | I | } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 225, + 577, + 771, + 747 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Suppose that $\\begin{array} { r } { | I | = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } \\end{array}$ , for some $k > 1$ . Then it follows that for some constant $C _ { 1 }$ , ", + "bbox": [ + 173, + 779, + 795, + 800 + ], + "page_idx": 9 + }, + { + "type": "equation", + "img_path": "images/c06bd9e11a4120e53c0b8d2cd822906d72211ae50589d30086bcbaef0151ca3a.jpg", + "text": "$$\n\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } = \\cfrac { k ^ { 2 } S A \\ln S A } { ( 1 - \\gamma ) ^ { 4 } \\eta } \\leq 2 \\cfrac { ( \\eta - \\epsilon _ { 1 } ) | I | } { 1 - \\gamma } } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { \\sqrt { S A | I | \\ln \\left( S A | I | \\right) } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { S A k } { \\eta ( 1 - \\gamma ) ^ { 4 } } \\sqrt { \\ln S A \\cdot ( \\ln S A + \\ln | I | ) } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 209, + 806, + 789, + 925 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Therefore ", + "bbox": [ + 173, + 103, + 240, + 118 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/bae465b19e019059a91e97016934af5f7bf0c47e9357a7ca6741dfa6dd715015.jpg", + "text": "$$\n\\begin{array} { r l } & { k ^ { 2 } \\ln ( S A ) \\le C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( k \\left( \\ln S A + \\ln | I | \\right) + \\eta ( 1 - \\gamma ) \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\ln S A + 2 \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( 3 \\ln S A + 4 \\ln k + 6 \\ln \\frac { 1 } { \\eta ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\le 6 k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\ln S A + \\ln \\epsilon k \\right) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 245, + 122, + 753, + 273 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Let $\\begin{array} { r } { C ^ { \\prime } = \\operatorname* { m a x } \\lbrace 2 , 6 C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\rbrace } \\end{array}$ . Then ", + "bbox": [ + 171, + 280, + 495, + 304 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/6648dce62ed22bd5032b6259b54c85d35efee676cf37255340a1199f339eda3d.jpg", + "text": "$$\nk \\leq C ^ { \\prime } ( 2 + \\ln k ) .\n$$", + "text_format": "latex", + "bbox": [ + 437, + 310, + 560, + 330 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "If $k \\geq 1 0 C ^ { \\prime } \\ln C ^ { \\prime }$ , then ", + "bbox": [ + 173, + 335, + 328, + 351 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/4ae996b10596e3c7079276b4b960d96f44d70fb21ef5f209a0ea7326b6d16c83.jpg", + "text": "$$\n\\begin{array} { c } { k - C ^ { \\prime } \\left( 2 + \\ln k \\right) \\geq 8 C ^ { \\prime } \\ln C ^ { \\prime } - ( 2 + \\ln 1 0 ) C ^ { \\prime } } \\\\ { \\geq 4 C ^ { \\prime } \\left( 2 \\ln C ^ { \\prime } - 4 \\right) \\geq 0 , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 341, + 356, + 655, + 397 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "which means violation of (9). Therefore, since $C ^ { \\prime } \\geq 2$ ", + "bbox": [ + 173, + 401, + 532, + 416 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/3659eb39bbdbcc0c60a416911c59460e57107ec3dbd0e6f62d3b7b93a525422e.jpg", + "text": "$$\nk \\leq 1 0 C ^ { \\prime } \\ln C ^ { \\prime } \\leq 3 6 0 C _ { 1 } ^ { 2 } \\operatorname* { m a x } \\{ \\ln ^ { 4 } { \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } } , 2 0 \\ln 2 \\} .\n$$", + "text_format": "latex", + "bbox": [ + 315, + 421, + 681, + 455 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "It immediately follows that ", + "bbox": [ + 173, + 462, + 354, + 476 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/ad528682451c2a80d59b5c6f1508b5b7b9f26c92e6092f8aeb71c929441cec39.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\vert I \\vert = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } } \\\\ & { \\leq \\frac { S A \\ln S A } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\ln \\frac { 1 } { \\delta } \\cdot \\mathcal { O } ( \\ln ^ { 8 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 343, + 482, + 655, + 555 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B PROOF OF LEMMA 2 ", + "text_level": 1, + "bbox": [ + 176, + 602, + 380, + 618 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lemma 2. For every $( C , w )$ -sequence $( w _ { t } ) _ { t \\geq 1 }$ , with probability $1 - \\delta / 2$ , the following holds: ", + "bbox": [ + 168, + 632, + 790, + 650 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/ff487b4def18b98c770f155f72c7d52760fabc4d6097e5e050552ffa5fb8baab.jpg", + "text": "$$\n\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 209, + 674, + 787, + 717 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "where $\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}$ is a log-factor. ", + "bbox": [ + 174, + 732, + 472, + 752 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fact 1. (1) The following statement holds throughout the algorithm, ", + "bbox": [ + 173, + 755, + 622, + 770 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/33f263fc3dd14df50727d4ef76ef6c4ffe1d222ffa11393ce94b14974de75afd.jpg", + "text": "$$\n\\hat { Q } _ { p + 1 } ( s , a ) \\leq Q _ { p + 1 } ( s , a ) .\n$$", + "text_format": "latex", + "bbox": [ + 410, + 776, + 586, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "(2) For any $p$ , there exists $p ^ { \\prime } \\leq p$ such that ", + "bbox": [ + 173, + 803, + 452, + 818 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/6fac3e0bba44afb720a787e2593dcafd19d5e03a7a0d6878331dbb6aecc146fa.jpg", + "text": "$$\n{ \\hat { Q } } _ { p + 1 } ( s , a ) \\geq Q _ { p ^ { \\prime } + 1 } ( s , a ) .\n$$", + "text_format": "latex", + "bbox": [ + 408, + 824, + 588, + 844 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Proof. Both properties are results of the update rule at line 11 of Algorithm 1. ", + "bbox": [ + 173, + 858, + 686, + 875 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Before proving lemma 2, we will prove two auxiliary lemmas. ", + "bbox": [ + 171, + 890, + 583, + 906 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Lemma 3. The following properties hold for $\\alpha _ { t } ^ { i }$ : ", + "bbox": [ + 174, + 909, + 498, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "1. $\\begin{array} { r } { \\sqrt { \\frac { 1 } { t } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { 1 } { i } } \\leq 2 \\sqrt { \\frac { 1 } { t } } } \\end{array}$ for every $t \\geq 1 , c > 0$ . ", + "bbox": [ + 210, + 99, + 576, + 127 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "2. $\\begin{array} { r } { \\operatorname* { m a x } _ { i \\in [ t ] } \\alpha _ { t } ^ { i } \\leq \\frac { 2 H } { t } } \\end{array}$ and $\\begin{array} { r } { \\sum _ { i = 1 } ^ { t } ( \\alpha _ { t } ^ { i } ) ^ { 2 } \\le \\frac { 2 H } { t } } \\end{array}$ for every $t \\geq 1$ ", + "bbox": [ + 210, + 133, + 611, + 155 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "3. $\\textstyle \\sum _ { t = i } ^ { \\infty } \\alpha _ { t } ^ { i } = 1 + 1 / H$ , for every $i \\geq 1$ ", + "bbox": [ + 210, + 161, + 480, + 179 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proof. Recall that ", + "bbox": [ + 173, + 227, + 295, + 242 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/242c5979af7b492fba1d80135236539576c26cc515f271d589d39be0ce0b496a.jpg", + "text": "$$\n\\alpha _ { t } = \\frac { H + 1 } { H + t } , \\quad \\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) .\n$$", + "text_format": "latex", + "bbox": [ + 292, + 246, + 705, + 290 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Properties 1-3 are proven by Jin et al. (2018). Now we prove the last property. ", + "bbox": [ + 173, + 294, + 686, + 309 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "On the one hand, ", + "bbox": [ + 173, + 315, + 287, + 330 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/07ae804a6b110033f8bf12b32de163cad5567ec4167d97952b977363794b7bc4.jpg", + "text": "$$\n\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( t ) } { i } } \\leq 2 \\sqrt { \\frac { \\iota ( t ) } { t } } ,\n$$", + "text_format": "latex", + "bbox": [ + 357, + 327, + 637, + 369 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where the last inequality follows from property 1. ", + "bbox": [ + 173, + 371, + 500, + 386 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The left-hand side is proven by induction on $t$ . For the base case, when $t = 1 , \\alpha _ { t } ^ { t } = 1$ . For $t \\geq 2$ , we have $\\alpha _ { t } ^ { i } = ( 1 - \\alpha _ { t } ) \\dot { \\alpha _ { t - 1 } ^ { i } }$ for $1 \\leq i \\leq t - 1$ . It follows that ", + "bbox": [ + 173, + 391, + 825, + 420 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/ab4e05f21481995ad7b6907c39b07eca8dcf1b1707af7599ec0fb0ba9984a116.jpg", + "text": "$$\n\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } = \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sum _ { i = 1 } ^ { t - 1 } \\alpha _ { t - 1 } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } .\n$$", + "text_format": "latex", + "bbox": [ + 199, + 425, + 795, + 469 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Since function $f ( t ) = \\iota ( t ) / t$ is monotonically decreasing for $t \\geq 1 , c \\geq 1$ , we have ", + "bbox": [ + 173, + 473, + 720, + 489 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/7a7308ee3372dff7c4a6c3cae36090077c647b119d15a408f894359ff305977f.jpg", + "text": "$$\n\\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t ) } { t } } \\geq \\sqrt { \\frac { \\iota ( t ) } { t } } .\n$$", + "text_format": "latex", + "bbox": [ + 253, + 492, + 743, + 529 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lemma 4. With probability at least $1 - \\delta / 2$ , for all $p \\geq 0$ and $( s , a )$ -pair, ", + "bbox": [ + 171, + 553, + 661, + 570 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/bf1a7588fbf4d42939d18d2d26129510541081c72f7d186c33706ba60d0c68e1.jpg", + "text": "$$\n\\begin{array} { r l } & { 0 \\le ( Q _ { p } - Q ^ { * } ) ( s , a ) \\le \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } _ { t _ { i } } - V ^ { * } ) ( s _ { t _ { i } + 1 } ) + \\beta _ { t } , } \\\\ & { 0 \\le ( { \\hat { Q } } _ { p } - Q ^ { * } ) ( s , a ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 277, + 574, + 717, + 638 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "where $t = N _ { p } ( s , a ) , t _ { i } = \\tau ( s , a , i )$ and $\\beta _ { t } = c _ { 3 } \\sqrt { H \\iota ( t ) / ( ( 1 - \\gamma ) ^ { 2 } t ) }$ . ", + "bbox": [ + 174, + 643, + 630, + 661 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Proof. Recall that ", + "bbox": [ + 173, + 674, + 295, + 689 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/88278a9f30fd7a368323e468f5e3ad316e4a238d906d20218881efa56614124b.jpg", + "text": "$$\n\\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) .\n$$", + "text_format": "latex", + "bbox": [ + 346, + 686, + 650, + 731 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "From the update rule, it can be seen that our algorithm maintains the following $Q ( s , a )$ : ", + "bbox": [ + 171, + 733, + 746, + 748 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/1b23fc9ed7edc8e4d294bbfadc8232c0e3a3f7ae514e3b0f9f7eb6f05a9180e4.jpg", + "text": "$$\nQ _ { p } ( s , a ) = \\alpha _ { t } ^ { 0 } \\frac { 1 } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + b _ { i } + \\gamma \\hat { V } _ { t _ { i } } ( s _ { t _ { i } + 1 } ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 297, + 752, + 700, + 795 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Bellman optimality equation gives: ", + "bbox": [ + 173, + 799, + 405, + 814 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/7f3af02c43fff8510d8c8a82457783c482e400b5602396a1f237a333d06d0211.jpg", + "text": "$$\nQ ^ { * } ( s , a ) = r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) = \\alpha _ { t } ^ { 0 } Q ^ { * } ( s , a ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 228, + 818, + 767, + 861 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Subtracting the two equations gives ", + "bbox": [ + 174, + 864, + 408, + 880 + ], + "page_idx": 11 + }, + { + "type": "equation", + "img_path": "images/4daae5a78519cf382ebb3c39463504ef4bdc6c481b0042a0fad9b5891adc01b0.jpg", + "text": "$$\nQ _ { p } - Q ^ { * } ) ( s , a ) = \\alpha _ { t } ^ { 0 } ( \\frac { 1 } { 1 - \\gamma } - Q ^ { * } ( s , a ) ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ b _ { i } + \\gamma \\left( V _ { t _ { i } } - V ^ { * } \\right) \\left( s _ { t _ { i } + 1 } \\right) + \\gamma \\left( V ^ { * } \\left( s _ { t _ { i } + 1 } \\right) - \\mathbb { P } V ^ { * } \\right) \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 181, + 883, + 839, + 926 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The identity above holds for arbitrary $p , s$ and $a$ . Now fix $s \\in S$ , $a \\in A$ and $p \\in \\mathbb N$ . Let $t = N _ { p } ( s , a )$ , $t _ { i } = \\tau ( s , a , i )$ . The $t = 0$ case is trivial; we assume $t \\geq 1$ below. Now consider an arbitrary fixed $k$ . Define ", + "bbox": [ + 173, + 103, + 826, + 145 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/7ed6137165be41b2ad3ac59dec10e06c5f86c9778d1953d9a7904ac5097d5be4.jpg", + "text": "$$\n\\Delta _ { i } = \\left( \\alpha _ { k } ^ { i } \\cdot I [ t _ { i } < \\infty ] \\cdot \\left( \\mathbb { P } V ^ { * } - \\mathbb { \\hat { P } } _ { t _ { i } } V ^ { * } \\right) ( s , a ) \\right)\n$$", + "text_format": "latex", + "bbox": [ + 336, + 145, + 660, + 172 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Let $F _ { i }$ be the $\\sigma$ -Field generated by random variables $( s _ { 1 } , a _ { 1 } , . . . , s _ { t _ { i } } , a _ { t _ { i } } )$ . It can be seen that $\\mathbb { E } \\left[ \\Delta _ { i } | F _ { i } \\right] = 0$ , while $\\Delta _ { i }$ is measurable in $F _ { i + 1 }$ . Also, since $\\textstyle 0 \\leq V ^ { \\ast } ( s , { \\bar { a } } ) \\leq { \\frac { 1 } { 1 - \\gamma } }$ , $\\begin{array} { r } { | \\Delta _ { i } | \\le \\frac { 2 } { 1 - \\gamma } } \\end{array}$ . Therefore, $\\Delta _ { i }$ is a martingale difference sequence; by the Azuma-Hoeffding inequality, ", + "bbox": [ + 173, + 174, + 828, + 219 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/e1fd0592f2fbfb43a68fa1b2348697afc8a3b51894eb01a3c12dfc2769dfd098.jpg", + "text": "$$\n\\mathrm { P r } \\left[ \\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| > \\eta \\right] \\leq 2 \\exp \\left\\{ - \\frac { \\eta ^ { 2 } } { 8 \\left( 1 - \\gamma \\right) ^ { - 2 } \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } } \\right\\} .\n$$", + "text_format": "latex", + "bbox": [ + 299, + 220, + 699, + 265 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "By choosing $\\eta$ , we can show that with probability $1 - \\delta / \\left[ S A ( k + 1 ) ( k + 2 ) \\right]$ , ", + "bbox": [ + 173, + 266, + 687, + 282 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/4635fc976e34fd23da44f8967bc69d839d0dafaecc85d492653e8f8184e7782e.jpg", + "text": "$$\n\\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| \\leq \\frac { 2 \\sqrt { 2 } } { 1 - \\gamma } \\cdot \\sqrt { \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } \\cdot \\ln \\frac { 2 ( k + 1 ) ( k + 2 ) S A } { \\delta } } \\leq \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } .\n$$", + "text_format": "latex", + "bbox": [ + 241, + 304, + 753, + 352 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Here $c _ { 2 } = 4 \\sqrt { 2 }$ , $\\begin{array} { r } { \\iota ( k ) = \\ln { \\frac { ( k + 1 ) ( k + 2 ) S A } { \\delta } } } \\end{array}$ . By a union bound for all $k$ , this holds for arbitrary $k > 0$ arbitrary $s \\in S , a \\in A$ simultaneously with probability ", + "bbox": [ + 176, + 354, + 826, + 386 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/d36f303c3c54ca211346b95b702c878d2e496acc0473679ca8912f1c49bc42a5.jpg", + "text": "$$\n1 - \\sum _ { s ^ { \\prime } \\in S , a ^ { \\prime } \\in A } \\sum _ { k = 1 } ^ { \\infty } \\frac { \\delta } { 2 S A ( k + 1 ) ( k + 2 ) } = 1 - \\frac { \\delta } { 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 334, + 387, + 660, + 430 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Therefore, we conclude that (16) holds for the random variable $t = N _ { p } ( s , a )$ and for all $p$ , with probability $1 - \\delta / 2$ as well. ", + "bbox": [ + 173, + 433, + 823, + 462 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof of the right hand side of (13): We also know that $\\begin{array} { r } { ( b _ { k } = \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } ) } \\end{array}$ ", + "bbox": [ + 173, + 470, + 683, + 493 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/881fb37a93a36ca33279a8cb7352a55121d9d63aa62473943680b180547b1336.jpg", + "text": "$$\n\\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } \\leq \\sum _ { i = 1 } ^ { k } \\alpha _ { k } ^ { i } b _ { i } \\leq \\frac { 2 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } .\n$$", + "text_format": "latex", + "bbox": [ + 339, + 496, + 655, + 539 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "It is implied by (16) that ", + "bbox": [ + 173, + 540, + 336, + 555 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/d2bbc005947d28dd95e009c07818fe62e6d9b1e5309a47331c5537b6fbe06992.jpg", + "text": "$$\n\\begin{array} { r l } & { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\gamma \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\frac { 3 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( t ) } { t } } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( \\hat { V } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 243, + 558, + 754, + 643 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/53eab3d1f26c37eff9f588636d51ccb07f492de053eae442ba9c0192a7489963.jpg", + "text": "$$\n\\leq \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + \\beta _ { t } .\n$$", + "text_format": "latex", + "bbox": [ + 356, + 660, + 658, + 702 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Note that $\\beta _ { t } = c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { H \\iota ( t ) / t }$ ; $c _ { 3 } = 3 c _ { 2 } = 1 2 { \\sqrt { 2 } } .$ . ", + "bbox": [ + 171, + 704, + 560, + 723 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof of the left hand side of (13): Now, we assume that event that (16) holds. We assert that $Q _ { p } \\geq Q ^ { * }$ for all $( s , a )$ and $p \\leq p ^ { \\prime }$ . This assertion is obviously true when $p ^ { \\prime } = 0$ . Then ", + "bbox": [ + 173, + 727, + 825, + 757 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/70ff20e4e31a1bb0299ebb34008753b3eaca5c5e3da766015eda50bc514d5a51.jpg", + "text": "$$\n\\begin{array} { c l } { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\geq - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ { \\geq \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } b _ { i } - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| \\geq 0 . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 266, + 758, + 730, + 847 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Therefore the assertion holds for $p ^ { \\prime } + 1$ as well. By induction, it holds for all $p$ . ", + "bbox": [ + 173, + 847, + 694, + 863 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We now see that (13) holds for probability $1 - \\delta / 2$ for all $p , s , a$ . Since $\\hat { Q } _ { p } ( s , a )$ is always greater than $Q _ { p ^ { \\prime } } ( s , a )$ for some $p ^ { \\prime } \\leq p$ , we know that $\\hat { Q } _ { p } ( s , a ) \\geq Q _ { p ^ { \\prime } } ( s , a ) \\geq Q ^ { * } ( s , a )$ , thus proving (14). ", + "bbox": [ + 173, + 869, + 825, + 905 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We now give a proof for lemma 2. Recall the definition for a $( C , w )$ -sequence. A sequence $( w _ { t } ) _ { t \\geq 1 }$ is said to be a $( C , w )$ -sequence for $C , w > 0$ , if $0 \\leq w _ { t } \\leq w$ for all $t \\geq 1$ , and $\\textstyle \\sum _ { t \\geq 1 } w _ { t } \\leq C$ . ", + "bbox": [ + 169, + 102, + 825, + 135 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proof. Let $n _ { t } = N _ { t } ( s _ { t } , a _ { t } )$ for simplicity; we have ", + "bbox": [ + 173, + 148, + 511, + 165 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/56de998ef0518be48301779db9712c42690034bbbfeba7d7ae47e786453e29e7.jpg", + "text": "$$\n\\begin{array} { r l } { { \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } ( Q _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } [ \\cfrac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } + \\beta _ { n _ { t } } + \\gamma \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } ( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } ) ( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } ) ] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 258, + 169, + 736, + 285 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "The last inequality is due to lemma 4. Note that $\\alpha _ { n _ { t } } ^ { 0 } = \\mathbb { I } [ n _ { t } = 0 ]$ , the first term in the summation can be bounded by, ", + "bbox": [ + 173, + 290, + 825, + 320 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/fb8a30f2f69b0a0ef551b8c5d16a0c30b4272df4f37761a5fa9bb9799022d5bd.jpg", + "text": "$$\n\\sum _ { t \\geq 1 } w _ { t } \\frac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } \\leq \\frac { S A w } { 1 - \\gamma } .\n$$", + "text_format": "latex", + "bbox": [ + 418, + 323, + 578, + 366 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For the second term, define $u ( s , a ) = \\operatorname* { s u p } _ { t } N _ { t } ( s , a )$ .2 It follows that, ", + "bbox": [ + 173, + 377, + 624, + 395 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/bce9df1dbdecfd18cd0fbd700e55e9d74ea6aeee786233ea436d7b29badc95fc.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } \\beta _ { n _ { t } } = \\sum _ { s , o } ^ { u ( s , o ) } w _ { \\tau ( s , a , i ) } \\beta _ { i } } \\\\ { \\displaystyle \\qquad \\leq \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sum _ { i = 1 } ^ { C _ { s , a } / w } \\sqrt { \\frac { H \\iota ( i ) } { i } } w } \\\\ { \\displaystyle \\qquad \\leq 2 \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sqrt { \\iota ( C ) H C _ { s , a } w } } \\\\ { \\displaystyle \\qquad \\leq 2 c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { w S A H C \\iota ( C ) } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 338, + 400, + 656, + 553 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Where $\\begin{array} { r } { C _ { s , a } = \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } ) = ( s , a ) } w _ { t } } \\end{array}$ . Inequality (19) follows from rearrangement inequality, since $\\iota ( x ) / x$ is monotonically decreasing. Inequality (21) follows from Jensen’s inequality. ", + "bbox": [ + 173, + 555, + 826, + 588 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For the third term of the summation, we have ", + "bbox": [ + 173, + 593, + 472, + 608 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/a3e0c01b442b6f3c208a66bd1e6c84604cad0908954b7748b2f4faaaf3ab9889.jpg", + "text": "$$\n\\begin{array} { c } { { \\displaystyle \\sum _ { t \\ge 1 } w _ { t } \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } \\left( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } \\right) \\left( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } \\right) } } \\\\ { { \\displaystyle \\le \\sum _ { t ^ { \\prime } \\ge 1 } \\left( \\hat { V } _ { t ^ { \\prime } } - V ^ { * } \\right) \\left( s _ { t ^ { \\prime } + 1 } \\right) \\left( \\sum _ { t = t ^ { \\prime } + 1 } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 323, + 612, + 676, + 723 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Define ", + "bbox": [ + 173, + 746, + 220, + 760 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/90d6cf339da0fc0bd898216b68464ffc95f01fd8012796a3a8c23865fc73b784.jpg", + "text": "$$\nw _ { t ^ { \\prime } + 1 } ^ { \\prime } = \\left( \\sum _ { \\stackrel { t = t ^ { \\prime } + 1 } { ( s _ { t } , a _ { t } ) = ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) } } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 380, + 755, + 616, + 821 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We claim that $w _ { t + 1 } ^ { \\prime }$ is a $\\textstyle ( C , ( 1 + { \\frac { 1 } { H } } ) w )$ -sequence. We now prove this claim. By lemma 3, for any $t ^ { \\prime } \\geq 0$ , ", + "bbox": [ + 173, + 832, + 823, + 861 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/4253341d3793835fab3da24d7c177c2e36a4c25bd994300d52241f505a8a3adc.jpg", + "text": "$$\nw _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq w \\sum _ { j = n _ { t ^ { \\prime } } } ^ { \\infty } \\alpha _ { j } ^ { n _ { t ^ { \\prime } } } = ( 1 + 1 / H ) w .\n$$", + "text_format": "latex", + "bbox": [ + 370, + 858, + 625, + 901 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "By $\\textstyle \\sum _ { j = 0 } ^ { i } \\alpha _ { i } ^ { j } = 1$ , we have $\\begin{array} { r } { \\sum _ { t ^ { \\prime } \\geq 1 } w _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq \\sum _ { t \\geq 1 } w _ { t } \\leq C . } \\end{array}$ . This proves the assertion. It follows from (22) that ", + "bbox": [ + 169, + 102, + 825, + 133 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/a1c63f8ce8866bfa935cc54ba0fa4060644052edbd98620aad79e7ef97f0f148.jpg", + "text": "$$\n\\begin{array} { r l } & { \\quad \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { = \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - \\hat { V } _ { t + 1 } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( 2 \\alpha _ { n _ { t } + 1 } \\frac { 1 } { 1 - \\gamma } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) \\left( s _ { t + 1 } , a _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 276, + 138, + 715, + 337 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Inequality (25) comes from the update rule of our algorithm. Inequality (26) comes from the fact that $\\dot { \\alpha } _ { t } = ( H + 1 ) / ( H + t ) \\overset { \\cdot } { \\leq } H / t$ and Jensen’s Inequality. More specifically, let $C _ { s , a } ^ { \\prime } =$ $\\scriptstyle \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } = s , a } w _ { t + 1 } ^ { \\prime }$ , $w ^ { \\prime } = w ( 1 + 1 / H )$ . Then ", + "bbox": [ + 176, + 340, + 821, + 387 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/9834ce9a1e7dc58cf0a708f75e6b5b858bcdc589ad24752ed45e1722a959c115.jpg", + "text": "$$\n\\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\alpha _ { n _ { t } + 1 } \\leq \\sum _ { s , a } \\sum _ { n = 1 } ^ { C _ { s , a } ^ { \\prime } / w ^ { \\prime } } w ^ { \\prime } \\frac { H } { n } \\leq \\sum _ { s , a } H w ^ { \\prime } \\ln ( C _ { s , a } ^ { \\prime } / w ) \\leq 2 S A H w \\ln C .\n$$", + "text_format": "latex", + "bbox": [ + 241, + 392, + 754, + 443 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Putting (18), (21) and (27) together, we have, ", + "bbox": [ + 173, + 454, + 472, + 469 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/fca33e0caf6dada43c874ac741c299779e61219671c7f66b39e75e1f2d6533ae.jpg", + "text": "$$\n\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w S A H C \\iota ( C ) } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln C \\right) + \\gamma \\displaystyle \\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) ( s _ { t + 1 } , a _ { t + 1 } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 191, + 491, + 779, + 571 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Observe that the third term is another weighted sum with the same form as (17). Therefore, we can unroll this term repetitively with changing weight sequences.Suppose that our original weight sequence is also denoted by $\\{ \\bar { w _ { t } ^ { ( 0 ) } } \\} _ { t \\ge 1 }$ , while $\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }$ denotes the weight sequence after unrolling for $k$ times. Let $w ^ { ( k ) }$ be $w \\cdot \\left( 1 + 1 / H \\right) ^ { k }$ . Then we can see that $\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }$ is a $( C , w ^ { ( k ) } )$ -sequence. Suppose that we unroll for $H$ times. Then ", + "bbox": [ + 173, + 575, + 828, + 654 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/4b3ee4cb41b96d3c4f223bc4146fc59bb343a3cac7f567861e3ed3154212ede1.jpg", + "text": "$$\n\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ^ { ( H ) } \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\frac { C } { 1 - \\gamma } . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 204, + 659, + 795, + 779 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We set $\\begin{array} { r } { H = \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma } , } \\end{array}$ . It follows that $w ^ { ( H ) } = ( 1 + 1 / H ) ^ { H } w ^ { ( 0 ) } \\leq e w ^ { ( 0 ) }$ , and that γH C1−γ $\\gamma ^ { H } \\frac { C } { 1 - \\gamma } \\leq C \\epsilon _ { 1 }$ . Also, let $\\ell ( C ) = \\iota ( C ) \\ln ( ( 1 - \\gamma ) ^ { - 1 } \\epsilon _ { 1 } ^ { - 1 } )$ . Therefore, ", + "bbox": [ + 173, + 799, + 826, + 839 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/11b4d3a7f4cf29588edd0397db246aa8a18f798a841609b619f1a546acf17d46.jpg", + "text": "$$\n\\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln C \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 184, + 861, + 787, + 904 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "C EXTENSION TO OTHER SETTINGS ", + "text_level": 1, + "bbox": [ + 176, + 102, + 482, + 118 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "First we define a mapping from a finite horizon MDP to an infinite horizon MDP so that our algorithm can be applied. For an arbitrary finite horizon MDP $\\mathcal { M } = ( S , A , H , r _ { h } ( s _ { \\perp } a ) , p _ { h } ( s ^ { \\prime } \\mid s , a ) )$ where $H$ is the length of episode, the corresponding infinite horizon MDP $\\mathcal { M } = ( S , A , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )$ is defined as, ", + "bbox": [ + 174, + 132, + 826, + 189 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "$\\begin{array} { l } { { \\bullet \\ { \\bar { S } } = S \\times H , { \\bar { A } } = A ; } } \\\\ { { \\bullet \\ \\gamma = ( 1 - 1 / H ) ; } } \\end{array}$ • for a state $s$ at step $h$ , let $\\bar { s } _ { s , h }$ be the corresponding state. For any action $a$ and next state $s ^ { \\prime }$ , define $\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )$ and $\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s ^ { \\prime } \\mid s , h )$ . And for $h = H$ , set $\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0$ and $\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]$ for a fixed starting state $s _ { 1 }$ . ", + "bbox": [ + 214, + 198, + 825, + 285 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Let $\\bar { V } _ { t }$ be the value function in $\\bar { \\mathcal { M } }$ at time $t$ and $V _ { h } ^ { k }$ the value function in $\\mathcal { M }$ at episode $k$ , step $h$ . It follows that $\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V _ { 1 } ^ { \\ast } ( s _ { 1 } ) } \\end{array}$ . And the policy mapping is defined as $\\pi _ { h } ( s ) = \\bar { \\pi } ( \\bar { s } _ { s , h } )$ for policy $\\bar { \\pi }$ in $\\bar { \\mathcal { M } }$ . Value functions in MDP $\\mathcal { M }$ and $\\bar { \\mathcal { M } }$ are closely related in a sense that, any $\\epsilon$ -optimal policy $\\bar { \\pi }$ of $\\bar { \\mathcal { M } }$ corresponding to an $( \\epsilon / \\gamma ^ { H } )$ -optimal policy $\\pi$ in $\\mathcal { M }$ (see section C.1 for proof). Note that here $\\gamma ^ { H } = ( 1 - \\dot { 1 } / H ) ^ { \\tilde { H } } = \\mathcal { O } ( \\dot { 1 } )$ is a constant. ", + "bbox": [ + 173, + 294, + 826, + 375 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "For any visited a $\\epsilon > 0$ g our algorithm on times, and at most $\\bar { M }$ for of $\\tilde { \\mathcal O } \\big ( \\frac { 3 S A H ^ { 9 } } { \\epsilon ^ { 2 } } \\big )$ tt me steps, the starting state -optimal. If we select the p $s _ { 1 }$ iscy $\\tilde { \\mathcal O } \\big ( { \\textstyle { \\frac { 3 S A H ^ { 8 } } { \\epsilon ^ { 2 } } } } \\big )$ $1 / 3$ $\\epsilon$ uniformly randomly from the policy $\\pi ^ { t H + 1 }$ for $0 \\leq t < T / H$ , with probability at least $2 / 3$ we can get an $\\epsilon$ -optimal policy. Therefore the PAC sample complexity is $\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)$ after hiding $S , A , H$ terms. ", + "bbox": [ + 173, + 381, + 826, + 450 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "On the other hand, we want to show that for any $K$ episodes, ", + "bbox": [ + 173, + 455, + 573, + 470 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/dfeb0feb59d694c97b8d3ad8daa0ee95cb21f9b5f8a16f568ba86cb50b2a11bd.jpg", + "text": "$$\n\\mathrm { R e g r e t } ( T ) = \\sum _ { k = 1 } ^ { T / H } \\left[ V ^ { * } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } ) \\right] \\propto T ^ { 1 / 2 } .\n$$", + "text_format": "latex", + "bbox": [ + 341, + 476, + 655, + 522 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The reason why our algorithm can have a better reduction from regret to PAC is that, after choosing $\\epsilon _ { 1 }$ , it follows from the argument of theorem 1 that for all $\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )$ , the number of $\\epsilon _ { 2 }$ -suboptimal steps is bounded by ", + "bbox": [ + 173, + 534, + 826, + 578 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/fc17daba551915b444155b637b9972b89541ce4691b2e2c14706d565e8c89847.jpg", + "text": "$$\n\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right)\n$$", + "text_format": "latex", + "bbox": [ + 348, + 583, + 648, + 618 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "with probability $1 - \\delta$ . In contrast, delayed Q-learning can only give an upper bound on $\\epsilon _ { 1 }$ -suboptimal steps after setting parameter $\\epsilon _ { 1 }$ . ", + "bbox": [ + 174, + 623, + 825, + 652 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Formally, let $X _ { k } = V ^ { \\ast } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } )$ be the regret of $k$ -th episode. For any $T$ , set $\\epsilon = \\sqrt { S A / T }$ and $\\epsilon _ { 2 } = \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )$ . Let $\\begin{array} { r } { M = \\lceil \\log _ { 2 } \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } \\rceil } \\end{array}$ . It follows that, ", + "bbox": [ + 171, + 659, + 825, + 695 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/1d7fa9ef3cc29e1e3ac2d0a868dec889d151fa6dfb8e0d2bbd8979285d9e7f25.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\operatorname { R e g r e t } ( T ) \\le T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } ( | k : \\{ X _ { k } \\ge \\epsilon _ { 2 } \\cdot 2 ^ { i - 1 } \\} | ) \\epsilon _ { 2 } \\cdot 2 ^ { i } } } \\\\ & { } & \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } \\frac { S A \\ln 1 / \\delta } { \\epsilon _ { 2 } \\cdot 2 ^ { i - 2 } } ) } \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( \\sqrt { S A T } \\ln 1 / \\delta ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 312, + 702, + 686, + 818 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "with probability $1 - \\delta$ . Note that the $\\tilde { \\mathcal { O } }$ notation hides the poly $( 1 / ( 1 - \\gamma ) , \\log { 1 / \\epsilon _ { 1 } } )$ which is, by our reduction, poly $( H , \\log T , \\log S , \\log A )$ . ", + "bbox": [ + 173, + 823, + 823, + 854 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C.1 CONNECTION BETWEEN VALUE FUNCTIONS ", + "text_level": 1, + "bbox": [ + 173, + 869, + 521, + 883 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Recall that our MDP mapping from $\\begin{array} { r c l } { \\mathcal { M } } & { = } & { ( S , A , H , r _ { h } ( s , a ) , p _ { h } ( s ^ { \\prime } \\mathrm { ~ ~ \\chi ~ } | \\mathrm { ~ ~ \\chi ~ } s , a ) ) } \\end{array}$ to $\\begin{array} { r l } { \\bar { \\mathcal { M } } } & { { } = } \\end{array}$ $( \\bar { S } , \\bar { A } , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )$ is defined as, ", + "bbox": [ + 173, + 893, + 825, + 924 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "• $\\bar { S } = S \\times H , \\bar { A } = A$ ; \n• $\\gamma = ( 1 - 1 / H )$ ; \n• for a state $s$ at step $h$ , let $\\bar { s } _ { s , h }$ be the corresponding state. For any action $a$ and next state $s ^ { \\prime }$ , define $\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )$ and $\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s , h )$ . And for $h = H$ , set $\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0$ and $\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]$ for a fixed starting state $s _ { 1 }$ . ", + "bbox": [ + 214, + 102, + 830, + 186 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "For a trajectory $\\left\\{ { \\left( { \\bar { s } } _ { s _ { 1 } , 1 } , { \\bar { a } } _ { 1 } \\right) } , { \\left( { \\bar { s } } _ { s _ { 2 } , 2 } , { \\bar { a } } _ { 2 } \\right) } , \\cdot \\cdot \\cdot \\right\\}$ in $\\bar { \\mathcal { M } }$ , let $\\{ ( s _ { 1 } , a _ { 1 } ) , ( s _ { 2 } , a _ { 2 } ) , \\cdot \\cdot \\cdot \\}$ be the corresponding trajectory in $\\mathcal { M }$ . Note that $\\mathcal { M }$ has a unique fixed starting state $s _ { 1 }$ , which means that $s _ { t H + 1 } = s _ { 1 }$ for all $t \\geq 0$ . Denote the corresponding policy of $\\bar { \\pi } ^ { t }$ as $\\pi ^ { t }$ (may be non-stationary), then we have ", + "bbox": [ + 176, + 195, + 828, + 239 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/52704ae56896c370e8ec3e398577b25911ebb487e2237bb93613270924e4d24c.jpg", + "text": "$$\n\\begin{array} { r l } & { \\bar { \\gamma } ^ { \\pi ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\mathbb { E } \\left[ \\bar { r } ( \\bar { s } _ { s _ { 1 } , 1 } , \\bar { a } _ { 1 } ) + \\gamma \\bar { r } ( \\bar { s } _ { s _ { 2 } , 2 } , \\bar { a } _ { 2 } ) + \\cdot \\cdot \\cdot + \\gamma ^ { H - 1 } \\bar { r } ( \\bar { s } _ { s _ { H - 1 } , H - 1 } , \\bar { a } _ { H - 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H - 1 } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } \\mathbb { E } \\left[ r _ { 1 } ( s _ { 1 } , a _ { 1 } ) + r _ { 2 } ( s _ { 2 } , a _ { 2 } ) + \\cdot \\cdot \\cdot + r _ { H - 1 } ( s _ { H - 1 } , a _ { H - 1 } ) + \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi + H } ( \\bar { s } _ { s _ { 1 } , 1 } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 244, + 839, + 309 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Then for a stationary policy $\\bar { \\pi }$ , we can conclude $\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\pi } ( s _ { 1 } ) } \\end{array}$ . Since the optimal policy $\\bar { \\pi } ^ { * }$ is stationary, we have $\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\ast } ( s _ { 1 } ) } \\end{array}$ . ", + "bbox": [ + 174, + 314, + 825, + 358 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "By definition, $\\bar { \\pi }$ is $\\epsilon$ -optimal at time step $t$ means that ", + "bbox": [ + 173, + 362, + 526, + 377 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/f086319de63ecc59068ffa365e9e62e72d387018083ec5144b0aa425136c465f.jpg", + "text": "$$\n\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 401, + 382, + 594, + 404 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "It follows that ", + "bbox": [ + 173, + 409, + 267, + 422 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/5cce7a222551bc6274bbc52ffdb272aeca3dced98abbbc8f80d2fab7858deadf.jpg", + "text": "$$\n\\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon ,\n$$", + "text_format": "latex", + "bbox": [ + 289, + 424, + 705, + 445 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "hence ", + "bbox": [ + 173, + 450, + 215, + 464 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/a3a5a1e7d68659f678d45d1177dbe36c95439ce8dceba0366c8e7ffef18cce5d.jpg", + "text": "$$\n^ { H } V ^ { \\pi ^ { \\varepsilon } } ( s _ { 1 } ) \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) + \\gamma ^ { H } ( \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) ) - \\epsilon \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon .\n$$", + "text_format": "latex", + "bbox": [ + 181, + 468, + 823, + 488 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Therefore we have ", + "bbox": [ + 174, + 492, + 299, + 507 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/1f1b44222a0c1087458d2ccf43ee9da0e84754c0602268ed6cbc91b7c6c5c549.jpg", + "text": "$$\n{ V ^ { \\pi } } ^ { t } ( s _ { 1 } ) \\ge \\frac { 1 - \\gamma ^ { H } } { \\gamma ^ { H } } \\bar { V } ^ { * } ( \\bar { s } _ { s 1 , 1 } ) - \\epsilon / \\gamma ^ { H } = V ^ { * } ( s _ { 1 } ) - \\epsilon / \\gamma ^ { H } ,\n$$", + "text_format": "latex", + "bbox": [ + 305, + 510, + 691, + 544 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "which means that $\\pi ^ { t }$ is an $( \\epsilon / \\gamma ^ { H } )$ -optimal policy. ", + "bbox": [ + 173, + 549, + 500, + 566 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D A HARD INSTANCE FOR DELAYED Q-LEARNING ", + "text_level": 1, + "bbox": [ + 173, + 583, + 607, + 602 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "In this section, we prove Theorem 2 regarding the performance of Delayed Q-learning. ", + "bbox": [ + 176, + 616, + 745, + 632 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Theorem 2. There exists a family of MDPs with constant $S$ and $A$ , in which with probability $1 - \\delta$ , Delayed $Q$ -learning incurs sample complexity of exploration of $\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)$ , assuming that $\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }$ . ", + "bbox": [ + 173, + 633, + 826, + 688 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/d5ea7466679255528080a8e623d46ab5efe02c60983c67939e2c04d8974b3167.jpg", + "image_caption": [ + "Figure 1: The MDP family. Actions are denoted by arrows. Actions with red color have reward 1, and reward 0 otherwise. " + ], + "image_footnote": [], + "bbox": [ + 397, + 705, + 599, + 829 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Proof. For each $0 < \\epsilon < \\frac { 1 } { 1 0 }$ , consider the following MDP (see also Fig. 1): state space is ${ \\boldsymbol { s } } =$ $\\{ a , b , c \\}$ while action set is $\\overset { \\vartriangle } { \\mathcal { A } } = \\{ \\ v { x } , \\ v { y } \\}$ ; transition probabilities are $P ( b | a , y ) = 1 - 1 0 \\epsilon$ , $P ( c | a , y ) =$ 10\u000f, $P ( b | a , x ) = 1$ , $P ( a | b , \\cdot ) = P ( a | c , \\cdot ) = 1$ . Rewards are all 1, except $R ( c , \\cdot ) = 0$ . ", + "bbox": [ + 174, + 881, + 825, + 925 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Assume that Delayed Q-learning is called for this MDP starting from state $a$ , with discount $\\begin{array} { r } { \\gamma > \\frac { 1 } { 2 } } \\end{array}$ and precision set as $\\epsilon$ . Denote the $Q$ value maintained by the algorithm by $\\hat { Q }$ . Without loss of generality, assume that the initial tie-breaking favors action $y$ when comparing $\\hat { Q } ( a , x )$ and $\\hat { Q } ( a , y )$ . In that case, unless $\\hat { Q } ( a , y )$ is updated, the agent will always choose $y$ in state $a$ . Since $Q ( a , x ) - Q ( a , y ) = 1 0 \\epsilon \\gamma > \\epsilon$ for any policy, choosing $y$ at state $a$ implies that the timestep is not $\\epsilon$ -optimal. In other words, sample complexity for exploration is at least the number of times the agent visits $a$ before the first update of $\\hat { Q } ( a , y )$ . ", + "bbox": [ + 173, + 102, + 825, + 213 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In the Delayed Q-learning algorithm, $\\hat { Q } ( \\cdot , \\cdot )$ are initialized to $1 / ( 1 - \\gamma )$ . Therefore, $\\hat { Q } ( a , y )$ could only be updated if max $\\hat { Q } ( c , \\cdot )$ is updated (and becomes smaller than $1 / ( 1 - \\gamma ) )$ . According to the algorithm, this can only happen if $c$ is visited $\\begin{array} { r } { m = \\Omega \\left( \\frac { 1 } { \\epsilon ^ { 2 } } \\right) } \\end{array}$ times. ", + "bbox": [ + 174, + 219, + 826, + 268 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "However, each time the agent visits $a$ , there is less than $1 0 \\epsilon$ probability of transiting to $c$ . Let $t _ { 0 } = m / ( 1 0 \\epsilon C )$ , where $\\begin{array} { r } { C ^ { ' } = 3 \\ln \\frac { 1 } { \\delta } + 1 } \\end{array}$ . $\\delta$ is chosen such that $C \\leq m$ . In the first $2 t _ { 0 }$ timesteps, $a$ will be visited $t _ { 0 }$ times. By Chernoff’s bound, with probability $1 - \\delta$ , state $c$ will be visited less than $m$ times. In that case, $\\hat { Q } ( a , y )$ will not be updated in the first $2 t _ { 0 }$ timesteps. Therefore, with probability $1 - \\delta$ , sample complexity of exploration is at least ", + "bbox": [ + 173, + 273, + 825, + 345 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/446057ad76737ed78dbeb6c14c391ec04ae8a21d0dea27c95bf135aaaf7cc43f.jpg", + "text": "$$\nt _ { 0 } = \\Omega \\left( { \\frac { 1 } { \\epsilon ^ { 3 } \\left( \\ln { 1 / \\delta } \\right) } } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 419, + 352, + 578, + 387 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "When $\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }$ , it can be seen that $\\begin{array} { r } { C = 3 \\ln \\frac { 1 } { \\delta } + 1 < \\frac { 4 } { \\epsilon ^ { 2 } } < m } \\end{array}$ ", + "bbox": [ + 173, + 393, + 620, + 411 + ], + "page_idx": 17 + } +] \ No newline at end of file diff --git a/parse/train/BkglSTNFDB/BkglSTNFDB_middle.json b/parse/train/BkglSTNFDB/BkglSTNFDB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..1112d3bc82a64774b1a91fc9feb603b851a7f818 --- /dev/null +++ b/parse/train/BkglSTNFDB/BkglSTNFDB_middle.json @@ -0,0 +1,60344 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "score": 1.0, + "content": "Q-LEARNING WITH UCB EXPLORATION IS SAMPLE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 411, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 411, + 117 + ], + "score": 1.0, + "content": "EFFICIENT FOR INFINITE-HORIZON MDP", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 134, + 340, + 179 + ], + "lines": [ + { + "bbox": [ + 112, + 134, + 243, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 134, + 243, + 147 + ], + "score": 1.0, + "content": "Kefan Dong*, Yuanhao Wang∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 145, + 320, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 320, + 159 + ], + "score": 1.0, + "content": "Institute for Interdisciplinary Information Sciences,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 156, + 196, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 196, + 169 + ], + "score": 1.0, + "content": "Tsinghua University", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 168, + 341, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 341, + 181 + ], + "score": 1.0, + "content": "{dkf16,yuanhao-16}@mails.tsinghua.edu.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 113, + 196, + 370, + 240 + ], + "lines": [ + { + "bbox": [ + 112, + 196, + 171, + 208 + ], + "spans": [ + { + "bbox": [ + 112, + 196, + 171, + 208 + ], + "score": 1.0, + "content": "Xiaoyu Chen", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 206, + 371, + 219 + ], + "spans": [ + { + "bbox": [ + 111, + 206, + 371, + 219 + ], + "score": 1.0, + "content": "Key Laboratory of Machine Perception, MOE, School of EECS,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 217, + 187, + 231 + ], + "spans": [ + { + "bbox": [ + 111, + 217, + 187, + 231 + ], + "score": 1.0, + "content": "Peking University", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 112, + 229, + 210, + 241 + ], + "spans": [ + { + "bbox": [ + 112, + 229, + 210, + 241 + ], + "score": 1.0, + "content": "cxy30@pku.edu.cn", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 112, + 257, + 368, + 301 + ], + "lines": [ + { + "bbox": [ + 111, + 255, + 167, + 271 + ], + "spans": [ + { + "bbox": [ + 111, + 255, + 167, + 271 + ], + "score": 1.0, + "content": "Liwei Wang", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 112, + 268, + 368, + 280 + ], + "spans": [ + { + "bbox": [ + 112, + 268, + 368, + 280 + ], + "score": 1.0, + "content": "Key Laboratory of Machine Perception, MOE, School of EECS", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 111, + 277, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 111, + 277, + 287, + 291 + ], + "score": 1.0, + "content": "Center for Data Science, Peking University", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 111, + 290, + 241, + 302 + ], + "spans": [ + { + "bbox": [ + 111, + 290, + 241, + 302 + ], + "score": 1.0, + "content": "wanglw@cis.pku.edu.cn", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 278, + 330, + 333, + 342 + ], + "lines": [ + { + "bbox": [ + 276, + 329, + 336, + 344 + ], + "spans": [ + { + "bbox": [ + 276, + 329, + 336, + 344 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 143, + 357, + 469, + 475 + ], + "lines": [ + { + "bbox": [ + 141, + 357, + 470, + 370 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 470, + 370 + ], + "score": 1.0, + "content": "A fundamental question in reinforcement learning is whether model-free algo-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 367, + 469, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 469, + 381 + ], + "score": 1.0, + "content": "rithms are sample efficient. Recently, Jin et al. (2018) proposed a Q-learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 379, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 379, + 469, + 392 + ], + "score": 1.0, + "content": "algorithm with UCB exploration policy, and proved it has nearly optimal regret", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 390, + 469, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 390, + 469, + 403 + ], + "score": 1.0, + "content": "bound for finite-horizon episodic MDP. In this paper, we adapt Q-learning with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 401, + 469, + 413 + ], + "spans": [ + { + "bbox": [ + 142, + 401, + 469, + 413 + ], + "score": 1.0, + "content": "UCB-exploration bonus to infinite-horizon MDP with discounted rewards without", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 412, + 469, + 426 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 469, + 426 + ], + "score": 1.0, + "content": "accessing a generative model. We show that the sample complexity of exploration", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 139, + 422, + 471, + 443 + ], + "spans": [ + { + "bbox": [ + 139, + 422, + 272, + 443 + ], + "score": 1.0, + "content": "of our algorithm is bounded by", + "type": "text" + }, + { + "bbox": [ + 273, + 423, + 323, + 439 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 422, + 471, + 443 + ], + "score": 1.0, + "content": ". This improves the previously best", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 136, + 432, + 474, + 460 + ], + "spans": [ + { + "bbox": [ + 136, + 432, + 206, + 460 + ], + "score": 1.0, + "content": "known result of", + "type": "text" + }, + { + "bbox": [ + 207, + 438, + 257, + 454 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 432, + 474, + 460 + ], + "score": 1.0, + "content": "in this setting achieved by delayed Q-learning (Strehl", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 452, + 469, + 466 + ], + "spans": [ + { + "bbox": [ + 141, + 452, + 362, + 466 + ], + "score": 1.0, + "content": "et al., 2006), and matches the lower bound in terms of", + "type": "text" + }, + { + "bbox": [ + 362, + 455, + 367, + 463 + ], + "score": 0.73, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 452, + 410, + 466 + ], + "score": 1.0, + "content": "as well as", + "type": "text" + }, + { + "bbox": [ + 419, + 452, + 436, + 466 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 446, + 452, + 469, + 466 + ], + "score": 1.0, + "content": "up to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 464, + 222, + 476 + ], + "spans": [ + { + "bbox": [ + 141, + 464, + 222, + 476 + ], + "score": 1.0, + "content": "logarithmic factors.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19.5 + }, + { + "type": "title", + "bbox": [ + 108, + 501, + 206, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 208, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 208, + 516 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "The goal of reinforcement learning (RL) is to construct efficient algorithms that learn and plan in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "sequential decision making tasks when the underlying system dynamics are unknown. A typical", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "score": 1.0, + "content": "model in RL is Markov Decision Process (MDP). At each time step, the environment is in a state", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 563, + 113, + 570 + ], + "score": 0.52, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 561, + 221, + 573 + ], + "score": 1.0, + "content": ". The agent takes an action", + "type": "text" + }, + { + "bbox": [ + 221, + 563, + 227, + 570 + ], + "score": 0.7, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 561, + 295, + 573 + ], + "score": 1.0, + "content": ", obtain a reward", + "type": "text" + }, + { + "bbox": [ + 295, + 563, + 301, + 571 + ], + "score": 0.75, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 561, + 506, + 573 + ], + "score": 1.0, + "content": ", and then the environment transits to another state.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "In reinforcement learning, the transition probability distribution is unknown. The algorithm needs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "to learn the transition dynamics of MDP, while aiming to maximize the cumulative reward. This", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "poses the exploration-exploitation dilemma: whether to act to gain new information (explore) or to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 605, + 375, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 375, + 617 + ], + "score": 1.0, + "content": "act consistently with past experience to maximize reward (exploit).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 709 + ], + "lines": [ + { + "bbox": [ + 105, + 620, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 635 + ], + "score": 1.0, + "content": "Theoretical analyses of reinforcement learning fall into two broad categories: those assuming a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "simulator (a.k.a. generative model), and those without a simulator. In the first category, the algorithm", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "is allowed to query the outcome of any state action pair from an oracle. The emphasis is on the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 265, + 667 + ], + "score": 1.0, + "content": "number of calls needed to estimate the", + "type": "text" + }, + { + "bbox": [ + 265, + 655, + 275, + 666 + ], + "score": 0.82, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "value or to output a near-optimal policy. There has been", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "extensive research in literature following this line of research, the majority of which focuses on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 675, + 507, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 507, + 689 + ], + "score": 1.0, + "content": "discounted infinite horizon MDPs (Azar et al., 2011; Even-Dar & Mansour, 2003; Sidford et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "score": 1.0, + "content": "2018b). The current results have achieved near-optimal time and sample complexities (Sidford et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 697, + 147, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 147, + 711 + ], + "score": 1.0, + "content": "2018b;a).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 262, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 263, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 263, + 734 + ], + "score": 1.0, + "content": "∗These two authors contributed equally.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 97 + ], + "score": 1.0, + "content": "Q-LEARNING WITH UCB EXPLORATION IS SAMPLE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 411, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 411, + 117 + ], + "score": 1.0, + "content": "EFFICIENT FOR INFINITE-HORIZON MDP", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 112, + 134, + 340, + 179 + ], + "lines": [ + { + "bbox": [ + 112, + 134, + 243, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 134, + 243, + 147 + ], + "score": 1.0, + "content": "Kefan Dong*, Yuanhao Wang∗", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 145, + 320, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 320, + 159 + ], + "score": 1.0, + "content": "Institute for Interdisciplinary Information Sciences,", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 156, + 196, + 169 + ], + "spans": [ + { + "bbox": [ + 111, + 156, + 196, + 169 + ], + "score": 1.0, + "content": "Tsinghua University", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 113, + 168, + 341, + 181 + ], + "spans": [ + { + "bbox": [ + 113, + 168, + 341, + 181 + ], + "score": 1.0, + "content": "{dkf16,yuanhao-16}@mails.tsinghua.edu.", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 196, + 171, + 208 + ], + "spans": [ + { + "bbox": [ + 112, + 196, + 171, + 208 + ], + "score": 1.0, + "content": "Xiaoyu Chen", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 206, + 371, + 219 + ], + "spans": [ + { + "bbox": [ + 111, + 206, + 371, + 219 + ], + "score": 1.0, + "content": "Key Laboratory of Machine Perception, MOE, School of EECS,", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 217, + 187, + 231 + ], + "spans": [ + { + "bbox": [ + 111, + 217, + 187, + 231 + ], + "score": 1.0, + "content": "Peking University", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 229, + 210, + 241 + ], + "spans": [ + { + "bbox": [ + 112, + 229, + 210, + 241 + ], + "score": 1.0, + "content": "cxy30@pku.edu.cn", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 255, + 167, + 271 + ], + "spans": [ + { + "bbox": [ + 111, + 255, + 167, + 271 + ], + "score": 1.0, + "content": "Liwei Wang", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 268, + 368, + 280 + ], + "spans": [ + { + "bbox": [ + 112, + 268, + 368, + 280 + ], + "score": 1.0, + "content": "Key Laboratory of Machine Perception, MOE, School of EECS", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 277, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 111, + 277, + 287, + 291 + ], + "score": 1.0, + "content": "Center for Data Science, Peking University", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 290, + 241, + 302 + ], + "spans": [ + { + "bbox": [ + 111, + 290, + 241, + 302 + ], + "score": 1.0, + "content": "wanglw@cis.pku.edu.cn", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + } + ], + "index": 3.5, + "bbox_fs": [ + 111, + 134, + 341, + 181 + ] + }, + { + "type": "list", + "bbox": [ + 113, + 196, + 370, + 240 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 111, + 196, + 371, + 241 + ], + "lines_deleted": true + }, + { + "type": "list", + "bbox": [ + 112, + 257, + 368, + 301 + ], + "lines": [], + "index": 11.5, + "bbox_fs": [ + 111, + 255, + 368, + 302 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 278, + 330, + 333, + 342 + ], + "lines": [ + { + "bbox": [ + 276, + 329, + 336, + 344 + ], + "spans": [ + { + "bbox": [ + 276, + 329, + 336, + 344 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 143, + 357, + 469, + 475 + ], + "lines": [ + { + "bbox": [ + 141, + 357, + 470, + 370 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 470, + 370 + ], + "score": 1.0, + "content": "A fundamental question in reinforcement learning is whether model-free algo-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 367, + 469, + 381 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 469, + 381 + ], + "score": 1.0, + "content": "rithms are sample efficient. Recently, Jin et al. (2018) proposed a Q-learning", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 379, + 469, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 379, + 469, + 392 + ], + "score": 1.0, + "content": "algorithm with UCB exploration policy, and proved it has nearly optimal regret", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 390, + 469, + 403 + ], + "spans": [ + { + "bbox": [ + 141, + 390, + 469, + 403 + ], + "score": 1.0, + "content": "bound for finite-horizon episodic MDP. In this paper, we adapt Q-learning with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 401, + 469, + 413 + ], + "spans": [ + { + "bbox": [ + 142, + 401, + 469, + 413 + ], + "score": 1.0, + "content": "UCB-exploration bonus to infinite-horizon MDP with discounted rewards without", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 412, + 469, + 426 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 469, + 426 + ], + "score": 1.0, + "content": "accessing a generative model. We show that the sample complexity of exploration", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 139, + 422, + 471, + 443 + ], + "spans": [ + { + "bbox": [ + 139, + 422, + 272, + 443 + ], + "score": 1.0, + "content": "of our algorithm is bounded by", + "type": "text" + }, + { + "bbox": [ + 273, + 423, + 323, + 439 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 422, + 471, + 443 + ], + "score": 1.0, + "content": ". This improves the previously best", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 136, + 432, + 474, + 460 + ], + "spans": [ + { + "bbox": [ + 136, + 432, + 206, + 460 + ], + "score": 1.0, + "content": "known result of", + "type": "text" + }, + { + "bbox": [ + 207, + 438, + 257, + 454 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 432, + 474, + 460 + ], + "score": 1.0, + "content": "in this setting achieved by delayed Q-learning (Strehl", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 452, + 469, + 466 + ], + "spans": [ + { + "bbox": [ + 141, + 452, + 362, + 466 + ], + "score": 1.0, + "content": "et al., 2006), and matches the lower bound in terms of", + "type": "text" + }, + { + "bbox": [ + 362, + 455, + 367, + 463 + ], + "score": 0.73, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 452, + 410, + 466 + ], + "score": 1.0, + "content": "as well as", + "type": "text" + }, + { + "bbox": [ + 419, + 452, + 436, + 466 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 446, + 452, + 469, + 466 + ], + "score": 1.0, + "content": "up to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 464, + 222, + 476 + ], + "spans": [ + { + "bbox": [ + 141, + 464, + 222, + 476 + ], + "score": 1.0, + "content": "logarithmic factors.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 19.5, + "bbox_fs": [ + 136, + 357, + 474, + 476 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 501, + 206, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 499, + 208, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 208, + 516 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 527, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "The goal of reinforcement learning (RL) is to construct efficient algorithms that learn and plan in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 551 + ], + "score": 1.0, + "content": "sequential decision making tasks when the underlying system dynamics are unknown. A typical", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 562 + ], + "score": 1.0, + "content": "model in RL is Markov Decision Process (MDP). At each time step, the environment is in a state", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 107, + 563, + 113, + 570 + ], + "score": 0.52, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 561, + 221, + 573 + ], + "score": 1.0, + "content": ". The agent takes an action", + "type": "text" + }, + { + "bbox": [ + 221, + 563, + 227, + 570 + ], + "score": 0.7, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 561, + 295, + 573 + ], + "score": 1.0, + "content": ", obtain a reward", + "type": "text" + }, + { + "bbox": [ + 295, + 563, + 301, + 571 + ], + "score": 0.75, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 561, + 506, + 573 + ], + "score": 1.0, + "content": ", and then the environment transits to another state.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "In reinforcement learning, the transition probability distribution is unknown. The algorithm needs", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 506, + 595 + ], + "score": 1.0, + "content": "to learn the transition dynamics of MDP, while aiming to maximize the cumulative reward. This", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "poses the exploration-exploitation dilemma: whether to act to gain new information (explore) or to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 605, + 375, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 375, + 617 + ], + "score": 1.0, + "content": "act consistently with past experience to maximize reward (exploit).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 528, + 506, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 709 + ], + "lines": [ + { + "bbox": [ + 105, + 620, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 506, + 635 + ], + "score": 1.0, + "content": "Theoretical analyses of reinforcement learning fall into two broad categories: those assuming a", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 632, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 505, + 645 + ], + "score": 1.0, + "content": "simulator (a.k.a. generative model), and those without a simulator. In the first category, the algorithm", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "is allowed to query the outcome of any state action pair from an oracle. The emphasis is on the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 265, + 667 + ], + "score": 1.0, + "content": "number of calls needed to estimate the", + "type": "text" + }, + { + "bbox": [ + 265, + 655, + 275, + 666 + ], + "score": 0.82, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "value or to output a near-optimal policy. There has been", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "extensive research in literature following this line of research, the majority of which focuses on", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 675, + 507, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 507, + 689 + ], + "score": 1.0, + "content": "discounted infinite horizon MDPs (Azar et al., 2011; Even-Dar & Mansour, 2003; Sidford et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 507, + 700 + ], + "score": 1.0, + "content": "2018b). The current results have achieved near-optimal time and sample complexities (Sidford et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 697, + 147, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 147, + 711 + ], + "score": 1.0, + "content": "2018b;a).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 620, + 507, + 711 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Without a simulator, there is a dichotomy between finite-horizon and infinite-horizon settings. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "score": 1.0, + "content": "finite-horizon settings, there are straightforward definitions for both regret and sample complexity;", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "the latter is defined as the number of samples needed before the policy becomes near optimal. In this", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "setting, extensive research in the past decade (Jin et al., 2018; Azar et al., 2017; Jaksch et al., 2010;", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "Dann et al., 2017) has achieved great progress, and established nearly-tight bounds for both regret", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 203, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 203, + 150 + ], + "score": 1.0, + "content": "and sample complexity.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "The infinite-horizon setting is a very different matter. First of all, the performance measure cannot be", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "a straightforward extension of the sample complexity defined above (See Strehl & Littman (2008)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "for detailed discussion). Instead, the measure of sample efficiency we adopt is the so-called sample", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "complexity of exploration (Kakade et al., 2003), which is also a widely-accepted definition. This", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 196, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 507, + 212 + ], + "score": 1.0, + "content": "measure counts the number of times that the algorithm “makes mistakes” along the whole trajectory.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 418, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 418, + 222 + ], + "score": 1.0, + "content": "See also (Strehl & Littman, 2008) for further discussions regarding this issue.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 507, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 507, + 239 + ], + "score": 1.0, + "content": "Several model based algorithms have been proposed for infinite horizon MDP, for example R-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 452, + 249 + ], + "score": 1.0, + "content": "max (Brafman & Tennenholtz, 2003), MoRmax (Szita & Szepesvári, 2010) and UCRL-", + "type": "text" + }, + { + "bbox": [ + 452, + 239, + 460, + 249 + ], + "score": 0.8, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "(Lattimore", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "& Hutter, 2012). It is noteworthy that there still exists a considerable gap between the state-of-the-art", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 499, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 428, + 272 + ], + "score": 1.0, + "content": "algorithm and the theoretical lower bound (Lattimore & Hutter, 2012) regarding", + "type": "text" + }, + { + "bbox": [ + 428, + 259, + 470, + 271 + ], + "score": 0.92, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 258, + 499, + 272 + ], + "score": 1.0, + "content": "factor.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 507, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 507, + 289 + ], + "score": 1.0, + "content": "Though model-based algorithms have been proved to be sample efficient in various MDP settings,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "score": 1.0, + "content": "most state-of-the-art RL algorithms are developed in the model-free paradigm (Schulman et al., 2015;", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "Mnih et al., 2013; 2016). Model-free algorithms are more flexible and require less space, which have", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "achieved remarkable performance on benchmarks such as Atari games and simulated robot control", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 319, + 150, + 332 + ], + "spans": [ + { + "bbox": [ + 104, + 319, + 150, + 332 + ], + "score": 1.0, + "content": "problems.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "For infinite horizon MDPs without access to simulator, the best model-free algorithm has a sample", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 103, + 347, + 507, + 366 + ], + "spans": [ + { + "bbox": [ + 103, + 347, + 215, + 366 + ], + "score": 1.0, + "content": "complexity of exploration", + "type": "text" + }, + { + "bbox": [ + 215, + 348, + 266, + 364 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 347, + 507, + 366 + ], + "score": 1.0, + "content": ", achieved by delayed Q-learning (Strehl et al., 2006). The", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "authors provide a novel strategy of argument when proving the upper bound for the sample complexity", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "of exploration, namely identifying a sufficient condition for optimality, and then bound the number of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 384, + 251, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 251, + 396 + ], + "score": 1.0, + "content": "times that this condition is violated.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 401, + 414 + ], + "score": 1.0, + "content": "However, the results of Delayed Q-learning still leave a quadratic gap in", + "type": "text" + }, + { + "bbox": [ + 402, + 401, + 417, + 413 + ], + "score": 0.86, + "content": "1 / \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "from the best-known", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "lower bound. This is partly because the updates in Q-value are made in an over-conservative way. In", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "score": 1.0, + "content": "fact, the loose sample complexity bound is a result of delayed Q-learning algorithm itself, as well as", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "the mathematical artifact in their analysis. To illustrate this, we construct a hard instance showing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 231, + 457 + ], + "score": 1.0, + "content": "that Delayed Q-learning incurs", + "type": "text" + }, + { + "bbox": [ + 231, + 444, + 266, + 457 + ], + "score": 0.92, + "content": "\\Omega ( 1 / \\epsilon ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "sample complexity. This observation, as well as the success", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "of the Q-learning with UCB algorithm (Jin et al., 2018) in proving a regret bound in finite-horizon", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 466, + 446, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 446, + 479 + ], + "score": 1.0, + "content": "settings, motivates us to incorporate a UCB-like exploration term into our algorithm.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "In this work, we propose a Q-learning algorithm with UCB exploration policy. We show the sample", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 507, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 312, + 510 + ], + "score": 1.0, + "content": "complexity of exploration bound of our algorithm is", + "type": "text" + }, + { + "bbox": [ + 312, + 495, + 363, + 510 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 495, + 507, + 510 + ], + "score": 1.0, + "content": ". This strictly improves the previous", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "best known result due to Delayed Q-learning. It also matches the lower bound in the dependence on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 519, + 255, + 533 + ], + "spans": [ + { + "bbox": [ + 107, + 522, + 112, + 529 + ], + "score": 0.41, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 519, + 115, + 533 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 116, + 520, + 123, + 530 + ], + "score": 0.72, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 519, + 141, + 533 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 142, + 520, + 150, + 529 + ], + "score": 0.68, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 519, + 255, + 533 + ], + "score": 1.0, + "content": "up to logarithmic factors.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "We point out here that the infinite-horizon setting cannot be solved by reducing to finite-horizon", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "setting. There are key technical differences between these two settings: the definition of sample", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "complexity of exploration, time-invariant policies and the error propagation structure in Q-learning.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "In particular, the analysis techniques developed in (Jin et al., 2018) do not directly apply here. We", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 430, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 430, + 594 + ], + "score": 1.0, + "content": "refer the readers to Section 3.2 for detailed explanations and a concrete example.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "The rest of the paper is organized as follows. After introducing the notation used in the paper in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "Section 2, we describe our infinite Q-learning with UCB algorithm in Section 3. We then state our", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "main theoretical results, which are in the form of PAC sample complexity bounds. In Section 4 we", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "score": 1.0, + "content": "present some interesting properties beyond sample complexity bound. Finally, we conclude the paper", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 642, + 159, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 159, + 653 + ], + "score": 1.0, + "content": "in Section 5.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + }, + { + "type": "title", + "bbox": [ + 108, + 678, + 198, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 200, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 200, + 693 + ], + "score": 1.0, + "content": "2 PRELIMINARY", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 364, + 722 + ], + "score": 1.0, + "content": "We consider a Markov Decision Process defined by a five tuple", + "type": "text" + }, + { + "bbox": [ + 364, + 710, + 419, + 721 + ], + "score": 0.91, + "content": "\\langle S , \\mathcal { A } , p , r , \\gamma \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 709, + 450, + 722 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 451, + 710, + 459, + 720 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "is the state", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 133, + 733 + ], + "score": 1.0, + "content": "space,", + "type": "text" + }, + { + "bbox": [ + 133, + 721, + 142, + 730 + ], + "score": 0.8, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 720, + 217, + 733 + ], + "score": 1.0, + "content": "is the action space,", + "type": "text" + }, + { + "bbox": [ + 218, + 720, + 255, + 732 + ], + "score": 0.94, + "content": "p ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 720, + 355, + 733 + ], + "score": 1.0, + "content": "is the transition function,", + "type": "text" + }, + { + "bbox": [ + 356, + 721, + 428, + 732 + ], + "score": 0.9, + "content": "r : S \\times \\mathcal { A } [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is the deterministic", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 94 + ], + "score": 1.0, + "content": "Without a simulator, there is a dichotomy between finite-horizon and infinite-horizon settings. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 506, + 107 + ], + "score": 1.0, + "content": "finite-horizon settings, there are straightforward definitions for both regret and sample complexity;", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "the latter is defined as the number of samples needed before the policy becomes near optimal. In this", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "setting, extensive research in the past decade (Jin et al., 2018; Azar et al., 2017; Jaksch et al., 2010;", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "Dann et al., 2017) has achieved great progress, and established nearly-tight bounds for both regret", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 203, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 203, + 150 + ], + "score": 1.0, + "content": "and sample complexity.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 82, + 506, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "The infinite-horizon setting is a very different matter. First of all, the performance measure cannot be", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 165, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 506, + 177 + ], + "score": 1.0, + "content": "a straightforward extension of the sample complexity defined above (See Strehl & Littman (2008)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "for detailed discussion). Instead, the measure of sample efficiency we adopt is the so-called sample", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 200 + ], + "score": 1.0, + "content": "complexity of exploration (Kakade et al., 2003), which is also a widely-accepted definition. This", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 196, + 507, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 507, + 212 + ], + "score": 1.0, + "content": "measure counts the number of times that the algorithm “makes mistakes” along the whole trajectory.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 418, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 418, + 222 + ], + "score": 1.0, + "content": "See also (Strehl & Littman, 2008) for further discussions regarding this issue.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 154, + 507, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 507, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 507, + 239 + ], + "score": 1.0, + "content": "Several model based algorithms have been proposed for infinite horizon MDP, for example R-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 452, + 249 + ], + "score": 1.0, + "content": "max (Brafman & Tennenholtz, 2003), MoRmax (Szita & Szepesvári, 2010) and UCRL-", + "type": "text" + }, + { + "bbox": [ + 452, + 239, + 460, + 249 + ], + "score": 0.8, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "(Lattimore", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "& Hutter, 2012). It is noteworthy that there still exists a considerable gap between the state-of-the-art", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 499, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 428, + 272 + ], + "score": 1.0, + "content": "algorithm and the theoretical lower bound (Lattimore & Hutter, 2012) regarding", + "type": "text" + }, + { + "bbox": [ + 428, + 259, + 470, + 271 + ], + "score": 0.92, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 258, + 499, + 272 + ], + "score": 1.0, + "content": "factor.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 226, + 507, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 275, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 507, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 507, + 289 + ], + "score": 1.0, + "content": "Though model-based algorithms have been proved to be sample efficient in various MDP settings,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 506, + 299 + ], + "score": 1.0, + "content": "most state-of-the-art RL algorithms are developed in the model-free paradigm (Schulman et al., 2015;", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "Mnih et al., 2013; 2016). Model-free algorithms are more flexible and require less space, which have", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "achieved remarkable performance on benchmarks such as Atari games and simulated robot control", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 319, + 150, + 332 + ], + "spans": [ + { + "bbox": [ + 104, + 319, + 150, + 332 + ], + "score": 1.0, + "content": "problems.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 275, + 507, + 332 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 336, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "For infinite horizon MDPs without access to simulator, the best model-free algorithm has a sample", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 103, + 347, + 507, + 366 + ], + "spans": [ + { + "bbox": [ + 103, + 347, + 215, + 366 + ], + "score": 1.0, + "content": "complexity of exploration", + "type": "text" + }, + { + "bbox": [ + 215, + 348, + 266, + 364 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 347, + 507, + 366 + ], + "score": 1.0, + "content": ", achieved by delayed Q-learning (Strehl et al., 2006). The", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "authors provide a novel strategy of argument when proving the upper bound for the sample complexity", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "of exploration, namely identifying a sufficient condition for optimality, and then bound the number of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 384, + 251, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 251, + 396 + ], + "score": 1.0, + "content": "times that this condition is violated.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 103, + 336, + 507, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 478 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 401, + 414 + ], + "score": 1.0, + "content": "However, the results of Delayed Q-learning still leave a quadratic gap in", + "type": "text" + }, + { + "bbox": [ + 402, + 401, + 417, + 413 + ], + "score": 0.86, + "content": "1 / \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 400, + 505, + 414 + ], + "score": 1.0, + "content": "from the best-known", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "lower bound. This is partly because the updates in Q-value are made in an over-conservative way. In", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 434 + ], + "score": 1.0, + "content": "fact, the loose sample complexity bound is a result of delayed Q-learning algorithm itself, as well as", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "the mathematical artifact in their analysis. To illustrate this, we construct a hard instance showing", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 444, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 231, + 457 + ], + "score": 1.0, + "content": "that Delayed Q-learning incurs", + "type": "text" + }, + { + "bbox": [ + 231, + 444, + 266, + 457 + ], + "score": 0.92, + "content": "\\Omega ( 1 / \\epsilon ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 444, + 505, + 457 + ], + "score": 1.0, + "content": "sample complexity. This observation, as well as the success", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "of the Q-learning with UCB algorithm (Jin et al., 2018) in proving a regret bound in finite-horizon", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 466, + 446, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 446, + 479 + ], + "score": 1.0, + "content": "settings, motivates us to incorporate a UCB-like exploration term into our algorithm.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 400, + 505, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 483, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "In this work, we propose a Q-learning algorithm with UCB exploration policy. We show the sample", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 507, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 312, + 510 + ], + "score": 1.0, + "content": "complexity of exploration bound of our algorithm is", + "type": "text" + }, + { + "bbox": [ + 312, + 495, + 363, + 510 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 495, + 507, + 510 + ], + "score": 1.0, + "content": ". This strictly improves the previous", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "best known result due to Delayed Q-learning. It also matches the lower bound in the dependence on", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 519, + 255, + 533 + ], + "spans": [ + { + "bbox": [ + 107, + 522, + 112, + 529 + ], + "score": 0.41, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 519, + 115, + 533 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 116, + 520, + 123, + 530 + ], + "score": 0.72, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 519, + 141, + 533 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 142, + 520, + 150, + 529 + ], + "score": 0.68, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 519, + 255, + 533 + ], + "score": 1.0, + "content": "up to logarithmic factors.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 483, + 507, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "We point out here that the infinite-horizon setting cannot be solved by reducing to finite-horizon", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "setting. There are key technical differences between these two settings: the definition of sample", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 572 + ], + "score": 1.0, + "content": "complexity of exploration, time-invariant policies and the error propagation structure in Q-learning.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "In particular, the analysis techniques developed in (Jin et al., 2018) do not directly apply here. We", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 430, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 430, + 594 + ], + "score": 1.0, + "content": "refer the readers to Section 3.2 for detailed explanations and a concrete example.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 536, + 506, + 594 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "The rest of the paper is organized as follows. After introducing the notation used in the paper in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "Section 2, we describe our infinite Q-learning with UCB algorithm in Section 3. We then state our", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 619, + 505, + 631 + ], + "score": 1.0, + "content": "main theoretical results, which are in the form of PAC sample complexity bounds. In Section 4 we", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "score": 1.0, + "content": "present some interesting properties beyond sample complexity bound. Finally, we conclude the paper", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 642, + 159, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 159, + 653 + ], + "score": 1.0, + "content": "in Section 5.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 596, + 505, + 653 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 678, + 198, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 200, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 200, + 693 + ], + "score": 1.0, + "content": "2 PRELIMINARY", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 364, + 722 + ], + "score": 1.0, + "content": "We consider a Markov Decision Process defined by a five tuple", + "type": "text" + }, + { + "bbox": [ + 364, + 710, + 419, + 721 + ], + "score": 0.91, + "content": "\\langle S , \\mathcal { A } , p , r , \\gamma \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 709, + 450, + 722 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 451, + 710, + 459, + 720 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "is the state", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 133, + 733 + ], + "score": 1.0, + "content": "space,", + "type": "text" + }, + { + "bbox": [ + 133, + 721, + 142, + 730 + ], + "score": 0.8, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 720, + 217, + 733 + ], + "score": 1.0, + "content": "is the action space,", + "type": "text" + }, + { + "bbox": [ + 218, + 720, + 255, + 732 + ], + "score": 0.94, + "content": "p ( s ^ { \\prime } | s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 720, + 355, + 733 + ], + "score": 1.0, + "content": "is the transition function,", + "type": "text" + }, + { + "bbox": [ + 356, + 721, + 428, + 732 + ], + "score": 0.9, + "content": "r : S \\times \\mathcal { A } [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "is the deterministic", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 191, + 95 + ], + "score": 1.0, + "content": "reward function, and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 191, + 83, + 235, + 94 + ], + "score": 0.93, + "content": "0 \\leq \\gamma < 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 235, + 82, + 388, + 95 + ], + "score": 1.0, + "content": "is the discount factor for rewards. Let", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 388, + 82, + 421, + 95 + ], + "score": 0.93, + "content": "S = | S |", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 421, + 82, + 439, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 439, + 82, + 475, + 95 + ], + "score": 0.92, + "content": "A = | { \\mathcal { A } } |", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 475, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "denote", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 347, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 347, + 106 + ], + "score": 1.0, + "content": "the number of states and the number of actions respectively.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 709, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 191, + 95 + ], + "score": 1.0, + "content": "reward function, and", + "type": "text" + }, + { + "bbox": [ + 191, + 83, + 235, + 94 + ], + "score": 0.93, + "content": "0 \\leq \\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 82, + 388, + 95 + ], + "score": 1.0, + "content": "is the discount factor for rewards. Let", + "type": "text" + }, + { + "bbox": [ + 388, + 82, + 421, + 95 + ], + "score": 0.93, + "content": "S = | S |", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 82, + 439, + 95 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 439, + 82, + 475, + 95 + ], + "score": 0.92, + "content": "A = | { \\mathcal { A } } |", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "denote", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 347, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 347, + 106 + ], + "score": 1.0, + "content": "the number of states and the number of actions respectively.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 191, + 124 + ], + "score": 1.0, + "content": "Starting from a state", + "type": "text" + }, + { + "bbox": [ + 191, + 113, + 201, + 122 + ], + "score": 0.86, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 110, + 506, + 124 + ], + "score": 1.0, + "content": ", the agent interacts with the environment for infinite number of time steps.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 278, + 133 + ], + "score": 1.0, + "content": "At each time step, the agent observes state", + "type": "text" + }, + { + "bbox": [ + 278, + 122, + 307, + 132 + ], + "score": 0.93, + "content": "s _ { t } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 121, + 361, + 133 + ], + "score": 1.0, + "content": ", picks action", + "type": "text" + }, + { + "bbox": [ + 361, + 122, + 391, + 132 + ], + "score": 0.91, + "content": "a _ { t } \\in \\mathcal A", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 121, + 477, + 133 + ], + "score": 1.0, + "content": ", and receives reward", + "type": "text" + }, + { + "bbox": [ + 477, + 123, + 486, + 132 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "; the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 262, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 238, + 146 + ], + "score": 1.0, + "content": "system then transits to next state", + "type": "text" + }, + { + "bbox": [ + 238, + 134, + 257, + 144 + ], + "score": 0.9, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 131, + 262, + 146 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 148, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 147, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 319, + 163 + ], + "score": 1.0, + "content": "Using the notations in Strehl et al. (2006), a policy", + "type": "text" + }, + { + "bbox": [ + 319, + 151, + 330, + 160 + ], + "score": 0.86, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 147, + 506, + 163 + ], + "score": 1.0, + "content": "refers to the non-stationary control policy", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 218, + 173 + ], + "score": 1.0, + "content": "of the algorithm since step", + "type": "text" + }, + { + "bbox": [ + 218, + 161, + 223, + 170 + ], + "score": 0.62, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 159, + 260, + 173 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 261, + 160, + 290, + 172 + ], + "score": 0.91, + "content": "V ^ { \\pi _ { t } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 159, + 463, + 173 + ], + "score": 1.0, + "content": "to denote the value function under policy", + "type": "text" + }, + { + "bbox": [ + 463, + 162, + 473, + 171 + ], + "score": 0.82, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 159, + 506, + 173 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 166, + 504, + 188 + ], + "spans": [ + { + "bbox": [ + 102, + 166, + 160, + 188 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + }, + { + "bbox": [ + 160, + 171, + 360, + 184 + ], + "score": 0.85, + "content": "\\begin{array} { r } { V ^ { \\pi _ { t } } ( s ) = \\mathbb { E } \\bigl [ \\sum _ { i = 1 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = s \\bigr ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 166, + 416, + 188 + ], + "score": 1.0, + "content": ". We also use", + "type": "text" + }, + { + "bbox": [ + 416, + 171, + 504, + 183 + ], + "score": 0.91, + "content": "V ^ { * } ( s ) = \\operatorname* { s u p } _ { \\pi } V ^ { \\pi } ( s )", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 181, + 509, + 211 + ], + "spans": [ + { + "bbox": [ + 104, + 181, + 414, + 195 + ], + "score": 1.0, + "content": "to denote the value function of the optimal policy. Accordingly, we define", + "type": "text" + }, + { + "bbox": [ + 106, + 192, + 296, + 206 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\mathbb { E } [ \\sum _ { i = 2 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = \\stackrel { \\cdot } { s } , a _ { 1 } = \\stackrel { \\cdot } { a } ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 187, + 322, + 211 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 323, + 194, + 332, + 204 + ], + "score": 0.26, + "content": "\\mathrm { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 187, + 420, + 211 + ], + "score": 1.0, + "content": "function under policy", + "type": "text" + }, + { + "bbox": [ + 415, + 183, + 506, + 194 + ], + "score": 0.86, + "content": "Q ^ { \\pi _ { t } } ( s , a ) = r ( s , a ) +", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 195, + 430, + 204 + ], + "score": 0.64, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 187, + 434, + 211 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 435, + 194, + 470, + 205 + ], + "score": 0.89, + "content": "Q ^ { * } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 187, + 509, + 211 + ], + "score": 1.0, + "content": "is the Q", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 243, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 227, + 217 + ], + "score": 1.0, + "content": "function under optimal policy", + "type": "text" + }, + { + "bbox": [ + 227, + 205, + 239, + 214 + ], + "score": 0.84, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 204, + 243, + 217 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 108, + 221, + 503, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 235 + ], + "score": 1.0, + "content": "We use the sample complexity of exploration defined in Kakade et al. (2003) to measure the learning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "efficiency of our algorithm. This sample complexity definition has been widely used in previous", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 243, + 427, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 351, + 255 + ], + "score": 1.0, + "content": "works Strehl et al. (2006); Lattimore & Hutter (2012); Strehl", + "type": "text" + }, + { + "bbox": [ + 351, + 244, + 360, + 253 + ], + "score": 0.28, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 243, + 427, + 255 + ], + "score": 1.0, + "content": "Littman (2008).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 256, + 506, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 380, + 269 + ], + "score": 1.0, + "content": "Definition 1. Sample complexity of Exploration of an algorithm", + "type": "text" + }, + { + "bbox": [ + 380, + 256, + 402, + 267 + ], + "score": 0.56, + "content": "\\mathcal { A L G }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is defined as the number", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 267, + 507, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 302, + 281 + ], + "score": 1.0, + "content": "of time steps t such that the non-stationary policy", + "type": "text" + }, + { + "bbox": [ + 303, + 269, + 313, + 278 + ], + "score": 0.84, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 267, + 344, + 281 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 344, + 269, + 349, + 277 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 267, + 374, + 281 + ], + "score": 1.0, + "content": "is not", + "type": "text" + }, + { + "bbox": [ + 374, + 270, + 379, + 277 + ], + "score": 0.47, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 267, + 479, + 281 + ], + "score": 1.0, + "content": "-optimal for current state", + "type": "text" + }, + { + "bbox": [ + 479, + 269, + 488, + 278 + ], + "score": 0.8, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 267, + 507, + 281 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 276, + 207, + 291 + ], + "spans": [ + { + "bbox": [ + 107, + 278, + 203, + 290 + ], + "score": 0.89, + "content": "\\dot { V } ^ { \\pi _ { t } } \\left( s _ { t } \\right) \\dot { < } V ^ { * } \\left( s _ { t } \\right) - \\epsilon .", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 276, + 207, + 291 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 297, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "score": 1.0, + "content": "Roughly speaking, this measure counts the number of mistakes along the whole trajectory. We use", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 307, + 339, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 339, + 320 + ], + "score": 1.0, + "content": "the following definition of PAC-MDP Strehl et al. (2006).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 223, + 335 + ], + "score": 1.0, + "content": "Definition 2. An algorithm", + "type": "text" + }, + { + "bbox": [ + 223, + 321, + 246, + 332 + ], + "score": 0.67, + "content": "\\mathcal { A L G }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "is said to be PAC-MDP (Probably Approximately Correct in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 225, + 345 + ], + "score": 1.0, + "content": "Markov Decision Processes)", + "type": "text" + }, + { + "bbox": [ + 226, + 333, + 235, + 344 + ], + "score": 0.67, + "content": "i f ,", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 332, + 294, + 345 + ], + "score": 1.0, + "content": "for any \u000f and", + "type": "text" + }, + { + "bbox": [ + 294, + 333, + 300, + 342 + ], + "score": 0.65, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 332, + 409, + 345 + ], + "score": 1.0, + "content": ", the sample complexity of", + "type": "text" + }, + { + "bbox": [ + 409, + 333, + 432, + 343 + ], + "score": 0.61, + "content": "\\mathcal { A L G }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "is less than some", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 493, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 255, + 357 + ], + "score": 1.0, + "content": "polynomial in the relevant quantities", + "type": "text" + }, + { + "bbox": [ + 255, + 343, + 363, + 356 + ], + "score": 0.89, + "content": "( S , A , 1 / \\epsilon , 1 / \\delta , 1 / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 343, + 465, + 357 + ], + "score": 1.0, + "content": ", with probability at least", + "type": "text" + }, + { + "bbox": [ + 466, + 344, + 488, + 354 + ], + "score": 0.85, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 343, + 493, + 357 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 363, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 365, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 365, + 376 + ], + "score": 1.0, + "content": "Finally, recall that Bellman equation is defined as the following:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 375, + 462, + 403 + ], + "lines": [ + { + "bbox": [ + 142, + 375, + 462, + 403 + ], + "spans": [ + { + "bbox": [ + 142, + 375, + 462, + 403 + ], + "score": 0.89, + "content": "\\left\\{ \\begin{array} { l l } { V ^ { \\pi _ { t } } ( s ) = Q ^ { \\pi _ { t } } \\left( s , \\pi _ { t } ( s ) \\right) } \\\\ { Q ^ { \\pi _ { t } } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { \\pi _ { t + 1 } } \\right) ( s , a ) , } \\end{array} \\right. \\quad \\left\\{ \\begin{array} { l l } { V ^ { * } ( s ) = Q ^ { * } \\left( s , \\pi ^ { * } ( s ) \\right) } \\\\ { Q ^ { * } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { * } \\right) ( s , a ) , } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "59cd62d5ec6d571803efed659b258dfba8d0141783a9422cc7a1cf15db20a44e.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 142, + 375, + 462, + 403 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 103, + 403, + 493, + 416 + ], + "lines": [ + { + "bbox": [ + 104, + 400, + 493, + 419 + ], + "spans": [ + { + "bbox": [ + 104, + 400, + 335, + 419 + ], + "score": 1.0, + "content": "which is frequently used in our analysis. Here we denote", + "type": "text" + }, + { + "bbox": [ + 335, + 404, + 493, + 416 + ], + "score": 0.87, + "content": "\\left[ \\mathbb { P } V ^ { \\pi _ { t } } \\right] ( s , a ) : = \\mathbb { E } _ { s ^ { \\prime } \\sim p ( \\cdot \\vert s , a ) } V ^ { \\pi _ { t + 1 } } ( s ^ { \\prime } ) .", + "type": "inline_equation" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 430, + 206, + 443 + ], + "lines": [ + { + "bbox": [ + 104, + 428, + 207, + 445 + ], + "spans": [ + { + "bbox": [ + 104, + 428, + 207, + 445 + ], + "score": 1.0, + "content": "3 MAIN RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 102, + 454, + 475, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 476, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 476, + 469 + ], + "score": 1.0, + "content": "In this section, we present the UCB Q-learning algorithm and the sample complexity bound.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 107, + 478, + 186, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 187, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 187, + 491 + ], + "score": 1.0, + "content": "3.1 ALGORITHM", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 108, + 502, + 279, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 281, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 281, + 516 + ], + "score": 1.0, + "content": "Algorithm 1 Infinite Q-learning with UCB", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 110, + 516, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 123, + 515, + 201, + 528 + ], + "spans": [ + { + "bbox": [ + 123, + 515, + 201, + 528 + ], + "score": 1.0, + "content": "Parameters: \u000f, γ, δ", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 122, + 526, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 122, + 526, + 155, + 543 + ], + "score": 1.0, + "content": "Initializ", + "type": "text" + }, + { + "bbox": [ + 155, + 526, + 479, + 562 + ], + "score": 0.69, + "content": "\\begin{array} { r l } & { \\textnormal { \\texttt { e } } Q ( s , a ) , \\hat { Q } ( s , a ) \\gets \\frac { 1 } { 1 - \\gamma } , N ( s , a ) \\gets 0 , \\epsilon _ { 1 } \\gets \\frac { \\epsilon } { 2 4 R M \\ln \\frac { 1 } { 1 - \\gamma } } , H \\gets \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } . } \\\\ & { \\cdot ( k ) = \\ln ( S A ( k + 1 ) ( k + 2 ) / \\delta ) , \\alpha _ { k } = \\frac { H + 1 } { H + k } . } \\end{array}", + "type": "inline_equation", + "image_path": "ad0c57b5739df10ef8c0d2046656de08d85d51ac78dc2acbfb1350115223ebf4.jpg" + } + ], + "index": 29 + }, + { + "bbox": [ + 122, + 542, + 153, + 560 + ], + "spans": [ + { + "bbox": [ + 122, + 542, + 153, + 560 + ], + "score": 1.0, + "content": "Define", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 123, + 557, + 197, + 570 + ], + "spans": [ + { + "bbox": [ + 123, + 557, + 138, + 570 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 138, + 559, + 197, + 569 + ], + "score": 0.66, + "content": "t = 1 , 2 , \\dots \\mathbf { d o }", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 110, + 569, + 293, + 583 + ], + "spans": [ + { + "bbox": [ + 110, + 570, + 123, + 583 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 137, + 569, + 186, + 583 + ], + "score": 1.0, + "content": "Take action", + "type": "text" + }, + { + "bbox": [ + 187, + 569, + 293, + 582 + ], + "score": 0.73, + "content": "a _ { t } \\gets \\arg \\operatorname* { m a x } _ { a ^ { \\prime } } \\hat { Q } ( s _ { t } , a ^ { \\prime } )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 137, + 580, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 137, + 580, + 288, + 594 + ], + "score": 1.0, + "content": "Receive reward rt and transit to st+1", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 137, + 590, + 504, + 670 + ], + "spans": [ + { + "bbox": [ + 137, + 590, + 411, + 670 + ], + "score": 0.48, + "content": "\\begin{array} { r l r } & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { k \\gets N ( s _ { t } , a _ { t } ) , b _ { k } \\gets \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H ( k ) } { k } } } & \\\\ & { \\hat { V } ( s _ { t + 1 } ) \\gets \\operatorname* { m a x } _ { a \\in A } \\hat { Q } ( s _ { t + 1 } , a ) } & \\\\ & { Q ( s _ { t } , a _ { t } ) \\gets ( 1 - \\alpha _ { k } ) Q ( s _ { t } , a _ { t } ) + \\alpha _ { k } \\left[ r ( s _ { t } , a _ { t } ) + b _ { k } + \\gamma \\hat { V } ( s _ { t + 1 } ) \\right] } \\\\ & { \\hat { Q } ( s _ { t } , a _ { t } ) \\gets \\operatorname* { m i n } ( \\hat { Q } ( s _ { t } , a _ { t } ) , Q ( s _ { t } , a _ { t } ) ) } & \\end{array}", + "type": "inline_equation", + "image_path": "d5388b49142af9a03fc552a4ec774a7df6c96f7f162f72703f3ded5df679058d.jpg" + }, + { + "bbox": [ + 411, + 608, + 484, + 621 + ], + "score": 1.0, + "content": "t and can be set to", + "type": "text" + }, + { + "bbox": [ + 484, + 608, + 504, + 621 + ], + "score": 0.88, + "content": "4 { \\sqrt { 2 } }", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 639, + 122, + 651 + ], + "spans": [ + { + "bbox": [ + 107, + 639, + 122, + 651 + ], + "score": 1.0, + "content": "10:", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 122, + 666, + 158, + 679 + ], + "spans": [ + { + "bbox": [ + 122, + 666, + 158, + 679 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 505, + 735 + ], + "lines": [ + { + "bbox": [ + 104, + 690, + 507, + 710 + ], + "spans": [ + { + "bbox": [ + 104, + 690, + 129, + 710 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 129, + 692, + 173, + 704 + ], + "score": 0.93, + "content": "c _ { 2 } = 4 \\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 690, + 232, + 710 + ], + "score": 1.0, + "content": "is a constant.", + "type": "text" + }, + { + "bbox": [ + 233, + 692, + 338, + 708 + ], + "score": 0.94, + "content": "\\begin{array} { r } { R = \\lceil \\ln \\frac { 3 } { \\epsilon ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 690, + 424, + 710 + ], + "score": 1.0, + "content": ", while the choice of", + "type": "text" + }, + { + "bbox": [ + 425, + 694, + 437, + 703 + ], + "score": 0.8, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 690, + 507, + 710 + ], + "score": 1.0, + "content": "can be found in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 705, + 505, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 162, + 719 + ], + "score": 1.0, + "content": "Section. 3.3.", + "type": "text" + }, + { + "bbox": [ + 162, + 707, + 270, + 719 + ], + "score": 0.75, + "content": "{ \\cal { \\left( M = 0 \\left( \\ln 1 / ( 1 - \\gamma ) \\epsilon \\right) \\right) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 705, + 401, + 719 + ], + "score": 1.0, + "content": "). The learning rate is defined as", + "type": "text" + }, + { + "bbox": [ + 402, + 706, + 501, + 719 + ], + "score": 0.91, + "content": "\\alpha _ { k } = ( H + 1 ) / ( H + k )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 705, + 505, + 719 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 714, + 366, + 738 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 117, + 730 + ], + "score": 0.79, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 714, + 168, + 738 + ], + "score": 1.0, + "content": "is chosen as", + "type": "text" + }, + { + "bbox": [ + 169, + 718, + 222, + 735 + ], + "score": 0.94, + "content": "\\frac { \\ln { 1 } / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln { 1 } / \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 714, + 287, + 738 + ], + "score": 1.0, + "content": ", which satisfies", + "type": "text" + }, + { + "bbox": [ + 287, + 718, + 363, + 735 + ], + "score": 0.93, + "content": "H \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 714, + 366, + 738 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 504, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 191, + 124 + ], + "score": 1.0, + "content": "Starting from a state", + "type": "text" + }, + { + "bbox": [ + 191, + 113, + 201, + 122 + ], + "score": 0.86, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 110, + 506, + 124 + ], + "score": 1.0, + "content": ", the agent interacts with the environment for infinite number of time steps.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 121, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 121, + 278, + 133 + ], + "score": 1.0, + "content": "At each time step, the agent observes state", + "type": "text" + }, + { + "bbox": [ + 278, + 122, + 307, + 132 + ], + "score": 0.93, + "content": "s _ { t } \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 121, + 361, + 133 + ], + "score": 1.0, + "content": ", picks action", + "type": "text" + }, + { + "bbox": [ + 361, + 122, + 391, + 132 + ], + "score": 0.91, + "content": "a _ { t } \\in \\mathcal A", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 121, + 477, + 133 + ], + "score": 1.0, + "content": ", and receives reward", + "type": "text" + }, + { + "bbox": [ + 477, + 123, + 486, + 132 + ], + "score": 0.85, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 121, + 505, + 133 + ], + "score": 1.0, + "content": "; the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 262, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 238, + 146 + ], + "score": 1.0, + "content": "system then transits to next state", + "type": "text" + }, + { + "bbox": [ + 238, + 134, + 257, + 144 + ], + "score": 0.9, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 131, + 262, + 146 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 110, + 506, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 148, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 147, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 319, + 163 + ], + "score": 1.0, + "content": "Using the notations in Strehl et al. (2006), a policy", + "type": "text" + }, + { + "bbox": [ + 319, + 151, + 330, + 160 + ], + "score": 0.86, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 147, + 506, + 163 + ], + "score": 1.0, + "content": "refers to the non-stationary control policy", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 218, + 173 + ], + "score": 1.0, + "content": "of the algorithm since step", + "type": "text" + }, + { + "bbox": [ + 218, + 161, + 223, + 170 + ], + "score": 0.62, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 159, + 260, + 173 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 261, + 160, + 290, + 172 + ], + "score": 0.91, + "content": "V ^ { \\pi _ { t } } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 159, + 463, + 173 + ], + "score": 1.0, + "content": "to denote the value function under policy", + "type": "text" + }, + { + "bbox": [ + 463, + 162, + 473, + 171 + ], + "score": 0.82, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 159, + 506, + 173 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 166, + 504, + 188 + ], + "spans": [ + { + "bbox": [ + 102, + 166, + 160, + 188 + ], + "score": 1.0, + "content": "is defined as", + "type": "text" + }, + { + "bbox": [ + 160, + 171, + 360, + 184 + ], + "score": 0.85, + "content": "\\begin{array} { r } { V ^ { \\pi _ { t } } ( s ) = \\mathbb { E } \\bigl [ \\sum _ { i = 1 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = s \\bigr ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 166, + 416, + 188 + ], + "score": 1.0, + "content": ". We also use", + "type": "text" + }, + { + "bbox": [ + 416, + 171, + 504, + 183 + ], + "score": 0.91, + "content": "V ^ { * } ( s ) = \\operatorname* { s u p } _ { \\pi } V ^ { \\pi } ( s )", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 181, + 509, + 211 + ], + "spans": [ + { + "bbox": [ + 104, + 181, + 414, + 195 + ], + "score": 1.0, + "content": "to denote the value function of the optimal policy. Accordingly, we define", + "type": "text" + }, + { + "bbox": [ + 106, + 192, + 296, + 206 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\mathbb { E } [ \\sum _ { i = 2 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = \\stackrel { \\cdot } { s } , a _ { 1 } = \\stackrel { \\cdot } { a } ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 187, + 322, + 211 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 323, + 194, + 332, + 204 + ], + "score": 0.26, + "content": "\\mathrm { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 187, + 420, + 211 + ], + "score": 1.0, + "content": "function under policy", + "type": "text" + }, + { + "bbox": [ + 415, + 183, + 506, + 194 + ], + "score": 0.86, + "content": "Q ^ { \\pi _ { t } } ( s , a ) = r ( s , a ) +", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 195, + 430, + 204 + ], + "score": 0.64, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 187, + 434, + 211 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 435, + 194, + 470, + 205 + ], + "score": 0.89, + "content": "Q ^ { * } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 187, + 509, + 211 + ], + "score": 1.0, + "content": "is the Q", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 243, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 227, + 217 + ], + "score": 1.0, + "content": "function under optimal policy", + "type": "text" + }, + { + "bbox": [ + 227, + 205, + 239, + 214 + ], + "score": 0.84, + "content": "\\pi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 204, + 243, + 217 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 102, + 147, + 509, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 221, + 503, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 235 + ], + "score": 1.0, + "content": "We use the sample complexity of exploration defined in Kakade et al. (2003) to measure the learning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "efficiency of our algorithm. This sample complexity definition has been widely used in previous", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 243, + 427, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 351, + 255 + ], + "score": 1.0, + "content": "works Strehl et al. (2006); Lattimore & Hutter (2012); Strehl", + "type": "text" + }, + { + "bbox": [ + 351, + 244, + 360, + 253 + ], + "score": 0.28, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 243, + 427, + 255 + ], + "score": 1.0, + "content": "Littman (2008).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 219, + 505, + 255 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 256, + 506, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 380, + 269 + ], + "score": 1.0, + "content": "Definition 1. Sample complexity of Exploration of an algorithm", + "type": "text" + }, + { + "bbox": [ + 380, + 256, + 402, + 267 + ], + "score": 0.56, + "content": "\\mathcal { A L G }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is defined as the number", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 267, + 507, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 302, + 281 + ], + "score": 1.0, + "content": "of time steps t such that the non-stationary policy", + "type": "text" + }, + { + "bbox": [ + 303, + 269, + 313, + 278 + ], + "score": 0.84, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 267, + 344, + 281 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 344, + 269, + 349, + 277 + ], + "score": 0.68, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 267, + 374, + 281 + ], + "score": 1.0, + "content": "is not", + "type": "text" + }, + { + "bbox": [ + 374, + 270, + 379, + 277 + ], + "score": 0.47, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 267, + 479, + 281 + ], + "score": 1.0, + "content": "-optimal for current state", + "type": "text" + }, + { + "bbox": [ + 479, + 269, + 488, + 278 + ], + "score": 0.8, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 267, + 507, + 281 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 276, + 207, + 291 + ], + "spans": [ + { + "bbox": [ + 107, + 278, + 203, + 290 + ], + "score": 0.89, + "content": "\\dot { V } ^ { \\pi _ { t } } \\left( s _ { t } \\right) \\dot { < } V ^ { * } \\left( s _ { t } \\right) - \\epsilon .", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 276, + 207, + 291 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 255, + 507, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 297, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "score": 1.0, + "content": "Roughly speaking, this measure counts the number of mistakes along the whole trajectory. We use", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 307, + 339, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 339, + 320 + ], + "score": 1.0, + "content": "the following definition of PAC-MDP Strehl et al. (2006).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 296, + 505, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 321, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 223, + 335 + ], + "score": 1.0, + "content": "Definition 2. An algorithm", + "type": "text" + }, + { + "bbox": [ + 223, + 321, + 246, + 332 + ], + "score": 0.67, + "content": "\\mathcal { A L G }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "is said to be PAC-MDP (Probably Approximately Correct in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 332, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 225, + 345 + ], + "score": 1.0, + "content": "Markov Decision Processes)", + "type": "text" + }, + { + "bbox": [ + 226, + 333, + 235, + 344 + ], + "score": 0.67, + "content": "i f ,", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 332, + 294, + 345 + ], + "score": 1.0, + "content": "for any \u000f and", + "type": "text" + }, + { + "bbox": [ + 294, + 333, + 300, + 342 + ], + "score": 0.65, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 332, + 409, + 345 + ], + "score": 1.0, + "content": ", the sample complexity of", + "type": "text" + }, + { + "bbox": [ + 409, + 333, + 432, + 343 + ], + "score": 0.61, + "content": "\\mathcal { A L G }", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 332, + 505, + 345 + ], + "score": 1.0, + "content": "is less than some", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 343, + 493, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 255, + 357 + ], + "score": 1.0, + "content": "polynomial in the relevant quantities", + "type": "text" + }, + { + "bbox": [ + 255, + 343, + 363, + 356 + ], + "score": 0.89, + "content": "( S , A , 1 / \\epsilon , 1 / \\delta , 1 / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 343, + 465, + 357 + ], + "score": 1.0, + "content": ", with probability at least", + "type": "text" + }, + { + "bbox": [ + 466, + 344, + 488, + 354 + ], + "score": 0.85, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 343, + 493, + 357 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 321, + 505, + 357 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 363, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 365, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 365, + 376 + ], + "score": 1.0, + "content": "Finally, recall that Bellman equation is defined as the following:", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 361, + 365, + 376 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 142, + 375, + 462, + 403 + ], + "lines": [ + { + "bbox": [ + 142, + 375, + 462, + 403 + ], + "spans": [ + { + "bbox": [ + 142, + 375, + 462, + 403 + ], + "score": 0.89, + "content": "\\left\\{ \\begin{array} { l l } { V ^ { \\pi _ { t } } ( s ) = Q ^ { \\pi _ { t } } \\left( s , \\pi _ { t } ( s ) \\right) } \\\\ { Q ^ { \\pi _ { t } } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { \\pi _ { t + 1 } } \\right) ( s , a ) , } \\end{array} \\right. \\quad \\left\\{ \\begin{array} { l l } { V ^ { * } ( s ) = Q ^ { * } \\left( s , \\pi ^ { * } ( s ) \\right) } \\\\ { Q ^ { * } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { * } \\right) ( s , a ) , } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "59cd62d5ec6d571803efed659b258dfba8d0141783a9422cc7a1cf15db20a44e.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 142, + 375, + 462, + 403 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 103, + 403, + 493, + 416 + ], + "lines": [ + { + "bbox": [ + 104, + 400, + 493, + 419 + ], + "spans": [ + { + "bbox": [ + 104, + 400, + 335, + 419 + ], + "score": 1.0, + "content": "which is frequently used in our analysis. Here we denote", + "type": "text" + }, + { + "bbox": [ + 335, + 404, + 493, + 416 + ], + "score": 0.87, + "content": "\\left[ \\mathbb { P } V ^ { \\pi _ { t } } \\right] ( s , a ) : = \\mathbb { E } _ { s ^ { \\prime } \\sim p ( \\cdot \\vert s , a ) } V ^ { \\pi _ { t + 1 } } ( s ^ { \\prime } ) .", + "type": "inline_equation" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 104, + 400, + 493, + 419 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 430, + 206, + 443 + ], + "lines": [ + { + "bbox": [ + 104, + 428, + 207, + 445 + ], + "spans": [ + { + "bbox": [ + 104, + 428, + 207, + 445 + ], + "score": 1.0, + "content": "3 MAIN RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 102, + 454, + 475, + 467 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 476, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 476, + 469 + ], + "score": 1.0, + "content": "In this section, we present the UCB Q-learning algorithm and the sample complexity bound.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 453, + 476, + 469 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 478, + 186, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 187, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 187, + 491 + ], + "score": 1.0, + "content": "3.1 ALGORITHM", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 108, + 502, + 279, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 281, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 281, + 516 + ], + "score": 1.0, + "content": "Algorithm 1 Infinite Q-learning with UCB", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "list", + "bbox": [ + 110, + 516, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 123, + 515, + 201, + 528 + ], + "spans": [ + { + "bbox": [ + 123, + 515, + 201, + 528 + ], + "score": 1.0, + "content": "Parameters: \u000f, γ, δ", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 122, + 526, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 122, + 526, + 155, + 543 + ], + "score": 1.0, + "content": "Initializ", + "type": "text" + }, + { + "bbox": [ + 155, + 526, + 479, + 562 + ], + "score": 0.69, + "content": "\\begin{array} { r l } & { \\textnormal { \\texttt { e } } Q ( s , a ) , \\hat { Q } ( s , a ) \\gets \\frac { 1 } { 1 - \\gamma } , N ( s , a ) \\gets 0 , \\epsilon _ { 1 } \\gets \\frac { \\epsilon } { 2 4 R M \\ln \\frac { 1 } { 1 - \\gamma } } , H \\gets \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } . } \\\\ & { \\cdot ( k ) = \\ln ( S A ( k + 1 ) ( k + 2 ) / \\delta ) , \\alpha _ { k } = \\frac { H + 1 } { H + k } . } \\end{array}", + "type": "inline_equation", + "image_path": "ad0c57b5739df10ef8c0d2046656de08d85d51ac78dc2acbfb1350115223ebf4.jpg" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 122, + 542, + 153, + 560 + ], + "spans": [ + { + "bbox": [ + 122, + 542, + 153, + 560 + ], + "score": 1.0, + "content": "Define", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 123, + 557, + 197, + 570 + ], + "spans": [ + { + "bbox": [ + 123, + 557, + 138, + 570 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 138, + 559, + 197, + 569 + ], + "score": 0.66, + "content": "t = 1 , 2 , \\dots \\mathbf { d o }", + "type": "inline_equation" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 569, + 293, + 583 + ], + "spans": [ + { + "bbox": [ + 110, + 570, + 123, + 583 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 137, + 569, + 186, + 583 + ], + "score": 1.0, + "content": "Take action", + "type": "text" + }, + { + "bbox": [ + 187, + 569, + 293, + 582 + ], + "score": 0.73, + "content": "a _ { t } \\gets \\arg \\operatorname* { m a x } _ { a ^ { \\prime } } \\hat { Q } ( s _ { t } , a ^ { \\prime } )", + "type": "inline_equation" + } + ], + "index": 32, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 137, + 580, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 137, + 580, + 288, + 594 + ], + "score": 1.0, + "content": "Receive reward rt and transit to st+1", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 137, + 590, + 504, + 670 + ], + "spans": [ + { + "bbox": [ + 137, + 590, + 411, + 670 + ], + "score": 0.48, + "content": "\\begin{array} { r l r } & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { k \\gets N ( s _ { t } , a _ { t } ) , b _ { k } \\gets \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H ( k ) } { k } } } & \\\\ & { \\hat { V } ( s _ { t + 1 } ) \\gets \\operatorname* { m a x } _ { a \\in A } \\hat { Q } ( s _ { t + 1 } , a ) } & \\\\ & { Q ( s _ { t } , a _ { t } ) \\gets ( 1 - \\alpha _ { k } ) Q ( s _ { t } , a _ { t } ) + \\alpha _ { k } \\left[ r ( s _ { t } , a _ { t } ) + b _ { k } + \\gamma \\hat { V } ( s _ { t + 1 } ) \\right] } \\\\ & { \\hat { Q } ( s _ { t } , a _ { t } ) \\gets \\operatorname* { m i n } ( \\hat { Q } ( s _ { t } , a _ { t } ) , Q ( s _ { t } , a _ { t } ) ) } & \\end{array}", + "type": "inline_equation", + "image_path": "d5388b49142af9a03fc552a4ec774a7df6c96f7f162f72703f3ded5df679058d.jpg" + }, + { + "bbox": [ + 411, + 608, + 484, + 621 + ], + "score": 1.0, + "content": "t and can be set to", + "type": "text" + }, + { + "bbox": [ + 484, + 608, + 504, + 621 + ], + "score": 0.88, + "content": "4 { \\sqrt { 2 } }", + "type": "inline_equation" + } + ], + "index": 35 + }, + { + "bbox": [ + 107, + 639, + 122, + 651 + ], + "spans": [ + { + "bbox": [ + 107, + 639, + 122, + 651 + ], + "score": 1.0, + "content": "10:", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 122, + 666, + 158, + 679 + ], + "spans": [ + { + "bbox": [ + 122, + 666, + 158, + 679 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + } + ], + "index": 32, + "bbox_fs": [ + 107, + 515, + 504, + 679 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 505, + 735 + ], + "lines": [ + { + "bbox": [ + 104, + 690, + 507, + 710 + ], + "spans": [ + { + "bbox": [ + 104, + 690, + 129, + 710 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 129, + 692, + 173, + 704 + ], + "score": 0.93, + "content": "c _ { 2 } = 4 \\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 690, + 232, + 710 + ], + "score": 1.0, + "content": "is a constant.", + "type": "text" + }, + { + "bbox": [ + 233, + 692, + 338, + 708 + ], + "score": 0.94, + "content": "\\begin{array} { r } { R = \\lceil \\ln \\frac { 3 } { \\epsilon ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 690, + 424, + 710 + ], + "score": 1.0, + "content": ", while the choice of", + "type": "text" + }, + { + "bbox": [ + 425, + 694, + 437, + 703 + ], + "score": 0.8, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 437, + 690, + 507, + 710 + ], + "score": 1.0, + "content": "can be found in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 705, + 505, + 719 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 162, + 719 + ], + "score": 1.0, + "content": "Section. 3.3.", + "type": "text" + }, + { + "bbox": [ + 162, + 707, + 270, + 719 + ], + "score": 0.75, + "content": "{ \\cal { \\left( M = 0 \\left( \\ln 1 / ( 1 - \\gamma ) \\epsilon \\right) \\right) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 705, + 401, + 719 + ], + "score": 1.0, + "content": "). The learning rate is defined as", + "type": "text" + }, + { + "bbox": [ + 402, + 706, + 501, + 719 + ], + "score": 0.91, + "content": "\\alpha _ { k } = ( H + 1 ) / ( H + k )", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 705, + 505, + 719 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 714, + 366, + 738 + ], + "spans": [ + { + "bbox": [ + 107, + 720, + 117, + 730 + ], + "score": 0.79, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 714, + 168, + 738 + ], + "score": 1.0, + "content": "is chosen as", + "type": "text" + }, + { + "bbox": [ + 169, + 718, + 222, + 735 + ], + "score": 0.94, + "content": "\\frac { \\ln { 1 } / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln { 1 } / \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 714, + 287, + 738 + ], + "score": 1.0, + "content": ", which satisfies", + "type": "text" + }, + { + "bbox": [ + 287, + 718, + 363, + 735 + ], + "score": 0.93, + "content": "H \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma }", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 714, + 366, + 738 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 104, + 690, + 507, + 738 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Our UCB Q-learning algorithm (Algorithm 1) maintains an optimistic estimation of action value", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 141, + 108 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 141, + 95, + 172, + 107 + ], + "score": 0.93, + "content": "Q ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 94, + 304, + 108 + ], + "score": 1.0, + "content": "and its historical minimum value", + "type": "text" + }, + { + "bbox": [ + 304, + 93, + 335, + 107 + ], + "score": 0.86, + "content": "\\hat { Q } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 94, + 339, + 108 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 339, + 95, + 374, + 107 + ], + "score": 0.85, + "content": "N _ { t } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "denotes the number of times that", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 128, + 118 + ], + "score": 0.91, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 106, + 258, + 119 + ], + "score": 1.0, + "content": "is experienced before time step", + "type": "text" + }, + { + "bbox": [ + 258, + 107, + 264, + 117 + ], + "score": 0.48, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 106, + 267, + 119 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 268, + 106, + 306, + 118 + ], + "score": 0.92, + "content": "\\tau ( s , a , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 106, + 395, + 119 + ], + "score": 1.0, + "content": "denotes the time step", + "type": "text" + }, + { + "bbox": [ + 396, + 107, + 401, + 116 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 106, + 439, + 119 + ], + "score": 1.0, + "content": "at which", + "type": "text" + }, + { + "bbox": [ + 440, + 106, + 505, + 118 + ], + "score": 0.91, + "content": "( s _ { t } , a _ { t } ) = ( s , a )", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 136, + 130 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 136, + 118, + 143, + 127 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 115, + 401, + 130 + ], + "score": 1.0, + "content": "-th time; if this state-action pair is not visited that many times,", + "type": "text" + }, + { + "bbox": [ + 402, + 117, + 464, + 129 + ], + "score": 0.89, + "content": "\\tau ( s , a , k ) = \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 115, + 470, + 130 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 470, + 118, + 505, + 129 + ], + "score": 0.84, + "content": "Q _ { t } ( s , a )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 124, + 144 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 128, + 158, + 142 + ], + "score": 0.93, + "content": "\\hat { Q } _ { t } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 128, + 208, + 144 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 209, + 130, + 218, + 142 + ], + "score": 0.85, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 128, + 236, + 144 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 128, + 246, + 142 + ], + "score": 0.87, + "content": "\\hat { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 128, + 283, + 144 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 284, + 130, + 306, + 142 + ], + "score": 0.91, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 128, + 494, + 144 + ], + "score": 1.0, + "content": "that the algorithm maintains when arriving at", + "type": "text" + }, + { + "bbox": [ + 495, + 132, + 504, + 141 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 159, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 159, + 155 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 106, + 169, + 306, + 181 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 307, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 307, + 183 + ], + "score": 1.0, + "content": "3.2 SAMPLE COMPLEXITY OF EXPLORATION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 402, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 402, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 402, + 205 + ], + "score": 1.0, + "content": "Our main result is the following sample complexity of exploration bound.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 507, + 241 + ], + "lines": [ + { + "bbox": [ + 105, + 206, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 194, + 221 + ], + "score": 1.0, + "content": "Theorem 1. For any", + "type": "text" + }, + { + "bbox": [ + 194, + 209, + 218, + 219 + ], + "score": 0.77, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 206, + 221, + 221 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 221, + 208, + 303, + 220 + ], + "score": 0.6, + "content": "\\delta > 0 , 1 / 2 < \\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 206, + 374, + 221 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + }, + { + "bbox": [ + 375, + 208, + 398, + 219 + ], + "score": 0.76, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 206, + 507, + 221 + ], + "score": 1.0, + "content": ", the sample complexity of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 219, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 322, + 231 + ], + "score": 1.0, + "content": "exploration (i.e., the number of time steps t such that", + "type": "text" + }, + { + "bbox": [ + 322, + 221, + 333, + 230 + ], + "score": 0.86, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 219, + 358, + 231 + ], + "score": 1.0, + "content": "is not", + "type": "text" + }, + { + "bbox": [ + 359, + 222, + 364, + 229 + ], + "score": 0.55, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 219, + 410, + 231 + ], + "score": 1.0, + "content": "-optimal at", + "type": "text" + }, + { + "bbox": [ + 410, + 221, + 420, + 230 + ], + "score": 0.82, + "content": "s _ { t . }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 219, + 506, + 231 + ], + "score": 1.0, + "content": ") of Algorithm 1 is at", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 129, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 129, + 242 + ], + "score": 1.0, + "content": "most", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 242, + 345, + 276 + ], + "lines": [ + { + "bbox": [ + 265, + 242, + 345, + 276 + ], + "spans": [ + { + "bbox": [ + 265, + 242, + 345, + 276 + ], + "score": 0.92, + "content": "\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln { 1 / \\delta } } { \\epsilon ^ { 2 } \\left( 1 - \\gamma \\right) ^ { 7 } } \\right) ,", + "type": "interline_equation", + "image_path": "cc2b5aa0f63178566f02139857adabb797f2f963feb35907fc51fd7501fb1693.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 265, + 242, + 345, + 259.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 265, + 259.0, + 345, + 276.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 375, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 376, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 133, + 298 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 283, + 143, + 295 + ], + "score": 0.79, + "content": "\\tilde { \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 282, + 278, + 298 + ], + "score": 1.0, + "content": "suppresses logarithmic factors of", + "type": "text" + }, + { + "bbox": [ + 279, + 284, + 339, + 297 + ], + "score": 0.9, + "content": "1 / \\epsilon , 1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 282, + 357, + 298 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 358, + 284, + 373, + 295 + ], + "score": 0.77, + "content": "S A", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 282, + 376, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "We first point out the obstacles for proving the theorem and reasons why the techniques in Jin et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "(2018) do not directly apply here. We then give a high level description of the ideas of our approach.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 506, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "One important issue is caused by the difference in the definition of sample complexity for finite and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "infinite horizon MDP. In finite horizon settings, sample complexity (and regret) is determined in the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 354, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 104, + 354, + 125, + 371 + ], + "score": 1.0, + "content": "first", + "type": "text" + }, + { + "bbox": [ + 125, + 357, + 134, + 367 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 354, + 402, + 371 + ], + "score": 1.0, + "content": "timesteps, and only measures the performance at the initial state", + "type": "text" + }, + { + "bbox": [ + 402, + 358, + 413, + 368 + ], + "score": 0.83, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 354, + 435, + 371 + ], + "score": 1.0, + "content": "(i.e.", + "type": "text" + }, + { + "bbox": [ + 435, + 357, + 502, + 369 + ], + "score": 0.91, + "content": "( V ^ { * } - V ^ { \\pi } ) ( s _ { 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 354, + 506, + 371 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "However, in the infinite horizon setting, the agent may enter under-explored regions at any time", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "score": 1.0, + "content": "period, and sample complexity of exploration characterizes the performance at all states the agent", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 390, + 136, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 136, + 402 + ], + "score": 1.0, + "content": "enters.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "score": 1.0, + "content": "The following example clearly illustrates the key difference between infinite-horizon and finite-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 104, + 416, + 296, + 431 + ], + "score": 1.0, + "content": "horizon. Consider an MDP with a starting state", + "type": "text" + }, + { + "bbox": [ + 296, + 419, + 307, + 429 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 416, + 437, + 431 + ], + "score": 1.0, + "content": "where the probability of leaving", + "type": "text" + }, + { + "bbox": [ + 438, + 419, + 448, + 429 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 416, + 458, + 431 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 459, + 417, + 489, + 430 + ], + "score": 0.92, + "content": "o ( T ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 416, + 506, + 431 + ], + "score": 1.0, + "content": ". In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 331, + 442 + ], + "score": 1.0, + "content": "this case, with high probability, it would take more than", + "type": "text" + }, + { + "bbox": [ + 331, + 429, + 340, + 439 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 428, + 414, + 442 + ], + "score": 1.0, + "content": "timesteps to leave", + "type": "text" + }, + { + "bbox": [ + 415, + 430, + 425, + 440 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 428, + 505, + 442 + ], + "score": 1.0, + "content": ". Hence, guarantees", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 440, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 222, + 452 + ], + "score": 1.0, + "content": "about the learning in the first", + "type": "text" + }, + { + "bbox": [ + 222, + 441, + 231, + 450 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 440, + 382, + 452 + ], + "score": 1.0, + "content": "timesteps or about the performance at", + "type": "text" + }, + { + "bbox": [ + 383, + 441, + 393, + 451 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 440, + 506, + 452 + ], + "score": 1.0, + "content": "imply almost nothing about", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "score": 1.0, + "content": "the number of mistakes the algorithm would make in the rest of the MDP (i.e. the sample complexity", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "of exploration of the algorithm). As a result, the analysis for finite horizon MDPs cannot be directly", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 244, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 244, + 486 + ], + "score": 1.0, + "content": "applied to infinite horizon setting.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "This calls for techniques for counting mistakes along the entire trajectory, such as those employed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "by Strehl et al. (2006). In particular, we need to establish convenient sufficient conditions for being", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 107, + 514, + 112, + 521 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 511, + 196, + 524 + ], + "score": 1.0, + "content": "-optimal at timestep", + "type": "text" + }, + { + "bbox": [ + 196, + 513, + 201, + 521 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 511, + 242, + 524 + ], + "score": 1.0, + "content": "and state", + "type": "text" + }, + { + "bbox": [ + 243, + 513, + 252, + 522 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 511, + 274, + 524 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 275, + 511, + 370, + 523 + ], + "score": 0.92, + "content": "V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 511, + 506, + 524 + ], + "score": 1.0, + "content": ". Then, bounding the number of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 522, + 373, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 373, + 536 + ], + "score": 1.0, + "content": "violations of such conditions gives a bound on sample complexity.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "Another technical reason why the proof in Jin et al. (2018) cannot be directly applied to our problem", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "is the following: In finite horizon settings, Jin et al. (2018) decomposed the learning error at episode", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 113, + 571 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 560, + 150, + 574 + ], + "score": 1.0, + "content": "and time", + "type": "text" + }, + { + "bbox": [ + 151, + 562, + 158, + 571 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 560, + 358, + 574 + ], + "score": 1.0, + "content": "as errors from a set of consecutive episodes before", + "type": "text" + }, + { + "bbox": [ + 358, + 562, + 365, + 571 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 560, + 395, + 574 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 396, + 561, + 420, + 572 + ], + "score": 0.91, + "content": "h + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "using a clever design", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "of learning rate. However, in the infinite horizon setting, this property does not hold. Suppose at", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 126, + 596 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 127, + 584, + 132, + 593 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 582, + 211, + 596 + ], + "score": 1.0, + "content": "the agent is at state", + "type": "text" + }, + { + "bbox": [ + 211, + 585, + 221, + 594 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 582, + 288, + 596 + ], + "score": 1.0, + "content": "and takes action", + "type": "text" + }, + { + "bbox": [ + 289, + 585, + 298, + 594 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 582, + 407, + 596 + ], + "score": 1.0, + "content": ". Then the learning error at", + "type": "text" + }, + { + "bbox": [ + 407, + 584, + 412, + 593 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "only depends on those", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 382, + 606 + ], + "score": 1.0, + "content": "previous time steps such that the agent encountered the same state as", + "type": "text" + }, + { + "bbox": [ + 382, + 596, + 392, + 605 + ], + "score": 0.86, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "and took the same action as", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 107, + 606, + 116, + 616 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 604, + 245, + 618 + ], + "score": 1.0, + "content": ". Thus the learning error at time", + "type": "text" + }, + { + "bbox": [ + 245, + 606, + 250, + 615 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "cannot be decomposed as errors from a set of consecutive time", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 157, + 629 + ], + "score": 1.0, + "content": "steps before", + "type": "text" + }, + { + "bbox": [ + 157, + 617, + 162, + 626 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 616, + 506, + 629 + ], + "score": 1.0, + "content": ", but errors from a set of non-consecutive time steps without any structure. Therefore,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "we have to control the sum of learning errors over an unstructured set of time steps. This makes the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 215, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 215, + 652 + ], + "score": 1.0, + "content": "analysis more challenging.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "Now we give a brief road map of the proof of Theorem 1. Our first goal is to establish a sufficient", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 174, + 678 + ], + "score": 1.0, + "content": "condition so that", + "type": "text" + }, + { + "bbox": [ + 175, + 667, + 185, + 677 + ], + "score": 0.85, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 666, + 245, + 678 + ], + "score": 1.0, + "content": "learned at step", + "type": "text" + }, + { + "bbox": [ + 246, + 667, + 250, + 676 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 666, + 261, + 678 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 261, + 668, + 266, + 676 + ], + "score": 0.75, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 666, + 335, + 678 + ], + "score": 1.0, + "content": "-optimal for state", + "type": "text" + }, + { + "bbox": [ + 335, + 668, + 344, + 677 + ], + "score": 0.81, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 666, + 506, + 678 + ], + "score": 1.0, + "content": ". As an intermediate step we show that a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 200, + 690 + ], + "score": 1.0, + "content": "sufficient condition for", + "type": "text" + }, + { + "bbox": [ + 201, + 677, + 292, + 689 + ], + "score": 0.91, + "content": "V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 676, + 320, + 690 + ], + "score": 1.0, + "content": "is that", + "type": "text" + }, + { + "bbox": [ + 320, + 677, + 411, + 689 + ], + "score": 0.92, + "content": "V ^ { \\ast } ( s _ { t ^ { \\prime } } ) - Q ^ { \\ast } ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "is small for a few time", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 128, + 700 + ], + "score": 1.0, + "content": "steps", + "type": "text" + }, + { + "bbox": [ + 128, + 687, + 136, + 698 + ], + "score": 0.83, + "content": "t ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 687, + 208, + 700 + ], + "score": 1.0, + "content": "within an interval", + "type": "text" + }, + { + "bbox": [ + 208, + 689, + 245, + 699 + ], + "score": 0.91, + "content": "[ t , t + R ]", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 687, + 333, + 700 + ], + "score": 1.0, + "content": "for a carefully chosen", + "type": "text" + }, + { + "bbox": [ + 333, + 689, + 342, + 698 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "(Condition 1). Then we show the desired", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "sufficient condition (Condition 2) implies Condition 1. We then bound the total number of bad time", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 167, + 722 + ], + "score": 1.0, + "content": "steps on which", + "type": "text" + }, + { + "bbox": [ + 168, + 710, + 251, + 722 + ], + "score": 0.91, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "is large for the whole MDP; this implies a bound on the number", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 450, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 450, + 733 + ], + "score": 1.0, + "content": "of violations of Condition 2. This in turn relies on a key technical lemma (Lemma 2).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 153 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "Our UCB Q-learning algorithm (Algorithm 1) maintains an optimistic estimation of action value", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 141, + 108 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 141, + 95, + 172, + 107 + ], + "score": 0.93, + "content": "Q ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 94, + 304, + 108 + ], + "score": 1.0, + "content": "and its historical minimum value", + "type": "text" + }, + { + "bbox": [ + 304, + 93, + 335, + 107 + ], + "score": 0.86, + "content": "\\hat { Q } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 94, + 339, + 108 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 339, + 95, + 374, + 107 + ], + "score": 0.85, + "content": "N _ { t } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "denotes the number of times that", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 106, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 128, + 118 + ], + "score": 0.91, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 106, + 258, + 119 + ], + "score": 1.0, + "content": "is experienced before time step", + "type": "text" + }, + { + "bbox": [ + 258, + 107, + 264, + 117 + ], + "score": 0.48, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 106, + 267, + 119 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 268, + 106, + 306, + 118 + ], + "score": 0.92, + "content": "\\tau ( s , a , k )", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 106, + 395, + 119 + ], + "score": 1.0, + "content": "denotes the time step", + "type": "text" + }, + { + "bbox": [ + 396, + 107, + 401, + 116 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 106, + 439, + 119 + ], + "score": 1.0, + "content": "at which", + "type": "text" + }, + { + "bbox": [ + 440, + 106, + 505, + 118 + ], + "score": 0.91, + "content": "( s _ { t } , a _ { t } ) = ( s , a )", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 136, + 130 + ], + "score": 1.0, + "content": "for the", + "type": "text" + }, + { + "bbox": [ + 136, + 118, + 143, + 127 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 115, + 401, + 130 + ], + "score": 1.0, + "content": "-th time; if this state-action pair is not visited that many times,", + "type": "text" + }, + { + "bbox": [ + 402, + 117, + 464, + 129 + ], + "score": 0.89, + "content": "\\tau ( s , a , k ) = \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 115, + 470, + 130 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 470, + 118, + 505, + 129 + ], + "score": 0.84, + "content": "Q _ { t } ( s , a )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 504, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 124, + 144 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 128, + 158, + 142 + ], + "score": 0.93, + "content": "\\hat { Q } _ { t } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 128, + 208, + 144 + ], + "score": 1.0, + "content": "denotes the", + "type": "text" + }, + { + "bbox": [ + 209, + 130, + 218, + 142 + ], + "score": 0.85, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 128, + 236, + 144 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 237, + 128, + 246, + 142 + ], + "score": 0.87, + "content": "\\hat { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 128, + 283, + 144 + ], + "score": 1.0, + "content": "value of", + "type": "text" + }, + { + "bbox": [ + 284, + 130, + 306, + 142 + ], + "score": 0.91, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 128, + 494, + 144 + ], + "score": 1.0, + "content": "that the algorithm maintains when arriving at", + "type": "text" + }, + { + "bbox": [ + 495, + 132, + 504, + 141 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 139, + 159, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 159, + 155 + ], + "score": 1.0, + "content": "respectively.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 81, + 506, + 155 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 169, + 306, + 181 + ], + "lines": [ + { + "bbox": [ + 105, + 168, + 307, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 307, + 183 + ], + "score": 1.0, + "content": "3.2 SAMPLE COMPLEXITY OF EXPLORATION", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 402, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 402, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 402, + 205 + ], + "score": 1.0, + "content": "Our main result is the following sample complexity of exploration bound.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 190, + 402, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 207, + 507, + 241 + ], + "lines": [ + { + "bbox": [ + 105, + 206, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 194, + 221 + ], + "score": 1.0, + "content": "Theorem 1. For any", + "type": "text" + }, + { + "bbox": [ + 194, + 209, + 218, + 219 + ], + "score": 0.77, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 206, + 221, + 221 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 221, + 208, + 303, + 220 + ], + "score": 0.6, + "content": "\\delta > 0 , 1 / 2 < \\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 206, + 374, + 221 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + }, + { + "bbox": [ + 375, + 208, + 398, + 219 + ], + "score": 0.76, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 206, + 507, + 221 + ], + "score": 1.0, + "content": ", the sample complexity of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 219, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 322, + 231 + ], + "score": 1.0, + "content": "exploration (i.e., the number of time steps t such that", + "type": "text" + }, + { + "bbox": [ + 322, + 221, + 333, + 230 + ], + "score": 0.86, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 219, + 358, + 231 + ], + "score": 1.0, + "content": "is not", + "type": "text" + }, + { + "bbox": [ + 359, + 222, + 364, + 229 + ], + "score": 0.55, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 219, + 410, + 231 + ], + "score": 1.0, + "content": "-optimal at", + "type": "text" + }, + { + "bbox": [ + 410, + 221, + 420, + 230 + ], + "score": 0.82, + "content": "s _ { t . }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 219, + 506, + 231 + ], + "score": 1.0, + "content": ") of Algorithm 1 is at", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 129, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 129, + 242 + ], + "score": 1.0, + "content": "most", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 206, + 507, + 242 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 265, + 242, + 345, + 276 + ], + "lines": [ + { + "bbox": [ + 265, + 242, + 345, + 276 + ], + "spans": [ + { + "bbox": [ + 265, + 242, + 345, + 276 + ], + "score": 0.92, + "content": "\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln { 1 / \\delta } } { \\epsilon ^ { 2 } \\left( 1 - \\gamma \\right) ^ { 7 } } \\right) ,", + "type": "interline_equation", + "image_path": "cc2b5aa0f63178566f02139857adabb797f2f963feb35907fc51fd7501fb1693.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 265, + 242, + 345, + 259.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 265, + 259.0, + 345, + 276.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 375, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 376, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 133, + 298 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 283, + 143, + 295 + ], + "score": 0.79, + "content": "\\tilde { \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 282, + 278, + 298 + ], + "score": 1.0, + "content": "suppresses logarithmic factors of", + "type": "text" + }, + { + "bbox": [ + 279, + 284, + 339, + 297 + ], + "score": 0.9, + "content": "1 / \\epsilon , 1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 282, + 357, + 298 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 358, + 284, + 373, + 295 + ], + "score": 0.77, + "content": "S A", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 282, + 376, + 298 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 282, + 376, + 298 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 307, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 106, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "We first point out the obstacles for proving the theorem and reasons why the techniques in Jin et al.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "(2018) do not directly apply here. We then give a high level description of the ideas of our approach.", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 307, + 506, + 331 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 506, + 401 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "One important issue is caused by the difference in the definition of sample complexity for finite and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 505, + 358 + ], + "score": 1.0, + "content": "infinite horizon MDP. In finite horizon settings, sample complexity (and regret) is determined in the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 354, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 104, + 354, + 125, + 371 + ], + "score": 1.0, + "content": "first", + "type": "text" + }, + { + "bbox": [ + 125, + 357, + 134, + 367 + ], + "score": 0.79, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 354, + 402, + 371 + ], + "score": 1.0, + "content": "timesteps, and only measures the performance at the initial state", + "type": "text" + }, + { + "bbox": [ + 402, + 358, + 413, + 368 + ], + "score": 0.83, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 354, + 435, + 371 + ], + "score": 1.0, + "content": "(i.e.", + "type": "text" + }, + { + "bbox": [ + 435, + 357, + 502, + 369 + ], + "score": 0.91, + "content": "( V ^ { * } - V ^ { \\pi } ) ( s _ { 1 } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 354, + 506, + 371 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "However, in the infinite horizon setting, the agent may enter under-explored regions at any time", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 391 + ], + "score": 1.0, + "content": "period, and sample complexity of exploration characterizes the performance at all states the agent", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 390, + 136, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 136, + 402 + ], + "score": 1.0, + "content": "enters.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 334, + 506, + 402 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 406, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 419 + ], + "score": 1.0, + "content": "The following example clearly illustrates the key difference between infinite-horizon and finite-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 416, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 104, + 416, + 296, + 431 + ], + "score": 1.0, + "content": "horizon. Consider an MDP with a starting state", + "type": "text" + }, + { + "bbox": [ + 296, + 419, + 307, + 429 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 416, + 437, + 431 + ], + "score": 1.0, + "content": "where the probability of leaving", + "type": "text" + }, + { + "bbox": [ + 438, + 419, + 448, + 429 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 416, + 458, + 431 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 459, + 417, + 489, + 430 + ], + "score": 0.92, + "content": "o ( T ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 416, + 506, + 431 + ], + "score": 1.0, + "content": ". In", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 331, + 442 + ], + "score": 1.0, + "content": "this case, with high probability, it would take more than", + "type": "text" + }, + { + "bbox": [ + 331, + 429, + 340, + 439 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 428, + 414, + 442 + ], + "score": 1.0, + "content": "timesteps to leave", + "type": "text" + }, + { + "bbox": [ + 415, + 430, + 425, + 440 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 428, + 505, + 442 + ], + "score": 1.0, + "content": ". Hence, guarantees", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 440, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 222, + 452 + ], + "score": 1.0, + "content": "about the learning in the first", + "type": "text" + }, + { + "bbox": [ + 222, + 441, + 231, + 450 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 440, + 382, + 452 + ], + "score": 1.0, + "content": "timesteps or about the performance at", + "type": "text" + }, + { + "bbox": [ + 383, + 441, + 393, + 451 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 440, + 506, + 452 + ], + "score": 1.0, + "content": "imply almost nothing about", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 464 + ], + "score": 1.0, + "content": "the number of mistakes the algorithm would make in the rest of the MDP (i.e. the sample complexity", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "of exploration of the algorithm). As a result, the analysis for finite horizon MDPs cannot be directly", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 472, + 244, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 244, + 486 + ], + "score": 1.0, + "content": "applied to infinite horizon setting.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25, + "bbox_fs": [ + 104, + 407, + 506, + 486 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 489, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 502 + ], + "score": 1.0, + "content": "This calls for techniques for counting mistakes along the entire trajectory, such as those employed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "by Strehl et al. (2006). In particular, we need to establish convenient sufficient conditions for being", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 107, + 514, + 112, + 521 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 511, + 196, + 524 + ], + "score": 1.0, + "content": "-optimal at timestep", + "type": "text" + }, + { + "bbox": [ + 196, + 513, + 201, + 521 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 511, + 242, + 524 + ], + "score": 1.0, + "content": "and state", + "type": "text" + }, + { + "bbox": [ + 243, + 513, + 252, + 522 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 511, + 274, + 524 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 275, + 511, + 370, + 523 + ], + "score": 0.92, + "content": "V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 511, + 506, + 524 + ], + "score": 1.0, + "content": ". Then, bounding the number of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 522, + 373, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 373, + 536 + ], + "score": 1.0, + "content": "violations of such conditions gives a bound on sample complexity.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 488, + 506, + 536 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "Another technical reason why the proof in Jin et al. (2018) cannot be directly applied to our problem", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 563 + ], + "score": 1.0, + "content": "is the following: In finite horizon settings, Jin et al. (2018) decomposed the learning error at episode", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 113, + 571 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 560, + 150, + 574 + ], + "score": 1.0, + "content": "and time", + "type": "text" + }, + { + "bbox": [ + 151, + 562, + 158, + 571 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 560, + 358, + 574 + ], + "score": 1.0, + "content": "as errors from a set of consecutive episodes before", + "type": "text" + }, + { + "bbox": [ + 358, + 562, + 365, + 571 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 560, + 395, + 574 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 396, + 561, + 420, + 572 + ], + "score": 0.91, + "content": "h + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "using a clever design", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "of learning rate. However, in the infinite horizon setting, this property does not hold. Suppose at", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 126, + 596 + ], + "score": 1.0, + "content": "time", + "type": "text" + }, + { + "bbox": [ + 127, + 584, + 132, + 593 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 582, + 211, + 596 + ], + "score": 1.0, + "content": "the agent is at state", + "type": "text" + }, + { + "bbox": [ + 211, + 585, + 221, + 594 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 582, + 288, + 596 + ], + "score": 1.0, + "content": "and takes action", + "type": "text" + }, + { + "bbox": [ + 289, + 585, + 298, + 594 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 582, + 407, + 596 + ], + "score": 1.0, + "content": ". Then the learning error at", + "type": "text" + }, + { + "bbox": [ + 407, + 584, + 412, + 593 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "only depends on those", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 382, + 606 + ], + "score": 1.0, + "content": "previous time steps such that the agent encountered the same state as", + "type": "text" + }, + { + "bbox": [ + 382, + 596, + 392, + 605 + ], + "score": 0.86, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "and took the same action as", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 107, + 606, + 116, + 616 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 604, + 245, + 618 + ], + "score": 1.0, + "content": ". Thus the learning error at time", + "type": "text" + }, + { + "bbox": [ + 245, + 606, + 250, + 615 + ], + "score": 0.64, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "cannot be decomposed as errors from a set of consecutive time", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 157, + 629 + ], + "score": 1.0, + "content": "steps before", + "type": "text" + }, + { + "bbox": [ + 157, + 617, + 162, + 626 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 616, + 506, + 629 + ], + "score": 1.0, + "content": ", but errors from a set of non-consecutive time steps without any structure. Therefore,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "we have to control the sum of learning errors over an unstructured set of time steps. This makes the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 215, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 215, + 652 + ], + "score": 1.0, + "content": "analysis more challenging.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 538, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 506, + 667 + ], + "score": 1.0, + "content": "Now we give a brief road map of the proof of Theorem 1. Our first goal is to establish a sufficient", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 174, + 678 + ], + "score": 1.0, + "content": "condition so that", + "type": "text" + }, + { + "bbox": [ + 175, + 667, + 185, + 677 + ], + "score": 0.85, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 666, + 245, + 678 + ], + "score": 1.0, + "content": "learned at step", + "type": "text" + }, + { + "bbox": [ + 246, + 667, + 250, + 676 + ], + "score": 0.8, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 666, + 261, + 678 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 261, + 668, + 266, + 676 + ], + "score": 0.75, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 666, + 335, + 678 + ], + "score": 1.0, + "content": "-optimal for state", + "type": "text" + }, + { + "bbox": [ + 335, + 668, + 344, + 677 + ], + "score": 0.81, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 666, + 506, + 678 + ], + "score": 1.0, + "content": ". As an intermediate step we show that a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 200, + 690 + ], + "score": 1.0, + "content": "sufficient condition for", + "type": "text" + }, + { + "bbox": [ + 201, + 677, + 292, + 689 + ], + "score": 0.91, + "content": "V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 676, + 320, + 690 + ], + "score": 1.0, + "content": "is that", + "type": "text" + }, + { + "bbox": [ + 320, + 677, + 411, + 689 + ], + "score": 0.92, + "content": "V ^ { \\ast } ( s _ { t ^ { \\prime } } ) - Q ^ { \\ast } ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "is small for a few time", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 128, + 700 + ], + "score": 1.0, + "content": "steps", + "type": "text" + }, + { + "bbox": [ + 128, + 687, + 136, + 698 + ], + "score": 0.83, + "content": "t ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 687, + 208, + 700 + ], + "score": 1.0, + "content": "within an interval", + "type": "text" + }, + { + "bbox": [ + 208, + 689, + 245, + 699 + ], + "score": 0.91, + "content": "[ t , t + R ]", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 687, + 333, + 700 + ], + "score": 1.0, + "content": "for a carefully chosen", + "type": "text" + }, + { + "bbox": [ + 333, + 689, + 342, + 698 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "(Condition 1). Then we show the desired", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "sufficient condition (Condition 2) implies Condition 1. We then bound the total number of bad time", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 708, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 167, + 722 + ], + "score": 1.0, + "content": "steps on which", + "type": "text" + }, + { + "bbox": [ + 168, + 710, + 251, + 722 + ], + "score": 0.91, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 708, + 506, + 722 + ], + "score": 1.0, + "content": "is large for the whole MDP; this implies a bound on the number", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 450, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 450, + 733 + ], + "score": 1.0, + "content": "of violations of Condition 2. This in turn relies on a key technical lemma (Lemma 2).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 655, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The remaining part of this section is organized as follows. We establish the sufficient condition for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 96, + 112, + 104 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "-optimality in Section 3.3. The key lemma is presented in Section 3.4. Finally we prove Theorem 1", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 167, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 167, + 116 + ], + "score": 1.0, + "content": "in Section 3.5.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 106, + 129, + 318, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 319, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 255, + 142 + ], + "score": 1.0, + "content": "3.3 SUFFICIENT CONDITION FOR", + "type": "text" + }, + { + "bbox": [ + 255, + 132, + 261, + 140 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 129, + 319, + 142 + ], + "score": 1.0, + "content": "-OPTIMALITY", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 483, + 163 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 484, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 376, + 165 + ], + "score": 1.0, + "content": "In this section, we establish a sufficient condition (Condition 2) for", + "type": "text" + }, + { + "bbox": [ + 376, + 153, + 381, + 160 + ], + "score": 0.65, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 148, + 475, + 165 + ], + "score": 1.0, + "content": "-optimality at time step", + "type": "text" + }, + { + "bbox": [ + 475, + 152, + 480, + 160 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 148, + 484, + 165 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 506, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 152, + 181 + ], + "score": 1.0, + "content": "For a fixed", + "type": "text" + }, + { + "bbox": [ + 153, + 169, + 162, + 178 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 165, + 204, + 181 + ], + "score": 1.0, + "content": ", let TRAJ", + "type": "text" + }, + { + "bbox": [ + 204, + 168, + 218, + 179 + ], + "score": 0.66, + "content": "( R )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 165, + 300, + 181 + ], + "score": 1.0, + "content": "be the set of length-", + "type": "text" + }, + { + "bbox": [ + 301, + 168, + 309, + 177 + ], + "score": 0.8, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 165, + 413, + 181 + ], + "score": 1.0, + "content": "trajectories starting from", + "type": "text" + }, + { + "bbox": [ + 414, + 170, + 423, + 178 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 165, + 506, + 181 + ], + "score": 1.0, + "content": ". Our goal is to give", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 225, + 191 + ], + "score": 1.0, + "content": "a sufficient condition so that", + "type": "text" + }, + { + "bbox": [ + 226, + 180, + 236, + 190 + ], + "score": 0.85, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 178, + 345, + 191 + ], + "score": 1.0, + "content": ", the policy learned at step", + "type": "text" + }, + { + "bbox": [ + 346, + 180, + 351, + 189 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 178, + 364, + 191 + ], + "score": 1.0, + "content": ", is", + "type": "text" + }, + { + "bbox": [ + 365, + 181, + 370, + 188 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 178, + 444, + 191 + ], + "score": 1.0, + "content": "-optimal. For any", + "type": "text" + }, + { + "bbox": [ + 444, + 179, + 473, + 190 + ], + "score": 0.91, + "content": "\\epsilon _ { 2 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 178, + 505, + 191 + ], + "score": 1.0, + "content": ", define", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 187, + 406, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 216, + 204 + ], + "score": 0.93, + "content": "\\begin{array} { r } { R : = \\lceil \\ln \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 187, + 252, + 206 + ], + "score": 1.0, + "content": ". Denote", + "type": "text" + }, + { + "bbox": [ + 253, + 189, + 335, + 201 + ], + "score": 0.94, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 187, + 349, + 206 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 349, + 190, + 362, + 200 + ], + "score": 0.88, + "content": "\\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 187, + 406, + 206 + ], + "score": 1.0, + "content": ". We have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 124, + 209, + 485, + 372 + ], + "lines": [ + { + "bbox": [ + 124, + 209, + 485, + 372 + ], + "spans": [ + { + "bbox": [ + 124, + 209, + 485, + 372 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\quad V ^ { * } ( s _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + Q ^ { * } ( s _ { t } , a _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { P } ( V ^ { * } - V ^ { \\pi } ) ( s _ { t } , \\pi _ { t } ( s _ { t } ) ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\sum _ { s + 1 } ^ { \\infty } p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\cdot [ V ^ { * } ( s _ { t + 1 } ) - Q ^ { * } ( s _ { t + 1 } , a _ { t + 1 } ) ] + } \\\\ & { \\quad \\Big . \\Big . \\Big . \\Big . \\Big . \\Big . } \\\\ & { \\quad \\quad \\Big . \\gamma _ { s + 1 , s ^ { * } + 2 } \\mathrm { ~ } p ( s _ { t + 2 } | s _ { t + 1 } , \\pi _ { t + 1 } ( s _ { t + 1 } ) ) \\cdot p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\big [ V ^ { * } ( s _ { t + 2 } ) - Q ^ { * } ( s _ { t + 2 } , a _ { t + 2 } ) \\Big ] } \\\\ & { \\quad \\quad \\cdot \\dots } \\\\ & { \\le \\epsilon _ { 2 } + \\displaystyle \\sum _ { s ^ { \\prime } = s ^ { \\prime } } p ( t r a _ { t } ) \\cdot [ \\frac { n - 1 } { \\lambda _ { 2 } } \\gamma ^ { 2 } \\Delta _ { t + 2 } ] , } \\\\ & { \\quad \\quad \\pi _ { s \\lambda ^ { ( 2 ) } ( L ) } ^ { R , a } } \\end{array}", + "type": "interline_equation", + "image_path": "0a9ed2938f048cb014418095c8bf22495e13c88841960e9512dce70e6eff25c2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 124, + 209, + 485, + 263.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 124, + 263.3333333333333, + 485, + 317.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 124, + 317.66666666666663, + 485, + 371.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 376, + 467, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 464, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 265, + 401 + ], + "score": 1.0, + "content": "where the last inequality holds because", + "type": "text" + }, + { + "bbox": [ + 265, + 376, + 304, + 394 + ], + "score": 0.94, + "content": "\\frac { \\gamma ^ { R } } { 1 - \\gamma } \\leq \\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 369, + 456, + 401 + ], + "score": 1.0, + "content": ", which follows from the definition of", + "type": "text" + }, + { + "bbox": [ + 456, + 381, + 464, + 390 + ], + "score": 0.84, + "content": "R", + "type": "inline_equation" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 503, + 428 + ], + "lines": [ + { + "bbox": [ + 103, + 398, + 504, + 418 + ], + "spans": [ + { + "bbox": [ + 103, + 398, + 239, + 418 + ], + "score": 1.0, + "content": "For any fixed trajectory of length", + "type": "text" + }, + { + "bbox": [ + 239, + 402, + 248, + 412 + ], + "score": 0.82, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 398, + 302, + 418 + ], + "score": 1.0, + "content": "starting from", + "type": "text" + }, + { + "bbox": [ + 302, + 404, + 311, + 413 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 398, + 404, + 418 + ], + "score": 1.0, + "content": ", consider the sequence", + "type": "text" + }, + { + "bbox": [ + 404, + 401, + 465, + 416 + ], + "score": 0.93, + "content": "\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 398, + 484, + 418 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 484, + 399, + 504, + 414 + ], + "score": 0.91, + "content": "X _ { t } ^ { ( i ) }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 413, + 402, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 133, + 430 + ], + "score": 1.0, + "content": "be the", + "type": "text" + }, + { + "bbox": [ + 133, + 416, + 138, + 425 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 413, + 210, + 430 + ], + "score": 1.0, + "content": "-th largest item of", + "type": "text" + }, + { + "bbox": [ + 210, + 415, + 271, + 429 + ], + "score": 0.93, + "content": "\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 413, + 402, + 430 + ], + "score": 1.0, + "content": ". Rearranging Eq. (1), we obtain", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 434, + 408, + 469 + ], + "lines": [ + { + "bbox": [ + 203, + 434, + 408, + 469 + ], + "spans": [ + { + "bbox": [ + 203, + 434, + 408, + 469 + ], + "score": 0.93, + "content": "V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E _ { t r a j } \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] .", + "type": "interline_equation", + "image_path": "353a318648b4e83bfc1d70cd63fedbd78d74dcde5d09d9484e43d06da944118d.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 434, + 408, + 451.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 203, + 451.5, + 408, + 469.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 432, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 430, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 263, + 494 + ], + "score": 1.0, + "content": "We first prove that Condition 1 implies", + "type": "text" + }, + { + "bbox": [ + 264, + 482, + 269, + 490 + ], + "score": 0.76, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 478, + 362, + 494 + ], + "score": 1.0, + "content": "-optimality at time step", + "type": "text" + }, + { + "bbox": [ + 363, + 481, + 368, + 490 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 478, + 393, + 494 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 393, + 480, + 430, + 492 + ], + "score": 0.92, + "content": "\\epsilon _ { 2 } = \\epsilon / 3", + "type": "inline_equation" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 391, + 517 + ], + "lines": [ + { + "bbox": [ + 102, + 494, + 387, + 518 + ], + "spans": [ + { + "bbox": [ + 102, + 494, + 179, + 518 + ], + "score": 1.0, + "content": "Condition 1. Let", + "type": "text" + }, + { + "bbox": [ + 180, + 495, + 281, + 517 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\xi _ { i } : = \\frac { 1 } { 2 ^ { i + 2 } } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 494, + 316, + 518 + ], + "score": 1.0, + "content": ". For all", + "type": "text" + }, + { + "bbox": [ + 316, + 501, + 387, + 514 + ], + "score": 0.88, + "content": "0 \\leq i \\leq \\lfloor \\log _ { 2 } R \\rfloor", + "type": "inline_equation" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 274, + 522, + 336, + 540 + ], + "lines": [ + { + "bbox": [ + 274, + 522, + 336, + 540 + ], + "spans": [ + { + "bbox": [ + 274, + 522, + 336, + 540 + ], + "score": 0.9, + "content": "E [ X _ { t } ^ { ( 2 ^ { i } ) } ] \\leq \\xi _ { i } .", + "type": "interline_equation", + "image_path": "4c5556442dd84c67ed2f8017702e406cf2d4d5acde09b8b3242d907ec73dda13.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 274, + 522, + 336, + 540 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 545, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 198, + 560 + ], + "score": 1.0, + "content": "Claim 1. If Condition", + "type": "text" + }, + { + "bbox": [ + 198, + 547, + 203, + 556 + ], + "score": 0.7, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 543, + 295, + 560 + ], + "score": 1.0, + "content": "is satisfied at time step", + "type": "text" + }, + { + "bbox": [ + 295, + 547, + 300, + 556 + ], + "score": 0.51, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 543, + 344, + 560 + ], + "score": 1.0, + "content": ", the policy", + "type": "text" + }, + { + "bbox": [ + 345, + 548, + 355, + 557 + ], + "score": 0.85, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 543, + 365, + 560 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 366, + 548, + 371, + 556 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 543, + 436, + 560 + ], + "score": 1.0, + "content": "-optimal at state", + "type": "text" + }, + { + "bbox": [ + 436, + 548, + 446, + 557 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 543, + 464, + 560 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 464, + 545, + 505, + 558 + ], + "score": 0.91, + "content": "V ^ { \\ast } ( s _ { t } ) -", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 554, + 163, + 571 + ], + "spans": [ + { + "bbox": [ + 107, + 557, + 157, + 569 + ], + "score": 0.91, + "content": "V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 554, + 163, + 571 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 108, + 581, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 103, + 579, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 103, + 579, + 180, + 600 + ], + "score": 1.0, + "content": "Proof. Note that", + "type": "text" + }, + { + "bbox": [ + 107, + 596, + 164, + 612 + ], + "score": 0.93, + "content": "E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 588, + 263, + 618 + ], + "score": 1.0, + "content": "t. Eq. (3) implies that for", + "type": "text" + }, + { + "bbox": [ + 180, + 582, + 200, + 597 + ], + "score": 0.93, + "content": "X _ { t } ^ { ( i ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 579, + 393, + 600 + ], + "score": 1.0, + "content": "is monotonically decreasing with respect to", + "type": "text" + }, + { + "bbox": [ + 263, + 600, + 317, + 612 + ], + "score": 0.91, + "content": "1 / 2 < \\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 588, + 327, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 393, + 586, + 398, + 595 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 579, + 456, + 600 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 456, + 581, + 505, + 597 + ], + "score": 0.93, + "content": "E [ X _ { t } ^ { ( i ) } ] \\ \\leq", + "type": "inline_equation" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 617, + 466, + 687 + ], + "lines": [ + { + "bbox": [ + 144, + 617, + 466, + 687 + ], + "spans": [ + { + "bbox": [ + 144, + 617, + 466, + 687 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { E \\left[ \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] = \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( i ) } ] \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ] } \\\\ & { \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } 2 ^ { - \\lfloor \\log _ { 2 } i \\rfloor - 2 } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\frac { \\gamma ^ { i - 1 } } { i } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq 2 \\epsilon _ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "62ba3bea0f084d11903c964028d906643070d34916bde88924637c5e08ea73d6.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 144, + 617, + 466, + 640.3333333333334 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 144, + 640.3333333333334, + 466, + 663.6666666666667 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 144, + 663.6666666666667, + 466, + 687.0000000000001 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 692, + 462, + 710 + ], + "lines": [ + { + "bbox": [ + 102, + 689, + 458, + 715 + ], + "spans": [ + { + "bbox": [ + 102, + 689, + 310, + 715 + ], + "score": 1.0, + "content": "where the last inequality follows from the fact that", + "type": "text" + }, + { + "bbox": [ + 310, + 692, + 405, + 711 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { \\infty } \\frac { \\gamma ^ { i - 1 } } { i } = \\frac { 1 } { \\gamma } \\ln \\frac { 1 } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 689, + 423, + 715 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 423, + 695, + 458, + 708 + ], + "score": 0.91, + "content": "\\gamma > 1 / 2", + "type": "inline_equation" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 102, + 714, + 476, + 735 + ], + "lines": [ + { + "bbox": [ + 103, + 712, + 475, + 737 + ], + "spans": [ + { + "bbox": [ + 103, + 712, + 238, + 737 + ], + "score": 1.0, + "content": "Combining with Eq. 2, we have,", + "type": "text" + }, + { + "bbox": [ + 238, + 715, + 475, + 736 + ], + "score": 0.87, + "content": "\\begin{array} { r } { V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] \\leq 3 \\epsilon _ { 2 } = \\epsilon . } \\end{array}", + "type": "inline_equation" + } + ], + "index": 26 + } + ], + "index": 26 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 719, + 505, + 730 + ], + "lines": [ + { + "bbox": [ + 496, + 721, + 504, + 729 + ], + "spans": [ + { + "bbox": [ + 496, + 721, + 504, + 729 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The remaining part of this section is organized as follows. We establish the sufficient condition for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 96, + 112, + 104 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "-optimality in Section 3.3. The key lemma is presented in Section 3.4. Finally we prove Theorem 1", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 167, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 167, + 116 + ], + "score": 1.0, + "content": "in Section 3.5.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 116 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 129, + 318, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 319, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 255, + 142 + ], + "score": 1.0, + "content": "3.3 SUFFICIENT CONDITION FOR", + "type": "text" + }, + { + "bbox": [ + 255, + 132, + 261, + 140 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 129, + 319, + 142 + ], + "score": 1.0, + "content": "-OPTIMALITY", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 483, + 163 + ], + "lines": [ + { + "bbox": [ + 105, + 148, + 484, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 376, + 165 + ], + "score": 1.0, + "content": "In this section, we establish a sufficient condition (Condition 2) for", + "type": "text" + }, + { + "bbox": [ + 376, + 153, + 381, + 160 + ], + "score": 0.65, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 148, + 475, + 165 + ], + "score": 1.0, + "content": "-optimality at time step", + "type": "text" + }, + { + "bbox": [ + 475, + 152, + 480, + 160 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 148, + 484, + 165 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 148, + 484, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 167, + 506, + 203 + ], + "lines": [ + { + "bbox": [ + 105, + 165, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 152, + 181 + ], + "score": 1.0, + "content": "For a fixed", + "type": "text" + }, + { + "bbox": [ + 153, + 169, + 162, + 178 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 165, + 204, + 181 + ], + "score": 1.0, + "content": ", let TRAJ", + "type": "text" + }, + { + "bbox": [ + 204, + 168, + 218, + 179 + ], + "score": 0.66, + "content": "( R )", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 165, + 300, + 181 + ], + "score": 1.0, + "content": "be the set of length-", + "type": "text" + }, + { + "bbox": [ + 301, + 168, + 309, + 177 + ], + "score": 0.8, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 165, + 413, + 181 + ], + "score": 1.0, + "content": "trajectories starting from", + "type": "text" + }, + { + "bbox": [ + 414, + 170, + 423, + 178 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 165, + 506, + 181 + ], + "score": 1.0, + "content": ". Our goal is to give", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 178, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 225, + 191 + ], + "score": 1.0, + "content": "a sufficient condition so that", + "type": "text" + }, + { + "bbox": [ + 226, + 180, + 236, + 190 + ], + "score": 0.85, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 178, + 345, + 191 + ], + "score": 1.0, + "content": ", the policy learned at step", + "type": "text" + }, + { + "bbox": [ + 346, + 180, + 351, + 189 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 178, + 364, + 191 + ], + "score": 1.0, + "content": ", is", + "type": "text" + }, + { + "bbox": [ + 365, + 181, + 370, + 188 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 178, + 444, + 191 + ], + "score": 1.0, + "content": "-optimal. For any", + "type": "text" + }, + { + "bbox": [ + 444, + 179, + 473, + 190 + ], + "score": 0.91, + "content": "\\epsilon _ { 2 } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 178, + 505, + 191 + ], + "score": 1.0, + "content": ", define", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 187, + 406, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 216, + 204 + ], + "score": 0.93, + "content": "\\begin{array} { r } { R : = \\lceil \\ln \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 187, + 252, + 206 + ], + "score": 1.0, + "content": ". Denote", + "type": "text" + }, + { + "bbox": [ + 253, + 189, + 335, + 201 + ], + "score": 0.94, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 187, + 349, + 206 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 349, + 190, + 362, + 200 + ], + "score": 0.88, + "content": "\\Delta _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 187, + 406, + 206 + ], + "score": 1.0, + "content": ". We have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 165, + 506, + 206 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 124, + 209, + 485, + 372 + ], + "lines": [ + { + "bbox": [ + 124, + 209, + 485, + 372 + ], + "spans": [ + { + "bbox": [ + 124, + 209, + 485, + 372 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\quad V ^ { * } ( s _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + Q ^ { * } ( s _ { t } , a _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { P } ( V ^ { * } - V ^ { \\pi } ) ( s _ { t } , \\pi _ { t } ( s _ { t } ) ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\sum _ { s + 1 } ^ { \\infty } p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\cdot [ V ^ { * } ( s _ { t + 1 } ) - Q ^ { * } ( s _ { t + 1 } , a _ { t + 1 } ) ] + } \\\\ & { \\quad \\Big . \\Big . \\Big . \\Big . \\Big . \\Big . } \\\\ & { \\quad \\quad \\Big . \\gamma _ { s + 1 , s ^ { * } + 2 } \\mathrm { ~ } p ( s _ { t + 2 } | s _ { t + 1 } , \\pi _ { t + 1 } ( s _ { t + 1 } ) ) \\cdot p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\big [ V ^ { * } ( s _ { t + 2 } ) - Q ^ { * } ( s _ { t + 2 } , a _ { t + 2 } ) \\Big ] } \\\\ & { \\quad \\quad \\cdot \\dots } \\\\ & { \\le \\epsilon _ { 2 } + \\displaystyle \\sum _ { s ^ { \\prime } = s ^ { \\prime } } p ( t r a _ { t } ) \\cdot [ \\frac { n - 1 } { \\lambda _ { 2 } } \\gamma ^ { 2 } \\Delta _ { t + 2 } ] , } \\\\ & { \\quad \\quad \\pi _ { s \\lambda ^ { ( 2 ) } ( L ) } ^ { R , a } } \\end{array}", + "type": "interline_equation", + "image_path": "0a9ed2938f048cb014418095c8bf22495e13c88841960e9512dce70e6eff25c2.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 124, + 209, + 485, + 263.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 124, + 263.3333333333333, + 485, + 317.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 124, + 317.66666666666663, + 485, + 371.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 376, + 467, + 394 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 464, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 265, + 401 + ], + "score": 1.0, + "content": "where the last inequality holds because", + "type": "text" + }, + { + "bbox": [ + 265, + 376, + 304, + 394 + ], + "score": 0.94, + "content": "\\frac { \\gamma ^ { R } } { 1 - \\gamma } \\leq \\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 369, + 456, + 401 + ], + "score": 1.0, + "content": ", which follows from the definition of", + "type": "text" + }, + { + "bbox": [ + 456, + 381, + 464, + 390 + ], + "score": 0.84, + "content": "R", + "type": "inline_equation" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 369, + 464, + 401 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 503, + 428 + ], + "lines": [ + { + "bbox": [ + 103, + 398, + 504, + 418 + ], + "spans": [ + { + "bbox": [ + 103, + 398, + 239, + 418 + ], + "score": 1.0, + "content": "For any fixed trajectory of length", + "type": "text" + }, + { + "bbox": [ + 239, + 402, + 248, + 412 + ], + "score": 0.82, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 398, + 302, + 418 + ], + "score": 1.0, + "content": "starting from", + "type": "text" + }, + { + "bbox": [ + 302, + 404, + 311, + 413 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 398, + 404, + 418 + ], + "score": 1.0, + "content": ", consider the sequence", + "type": "text" + }, + { + "bbox": [ + 404, + 401, + 465, + 416 + ], + "score": 0.93, + "content": "\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 398, + 484, + 418 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 484, + 399, + 504, + 414 + ], + "score": 0.91, + "content": "X _ { t } ^ { ( i ) }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 413, + 402, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 133, + 430 + ], + "score": 1.0, + "content": "be the", + "type": "text" + }, + { + "bbox": [ + 133, + 416, + 138, + 425 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 413, + 210, + 430 + ], + "score": 1.0, + "content": "-th largest item of", + "type": "text" + }, + { + "bbox": [ + 210, + 415, + 271, + 429 + ], + "score": 0.93, + "content": "\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 413, + 402, + 430 + ], + "score": 1.0, + "content": ". Rearranging Eq. (1), we obtain", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 103, + 398, + 504, + 430 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 434, + 408, + 469 + ], + "lines": [ + { + "bbox": [ + 203, + 434, + 408, + 469 + ], + "spans": [ + { + "bbox": [ + 203, + 434, + 408, + 469 + ], + "score": 0.93, + "content": "V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E _ { t r a j } \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] .", + "type": "interline_equation", + "image_path": "353a318648b4e83bfc1d70cd63fedbd78d74dcde5d09d9484e43d06da944118d.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 434, + 408, + 451.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 203, + 451.5, + 408, + 469.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 432, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 430, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 263, + 494 + ], + "score": 1.0, + "content": "We first prove that Condition 1 implies", + "type": "text" + }, + { + "bbox": [ + 264, + 482, + 269, + 490 + ], + "score": 0.76, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 478, + 362, + 494 + ], + "score": 1.0, + "content": "-optimality at time step", + "type": "text" + }, + { + "bbox": [ + 363, + 481, + 368, + 490 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 478, + 393, + 494 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 393, + 480, + 430, + 492 + ], + "score": 0.92, + "content": "\\epsilon _ { 2 } = \\epsilon / 3", + "type": "inline_equation" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 478, + 430, + 494 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 496, + 391, + 517 + ], + "lines": [ + { + "bbox": [ + 102, + 494, + 387, + 518 + ], + "spans": [ + { + "bbox": [ + 102, + 494, + 179, + 518 + ], + "score": 1.0, + "content": "Condition 1. Let", + "type": "text" + }, + { + "bbox": [ + 180, + 495, + 281, + 517 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\xi _ { i } : = \\frac { 1 } { 2 ^ { i + 2 } } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 494, + 316, + 518 + ], + "score": 1.0, + "content": ". For all", + "type": "text" + }, + { + "bbox": [ + 316, + 501, + 387, + 514 + ], + "score": 0.88, + "content": "0 \\leq i \\leq \\lfloor \\log _ { 2 } R \\rfloor", + "type": "inline_equation" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 102, + 494, + 387, + 518 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 274, + 522, + 336, + 540 + ], + "lines": [ + { + "bbox": [ + 274, + 522, + 336, + 540 + ], + "spans": [ + { + "bbox": [ + 274, + 522, + 336, + 540 + ], + "score": 0.9, + "content": "E [ X _ { t } ^ { ( 2 ^ { i } ) } ] \\leq \\xi _ { i } .", + "type": "interline_equation", + "image_path": "4c5556442dd84c67ed2f8017702e406cf2d4d5acde09b8b3242d907ec73dda13.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 274, + 522, + 336, + 540 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "list", + "bbox": [ + 106, + 545, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 198, + 560 + ], + "score": 1.0, + "content": "Claim 1. If Condition", + "type": "text" + }, + { + "bbox": [ + 198, + 547, + 203, + 556 + ], + "score": 0.7, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 543, + 295, + 560 + ], + "score": 1.0, + "content": "is satisfied at time step", + "type": "text" + }, + { + "bbox": [ + 295, + 547, + 300, + 556 + ], + "score": 0.51, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 543, + 344, + 560 + ], + "score": 1.0, + "content": ", the policy", + "type": "text" + }, + { + "bbox": [ + 345, + 548, + 355, + 557 + ], + "score": 0.85, + "content": "\\pi _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 543, + 365, + 560 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 366, + 548, + 371, + 556 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 543, + 436, + 560 + ], + "score": 1.0, + "content": "-optimal at state", + "type": "text" + }, + { + "bbox": [ + 436, + 548, + 446, + 557 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 543, + 464, + 560 + ], + "score": 1.0, + "content": ", i.e.", + "type": "text" + }, + { + "bbox": [ + 464, + 545, + 505, + 558 + ], + "score": 0.91, + "content": "V ^ { \\ast } ( s _ { t } ) -", + "type": "inline_equation" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 554, + 163, + 571 + ], + "spans": [ + { + "bbox": [ + 107, + 557, + 157, + 569 + ], + "score": 0.91, + "content": "V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 554, + 163, + 571 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 543, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 581, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 103, + 579, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 103, + 579, + 180, + 600 + ], + "score": 1.0, + "content": "Proof. Note that", + "type": "text" + }, + { + "bbox": [ + 107, + 596, + 164, + 612 + ], + "score": 0.93, + "content": "E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 588, + 263, + 618 + ], + "score": 1.0, + "content": "t. Eq. (3) implies that for", + "type": "text" + }, + { + "bbox": [ + 180, + 582, + 200, + 597 + ], + "score": 0.93, + "content": "X _ { t } ^ { ( i ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 579, + 393, + 600 + ], + "score": 1.0, + "content": "is monotonically decreasing with respect to", + "type": "text" + }, + { + "bbox": [ + 263, + 600, + 317, + 612 + ], + "score": 0.91, + "content": "1 / 2 < \\gamma < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 588, + 327, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 393, + 586, + 398, + 595 + ], + "score": 0.73, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 579, + 456, + 600 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 456, + 581, + 505, + 597 + ], + "score": 0.93, + "content": "E [ X _ { t } ^ { ( i ) } ] \\ \\leq", + "type": "inline_equation" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 103, + 579, + 505, + 618 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 144, + 617, + 466, + 687 + ], + "lines": [ + { + "bbox": [ + 144, + 617, + 466, + 687 + ], + "spans": [ + { + "bbox": [ + 144, + 617, + 466, + 687 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { E \\left[ \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] = \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( i ) } ] \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ] } \\\\ & { \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } 2 ^ { - \\lfloor \\log _ { 2 } i \\rfloor - 2 } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\frac { \\gamma ^ { i - 1 } } { i } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq 2 \\epsilon _ { 2 } , } \\end{array}", + "type": "interline_equation", + "image_path": "62ba3bea0f084d11903c964028d906643070d34916bde88924637c5e08ea73d6.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 144, + 617, + 466, + 640.3333333333334 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 144, + 640.3333333333334, + 466, + 663.6666666666667 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 144, + 663.6666666666667, + 466, + 687.0000000000001 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 692, + 462, + 710 + ], + "lines": [ + { + "bbox": [ + 102, + 689, + 458, + 715 + ], + "spans": [ + { + "bbox": [ + 102, + 689, + 310, + 715 + ], + "score": 1.0, + "content": "where the last inequality follows from the fact that", + "type": "text" + }, + { + "bbox": [ + 310, + 692, + 405, + 711 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { \\infty } \\frac { \\gamma ^ { i - 1 } } { i } = \\frac { 1 } { \\gamma } \\ln \\frac { 1 } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 689, + 423, + 715 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 423, + 695, + 458, + 708 + ], + "score": 0.91, + "content": "\\gamma > 1 / 2", + "type": "inline_equation" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 102, + 689, + 458, + 715 + ] + }, + { + "type": "text", + "bbox": [ + 102, + 714, + 476, + 735 + ], + "lines": [ + { + "bbox": [ + 103, + 712, + 475, + 737 + ], + "spans": [ + { + "bbox": [ + 103, + 712, + 238, + 737 + ], + "score": 1.0, + "content": "Combining with Eq. 2, we have,", + "type": "text" + }, + { + "bbox": [ + 238, + 715, + 475, + 736 + ], + "score": 0.87, + "content": "\\begin{array} { r } { V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] \\leq 3 \\epsilon _ { 2 } = \\epsilon . } \\end{array}", + "type": "inline_equation" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 103, + 712, + 475, + 737 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 338, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 339, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 207, + 97 + ], + "score": 1.0, + "content": "Next we show that given", + "type": "text" + }, + { + "bbox": [ + 207, + 83, + 220, + 94 + ], + "score": 0.82, + "content": "i , t", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 79, + 339, + 97 + ], + "score": 1.0, + "content": ", Condition 2 implies Eq. (3).", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 97, + 506, + 134 + ], + "lines": [ + { + "bbox": [ + 103, + 95, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 103, + 95, + 194, + 120 + ], + "score": 1.0, + "content": "Condition 2. Define", + "type": "text" + }, + { + "bbox": [ + 194, + 101, + 253, + 114 + ], + "score": 0.91, + "content": "L = \\lfloor \\log _ { 2 } R \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 95, + 273, + 120 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 273, + 97, + 416, + 118 + ], + "score": 0.87, + "content": "\\begin{array} { r } { M = \\operatorname* { m a x } \\left\\{ \\lceil 2 \\log _ { 2 } \\frac { 1 } { \\xi _ { L } ( 1 - \\gamma ) } \\rceil , 1 0 \\right\\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 95, + 439, + 120 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 439, + 99, + 502, + 115 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\eta _ { j } = \\frac { \\xi _ { i } } { M } \\cdot 2 ^ { j - 1 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 95, + 507, + 120 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 115, + 300, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 135, + 136 + ], + "score": 1.0, + "content": "For all", + "type": "text" + }, + { + "bbox": [ + 140, + 117, + 296, + 134 + ], + "score": 0.75, + "content": "\\begin{array} { r } { \\because j \\le M , \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] \\le \\frac { \\xi _ { i } } { M } , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 115, + 300, + 136 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 136, + 350, + 148 + ], + "lines": [ + { + "bbox": [ + 107, + 136, + 349, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 136, + 349, + 149 + ], + "score": 1.0, + "content": "Claim 2. Given i, t, Eq. (3) holds if Condition 2 is satisfied.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 104, + 161, + 506, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 273, + 175 + ], + "score": 1.0, + "content": "Proof. The reason behind the choice of", + "type": "text" + }, + { + "bbox": [ + 273, + 163, + 285, + 173 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 162, + 356, + 175 + ], + "score": 1.0, + "content": "is to ensure that", + "type": "text" + }, + { + "bbox": [ + 357, + 162, + 433, + 175 + ], + "score": 0.91, + "content": "\\eta _ { M } > 1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "1. It follows that,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 173, + 290, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 237, + 186 + ], + "score": 1.0, + "content": "assuming Condition 2 holds, for", + "type": "text" + }, + { + "bbox": [ + 237, + 174, + 285, + 186 + ], + "score": 0.92, + "content": "1 \\leq j \\leq M", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 173, + 290, + 186 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 139, + 191, + 471, + 227 + ], + "lines": [ + { + "bbox": [ + 139, + 191, + 471, + 227 + ], + "spans": [ + { + "bbox": [ + 139, + 191, + 471, + 227 + ], + "score": 0.93, + "content": "E \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } \\right] = \\int _ { 0 } ^ { 1 / ( 1 - \\gamma ) } \\operatorname* { P r } \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } > x \\right] d x \\le \\eta _ { 1 } + \\sum _ { j = 2 } ^ { M } \\eta _ { j } \\operatorname* { P r } [ { X _ { t } ^ { ( 2 ^ { i } ) } } > \\eta _ { j - 1 } ] \\le \\xi _ { i } .", + "type": "interline_equation", + "image_path": "b1e98198c486d877466252bbfdd5d0afb8e4019bc37fc19f80c274168b0bb6a3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 139, + 191, + 471, + 203.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 139, + 203.0, + 471, + 215.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 139, + 215.0, + 471, + 227.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 259, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 204, + 273 + ], + "score": 1.0, + "content": "Therefore, if a time step", + "type": "text" + }, + { + "bbox": [ + 204, + 262, + 209, + 270 + ], + "score": 0.79, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 258, + 234, + 273 + ], + "score": 1.0, + "content": "is not", + "type": "text" + }, + { + "bbox": [ + 234, + 262, + 244, + 271 + ], + "score": 0.85, + "content": "\\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 258, + 328, + 273 + ], + "score": 1.0, + "content": "-optimal, there exists", + "type": "text" + }, + { + "bbox": [ + 329, + 260, + 399, + 272 + ], + "score": 0.93, + "content": "0 \\leq i < \\lfloor \\log _ { 2 } R \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 258, + 418, + 273 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 418, + 260, + 466, + 272 + ], + "score": 0.91, + "content": "2 \\leq j \\leq M", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 277, + 363, + 302 + ], + "lines": [ + { + "bbox": [ + 248, + 277, + 363, + 302 + ], + "spans": [ + { + "bbox": [ + 248, + 277, + 363, + 302 + ], + "score": 0.93, + "content": "\\eta _ { j } \\mathrm { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } .", + "type": "interline_equation", + "image_path": "7d918d021bc39256ac2de308ee455c6a54396bc36335cd0f685f840f0c5005d1.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 248, + 277, + 363, + 302 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 506, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 360, + 321 + ], + "score": 1.0, + "content": "Now, the sample complexity can be bounded by the number of", + "type": "text" + }, + { + "bbox": [ + 360, + 308, + 388, + 320 + ], + "score": 0.92, + "content": "( t , i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "pairs that Eq. (4) is violated.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 336, + 332 + ], + "score": 1.0, + "content": "Following the approach of Strehl et al. (2006), for a fixed", + "type": "text" + }, + { + "bbox": [ + 336, + 319, + 357, + 331 + ], + "score": 0.92, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "-pair, instead of directly counting the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 101, + 330, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 101, + 330, + 193, + 351 + ], + "score": 1.0, + "content": "number of time steps", + "type": "text" + }, + { + "bbox": [ + 194, + 335, + 199, + 344 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 330, + 238, + 351 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 238, + 330, + 343, + 349 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 330, + 505, + 351 + ], + "score": 1.0, + "content": ", we count the number of time steps that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 348, + 409, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 163, + 365 + ], + "score": 0.92, + "content": "X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 349, + 399, + 366 + ], + "score": 1.0, + "content": ". Lemma 1 provides an upper bound of the number of such", + "type": "text" + }, + { + "bbox": [ + 399, + 353, + 404, + 362 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 349, + 409, + 366 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 107, + 378, + 192, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 193, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 193, + 391 + ], + "score": 1.0, + "content": "3.4 KEY LEMMAS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 398, + 506, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 507, + 412 + ], + "score": 1.0, + "content": "In this section, we present two key lemmas. Lemma 1 bounds the number of sub-optimal actions,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "score": 1.0, + "content": "which in turn, bounds the sample complexity of our algorithm. Lemma 2 bounds the weighted sum", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 192, + 435 + ], + "score": 1.0, + "content": "of learning error, i.e.", + "type": "text" + }, + { + "bbox": [ + 192, + 421, + 259, + 434 + ], + "score": 0.92, + "content": "( \\hat { Q } _ { t } - Q ^ { * } ) ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 421, + 506, + 435 + ], + "score": 1.0, + "content": ", with the sum and maximum of weights. Then, we show that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 433, + 244, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 244, + 445 + ], + "score": 1.0, + "content": "Lemma 1 follows from Lemma 2.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 105, + 446, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 104, + 443, + 507, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 191, + 465 + ], + "score": 1.0, + "content": "Lemma 1. For fixed", + "type": "text" + }, + { + "bbox": [ + 191, + 450, + 196, + 459 + ], + "score": 0.25, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 443, + 214, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 215, + 449, + 240, + 461 + ], + "score": 0.89, + "content": "\\eta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 443, + 255, + 465 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 256, + 446, + 275, + 461 + ], + "score": 0.91, + "content": "{ B } _ { \\eta } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 443, + 344, + 465 + ], + "score": 1.0, + "content": "be the event that", + "type": "text" + }, + { + "bbox": [ + 344, + 448, + 457, + 463 + ], + "score": 0.91, + "content": "\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 443, + 507, + 465 + ], + "score": 1.0, + "content": "in step t. If", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 461, + 298, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 140, + 474 + ], + "score": 0.89, + "content": "\\eta > 2 \\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 461, + 261, + 475 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 261, + 462, + 293, + 474 + ], + "score": 0.9, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 461, + 298, + 475 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 479, + 423, + 514 + ], + "lines": [ + { + "bbox": [ + 186, + 479, + 423, + 514 + ], + "spans": [ + { + "bbox": [ + 186, + 479, + 423, + 514 + ], + "score": 0.92, + "content": "\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ,", + "type": "interline_equation", + "image_path": "8e6137e01a7beea297bf95283d863b21a785d17463e6944e2efedb515bf7d808.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 479, + 423, + 496.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 186, + 496.5, + 423, + 514.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 248, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 249, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 133, + 533 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 520, + 147, + 532 + ], + "score": 0.87, + "content": "I [ \\cdot ]", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 520, + 249, + 533 + ], + "score": 1.0, + "content": "is the indicator function.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 541, + 440, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 440, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 440, + 555 + ], + "score": 1.0, + "content": "Before presenting Lemma 2, we define a class of sequence that occurs in the proof.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 556, + 504, + 581 + ], + "lines": [ + { + "bbox": [ + 104, + 554, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 104, + 554, + 212, + 571 + ], + "score": 1.0, + "content": "Definition 3. A sequence", + "type": "text" + }, + { + "bbox": [ + 212, + 557, + 245, + 569 + ], + "score": 0.91, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 554, + 305, + 571 + ], + "score": 1.0, + "content": "is said to be a", + "type": "text" + }, + { + "bbox": [ + 305, + 557, + 332, + 569 + ], + "score": 0.9, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 554, + 388, + 571 + ], + "score": 1.0, + "content": "-sequence for", + "type": "text" + }, + { + "bbox": [ + 389, + 557, + 427, + 568 + ], + "score": 0.82, + "content": "C , w > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 554, + 431, + 571 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 432, + 557, + 489, + 569 + ], + "score": 0.79, + "content": "i f 0 \\le w _ { t } \\le w", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 554, + 506, + 571 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 567, + 227, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 119, + 582 + ], + "score": 1.0, + "content": "all", + "type": "text" + }, + { + "bbox": [ + 120, + 568, + 142, + 579 + ], + "score": 0.88, + "content": "t \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 567, + 164, + 582 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 164, + 568, + 222, + 582 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\sum _ { t \\geq 1 } w _ { t } \\leq C } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 567, + 227, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 105, + 584, + 485, + 598 + ], + "lines": [ + { + "bbox": [ + 104, + 583, + 486, + 599 + ], + "spans": [ + { + "bbox": [ + 104, + 583, + 195, + 599 + ], + "score": 1.0, + "content": "Lemma 2. For every", + "type": "text" + }, + { + "bbox": [ + 195, + 585, + 222, + 597 + ], + "score": 0.91, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 583, + 264, + 599 + ], + "score": 1.0, + "content": "-sequence", + "type": "text" + }, + { + "bbox": [ + 264, + 585, + 297, + 597 + ], + "score": 0.92, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 583, + 366, + 599 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + }, + { + "bbox": [ + 367, + 585, + 400, + 597 + ], + "score": 0.9, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 583, + 486, + 599 + ], + "score": 1.0, + "content": ", the following holds:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 617, + 483, + 651 + ], + "lines": [ + { + "bbox": [ + 127, + 617, + 483, + 651 + ], + "spans": [ + { + "bbox": [ + 127, + 617, + 483, + 651 + ], + "score": 0.92, + "content": "\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "576704e0ba7bc113c9eb1c2218a2bd6ace18a8d09188a27f397409e00cdaad49.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 127, + 617, + 483, + 628.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 127, + 628.3333333333334, + 483, + 639.6666666666667 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 127, + 639.6666666666667, + 483, + 651.0000000000001 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 664, + 289, + 681 + ], + "lines": [ + { + "bbox": [ + 133, + 665, + 228, + 681 + ], + "spans": [ + { + "bbox": [ + 133, + 665, + 228, + 681 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 104, + 689, + 468, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 468, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 468, + 703 + ], + "score": 1.0, + "content": "Proof of Lemma 2 is quite technical, and is therefore deferred to supplementary materials.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 233, + 505, + 244 + ], + "lines": [ + { + "bbox": [ + 495, + 234, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 495, + 234, + 505, + 245 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 710, + 507, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 707, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 119, + 707, + 124, + 724 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 125, + 711, + 186, + 722 + ], + "score": 0.88, + "content": "\\eta _ { M } > 1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 707, + 334, + 724 + ], + "score": 1.0, + "content": "can be verified by combining inequalities", + "type": "text" + }, + { + "bbox": [ + 334, + 710, + 415, + 723 + ], + "score": 0.92, + "content": "\\xi _ { i } \\cdot 2 ^ { M / 2 } \\geq 1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 707, + 431, + 724 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 431, + 709, + 505, + 722 + ], + "score": 0.93, + "content": "2 ^ { M / 2 - 1 } > ( M + 1 )", + "type": "inline_equation" + } + ] + }, + { + "bbox": [ + 106, + 721, + 183, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 168, + 732 + ], + "score": 1.0, + "content": "for large enough", + "type": "text" + }, + { + "bbox": [ + 168, + 722, + 179, + 730 + ], + "score": 0.8, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 721, + 183, + 732 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 338, + 95 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 339, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 207, + 97 + ], + "score": 1.0, + "content": "Next we show that given", + "type": "text" + }, + { + "bbox": [ + 207, + 83, + 220, + 94 + ], + "score": 0.82, + "content": "i , t", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 79, + 339, + 97 + ], + "score": 1.0, + "content": ", Condition 2 implies Eq. (3).", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 79, + 339, + 97 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 97, + 506, + 134 + ], + "lines": [ + { + "bbox": [ + 103, + 95, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 103, + 95, + 194, + 120 + ], + "score": 1.0, + "content": "Condition 2. Define", + "type": "text" + }, + { + "bbox": [ + 194, + 101, + 253, + 114 + ], + "score": 0.91, + "content": "L = \\lfloor \\log _ { 2 } R \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 95, + 273, + 120 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 273, + 97, + 416, + 118 + ], + "score": 0.87, + "content": "\\begin{array} { r } { M = \\operatorname* { m a x } \\left\\{ \\lceil 2 \\log _ { 2 } \\frac { 1 } { \\xi _ { L } ( 1 - \\gamma ) } \\rceil , 1 0 \\right\\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 95, + 439, + 120 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 439, + 99, + 502, + 115 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\eta _ { j } = \\frac { \\xi _ { i } } { M } \\cdot 2 ^ { j - 1 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 95, + 507, + 120 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 115, + 300, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 135, + 136 + ], + "score": 1.0, + "content": "For all", + "type": "text" + }, + { + "bbox": [ + 140, + 117, + 296, + 134 + ], + "score": 0.75, + "content": "\\begin{array} { r } { \\because j \\le M , \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] \\le \\frac { \\xi _ { i } } { M } , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 115, + 300, + 136 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 1.5, + "bbox_fs": [ + 103, + 95, + 507, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 136, + 350, + 148 + ], + "lines": [ + { + "bbox": [ + 107, + 136, + 349, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 136, + 349, + 149 + ], + "score": 1.0, + "content": "Claim 2. Given i, t, Eq. (3) holds if Condition 2 is satisfied.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 107, + 136, + 349, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 161, + 506, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 273, + 175 + ], + "score": 1.0, + "content": "Proof. The reason behind the choice of", + "type": "text" + }, + { + "bbox": [ + 273, + 163, + 285, + 173 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 162, + 356, + 175 + ], + "score": 1.0, + "content": "is to ensure that", + "type": "text" + }, + { + "bbox": [ + 357, + 162, + 433, + 175 + ], + "score": 0.91, + "content": "\\eta _ { M } > 1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 162, + 506, + 175 + ], + "score": 1.0, + "content": "1. It follows that,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 173, + 290, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 237, + 186 + ], + "score": 1.0, + "content": "assuming Condition 2 holds, for", + "type": "text" + }, + { + "bbox": [ + 237, + 174, + 285, + 186 + ], + "score": 0.92, + "content": "1 \\leq j \\leq M", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 173, + 290, + 186 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 162, + 506, + 186 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 139, + 191, + 471, + 227 + ], + "lines": [ + { + "bbox": [ + 139, + 191, + 471, + 227 + ], + "spans": [ + { + "bbox": [ + 139, + 191, + 471, + 227 + ], + "score": 0.93, + "content": "E \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } \\right] = \\int _ { 0 } ^ { 1 / ( 1 - \\gamma ) } \\operatorname* { P r } \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } > x \\right] d x \\le \\eta _ { 1 } + \\sum _ { j = 2 } ^ { M } \\eta _ { j } \\operatorname* { P r } [ { X _ { t } ^ { ( 2 ^ { i } ) } } > \\eta _ { j - 1 } ] \\le \\xi _ { i } .", + "type": "interline_equation", + "image_path": "b1e98198c486d877466252bbfdd5d0afb8e4019bc37fc19f80c274168b0bb6a3.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 139, + 191, + 471, + 203.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 139, + 203.0, + 471, + 215.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 139, + 215.0, + 471, + 227.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 259, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 204, + 273 + ], + "score": 1.0, + "content": "Therefore, if a time step", + "type": "text" + }, + { + "bbox": [ + 204, + 262, + 209, + 270 + ], + "score": 0.79, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 258, + 234, + 273 + ], + "score": 1.0, + "content": "is not", + "type": "text" + }, + { + "bbox": [ + 234, + 262, + 244, + 271 + ], + "score": 0.85, + "content": "\\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 258, + 328, + 273 + ], + "score": 1.0, + "content": "-optimal, there exists", + "type": "text" + }, + { + "bbox": [ + 329, + 260, + 399, + 272 + ], + "score": 0.93, + "content": "0 \\leq i < \\lfloor \\log _ { 2 } R \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 258, + 418, + 273 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 418, + 260, + 466, + 272 + ], + "score": 0.91, + "content": "2 \\leq j \\leq M", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 258, + 506, + 273 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 258, + 506, + 273 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 248, + 277, + 363, + 302 + ], + "lines": [ + { + "bbox": [ + 248, + 277, + 363, + 302 + ], + "spans": [ + { + "bbox": [ + 248, + 277, + 363, + 302 + ], + "score": 0.93, + "content": "\\eta _ { j } \\mathrm { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } .", + "type": "interline_equation", + "image_path": "7d918d021bc39256ac2de308ee455c6a54396bc36335cd0f685f840f0c5005d1.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 248, + 277, + 363, + 302 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 506, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 360, + 321 + ], + "score": 1.0, + "content": "Now, the sample complexity can be bounded by the number of", + "type": "text" + }, + { + "bbox": [ + 360, + 308, + 388, + 320 + ], + "score": 0.92, + "content": "( t , i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "pairs that Eq. (4) is violated.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 336, + 332 + ], + "score": 1.0, + "content": "Following the approach of Strehl et al. (2006), for a fixed", + "type": "text" + }, + { + "bbox": [ + 336, + 319, + 357, + 331 + ], + "score": 0.92, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "-pair, instead of directly counting the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 101, + 330, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 101, + 330, + 193, + 351 + ], + "score": 1.0, + "content": "number of time steps", + "type": "text" + }, + { + "bbox": [ + 194, + 335, + 199, + 344 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 330, + 238, + 351 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 238, + 330, + 343, + 349 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 330, + 505, + 351 + ], + "score": 1.0, + "content": ", we count the number of time steps that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 348, + 409, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 163, + 365 + ], + "score": 0.92, + "content": "X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 349, + 399, + 366 + ], + "score": 1.0, + "content": ". Lemma 1 provides an upper bound of the number of such", + "type": "text" + }, + { + "bbox": [ + 399, + 353, + 404, + 362 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 349, + 409, + 366 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 101, + 307, + 506, + 366 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 378, + 192, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 193, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 193, + 391 + ], + "score": 1.0, + "content": "3.4 KEY LEMMAS", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 398, + 506, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 507, + 412 + ], + "score": 1.0, + "content": "In this section, we present two key lemmas. Lemma 1 bounds the number of sub-optimal actions,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 422 + ], + "score": 1.0, + "content": "which in turn, bounds the sample complexity of our algorithm. Lemma 2 bounds the weighted sum", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 192, + 435 + ], + "score": 1.0, + "content": "of learning error, i.e.", + "type": "text" + }, + { + "bbox": [ + 192, + 421, + 259, + 434 + ], + "score": 0.92, + "content": "( \\hat { Q } _ { t } - Q ^ { * } ) ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 421, + 506, + 435 + ], + "score": 1.0, + "content": ", with the sum and maximum of weights. Then, we show that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 433, + 244, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 244, + 445 + ], + "score": 1.0, + "content": "Lemma 1 follows from Lemma 2.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 397, + 507, + 445 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 446, + 505, + 474 + ], + "lines": [ + { + "bbox": [ + 104, + 443, + 507, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 191, + 465 + ], + "score": 1.0, + "content": "Lemma 1. For fixed", + "type": "text" + }, + { + "bbox": [ + 191, + 450, + 196, + 459 + ], + "score": 0.25, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 443, + 214, + 465 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 215, + 449, + 240, + 461 + ], + "score": 0.89, + "content": "\\eta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 443, + 255, + 465 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 256, + 446, + 275, + 461 + ], + "score": 0.91, + "content": "{ B } _ { \\eta } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 443, + 344, + 465 + ], + "score": 1.0, + "content": "be the event that", + "type": "text" + }, + { + "bbox": [ + 344, + 448, + 457, + 463 + ], + "score": 0.91, + "content": "\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 443, + 507, + 465 + ], + "score": 1.0, + "content": "in step t. If", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 461, + 298, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 140, + 474 + ], + "score": 0.89, + "content": "\\eta > 2 \\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 461, + 261, + 475 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 261, + 462, + 293, + 474 + ], + "score": 0.9, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 461, + 298, + 475 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 443, + 507, + 475 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 479, + 423, + 514 + ], + "lines": [ + { + "bbox": [ + 186, + 479, + 423, + 514 + ], + "spans": [ + { + "bbox": [ + 186, + 479, + 423, + 514 + ], + "score": 0.92, + "content": "\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ,", + "type": "interline_equation", + "image_path": "8e6137e01a7beea297bf95283d863b21a785d17463e6944e2efedb515bf7d808.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 479, + 423, + 496.5 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 186, + 496.5, + 423, + 514.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 248, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 249, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 133, + 533 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 520, + 147, + 532 + ], + "score": 0.87, + "content": "I [ \\cdot ]", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 520, + 249, + 533 + ], + "score": 1.0, + "content": "is the indicator function.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 520, + 249, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 541, + 440, + 553 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 440, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 440, + 555 + ], + "score": 1.0, + "content": "Before presenting Lemma 2, we define a class of sequence that occurs in the proof.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 540, + 440, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 556, + 504, + 581 + ], + "lines": [ + { + "bbox": [ + 104, + 554, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 104, + 554, + 212, + 571 + ], + "score": 1.0, + "content": "Definition 3. A sequence", + "type": "text" + }, + { + "bbox": [ + 212, + 557, + 245, + 569 + ], + "score": 0.91, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 554, + 305, + 571 + ], + "score": 1.0, + "content": "is said to be a", + "type": "text" + }, + { + "bbox": [ + 305, + 557, + 332, + 569 + ], + "score": 0.9, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 554, + 388, + 571 + ], + "score": 1.0, + "content": "-sequence for", + "type": "text" + }, + { + "bbox": [ + 389, + 557, + 427, + 568 + ], + "score": 0.82, + "content": "C , w > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 554, + 431, + 571 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 432, + 557, + 489, + 569 + ], + "score": 0.79, + "content": "i f 0 \\le w _ { t } \\le w", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 554, + 506, + 571 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 567, + 227, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 119, + 582 + ], + "score": 1.0, + "content": "all", + "type": "text" + }, + { + "bbox": [ + 120, + 568, + 142, + 579 + ], + "score": 0.88, + "content": "t \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 567, + 164, + 582 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 164, + 568, + 222, + 582 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\sum _ { t \\geq 1 } w _ { t } \\leq C } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 567, + 227, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 554, + 506, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 584, + 485, + 598 + ], + "lines": [ + { + "bbox": [ + 104, + 583, + 486, + 599 + ], + "spans": [ + { + "bbox": [ + 104, + 583, + 195, + 599 + ], + "score": 1.0, + "content": "Lemma 2. For every", + "type": "text" + }, + { + "bbox": [ + 195, + 585, + 222, + 597 + ], + "score": 0.91, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 583, + 264, + 599 + ], + "score": 1.0, + "content": "-sequence", + "type": "text" + }, + { + "bbox": [ + 264, + 585, + 297, + 597 + ], + "score": 0.92, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 583, + 366, + 599 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + }, + { + "bbox": [ + 367, + 585, + 400, + 597 + ], + "score": 0.9, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 583, + 486, + 599 + ], + "score": 1.0, + "content": ", the following holds:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 583, + 486, + 599 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 127, + 617, + 483, + 651 + ], + "lines": [ + { + "bbox": [ + 127, + 617, + 483, + 651 + ], + "spans": [ + { + "bbox": [ + 127, + 617, + 483, + 651 + ], + "score": 0.92, + "content": "\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "576704e0ba7bc113c9eb1c2218a2bd6ace18a8d09188a27f397409e00cdaad49.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 127, + 617, + 483, + 628.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 127, + 628.3333333333334, + 483, + 639.6666666666667 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 127, + 639.6666666666667, + 483, + 651.0000000000001 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 664, + 289, + 681 + ], + "lines": [ + { + "bbox": [ + 133, + 665, + 228, + 681 + ], + "spans": [ + { + "bbox": [ + 133, + 665, + 228, + 681 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 133, + 665, + 228, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 689, + 468, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 468, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 468, + 703 + ], + "score": 1.0, + "content": "Proof of Lemma 2 is quite technical, and is therefore deferred to supplementary materials.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 688, + 468, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 107 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Now, we briefly explain how to prove Lemma 1 with Lemma 2. (Full proof can be found in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 437, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 275, + 109 + ], + "score": 1.0, + "content": "supplementary materials.) Note that since", + "type": "text" + }, + { + "bbox": [ + 276, + 93, + 313, + 107 + ], + "score": 0.93, + "content": "\\hat { Q } _ { t } \\geq Q ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 94, + 331, + 109 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 332, + 93, + 433, + 107 + ], + "score": 0.89, + "content": "a _ { t } = \\arg \\operatorname* { m a x } _ { a } \\hat { Q } _ { t } ( s _ { t } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 94, + 437, + 109 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 109, + 400, + 124 + ], + "lines": [ + { + "bbox": [ + 207, + 109, + 400, + 124 + ], + "spans": [ + { + "bbox": [ + 207, + 109, + 400, + 124 + ], + "score": 0.83, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\leq \\hat { Q } _ { t } ( s _ { t } , a _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) .", + "type": "interline_equation", + "image_path": "e36c4bae9f64d8b170464eae13de7df006383d23275ffcf8d31e0711f7c4678a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 207, + 109, + 400, + 124 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 125, + 506, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 124, + 507, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 201, + 139 + ], + "score": 1.0, + "content": "We now consider a set", + "type": "text" + }, + { + "bbox": [ + 202, + 125, + 397, + 138 + ], + "score": 0.85, + "content": "J = \\{ t : \\ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta ( 1 - \\gamma ) ^ { - 1 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 124, + 472, + 139 + ], + "score": 1.0, + "content": ", and consider the", + "type": "text" + }, + { + "bbox": [ + 472, + 126, + 502, + 138 + ], + "score": 0.91, + "content": "( | J | , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 124, + 507, + 139 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 228, + 149 + ], + "score": 1.0, + "content": "weight sequence defined by", + "type": "text" + }, + { + "bbox": [ + 228, + 137, + 294, + 149 + ], + "score": 0.91, + "content": "w _ { t } ~ = ~ I \\left[ t \\in J \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 135, + 506, + 149 + ], + "score": 1.0, + "content": ". We can now apply Lemma 2 to weighted sum", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 146, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 236, + 162 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\sum _ { t \\geq 1 } \\dot { w } _ { t } \\left[ \\bar { V } ^ { \\ast } ( s _ { t } ) - Q ^ { \\ast } ( s _ { t } , \\dot { a } _ { t } ) \\right] , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 146, + 443, + 162 + ], + "score": 1.0, + "content": ". On the one hand, this quantity is obviously at least", + "type": "text" + }, + { + "bbox": [ + 443, + 147, + 502, + 160 + ], + "score": 0.92, + "content": "| J | \\bar { \\eta ( 1 - \\gamma ) } ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 146, + 506, + 162 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 161, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 408, + 175 + ], + "score": 1.0, + "content": "On the other hand, by lemma 2, it is upper bounded by the weighted sum of", + "type": "text" + }, + { + "bbox": [ + 408, + 161, + 478, + 174 + ], + "score": 0.92, + "content": "( \\hat { Q } - Q ^ { * } ) ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 162, + 505, + 175 + ], + "score": 1.0, + "content": ". Thus", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 174, + 136, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 174, + 136, + 186 + ], + "score": 1.0, + "content": "we get", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 146, + 186, + 464, + 219 + ], + "lines": [ + { + "bbox": [ + 146, + 186, + 464, + 219 + ], + "spans": [ + { + "bbox": [ + 146, + 186, + 464, + 219 + ], + "score": 0.93, + "content": "| J | \\eta ( 1 - \\gamma ) ^ { - 1 } \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { S A | J | \\ell ( | J | ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln | J | } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "dc3a723cc9eab16cf16971abf7bbf69938ee49002e529fdd61110b29bb930e71.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 146, + 186, + 464, + 197.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 146, + 197.0, + 464, + 208.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 146, + 208.0, + 464, + 219.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 221, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 241, + 234 + ], + "score": 1.0, + "content": "Now focus on the dependence on", + "type": "text" + }, + { + "bbox": [ + 242, + 221, + 254, + 233 + ], + "score": 0.89, + "content": "| J |", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 221, + 438, + 234 + ], + "score": 1.0, + "content": ". The left-hand-side has linear dependence on", + "type": "text" + }, + { + "bbox": [ + 438, + 221, + 451, + 234 + ], + "score": 0.88, + "content": "| J |", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 221, + 505, + 234 + ], + "score": 1.0, + "content": ", whereas the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 232, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 104, + 233, + 187, + 252 + ], + "score": 1.0, + "content": "left-hand-side has a", + "type": "text" + }, + { + "bbox": [ + 187, + 232, + 231, + 252 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal { O } } \\left( \\sqrt { | J | } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 233, + 470, + 252 + ], + "score": 1.0, + "content": "dependence. This allows us to solve out an upper bound on", + "type": "text" + }, + { + "bbox": [ + 471, + 236, + 483, + 248 + ], + "score": 0.91, + "content": "| J |", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 233, + 506, + 252 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 251, + 228, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 208, + 265 + ], + "score": 1.0, + "content": "quadratic dependence on", + "type": "text" + }, + { + "bbox": [ + 208, + 252, + 225, + 263 + ], + "score": 0.87, + "content": "1 / \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 251, + 228, + 265 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 107, + 275, + 234, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 235, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 235, + 287 + ], + "score": 1.0, + "content": "3.5 PROOF FOR THEOREM 1", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 354, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 355, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 355, + 309 + ], + "score": 1.0, + "content": "We prove the theorem by stitching Lemma 1 and Condition 2.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 319, + 228, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 319, + 227, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 227, + 333 + ], + "score": 1.0, + "content": "Proof. (Proof for Theorem 1)", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 449, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 450, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 191, + 351 + ], + "score": 1.0, + "content": "By lemma 1, for any", + "type": "text" + }, + { + "bbox": [ + 191, + 336, + 419, + 349 + ], + "score": 0.85, + "content": "2 \\leq j \\leq M , \\sum _ { t = 1 } ^ { \\infty } I \\left[ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta _ { j - 1 } \\right] \\leq C", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 333, + 450, + 351 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 250, + 351, + 360, + 379 + ], + "spans": [ + { + "bbox": [ + 250, + 351, + 360, + 379 + ], + "score": 0.89, + "content": "C = \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta _ { j - 1 } ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\tilde { P } .", + "type": "inline_equation", + "image_path": "ffd9a755f0215c6d325747e98ccee281d70f6c17fc6631e7e6fc4606d3a5e751.jpg" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 291, + 401 + ], + "lines": [ + { + "bbox": [ + 101, + 378, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 101, + 378, + 128, + 404 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 128, + 383, + 138, + 395 + ], + "score": 0.86, + "content": "\\tilde { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 378, + 243, + 404 + ], + "score": 1.0, + "content": "is a shorthand for polylog", + "type": "text" + }, + { + "bbox": [ + 243, + 381, + 288, + 401 + ], + "score": 0.69, + "content": "\\begin{array} { r } { \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 105, + 406, + 506, + 473 + ], + "lines": [ + { + "bbox": [ + 103, + 403, + 508, + 426 + ], + "spans": [ + { + "bbox": [ + 103, + 403, + 122, + 426 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 406, + 214, + 422 + ], + "score": 0.93, + "content": "A _ { t } = I [ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 403, + 363, + 426 + ], + "score": 1.0, + "content": "be a Bernoulli random variable, and", + "type": "text" + }, + { + "bbox": [ + 363, + 410, + 398, + 422 + ], + "score": 0.92, + "content": "\\{ \\mathcal { F } _ { t } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 403, + 508, + 426 + ], + "score": 1.0, + "content": "be the filtration generated", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 419, + 507, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 192, + 434 + ], + "score": 1.0, + "content": "by random variables", + "type": "text" + }, + { + "bbox": [ + 193, + 421, + 287, + 433 + ], + "score": 0.88, + "content": "\\bar { \\{ ( s _ { \\tau } , a _ { \\tau } ) : 1 \\leq \\tau \\leq t \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 419, + 317, + 434 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 318, + 421, + 330, + 432 + ], + "score": 0.88, + "content": "A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 419, + 340, + 434 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 341, + 421, + 365, + 433 + ], + "score": 0.8, + "content": "\\mathcal { F } _ { t + R }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 419, + 455, + 434 + ], + "score": 1.0, + "content": "−measurable, for any", + "type": "text" + }, + { + "bbox": [ + 455, + 421, + 502, + 432 + ], + "score": 0.9, + "content": "0 \\leq k < R", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 419, + 507, + 434 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 430, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 107, + 432, + 248, + 444 + ], + "score": 0.88, + "content": "\\{ \\dot { A } _ { k + t R } - E [ A _ { k + t R } \\ | \\ \\mathcal { F } _ { k + t R } ] \\} _ { t \\geq 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 430, + 506, + 446 + ], + "score": 1.0, + "content": "is a martingale difference sequence. For now, consider a fixed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 442, + 507, + 464 + ], + "spans": [ + { + "bbox": [ + 107, + 447, + 153, + 458 + ], + "score": 0.9, + "content": "0 \\leq k < R", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 443, + 318, + 462 + ], + "score": 1.0, + "content": ". By Azuma-Hoeffiding inequality, after", + "type": "text" + }, + { + "bbox": [ + 318, + 443, + 439, + 464 + ], + "score": 0.93, + "content": "\\begin{array} { r } { T = \\mathcal { O } \\left( \\frac { C } { 2 ^ { i } } \\cdot \\frac { M \\eta _ { j } } { \\xi _ { i } } \\ln ( R M L ) \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 442, + 507, + 462 + ], + "score": 1.0, + "content": "· M ηjξ ln(RM L)\u0011 time steps (if it", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 460, + 231, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 231, + 473 + ], + "score": 1.0, + "content": "happens that many times) with", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 474, + 393, + 501 + ], + "lines": [ + { + "bbox": [ + 217, + 474, + 393, + 501 + ], + "spans": [ + { + "bbox": [ + 217, + 474, + 393, + 501 + ], + "score": 0.93, + "content": "\\operatorname* { P r } \\left[ X _ { k + t R } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] = \\mathbb { E } [ A _ { k + t R } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } ,", + "type": "interline_equation", + "image_path": "b83b246b4683938705ff27511cfa875c502578f40906ac3c2c49408d5a5881f3.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 217, + 474, + 393, + 501 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 387, + 516 + ], + "lines": [ + { + "bbox": [ + 104, + 501, + 388, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 141, + 518 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 141, + 502, + 219, + 516 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { t } A _ { k + t R } \\geq C / 2 ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 501, + 318, + 518 + ], + "score": 1.0, + "content": "with probability at least", + "type": "text" + }, + { + "bbox": [ + 318, + 503, + 384, + 515 + ], + "score": 0.89, + "content": "1 - \\delta / ( 2 M R L )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 501, + 388, + 518 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 519, + 506, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 198, + 534 + ], + "score": 1.0, + "content": "On the other hand, if", + "type": "text" + }, + { + "bbox": [ + 198, + 521, + 226, + 532 + ], + "score": 0.9, + "content": "A _ { k + t R }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 519, + 298, + 534 + ], + "score": 1.0, + "content": "happens, within", + "type": "text" + }, + { + "bbox": [ + 298, + 520, + 408, + 532 + ], + "score": 0.92, + "content": "[ k + t R , k + t R + R - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 519, + 506, + 534 + ], + "score": 1.0, + "content": ", there must be at least", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 529, + 508, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 116, + 542 + ], + "score": 0.83, + "content": "2 ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 529, + 202, + 545 + ], + "score": 1.0, + "content": "time steps at which", + "type": "text" + }, + { + "bbox": [ + 202, + 532, + 322, + 543 + ], + "score": 0.91, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > ^ { \\mathsf { ^ { * } } } \\eta _ { j - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 529, + 468, + 545 + ], + "score": 1.0, + "content": ". The latter event happens at most", + "type": "text" + }, + { + "bbox": [ + 468, + 532, + 477, + 541 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 529, + 508, + 545 + ], + "score": 1.0, + "content": "times,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 102, + 537, + 510, + 567 + ], + "spans": [ + { + "bbox": [ + 102, + 537, + 124, + 567 + ], + "score": 1.0, + "content": "and the", + "type": "text" + }, + { + "bbox": [ + 125, + 543, + 233, + 555 + ], + "score": 0.9, + "content": "[ k + t R , k + t R + R - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 537, + 293, + 567 + ], + "score": 1.0, + "content": "are disjoint. Tppens at most", + "type": "text" + }, + { + "bbox": [ + 302, + 537, + 334, + 567 + ], + "score": 1.0, + "content": "erefore, times f", + "type": "text" + }, + { + "bbox": [ + 334, + 543, + 424, + 556 + ], + "score": 0.92, + "content": "\\textstyle \\sum _ { t = 0 } ^ { \\infty } A _ { k + t R } \\leq C / 2 ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 537, + 510, + 567 + ], + "score": 1.0, + "content": ". This suggests that a union bound on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 293, + 555, + 401, + 566 + ], + "spans": [ + { + "bbox": [ + 293, + 555, + 302, + 564 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 556, + 375, + 564 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 556, + 401, + 566 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 107, + 565, + 153, + 576 + ], + "score": 0.9, + "content": "0 \\leq k < R", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 564, + 291, + 578 + ], + "score": 1.0, + "content": ", we can show that with probability", + "type": "text" + }, + { + "bbox": [ + 291, + 565, + 348, + 577 + ], + "score": 0.92, + "content": "1 - \\delta / ( 2 M L )", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 564, + 420, + 578 + ], + "score": 1.0, + "content": ", there are at most", + "type": "text" + }, + { + "bbox": [ + 420, + 565, + 436, + 576 + ], + "score": 0.78, + "content": "R T", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 564, + 506, + 578 + ], + "score": 1.0, + "content": "time steps where", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 575, + 461, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 239, + 596 + ], + "score": 0.85, + "content": "\\mathrm { P r } \\left[ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] > \\xi _ { i } / ( M \\eta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 575, + 461, + 597 + ], + "score": 1.0, + "content": ". Thus, the number of sub-optimal steps is bounded by,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 604, + 512, + 734 + ], + "lines": [ + { + "bbox": [ + 113, + 604, + 512, + 734 + ], + "spans": [ + { + "bbox": [ + 113, + 604, + 512, + 734 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { \\sum _ { t = 1 } ^ { \\ell - 1 } I [ V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) > \\epsilon ] } } \\\\ & { \\leq \\sum _ { t = 1 } ^ { \\infty } \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } I [ \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } ] = \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\sum _ { t = 1 } ^ { \\infty } I [ \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { \\eta _ { j } M } ] } \\\\ & { \\leq \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\frac { S A M R \\ln 1 / \\delta \\ln S A } { \\eta _ { j } \\xi _ { i } \\cdot 2 ^ { i } ( 1 - \\gamma ) ^ { 5 } } \\tilde { P } \\leq \\sum _ { i = 0 } ^ { L } \\frac { S A \\cdot 2 ^ { i + 4 } \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\xi _ { i } \\mathrm { ~ a n d } \\eta _ { j } ) } \\\\ & { \\leq \\frac { S A R \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\tilde { P } . } & { \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\hbar ) } \\end{array}", + "type": "interline_equation", + "image_path": "19060e0af63eb5746ff2fa8e81a511f198a258681b5fb178d09da3690eab476b.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 113, + 604, + 512, + 647.3333333333334 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 113, + 647.3333333333334, + 512, + 690.6666666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 113, + 690.6666666666667, + 512, + 734.0000000000001 + ], + "spans": [], + "index": 35 + } + ] + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 750, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 107 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Now, we briefly explain how to prove Lemma 1 with Lemma 2. (Full proof can be found in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 437, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 275, + 109 + ], + "score": 1.0, + "content": "supplementary materials.) Note that since", + "type": "text" + }, + { + "bbox": [ + 276, + 93, + 313, + 107 + ], + "score": 0.93, + "content": "\\hat { Q } _ { t } \\geq Q ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 94, + 331, + 109 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 332, + 93, + 433, + 107 + ], + "score": 0.89, + "content": "a _ { t } = \\arg \\operatorname* { m a x } _ { a } \\hat { Q } _ { t } ( s _ { t } , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 94, + 437, + 109 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 109 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 109, + 400, + 124 + ], + "lines": [ + { + "bbox": [ + 207, + 109, + 400, + 124 + ], + "spans": [ + { + "bbox": [ + 207, + 109, + 400, + 124 + ], + "score": 0.83, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\leq \\hat { Q } _ { t } ( s _ { t } , a _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) .", + "type": "interline_equation", + "image_path": "e36c4bae9f64d8b170464eae13de7df006383d23275ffcf8d31e0711f7c4678a.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 207, + 109, + 400, + 124 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 125, + 506, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 124, + 507, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 201, + 139 + ], + "score": 1.0, + "content": "We now consider a set", + "type": "text" + }, + { + "bbox": [ + 202, + 125, + 397, + 138 + ], + "score": 0.85, + "content": "J = \\{ t : \\ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta ( 1 - \\gamma ) ^ { - 1 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 124, + 472, + 139 + ], + "score": 1.0, + "content": ", and consider the", + "type": "text" + }, + { + "bbox": [ + 472, + 126, + 502, + 138 + ], + "score": 0.91, + "content": "( | J | , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 124, + 507, + 139 + ], + "score": 1.0, + "content": "-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 228, + 149 + ], + "score": 1.0, + "content": "weight sequence defined by", + "type": "text" + }, + { + "bbox": [ + 228, + 137, + 294, + 149 + ], + "score": 0.91, + "content": "w _ { t } ~ = ~ I \\left[ t \\in J \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 135, + 506, + 149 + ], + "score": 1.0, + "content": ". We can now apply Lemma 2 to weighted sum", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 146, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 236, + 162 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\sum _ { t \\geq 1 } \\dot { w } _ { t } \\left[ \\bar { V } ^ { \\ast } ( s _ { t } ) - Q ^ { \\ast } ( s _ { t } , \\dot { a } _ { t } ) \\right] , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 146, + 443, + 162 + ], + "score": 1.0, + "content": ". On the one hand, this quantity is obviously at least", + "type": "text" + }, + { + "bbox": [ + 443, + 147, + 502, + 160 + ], + "score": 0.92, + "content": "| J | \\bar { \\eta ( 1 - \\gamma ) } ^ { - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 146, + 506, + 162 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 161, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 408, + 175 + ], + "score": 1.0, + "content": "On the other hand, by lemma 2, it is upper bounded by the weighted sum of", + "type": "text" + }, + { + "bbox": [ + 408, + 161, + 478, + 174 + ], + "score": 0.92, + "content": "( \\hat { Q } - Q ^ { * } ) ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 162, + 505, + 175 + ], + "score": 1.0, + "content": ". Thus", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 174, + 136, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 174, + 136, + 186 + ], + "score": 1.0, + "content": "we get", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 124, + 507, + 186 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 146, + 186, + 464, + 219 + ], + "lines": [ + { + "bbox": [ + 146, + 186, + 464, + 219 + ], + "spans": [ + { + "bbox": [ + 146, + 186, + 464, + 219 + ], + "score": 0.93, + "content": "| J | \\eta ( 1 - \\gamma ) ^ { - 1 } \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { S A | J | \\ell ( | J | ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln | J | } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "dc3a723cc9eab16cf16971abf7bbf69938ee49002e529fdd61110b29bb930e71.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 146, + 186, + 464, + 197.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 146, + 197.0, + 464, + 208.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 146, + 208.0, + 464, + 219.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 221, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 241, + 234 + ], + "score": 1.0, + "content": "Now focus on the dependence on", + "type": "text" + }, + { + "bbox": [ + 242, + 221, + 254, + 233 + ], + "score": 0.89, + "content": "| J |", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 221, + 438, + 234 + ], + "score": 1.0, + "content": ". The left-hand-side has linear dependence on", + "type": "text" + }, + { + "bbox": [ + 438, + 221, + 451, + 234 + ], + "score": 0.88, + "content": "| J |", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 221, + 505, + 234 + ], + "score": 1.0, + "content": ", whereas the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 232, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 104, + 233, + 187, + 252 + ], + "score": 1.0, + "content": "left-hand-side has a", + "type": "text" + }, + { + "bbox": [ + 187, + 232, + 231, + 252 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal { O } } \\left( \\sqrt { | J | } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 233, + 470, + 252 + ], + "score": 1.0, + "content": "dependence. This allows us to solve out an upper bound on", + "type": "text" + }, + { + "bbox": [ + 471, + 236, + 483, + 248 + ], + "score": 0.91, + "content": "| J |", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 233, + 506, + 252 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 251, + 228, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 208, + 265 + ], + "score": 1.0, + "content": "quadratic dependence on", + "type": "text" + }, + { + "bbox": [ + 208, + 252, + 225, + 263 + ], + "score": 0.87, + "content": "1 / \\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 251, + 228, + 265 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 104, + 221, + 506, + 265 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 275, + 234, + 286 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 235, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 235, + 287 + ], + "score": 1.0, + "content": "3.5 PROOF FOR THEOREM 1", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 354, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 355, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 355, + 309 + ], + "score": 1.0, + "content": "We prove the theorem by stitching Lemma 1 and Condition 2.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 295, + 355, + 309 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 319, + 228, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 319, + 227, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 227, + 333 + ], + "score": 1.0, + "content": "Proof. (Proof for Theorem 1)", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 319, + 227, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 335, + 449, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 450, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 191, + 351 + ], + "score": 1.0, + "content": "By lemma 1, for any", + "type": "text" + }, + { + "bbox": [ + 191, + 336, + 419, + 349 + ], + "score": 0.85, + "content": "2 \\leq j \\leq M , \\sum _ { t = 1 } ^ { \\infty } I \\left[ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta _ { j - 1 } \\right] \\leq C", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 333, + 450, + 351 + ], + "score": 1.0, + "content": ", where", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 250, + 351, + 360, + 379 + ], + "spans": [ + { + "bbox": [ + 250, + 351, + 360, + 379 + ], + "score": 0.89, + "content": "C = \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta _ { j - 1 } ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\tilde { P } .", + "type": "inline_equation", + "image_path": "ffd9a755f0215c6d325747e98ccee281d70f6c17fc6631e7e6fc4606d3a5e751.jpg" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 333, + 450, + 379 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 381, + 291, + 401 + ], + "lines": [ + { + "bbox": [ + 101, + 378, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 101, + 378, + 128, + 404 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 128, + 383, + 138, + 395 + ], + "score": 0.86, + "content": "\\tilde { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 378, + 243, + 404 + ], + "score": 1.0, + "content": "is a shorthand for polylog", + "type": "text" + }, + { + "bbox": [ + 243, + 381, + 288, + 401 + ], + "score": 0.69, + "content": "\\begin{array} { r } { \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 101, + 378, + 288, + 404 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 406, + 506, + 473 + ], + "lines": [ + { + "bbox": [ + 103, + 403, + 508, + 426 + ], + "spans": [ + { + "bbox": [ + 103, + 403, + 122, + 426 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 406, + 214, + 422 + ], + "score": 0.93, + "content": "A _ { t } = I [ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 403, + 363, + 426 + ], + "score": 1.0, + "content": "be a Bernoulli random variable, and", + "type": "text" + }, + { + "bbox": [ + 363, + 410, + 398, + 422 + ], + "score": 0.92, + "content": "\\{ \\mathcal { F } _ { t } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 403, + 508, + 426 + ], + "score": 1.0, + "content": "be the filtration generated", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 419, + 507, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 192, + 434 + ], + "score": 1.0, + "content": "by random variables", + "type": "text" + }, + { + "bbox": [ + 193, + 421, + 287, + 433 + ], + "score": 0.88, + "content": "\\bar { \\{ ( s _ { \\tau } , a _ { \\tau } ) : 1 \\leq \\tau \\leq t \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 419, + 317, + 434 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 318, + 421, + 330, + 432 + ], + "score": 0.88, + "content": "A _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 419, + 340, + 434 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 341, + 421, + 365, + 433 + ], + "score": 0.8, + "content": "\\mathcal { F } _ { t + R }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 419, + 455, + 434 + ], + "score": 1.0, + "content": "−measurable, for any", + "type": "text" + }, + { + "bbox": [ + 455, + 421, + 502, + 432 + ], + "score": 0.9, + "content": "0 \\leq k < R", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 419, + 507, + 434 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 430, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 107, + 432, + 248, + 444 + ], + "score": 0.88, + "content": "\\{ \\dot { A } _ { k + t R } - E [ A _ { k + t R } \\ | \\ \\mathcal { F } _ { k + t R } ] \\} _ { t \\geq 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 430, + 506, + 446 + ], + "score": 1.0, + "content": "is a martingale difference sequence. For now, consider a fixed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 442, + 507, + 464 + ], + "spans": [ + { + "bbox": [ + 107, + 447, + 153, + 458 + ], + "score": 0.9, + "content": "0 \\leq k < R", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 443, + 318, + 462 + ], + "score": 1.0, + "content": ". By Azuma-Hoeffiding inequality, after", + "type": "text" + }, + { + "bbox": [ + 318, + 443, + 439, + 464 + ], + "score": 0.93, + "content": "\\begin{array} { r } { T = \\mathcal { O } \\left( \\frac { C } { 2 ^ { i } } \\cdot \\frac { M \\eta _ { j } } { \\xi _ { i } } \\ln ( R M L ) \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 442, + 507, + 462 + ], + "score": 1.0, + "content": "· M ηjξ ln(RM L)\u0011 time steps (if it", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 460, + 231, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 231, + 473 + ], + "score": 1.0, + "content": "happens that many times) with", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 103, + 403, + 508, + 473 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 474, + 393, + 501 + ], + "lines": [ + { + "bbox": [ + 217, + 474, + 393, + 501 + ], + "spans": [ + { + "bbox": [ + 217, + 474, + 393, + 501 + ], + "score": 0.93, + "content": "\\operatorname* { P r } \\left[ X _ { k + t R } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] = \\mathbb { E } [ A _ { k + t R } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } ,", + "type": "interline_equation", + "image_path": "b83b246b4683938705ff27511cfa875c502578f40906ac3c2c49408d5a5881f3.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 217, + 474, + 393, + 501 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 387, + 516 + ], + "lines": [ + { + "bbox": [ + 104, + 501, + 388, + 518 + ], + "spans": [ + { + "bbox": [ + 104, + 501, + 141, + 518 + ], + "score": 1.0, + "content": "we have", + "type": "text" + }, + { + "bbox": [ + 141, + 502, + 219, + 516 + ], + "score": 0.93, + "content": "\\textstyle \\sum _ { t } A _ { k + t R } \\geq C / 2 ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 501, + 318, + 518 + ], + "score": 1.0, + "content": "with probability at least", + "type": "text" + }, + { + "bbox": [ + 318, + 503, + 384, + 515 + ], + "score": 0.89, + "content": "1 - \\delta / ( 2 M R L )", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 501, + 388, + 518 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 501, + 388, + 518 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 519, + 506, + 596 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 198, + 534 + ], + "score": 1.0, + "content": "On the other hand, if", + "type": "text" + }, + { + "bbox": [ + 198, + 521, + 226, + 532 + ], + "score": 0.9, + "content": "A _ { k + t R }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 519, + 298, + 534 + ], + "score": 1.0, + "content": "happens, within", + "type": "text" + }, + { + "bbox": [ + 298, + 520, + 408, + 532 + ], + "score": 0.92, + "content": "[ k + t R , k + t R + R - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 519, + 506, + 534 + ], + "score": 1.0, + "content": ", there must be at least", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 529, + 508, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 116, + 542 + ], + "score": 0.83, + "content": "2 ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 529, + 202, + 545 + ], + "score": 1.0, + "content": "time steps at which", + "type": "text" + }, + { + "bbox": [ + 202, + 532, + 322, + 543 + ], + "score": 0.91, + "content": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > ^ { \\mathsf { ^ { * } } } \\eta _ { j - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 529, + 468, + 545 + ], + "score": 1.0, + "content": ". The latter event happens at most", + "type": "text" + }, + { + "bbox": [ + 468, + 532, + 477, + 541 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 529, + 508, + 545 + ], + "score": 1.0, + "content": "times,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 102, + 537, + 510, + 567 + ], + "spans": [ + { + "bbox": [ + 102, + 537, + 124, + 567 + ], + "score": 1.0, + "content": "and the", + "type": "text" + }, + { + "bbox": [ + 125, + 543, + 233, + 555 + ], + "score": 0.9, + "content": "[ k + t R , k + t R + R - 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 537, + 293, + 567 + ], + "score": 1.0, + "content": "are disjoint. Tppens at most", + "type": "text" + }, + { + "bbox": [ + 302, + 537, + 334, + 567 + ], + "score": 1.0, + "content": "erefore, times f", + "type": "text" + }, + { + "bbox": [ + 334, + 543, + 424, + 556 + ], + "score": 0.92, + "content": "\\textstyle \\sum _ { t = 0 } ^ { \\infty } A _ { k + t R } \\leq C / 2 ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 537, + 510, + 567 + ], + "score": 1.0, + "content": ". This suggests that a union bound on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 293, + 555, + 401, + 566 + ], + "spans": [ + { + "bbox": [ + 293, + 555, + 302, + 564 + ], + "score": 0.83, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 556, + 375, + 564 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 556, + 401, + 566 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 107, + 565, + 153, + 576 + ], + "score": 0.9, + "content": "0 \\leq k < R", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 564, + 291, + 578 + ], + "score": 1.0, + "content": ", we can show that with probability", + "type": "text" + }, + { + "bbox": [ + 291, + 565, + 348, + 577 + ], + "score": 0.92, + "content": "1 - \\delta / ( 2 M L )", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 564, + 420, + 578 + ], + "score": 1.0, + "content": ", there are at most", + "type": "text" + }, + { + "bbox": [ + 420, + 565, + 436, + 576 + ], + "score": 0.78, + "content": "R T", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 564, + 506, + 578 + ], + "score": 1.0, + "content": "time steps where", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 575, + 461, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 239, + 596 + ], + "score": 0.85, + "content": "\\mathrm { P r } \\left[ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] > \\xi _ { i } / ( M \\eta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 575, + 461, + 597 + ], + "score": 1.0, + "content": ". Thus, the number of sub-optimal steps is bounded by,", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 102, + 519, + 510, + 597 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 604, + 512, + 734 + ], + "lines": [ + { + "bbox": [ + 113, + 604, + 512, + 734 + ], + "spans": [ + { + "bbox": [ + 113, + 604, + 512, + 734 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { \\sum _ { t = 1 } ^ { \\ell - 1 } I [ V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) > \\epsilon ] } } \\\\ & { \\leq \\sum _ { t = 1 } ^ { \\infty } \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } I [ \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } ] = \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\sum _ { t = 1 } ^ { \\infty } I [ \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { \\eta _ { j } M } ] } \\\\ & { \\leq \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\frac { S A M R \\ln 1 / \\delta \\ln S A } { \\eta _ { j } \\xi _ { i } \\cdot 2 ^ { i } ( 1 - \\gamma ) ^ { 5 } } \\tilde { P } \\leq \\sum _ { i = 0 } ^ { L } \\frac { S A \\cdot 2 ^ { i + 4 } \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\xi _ { i } \\mathrm { ~ a n d } \\eta _ { j } ) } \\\\ & { \\leq \\frac { S A R \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\tilde { P } . } & { \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\hbar ) } \\end{array}", + "type": "interline_equation", + "image_path": "19060e0af63eb5746ff2fa8e81a511f198a258681b5fb178d09da3690eab476b.jpg" + } + ] + } + ], + "index": 34, + "virtual_lines": [ + { + "bbox": [ + 113, + 604, + 512, + 647.3333333333334 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 113, + 647.3333333333334, + 512, + 690.6666666666667 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 113, + 690.6666666666667, + 512, + 734.0000000000001 + ], + "spans": [], + "index": 35 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 136 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 289, + 96 + ], + "score": 1.0, + "content": "It should be stressed that throughout the lines,", + "type": "text" + }, + { + "bbox": [ + 290, + 81, + 299, + 92 + ], + "score": 0.87, + "content": "\\tilde { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "is a shorthand for an asymptotic expression, instead", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 91, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 104, + 91, + 434, + 109 + ], + "score": 1.0, + "content": "of an exact value. Our final choice of \u000f2 and \u000f1 are \u000f2 = \u000f3 , and \u000f1 = \u000f24RM ln", + "type": "text" + }, + { + "bbox": [ + 446, + 93, + 505, + 105 + ], + "score": 1.0, + "content": ". It is not hard", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 428, + 102, + 442, + 110 + ], + "spans": [ + { + "bbox": [ + 428, + 102, + 442, + 110 + ], + "score": 1.0, + "content": "1−γ", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 107, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 104, + 107, + 150, + 127 + ], + "score": 1.0, + "content": "to see that", + "type": "text" + }, + { + "bbox": [ + 150, + 109, + 266, + 124 + ], + "score": 0.92, + "content": "\\ln 1 / \\epsilon _ { 1 } = \\mathrm { p o l y } ( \\ln \\textstyle \\frac { 1 } { \\epsilon } , \\ln \\frac { 1 } { 1 - \\gamma } )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 107, + 463, + 127 + ], + "score": 1.0, + "content": ". This immediately implies that with probability", + "type": "text" + }, + { + "bbox": [ + 463, + 110, + 486, + 121 + ], + "score": 0.79, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 107, + 506, + 127 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 324, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 231, + 137 + ], + "score": 1.0, + "content": "number of time steps such that", + "type": "text" + }, + { + "bbox": [ + 231, + 123, + 313, + 136 + ], + "score": 0.9, + "content": "\\left( V ^ { * } - \\dot { V } ^ { \\pi } \\right) \\left( s _ { t } \\right) > \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 122, + 324, + 137 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 140, + 343, + 168 + ], + "lines": [ + { + "bbox": [ + 266, + 140, + 343, + 168 + ], + "spans": [ + { + "bbox": [ + 266, + 140, + 343, + 168 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln 1 / \\delta } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\right) ,", + "type": "interline_equation", + "image_path": "112b2a0ff239fa280515807bb22278ab6c902bb4dd2bada09b6a5e780e0dcd78.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 266, + 140, + 343, + 168 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 313, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 314, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 208, + 190 + ], + "score": 1.0, + "content": "where hidden factors are", + "type": "text" + }, + { + "bbox": [ + 208, + 173, + 309, + 189 + ], + "score": 0.79, + "content": "\\begin{array} { r } { \\operatorname { p o l y } ( \\ln { \\frac { 1 } { \\epsilon } } , \\ln { \\frac { 1 } { 1 - \\gamma } } , \\ln S A ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 171, + 314, + 190 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 107, + 201, + 191, + 215 + ], + "lines": [ + { + "bbox": [ + 104, + 200, + 192, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 192, + 217 + ], + "score": 1.0, + "content": "4 DISCUSSION", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "In this section, we discuss the implication of our results, and present some interesting properties of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 238, + 315, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 315, + 251 + ], + "score": 1.0, + "content": "our algorithm beyond its sample complexity bound.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "title", + "bbox": [ + 107, + 263, + 300, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 301, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 301, + 276 + ], + "score": 1.0, + "content": "4.1 COMPARISON WITH PREVIOUS RESULTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 506, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "Lower bound To the best of our knowledge, the current best lower bound for worst-case sample", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 508, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 164, + 312 + ], + "score": 1.0, + "content": "complexity is", + "type": "text" + }, + { + "bbox": [ + 173, + 293, + 508, + 315 + ], + "score": 1.0, + "content": "\u0010 SA\u000f2(1−γ)3 ln 1/δ\u0011 due to Lattimore & Hutter (2012). The gap between our results", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 309, + 326 + ], + "score": 1.0, + "content": "and this lower bound lies only in the dependence on", + "type": "text" + }, + { + "bbox": [ + 309, + 313, + 348, + 325 + ], + "score": 0.93, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 313, + 450, + 326 + ], + "score": 1.0, + "content": "and logarithmic terms of", + "type": "text" + }, + { + "bbox": [ + 451, + 312, + 505, + 325 + ], + "score": 0.88, + "content": "\\mathrm { { } } ^ { \\mathrm { { 3 } } A , 1 / ( 1 - \\gamma ) }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 323, + 144, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 123, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 325, + 138, + 336 + ], + "score": 0.87, + "content": "1 / \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 323, + 144, + 337 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 506, + 408 + ], + "lines": [ + { + "bbox": [ + 104, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Model-free algorithms Previously, the best sample complexity bound for a model-free algorithm", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 352, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 116, + 369 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 117, + 352, + 173, + 372 + ], + "score": 0.95, + "content": "\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 355, + 506, + 369 + ], + "score": 1.0, + "content": "(suppressing all logarithmic terms), achieved by Delayed Q-learning Strehl et al.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 102, + 369, + 508, + 390 + ], + "spans": [ + { + "bbox": [ + 102, + 369, + 345, + 390 + ], + "score": 1.0, + "content": "(2006). Our results improve this upper bound by a factor of", + "type": "text" + }, + { + "bbox": [ + 345, + 370, + 376, + 387 + ], + "score": 0.93, + "content": "\\frac { 1 } { \\epsilon ^ { 2 } ( 1 - \\gamma ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 369, + 508, + 390 + ], + "score": 1.0, + "content": ", and closes the quadratic gap in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 122, + 397 + ], + "score": 0.87, + "content": "1 / \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "between Delayed Q-learning’s result and the lower bound. In fact, the following theorem shows", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 396, + 369, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 369, + 410 + ], + "score": 1.0, + "content": "that UCB Q-learning can indeed outperform Delayed Q-learning.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 353, + 424 + ], + "score": 1.0, + "content": "Theorem 2. There exists a family of MDPs with constant", + "type": "text" + }, + { + "bbox": [ + 354, + 412, + 362, + 421 + ], + "score": 0.77, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 409, + 383, + 424 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 412, + 392, + 421 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 409, + 505, + 424 + ], + "score": 1.0, + "content": ", in which with probability", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 422, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 129, + 437 + ], + "score": 0.68, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 423, + 169, + 439 + ], + "score": 1.0, + "content": ", Delayed", + "type": "text" + }, + { + "bbox": [ + 170, + 426, + 178, + 437 + ], + "score": 0.66, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 423, + 393, + 439 + ], + "score": 1.0, + "content": "-learning incurs sample complexity of exploration of", + "type": "text" + }, + { + "bbox": [ + 393, + 422, + 441, + 442 + ], + "score": 0.94, + "content": "\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 425, + 505, + 438 + ], + "score": 1.0, + "content": ", assuming that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 440, + 170, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 167, + 453 + ], + "score": 0.89, + "content": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 440, + 170, + 453 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 432, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 433, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 433, + 475 + ], + "score": 1.0, + "content": "The construction of this hard MDP family is given in the supplementary material.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 478, + 506, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Model-based algorithms For model-based algorithms, better sample complexity results in infinite", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "horizon settings have been claimed Szita & Szepesvári (2010). To the best of our knowledge, the best", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 501, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 325, + 517 + ], + "score": 1.0, + "content": "published result without further restrictions on MDPs is", + "type": "text" + }, + { + "bbox": [ + 326, + 501, + 382, + 521 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "claimed by Szita & Szepesvári", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 176, + 533 + ], + "score": 1.0, + "content": "(2010), which is", + "type": "text" + }, + { + "bbox": [ + 176, + 520, + 207, + 532 + ], + "score": 0.91, + "content": "( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "smaller than our upper bound. From the space complexity point of view,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 389, + 543 + ], + "score": 1.0, + "content": "our algorithm is much more memory-efficient. Our algorithm stores", + "type": "text" + }, + { + "bbox": [ + 390, + 531, + 421, + 543 + ], + "score": 0.92, + "content": "O ( S A )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "values, whereas the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 541, + 479, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 290, + 554 + ], + "score": 1.0, + "content": "algorithm in Szita & Szepesvári (2010) needs", + "type": "text" + }, + { + "bbox": [ + 291, + 541, + 325, + 554 + ], + "score": 0.93, + "content": "\\Omega ( S ^ { 2 } A )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 542, + 479, + 554 + ], + "score": 1.0, + "content": "memory to store the transition model.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 271, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 272, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 272, + 579 + ], + "score": 1.0, + "content": "4.2 EXTENSION TO OTHER SETTINGS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 586, + 484, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 485, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 485, + 601 + ], + "score": 1.0, + "content": "Due to length limits, detailed discussion in this section is deferred to supplementary materials.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "score": 1.0, + "content": "Finite horizon MDP The sample complexity of exploration bounds of UCB Q-learning implies", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 613, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 107, + 615, + 141, + 629 + ], + "score": 0.91, + "content": "\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 613, + 265, + 631 + ], + "score": 1.0, + "content": "PAC sample complexity and a", + "type": "text" + }, + { + "bbox": [ + 266, + 615, + 304, + 629 + ], + "score": 0.9, + "content": "\\bar { \\mathcal { O } } \\left( T ^ { 1 / 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 613, + 506, + 631 + ], + "score": 1.0, + "content": "regret bound in finite horizon MDPs. That is, our", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "algorithm implies a PAC algorithm for finite horizon MDPs. We are not aware of reductions of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "score": 1.0, + "content": "opposite direction (from finite horizon sample complexity to infinite horizon sample complexity of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 649, + 160, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 160, + 663 + ], + "score": 1.0, + "content": "exploration).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 304, + 681 + ], + "score": 1.0, + "content": "Regret The reason why our results can imply an", + "type": "text" + }, + { + "bbox": [ + 304, + 666, + 336, + 680 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal { O } } ( { \\sqrt { T } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 666, + 451, + 681 + ], + "score": 1.0, + "content": "regret is that, after choosing", + "type": "text" + }, + { + "bbox": [ + 451, + 670, + 461, + 679 + ], + "score": 0.82, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 666, + 505, + 681 + ], + "score": 1.0, + "content": ", it follows", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 319, + 693 + ], + "score": 1.0, + "content": "from the argument of Theorem 1 that with probability", + "type": "text" + }, + { + "bbox": [ + 320, + 681, + 342, + 691 + ], + "score": 0.86, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 680, + 372, + 693 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 372, + 679, + 454, + 693 + ], + "score": 0.93, + "content": "\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 680, + 505, + 693 + ], + "score": 1.0, + "content": ", the number", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 691, + 256, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 117, + 704 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 693, + 127, + 703 + ], + "score": 0.85, + "content": "\\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 691, + 256, + 704 + ], + "score": 1.0, + "content": "-suboptimal steps is bounded by", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 708, + 399, + 736 + ], + "lines": [ + { + "bbox": [ + 211, + 708, + 399, + 736 + ], + "spans": [ + { + "bbox": [ + 211, + 708, + 399, + 736 + ], + "score": 0.94, + "content": "\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right) .", + "type": "interline_equation", + "image_path": "4312c45d17cd6e06d125f0e95014a5792379fa26cd13205bd4de52e4ac033fac.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 211, + 708, + 399, + 736 + ], + "spans": [], + "index": 40 + } + ] + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 174, + 505, + 185 + ], + "lines": [ + { + "bbox": [ + 496, + 176, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 496, + 176, + 505, + 185 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 136 + ], + "lines": [ + { + "bbox": [ + 104, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 80, + 289, + 96 + ], + "score": 1.0, + "content": "It should be stressed that throughout the lines,", + "type": "text" + }, + { + "bbox": [ + 290, + 81, + 299, + 92 + ], + "score": 0.87, + "content": "\\tilde { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "is a shorthand for an asymptotic expression, instead", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 91, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 104, + 91, + 434, + 109 + ], + "score": 1.0, + "content": "of an exact value. Our final choice of \u000f2 and \u000f1 are \u000f2 = \u000f3 , and \u000f1 = \u000f24RM ln", + "type": "text" + }, + { + "bbox": [ + 446, + 93, + 505, + 105 + ], + "score": 1.0, + "content": ". It is not hard", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 428, + 102, + 442, + 110 + ], + "spans": [ + { + "bbox": [ + 428, + 102, + 442, + 110 + ], + "score": 1.0, + "content": "1−γ", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 107, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 104, + 107, + 150, + 127 + ], + "score": 1.0, + "content": "to see that", + "type": "text" + }, + { + "bbox": [ + 150, + 109, + 266, + 124 + ], + "score": 0.92, + "content": "\\ln 1 / \\epsilon _ { 1 } = \\mathrm { p o l y } ( \\ln \\textstyle \\frac { 1 } { \\epsilon } , \\ln \\frac { 1 } { 1 - \\gamma } )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 107, + 463, + 127 + ], + "score": 1.0, + "content": ". This immediately implies that with probability", + "type": "text" + }, + { + "bbox": [ + 463, + 110, + 486, + 121 + ], + "score": 0.79, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 107, + 506, + 127 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 324, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 231, + 137 + ], + "score": 1.0, + "content": "number of time steps such that", + "type": "text" + }, + { + "bbox": [ + 231, + 123, + 313, + 136 + ], + "score": 0.9, + "content": "\\left( V ^ { * } - \\dot { V } ^ { \\pi } \\right) \\left( s _ { t } \\right) > \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 122, + 324, + 137 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 104, + 80, + 506, + 137 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 266, + 140, + 343, + 168 + ], + "lines": [ + { + "bbox": [ + 266, + 140, + 343, + 168 + ], + "spans": [ + { + "bbox": [ + 266, + 140, + 343, + 168 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln 1 / \\delta } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\right) ,", + "type": "interline_equation", + "image_path": "112b2a0ff239fa280515807bb22278ab6c902bb4dd2bada09b6a5e780e0dcd78.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 266, + 140, + 343, + 168 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 313, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 314, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 208, + 190 + ], + "score": 1.0, + "content": "where hidden factors are", + "type": "text" + }, + { + "bbox": [ + 208, + 173, + 309, + 189 + ], + "score": 0.79, + "content": "\\begin{array} { r } { \\operatorname { p o l y } ( \\ln { \\frac { 1 } { \\epsilon } } , \\ln { \\frac { 1 } { 1 - \\gamma } } , \\ln S A ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 171, + 314, + 190 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 171, + 314, + 190 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 201, + 191, + 215 + ], + "lines": [ + { + "bbox": [ + 104, + 200, + 192, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 192, + 217 + ], + "score": 1.0, + "content": "4 DISCUSSION", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 506, + 240 + ], + "score": 1.0, + "content": "In this section, we discuss the implication of our results, and present some interesting properties of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 238, + 315, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 315, + 251 + ], + "score": 1.0, + "content": "our algorithm beyond its sample complexity bound.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 226, + 506, + 251 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 263, + 300, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 263, + 301, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 301, + 276 + ], + "score": 1.0, + "content": "4.1 COMPARISON WITH PREVIOUS RESULTS", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 283, + 506, + 336 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "Lower bound To the best of our knowledge, the current best lower bound for worst-case sample", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 508, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 164, + 312 + ], + "score": 1.0, + "content": "complexity is", + "type": "text" + }, + { + "bbox": [ + 173, + 293, + 508, + 315 + ], + "score": 1.0, + "content": "\u0010 SA\u000f2(1−γ)3 ln 1/δ\u0011 due to Lattimore & Hutter (2012). The gap between our results", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 312, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 309, + 326 + ], + "score": 1.0, + "content": "and this lower bound lies only in the dependence on", + "type": "text" + }, + { + "bbox": [ + 309, + 313, + 348, + 325 + ], + "score": 0.93, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 313, + 450, + 326 + ], + "score": 1.0, + "content": "and logarithmic terms of", + "type": "text" + }, + { + "bbox": [ + 451, + 312, + 505, + 325 + ], + "score": 0.88, + "content": "\\mathrm { { } } ^ { \\mathrm { { 3 } } A , 1 / ( 1 - \\gamma ) }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 323, + 144, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 123, + 337 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 325, + 138, + 336 + ], + "score": 0.87, + "content": "1 / \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 323, + 144, + 337 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 283, + 508, + 337 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 506, + 408 + ], + "lines": [ + { + "bbox": [ + 104, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 104, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "Model-free algorithms Previously, the best sample complexity bound for a model-free algorithm", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 352, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 116, + 369 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 117, + 352, + 173, + 372 + ], + "score": 0.95, + "content": "\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 355, + 506, + 369 + ], + "score": 1.0, + "content": "(suppressing all logarithmic terms), achieved by Delayed Q-learning Strehl et al.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 102, + 369, + 508, + 390 + ], + "spans": [ + { + "bbox": [ + 102, + 369, + 345, + 390 + ], + "score": 1.0, + "content": "(2006). Our results improve this upper bound by a factor of", + "type": "text" + }, + { + "bbox": [ + 345, + 370, + 376, + 387 + ], + "score": 0.93, + "content": "\\frac { 1 } { \\epsilon ^ { 2 } ( 1 - \\gamma ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 369, + 508, + 390 + ], + "score": 1.0, + "content": ", and closes the quadratic gap in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 122, + 397 + ], + "score": 0.87, + "content": "1 / \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "between Delayed Q-learning’s result and the lower bound. In fact, the following theorem shows", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 396, + 369, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 369, + 410 + ], + "score": 1.0, + "content": "that UCB Q-learning can indeed outperform Delayed Q-learning.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 102, + 340, + 508, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 409, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 353, + 424 + ], + "score": 1.0, + "content": "Theorem 2. There exists a family of MDPs with constant", + "type": "text" + }, + { + "bbox": [ + 354, + 412, + 362, + 421 + ], + "score": 0.77, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 409, + 383, + 424 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 412, + 392, + 421 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 409, + 505, + 424 + ], + "score": 1.0, + "content": ", in which with probability", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 422, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 129, + 437 + ], + "score": 0.68, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 423, + 169, + 439 + ], + "score": 1.0, + "content": ", Delayed", + "type": "text" + }, + { + "bbox": [ + 170, + 426, + 178, + 437 + ], + "score": 0.66, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 423, + 393, + 439 + ], + "score": 1.0, + "content": "-learning incurs sample complexity of exploration of", + "type": "text" + }, + { + "bbox": [ + 393, + 422, + 441, + 442 + ], + "score": 0.94, + "content": "\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 425, + 505, + 438 + ], + "score": 1.0, + "content": ", assuming that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 440, + 170, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 167, + 453 + ], + "score": 0.89, + "content": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 440, + 170, + 453 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 409, + 505, + 453 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 461, + 432, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 433, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 433, + 475 + ], + "score": 1.0, + "content": "The construction of this hard MDP family is given in the supplementary material.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 461, + 433, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 478, + 506, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "Model-based algorithms For model-based algorithms, better sample complexity results in infinite", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 506, + 502 + ], + "score": 1.0, + "content": "horizon settings have been claimed Szita & Szepesvári (2010). To the best of our knowledge, the best", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 501, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 325, + 517 + ], + "score": 1.0, + "content": "published result without further restrictions on MDPs is", + "type": "text" + }, + { + "bbox": [ + 326, + 501, + 382, + 521 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "claimed by Szita & Szepesvári", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 176, + 533 + ], + "score": 1.0, + "content": "(2010), which is", + "type": "text" + }, + { + "bbox": [ + 176, + 520, + 207, + 532 + ], + "score": 0.91, + "content": "( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "smaller than our upper bound. From the space complexity point of view,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 389, + 543 + ], + "score": 1.0, + "content": "our algorithm is much more memory-efficient. Our algorithm stores", + "type": "text" + }, + { + "bbox": [ + 390, + 531, + 421, + 543 + ], + "score": 0.92, + "content": "O ( S A )", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "values, whereas the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 541, + 479, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 290, + 554 + ], + "score": 1.0, + "content": "algorithm in Szita & Szepesvári (2010) needs", + "type": "text" + }, + { + "bbox": [ + 291, + 541, + 325, + 554 + ], + "score": 0.93, + "content": "\\Omega ( S ^ { 2 } A )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 542, + 479, + 554 + ], + "score": 1.0, + "content": "memory to store the transition model.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 479, + 506, + 554 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 271, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 566, + 272, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 272, + 579 + ], + "score": 1.0, + "content": "4.2 EXTENSION TO OTHER SETTINGS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 586, + 484, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 485, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 485, + 601 + ], + "score": 1.0, + "content": "Due to length limits, detailed discussion in this section is deferred to supplementary materials.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 585, + 485, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 603, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 618 + ], + "score": 1.0, + "content": "Finite horizon MDP The sample complexity of exploration bounds of UCB Q-learning implies", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 613, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 107, + 615, + 141, + 629 + ], + "score": 0.91, + "content": "\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 613, + 265, + 631 + ], + "score": 1.0, + "content": "PAC sample complexity and a", + "type": "text" + }, + { + "bbox": [ + 266, + 615, + 304, + 629 + ], + "score": 0.9, + "content": "\\bar { \\mathcal { O } } \\left( T ^ { 1 / 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 613, + 506, + 631 + ], + "score": 1.0, + "content": "regret bound in finite horizon MDPs. That is, our", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "algorithm implies a PAC algorithm for finite horizon MDPs. We are not aware of reductions of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 653 + ], + "score": 1.0, + "content": "opposite direction (from finite horizon sample complexity to infinite horizon sample complexity of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 649, + 160, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 160, + 663 + ], + "score": 1.0, + "content": "exploration).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 104, + 603, + 506, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 304, + 681 + ], + "score": 1.0, + "content": "Regret The reason why our results can imply an", + "type": "text" + }, + { + "bbox": [ + 304, + 666, + 336, + 680 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal { O } } ( { \\sqrt { T } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 666, + 451, + 681 + ], + "score": 1.0, + "content": "regret is that, after choosing", + "type": "text" + }, + { + "bbox": [ + 451, + 670, + 461, + 679 + ], + "score": 0.82, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 666, + 505, + 681 + ], + "score": 1.0, + "content": ", it follows", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 679, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 319, + 693 + ], + "score": 1.0, + "content": "from the argument of Theorem 1 that with probability", + "type": "text" + }, + { + "bbox": [ + 320, + 681, + 342, + 691 + ], + "score": 0.86, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 680, + 372, + 693 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 372, + 679, + 454, + 693 + ], + "score": 0.93, + "content": "\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 680, + 505, + 693 + ], + "score": 1.0, + "content": ", the number", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 691, + 256, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 691, + 117, + 704 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 117, + 693, + 127, + 703 + ], + "score": 0.85, + "content": "\\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 691, + 256, + 704 + ], + "score": 1.0, + "content": "-suboptimal steps is bounded by", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 666, + 505, + 704 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 708, + 399, + 736 + ], + "lines": [ + { + "bbox": [ + 211, + 708, + 399, + 736 + ], + "spans": [ + { + "bbox": [ + 211, + 708, + 399, + 736 + ], + "score": 0.94, + "content": "\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right) .", + "type": "interline_equation", + "image_path": "4312c45d17cd6e06d125f0e95014a5792379fa26cd13205bd4de52e4ac033fac.jpg" + } + ] + } + ], + "index": 40, + "virtual_lines": [ + { + "bbox": [ + 211, + 708, + 399, + 736 + ], + "spans": [], + "index": 40 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 446, + 96 + ], + "score": 1.0, + "content": "In contrast, Delayed Q-learning Strehl et al. (2006) can only give an upper bound on", + "type": "text" + }, + { + "bbox": [ + 447, + 84, + 456, + 93 + ], + "score": 0.84, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "-suboptimal", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 235, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 221, + 107 + ], + "score": 1.0, + "content": "steps after setting parameter", + "type": "text" + }, + { + "bbox": [ + 221, + 96, + 231, + 105 + ], + "score": 0.84, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 93, + 235, + 107 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 108, + 120, + 195, + 133 + ], + "lines": [ + { + "bbox": [ + 104, + 119, + 198, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 119, + 198, + 136 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 106, + 146, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 505, + 157 + ], + "score": 1.0, + "content": "Infinite-horizon MDP with discounted reward is a setting that is arguably more difficult than other", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 157, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 101, + 157, + 383, + 190 + ], + "score": 1.0, + "content": "popular settings, such as finite-horizon MDP. Previously, the best samby model-free reinforcement learning algorithms in this setting is", + "type": "text" + }, + { + "bbox": [ + 383, + 168, + 434, + 184 + ], + "score": 0.94, + "content": "\\tilde { \\cal O } \\big ( { \\textstyle \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma _ { _ { - } } ) ^ { 8 } } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 157, + 506, + 190 + ], + "score": 1.0, + "content": "y bound achieved, due to Delayed", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 182, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 506, + 195 + ], + "score": 1.0, + "content": "Q-learning Strehl et al. (2006). In this paper, we propose a variant of Q-learning that incorporates", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 188, + 509, + 214 + ], + "spans": [ + { + "bbox": [ + 102, + 188, + 380, + 214 + ], + "score": 1.0, + "content": "upper confidence bound, and show that it has a sample complexity of", + "type": "text" + }, + { + "bbox": [ + 381, + 193, + 431, + 209 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 188, + 509, + 214 + ], + "score": 1.0, + "content": ". This matches the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 207, + 419, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 278, + 221 + ], + "score": 1.0, + "content": "best lower bound except in dependence on", + "type": "text" + }, + { + "bbox": [ + 278, + 208, + 320, + 220 + ], + "score": 0.93, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 207, + 419, + 221 + ], + "score": 1.0, + "content": "and logarithmic factors.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 236, + 242, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 243, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 243, + 250 + ], + "score": 1.0, + "content": "6 ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 260, + 504, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "The authors thank Chi Jin and Chongjie Zhang for helpful discussions. This work is supported", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 270, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 284 + ], + "score": 1.0, + "content": "by National Basic Research Program of China (973 Program) (grant no. 2015CB352502), NSFC", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 282, + 425, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 425, + 296 + ], + "score": 1.0, + "content": "(61573026), BJNSF (L172037) and Beijing Acedemy of Artificial Intelligence.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "title", + "bbox": [ + 107, + 310, + 176, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 176, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 176, + 323 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 329, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 327, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 343 + ], + "score": 1.0, + "content": "Mohammad Gheshlaghi Azar, Remi Munos, Mohammad Ghavamzadeh, and Hilbert Kappen. Speedy", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 339, + 409, + 352 + ], + "spans": [ + { + "bbox": [ + 116, + 339, + 409, + 352 + ], + "score": 1.0, + "content": "q-learning. In Advances in neural information processing systems, 2011.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "score": 1.0, + "content": "Mohammad Gheshlaghi Azar, Ian Osband, and Rémi Munos. Minimax regret bounds for reinforce-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 369, + 340, + 382 + ], + "spans": [ + { + "bbox": [ + 116, + 369, + 340, + 382 + ], + "score": 1.0, + "content": "ment learning. arXiv preprint arXiv:1703.05449, 2017.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 506, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "Ronen I. Brafman and Moshe Tennenholtz. R-max - a general polynomial time algorithm for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 399, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 114, + 399, + 506, + 411 + ], + "score": 1.0, + "content": "near-optimal reinforcement learning. J. Mach. Learn. Res., 3:213–231, March 2003. ISSN", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 410, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 506, + 422 + ], + "score": 1.0, + "content": "1532-4435. doi: 10.1162/153244303765208377. URL https://doi.org/10.1162/", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 421, + 230, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 230, + 433 + ], + "score": 1.0, + "content": "153244303765208377.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "Christoph Dann, Tor Lattimore, and Emma Brunskill. Unifying pac and regret: Uniform pac bounds", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 449, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 449, + 506, + 464 + ], + "score": 1.0, + "content": "for episodic reinforcement learning. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 462, + 191, + 473 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 191, + 473 + ], + "score": 1.0, + "content": "5713–5723, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 504, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 494 + ], + "score": 1.0, + "content": "Eyal Even-Dar and Yishay Mansour. Learning rates for q-learning. Journal of Machine Learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 491, + 237, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 237, + 502 + ], + "score": 1.0, + "content": "Research, 5(Dec):1–25, 2003.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "Thomas Jaksch, Ronald Ortner, and Peter Auer. Near-optimal regret bounds for reinforcement", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 520, + 426, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 520, + 426, + 533 + ], + "score": 1.0, + "content": "learning. Journal of Machine Learning Research, 11(Apr):1563–1600, 2010.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "Chi Jin, Zeyuan Allen-Zhu, Sebastien Bubeck, and Michael I Jordan. Is q-learning provably efficient?", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 550, + 430, + 562 + ], + "spans": [ + { + "bbox": [ + 116, + 550, + 430, + 562 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 4864–4874, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "Sham Machandranath Kakade et al. On the sample complexity of reinforcement learning. PhD thesis,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 580, + 303, + 591 + ], + "spans": [ + { + "bbox": [ + 117, + 580, + 303, + 591 + ], + "score": 1.0, + "content": "University of London London, England, 2003.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 105, + 598, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 612 + ], + "score": 1.0, + "content": "Tor Lattimore and Marcus Hutter. Pac bounds for discounted mdps. In International Conference on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 609, + 357, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 357, + 622 + ], + "score": 1.0, + "content": "Algorithmic Learning Theory, pp. 320–334. Springer, 2012.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 108, + 627, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 651, + 214, + 661 + ], + "spans": [ + { + "bbox": [ + 116, + 651, + 214, + 661 + ], + "score": 1.0, + "content": "arXiv:1312.5602, 2013.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 680, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 505, + 692 + ], + "score": 1.0, + "content": "Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 690, + 444, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 444, + 703 + ], + "score": 1.0, + "content": "learning. In International conference on machine learning, pp. 1928–1937, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 720, + 495, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 495, + 732 + ], + "score": 1.0, + "content": "policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 446, + 96 + ], + "score": 1.0, + "content": "In contrast, Delayed Q-learning Strehl et al. (2006) can only give an upper bound on", + "type": "text" + }, + { + "bbox": [ + 447, + 84, + 456, + 93 + ], + "score": 0.84, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "-suboptimal", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 235, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 221, + 107 + ], + "score": 1.0, + "content": "steps after setting parameter", + "type": "text" + }, + { + "bbox": [ + 221, + 96, + 231, + 105 + ], + "score": 0.84, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 93, + 235, + 107 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 107 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 120, + 195, + 133 + ], + "lines": [ + { + "bbox": [ + 104, + 119, + 198, + 136 + ], + "spans": [ + { + "bbox": [ + 104, + 119, + 198, + 136 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 106, + 146, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 505, + 157 + ], + "score": 1.0, + "content": "Infinite-horizon MDP with discounted reward is a setting that is arguably more difficult than other", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 157, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 101, + 157, + 383, + 190 + ], + "score": 1.0, + "content": "popular settings, such as finite-horizon MDP. Previously, the best samby model-free reinforcement learning algorithms in this setting is", + "type": "text" + }, + { + "bbox": [ + 383, + 168, + 434, + 184 + ], + "score": 0.94, + "content": "\\tilde { \\cal O } \\big ( { \\textstyle \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma _ { _ { - } } ) ^ { 8 } } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 157, + 506, + 190 + ], + "score": 1.0, + "content": "y bound achieved, due to Delayed", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 182, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 506, + 195 + ], + "score": 1.0, + "content": "Q-learning Strehl et al. (2006). In this paper, we propose a variant of Q-learning that incorporates", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 188, + 509, + 214 + ], + "spans": [ + { + "bbox": [ + 102, + 188, + 380, + 214 + ], + "score": 1.0, + "content": "upper confidence bound, and show that it has a sample complexity of", + "type": "text" + }, + { + "bbox": [ + 381, + 193, + 431, + 209 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 188, + 509, + 214 + ], + "score": 1.0, + "content": ". This matches the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 207, + 419, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 278, + 221 + ], + "score": 1.0, + "content": "best lower bound except in dependence on", + "type": "text" + }, + { + "bbox": [ + 278, + 208, + 320, + 220 + ], + "score": 0.93, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 207, + 419, + 221 + ], + "score": 1.0, + "content": "and logarithmic factors.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 101, + 146, + 509, + 221 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 236, + 242, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 243, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 243, + 250 + ], + "score": 1.0, + "content": "6 ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 260, + 504, + 294 + ], + "lines": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "The authors thank Chi Jin and Chongjie Zhang for helpful discussions. This work is supported", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 270, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 284 + ], + "score": 1.0, + "content": "by National Basic Research Program of China (973 Program) (grant no. 2015CB352502), NSFC", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 282, + 425, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 425, + 296 + ], + "score": 1.0, + "content": "(61573026), BJNSF (L172037) and Beijing Acedemy of Artificial Intelligence.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 260, + 506, + 296 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 310, + 176, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 176, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 176, + 323 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 329, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 327, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 505, + 343 + ], + "score": 1.0, + "content": "Mohammad Gheshlaghi Azar, Remi Munos, Mohammad Ghavamzadeh, and Hilbert Kappen. Speedy", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 339, + 409, + 352 + ], + "spans": [ + { + "bbox": [ + 116, + 339, + 409, + 352 + ], + "score": 1.0, + "content": "q-learning. In Advances in neural information processing systems, 2011.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 327, + 505, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 358, + 504, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 371 + ], + "score": 1.0, + "content": "Mohammad Gheshlaghi Azar, Ian Osband, and Rémi Munos. Minimax regret bounds for reinforce-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 369, + 340, + 382 + ], + "spans": [ + { + "bbox": [ + 116, + 369, + 340, + 382 + ], + "score": 1.0, + "content": "ment learning. arXiv preprint arXiv:1703.05449, 2017.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 356, + 506, + 382 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 387, + 506, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "Ronen I. Brafman and Moshe Tennenholtz. R-max - a general polynomial time algorithm for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 399, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 114, + 399, + 506, + 411 + ], + "score": 1.0, + "content": "near-optimal reinforcement learning. J. Mach. Learn. Res., 3:213–231, March 2003. ISSN", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 410, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 506, + 422 + ], + "score": 1.0, + "content": "1532-4435. doi: 10.1162/153244303765208377. URL https://doi.org/10.1162/", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 421, + 230, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 230, + 433 + ], + "score": 1.0, + "content": "153244303765208377.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 387, + 506, + 433 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 439, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 506, + 452 + ], + "score": 1.0, + "content": "Christoph Dann, Tor Lattimore, and Emma Brunskill. Unifying pac and regret: Uniform pac bounds", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 449, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 115, + 449, + 506, + 464 + ], + "score": 1.0, + "content": "for episodic reinforcement learning. In Advances in Neural Information Processing Systems, pp.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 462, + 191, + 473 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 191, + 473 + ], + "score": 1.0, + "content": "5713–5723, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 439, + 506, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 504, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 494 + ], + "score": 1.0, + "content": "Eyal Even-Dar and Yishay Mansour. Learning rates for q-learning. Journal of Machine Learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 491, + 237, + 502 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 237, + 502 + ], + "score": 1.0, + "content": "Research, 5(Dec):1–25, 2003.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 478, + 505, + 502 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "Thomas Jaksch, Ronald Ortner, and Peter Auer. Near-optimal regret bounds for reinforcement", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 520, + 426, + 533 + ], + "spans": [ + { + "bbox": [ + 115, + 520, + 426, + 533 + ], + "score": 1.0, + "content": "learning. Journal of Machine Learning Research, 11(Apr):1563–1600, 2010.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 509, + 505, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 539, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "Chi Jin, Zeyuan Allen-Zhu, Sebastien Bubeck, and Michael I Jordan. Is q-learning provably efficient?", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 550, + 430, + 562 + ], + "spans": [ + { + "bbox": [ + 116, + 550, + 430, + 562 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems, pp. 4864–4874, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 539, + 505, + 562 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "Sham Machandranath Kakade et al. On the sample complexity of reinforcement learning. PhD thesis,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 117, + 580, + 303, + 591 + ], + "spans": [ + { + "bbox": [ + 117, + 580, + 303, + 591 + ], + "score": 1.0, + "content": "University of London London, England, 2003.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 106, + 568, + 506, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 598, + 504, + 622 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 612 + ], + "score": 1.0, + "content": "Tor Lattimore and Marcus Hutter. Pac bounds for discounted mdps. In International Conference on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 609, + 357, + 622 + ], + "spans": [ + { + "bbox": [ + 115, + 609, + 357, + 622 + ], + "score": 1.0, + "content": "Algorithmic Learning Theory, pp. 320–334. Springer, 2012.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 597, + 505, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 627, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 651, + 214, + 661 + ], + "spans": [ + { + "bbox": [ + 116, + 651, + 214, + 661 + ], + "score": 1.0, + "content": "arXiv:1312.5602, 2013.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 627, + 505, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 680, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 505, + 692 + ], + "score": 1.0, + "content": "Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 690, + 444, + 703 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 444, + 703 + ], + "score": 1.0, + "content": "learning. In International conference on machine learning, pp. 1928–1937, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 668, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 720, + 495, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 495, + 732 + ], + "score": 1.0, + "content": "policy optimization. In International Conference on Machine Learning, pp. 1889–1897, 2015.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Aaron Sidford, Mengdi Wang, Xian Wu, Lin Yang, and Yinyu Ye. Near-optimal time and sample", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 95, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 117, + 95, + 505, + 105 + ], + "score": 1.0, + "content": "complexities for solving markov decision processes with a generative model. In Advances in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 374, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 374, + 118 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 5186–5196, 2018a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "Aaron Sidford, Mengdi Wang, Xian Wu, and Yinyu Ye. Variance reduced value iteration and faster", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "algorithms for solving markov decision processes. In Proceedings of the Twenty-Ninth Annual", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 145, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 114, + 145, + 505, + 159 + ], + "score": 1.0, + "content": "ACM-SIAM Symposium on Discrete Algorithms, pp. 770–787. Society for Industrial and Applied", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 202, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 202, + 168 + ], + "score": 1.0, + "content": "Mathematics, 2018b.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 105, + 174, + 503, + 198 + ], + "lines": [ + { + "bbox": [ + 107, + 176, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 107, + 176, + 505, + 187 + ], + "score": 1.0, + "content": "Alexander L Strehl and Michael L Littman. An analysis of model-based interval estimation for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 501, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 501, + 198 + ], + "score": 1.0, + "content": "markov decision processes. Journal of Computer and System Sciences, 74(8):1309–1331, 2008.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 506, + 239 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 506, + 217 + ], + "score": 1.0, + "content": "Alexander L Strehl, Lihong Li, Eric Wiewiora, John Langford, and Michael L Littman. Pac model-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 116, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "free reinforcement learning. In Proceedings of the 23rd international conference on Machine", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 227, + 262, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 227, + 262, + 239 + ], + "score": 1.0, + "content": "learning, pp. 881–888. ACM, 2006.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 504, + 258 + ], + "score": 1.0, + "content": "István Szita and Csaba Szepesvári. Model-based reinforcement learning with nearly tight exploration", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 256, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 115, + 256, + 505, + 271 + ], + "score": 1.0, + "content": "complexity bounds. In Proceedings of the 27th International Conference on Machine Learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 268, + 254, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 268, + 254, + 281 + ], + "score": 1.0, + "content": "(ICML-10), pp. 1031–1038, 2010.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 300, + 232, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 233, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 233, + 315 + ], + "score": 1.0, + "content": "A PROOF OF LEMMA 1", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 105, + 323, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 104, + 321, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 321, + 215, + 342 + ], + "score": 1.0, + "content": "Lemma 1. For fixed t and", + "type": "text" + }, + { + "bbox": [ + 215, + 326, + 240, + 338 + ], + "score": 0.92, + "content": "\\eta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 321, + 255, + 342 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 256, + 323, + 275, + 338 + ], + "score": 0.92, + "content": "B _ { \\eta } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 321, + 344, + 342 + ], + "score": 1.0, + "content": "be the event that", + "type": "text" + }, + { + "bbox": [ + 344, + 325, + 457, + 340 + ], + "score": 0.93, + "content": "\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 321, + 507, + 342 + ], + "score": 1.0, + "content": "in step t. If", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 338, + 297, + 353 + ], + "spans": [ + { + "bbox": [ + 107, + 339, + 139, + 351 + ], + "score": 0.9, + "content": "\\eta > 2 \\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 338, + 261, + 353 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 261, + 339, + 294, + 351 + ], + "score": 0.9, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 338, + 297, + 353 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 356, + 423, + 390 + ], + "lines": [ + { + "bbox": [ + 186, + 356, + 423, + 390 + ], + "spans": [ + { + "bbox": [ + 186, + 356, + 423, + 390 + ], + "score": 0.92, + "content": "\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ,", + "type": "interline_equation", + "image_path": "0b9f09a826319782774c08f51f81a7e3a19a570518937c623163e273091ef961.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 356, + 423, + 373.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 186, + 373.0, + 423, + 390.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 248, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 249, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 133, + 408 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 395, + 147, + 407 + ], + "score": 0.87, + "content": "I [ \\cdot ]", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 395, + 249, + 408 + ], + "score": 1.0, + "content": "is the indicator function.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 418, + 431, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 432, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 162, + 433 + ], + "score": 1.0, + "content": "Proof. When", + "type": "text" + }, + { + "bbox": [ + 163, + 420, + 187, + 431 + ], + "score": 0.91, + "content": "\\eta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 418, + 403, + 433 + ], + "score": 1.0, + "content": "the lemma holds trivially. Now consider the case that", + "type": "text" + }, + { + "bbox": [ + 403, + 420, + 428, + 431 + ], + "score": 0.91, + "content": "\\eta \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 418, + 432, + 433 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 430, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 427, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 122, + 452 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 435, + 279, + 451 + ], + "score": 0.92, + "content": "\\begin{array} { r } { I = \\{ t \\colon V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } \\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 433, + 404, + 452 + ], + "score": 1.0, + "content": ". By lemma 2, with probability", + "type": "text" + }, + { + "bbox": [ + 404, + 437, + 427, + 447 + ], + "score": 0.84, + "content": "1 - \\delta", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 457, + 472, + 592 + ], + "lines": [ + { + "bbox": [ + 138, + 457, + 472, + 592 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 472, + 592 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } \\leq \\displaystyle \\sum _ { t \\in I } \\left( V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\right) \\leq \\displaystyle \\sum _ { t \\in I } \\left[ \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { 1 } { ( 1 - \\gamma ) ^ { 5 / 2 } } \\sqrt { S A | I | \\ell | ( I | ) } + \\frac { S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln | I | \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln \\frac { S A | I | } { \\delta } } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln S A | I | } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\end{array}", + "type": "interline_equation", + "image_path": "c3600e24303e29ac9ff643f0f4892a82d147d38d65c5111e6d8ac103703617e6.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 138, + 457, + 472, + 502.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 138, + 502.0, + 472, + 547.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 138, + 547.0, + 472, + 592.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 487, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 488, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 160, + 635 + ], + "score": 1.0, + "content": "Suppose that", + "type": "text" + }, + { + "bbox": [ + 160, + 617, + 247, + 634 + ], + "score": 0.94, + "content": "\\begin{array} { r } { | I | = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 616, + 288, + 635 + ], + "score": 1.0, + "content": ", for some", + "type": "text" + }, + { + "bbox": [ + 288, + 619, + 313, + 630 + ], + "score": 0.9, + "content": "k > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 616, + 471, + 635 + ], + "score": 1.0, + "content": ". Then it follows that for some constant", + "type": "text" + }, + { + "bbox": [ + 472, + 620, + 484, + 631 + ], + "score": 0.88, + "content": "C _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 616, + 488, + 635 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 639, + 483, + 733 + ], + "lines": [ + { + "bbox": [ + 128, + 639, + 483, + 733 + ], + "spans": [ + { + "bbox": [ + 128, + 639, + 483, + 733 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } = \\cfrac { k ^ { 2 } S A \\ln S A } { ( 1 - \\gamma ) ^ { 4 } \\eta } \\leq 2 \\cfrac { ( \\eta - \\epsilon _ { 1 } ) | I | } { 1 - \\gamma } } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { \\sqrt { S A | I | \\ln \\left( S A | I | \\right) } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { S A k } { \\eta ( 1 - \\gamma ) ^ { 4 } } \\sqrt { \\ln S A \\cdot ( \\ln S A + \\ln | I | ) } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "c06bd9e11a4120e53c0b8d2cd822906d72211ae50589d30086bcbaef0151ca3a.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 128, + 639, + 483, + 670.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 128, + 670.3333333333334, + 483, + 701.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 128, + 701.6666666666667, + 483, + 733.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Aaron Sidford, Mengdi Wang, Xian Wu, Lin Yang, and Yinyu Ye. Near-optimal time and sample", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 95, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 117, + 95, + 505, + 105 + ], + "score": 1.0, + "content": "complexities for solving markov decision processes with a generative model. In Advances in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 374, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 374, + 118 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 5186–5196, 2018a.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 81, + 506, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 124, + 505, + 136 + ], + "score": 1.0, + "content": "Aaron Sidford, Mengdi Wang, Xian Wu, and Yinyu Ye. Variance reduced value iteration and faster", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "algorithms for solving markov decision processes. In Proceedings of the Twenty-Ninth Annual", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 145, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 114, + 145, + 505, + 159 + ], + "score": 1.0, + "content": "ACM-SIAM Symposium on Discrete Algorithms, pp. 770–787. Society for Industrial and Applied", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 202, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 202, + 168 + ], + "score": 1.0, + "content": "Mathematics, 2018b.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 124, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 174, + 503, + 198 + ], + "lines": [ + { + "bbox": [ + 107, + 176, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 107, + 176, + 505, + 187 + ], + "score": 1.0, + "content": "Alexander L Strehl and Michael L Littman. An analysis of model-based interval estimation for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 501, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 501, + 198 + ], + "score": 1.0, + "content": "markov decision processes. Journal of Computer and System Sciences, 74(8):1309–1331, 2008.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 107, + 176, + 505, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 204, + 506, + 239 + ], + "lines": [ + { + "bbox": [ + 106, + 205, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 506, + 217 + ], + "score": 1.0, + "content": "Alexander L Strehl, Lihong Li, Eric Wiewiora, John Langford, and Michael L Littman. Pac model-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 116, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "free reinforcement learning. In Proceedings of the 23rd international conference on Machine", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 227, + 262, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 227, + 262, + 239 + ], + "score": 1.0, + "content": "learning, pp. 881–888. ACM, 2006.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 205, + 506, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 245, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 106, + 246, + 504, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 504, + 258 + ], + "score": 1.0, + "content": "István Szita and Csaba Szepesvári. Model-based reinforcement learning with nearly tight exploration", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 256, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 115, + 256, + 505, + 271 + ], + "score": 1.0, + "content": "complexity bounds. In Proceedings of the 27th International Conference on Machine Learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 268, + 254, + 281 + ], + "spans": [ + { + "bbox": [ + 115, + 268, + 254, + 281 + ], + "score": 1.0, + "content": "(ICML-10), pp. 1031–1038, 2010.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 246, + 505, + 281 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 300, + 232, + 313 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 233, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 233, + 315 + ], + "score": 1.0, + "content": "A PROOF OF LEMMA 1", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 105, + 323, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 104, + 321, + 507, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 321, + 215, + 342 + ], + "score": 1.0, + "content": "Lemma 1. For fixed t and", + "type": "text" + }, + { + "bbox": [ + 215, + 326, + 240, + 338 + ], + "score": 0.92, + "content": "\\eta > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 321, + 255, + 342 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 256, + 323, + 275, + 338 + ], + "score": 0.92, + "content": "B _ { \\eta } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 321, + 344, + 342 + ], + "score": 1.0, + "content": "be the event that", + "type": "text" + }, + { + "bbox": [ + 344, + 325, + 457, + 340 + ], + "score": 0.93, + "content": "\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 321, + 507, + 342 + ], + "score": 1.0, + "content": "in step t. If", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 338, + 297, + 353 + ], + "spans": [ + { + "bbox": [ + 107, + 339, + 139, + 351 + ], + "score": 0.9, + "content": "\\eta > 2 \\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 338, + 261, + 353 + ], + "score": 1.0, + "content": ", then with probability at least", + "type": "text" + }, + { + "bbox": [ + 261, + 339, + 294, + 351 + ], + "score": 0.9, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 338, + 297, + 353 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 321, + 507, + 353 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 356, + 423, + 390 + ], + "lines": [ + { + "bbox": [ + 186, + 356, + 423, + 390 + ], + "spans": [ + { + "bbox": [ + 186, + 356, + 423, + 390 + ], + "score": 0.92, + "content": "\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ,", + "type": "interline_equation", + "image_path": "0b9f09a826319782774c08f51f81a7e3a19a570518937c623163e273091ef961.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 186, + 356, + 423, + 373.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 186, + 373.0, + 423, + 390.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 248, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 249, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 133, + 408 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 395, + 147, + 407 + ], + "score": 0.87, + "content": "I [ \\cdot ]", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 395, + 249, + 408 + ], + "score": 1.0, + "content": "is the indicator function.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 395, + 249, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 418, + 431, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 418, + 432, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 162, + 433 + ], + "score": 1.0, + "content": "Proof. When", + "type": "text" + }, + { + "bbox": [ + 163, + 420, + 187, + 431 + ], + "score": 0.91, + "content": "\\eta > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 418, + 403, + 433 + ], + "score": 1.0, + "content": "the lemma holds trivially. Now consider the case that", + "type": "text" + }, + { + "bbox": [ + 403, + 420, + 428, + 431 + ], + "score": 0.91, + "content": "\\eta \\leq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 418, + 432, + 433 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 418, + 432, + 433 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 430, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 427, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 122, + 452 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 435, + 279, + 451 + ], + "score": 0.92, + "content": "\\begin{array} { r } { I = \\{ t \\colon V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } \\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 433, + 404, + 452 + ], + "score": 1.0, + "content": ". By lemma 2, with probability", + "type": "text" + }, + { + "bbox": [ + 404, + 437, + 427, + 447 + ], + "score": 0.84, + "content": "1 - \\delta", + "type": "inline_equation" + } + ], + "index": 22 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 433, + 427, + 452 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 457, + 472, + 592 + ], + "lines": [ + { + "bbox": [ + 138, + 457, + 472, + 592 + ], + "spans": [ + { + "bbox": [ + 138, + 457, + 472, + 592 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } \\leq \\displaystyle \\sum _ { t \\in I } \\left( V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\right) \\leq \\displaystyle \\sum _ { t \\in I } \\left[ \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { 1 } { ( 1 - \\gamma ) ^ { 5 / 2 } } \\sqrt { S A | I | \\ell | ( I | ) } + \\frac { S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln | I | \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln \\frac { S A | I | } { \\delta } } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln S A | I | } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\end{array}", + "type": "interline_equation", + "image_path": "c3600e24303e29ac9ff643f0f4892a82d147d38d65c5111e6d8ac103703617e6.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 138, + 457, + 472, + 502.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 138, + 502.0, + 472, + 547.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 138, + 547.0, + 472, + 592.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 617, + 487, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 488, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 160, + 635 + ], + "score": 1.0, + "content": "Suppose that", + "type": "text" + }, + { + "bbox": [ + 160, + 617, + 247, + 634 + ], + "score": 0.94, + "content": "\\begin{array} { r } { | I | = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 616, + 288, + 635 + ], + "score": 1.0, + "content": ", for some", + "type": "text" + }, + { + "bbox": [ + 288, + 619, + 313, + 630 + ], + "score": 0.9, + "content": "k > 1", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 616, + 471, + 635 + ], + "score": 1.0, + "content": ". Then it follows that for some constant", + "type": "text" + }, + { + "bbox": [ + 472, + 620, + 484, + 631 + ], + "score": 0.88, + "content": "C _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 616, + 488, + 635 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 616, + 488, + 635 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 639, + 483, + 733 + ], + "lines": [ + { + "bbox": [ + 128, + 639, + 483, + 733 + ], + "spans": [ + { + "bbox": [ + 128, + 639, + 483, + 733 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } = \\cfrac { k ^ { 2 } S A \\ln S A } { ( 1 - \\gamma ) ^ { 4 } \\eta } \\leq 2 \\cfrac { ( \\eta - \\epsilon _ { 1 } ) | I | } { 1 - \\gamma } } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { \\sqrt { S A | I | \\ln \\left( S A | I | \\right) } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { S A k } { \\eta ( 1 - \\gamma ) ^ { 4 } } \\sqrt { \\ln S A \\cdot ( \\ln S A + \\ln | I | ) } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "c06bd9e11a4120e53c0b8d2cd822906d72211ae50589d30086bcbaef0151ca3a.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 128, + 639, + 483, + 670.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 128, + 670.3333333333334, + 483, + 701.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 128, + 701.6666666666667, + 483, + 733.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 147, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 148, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 148, + 94 + ], + "score": 1.0, + "content": "Therefore", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 97, + 461, + 217 + ], + "lines": [ + { + "bbox": [ + 150, + 97, + 461, + 217 + ], + "spans": [ + { + "bbox": [ + 150, + 97, + 461, + 217 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { k ^ { 2 } \\ln ( S A ) \\le C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( k \\left( \\ln S A + \\ln | I | \\right) + \\eta ( 1 - \\gamma ) \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\ln S A + 2 \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( 3 \\ln S A + 4 \\ln k + 6 \\ln \\frac { 1 } { \\eta ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\le 6 k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\ln S A + \\ln \\epsilon k \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "bae465b19e019059a91e97016934af5f7bf0c47e9357a7ca6741dfa6dd715015.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 150, + 97, + 461, + 137.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 150, + 137.0, + 461, + 177.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 150, + 177.0, + 461, + 217.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 222, + 303, + 241 + ], + "lines": [ + { + "bbox": [ + 100, + 220, + 307, + 246 + ], + "spans": [ + { + "bbox": [ + 100, + 220, + 122, + 246 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 222, + 275, + 243 + ], + "score": 0.94, + "content": "\\begin{array} { r } { C ^ { \\prime } = \\operatorname* { m a x } \\lbrace 2 , 6 C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\rbrace } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 220, + 307, + 246 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "interline_equation", + "bbox": [ + 268, + 246, + 343, + 262 + ], + "lines": [ + { + "bbox": [ + 268, + 246, + 343, + 262 + ], + "spans": [ + { + "bbox": [ + 268, + 246, + 343, + 262 + ], + "score": 0.89, + "content": "k \\leq C ^ { \\prime } ( 2 + \\ln k ) .", + "type": "interline_equation", + "image_path": "6648dce62ed22bd5032b6259b54c85d35efee676cf37255340a1199f339eda3d.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 268, + 246, + 343, + 262 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 201, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 202, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 116, + 280 + ], + "score": 1.0, + "content": "If", + "type": "text" + }, + { + "bbox": [ + 116, + 267, + 178, + 278 + ], + "score": 0.91, + "content": "k \\geq 1 0 C ^ { \\prime } \\ln C ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 265, + 202, + 280 + ], + "score": 1.0, + "content": ", then", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 282, + 401, + 315 + ], + "lines": [ + { + "bbox": [ + 209, + 282, + 401, + 315 + ], + "spans": [ + { + "bbox": [ + 209, + 282, + 401, + 315 + ], + "score": 0.9, + "content": "\\begin{array} { c } { k - C ^ { \\prime } \\left( 2 + \\ln k \\right) \\geq 8 C ^ { \\prime } \\ln C ^ { \\prime } - ( 2 + \\ln 1 0 ) C ^ { \\prime } } \\\\ { \\geq 4 C ^ { \\prime } \\left( 2 \\ln C ^ { \\prime } - 4 \\right) \\geq 0 , } \\end{array}", + "type": "interline_equation", + "image_path": "4ae996b10596e3c7079276b4b960d96f44d70fb21ef5f209a0ea7326b6d16c83.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 282, + 401, + 298.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 209, + 298.5, + 401, + 315.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 326, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 325, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 294, + 332 + ], + "score": 1.0, + "content": "which means violation of (9). Therefore, since", + "type": "text" + }, + { + "bbox": [ + 294, + 318, + 325, + 330 + ], + "score": 0.91, + "content": "C ^ { \\prime } \\geq 2", + "type": "inline_equation" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 334, + 417, + 361 + ], + "lines": [ + { + "bbox": [ + 193, + 334, + 417, + 361 + ], + "spans": [ + { + "bbox": [ + 193, + 334, + 417, + 361 + ], + "score": 0.9, + "content": "k \\leq 1 0 C ^ { \\prime } \\ln C ^ { \\prime } \\leq 3 6 0 C _ { 1 } ^ { 2 } \\operatorname* { m a x } \\{ \\ln ^ { 4 } { \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } } , 2 0 \\ln 2 \\} .", + "type": "interline_equation", + "image_path": "3659eb39bbdbcc0c60a416911c59460e57107ec3dbd0e6f62d3b7b93a525422e.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 193, + 334, + 417, + 361 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 217, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 217, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 217, + 379 + ], + "score": 1.0, + "content": "It immediately follows that", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 382, + 401, + 440 + ], + "lines": [ + { + "bbox": [ + 210, + 382, + 401, + 440 + ], + "spans": [ + { + "bbox": [ + 210, + 382, + 401, + 440 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { \\vert I \\vert = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } } \\\\ & { \\leq \\frac { S A \\ln S A } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\ln \\frac { 1 } { \\delta } \\cdot \\mathcal { O } ( \\ln ^ { 8 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ad528682451c2a80d59b5c6f1508b5b7b9f26c92e6092f8aeb71c929441cec39.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 210, + 382, + 401, + 401.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 210, + 401.3333333333333, + 401, + 420.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 210, + 420.66666666666663, + 401, + 439.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 477, + 233, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 233, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 233, + 492 + ], + "score": 1.0, + "content": "B PROOF OF LEMMA 2", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 103, + 501, + 484, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 486, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 195, + 516 + ], + "score": 1.0, + "content": "Lemma 2. For every", + "type": "text" + }, + { + "bbox": [ + 195, + 502, + 222, + 514 + ], + "score": 0.91, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 501, + 264, + 516 + ], + "score": 1.0, + "content": "-sequence", + "type": "text" + }, + { + "bbox": [ + 264, + 502, + 297, + 515 + ], + "score": 0.91, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 501, + 366, + 516 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + }, + { + "bbox": [ + 367, + 502, + 399, + 514 + ], + "score": 0.89, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 501, + 486, + 516 + ], + "score": 1.0, + "content": ", the following holds:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 534, + 482, + 568 + ], + "lines": [ + { + "bbox": [ + 128, + 534, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 128, + 534, + 482, + 568 + ], + "score": 0.93, + "content": "\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "ff487b4def18b98c770f155f72c7d52760fabc4d6097e5e050552ffa5fb8baab.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 128, + 534, + 482, + 545.3333333333334 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 128, + 545.3333333333334, + 482, + 556.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 128, + 556.6666666666667, + 482, + 568.0000000000001 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 580, + 289, + 596 + ], + "lines": [ + { + "bbox": [ + 102, + 577, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 102, + 577, + 133, + 601 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 580, + 228, + 596 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 577, + 294, + 601 + ], + "score": 1.0, + "content": "is a log-factor.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 381, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 381, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 381, + 613 + ], + "score": 1.0, + "content": "Fact 1. (1) The following statement holds throughout the algorithm,", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 251, + 615, + 359, + 632 + ], + "lines": [ + { + "bbox": [ + 251, + 615, + 359, + 632 + ], + "spans": [ + { + "bbox": [ + 251, + 615, + 359, + 632 + ], + "score": 0.91, + "content": "\\hat { Q } _ { p + 1 } ( s , a ) \\leq Q _ { p + 1 } ( s , a ) .", + "type": "interline_equation", + "image_path": "33f263fc3dd14df50727d4ef76ef6c4ffe1d222ffa11393ce94b14974de75afd.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 251, + 615, + 359, + 632 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 277, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 278, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 153, + 649 + ], + "score": 1.0, + "content": "(2) For any", + "type": "text" + }, + { + "bbox": [ + 153, + 639, + 159, + 648 + ], + "score": 0.54, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 635, + 210, + 649 + ], + "score": 1.0, + "content": ", there exists", + "type": "text" + }, + { + "bbox": [ + 210, + 636, + 238, + 648 + ], + "score": 0.92, + "content": "p ^ { \\prime } \\leq p", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 635, + 278, + 649 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 653, + 360, + 669 + ], + "lines": [ + { + "bbox": [ + 250, + 653, + 360, + 669 + ], + "spans": [ + { + "bbox": [ + 250, + 653, + 360, + 669 + ], + "score": 0.91, + "content": "{ \\hat { Q } } _ { p + 1 } ( s , a ) \\geq Q _ { p ^ { \\prime } + 1 } ( s , a ) .", + "type": "interline_equation", + "image_path": "6fac3e0bba44afb720a787e2593dcafd19d5e03a7a0d6878331dbb6aecc146fa.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 250, + 653, + 360, + 669 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 420, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 420, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 420, + 694 + ], + "score": 1.0, + "content": "Proof. Both properties are results of the update rule at line 11 of Algorithm 1.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 705, + 357, + 718 + ], + "lines": [ + { + "bbox": [ + 105, + 704, + 357, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 357, + 720 + ], + "score": 1.0, + "content": "Before proving lemma 2, we will prove two auxiliary lemmas.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 305, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 307, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 289, + 735 + ], + "score": 1.0, + "content": "Lemma 3. The following properties hold for", + "type": "text" + }, + { + "bbox": [ + 289, + 720, + 300, + 732 + ], + "score": 0.86, + "content": "\\alpha _ { t } ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 719, + 307, + 735 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "11", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 681, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 496, + 683, + 504, + 691 + ], + "spans": [ + { + "bbox": [ + 496, + 683, + 504, + 691 + ], + "score": 0.997, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 449, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 496, + 451, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 496, + 451, + 505, + 461 + ], + "score": 1.0, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 147, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 148, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 148, + 94 + ], + "score": 1.0, + "content": "Therefore", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 106, + 82, + 148, + 94 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 150, + 97, + 461, + 217 + ], + "lines": [ + { + "bbox": [ + 150, + 97, + 461, + 217 + ], + "spans": [ + { + "bbox": [ + 150, + 97, + 461, + 217 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { k ^ { 2 } \\ln ( S A ) \\le C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( k \\left( \\ln S A + \\ln | I | \\right) + \\eta ( 1 - \\gamma ) \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\ln S A + 2 \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( 3 \\ln S A + 4 \\ln k + 6 \\ln \\frac { 1 } { \\eta ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\le 6 k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\ln S A + \\ln \\epsilon k \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "bae465b19e019059a91e97016934af5f7bf0c47e9357a7ca6741dfa6dd715015.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 150, + 97, + 461, + 137.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 150, + 137.0, + 461, + 177.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 150, + 177.0, + 461, + 217.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 222, + 303, + 241 + ], + "lines": [ + { + "bbox": [ + 100, + 220, + 307, + 246 + ], + "spans": [ + { + "bbox": [ + 100, + 220, + 122, + 246 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 123, + 222, + 275, + 243 + ], + "score": 0.94, + "content": "\\begin{array} { r } { C ^ { \\prime } = \\operatorname* { m a x } \\lbrace 2 , 6 C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\rbrace } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 220, + 307, + 246 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 100, + 220, + 307, + 246 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 268, + 246, + 343, + 262 + ], + "lines": [ + { + "bbox": [ + 268, + 246, + 343, + 262 + ], + "spans": [ + { + "bbox": [ + 268, + 246, + 343, + 262 + ], + "score": 0.89, + "content": "k \\leq C ^ { \\prime } ( 2 + \\ln k ) .", + "type": "interline_equation", + "image_path": "6648dce62ed22bd5032b6259b54c85d35efee676cf37255340a1199f339eda3d.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 268, + 246, + 343, + 262 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 201, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 202, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 116, + 280 + ], + "score": 1.0, + "content": "If", + "type": "text" + }, + { + "bbox": [ + 116, + 267, + 178, + 278 + ], + "score": 0.91, + "content": "k \\geq 1 0 C ^ { \\prime } \\ln C ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 265, + 202, + 280 + ], + "score": 1.0, + "content": ", then", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 265, + 202, + 280 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 282, + 401, + 315 + ], + "lines": [ + { + "bbox": [ + 209, + 282, + 401, + 315 + ], + "spans": [ + { + "bbox": [ + 209, + 282, + 401, + 315 + ], + "score": 0.9, + "content": "\\begin{array} { c } { k - C ^ { \\prime } \\left( 2 + \\ln k \\right) \\geq 8 C ^ { \\prime } \\ln C ^ { \\prime } - ( 2 + \\ln 1 0 ) C ^ { \\prime } } \\\\ { \\geq 4 C ^ { \\prime } \\left( 2 \\ln C ^ { \\prime } - 4 \\right) \\geq 0 , } \\end{array}", + "type": "interline_equation", + "image_path": "4ae996b10596e3c7079276b4b960d96f44d70fb21ef5f209a0ea7326b6d16c83.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 282, + 401, + 298.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 209, + 298.5, + 401, + 315.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 326, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 325, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 294, + 332 + ], + "score": 1.0, + "content": "which means violation of (9). Therefore, since", + "type": "text" + }, + { + "bbox": [ + 294, + 318, + 325, + 330 + ], + "score": 0.91, + "content": "C ^ { \\prime } \\geq 2", + "type": "inline_equation" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 317, + 325, + 332 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 334, + 417, + 361 + ], + "lines": [ + { + "bbox": [ + 193, + 334, + 417, + 361 + ], + "spans": [ + { + "bbox": [ + 193, + 334, + 417, + 361 + ], + "score": 0.9, + "content": "k \\leq 1 0 C ^ { \\prime } \\ln C ^ { \\prime } \\leq 3 6 0 C _ { 1 } ^ { 2 } \\operatorname* { m a x } \\{ \\ln ^ { 4 } { \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } } , 2 0 \\ln 2 \\} .", + "type": "interline_equation", + "image_path": "3659eb39bbdbcc0c60a416911c59460e57107ec3dbd0e6f62d3b7b93a525422e.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 193, + 334, + 417, + 361 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 217, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 217, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 217, + 379 + ], + "score": 1.0, + "content": "It immediately follows that", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 364, + 217, + 379 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 382, + 401, + 440 + ], + "lines": [ + { + "bbox": [ + 210, + 382, + 401, + 440 + ], + "spans": [ + { + "bbox": [ + 210, + 382, + 401, + 440 + ], + "score": 0.93, + "content": "\\begin{array} { r l r } { { \\vert I \\vert = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } } \\\\ & { \\leq \\frac { S A \\ln S A } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\ln \\frac { 1 } { \\delta } \\cdot \\mathcal { O } ( \\ln ^ { 8 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ad528682451c2a80d59b5c6f1508b5b7b9f26c92e6092f8aeb71c929441cec39.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 210, + 382, + 401, + 401.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 210, + 401.3333333333333, + 401, + 420.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 210, + 420.66666666666663, + 401, + 439.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "title", + "bbox": [ + 108, + 477, + 233, + 490 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 233, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 233, + 492 + ], + "score": 1.0, + "content": "B PROOF OF LEMMA 2", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 103, + 501, + 484, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 486, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 195, + 516 + ], + "score": 1.0, + "content": "Lemma 2. For every", + "type": "text" + }, + { + "bbox": [ + 195, + 502, + 222, + 514 + ], + "score": 0.91, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 501, + 264, + 516 + ], + "score": 1.0, + "content": "-sequence", + "type": "text" + }, + { + "bbox": [ + 264, + 502, + 297, + 515 + ], + "score": 0.91, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 501, + 366, + 516 + ], + "score": 1.0, + "content": ", with probability", + "type": "text" + }, + { + "bbox": [ + 367, + 502, + 399, + 514 + ], + "score": 0.89, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 501, + 486, + 516 + ], + "score": 1.0, + "content": ", the following holds:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 501, + 486, + 516 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 128, + 534, + 482, + 568 + ], + "lines": [ + { + "bbox": [ + 128, + 534, + 482, + 568 + ], + "spans": [ + { + "bbox": [ + 128, + 534, + 482, + 568 + ], + "score": 0.93, + "content": "\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "ff487b4def18b98c770f155f72c7d52760fabc4d6097e5e050552ffa5fb8baab.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 128, + 534, + 482, + 545.3333333333334 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 128, + 545.3333333333334, + 482, + 556.6666666666667 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 128, + 556.6666666666667, + 482, + 568.0000000000001 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 580, + 289, + 596 + ], + "lines": [ + { + "bbox": [ + 102, + 577, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 102, + 577, + 133, + 601 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 580, + 228, + 596 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 577, + 294, + 601 + ], + "score": 1.0, + "content": "is a log-factor.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 102, + 577, + 294, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 381, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 381, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 381, + 613 + ], + "score": 1.0, + "content": "Fact 1. (1) The following statement holds throughout the algorithm,", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 597, + 381, + 613 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 251, + 615, + 359, + 632 + ], + "lines": [ + { + "bbox": [ + 251, + 615, + 359, + 632 + ], + "spans": [ + { + "bbox": [ + 251, + 615, + 359, + 632 + ], + "score": 0.91, + "content": "\\hat { Q } _ { p + 1 } ( s , a ) \\leq Q _ { p + 1 } ( s , a ) .", + "type": "interline_equation", + "image_path": "33f263fc3dd14df50727d4ef76ef6c4ffe1d222ffa11393ce94b14974de75afd.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 251, + 615, + 359, + 632 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 277, + 648 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 278, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 153, + 649 + ], + "score": 1.0, + "content": "(2) For any", + "type": "text" + }, + { + "bbox": [ + 153, + 639, + 159, + 648 + ], + "score": 0.54, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 635, + 210, + 649 + ], + "score": 1.0, + "content": ", there exists", + "type": "text" + }, + { + "bbox": [ + 210, + 636, + 238, + 648 + ], + "score": 0.92, + "content": "p ^ { \\prime } \\leq p", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 635, + 278, + 649 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 635, + 278, + 649 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 250, + 653, + 360, + 669 + ], + "lines": [ + { + "bbox": [ + 250, + 653, + 360, + 669 + ], + "spans": [ + { + "bbox": [ + 250, + 653, + 360, + 669 + ], + "score": 0.91, + "content": "{ \\hat { Q } } _ { p + 1 } ( s , a ) \\geq Q _ { p ^ { \\prime } + 1 } ( s , a ) .", + "type": "interline_equation", + "image_path": "6fac3e0bba44afb720a787e2593dcafd19d5e03a7a0d6878331dbb6aecc146fa.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 250, + 653, + 360, + 669 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 420, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 420, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 420, + 694 + ], + "score": 1.0, + "content": "Proof. Both properties are results of the update rule at line 11 of Algorithm 1.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 680, + 420, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 705, + 357, + 718 + ], + "lines": [ + { + "bbox": [ + 105, + 704, + 357, + 720 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 357, + 720 + ], + "score": 1.0, + "content": "Before proving lemma 2, we will prove two auxiliary lemmas.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 704, + 357, + 720 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 720, + 305, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 307, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 289, + 735 + ], + "score": 1.0, + "content": "Lemma 3. The following properties hold for", + "type": "text" + }, + { + "bbox": [ + 289, + 720, + 300, + 732 + ], + "score": 0.86, + "content": "\\alpha _ { t } ^ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 719, + 307, + 735 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 719, + 307, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 129, + 79, + 353, + 101 + ], + "lines": [ + { + "bbox": [ + 127, + 78, + 356, + 102 + ], + "spans": [ + { + "bbox": [ + 127, + 78, + 142, + 102 + ], + "score": 1.0, + "content": "1.", + "type": "text" + }, + { + "bbox": [ + 142, + 80, + 259, + 101 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\sqrt { \\frac { 1 } { t } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { 1 } { i } } \\leq 2 \\sqrt { \\frac { 1 } { t } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 78, + 298, + 102 + ], + "score": 1.0, + "content": "for every", + "type": "text" + }, + { + "bbox": [ + 299, + 84, + 349, + 97 + ], + "score": 0.58, + "content": "t \\geq 1 , c > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 78, + 356, + 102 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 129, + 106, + 374, + 123 + ], + "lines": [ + { + "bbox": [ + 126, + 104, + 372, + 125 + ], + "spans": [ + { + "bbox": [ + 126, + 104, + 142, + 125 + ], + "score": 1.0, + "content": "2.", + "type": "text" + }, + { + "bbox": [ + 142, + 108, + 216, + 122 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\operatorname* { m a x } _ { i \\in [ t ] } \\alpha _ { t } ^ { i } \\leq \\frac { 2 H } { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 104, + 235, + 125 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 107, + 309, + 122 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { t } ( \\alpha _ { t } ^ { i } ) ^ { 2 } \\le \\frac { 2 H } { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 104, + 348, + 125 + ], + "score": 1.0, + "content": "for every", + "type": "text" + }, + { + "bbox": [ + 348, + 109, + 372, + 121 + ], + "score": 0.68, + "content": "t \\geq 1", + "type": "inline_equation" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 129, + 128, + 294, + 142 + ], + "lines": [ + { + "bbox": [ + 127, + 126, + 292, + 145 + ], + "spans": [ + { + "bbox": [ + 127, + 126, + 142, + 145 + ], + "score": 1.0, + "content": "3.", + "type": "text" + }, + { + "bbox": [ + 142, + 129, + 228, + 142 + ], + "score": 0.9, + "content": "\\textstyle \\sum _ { t = i } ^ { \\infty } \\alpha _ { t } ^ { i } = 1 + 1 / H", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 126, + 269, + 145 + ], + "score": 1.0, + "content": ", for every", + "type": "text" + }, + { + "bbox": [ + 269, + 130, + 292, + 141 + ], + "score": 0.83, + "content": "i \\geq 1", + "type": "inline_equation" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 181, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 182, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 182, + 192 + ], + "score": 1.0, + "content": "Proof. Recall that", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 195, + 432, + 230 + ], + "lines": [ + { + "bbox": [ + 179, + 195, + 432, + 230 + ], + "spans": [ + { + "bbox": [ + 179, + 195, + 432, + 230 + ], + "score": 0.94, + "content": "\\alpha _ { t } = \\frac { H + 1 } { H + t } , \\quad \\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) .", + "type": "interline_equation", + "image_path": "242c5979af7b492fba1d80135236539576c26cc515f271d589d39be0ce0b496a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 179, + 195, + 432, + 206.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 179, + 206.66666666666666, + 432, + 218.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 179, + 218.33333333333331, + 432, + 229.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 420, + 245 + ], + "lines": [ + { + "bbox": [ + 104, + 230, + 421, + 249 + ], + "spans": [ + { + "bbox": [ + 104, + 230, + 421, + 249 + ], + "score": 1.0, + "content": "Properties 1-3 are proven by Jin et al. (2018). Now we prove the last property.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 176, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 177, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 177, + 263 + ], + "score": 1.0, + "content": "On the one hand,", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 259, + 390, + 293 + ], + "lines": [ + { + "bbox": [ + 219, + 259, + 390, + 293 + ], + "spans": [ + { + "bbox": [ + 219, + 259, + 390, + 293 + ], + "score": 0.95, + "content": "\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( t ) } { i } } \\leq 2 \\sqrt { \\frac { \\iota ( t ) } { t } } ,", + "type": "interline_equation", + "image_path": "07ae804a6b110033f8bf12b32de163cad5567ec4167d97952b977363794b7bc4.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 259, + 390, + 276.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 219, + 276.0, + 390, + 293.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 306, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 292, + 307, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 307, + 308 + ], + "score": 1.0, + "content": "where the last inequality follows from property 1.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 282, + 324 + ], + "score": 1.0, + "content": "The left-hand side is proven by induction on", + "type": "text" + }, + { + "bbox": [ + 283, + 312, + 288, + 321 + ], + "score": 0.7, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 308, + 388, + 324 + ], + "score": 1.0, + "content": ". For the base case, when", + "type": "text" + }, + { + "bbox": [ + 388, + 311, + 444, + 323 + ], + "score": 0.91, + "content": "t = 1 , \\alpha _ { t } ^ { t } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 308, + 464, + 324 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 464, + 311, + 487, + 322 + ], + "score": 0.88, + "content": "t \\geq 2", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 308, + 505, + 324 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 320, + 342, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 128, + 335 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 128, + 321, + 207, + 335 + ], + "score": 0.93, + "content": "\\alpha _ { t } ^ { i } = ( 1 - \\alpha _ { t } ) \\dot { \\alpha _ { t - 1 } ^ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 320, + 222, + 335 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 222, + 322, + 279, + 333 + ], + "score": 0.9, + "content": "1 \\leq i \\leq t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 320, + 342, + 335 + ], + "score": 1.0, + "content": ". It follows that", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "interline_equation", + "bbox": [ + 122, + 337, + 487, + 372 + ], + "lines": [ + { + "bbox": [ + 122, + 337, + 487, + 372 + ], + "spans": [ + { + "bbox": [ + 122, + 337, + 487, + 372 + ], + "score": 0.93, + "content": "\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } = \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sum _ { i = 1 } ^ { t - 1 } \\alpha _ { t - 1 } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } .", + "type": "interline_equation", + "image_path": "ab4e05f21481995ad7b6907c39b07eca8dcf1b1707af7599ec0fb0ba9984a116.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 122, + 337, + 487, + 348.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 122, + 348.6666666666667, + 487, + 360.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 122, + 360.33333333333337, + 487, + 372.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 441, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 441, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 167, + 389 + ], + "score": 1.0, + "content": "Since function", + "type": "text" + }, + { + "bbox": [ + 167, + 375, + 222, + 388 + ], + "score": 0.94, + "content": "f ( t ) = \\iota ( t ) / t", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 375, + 352, + 389 + ], + "score": 1.0, + "content": "is monotonically decreasing for", + "type": "text" + }, + { + "bbox": [ + 352, + 376, + 402, + 387 + ], + "score": 0.92, + "content": "t \\geq 1 , c \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 375, + 441, + 389 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 390, + 455, + 419 + ], + "lines": [ + { + "bbox": [ + 155, + 390, + 455, + 419 + ], + "spans": [ + { + "bbox": [ + 155, + 390, + 455, + 419 + ], + "score": 0.93, + "content": "\\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t ) } { t } } \\geq \\sqrt { \\frac { \\iota ( t ) } { t } } .", + "type": "interline_equation", + "image_path": "7a7308ee3372dff7c4a6c3cae36090077c647b119d15a408f894359ff305977f.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 155, + 390, + 455, + 419 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 438, + 405, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 406, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 253, + 454 + ], + "score": 1.0, + "content": "Lemma 4. With probability at least", + "type": "text" + }, + { + "bbox": [ + 253, + 439, + 286, + 452 + ], + "score": 0.91, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 438, + 317, + 454 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 317, + 440, + 342, + 451 + ], + "score": 0.92, + "content": "p \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 438, + 360, + 454 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 361, + 440, + 383, + 452 + ], + "score": 0.91, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 438, + 406, + 454 + ], + "score": 1.0, + "content": "-pair,", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 170, + 455, + 439, + 506 + ], + "lines": [ + { + "bbox": [ + 170, + 455, + 439, + 506 + ], + "spans": [ + { + "bbox": [ + 170, + 455, + 439, + 506 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { 0 \\le ( Q _ { p } - Q ^ { * } ) ( s , a ) \\le \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } _ { t _ { i } } - V ^ { * } ) ( s _ { t _ { i } + 1 } ) + \\beta _ { t } , } \\\\ & { 0 \\le ( { \\hat { Q } } _ { p } - Q ^ { * } ) ( s , a ) , } \\end{array}", + "type": "interline_equation", + "image_path": "bf1a7588fbf4d42939d18d2d26129510541081c72f7d186c33706ba60d0c68e1.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 170, + 455, + 439, + 472.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 170, + 472.0, + 439, + 489.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 170, + 489.0, + 439, + 506.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 386, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 387, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 133, + 527 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 511, + 246, + 524 + ], + "score": 0.89, + "content": "t = N _ { p } ( s , a ) , t _ { i } = \\tau ( s , a , i )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 507, + 264, + 527 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 265, + 509, + 383, + 524 + ], + "score": 0.89, + "content": "\\beta _ { t } = c _ { 3 } \\sqrt { H \\iota ( t ) / ( ( 1 - \\gamma ) ^ { 2 } t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 507, + 387, + 527 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 181, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 182, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 182, + 547 + ], + "score": 1.0, + "content": "Proof. Recall that", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 544, + 398, + 579 + ], + "lines": [ + { + "bbox": [ + 212, + 544, + 398, + 579 + ], + "spans": [ + { + "bbox": [ + 212, + 544, + 398, + 579 + ], + "score": 0.93, + "content": "\\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) .", + "type": "interline_equation", + "image_path": "88278a9f30fd7a368323e468f5e3ad316e4a238d906d20218881efa56614124b.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 544, + 398, + 561.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 212, + 561.5, + 398, + 579.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 581, + 457, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 458, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 423, + 595 + ], + "score": 1.0, + "content": "From the update rule, it can be seen that our algorithm maintains the following", + "type": "text" + }, + { + "bbox": [ + 423, + 581, + 453, + 593 + ], + "score": 0.92, + "content": "Q ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 580, + 458, + 595 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 596, + 429, + 630 + ], + "lines": [ + { + "bbox": [ + 182, + 596, + 429, + 630 + ], + "spans": [ + { + "bbox": [ + 182, + 596, + 429, + 630 + ], + "score": 0.94, + "content": "Q _ { p } ( s , a ) = \\alpha _ { t } ^ { 0 } \\frac { 1 } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + b _ { i } + \\gamma \\hat { V } _ { t _ { i } } ( s _ { t _ { i } + 1 } ) \\right] .", + "type": "interline_equation", + "image_path": "1b23fc9ed7edc8e4d294bbfadc8232c0e3a3f7ae514e3b0f9f7eb6f05a9180e4.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 182, + 596, + 429, + 607.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 182, + 607.3333333333334, + 429, + 618.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 182, + 618.6666666666667, + 429, + 630.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 248, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 249, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 249, + 647 + ], + "score": 1.0, + "content": "Bellman optimality equation gives:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 648, + 470, + 682 + ], + "lines": [ + { + "bbox": [ + 140, + 648, + 470, + 682 + ], + "spans": [ + { + "bbox": [ + 140, + 648, + 470, + 682 + ], + "score": 0.93, + "content": "Q ^ { * } ( s , a ) = r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) = \\alpha _ { t } ^ { 0 } Q ^ { * } ( s , a ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) \\right] .", + "type": "interline_equation", + "image_path": "7f3af02c43fff8510d8c8a82457783c482e400b5602396a1f237a333d06d0211.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 140, + 648, + 470, + 659.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 140, + 659.3333333333334, + 470, + 670.6666666666667 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 140, + 670.6666666666667, + 470, + 682.0000000000001 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 685, + 250, + 697 + ], + "lines": [ + { + "bbox": [ + 106, + 683, + 250, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 683, + 250, + 699 + ], + "score": 1.0, + "content": "Subtracting the two equations gives", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 700, + 514, + 734 + ], + "lines": [ + { + "bbox": [ + 111, + 700, + 514, + 734 + ], + "spans": [ + { + "bbox": [ + 111, + 700, + 514, + 734 + ], + "score": 0.9, + "content": "Q _ { p } - Q ^ { * } ) ( s , a ) = \\alpha _ { t } ^ { 0 } ( \\frac { 1 } { 1 - \\gamma } - Q ^ { * } ( s , a ) ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ b _ { i } + \\gamma \\left( V _ { t _ { i } } - V ^ { * } \\right) \\left( s _ { t _ { i } + 1 } \\right) + \\gamma \\left( V ^ { * } \\left( s _ { t _ { i } + 1 } \\right) - \\mathbb { P } V ^ { * } \\right) \\right] ,", + "type": "interline_equation", + "image_path": "4daae5a78519cf382ebb3c39463504ef4bdc6c481b0042a0fad9b5891adc01b0.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 111, + 700, + 514, + 711.3333333333334 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 111, + 711.3333333333334, + 514, + 722.6666666666667 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 111, + 722.6666666666667, + 514, + 734.0000000000001 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 422, + 505, + 432 + ], + "lines": [ + { + "bbox": [ + 495, + 423, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 495, + 423, + 505, + 433 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 761 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 129, + 79, + 353, + 101 + ], + "lines": [ + { + "bbox": [ + 127, + 78, + 356, + 102 + ], + "spans": [ + { + "bbox": [ + 127, + 78, + 142, + 102 + ], + "score": 1.0, + "content": "1.", + "type": "text" + }, + { + "bbox": [ + 142, + 80, + 259, + 101 + ], + "score": 0.89, + "content": "\\begin{array} { r } { \\sqrt { \\frac { 1 } { t } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { 1 } { i } } \\leq 2 \\sqrt { \\frac { 1 } { t } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 78, + 298, + 102 + ], + "score": 1.0, + "content": "for every", + "type": "text" + }, + { + "bbox": [ + 299, + 84, + 349, + 97 + ], + "score": 0.58, + "content": "t \\geq 1 , c > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 78, + 356, + 102 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 127, + 78, + 356, + 102 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 106, + 374, + 123 + ], + "lines": [ + { + "bbox": [ + 126, + 104, + 372, + 125 + ], + "spans": [ + { + "bbox": [ + 126, + 104, + 142, + 125 + ], + "score": 1.0, + "content": "2.", + "type": "text" + }, + { + "bbox": [ + 142, + 108, + 216, + 122 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\operatorname* { m a x } _ { i \\in [ t ] } \\alpha _ { t } ^ { i } \\leq \\frac { 2 H } { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 104, + 235, + 125 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 107, + 309, + 122 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { t } ( \\alpha _ { t } ^ { i } ) ^ { 2 } \\le \\frac { 2 H } { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 104, + 348, + 125 + ], + "score": 1.0, + "content": "for every", + "type": "text" + }, + { + "bbox": [ + 348, + 109, + 372, + 121 + ], + "score": 0.68, + "content": "t \\geq 1", + "type": "inline_equation" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 126, + 104, + 372, + 125 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 128, + 294, + 142 + ], + "lines": [ + { + "bbox": [ + 127, + 126, + 292, + 145 + ], + "spans": [ + { + "bbox": [ + 127, + 126, + 142, + 145 + ], + "score": 1.0, + "content": "3.", + "type": "text" + }, + { + "bbox": [ + 142, + 129, + 228, + 142 + ], + "score": 0.9, + "content": "\\textstyle \\sum _ { t = i } ^ { \\infty } \\alpha _ { t } ^ { i } = 1 + 1 / H", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 126, + 269, + 145 + ], + "score": 1.0, + "content": ", for every", + "type": "text" + }, + { + "bbox": [ + 269, + 130, + 292, + 141 + ], + "score": 0.83, + "content": "i \\geq 1", + "type": "inline_equation" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 127, + 126, + 292, + 145 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 180, + 181, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 179, + 182, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 182, + 192 + ], + "score": 1.0, + "content": "Proof. Recall that", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 179, + 182, + 192 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 195, + 432, + 230 + ], + "lines": [ + { + "bbox": [ + 179, + 195, + 432, + 230 + ], + "spans": [ + { + "bbox": [ + 179, + 195, + 432, + 230 + ], + "score": 0.94, + "content": "\\alpha _ { t } = \\frac { H + 1 } { H + t } , \\quad \\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) .", + "type": "interline_equation", + "image_path": "242c5979af7b492fba1d80135236539576c26cc515f271d589d39be0ce0b496a.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 179, + 195, + 432, + 206.66666666666666 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 179, + 206.66666666666666, + 432, + 218.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 179, + 218.33333333333331, + 432, + 229.99999999999997 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 420, + 245 + ], + "lines": [ + { + "bbox": [ + 104, + 230, + 421, + 249 + ], + "spans": [ + { + "bbox": [ + 104, + 230, + 421, + 249 + ], + "score": 1.0, + "content": "Properties 1-3 are proven by Jin et al. (2018). Now we prove the last property.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 104, + 230, + 421, + 249 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 250, + 176, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 177, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 177, + 263 + ], + "score": 1.0, + "content": "On the one hand,", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 249, + 177, + 263 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 259, + 390, + 293 + ], + "lines": [ + { + "bbox": [ + 219, + 259, + 390, + 293 + ], + "spans": [ + { + "bbox": [ + 219, + 259, + 390, + 293 + ], + "score": 0.95, + "content": "\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( t ) } { i } } \\leq 2 \\sqrt { \\frac { \\iota ( t ) } { t } } ,", + "type": "interline_equation", + "image_path": "07ae804a6b110033f8bf12b32de163cad5567ec4167d97952b977363794b7bc4.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 259, + 390, + 276.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 219, + 276.0, + 390, + 293.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 294, + 306, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 292, + 307, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 307, + 308 + ], + "score": 1.0, + "content": "where the last inequality follows from property 1.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 292, + 307, + 308 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 282, + 324 + ], + "score": 1.0, + "content": "The left-hand side is proven by induction on", + "type": "text" + }, + { + "bbox": [ + 283, + 312, + 288, + 321 + ], + "score": 0.7, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 308, + 388, + 324 + ], + "score": 1.0, + "content": ". For the base case, when", + "type": "text" + }, + { + "bbox": [ + 388, + 311, + 444, + 323 + ], + "score": 0.91, + "content": "t = 1 , \\alpha _ { t } ^ { t } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 308, + 464, + 324 + ], + "score": 1.0, + "content": ". For", + "type": "text" + }, + { + "bbox": [ + 464, + 311, + 487, + 322 + ], + "score": 0.88, + "content": "t \\geq 2", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 308, + 505, + 324 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 320, + 342, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 128, + 335 + ], + "score": 1.0, + "content": "have", + "type": "text" + }, + { + "bbox": [ + 128, + 321, + 207, + 335 + ], + "score": 0.93, + "content": "\\alpha _ { t } ^ { i } = ( 1 - \\alpha _ { t } ) \\dot { \\alpha _ { t - 1 } ^ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 320, + 222, + 335 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 222, + 322, + 279, + 333 + ], + "score": 0.9, + "content": "1 \\leq i \\leq t - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 320, + 342, + 335 + ], + "score": 1.0, + "content": ". It follows that", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 308, + 505, + 335 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 122, + 337, + 487, + 372 + ], + "lines": [ + { + "bbox": [ + 122, + 337, + 487, + 372 + ], + "spans": [ + { + "bbox": [ + 122, + 337, + 487, + 372 + ], + "score": 0.93, + "content": "\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } = \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sum _ { i = 1 } ^ { t - 1 } \\alpha _ { t - 1 } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } .", + "type": "interline_equation", + "image_path": "ab4e05f21481995ad7b6907c39b07eca8dcf1b1707af7599ec0fb0ba9984a116.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 122, + 337, + 487, + 348.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 122, + 348.6666666666667, + 487, + 360.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 122, + 360.33333333333337, + 487, + 372.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 441, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 375, + 441, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 167, + 389 + ], + "score": 1.0, + "content": "Since function", + "type": "text" + }, + { + "bbox": [ + 167, + 375, + 222, + 388 + ], + "score": 0.94, + "content": "f ( t ) = \\iota ( t ) / t", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 375, + 352, + 389 + ], + "score": 1.0, + "content": "is monotonically decreasing for", + "type": "text" + }, + { + "bbox": [ + 352, + 376, + 402, + 387 + ], + "score": 0.92, + "content": "t \\geq 1 , c \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 375, + 441, + 389 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 375, + 441, + 389 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 390, + 455, + 419 + ], + "lines": [ + { + "bbox": [ + 155, + 390, + 455, + 419 + ], + "spans": [ + { + "bbox": [ + 155, + 390, + 455, + 419 + ], + "score": 0.93, + "content": "\\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t ) } { t } } \\geq \\sqrt { \\frac { \\iota ( t ) } { t } } .", + "type": "interline_equation", + "image_path": "7a7308ee3372dff7c4a6c3cae36090077c647b119d15a408f894359ff305977f.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 155, + 390, + 455, + 419 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 438, + 405, + 452 + ], + "lines": [ + { + "bbox": [ + 106, + 438, + 406, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 253, + 454 + ], + "score": 1.0, + "content": "Lemma 4. With probability at least", + "type": "text" + }, + { + "bbox": [ + 253, + 439, + 286, + 452 + ], + "score": 0.91, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 438, + 317, + 454 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 317, + 440, + 342, + 451 + ], + "score": 0.92, + "content": "p \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 438, + 360, + 454 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 361, + 440, + 383, + 452 + ], + "score": 0.91, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 438, + 406, + 454 + ], + "score": 1.0, + "content": "-pair,", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 438, + 406, + 454 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 170, + 455, + 439, + 506 + ], + "lines": [ + { + "bbox": [ + 170, + 455, + 439, + 506 + ], + "spans": [ + { + "bbox": [ + 170, + 455, + 439, + 506 + ], + "score": 0.91, + "content": "\\begin{array} { r l } & { 0 \\le ( Q _ { p } - Q ^ { * } ) ( s , a ) \\le \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } _ { t _ { i } } - V ^ { * } ) ( s _ { t _ { i } + 1 } ) + \\beta _ { t } , } \\\\ & { 0 \\le ( { \\hat { Q } } _ { p } - Q ^ { * } ) ( s , a ) , } \\end{array}", + "type": "interline_equation", + "image_path": "bf1a7588fbf4d42939d18d2d26129510541081c72f7d186c33706ba60d0c68e1.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 170, + 455, + 439, + 472.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 170, + 472.0, + 439, + 489.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 170, + 489.0, + 439, + 506.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 386, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 387, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 133, + 527 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 511, + 246, + 524 + ], + "score": 0.89, + "content": "t = N _ { p } ( s , a ) , t _ { i } = \\tau ( s , a , i )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 507, + 264, + 527 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 265, + 509, + 383, + 524 + ], + "score": 0.89, + "content": "\\beta _ { t } = c _ { 3 } \\sqrt { H \\iota ( t ) / ( ( 1 - \\gamma ) ^ { 2 } t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 507, + 387, + 527 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 507, + 387, + 527 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 181, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 182, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 182, + 547 + ], + "score": 1.0, + "content": "Proof. Recall that", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 533, + 182, + 547 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 212, + 544, + 398, + 579 + ], + "lines": [ + { + "bbox": [ + 212, + 544, + 398, + 579 + ], + "spans": [ + { + "bbox": [ + 212, + 544, + 398, + 579 + ], + "score": 0.93, + "content": "\\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) .", + "type": "interline_equation", + "image_path": "88278a9f30fd7a368323e468f5e3ad316e4a238d906d20218881efa56614124b.jpg" + } + ] + } + ], + "index": 25.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 544, + 398, + 561.5 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 212, + 561.5, + 398, + 579.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 581, + 457, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 458, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 423, + 595 + ], + "score": 1.0, + "content": "From the update rule, it can be seen that our algorithm maintains the following", + "type": "text" + }, + { + "bbox": [ + 423, + 581, + 453, + 593 + ], + "score": 0.92, + "content": "Q ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 580, + 458, + 595 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 580, + 458, + 595 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 596, + 429, + 630 + ], + "lines": [ + { + "bbox": [ + 182, + 596, + 429, + 630 + ], + "spans": [ + { + "bbox": [ + 182, + 596, + 429, + 630 + ], + "score": 0.94, + "content": "Q _ { p } ( s , a ) = \\alpha _ { t } ^ { 0 } \\frac { 1 } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + b _ { i } + \\gamma \\hat { V } _ { t _ { i } } ( s _ { t _ { i } + 1 } ) \\right] .", + "type": "interline_equation", + "image_path": "1b23fc9ed7edc8e4d294bbfadc8232c0e3a3f7ae514e3b0f9f7eb6f05a9180e4.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 182, + 596, + 429, + 607.3333333333334 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 182, + 607.3333333333334, + 429, + 618.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 182, + 618.6666666666667, + 429, + 630.0000000000001 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 248, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 249, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 249, + 647 + ], + "score": 1.0, + "content": "Bellman optimality equation gives:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 632, + 249, + 647 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 140, + 648, + 470, + 682 + ], + "lines": [ + { + "bbox": [ + 140, + 648, + 470, + 682 + ], + "spans": [ + { + "bbox": [ + 140, + 648, + 470, + 682 + ], + "score": 0.93, + "content": "Q ^ { * } ( s , a ) = r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) = \\alpha _ { t } ^ { 0 } Q ^ { * } ( s , a ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) \\right] .", + "type": "interline_equation", + "image_path": "7f3af02c43fff8510d8c8a82457783c482e400b5602396a1f237a333d06d0211.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 140, + 648, + 470, + 659.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 140, + 659.3333333333334, + 470, + 670.6666666666667 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 140, + 670.6666666666667, + 470, + 682.0000000000001 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 685, + 250, + 697 + ], + "lines": [ + { + "bbox": [ + 106, + 683, + 250, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 683, + 250, + 699 + ], + "score": 1.0, + "content": "Subtracting the two equations gives", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 683, + 250, + 699 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 700, + 514, + 734 + ], + "lines": [ + { + "bbox": [ + 111, + 700, + 514, + 734 + ], + "spans": [ + { + "bbox": [ + 111, + 700, + 514, + 734 + ], + "score": 0.9, + "content": "Q _ { p } - Q ^ { * } ) ( s , a ) = \\alpha _ { t } ^ { 0 } ( \\frac { 1 } { 1 - \\gamma } - Q ^ { * } ( s , a ) ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ b _ { i } + \\gamma \\left( V _ { t _ { i } } - V ^ { * } \\right) \\left( s _ { t _ { i } + 1 } \\right) + \\gamma \\left( V ^ { * } \\left( s _ { t _ { i } + 1 } \\right) - \\mathbb { P } V ^ { * } \\right) \\right] ,", + "type": "interline_equation", + "image_path": "4daae5a78519cf382ebb3c39463504ef4bdc6c481b0042a0fad9b5891adc01b0.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 111, + 700, + 514, + 711.3333333333334 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 111, + 711.3333333333334, + 514, + 722.6666666666667 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 111, + 722.6666666666667, + 514, + 734.0000000000001 + ], + "spans": [], + "index": 38 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 254, + 96 + ], + "score": 1.0, + "content": "The identity above holds for arbitrary", + "type": "text" + }, + { + "bbox": [ + 254, + 85, + 271, + 93 + ], + "score": 0.31, + "content": "p , s", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 81, + 288, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 85, + 294, + 92 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 81, + 332, + 96 + ], + "score": 1.0, + "content": ". Now fix", + "type": "text" + }, + { + "bbox": [ + 333, + 83, + 357, + 93 + ], + "score": 0.89, + "content": "s \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 81, + 360, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 361, + 83, + 387, + 93 + ], + "score": 0.88, + "content": "a \\in A", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 81, + 404, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 83, + 430, + 94 + ], + "score": 0.9, + "content": "p \\in \\mathbb N", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 81, + 450, + 96 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 450, + 83, + 502, + 95 + ], + "score": 0.93, + "content": "t = N _ { p } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 163, + 106 + ], + "score": 0.93, + "content": "t _ { i } = \\tau ( s , a , i )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 93, + 185, + 106 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 186, + 94, + 209, + 104 + ], + "score": 0.9, + "content": "t = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 93, + 313, + 106 + ], + "score": 1.0, + "content": "case is trivial; we assume", + "type": "text" + }, + { + "bbox": [ + 314, + 94, + 337, + 105 + ], + "score": 0.9, + "content": "t \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 93, + 496, + 106 + ], + "score": 1.0, + "content": "below. Now consider an arbitrary fixed", + "type": "text" + }, + { + "bbox": [ + 496, + 95, + 503, + 104 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 93, + 507, + 106 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 137, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 137, + 117 + ], + "score": 1.0, + "content": "Define", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 115, + 404, + 137 + ], + "lines": [ + { + "bbox": [ + 206, + 115, + 404, + 137 + ], + "spans": [ + { + "bbox": [ + 206, + 115, + 404, + 137 + ], + "score": 0.92, + "content": "\\Delta _ { i } = \\left( \\alpha _ { k } ^ { i } \\cdot I [ t _ { i } < \\infty ] \\cdot \\left( \\mathbb { P } V ^ { * } - \\mathbb { \\hat { P } } _ { t _ { i } } V ^ { * } \\right) ( s , a ) \\right)", + "type": "interline_equation", + "image_path": "7ed6137165be41b2ad3ac59dec10e06c5f86c9778d1953d9a7904ac5097d5be4.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 206, + 115, + 404, + 137 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 138, + 507, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 137, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 124, + 152 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 124, + 139, + 135, + 149 + ], + "score": 0.86, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 137, + 168, + 152 + ], + "score": 1.0, + "content": "be the", + "type": "text" + }, + { + "bbox": [ + 168, + 140, + 176, + 149 + ], + "score": 0.81, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 137, + 336, + 152 + ], + "score": 1.0, + "content": "-Field generated by random variables", + "type": "text" + }, + { + "bbox": [ + 337, + 139, + 413, + 150 + ], + "score": 0.88, + "content": "( s _ { 1 } , a _ { 1 } , . . . , s _ { t _ { i } } , a _ { t _ { i } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 137, + 506, + 152 + ], + "score": 1.0, + "content": ". It can be seen that", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 507, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 166, + 162 + ], + "score": 0.92, + "content": "\\mathbb { E } \\left[ \\Delta _ { i } | F _ { i } \\right] = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 144, + 196, + 167 + ], + "score": 1.0, + "content": ", while", + "type": "text" + }, + { + "bbox": [ + 197, + 150, + 209, + 160 + ], + "score": 0.89, + "content": "\\Delta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 144, + 281, + 167 + ], + "score": 1.0, + "content": "is measurable in", + "type": "text" + }, + { + "bbox": [ + 281, + 150, + 302, + 162 + ], + "score": 0.92, + "content": "F _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 144, + 358, + 167 + ], + "score": 1.0, + "content": ". Also, since", + "type": "text" + }, + { + "bbox": [ + 358, + 150, + 447, + 163 + ], + "score": 0.85, + "content": "\\textstyle 0 \\leq V ^ { \\ast } ( s , { \\bar { a } } ) \\leq { \\frac { 1 } { 1 - \\gamma } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 144, + 452, + 167 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 452, + 149, + 502, + 164 + ], + "score": 0.9, + "content": "\\begin{array} { r } { | \\Delta _ { i } | \\le \\frac { 2 } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 144, + 507, + 167 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 150, + 175 + ], + "score": 1.0, + "content": "Therefore,", + "type": "text" + }, + { + "bbox": [ + 151, + 163, + 163, + 173 + ], + "score": 0.89, + "content": "\\Delta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 160, + 457, + 175 + ], + "score": 1.0, + "content": "is a martingale difference sequence; by the Azuma-Hoeffding inequality,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 175, + 428, + 210 + ], + "lines": [ + { + "bbox": [ + 183, + 175, + 428, + 210 + ], + "spans": [ + { + "bbox": [ + 183, + 175, + 428, + 210 + ], + "score": 0.93, + "content": "\\mathrm { P r } \\left[ \\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| > \\eta \\right] \\leq 2 \\exp \\left\\{ - \\frac { \\eta ^ { 2 } } { 8 \\left( 1 - \\gamma \\right) ^ { - 2 } \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } } \\right\\} .", + "type": "interline_equation", + "image_path": "e1fd0592f2fbfb43a68fa1b2348697afc8a3b51894eb01a3c12dfc2769dfd098.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 183, + 175, + 428, + 186.66666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 183, + 186.66666666666666, + 428, + 198.33333333333331 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 183, + 198.33333333333331, + 428, + 209.99999999999997 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 421, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 421, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 159, + 225 + ], + "score": 1.0, + "content": "By choosing", + "type": "text" + }, + { + "bbox": [ + 159, + 214, + 165, + 223 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 210, + 307, + 225 + ], + "score": 1.0, + "content": ", we can show that with probability", + "type": "text" + }, + { + "bbox": [ + 308, + 212, + 416, + 223 + ], + "score": 0.9, + "content": "1 - \\delta / \\left[ S A ( k + 1 ) ( k + 2 ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 210, + 421, + 225 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 241, + 461, + 279 + ], + "lines": [ + { + "bbox": [ + 148, + 241, + 461, + 279 + ], + "spans": [ + { + "bbox": [ + 148, + 241, + 461, + 279 + ], + "score": 0.93, + "content": "\\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| \\leq \\frac { 2 \\sqrt { 2 } } { 1 - \\gamma } \\cdot \\sqrt { \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } \\cdot \\ln \\frac { 2 ( k + 1 ) ( k + 2 ) S A } { \\delta } } \\leq \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } .", + "type": "interline_equation", + "image_path": "4635fc976e34fd23da44f8967bc69d839d0dafaecc85d492653e8f8184e7782e.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 148, + 241, + 461, + 253.66666666666666 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 148, + 253.66666666666666, + 461, + 266.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 148, + 266.3333333333333, + 461, + 279.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 281, + 506, + 306 + ], + "lines": [ + { + "bbox": [ + 104, + 276, + 502, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 276, + 128, + 299 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 128, + 282, + 170, + 294 + ], + "score": 0.89, + "content": "c _ { 2 } = 4 \\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 276, + 174, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 174, + 281, + 270, + 296 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\iota ( k ) = \\ln { \\frac { ( k + 1 ) ( k + 2 ) S A } { \\delta } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 276, + 375, + 299 + ], + "score": 1.0, + "content": ". By a union bound for all", + "type": "text" + }, + { + "bbox": [ + 375, + 284, + 381, + 293 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 276, + 478, + 299 + ], + "score": 1.0, + "content": ", this holds for arbitrary", + "type": "text" + }, + { + "bbox": [ + 478, + 284, + 502, + 294 + ], + "score": 0.89, + "content": "k > 0", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 293, + 328, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 143, + 307 + ], + "score": 1.0, + "content": "arbitrary", + "type": "text" + }, + { + "bbox": [ + 143, + 295, + 198, + 305 + ], + "score": 0.33, + "content": "s \\in S , a \\in A", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 293, + 328, + 307 + ], + "score": 1.0, + "content": "simultaneously with probability", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 307, + 404, + 341 + ], + "lines": [ + { + "bbox": [ + 205, + 307, + 404, + 341 + ], + "spans": [ + { + "bbox": [ + 205, + 307, + 404, + 341 + ], + "score": 0.93, + "content": "1 - \\sum _ { s ^ { \\prime } \\in S , a ^ { \\prime } \\in A } \\sum _ { k = 1 } ^ { \\infty } \\frac { \\delta } { 2 S A ( k + 1 ) ( k + 2 ) } = 1 - \\frac { \\delta } { 2 } .", + "type": "interline_equation", + "image_path": "d36f303c3c54ca211346b95b702c878d2e496acc0473679ca8912f1c49bc42a5.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 205, + 307, + 404, + 324.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 205, + 324.0, + 404, + 341.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 343, + 504, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 341, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 371, + 356 + ], + "score": 1.0, + "content": "Therefore, we conclude that (16) holds for the random variable", + "type": "text" + }, + { + "bbox": [ + 371, + 343, + 425, + 356 + ], + "score": 0.93, + "content": "t = N _ { p } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 341, + 473, + 356 + ], + "score": 1.0, + "content": "and for all", + "type": "text" + }, + { + "bbox": [ + 474, + 346, + 480, + 355 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 341, + 505, + 356 + ], + "score": 1.0, + "content": ", with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 354, + 220, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 153, + 366 + ], + "score": 1.0, + "content": "probability", + "type": "text" + }, + { + "bbox": [ + 153, + 354, + 186, + 366 + ], + "score": 0.91, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 354, + 220, + 366 + ], + "score": 1.0, + "content": "as well.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 418, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 416, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 336, + 389 + ], + "score": 1.0, + "content": "Proof of the right hand side of (13): We also know that", + "type": "text" + }, + { + "bbox": [ + 337, + 371, + 416, + 392 + ], + "score": 0.68, + "content": "\\begin{array} { r } { ( b _ { k } = \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 393, + 401, + 427 + ], + "lines": [ + { + "bbox": [ + 208, + 393, + 401, + 427 + ], + "spans": [ + { + "bbox": [ + 208, + 393, + 401, + 427 + ], + "score": 0.86, + "content": "\\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } \\leq \\sum _ { i = 1 } ^ { k } \\alpha _ { k } ^ { i } b _ { i } \\leq \\frac { 2 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } .", + "type": "interline_equation", + "image_path": "881fb37a93a36ca33279a8cb7352a55121d9d63aa62473943680b180547b1336.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 393, + 401, + 410.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 208, + 410.0, + 401, + 427.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 428, + 206, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 207, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 207, + 442 + ], + "score": 1.0, + "content": "It is implied by (16) that", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 442, + 462, + 510 + ], + "lines": [ + { + "bbox": [ + 149, + 442, + 462, + 510 + ], + "spans": [ + { + "bbox": [ + 149, + 442, + 462, + 510 + ], + "score": 0.88, + "content": "\\begin{array} { r l } & { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\gamma \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\frac { 3 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( t ) } { t } } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( \\hat { V } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "d2bbc005947d28dd95e009c07818fe62e6d9b1e5309a47331c5537b6fbe06992.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 149, + 442, + 462, + 464.6666666666667 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 149, + 464.6666666666667, + 462, + 487.33333333333337 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 149, + 487.33333333333337, + 462, + 510.00000000000006 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 523, + 403, + 556 + ], + "lines": [ + { + "bbox": [ + 218, + 523, + 403, + 556 + ], + "spans": [ + { + "bbox": [ + 218, + 523, + 403, + 556 + ], + "score": 0.38, + "content": "\\leq \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + \\beta _ { t } .", + "type": "interline_equation", + "image_path": "53eab3d1f26c37eff9f588636d51ccb07f492de053eae442ba9c0192a7489963.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 523, + 403, + 539.5 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 218, + 539.5, + 403, + 556.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 558, + 343, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 343, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 146, + 574 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 146, + 559, + 262, + 573 + ], + "score": 0.85, + "content": "\\beta _ { t } = c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { H \\iota ( t ) / t }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 557, + 266, + 574 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 266, + 558, + 341, + 572 + ], + "score": 0.73, + "content": "c _ { 3 } = 3 c _ { 2 } = 1 2 { \\sqrt { 2 } } .", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 557, + 343, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "Proof of the left hand side of (13): Now, we assume that event that (16) holds. We assert that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 587, + 453, + 601 + ], + "spans": [ + { + "bbox": [ + 107, + 588, + 146, + 600 + ], + "score": 0.92, + "content": "Q _ { p } \\geq Q ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 587, + 174, + 601 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 174, + 588, + 197, + 600 + ], + "score": 0.92, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 587, + 215, + 601 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 215, + 588, + 242, + 600 + ], + "score": 0.92, + "content": "p \\leq p ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 587, + 398, + 601 + ], + "score": 1.0, + "content": ". This assertion is obviously true when", + "type": "text" + }, + { + "bbox": [ + 398, + 588, + 425, + 599 + ], + "score": 0.93, + "content": "p ^ { \\prime } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 587, + 453, + 601 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 163, + 601, + 447, + 671 + ], + "lines": [ + { + "bbox": [ + 163, + 601, + 447, + 671 + ], + "spans": [ + { + "bbox": [ + 163, + 601, + 447, + 671 + ], + "score": 0.94, + "content": "\\begin{array} { c l } { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\geq - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ { \\geq \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } b _ { i } - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| \\geq 0 . } \\end{array}", + "type": "interline_equation", + "image_path": "70ff20e4e31a1bb0299ebb34008753b3eaca5c5e3da766015eda50bc514d5a51.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 163, + 601, + 447, + 624.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 163, + 624.3333333333334, + 447, + 647.6666666666667 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 163, + 647.6666666666667, + 447, + 671.0000000000001 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 671, + 425, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 425, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 238, + 685 + ], + "score": 1.0, + "content": "Therefore the assertion holds for", + "type": "text" + }, + { + "bbox": [ + 239, + 672, + 265, + 683 + ], + "score": 0.91, + "content": "p ^ { \\prime } + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 670, + 415, + 685 + ], + "score": 1.0, + "content": "as well. By induction, it holds for all", + "type": "text" + }, + { + "bbox": [ + 415, + 674, + 421, + 683 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 670, + 425, + 685 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 689, + 505, + 717 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 280, + 703 + ], + "score": 1.0, + "content": "We now see that (13) holds for probability", + "type": "text" + }, + { + "bbox": [ + 281, + 690, + 313, + 702 + ], + "score": 0.91, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 688, + 342, + 703 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 342, + 691, + 368, + 702 + ], + "score": 0.77, + "content": "p , s , a", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 688, + 397, + 703 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 398, + 688, + 433, + 703 + ], + "score": 0.93, + "content": "\\hat { Q } _ { p } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 688, + 505, + 703 + ], + "score": 1.0, + "content": "is always greater", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 702, + 503, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 126, + 718 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 126, + 703, + 164, + 716 + ], + "score": 0.93, + "content": "Q _ { p ^ { \\prime } } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 702, + 203, + 718 + ], + "score": 1.0, + "content": "for some", + "type": "text" + }, + { + "bbox": [ + 204, + 703, + 231, + 715 + ], + "score": 0.92, + "content": "p ^ { \\prime } \\leq p", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 702, + 290, + 718 + ], + "score": 1.0, + "content": ", we know that", + "type": "text" + }, + { + "bbox": [ + 291, + 703, + 425, + 716 + ], + "score": 0.92, + "content": "\\hat { Q } _ { p } ( s , a ) \\geq Q _ { p ^ { \\prime } } ( s , a ) \\geq Q ^ { * } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 702, + 503, + 718 + ], + "score": 1.0, + "content": ", thus proving (14).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 721, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 497, + 723, + 504, + 730 + ], + "spans": [ + { + "bbox": [ + 497, + 723, + 504, + 730 + ], + "score": 0.996, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 506, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 254, + 96 + ], + "score": 1.0, + "content": "The identity above holds for arbitrary", + "type": "text" + }, + { + "bbox": [ + 254, + 85, + 271, + 93 + ], + "score": 0.31, + "content": "p , s", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 81, + 288, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 85, + 294, + 92 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 81, + 332, + 96 + ], + "score": 1.0, + "content": ". Now fix", + "type": "text" + }, + { + "bbox": [ + 333, + 83, + 357, + 93 + ], + "score": 0.89, + "content": "s \\in S", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 81, + 360, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 361, + 83, + 387, + 93 + ], + "score": 0.88, + "content": "a \\in A", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 81, + 404, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 405, + 83, + 430, + 94 + ], + "score": 0.9, + "content": "p \\in \\mathbb N", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 81, + 450, + 96 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 450, + 83, + 502, + 95 + ], + "score": 0.93, + "content": "t = N _ { p } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 507, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 163, + 106 + ], + "score": 0.93, + "content": "t _ { i } = \\tau ( s , a , i )", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 93, + 185, + 106 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 186, + 94, + 209, + 104 + ], + "score": 0.9, + "content": "t = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 93, + 313, + 106 + ], + "score": 1.0, + "content": "case is trivial; we assume", + "type": "text" + }, + { + "bbox": [ + 314, + 94, + 337, + 105 + ], + "score": 0.9, + "content": "t \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 93, + 496, + 106 + ], + "score": 1.0, + "content": "below. Now consider an arbitrary fixed", + "type": "text" + }, + { + "bbox": [ + 496, + 95, + 503, + 104 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 93, + 507, + 106 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 137, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 137, + 117 + ], + "score": 1.0, + "content": "Define", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 507, + 117 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 115, + 404, + 137 + ], + "lines": [ + { + "bbox": [ + 206, + 115, + 404, + 137 + ], + "spans": [ + { + "bbox": [ + 206, + 115, + 404, + 137 + ], + "score": 0.92, + "content": "\\Delta _ { i } = \\left( \\alpha _ { k } ^ { i } \\cdot I [ t _ { i } < \\infty ] \\cdot \\left( \\mathbb { P } V ^ { * } - \\mathbb { \\hat { P } } _ { t _ { i } } V ^ { * } \\right) ( s , a ) \\right)", + "type": "interline_equation", + "image_path": "7ed6137165be41b2ad3ac59dec10e06c5f86c9778d1953d9a7904ac5097d5be4.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 206, + 115, + 404, + 137 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 138, + 507, + 174 + ], + "lines": [ + { + "bbox": [ + 105, + 137, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 124, + 152 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 124, + 139, + 135, + 149 + ], + "score": 0.86, + "content": "F _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 137, + 168, + 152 + ], + "score": 1.0, + "content": "be the", + "type": "text" + }, + { + "bbox": [ + 168, + 140, + 176, + 149 + ], + "score": 0.81, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 137, + 336, + 152 + ], + "score": 1.0, + "content": "-Field generated by random variables", + "type": "text" + }, + { + "bbox": [ + 337, + 139, + 413, + 150 + ], + "score": 0.88, + "content": "( s _ { 1 } , a _ { 1 } , . . . , s _ { t _ { i } } , a _ { t _ { i } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 137, + 506, + 152 + ], + "score": 1.0, + "content": ". It can be seen that", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 507, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 150, + 166, + 162 + ], + "score": 0.92, + "content": "\\mathbb { E } \\left[ \\Delta _ { i } | F _ { i } \\right] = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 144, + 196, + 167 + ], + "score": 1.0, + "content": ", while", + "type": "text" + }, + { + "bbox": [ + 197, + 150, + 209, + 160 + ], + "score": 0.89, + "content": "\\Delta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 144, + 281, + 167 + ], + "score": 1.0, + "content": "is measurable in", + "type": "text" + }, + { + "bbox": [ + 281, + 150, + 302, + 162 + ], + "score": 0.92, + "content": "F _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 144, + 358, + 167 + ], + "score": 1.0, + "content": ". Also, since", + "type": "text" + }, + { + "bbox": [ + 358, + 150, + 447, + 163 + ], + "score": 0.85, + "content": "\\textstyle 0 \\leq V ^ { \\ast } ( s , { \\bar { a } } ) \\leq { \\frac { 1 } { 1 - \\gamma } }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 144, + 452, + 167 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 452, + 149, + 502, + 164 + ], + "score": 0.9, + "content": "\\begin{array} { r } { | \\Delta _ { i } | \\le \\frac { 2 } { 1 - \\gamma } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 144, + 507, + 167 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 150, + 175 + ], + "score": 1.0, + "content": "Therefore,", + "type": "text" + }, + { + "bbox": [ + 151, + 163, + 163, + 173 + ], + "score": 0.89, + "content": "\\Delta _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 160, + 457, + 175 + ], + "score": 1.0, + "content": "is a martingale difference sequence; by the Azuma-Hoeffding inequality,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 137, + 507, + 175 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 175, + 428, + 210 + ], + "lines": [ + { + "bbox": [ + 183, + 175, + 428, + 210 + ], + "spans": [ + { + "bbox": [ + 183, + 175, + 428, + 210 + ], + "score": 0.93, + "content": "\\mathrm { P r } \\left[ \\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| > \\eta \\right] \\leq 2 \\exp \\left\\{ - \\frac { \\eta ^ { 2 } } { 8 \\left( 1 - \\gamma \\right) ^ { - 2 } \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } } \\right\\} .", + "type": "interline_equation", + "image_path": "e1fd0592f2fbfb43a68fa1b2348697afc8a3b51894eb01a3c12dfc2769dfd098.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 183, + 175, + 428, + 186.66666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 183, + 186.66666666666666, + 428, + 198.33333333333331 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 183, + 198.33333333333331, + 428, + 209.99999999999997 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 421, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 421, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 159, + 225 + ], + "score": 1.0, + "content": "By choosing", + "type": "text" + }, + { + "bbox": [ + 159, + 214, + 165, + 223 + ], + "score": 0.76, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 210, + 307, + 225 + ], + "score": 1.0, + "content": ", we can show that with probability", + "type": "text" + }, + { + "bbox": [ + 308, + 212, + 416, + 223 + ], + "score": 0.9, + "content": "1 - \\delta / \\left[ S A ( k + 1 ) ( k + 2 ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 210, + 421, + 225 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 210, + 421, + 225 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 241, + 461, + 279 + ], + "lines": [ + { + "bbox": [ + 148, + 241, + 461, + 279 + ], + "spans": [ + { + "bbox": [ + 148, + 241, + 461, + 279 + ], + "score": 0.93, + "content": "\\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| \\leq \\frac { 2 \\sqrt { 2 } } { 1 - \\gamma } \\cdot \\sqrt { \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } \\cdot \\ln \\frac { 2 ( k + 1 ) ( k + 2 ) S A } { \\delta } } \\leq \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } .", + "type": "interline_equation", + "image_path": "4635fc976e34fd23da44f8967bc69d839d0dafaecc85d492653e8f8184e7782e.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 148, + 241, + 461, + 253.66666666666666 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 148, + 253.66666666666666, + 461, + 266.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 148, + 266.3333333333333, + 461, + 279.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 281, + 506, + 306 + ], + "lines": [ + { + "bbox": [ + 104, + 276, + 502, + 299 + ], + "spans": [ + { + "bbox": [ + 104, + 276, + 128, + 299 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 128, + 282, + 170, + 294 + ], + "score": 0.89, + "content": "c _ { 2 } = 4 \\sqrt { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 276, + 174, + 299 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 174, + 281, + 270, + 296 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\iota ( k ) = \\ln { \\frac { ( k + 1 ) ( k + 2 ) S A } { \\delta } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 276, + 375, + 299 + ], + "score": 1.0, + "content": ". By a union bound for all", + "type": "text" + }, + { + "bbox": [ + 375, + 284, + 381, + 293 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 382, + 276, + 478, + 299 + ], + "score": 1.0, + "content": ", this holds for arbitrary", + "type": "text" + }, + { + "bbox": [ + 478, + 284, + 502, + 294 + ], + "score": 0.89, + "content": "k > 0", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 293, + 328, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 143, + 307 + ], + "score": 1.0, + "content": "arbitrary", + "type": "text" + }, + { + "bbox": [ + 143, + 295, + 198, + 305 + ], + "score": 0.33, + "content": "s \\in S , a \\in A", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 293, + 328, + 307 + ], + "score": 1.0, + "content": "simultaneously with probability", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 276, + 502, + 307 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 307, + 404, + 341 + ], + "lines": [ + { + "bbox": [ + 205, + 307, + 404, + 341 + ], + "spans": [ + { + "bbox": [ + 205, + 307, + 404, + 341 + ], + "score": 0.93, + "content": "1 - \\sum _ { s ^ { \\prime } \\in S , a ^ { \\prime } \\in A } \\sum _ { k = 1 } ^ { \\infty } \\frac { \\delta } { 2 S A ( k + 1 ) ( k + 2 ) } = 1 - \\frac { \\delta } { 2 } .", + "type": "interline_equation", + "image_path": "d36f303c3c54ca211346b95b702c878d2e496acc0473679ca8912f1c49bc42a5.jpg" + } + ] + } + ], + "index": 16.5, + "virtual_lines": [ + { + "bbox": [ + 205, + 307, + 404, + 324.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 205, + 324.0, + 404, + 341.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 343, + 504, + 366 + ], + "lines": [ + { + "bbox": [ + 105, + 341, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 371, + 356 + ], + "score": 1.0, + "content": "Therefore, we conclude that (16) holds for the random variable", + "type": "text" + }, + { + "bbox": [ + 371, + 343, + 425, + 356 + ], + "score": 0.93, + "content": "t = N _ { p } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 341, + 473, + 356 + ], + "score": 1.0, + "content": "and for all", + "type": "text" + }, + { + "bbox": [ + 474, + 346, + 480, + 355 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 341, + 505, + 356 + ], + "score": 1.0, + "content": ", with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 354, + 220, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 153, + 366 + ], + "score": 1.0, + "content": "probability", + "type": "text" + }, + { + "bbox": [ + 153, + 354, + 186, + 366 + ], + "score": 0.91, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 354, + 220, + 366 + ], + "score": 1.0, + "content": "as well.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 341, + 505, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 418, + 391 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 416, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 336, + 389 + ], + "score": 1.0, + "content": "Proof of the right hand side of (13): We also know that", + "type": "text" + }, + { + "bbox": [ + 337, + 371, + 416, + 392 + ], + "score": 0.68, + "content": "\\begin{array} { r } { ( b _ { k } = \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 371, + 416, + 392 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 393, + 401, + 427 + ], + "lines": [ + { + "bbox": [ + 208, + 393, + 401, + 427 + ], + "spans": [ + { + "bbox": [ + 208, + 393, + 401, + 427 + ], + "score": 0.86, + "content": "\\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } \\leq \\sum _ { i = 1 } ^ { k } \\alpha _ { k } ^ { i } b _ { i } \\leq \\frac { 2 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } .", + "type": "interline_equation", + "image_path": "881fb37a93a36ca33279a8cb7352a55121d9d63aa62473943680b180547b1336.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 393, + 401, + 410.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 208, + 410.0, + 401, + 427.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 428, + 206, + 440 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 207, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 207, + 442 + ], + "score": 1.0, + "content": "It is implied by (16) that", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 427, + 207, + 442 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 149, + 442, + 462, + 510 + ], + "lines": [ + { + "bbox": [ + 149, + 442, + 462, + 510 + ], + "spans": [ + { + "bbox": [ + 149, + 442, + 462, + 510 + ], + "score": 0.88, + "content": "\\begin{array} { r l } & { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\gamma \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\frac { 3 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( t ) } { t } } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( \\hat { V } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) } \\end{array}", + "type": "interline_equation", + "image_path": "d2bbc005947d28dd95e009c07818fe62e6d9b1e5309a47331c5537b6fbe06992.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 149, + 442, + 462, + 464.6666666666667 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 149, + 464.6666666666667, + 462, + 487.33333333333337 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 149, + 487.33333333333337, + 462, + 510.00000000000006 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "interline_equation", + "bbox": [ + 218, + 523, + 403, + 556 + ], + "lines": [ + { + "bbox": [ + 218, + 523, + 403, + 556 + ], + "spans": [ + { + "bbox": [ + 218, + 523, + 403, + 556 + ], + "score": 0.38, + "content": "\\leq \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + \\beta _ { t } .", + "type": "interline_equation", + "image_path": "53eab3d1f26c37eff9f588636d51ccb07f492de053eae442ba9c0192a7489963.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 218, + 523, + 403, + 539.5 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 218, + 539.5, + 403, + 556.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 558, + 343, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 557, + 343, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 146, + 574 + ], + "score": 1.0, + "content": "Note that", + "type": "text" + }, + { + "bbox": [ + 146, + 559, + 262, + 573 + ], + "score": 0.85, + "content": "\\beta _ { t } = c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { H \\iota ( t ) / t }", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 557, + 266, + 574 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 266, + 558, + 341, + 572 + ], + "score": 0.73, + "content": "c _ { 3 } = 3 c _ { 2 } = 1 2 { \\sqrt { 2 } } .", + "type": "inline_equation" + }, + { + "bbox": [ + 341, + 557, + 343, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 557, + 343, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "Proof of the left hand side of (13): Now, we assume that event that (16) holds. We assert that", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 587, + 453, + 601 + ], + "spans": [ + { + "bbox": [ + 107, + 588, + 146, + 600 + ], + "score": 0.92, + "content": "Q _ { p } \\geq Q ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 587, + 174, + 601 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 174, + 588, + 197, + 600 + ], + "score": 0.92, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 587, + 215, + 601 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 215, + 588, + 242, + 600 + ], + "score": 0.92, + "content": "p \\leq p ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 587, + 398, + 601 + ], + "score": 1.0, + "content": ". This assertion is obviously true when", + "type": "text" + }, + { + "bbox": [ + 398, + 588, + 425, + 599 + ], + "score": 0.93, + "content": "p ^ { \\prime } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 587, + 453, + 601 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 576, + 506, + 601 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 163, + 601, + 447, + 671 + ], + "lines": [ + { + "bbox": [ + 163, + 601, + 447, + 671 + ], + "spans": [ + { + "bbox": [ + 163, + 601, + 447, + 671 + ], + "score": 0.94, + "content": "\\begin{array} { c l } { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\geq - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ { \\geq \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } b _ { i } - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| \\geq 0 . } \\end{array}", + "type": "interline_equation", + "image_path": "70ff20e4e31a1bb0299ebb34008753b3eaca5c5e3da766015eda50bc514d5a51.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 163, + 601, + 447, + 624.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 163, + 624.3333333333334, + 447, + 647.6666666666667 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 163, + 647.6666666666667, + 447, + 671.0000000000001 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 671, + 425, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 425, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 238, + 685 + ], + "score": 1.0, + "content": "Therefore the assertion holds for", + "type": "text" + }, + { + "bbox": [ + 239, + 672, + 265, + 683 + ], + "score": 0.91, + "content": "p ^ { \\prime } + 1", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 670, + 415, + 685 + ], + "score": 1.0, + "content": "as well. By induction, it holds for all", + "type": "text" + }, + { + "bbox": [ + 415, + 674, + 421, + 683 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 670, + 425, + 685 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 670, + 425, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 689, + 505, + 717 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 280, + 703 + ], + "score": 1.0, + "content": "We now see that (13) holds for probability", + "type": "text" + }, + { + "bbox": [ + 281, + 690, + 313, + 702 + ], + "score": 0.91, + "content": "1 - \\delta / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 688, + 342, + 703 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 342, + 691, + 368, + 702 + ], + "score": 0.77, + "content": "p , s , a", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 688, + 397, + 703 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 398, + 688, + 433, + 703 + ], + "score": 0.93, + "content": "\\hat { Q } _ { p } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 688, + 505, + 703 + ], + "score": 1.0, + "content": "is always greater", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 702, + 503, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 126, + 718 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 126, + 703, + 164, + 716 + ], + "score": 0.93, + "content": "Q _ { p ^ { \\prime } } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 702, + 203, + 718 + ], + "score": 1.0, + "content": "for some", + "type": "text" + }, + { + "bbox": [ + 204, + 703, + 231, + 715 + ], + "score": 0.92, + "content": "p ^ { \\prime } \\leq p", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 702, + 290, + 718 + ], + "score": 1.0, + "content": ", we know that", + "type": "text" + }, + { + "bbox": [ + 291, + 703, + 425, + 716 + ], + "score": 0.92, + "content": "\\hat { Q } _ { p } ( s , a ) \\geq Q _ { p ^ { \\prime } } ( s , a ) \\geq Q ^ { * } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 702, + 503, + 718 + ], + "score": 1.0, + "content": ", thus proving (14).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 688, + 505, + 718 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 350, + 97 + ], + "score": 1.0, + "content": "We now give a proof for lemma 2. Recall the definition for a", + "type": "text" + }, + { + "bbox": [ + 351, + 83, + 378, + 94 + ], + "score": 0.92, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 79, + 471, + 97 + ], + "score": 1.0, + "content": "-sequence. A sequence", + "type": "text" + }, + { + "bbox": [ + 472, + 82, + 505, + 95 + ], + "score": 0.91, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 483, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 164, + 108 + ], + "score": 1.0, + "content": "is said to be a", + "type": "text" + }, + { + "bbox": [ + 164, + 94, + 191, + 106 + ], + "score": 0.95, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 92, + 246, + 108 + ], + "score": 1.0, + "content": "-sequence for", + "type": "text" + }, + { + "bbox": [ + 246, + 94, + 285, + 105 + ], + "score": 0.92, + "content": "C , w > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 92, + 297, + 108 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 298, + 94, + 348, + 105 + ], + "score": 0.92, + "content": "0 \\leq w _ { t } \\leq w", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 92, + 376, + 108 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 376, + 94, + 399, + 105 + ], + "score": 0.89, + "content": "t \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 92, + 420, + 108 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 420, + 93, + 478, + 108 + ], + "score": 0.92, + "content": "\\textstyle \\sum _ { t \\geq 1 } w _ { t } \\leq C", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 92, + 483, + 108 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 118, + 313, + 131 + ], + "lines": [ + { + "bbox": [ + 106, + 118, + 312, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 152, + 132 + ], + "score": 1.0, + "content": "Proof. Let", + "type": "text" + }, + { + "bbox": [ + 152, + 119, + 216, + 131 + ], + "score": 0.95, + "content": "n _ { t } = N _ { t } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 118, + 312, + 132 + ], + "score": 1.0, + "content": "for simplicity; we have", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 134, + 451, + 226 + ], + "lines": [ + { + "bbox": [ + 158, + 134, + 451, + 226 + ], + "spans": [ + { + "bbox": [ + 158, + 134, + 451, + 226 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { { \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } ( Q _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } [ \\cfrac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } + \\beta _ { n _ { t } } + \\gamma \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } ( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } ) ( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "56de998ef0518be48301779db9712c42690034bbbfeba7d7ae47e786453e29e7.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 158, + 134, + 451, + 164.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 158, + 164.66666666666666, + 451, + 195.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 158, + 195.33333333333331, + 451, + 225.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 104, + 228, + 507, + 246 + ], + "spans": [ + { + "bbox": [ + 104, + 228, + 295, + 246 + ], + "score": 1.0, + "content": "The last inequality is due to lemma 4. Note that", + "type": "text" + }, + { + "bbox": [ + 296, + 230, + 362, + 244 + ], + "score": 0.93, + "content": "\\alpha _ { n _ { t } } ^ { 0 } = \\mathbb { I } [ n _ { t } = 0 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 228, + 507, + 246 + ], + "score": 1.0, + "content": ", the first term in the summation can", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 240, + 170, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 170, + 255 + ], + "score": 1.0, + "content": "be bounded by,", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 256, + 354, + 290 + ], + "lines": [ + { + "bbox": [ + 256, + 256, + 354, + 290 + ], + "spans": [ + { + "bbox": [ + 256, + 256, + 354, + 290 + ], + "score": 0.93, + "content": "\\sum _ { t \\geq 1 } w _ { t } \\frac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } \\leq \\frac { S A w } { 1 - \\gamma } .", + "type": "interline_equation", + "image_path": "fb8a30f2f69b0a0ef551b8c5d16a0c30b4272df4f37761a5fa9bb9799022d5bd.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 256, + 256, + 354, + 273.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 256, + 273.0, + 354, + 290.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 382, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 383, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 217, + 313 + ], + "score": 1.0, + "content": "For the second term, define", + "type": "text" + }, + { + "bbox": [ + 218, + 300, + 314, + 312 + ], + "score": 0.93, + "content": "u ( s , a ) = \\operatorname* { s u p } _ { t } N _ { t } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 299, + 383, + 313 + ], + "score": 1.0, + "content": ".2 It follows that,", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 317, + 402, + 438 + ], + "lines": [ + { + "bbox": [ + 207, + 317, + 402, + 438 + ], + "spans": [ + { + "bbox": [ + 207, + 317, + 402, + 438 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } \\beta _ { n _ { t } } = \\sum _ { s , o } ^ { u ( s , o ) } w _ { \\tau ( s , a , i ) } \\beta _ { i } } \\\\ { \\displaystyle \\qquad \\leq \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sum _ { i = 1 } ^ { C _ { s , a } / w } \\sqrt { \\frac { H \\iota ( i ) } { i } } w } \\\\ { \\displaystyle \\qquad \\leq 2 \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sqrt { \\iota ( C ) H C _ { s , a } w } } \\\\ { \\displaystyle \\qquad \\leq 2 c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { w S A H C \\iota ( C ) } . } \\end{array}", + "type": "interline_equation", + "image_path": "bce9df1dbdecfd18cd0fbd700e55e9d74ea6aeee786233ea436d7b29badc95fc.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 207, + 317, + 402, + 334.2857142857143 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 207, + 334.2857142857143, + 402, + 351.57142857142856 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 207, + 351.57142857142856, + 402, + 368.85714285714283 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 207, + 368.85714285714283, + 402, + 386.1428571428571 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 207, + 386.1428571428571, + 402, + 403.4285714285714 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 207, + 403.4285714285714, + 402, + 420.71428571428567 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 207, + 420.71428571428567, + 402, + 437.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 506, + 466 + ], + "lines": [ + { + "bbox": [ + 104, + 436, + 507, + 458 + ], + "spans": [ + { + "bbox": [ + 104, + 436, + 136, + 458 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 136, + 441, + 254, + 455 + ], + "score": 0.91, + "content": "\\begin{array} { r } { C _ { s , a } = \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } ) = ( s , a ) } w _ { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 436, + 507, + 458 + ], + "score": 1.0, + "content": ". Inequality (19) follows from rearrangement inequality, since", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 451, + 450, + 468 + ], + "spans": [ + { + "bbox": [ + 107, + 453, + 135, + 466 + ], + "score": 0.9, + "content": "\\iota ( x ) / x", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 451, + 450, + 468 + ], + "score": 1.0, + "content": "is monotonically decreasing. Inequality (21) follows from Jensen’s inequality.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 289, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 289, + 483 + ], + "score": 1.0, + "content": "For the third term of the summation, we have", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 485, + 414, + 573 + ], + "lines": [ + { + "bbox": [ + 198, + 485, + 414, + 573 + ], + "spans": [ + { + "bbox": [ + 198, + 485, + 414, + 573 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { \\displaystyle \\sum _ { t \\ge 1 } w _ { t } \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } \\left( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } \\right) \\left( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } \\right) } } \\\\ { { \\displaystyle \\le \\sum _ { t ^ { \\prime } \\ge 1 } \\left( \\hat { V } _ { t ^ { \\prime } } - V ^ { * } \\right) \\left( s _ { t ^ { \\prime } + 1 } \\right) \\left( \\sum _ { t = t ^ { \\prime } + 1 } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) . } } \\end{array}", + "type": "interline_equation", + "image_path": "a3e0c01b442b6f3c208a66bd1e6c84604cad0908954b7748b2f4faaaf3ab9889.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 198, + 485, + 414, + 502.6 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 198, + 502.6, + 414, + 520.2 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 198, + 520.2, + 414, + 537.8000000000001 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 198, + 537.8000000000001, + 414, + 555.4000000000001 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 198, + 555.4000000000001, + 414, + 573.0000000000001 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 135, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 136, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 136, + 604 + ], + "score": 1.0, + "content": "Define", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 598, + 377, + 651 + ], + "lines": [ + { + "bbox": [ + 233, + 598, + 377, + 651 + ], + "spans": [ + { + "bbox": [ + 233, + 598, + 377, + 651 + ], + "score": 0.94, + "content": "w _ { t ^ { \\prime } + 1 } ^ { \\prime } = \\left( \\sum _ { \\stackrel { t = t ^ { \\prime } + 1 } { ( s _ { t } , a _ { t } ) = ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) } } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) .", + "type": "interline_equation", + "image_path": "90d6cf339da0fc0bd898216b68464ffc95f01fd8012796a3a8c23865fc73b784.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 598, + 377, + 624.5 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 233, + 624.5, + 377, + 651.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 659, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 104, + 657, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 165, + 675 + ], + "score": 1.0, + "content": "We claim that", + "type": "text" + }, + { + "bbox": [ + 165, + 660, + 187, + 673 + ], + "score": 0.92, + "content": "w _ { t + 1 } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 657, + 205, + 675 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 205, + 659, + 267, + 673 + ], + "score": 0.92, + "content": "\\textstyle ( C , ( 1 + { \\frac { 1 } { H } } ) w )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 657, + 506, + 675 + ], + "score": 1.0, + "content": "-sequence. We now prove this claim. By lemma 3, for any", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 668, + 138, + 685 + ], + "spans": [ + { + "bbox": [ + 107, + 671, + 132, + 682 + ], + "score": 0.88, + "content": "t ^ { \\prime } \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 668, + 138, + 685 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 680, + 383, + 714 + ], + "lines": [ + { + "bbox": [ + 227, + 680, + 383, + 714 + ], + "spans": [ + { + "bbox": [ + 227, + 680, + 383, + 714 + ], + "score": 0.94, + "content": "w _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq w \\sum _ { j = n _ { t ^ { \\prime } } } ^ { \\infty } \\alpha _ { j } ^ { n _ { t ^ { \\prime } } } = ( 1 + 1 / H ) w .", + "type": "interline_equation", + "image_path": "4253341d3793835fab3da24d7c177c2e36a4c25bd994300d52241f505a8a3adc.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 227, + 680, + 383, + 697.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 227, + 697.0, + 383, + 714.0 + ], + "spans": [], + "index": 32 + } + ] + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 118, + 720, + 394, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 394, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 721, + 149, + 732 + ], + "score": 0.89, + "content": "^ 2 u ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 720, + 234, + 733 + ], + "score": 1.0, + "content": "could be infinity when", + "type": "text" + }, + { + "bbox": [ + 234, + 722, + 255, + 732 + ], + "score": 0.89, + "content": "( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 720, + 394, + 733 + ], + "score": 1.0, + "content": "is visited for infinite number of times.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 350, + 97 + ], + "score": 1.0, + "content": "We now give a proof for lemma 2. Recall the definition for a", + "type": "text" + }, + { + "bbox": [ + 351, + 83, + 378, + 94 + ], + "score": 0.92, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 79, + 471, + 97 + ], + "score": 1.0, + "content": "-sequence. A sequence", + "type": "text" + }, + { + "bbox": [ + 472, + 82, + 505, + 95 + ], + "score": 0.91, + "content": "( w _ { t } ) _ { t \\geq 1 }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 483, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 164, + 108 + ], + "score": 1.0, + "content": "is said to be a", + "type": "text" + }, + { + "bbox": [ + 164, + 94, + 191, + 106 + ], + "score": 0.95, + "content": "( C , w )", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 92, + 246, + 108 + ], + "score": 1.0, + "content": "-sequence for", + "type": "text" + }, + { + "bbox": [ + 246, + 94, + 285, + 105 + ], + "score": 0.92, + "content": "C , w > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 92, + 297, + 108 + ], + "score": 1.0, + "content": ", if", + "type": "text" + }, + { + "bbox": [ + 298, + 94, + 348, + 105 + ], + "score": 0.92, + "content": "0 \\leq w _ { t } \\leq w", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 92, + 376, + 108 + ], + "score": 1.0, + "content": "for all", + "type": "text" + }, + { + "bbox": [ + 376, + 94, + 399, + 105 + ], + "score": 0.89, + "content": "t \\geq 1", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 92, + 420, + 108 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 420, + 93, + 478, + 108 + ], + "score": 0.92, + "content": "\\textstyle \\sum _ { t \\geq 1 } w _ { t } \\leq C", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 92, + 483, + 108 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 79, + 505, + 108 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 118, + 313, + 131 + ], + "lines": [ + { + "bbox": [ + 106, + 118, + 312, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 118, + 152, + 132 + ], + "score": 1.0, + "content": "Proof. Let", + "type": "text" + }, + { + "bbox": [ + 152, + 119, + 216, + 131 + ], + "score": 0.95, + "content": "n _ { t } = N _ { t } ( s _ { t } , a _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 118, + 312, + 132 + ], + "score": 1.0, + "content": "for simplicity; we have", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 118, + 312, + 132 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 134, + 451, + 226 + ], + "lines": [ + { + "bbox": [ + 158, + 134, + 451, + 226 + ], + "spans": [ + { + "bbox": [ + 158, + 134, + 451, + 226 + ], + "score": 0.94, + "content": "\\begin{array} { r l } { { \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } ( Q _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } [ \\cfrac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } + \\beta _ { n _ { t } } + \\gamma \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } ( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } ) ( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } ) ] } \\end{array}", + "type": "interline_equation", + "image_path": "56de998ef0518be48301779db9712c42690034bbbfeba7d7ae47e786453e29e7.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 158, + 134, + 451, + 164.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 158, + 164.66666666666666, + 451, + 195.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 158, + 195.33333333333331, + 451, + 225.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 104, + 228, + 507, + 246 + ], + "spans": [ + { + "bbox": [ + 104, + 228, + 295, + 246 + ], + "score": 1.0, + "content": "The last inequality is due to lemma 4. Note that", + "type": "text" + }, + { + "bbox": [ + 296, + 230, + 362, + 244 + ], + "score": 0.93, + "content": "\\alpha _ { n _ { t } } ^ { 0 } = \\mathbb { I } [ n _ { t } = 0 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 228, + 507, + 246 + ], + "score": 1.0, + "content": ", the first term in the summation can", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 240, + 170, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 170, + 255 + ], + "score": 1.0, + "content": "be bounded by,", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 104, + 228, + 507, + 255 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 256, + 354, + 290 + ], + "lines": [ + { + "bbox": [ + 256, + 256, + 354, + 290 + ], + "spans": [ + { + "bbox": [ + 256, + 256, + 354, + 290 + ], + "score": 0.93, + "content": "\\sum _ { t \\geq 1 } w _ { t } \\frac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } \\leq \\frac { S A w } { 1 - \\gamma } .", + "type": "interline_equation", + "image_path": "fb8a30f2f69b0a0ef551b8c5d16a0c30b4272df4f37761a5fa9bb9799022d5bd.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 256, + 256, + 354, + 273.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 256, + 273.0, + 354, + 290.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 382, + 313 + ], + "lines": [ + { + "bbox": [ + 105, + 299, + 383, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 217, + 313 + ], + "score": 1.0, + "content": "For the second term, define", + "type": "text" + }, + { + "bbox": [ + 218, + 300, + 314, + 312 + ], + "score": 0.93, + "content": "u ( s , a ) = \\operatorname* { s u p } _ { t } N _ { t } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 299, + 383, + 313 + ], + "score": 1.0, + "content": ".2 It follows that,", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 299, + 383, + 313 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 317, + 402, + 438 + ], + "lines": [ + { + "bbox": [ + 207, + 317, + 402, + 438 + ], + "spans": [ + { + "bbox": [ + 207, + 317, + 402, + 438 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } \\beta _ { n _ { t } } = \\sum _ { s , o } ^ { u ( s , o ) } w _ { \\tau ( s , a , i ) } \\beta _ { i } } \\\\ { \\displaystyle \\qquad \\leq \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sum _ { i = 1 } ^ { C _ { s , a } / w } \\sqrt { \\frac { H \\iota ( i ) } { i } } w } \\\\ { \\displaystyle \\qquad \\leq 2 \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sqrt { \\iota ( C ) H C _ { s , a } w } } \\\\ { \\displaystyle \\qquad \\leq 2 c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { w S A H C \\iota ( C ) } . } \\end{array}", + "type": "interline_equation", + "image_path": "bce9df1dbdecfd18cd0fbd700e55e9d74ea6aeee786233ea436d7b29badc95fc.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 207, + 317, + 402, + 334.2857142857143 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 207, + 334.2857142857143, + 402, + 351.57142857142856 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 207, + 351.57142857142856, + 402, + 368.85714285714283 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 207, + 368.85714285714283, + 402, + 386.1428571428571 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 207, + 386.1428571428571, + 402, + 403.4285714285714 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 207, + 403.4285714285714, + 402, + 420.71428571428567 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 207, + 420.71428571428567, + 402, + 437.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 440, + 506, + 466 + ], + "lines": [ + { + "bbox": [ + 104, + 436, + 507, + 458 + ], + "spans": [ + { + "bbox": [ + 104, + 436, + 136, + 458 + ], + "score": 1.0, + "content": "Where", + "type": "text" + }, + { + "bbox": [ + 136, + 441, + 254, + 455 + ], + "score": 0.91, + "content": "\\begin{array} { r } { C _ { s , a } = \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } ) = ( s , a ) } w _ { t } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 436, + 507, + 458 + ], + "score": 1.0, + "content": ". Inequality (19) follows from rearrangement inequality, since", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 451, + 450, + 468 + ], + "spans": [ + { + "bbox": [ + 107, + 453, + 135, + 466 + ], + "score": 0.9, + "content": "\\iota ( x ) / x", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 451, + 450, + 468 + ], + "score": 1.0, + "content": "is monotonically decreasing. Inequality (21) follows from Jensen’s inequality.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 104, + 436, + 507, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 289, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 469, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 289, + 483 + ], + "score": 1.0, + "content": "For the third term of the summation, we have", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 469, + 289, + 483 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 485, + 414, + 573 + ], + "lines": [ + { + "bbox": [ + 198, + 485, + 414, + 573 + ], + "spans": [ + { + "bbox": [ + 198, + 485, + 414, + 573 + ], + "score": 0.93, + "content": "\\begin{array} { c } { { \\displaystyle \\sum _ { t \\ge 1 } w _ { t } \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } \\left( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } \\right) \\left( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } \\right) } } \\\\ { { \\displaystyle \\le \\sum _ { t ^ { \\prime } \\ge 1 } \\left( \\hat { V } _ { t ^ { \\prime } } - V ^ { * } \\right) \\left( s _ { t ^ { \\prime } + 1 } \\right) \\left( \\sum _ { t = t ^ { \\prime } + 1 } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) . } } \\end{array}", + "type": "interline_equation", + "image_path": "a3e0c01b442b6f3c208a66bd1e6c84604cad0908954b7748b2f4faaaf3ab9889.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 198, + 485, + 414, + 502.6 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 198, + 502.6, + 414, + 520.2 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 198, + 520.2, + 414, + 537.8000000000001 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 198, + 537.8000000000001, + 414, + 555.4000000000001 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 198, + 555.4000000000001, + 414, + 573.0000000000001 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 591, + 135, + 602 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 136, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 136, + 604 + ], + "score": 1.0, + "content": "Define", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 589, + 136, + 604 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 598, + 377, + 651 + ], + "lines": [ + { + "bbox": [ + 233, + 598, + 377, + 651 + ], + "spans": [ + { + "bbox": [ + 233, + 598, + 377, + 651 + ], + "score": 0.94, + "content": "w _ { t ^ { \\prime } + 1 } ^ { \\prime } = \\left( \\sum _ { \\stackrel { t = t ^ { \\prime } + 1 } { ( s _ { t } , a _ { t } ) = ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) } } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) .", + "type": "interline_equation", + "image_path": "90d6cf339da0fc0bd898216b68464ffc95f01fd8012796a3a8c23865fc73b784.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 598, + 377, + 624.5 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 233, + 624.5, + 377, + 651.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 659, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 104, + 657, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 165, + 675 + ], + "score": 1.0, + "content": "We claim that", + "type": "text" + }, + { + "bbox": [ + 165, + 660, + 187, + 673 + ], + "score": 0.92, + "content": "w _ { t + 1 } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 657, + 205, + 675 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 205, + 659, + 267, + 673 + ], + "score": 0.92, + "content": "\\textstyle ( C , ( 1 + { \\frac { 1 } { H } } ) w )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 657, + 506, + 675 + ], + "score": 1.0, + "content": "-sequence. We now prove this claim. By lemma 3, for any", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 668, + 138, + 685 + ], + "spans": [ + { + "bbox": [ + 107, + 671, + 132, + 682 + ], + "score": 0.88, + "content": "t ^ { \\prime } \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 668, + 138, + 685 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 657, + 506, + 685 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 227, + 680, + 383, + 714 + ], + "lines": [ + { + "bbox": [ + 227, + 680, + 383, + 714 + ], + "spans": [ + { + "bbox": [ + 227, + 680, + 383, + 714 + ], + "score": 0.94, + "content": "w _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq w \\sum _ { j = n _ { t ^ { \\prime } } } ^ { \\infty } \\alpha _ { j } ^ { n _ { t ^ { \\prime } } } = ( 1 + 1 / H ) w .", + "type": "interline_equation", + "image_path": "4253341d3793835fab3da24d7c177c2e36a4c25bd994300d52241f505a8a3adc.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 227, + 680, + 383, + 697.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 227, + 697.0, + 383, + 714.0 + ], + "spans": [], + "index": 32 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 101, + 74, + 509, + 101 + ], + "spans": [ + { + "bbox": [ + 101, + 74, + 120, + 101 + ], + "score": 1.0, + "content": "By", + "type": "text" + }, + { + "bbox": [ + 120, + 80, + 177, + 96 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { j = 0 } ^ { i } \\alpha _ { i } ^ { j } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 74, + 214, + 101 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 214, + 82, + 339, + 96 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sum _ { t ^ { \\prime } \\geq 1 } w _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq \\sum _ { t \\geq 1 } w _ { t } \\leq C . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 74, + 509, + 101 + ], + "score": 1.0, + "content": ". This proves the assertion. It follows from", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 145, + 109 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 145, + 109 + ], + "score": 1.0, + "content": "(22) that", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 110, + 438, + 267 + ], + "lines": [ + { + "bbox": [ + 169, + 110, + 438, + 267 + ], + "spans": [ + { + "bbox": [ + 169, + 110, + 438, + 267 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\quad \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { = \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - \\hat { V } _ { t + 1 } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( 2 \\alpha _ { n _ { t } + 1 } \\frac { 1 } { 1 - \\gamma } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) \\left( s _ { t + 1 } , a _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\end{array}", + "type": "interline_equation", + "image_path": "a1c63f8ce8866bfa935cc54ba0fa4060644052edbd98620aad79e7ef97f0f148.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 169, + 110, + 438, + 162.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 169, + 162.33333333333334, + 438, + 214.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 169, + 214.66666666666669, + 438, + 267.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 270, + 503, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 270, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 504, + 282 + ], + "score": 1.0, + "content": "Inequality (25) comes from the update rule of our algorithm. Inequality (26) comes from the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 277, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 104, + 277, + 144, + 297 + ], + "score": 1.0, + "content": "fact that", + "type": "text" + }, + { + "bbox": [ + 144, + 281, + 277, + 293 + ], + "score": 0.91, + "content": "\\dot { \\alpha } _ { t } = ( H + 1 ) / ( H + t ) \\overset { \\cdot } { \\leq } H / t", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 277, + 473, + 297 + ], + "score": 1.0, + "content": "and Jensen’s Inequality. More specifically, let", + "type": "text" + }, + { + "bbox": [ + 473, + 281, + 505, + 295 + ], + "score": 0.91, + "content": "C _ { s , a } ^ { \\prime } =", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 292, + 303, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 193, + 308 + ], + "score": 0.85, + "content": "\\scriptstyle \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } = s , a } w _ { t + 1 } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 292, + 197, + 309 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 294, + 273, + 306 + ], + "score": 0.83, + "content": "w ^ { \\prime } = w ( 1 + 1 / H )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 292, + 303, + 309 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 311, + 462, + 351 + ], + "lines": [ + { + "bbox": [ + 148, + 311, + 462, + 351 + ], + "spans": [ + { + "bbox": [ + 148, + 311, + 462, + 351 + ], + "score": 0.92, + "content": "\\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\alpha _ { n _ { t } + 1 } \\leq \\sum _ { s , a } \\sum _ { n = 1 } ^ { C _ { s , a } ^ { \\prime } / w ^ { \\prime } } w ^ { \\prime } \\frac { H } { n } \\leq \\sum _ { s , a } H w ^ { \\prime } \\ln ( C _ { s , a } ^ { \\prime } / w ) \\leq 2 S A H w \\ln C .", + "type": "interline_equation", + "image_path": "9834ce9a1e7dc58cf0a708f75e6b5b858bcdc589ad24752ed45e1722a959c115.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 148, + 311, + 462, + 324.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 148, + 324.3333333333333, + 462, + 337.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 148, + 337.66666666666663, + 462, + 350.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 289, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 290, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 290, + 375 + ], + "score": 1.0, + "content": "Putting (18), (21) and (27) together, we have,", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 389, + 477, + 453 + ], + "lines": [ + { + "bbox": [ + 117, + 389, + 477, + 453 + ], + "spans": [ + { + "bbox": [ + 117, + 389, + 477, + 453 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w S A H C \\iota ( C ) } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln C \\right) + \\gamma \\displaystyle \\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) ( s _ { t + 1 } , a _ { t + 1 } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "fca33e0caf6dada43c874ac741c299779e61219671c7f66b39e75e1f2d6533ae.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 117, + 389, + 477, + 410.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 117, + 410.3333333333333, + 477, + 431.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 117, + 431.66666666666663, + 477, + 452.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 507, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 469 + ], + "score": 1.0, + "content": "Observe that the third term is another weighted sum with the same form as (17). Therefore, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "can unroll this term repetitively with changing weight sequences.Suppose that our original weight", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 102, + 475, + 508, + 497 + ], + "spans": [ + { + "bbox": [ + 102, + 475, + 217, + 497 + ], + "score": 1.0, + "content": "sequence is also denoted by", + "type": "text" + }, + { + "bbox": [ + 218, + 478, + 260, + 493 + ], + "score": 0.93, + "content": "\\{ \\bar { w _ { t } ^ { ( 0 ) } } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 475, + 288, + 497 + ], + "score": 1.0, + "content": ", while", + "type": "text" + }, + { + "bbox": [ + 289, + 479, + 331, + 494 + ], + "score": 0.93, + "content": "\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 475, + 508, + 497 + ], + "score": 1.0, + "content": "denotes the weight sequence after unrolling", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 491, + 508, + 510 + ], + "spans": [ + { + "bbox": [ + 104, + 491, + 120, + 510 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 496, + 128, + 505 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 491, + 172, + 510 + ], + "score": 1.0, + "content": "times. Let", + "type": "text" + }, + { + "bbox": [ + 173, + 494, + 192, + 506 + ], + "score": 0.9, + "content": "w ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 491, + 205, + 510 + ], + "score": 1.0, + "content": "be", + "type": "text" + }, + { + "bbox": [ + 206, + 493, + 270, + 507 + ], + "score": 0.93, + "content": "w \\cdot \\left( 1 + 1 / H \\right) ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 491, + 362, + 510 + ], + "score": 1.0, + "content": ". Then we can see that", + "type": "text" + }, + { + "bbox": [ + 363, + 492, + 406, + 508 + ], + "score": 0.94, + "content": "\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 491, + 424, + 510 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 424, + 494, + 462, + 507 + ], + "score": 0.91, + "content": "( C , w ^ { ( k ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 491, + 508, + 510 + ], + "score": 1.0, + "content": "-sequence.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 506, + 277, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 214, + 518 + ], + "score": 1.0, + "content": "Suppose that we unroll for", + "type": "text" + }, + { + "bbox": [ + 215, + 507, + 225, + 516 + ], + "score": 0.81, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 506, + 277, + 518 + ], + "score": 1.0, + "content": "times. Then", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 125, + 522, + 487, + 617 + ], + "lines": [ + { + "bbox": [ + 125, + 522, + 487, + 617 + ], + "spans": [ + { + "bbox": [ + 125, + 522, + 487, + 617 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ^ { ( H ) } \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\frac { C } { 1 - \\gamma } . } \\end{array}", + "type": "interline_equation", + "image_path": "4b3ee4cb41b96d3c4f223bc4146fc59bb343a3cac7f567861e3ed3154212ede1.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 125, + 522, + 487, + 553.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 125, + 553.6666666666666, + 487, + 585.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 125, + 585.3333333333333, + 487, + 616.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 506, + 665 + ], + "lines": [ + { + "bbox": [ + 102, + 629, + 508, + 653 + ], + "spans": [ + { + "bbox": [ + 102, + 629, + 136, + 653 + ], + "score": 1.0, + "content": "We set", + "type": "text" + }, + { + "bbox": [ + 136, + 633, + 280, + 651 + ], + "score": 0.94, + "content": "\\begin{array} { r } { H = \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma } , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 629, + 342, + 653 + ], + "score": 1.0, + "content": ". It follows that", + "type": "text" + }, + { + "bbox": [ + 342, + 635, + 483, + 649 + ], + "score": 0.92, + "content": "w ^ { ( H ) } = ( 1 + 1 / H ) ^ { H } w ^ { ( 0 ) } \\leq e w ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 629, + 508, + 653 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 102, + 646, + 402, + 668 + ], + "spans": [ + { + "bbox": [ + 102, + 646, + 159, + 668 + ], + "score": 1.0, + "content": "that γH C1−γ", + "type": "text" + }, + { + "bbox": [ + 124, + 650, + 185, + 666 + ], + "score": 0.92, + "content": "\\gamma ^ { H } \\frac { C } { 1 - \\gamma } \\leq C \\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 649, + 226, + 665 + ], + "score": 1.0, + "content": ". Also, let", + "type": "text" + }, + { + "bbox": [ + 226, + 650, + 352, + 664 + ], + "score": 0.9, + "content": "\\ell ( C ) = \\iota ( C ) \\ln ( ( 1 - \\gamma ) ^ { - 1 } \\epsilon _ { 1 } ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 649, + 402, + 665 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 682, + 482, + 716 + ], + "lines": [ + { + "bbox": [ + 113, + 682, + 482, + 716 + ], + "spans": [ + { + "bbox": [ + 113, + 682, + 482, + 716 + ], + "score": 0.93, + "content": "\\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln C \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "11b4d3a7f4cf29588edd0397db246aa8a18f798a841609b619f1a546acf17d46.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 113, + 682, + 482, + 716 + ], + "spans": [], + "index": 25 + } + ] + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 721, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 496, + 723, + 504, + 730 + ], + "spans": [ + { + "bbox": [ + 496, + 723, + 504, + 730 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 101, + 74, + 509, + 101 + ], + "spans": [ + { + "bbox": [ + 101, + 74, + 120, + 101 + ], + "score": 1.0, + "content": "By", + "type": "text" + }, + { + "bbox": [ + 120, + 80, + 177, + 96 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { j = 0 } ^ { i } \\alpha _ { i } ^ { j } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 74, + 214, + 101 + ], + "score": 1.0, + "content": ", we have", + "type": "text" + }, + { + "bbox": [ + 214, + 82, + 339, + 96 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\sum _ { t ^ { \\prime } \\geq 1 } w _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq \\sum _ { t \\geq 1 } w _ { t } \\leq C . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 74, + 509, + 101 + ], + "score": 1.0, + "content": ". This proves the assertion. It follows from", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 92, + 145, + 109 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 145, + 109 + ], + "score": 1.0, + "content": "(22) that", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 101, + 74, + 509, + 109 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 110, + 438, + 267 + ], + "lines": [ + { + "bbox": [ + 169, + 110, + 438, + 267 + ], + "spans": [ + { + "bbox": [ + 169, + 110, + 438, + 267 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\quad \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { = \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - \\hat { V } _ { t + 1 } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( 2 \\alpha _ { n _ { t } + 1 } \\frac { 1 } { 1 - \\gamma } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) \\left( s _ { t + 1 } , a _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\end{array}", + "type": "interline_equation", + "image_path": "a1c63f8ce8866bfa935cc54ba0fa4060644052edbd98620aad79e7ef97f0f148.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 169, + 110, + 438, + 162.33333333333334 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 169, + 162.33333333333334, + 438, + 214.66666666666669 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 169, + 214.66666666666669, + 438, + 267.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 270, + 503, + 307 + ], + "lines": [ + { + "bbox": [ + 106, + 270, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 504, + 282 + ], + "score": 1.0, + "content": "Inequality (25) comes from the update rule of our algorithm. Inequality (26) comes from the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 277, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 104, + 277, + 144, + 297 + ], + "score": 1.0, + "content": "fact that", + "type": "text" + }, + { + "bbox": [ + 144, + 281, + 277, + 293 + ], + "score": 0.91, + "content": "\\dot { \\alpha } _ { t } = ( H + 1 ) / ( H + t ) \\overset { \\cdot } { \\leq } H / t", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 277, + 473, + 297 + ], + "score": 1.0, + "content": "and Jensen’s Inequality. More specifically, let", + "type": "text" + }, + { + "bbox": [ + 473, + 281, + 505, + 295 + ], + "score": 0.91, + "content": "C _ { s , a } ^ { \\prime } =", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 292, + 303, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 193, + 308 + ], + "score": 0.85, + "content": "\\scriptstyle \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } = s , a } w _ { t + 1 } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 292, + 197, + 309 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 198, + 294, + 273, + 306 + ], + "score": 0.83, + "content": "w ^ { \\prime } = w ( 1 + 1 / H )", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 292, + 303, + 309 + ], + "score": 1.0, + "content": ". Then", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 104, + 270, + 505, + 309 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 311, + 462, + 351 + ], + "lines": [ + { + "bbox": [ + 148, + 311, + 462, + 351 + ], + "spans": [ + { + "bbox": [ + 148, + 311, + 462, + 351 + ], + "score": 0.92, + "content": "\\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\alpha _ { n _ { t } + 1 } \\leq \\sum _ { s , a } \\sum _ { n = 1 } ^ { C _ { s , a } ^ { \\prime } / w ^ { \\prime } } w ^ { \\prime } \\frac { H } { n } \\leq \\sum _ { s , a } H w ^ { \\prime } \\ln ( C _ { s , a } ^ { \\prime } / w ) \\leq 2 S A H w \\ln C .", + "type": "interline_equation", + "image_path": "9834ce9a1e7dc58cf0a708f75e6b5b858bcdc589ad24752ed45e1722a959c115.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 148, + 311, + 462, + 324.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 148, + 324.3333333333333, + 462, + 337.66666666666663 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 148, + 337.66666666666663, + 462, + 350.99999999999994 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 289, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 290, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 290, + 375 + ], + "score": 1.0, + "content": "Putting (18), (21) and (27) together, we have,", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 358, + 290, + 375 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 389, + 477, + 453 + ], + "lines": [ + { + "bbox": [ + 117, + 389, + 477, + 453 + ], + "spans": [ + { + "bbox": [ + 117, + 389, + 477, + 453 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w S A H C \\iota ( C ) } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln C \\right) + \\gamma \\displaystyle \\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) ( s _ { t + 1 } , a _ { t + 1 } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "fca33e0caf6dada43c874ac741c299779e61219671c7f66b39e75e1f2d6533ae.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 117, + 389, + 477, + 410.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 117, + 410.3333333333333, + 477, + 431.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 117, + 431.66666666666663, + 477, + 452.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 456, + 507, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 455, + 506, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 469 + ], + "score": 1.0, + "content": "Observe that the third term is another weighted sum with the same form as (17). Therefore, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 468, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 480 + ], + "score": 1.0, + "content": "can unroll this term repetitively with changing weight sequences.Suppose that our original weight", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 102, + 475, + 508, + 497 + ], + "spans": [ + { + "bbox": [ + 102, + 475, + 217, + 497 + ], + "score": 1.0, + "content": "sequence is also denoted by", + "type": "text" + }, + { + "bbox": [ + 218, + 478, + 260, + 493 + ], + "score": 0.93, + "content": "\\{ \\bar { w _ { t } ^ { ( 0 ) } } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 475, + 288, + 497 + ], + "score": 1.0, + "content": ", while", + "type": "text" + }, + { + "bbox": [ + 289, + 479, + 331, + 494 + ], + "score": 0.93, + "content": "\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 475, + 508, + 497 + ], + "score": 1.0, + "content": "denotes the weight sequence after unrolling", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 491, + 508, + 510 + ], + "spans": [ + { + "bbox": [ + 104, + 491, + 120, + 510 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 121, + 496, + 128, + 505 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 491, + 172, + 510 + ], + "score": 1.0, + "content": "times. Let", + "type": "text" + }, + { + "bbox": [ + 173, + 494, + 192, + 506 + ], + "score": 0.9, + "content": "w ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 491, + 205, + 510 + ], + "score": 1.0, + "content": "be", + "type": "text" + }, + { + "bbox": [ + 206, + 493, + 270, + 507 + ], + "score": 0.93, + "content": "w \\cdot \\left( 1 + 1 / H \\right) ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 491, + 362, + 510 + ], + "score": 1.0, + "content": ". Then we can see that", + "type": "text" + }, + { + "bbox": [ + 363, + 492, + 406, + 508 + ], + "score": 0.94, + "content": "\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 491, + 424, + 510 + ], + "score": 1.0, + "content": "is a", + "type": "text" + }, + { + "bbox": [ + 424, + 494, + 462, + 507 + ], + "score": 0.91, + "content": "( C , w ^ { ( k ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 491, + 508, + 510 + ], + "score": 1.0, + "content": "-sequence.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 506, + 277, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 214, + 518 + ], + "score": 1.0, + "content": "Suppose that we unroll for", + "type": "text" + }, + { + "bbox": [ + 215, + 507, + 225, + 516 + ], + "score": 0.81, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 506, + 277, + 518 + ], + "score": 1.0, + "content": "times. Then", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 102, + 455, + 508, + 518 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 125, + 522, + 487, + 617 + ], + "lines": [ + { + "bbox": [ + 125, + 522, + 487, + 617 + ], + "spans": [ + { + "bbox": [ + 125, + 522, + 487, + 617 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ^ { ( H ) } \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\frac { C } { 1 - \\gamma } . } \\end{array}", + "type": "interline_equation", + "image_path": "4b3ee4cb41b96d3c4f223bc4146fc59bb343a3cac7f567861e3ed3154212ede1.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 125, + 522, + 487, + 553.6666666666666 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 125, + 553.6666666666666, + 487, + 585.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 125, + 585.3333333333333, + 487, + 616.9999999999999 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 506, + 665 + ], + "lines": [ + { + "bbox": [ + 102, + 629, + 508, + 653 + ], + "spans": [ + { + "bbox": [ + 102, + 629, + 136, + 653 + ], + "score": 1.0, + "content": "We set", + "type": "text" + }, + { + "bbox": [ + 136, + 633, + 280, + 651 + ], + "score": 0.94, + "content": "\\begin{array} { r } { H = \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma } , } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 629, + 342, + 653 + ], + "score": 1.0, + "content": ". It follows that", + "type": "text" + }, + { + "bbox": [ + 342, + 635, + 483, + 649 + ], + "score": 0.92, + "content": "w ^ { ( H ) } = ( 1 + 1 / H ) ^ { H } w ^ { ( 0 ) } \\leq e w ^ { ( 0 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 629, + 508, + 653 + ], + "score": 1.0, + "content": ", and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 102, + 646, + 402, + 668 + ], + "spans": [ + { + "bbox": [ + 102, + 646, + 159, + 668 + ], + "score": 1.0, + "content": "that γH C1−γ", + "type": "text" + }, + { + "bbox": [ + 124, + 650, + 185, + 666 + ], + "score": 0.92, + "content": "\\gamma ^ { H } \\frac { C } { 1 - \\gamma } \\leq C \\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 649, + 226, + 665 + ], + "score": 1.0, + "content": ". Also, let", + "type": "text" + }, + { + "bbox": [ + 226, + 650, + 352, + 664 + ], + "score": 0.9, + "content": "\\ell ( C ) = \\iota ( C ) \\ln ( ( 1 - \\gamma ) ^ { - 1 } \\epsilon _ { 1 } ^ { - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 649, + 402, + 665 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 102, + 629, + 508, + 668 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 113, + 682, + 482, + 716 + ], + "lines": [ + { + "bbox": [ + 113, + 682, + 482, + 716 + ], + "spans": [ + { + "bbox": [ + 113, + 682, + 482, + 716 + ], + "score": 0.93, + "content": "\\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln C \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) .", + "type": "interline_equation", + "image_path": "11b4d3a7f4cf29588edd0397db246aa8a18f798a841609b619f1a546acf17d46.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 113, + 682, + 482, + 716 + ], + "spans": [], + "index": 25 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 297, + 95 + ], + "score": 1.0, + "content": "C EXTENSION TO OTHER SETTINGS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 506, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "First we define a mapping from a finite horizon MDP to an infinite horizon MDP so that our algorithm", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 504, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 312, + 129 + ], + "score": 1.0, + "content": "can be applied. For an arbitrary finite horizon MDP", + "type": "text" + }, + { + "bbox": [ + 312, + 117, + 465, + 129 + ], + "score": 0.91, + "content": "\\mathcal { M } = ( S , A , H , r _ { h } ( s _ { \\perp } a ) , p _ { h } ( s ^ { \\prime } \\mid s , a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 116, + 493, + 129 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 494, + 117, + 504, + 127 + ], + "score": 0.79, + "content": "H", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 363, + 140 + ], + "score": 1.0, + "content": "is the length of episode, the corresponding infinite horizon MDP", + "type": "text" + }, + { + "bbox": [ + 364, + 129, + 504, + 140 + ], + "score": 0.88, + "content": "\\mathcal { M } = ( S , A , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 138, + 162, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 138, + 162, + 151 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 131, + 157, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 131, + 157, + 226, + 187 + ], + "spans": [ + { + "bbox": [ + 131, + 157, + 226, + 187 + ], + "score": 0.37, + "content": "\\begin{array} { l } { { \\bullet \\ { \\bar { S } } = S \\times H , { \\bar { A } } = A ; } } \\\\ { { \\bullet \\ \\gamma = ( 1 - 1 / H ) ; } } \\end{array}", + "type": "inline_equation", + "image_path": "6163bc4f633c8d81c5759aacc60e20932b8157511f7dd0e021ff4e1df174dd54.jpg" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 188, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 185, + 201 + ], + "score": 1.0, + "content": "• for a state", + "type": "text" + }, + { + "bbox": [ + 186, + 192, + 191, + 199 + ], + "score": 0.76, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 188, + 221, + 201 + ], + "score": 1.0, + "content": "at step", + "type": "text" + }, + { + "bbox": [ + 222, + 190, + 228, + 199 + ], + "score": 0.83, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 188, + 245, + 201 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 246, + 190, + 263, + 201 + ], + "score": 0.88, + "content": "\\bar { s } _ { s , h }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 188, + 438, + 201 + ], + "score": 1.0, + "content": "be the corresponding state. For any action", + "type": "text" + }, + { + "bbox": [ + 439, + 192, + 445, + 199 + ], + "score": 0.77, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 188, + 505, + 201 + ], + "score": 1.0, + "content": "and next state", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 198, + 507, + 216 + ], + "spans": [ + { + "bbox": [ + 142, + 201, + 151, + 211 + ], + "score": 0.84, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 198, + 183, + 216 + ], + "score": 1.0, + "content": ", define", + "type": "text" + }, + { + "bbox": [ + 183, + 201, + 303, + 213 + ], + "score": 0.91, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 198, + 322, + 216 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 323, + 201, + 466, + 213 + ], + "score": 0.91, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s ^ { \\prime } \\mid s , h )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 198, + 507, + 216 + ], + "score": 1.0, + "content": ". And for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 211, + 497, + 226 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 172, + 223 + ], + "score": 0.88, + "content": "h = H", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 211, + 189, + 226 + ], + "score": 1.0, + "content": ", set", + "type": "text" + }, + { + "bbox": [ + 190, + 213, + 247, + 225 + ], + "score": 0.91, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 211, + 265, + 226 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 265, + 213, + 384, + 225 + ], + "score": 0.92, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 211, + 482, + 226 + ], + "score": 1.0, + "content": "for a fixed starting state", + "type": "text" + }, + { + "bbox": [ + 482, + 214, + 492, + 223 + ], + "score": 0.84, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 211, + 497, + 226 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 506, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 122, + 247 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 234, + 133, + 245 + ], + "score": 0.88, + "content": "\\bar { V } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 233, + 231, + 247 + ], + "score": 1.0, + "content": "be the value function in", + "type": "text" + }, + { + "bbox": [ + 231, + 234, + 244, + 245 + ], + "score": 0.87, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 233, + 275, + 247 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 275, + 235, + 281, + 244 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 233, + 298, + 247 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 234, + 312, + 247 + ], + "score": 0.91, + "content": "V _ { h } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 233, + 398, + 247 + ], + "score": 1.0, + "content": "the value function in", + "type": "text" + }, + { + "bbox": [ + 399, + 235, + 412, + 245 + ], + "score": 0.82, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 233, + 456, + 247 + ], + "score": 1.0, + "content": "at episode", + "type": "text" + }, + { + "bbox": [ + 456, + 235, + 462, + 244 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 233, + 485, + 247 + ], + "score": 1.0, + "content": ", step", + "type": "text" + }, + { + "bbox": [ + 485, + 235, + 492, + 244 + ], + "score": 0.76, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 233, + 506, + 247 + ], + "score": 1.0, + "content": ". It", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 245, + 507, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 245, + 157, + 265 + ], + "score": 1.0, + "content": "follows that", + "type": "text" + }, + { + "bbox": [ + 158, + 246, + 264, + 263 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V _ { 1 } ^ { \\ast } ( s _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 245, + 422, + 265 + ], + "score": 1.0, + "content": ". And the policy mapping is defined as", + "type": "text" + }, + { + "bbox": [ + 422, + 249, + 489, + 262 + ], + "score": 0.93, + "content": "\\pi _ { h } ( s ) = \\bar { \\pi } ( \\bar { s } _ { s , h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 245, + 507, + 265 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 262, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 133, + 276 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 133, + 264, + 141, + 273 + ], + "score": 0.78, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 262, + 152, + 276 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 153, + 262, + 165, + 273 + ], + "score": 0.85, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 262, + 268, + 276 + ], + "score": 1.0, + "content": ". Value functions in MDP", + "type": "text" + }, + { + "bbox": [ + 268, + 263, + 281, + 273 + ], + "score": 0.78, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 262, + 299, + 276 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 262, + 312, + 273 + ], + "score": 0.86, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 262, + 465, + 276 + ], + "score": 1.0, + "content": "are closely related in a sense that, any", + "type": "text" + }, + { + "bbox": [ + 465, + 266, + 470, + 273 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 262, + 506, + 276 + ], + "score": 1.0, + "content": "-optimal", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 273, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 133, + 287 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 133, + 275, + 141, + 284 + ], + "score": 0.79, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 273, + 153, + 287 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 153, + 273, + 166, + 284 + ], + "score": 0.82, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 273, + 248, + 287 + ], + "score": 1.0, + "content": "corresponding to an", + "type": "text" + }, + { + "bbox": [ + 249, + 273, + 279, + 286 + ], + "score": 0.91, + "content": "( \\epsilon / \\gamma ^ { H } )", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 273, + 340, + 287 + ], + "score": 1.0, + "content": "-optimal policy", + "type": "text" + }, + { + "bbox": [ + 341, + 276, + 348, + 284 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 273, + 360, + 287 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 360, + 275, + 373, + 284 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 273, + 506, + 287 + ], + "score": 1.0, + "content": "(see section C.1 for proof). Note", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 283, + 313, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 144, + 298 + ], + "score": 1.0, + "content": "that here", + "type": "text" + }, + { + "bbox": [ + 144, + 284, + 257, + 297 + ], + "score": 0.93, + "content": "\\gamma ^ { H } = ( 1 - \\dot { 1 } / H ) ^ { \\tilde { H } } = \\mathcal { O } ( \\dot { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 283, + 313, + 298 + ], + "score": 1.0, + "content": "is a constant.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 506, + 357 + ], + "lines": [ + { + "bbox": [ + 104, + 299, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 141, + 335 + ], + "score": 1.0, + "content": "For any visited a", + "type": "text" + }, + { + "bbox": [ + 141, + 305, + 166, + 315 + ], + "score": 0.88, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 299, + 289, + 335 + ], + "score": 1.0, + "content": "g our algorithm on times, and at most", + "type": "text" + }, + { + "bbox": [ + 291, + 303, + 302, + 315 + ], + "score": 0.85, + "content": "\\bar { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 299, + 319, + 335 + ], + "score": 1.0, + "content": "for of", + "type": "text" + }, + { + "bbox": [ + 319, + 302, + 364, + 317 + ], + "score": 0.94, + "content": "\\tilde { \\mathcal O } \\big ( \\frac { 3 S A H ^ { 9 } } { \\epsilon ^ { 2 } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 299, + 370, + 335 + ], + "score": 1.0, + "content": "tt", + "type": "text" + }, + { + "bbox": [ + 376, + 299, + 483, + 335 + ], + "score": 1.0, + "content": "me steps, the starting state -optimal. If we select the p", + "type": "text" + }, + { + "bbox": [ + 484, + 307, + 494, + 316 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 299, + 507, + 335 + ], + "score": 1.0, + "content": "iscy", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 167, + 317, + 376, + 332 + ], + "spans": [ + { + "bbox": [ + 167, + 317, + 212, + 332 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal O } \\big ( { \\textstyle { \\frac { 3 S A H ^ { 8 } } { \\epsilon ^ { 2 } } } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 319, + 306, + 331 + ], + "score": 0.72, + "content": "1 / 3", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 321, + 376, + 329 + ], + "score": 0.71, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 329, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 254, + 346 + ], + "score": 1.0, + "content": "uniformly randomly from the policy", + "type": "text" + }, + { + "bbox": [ + 255, + 331, + 282, + 342 + ], + "score": 0.9, + "content": "\\pi ^ { t H + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 329, + 298, + 346 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 298, + 331, + 355, + 344 + ], + "score": 0.93, + "content": "0 \\leq t < T / H", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 329, + 457, + 346 + ], + "score": 1.0, + "content": ", with probability at least", + "type": "text" + }, + { + "bbox": [ + 457, + 331, + 473, + 344 + ], + "score": 0.44, + "content": "2 / 3", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 329, + 506, + 346 + ], + "score": 1.0, + "content": "we can", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 343, + 507, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 133, + 359 + ], + "score": 1.0, + "content": "get an", + "type": "text" + }, + { + "bbox": [ + 133, + 347, + 138, + 354 + ], + "score": 0.65, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 343, + 362, + 359 + ], + "score": 1.0, + "content": "-optimal policy. Therefore the PAC sample complexity is", + "type": "text" + }, + { + "bbox": [ + 362, + 343, + 396, + 358 + ], + "score": 0.94, + "content": "\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 343, + 446, + 359 + ], + "score": 1.0, + "content": "after hiding", + "type": "text" + }, + { + "bbox": [ + 446, + 344, + 479, + 356 + ], + "score": 0.91, + "content": "S , A , H", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 343, + 507, + 359 + ], + "score": 1.0, + "content": "terms.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 351, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 352, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 301, + 376 + ], + "score": 1.0, + "content": "On the other hand, we want to show that for any", + "type": "text" + }, + { + "bbox": [ + 301, + 362, + 311, + 371 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 358, + 352, + 376 + ], + "score": 1.0, + "content": "episodes,", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 377, + 401, + 414 + ], + "lines": [ + { + "bbox": [ + 209, + 377, + 401, + 414 + ], + "spans": [ + { + "bbox": [ + 209, + 377, + 401, + 414 + ], + "score": 0.94, + "content": "\\mathrm { R e g r e t } ( T ) = \\sum _ { k = 1 } ^ { T / H } \\left[ V ^ { * } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } ) \\right] \\propto T ^ { 1 / 2 } .", + "type": "interline_equation", + "image_path": "dfeb0feb59d694c97b8d3ad8daa0ee95cb21f9b5f8a16f568ba86cb50b2a11bd.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 377, + 401, + 395.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 209, + 395.5, + 401, + 414.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 506, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "The reason why our algorithm can have a better reduction from regret to PAC is that, after choosing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 107, + 438, + 116, + 447 + ], + "score": 0.79, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 435, + 350, + 448 + ], + "score": 1.0, + "content": ", it follows from the argument of theorem 1 that for all", + "type": "text" + }, + { + "bbox": [ + 351, + 434, + 438, + 448 + ], + "score": 0.93, + "content": "\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 435, + 506, + 448 + ], + "score": 1.0, + "content": ", the number of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 446, + 246, + 460 + ], + "spans": [ + { + "bbox": [ + 107, + 448, + 116, + 458 + ], + "score": 0.85, + "content": "\\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 446, + 246, + 460 + ], + "score": 1.0, + "content": "-suboptimal steps is bounded by", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 462, + 397, + 490 + ], + "lines": [ + { + "bbox": [ + 213, + 462, + 397, + 490 + ], + "spans": [ + { + "bbox": [ + 213, + 462, + 397, + 490 + ], + "score": 0.94, + "content": "\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right)", + "type": "interline_equation", + "image_path": "fc17daba551915b444155b637b9972b89541ce4691b2e2c14706d565e8c89847.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 213, + 462, + 397, + 490 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 171, + 507 + ], + "score": 1.0, + "content": "with probability", + "type": "text" + }, + { + "bbox": [ + 171, + 494, + 192, + 505 + ], + "score": 0.77, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 493, + 447, + 507 + ], + "score": 1.0, + "content": ". In contrast, delayed Q-learning can only give an upper bound on", + "type": "text" + }, + { + "bbox": [ + 447, + 496, + 457, + 505 + ], + "score": 0.84, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 493, + 505, + 507 + ], + "score": 1.0, + "content": "-suboptimal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 504, + 235, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 221, + 519 + ], + "score": 1.0, + "content": "steps after setting parameter", + "type": "text" + }, + { + "bbox": [ + 221, + 507, + 230, + 516 + ], + "score": 0.85, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 504, + 235, + 519 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 105, + 522, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 159, + 537 + ], + "score": 1.0, + "content": "Formally, let", + "type": "text" + }, + { + "bbox": [ + 159, + 523, + 258, + 536 + ], + "score": 0.91, + "content": "X _ { k } = V ^ { \\ast } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 521, + 320, + 537 + ], + "score": 1.0, + "content": "be the regret of", + "type": "text" + }, + { + "bbox": [ + 321, + 524, + 327, + 533 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 521, + 407, + 537 + ], + "score": 1.0, + "content": "-th episode. For any", + "type": "text" + }, + { + "bbox": [ + 407, + 524, + 415, + 533 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 521, + 432, + 537 + ], + "score": 1.0, + "content": ", set", + "type": "text" + }, + { + "bbox": [ + 433, + 522, + 487, + 536 + ], + "score": 0.94, + "content": "\\epsilon = \\sqrt { S A / T }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 521, + 505, + 537 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 534, + 360, + 554 + ], + "spans": [ + { + "bbox": [ + 107, + 535, + 189, + 549 + ], + "score": 0.93, + "content": "\\epsilon _ { 2 } = \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 534, + 209, + 554 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 209, + 536, + 293, + 551 + ], + "score": 0.91, + "content": "\\begin{array} { r } { M = \\lceil \\log _ { 2 } \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } \\rceil } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 534, + 360, + 554 + ], + "score": 1.0, + "content": ". It follows that,", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 556, + 420, + 648 + ], + "lines": [ + { + "bbox": [ + 191, + 556, + 420, + 648 + ], + "spans": [ + { + "bbox": [ + 191, + 556, + 420, + 648 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { \\operatorname { R e g r e t } ( T ) \\le T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } ( | k : \\{ X _ { k } \\ge \\epsilon _ { 2 } \\cdot 2 ^ { i - 1 } \\} | ) \\epsilon _ { 2 } \\cdot 2 ^ { i } } } \\\\ & { } & \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } \\frac { S A \\ln 1 / \\delta } { \\epsilon _ { 2 } \\cdot 2 ^ { i - 2 } } ) } \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( \\sqrt { S A T } \\ln 1 / \\delta ) } \\end{array}", + "type": "interline_equation", + "image_path": "1d7fa9ef3cc29e1e3ac2d0a868dec889d151fa6dfb8e0d2bbd8979285d9e7f25.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 556, + 420, + 571.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 191, + 571.3333333333334, + 420, + 586.6666666666667 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 191, + 586.6666666666667, + 420, + 602.0000000000001 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 191, + 602.0000000000001, + 420, + 617.3333333333335 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 191, + 617.3333333333335, + 420, + 632.6666666666669 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 191, + 632.6666666666669, + 420, + 648.0000000000002 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 504, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 171, + 666 + ], + "score": 1.0, + "content": "with probability", + "type": "text" + }, + { + "bbox": [ + 171, + 653, + 194, + 663 + ], + "score": 0.71, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 650, + 250, + 666 + ], + "score": 1.0, + "content": ". Note that the", + "type": "text" + }, + { + "bbox": [ + 250, + 651, + 260, + 663 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 650, + 353, + 666 + ], + "score": 1.0, + "content": "notation hides the poly", + "type": "text" + }, + { + "bbox": [ + 353, + 653, + 438, + 666 + ], + "score": 0.87, + "content": "( 1 / ( 1 - \\gamma ) , \\log { 1 / \\epsilon _ { 1 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 650, + 505, + 666 + ], + "score": 1.0, + "content": "which is, by our", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 663, + 267, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 169, + 677 + ], + "score": 1.0, + "content": "reduction, poly", + "type": "text" + }, + { + "bbox": [ + 169, + 664, + 262, + 676 + ], + "score": 0.9, + "content": "( H , \\log T , \\log S , \\log A )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 663, + 267, + 677 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 106, + 689, + 319, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 319, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 319, + 702 + ], + "score": 1.0, + "content": "C.1 CONNECTION BETWEEN VALUE FUNCTIONS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 275, + 723 + ], + "score": 1.0, + "content": "Recall that our MDP mapping from", + "type": "text" + }, + { + "bbox": [ + 275, + 709, + 454, + 722 + ], + "score": 0.91, + "content": "\\begin{array} { r c l } { \\mathcal { M } } & { = } & { ( S , A , H , r _ { h } ( s , a ) , p _ { h } ( s ^ { \\prime } \\mathrm { ~ ~ \\chi ~ } | \\mathrm { ~ ~ \\chi ~ } s , a ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 707, + 473, + 723 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 474, + 708, + 505, + 721 + ], + "score": 0.8, + "content": "\\begin{array} { r l } { \\bar { \\mathcal { M } } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 719, + 279, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 221, + 732 + ], + "score": 0.9, + "content": "( \\bar { S } , \\bar { A } , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 719, + 279, + 734 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 297, + 95 + ], + "score": 1.0, + "content": "C EXTENSION TO OTHER SETTINGS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 506, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "First we define a mapping from a finite horizon MDP to an infinite horizon MDP so that our algorithm", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 504, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 312, + 129 + ], + "score": 1.0, + "content": "can be applied. For an arbitrary finite horizon MDP", + "type": "text" + }, + { + "bbox": [ + 312, + 117, + 465, + 129 + ], + "score": 0.91, + "content": "\\mathcal { M } = ( S , A , H , r _ { h } ( s _ { \\perp } a ) , p _ { h } ( s ^ { \\prime } \\mid s , a ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 116, + 493, + 129 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 494, + 117, + 504, + 127 + ], + "score": 0.79, + "content": "H", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 363, + 140 + ], + "score": 1.0, + "content": "is the length of episode, the corresponding infinite horizon MDP", + "type": "text" + }, + { + "bbox": [ + 364, + 129, + 504, + 140 + ], + "score": 0.88, + "content": "\\mathcal { M } = ( S , A , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 138, + 162, + 151 + ], + "spans": [ + { + "bbox": [ + 104, + 138, + 162, + 151 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 105, + 505, + 151 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 157, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 131, + 157, + 226, + 187 + ], + "spans": [ + { + "bbox": [ + 131, + 157, + 226, + 187 + ], + "score": 0.37, + "content": "\\begin{array} { l } { { \\bullet \\ { \\bar { S } } = S \\times H , { \\bar { A } } = A ; } } \\\\ { { \\bullet \\ \\gamma = ( 1 - 1 / H ) ; } } \\end{array}", + "type": "inline_equation", + "image_path": "6163bc4f633c8d81c5759aacc60e20932b8157511f7dd0e021ff4e1df174dd54.jpg" + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 188, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 131, + 188, + 185, + 201 + ], + "score": 1.0, + "content": "• for a state", + "type": "text" + }, + { + "bbox": [ + 186, + 192, + 191, + 199 + ], + "score": 0.76, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 188, + 221, + 201 + ], + "score": 1.0, + "content": "at step", + "type": "text" + }, + { + "bbox": [ + 222, + 190, + 228, + 199 + ], + "score": 0.83, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 188, + 245, + 201 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 246, + 190, + 263, + 201 + ], + "score": 0.88, + "content": "\\bar { s } _ { s , h }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 188, + 438, + 201 + ], + "score": 1.0, + "content": "be the corresponding state. For any action", + "type": "text" + }, + { + "bbox": [ + 439, + 192, + 445, + 199 + ], + "score": 0.77, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 188, + 505, + 201 + ], + "score": 1.0, + "content": "and next state", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 198, + 507, + 216 + ], + "spans": [ + { + "bbox": [ + 142, + 201, + 151, + 211 + ], + "score": 0.84, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 198, + 183, + 216 + ], + "score": 1.0, + "content": ", define", + "type": "text" + }, + { + "bbox": [ + 183, + 201, + 303, + 213 + ], + "score": 0.91, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 198, + 322, + 216 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 323, + 201, + 466, + 213 + ], + "score": 0.91, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s ^ { \\prime } \\mid s , h )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 198, + 507, + 216 + ], + "score": 1.0, + "content": ". And for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 211, + 497, + 226 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 172, + 223 + ], + "score": 0.88, + "content": "h = H", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 211, + 189, + 226 + ], + "score": 1.0, + "content": ", set", + "type": "text" + }, + { + "bbox": [ + 190, + 213, + 247, + 225 + ], + "score": 0.91, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 211, + 265, + 226 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 265, + 213, + 384, + 225 + ], + "score": 0.92, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 211, + 482, + 226 + ], + "score": 1.0, + "content": "for a fixed starting state", + "type": "text" + }, + { + "bbox": [ + 482, + 214, + 492, + 223 + ], + "score": 0.84, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 211, + 497, + 226 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 131, + 157, + 507, + 226 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 506, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 506, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 122, + 247 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 234, + 133, + 245 + ], + "score": 0.88, + "content": "\\bar { V } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 233, + 231, + 247 + ], + "score": 1.0, + "content": "be the value function in", + "type": "text" + }, + { + "bbox": [ + 231, + 234, + 244, + 245 + ], + "score": 0.87, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 233, + 275, + 247 + ], + "score": 1.0, + "content": "at time", + "type": "text" + }, + { + "bbox": [ + 275, + 235, + 281, + 244 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 233, + 298, + 247 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 234, + 312, + 247 + ], + "score": 0.91, + "content": "V _ { h } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 233, + 398, + 247 + ], + "score": 1.0, + "content": "the value function in", + "type": "text" + }, + { + "bbox": [ + 399, + 235, + 412, + 245 + ], + "score": 0.82, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 233, + 456, + 247 + ], + "score": 1.0, + "content": "at episode", + "type": "text" + }, + { + "bbox": [ + 456, + 235, + 462, + 244 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 233, + 485, + 247 + ], + "score": 1.0, + "content": ", step", + "type": "text" + }, + { + "bbox": [ + 485, + 235, + 492, + 244 + ], + "score": 0.76, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 233, + 506, + 247 + ], + "score": 1.0, + "content": ". It", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 245, + 507, + 265 + ], + "spans": [ + { + "bbox": [ + 104, + 245, + 157, + 265 + ], + "score": 1.0, + "content": "follows that", + "type": "text" + }, + { + "bbox": [ + 158, + 246, + 264, + 263 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V _ { 1 } ^ { \\ast } ( s _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 245, + 422, + 265 + ], + "score": 1.0, + "content": ". And the policy mapping is defined as", + "type": "text" + }, + { + "bbox": [ + 422, + 249, + 489, + 262 + ], + "score": 0.93, + "content": "\\pi _ { h } ( s ) = \\bar { \\pi } ( \\bar { s } _ { s , h } )", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 245, + 507, + 265 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 262, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 133, + 276 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 133, + 264, + 141, + 273 + ], + "score": 0.78, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 262, + 152, + 276 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 153, + 262, + 165, + 273 + ], + "score": 0.85, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 262, + 268, + 276 + ], + "score": 1.0, + "content": ". Value functions in MDP", + "type": "text" + }, + { + "bbox": [ + 268, + 263, + 281, + 273 + ], + "score": 0.78, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 262, + 299, + 276 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 299, + 262, + 312, + 273 + ], + "score": 0.86, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 262, + 465, + 276 + ], + "score": 1.0, + "content": "are closely related in a sense that, any", + "type": "text" + }, + { + "bbox": [ + 465, + 266, + 470, + 273 + ], + "score": 0.68, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 262, + 506, + 276 + ], + "score": 1.0, + "content": "-optimal", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 273, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 133, + 287 + ], + "score": 1.0, + "content": "policy", + "type": "text" + }, + { + "bbox": [ + 133, + 275, + 141, + 284 + ], + "score": 0.79, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 273, + 153, + 287 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 153, + 273, + 166, + 284 + ], + "score": 0.82, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 273, + 248, + 287 + ], + "score": 1.0, + "content": "corresponding to an", + "type": "text" + }, + { + "bbox": [ + 249, + 273, + 279, + 286 + ], + "score": 0.91, + "content": "( \\epsilon / \\gamma ^ { H } )", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 273, + 340, + 287 + ], + "score": 1.0, + "content": "-optimal policy", + "type": "text" + }, + { + "bbox": [ + 341, + 276, + 348, + 284 + ], + "score": 0.75, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 273, + 360, + 287 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 360, + 275, + 373, + 284 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 273, + 506, + 287 + ], + "score": 1.0, + "content": "(see section C.1 for proof). Note", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 283, + 313, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 144, + 298 + ], + "score": 1.0, + "content": "that here", + "type": "text" + }, + { + "bbox": [ + 144, + 284, + 257, + 297 + ], + "score": 0.93, + "content": "\\gamma ^ { H } = ( 1 - \\dot { 1 } / H ) ^ { \\tilde { H } } = \\mathcal { O } ( \\dot { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 283, + 313, + 298 + ], + "score": 1.0, + "content": "is a constant.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 104, + 233, + 507, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 506, + 357 + ], + "lines": [ + { + "bbox": [ + 104, + 299, + 507, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 299, + 141, + 335 + ], + "score": 1.0, + "content": "For any visited a", + "type": "text" + }, + { + "bbox": [ + 141, + 305, + 166, + 315 + ], + "score": 0.88, + "content": "\\epsilon > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 299, + 289, + 335 + ], + "score": 1.0, + "content": "g our algorithm on times, and at most", + "type": "text" + }, + { + "bbox": [ + 291, + 303, + 302, + 315 + ], + "score": 0.85, + "content": "\\bar { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 299, + 319, + 335 + ], + "score": 1.0, + "content": "for of", + "type": "text" + }, + { + "bbox": [ + 319, + 302, + 364, + 317 + ], + "score": 0.94, + "content": "\\tilde { \\mathcal O } \\big ( \\frac { 3 S A H ^ { 9 } } { \\epsilon ^ { 2 } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 364, + 299, + 370, + 335 + ], + "score": 1.0, + "content": "tt", + "type": "text" + }, + { + "bbox": [ + 376, + 299, + 483, + 335 + ], + "score": 1.0, + "content": "me steps, the starting state -optimal. If we select the p", + "type": "text" + }, + { + "bbox": [ + 484, + 307, + 494, + 316 + ], + "score": 0.85, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 299, + 507, + 335 + ], + "score": 1.0, + "content": "iscy", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 167, + 317, + 376, + 332 + ], + "spans": [ + { + "bbox": [ + 167, + 317, + 212, + 332 + ], + "score": 0.93, + "content": "\\tilde { \\mathcal O } \\big ( { \\textstyle { \\frac { 3 S A H ^ { 8 } } { \\epsilon ^ { 2 } } } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 319, + 306, + 331 + ], + "score": 0.72, + "content": "1 / 3", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 321, + 376, + 329 + ], + "score": 0.71, + "content": "\\epsilon", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 329, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 254, + 346 + ], + "score": 1.0, + "content": "uniformly randomly from the policy", + "type": "text" + }, + { + "bbox": [ + 255, + 331, + 282, + 342 + ], + "score": 0.9, + "content": "\\pi ^ { t H + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 329, + 298, + 346 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 298, + 331, + 355, + 344 + ], + "score": 0.93, + "content": "0 \\leq t < T / H", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 329, + 457, + 346 + ], + "score": 1.0, + "content": ", with probability at least", + "type": "text" + }, + { + "bbox": [ + 457, + 331, + 473, + 344 + ], + "score": 0.44, + "content": "2 / 3", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 329, + 506, + 346 + ], + "score": 1.0, + "content": "we can", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 343, + 507, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 133, + 359 + ], + "score": 1.0, + "content": "get an", + "type": "text" + }, + { + "bbox": [ + 133, + 347, + 138, + 354 + ], + "score": 0.65, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 343, + 362, + 359 + ], + "score": 1.0, + "content": "-optimal policy. Therefore the PAC sample complexity is", + "type": "text" + }, + { + "bbox": [ + 362, + 343, + 396, + 358 + ], + "score": 0.94, + "content": "\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 343, + 446, + 359 + ], + "score": 1.0, + "content": "after hiding", + "type": "text" + }, + { + "bbox": [ + 446, + 344, + 479, + 356 + ], + "score": 0.91, + "content": "S , A , H", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 343, + 507, + 359 + ], + "score": 1.0, + "content": "terms.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 104, + 299, + 507, + 359 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 351, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 352, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 301, + 376 + ], + "score": 1.0, + "content": "On the other hand, we want to show that for any", + "type": "text" + }, + { + "bbox": [ + 301, + 362, + 311, + 371 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 358, + 352, + 376 + ], + "score": 1.0, + "content": "episodes,", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 358, + 352, + 376 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 377, + 401, + 414 + ], + "lines": [ + { + "bbox": [ + 209, + 377, + 401, + 414 + ], + "spans": [ + { + "bbox": [ + 209, + 377, + 401, + 414 + ], + "score": 0.94, + "content": "\\mathrm { R e g r e t } ( T ) = \\sum _ { k = 1 } ^ { T / H } \\left[ V ^ { * } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } ) \\right] \\propto T ^ { 1 / 2 } .", + "type": "interline_equation", + "image_path": "dfeb0feb59d694c97b8d3ad8daa0ee95cb21f9b5f8a16f568ba86cb50b2a11bd.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 377, + 401, + 395.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 209, + 395.5, + 401, + 414.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 423, + 506, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "The reason why our algorithm can have a better reduction from regret to PAC is that, after choosing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 107, + 438, + 116, + 447 + ], + "score": 0.79, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 435, + 350, + 448 + ], + "score": 1.0, + "content": ", it follows from the argument of theorem 1 that for all", + "type": "text" + }, + { + "bbox": [ + 351, + 434, + 438, + 448 + ], + "score": 0.93, + "content": "\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 435, + 506, + 448 + ], + "score": 1.0, + "content": ", the number of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 446, + 246, + 460 + ], + "spans": [ + { + "bbox": [ + 107, + 448, + 116, + 458 + ], + "score": 0.85, + "content": "\\epsilon _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 446, + 246, + 460 + ], + "score": 1.0, + "content": "-suboptimal steps is bounded by", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 422, + 506, + 460 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 462, + 397, + 490 + ], + "lines": [ + { + "bbox": [ + 213, + 462, + 397, + 490 + ], + "spans": [ + { + "bbox": [ + 213, + 462, + 397, + 490 + ], + "score": 0.94, + "content": "\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right)", + "type": "interline_equation", + "image_path": "fc17daba551915b444155b637b9972b89541ce4691b2e2c14706d565e8c89847.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 213, + 462, + 397, + 490 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 171, + 507 + ], + "score": 1.0, + "content": "with probability", + "type": "text" + }, + { + "bbox": [ + 171, + 494, + 192, + 505 + ], + "score": 0.77, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 493, + 447, + 507 + ], + "score": 1.0, + "content": ". In contrast, delayed Q-learning can only give an upper bound on", + "type": "text" + }, + { + "bbox": [ + 447, + 496, + 457, + 505 + ], + "score": 0.84, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 493, + 505, + 507 + ], + "score": 1.0, + "content": "-suboptimal", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 504, + 235, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 221, + 519 + ], + "score": 1.0, + "content": "steps after setting parameter", + "type": "text" + }, + { + "bbox": [ + 221, + 507, + 230, + 516 + ], + "score": 0.85, + "content": "\\epsilon _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 504, + 235, + 519 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 493, + 505, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 522, + 505, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 159, + 537 + ], + "score": 1.0, + "content": "Formally, let", + "type": "text" + }, + { + "bbox": [ + 159, + 523, + 258, + 536 + ], + "score": 0.91, + "content": "X _ { k } = V ^ { \\ast } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 521, + 320, + 537 + ], + "score": 1.0, + "content": "be the regret of", + "type": "text" + }, + { + "bbox": [ + 321, + 524, + 327, + 533 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 521, + 407, + 537 + ], + "score": 1.0, + "content": "-th episode. For any", + "type": "text" + }, + { + "bbox": [ + 407, + 524, + 415, + 533 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 521, + 432, + 537 + ], + "score": 1.0, + "content": ", set", + "type": "text" + }, + { + "bbox": [ + 433, + 522, + 487, + 536 + ], + "score": 0.94, + "content": "\\epsilon = \\sqrt { S A / T }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 521, + 505, + 537 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 534, + 360, + 554 + ], + "spans": [ + { + "bbox": [ + 107, + 535, + 189, + 549 + ], + "score": 0.93, + "content": "\\epsilon _ { 2 } = \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 534, + 209, + 554 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 209, + 536, + 293, + 551 + ], + "score": 0.91, + "content": "\\begin{array} { r } { M = \\lceil \\log _ { 2 } \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } \\rceil } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 534, + 360, + 554 + ], + "score": 1.0, + "content": ". It follows that,", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 521, + 505, + 554 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 556, + 420, + 648 + ], + "lines": [ + { + "bbox": [ + 191, + 556, + 420, + 648 + ], + "spans": [ + { + "bbox": [ + 191, + 556, + 420, + 648 + ], + "score": 0.94, + "content": "\\begin{array} { r l r } { { \\operatorname { R e g r e t } ( T ) \\le T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } ( | k : \\{ X _ { k } \\ge \\epsilon _ { 2 } \\cdot 2 ^ { i - 1 } \\} | ) \\epsilon _ { 2 } \\cdot 2 ^ { i } } } \\\\ & { } & \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } \\frac { S A \\ln 1 / \\delta } { \\epsilon _ { 2 } \\cdot 2 ^ { i - 2 } } ) } \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( \\sqrt { S A T } \\ln 1 / \\delta ) } \\end{array}", + "type": "interline_equation", + "image_path": "1d7fa9ef3cc29e1e3ac2d0a868dec889d151fa6dfb8e0d2bbd8979285d9e7f25.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 556, + 420, + 571.3333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 191, + 571.3333333333334, + 420, + 586.6666666666667 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 191, + 586.6666666666667, + 420, + 602.0000000000001 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 191, + 602.0000000000001, + 420, + 617.3333333333335 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 191, + 617.3333333333335, + 420, + 632.6666666666669 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 191, + 632.6666666666669, + 420, + 648.0000000000002 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 504, + 677 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 171, + 666 + ], + "score": 1.0, + "content": "with probability", + "type": "text" + }, + { + "bbox": [ + 171, + 653, + 194, + 663 + ], + "score": 0.71, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 650, + 250, + 666 + ], + "score": 1.0, + "content": ". Note that the", + "type": "text" + }, + { + "bbox": [ + 250, + 651, + 260, + 663 + ], + "score": 0.85, + "content": "\\tilde { \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 650, + 353, + 666 + ], + "score": 1.0, + "content": "notation hides the poly", + "type": "text" + }, + { + "bbox": [ + 353, + 653, + 438, + 666 + ], + "score": 0.87, + "content": "( 1 / ( 1 - \\gamma ) , \\log { 1 / \\epsilon _ { 1 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 650, + 505, + 666 + ], + "score": 1.0, + "content": "which is, by our", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 663, + 267, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 169, + 677 + ], + "score": 1.0, + "content": "reduction, poly", + "type": "text" + }, + { + "bbox": [ + 169, + 664, + 262, + 676 + ], + "score": 0.9, + "content": "( H , \\log T , \\log S , \\log A )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 663, + 267, + 677 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 650, + 505, + 677 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 689, + 319, + 700 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 319, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 319, + 702 + ], + "score": 1.0, + "content": "C.1 CONNECTION BETWEEN VALUE FUNCTIONS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 708, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 275, + 723 + ], + "score": 1.0, + "content": "Recall that our MDP mapping from", + "type": "text" + }, + { + "bbox": [ + 275, + 709, + 454, + 722 + ], + "score": 0.91, + "content": "\\begin{array} { r c l } { \\mathcal { M } } & { = } & { ( S , A , H , r _ { h } ( s , a ) , p _ { h } ( s ^ { \\prime } \\mathrm { ~ ~ \\chi ~ } | \\mathrm { ~ ~ \\chi ~ } s , a ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 707, + 473, + 723 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 474, + 708, + 505, + 721 + ], + "score": 0.8, + "content": "\\begin{array} { r l } { \\bar { \\mathcal { M } } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 719, + 279, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 221, + 732 + ], + "score": 0.9, + "content": "( \\bar { S } , \\bar { A } , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 719, + 279, + 734 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 707, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 131, + 81, + 508, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 79, + 228, + 95 + ], + "spans": [ + { + "bbox": [ + 131, + 79, + 141, + 95 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 81, + 224, + 94 + ], + "score": 0.81, + "content": "\\bar { S } = S \\times H , \\bar { A } = A", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 79, + 228, + 95 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 131, + 97, + 211, + 109 + ], + "spans": [ + { + "bbox": [ + 131, + 97, + 142, + 109 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 97, + 207, + 109 + ], + "score": 0.73, + "content": "\\gamma = ( 1 - 1 / H )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 97, + 211, + 109 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 111, + 507, + 125 + ], + "spans": [ + { + "bbox": [ + 131, + 111, + 184, + 125 + ], + "score": 1.0, + "content": "• for a state", + "type": "text" + }, + { + "bbox": [ + 184, + 115, + 190, + 122 + ], + "score": 0.76, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 111, + 219, + 125 + ], + "score": 1.0, + "content": "at step", + "type": "text" + }, + { + "bbox": [ + 219, + 113, + 226, + 122 + ], + "score": 0.78, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 111, + 242, + 125 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 242, + 113, + 259, + 124 + ], + "score": 0.88, + "content": "\\bar { s } _ { s , h }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 111, + 429, + 125 + ], + "score": 1.0, + "content": "be the corresponding state. For any action", + "type": "text" + }, + { + "bbox": [ + 429, + 115, + 436, + 122 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 111, + 494, + 125 + ], + "score": 1.0, + "content": "and next state", + "type": "text" + }, + { + "bbox": [ + 494, + 113, + 502, + 122 + ], + "score": 0.85, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 111, + 507, + 125 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 122, + 509, + 138 + ], + "spans": [ + { + "bbox": [ + 140, + 122, + 170, + 138 + ], + "score": 1.0, + "content": "define", + "type": "text" + }, + { + "bbox": [ + 170, + 123, + 289, + 136 + ], + "score": 0.9, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 122, + 308, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 309, + 124, + 433, + 136 + ], + "score": 0.91, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s , h )", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 122, + 472, + 138 + ], + "score": 1.0, + "content": ". And for", + "type": "text" + }, + { + "bbox": [ + 473, + 124, + 502, + 135 + ], + "score": 0.89, + "content": "h = H", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 122, + 509, + 138 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 134, + 464, + 149 + ], + "spans": [ + { + "bbox": [ + 140, + 134, + 156, + 149 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 156, + 136, + 214, + 147 + ], + "score": 0.92, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 134, + 232, + 149 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 232, + 136, + 352, + 147 + ], + "score": 0.89, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 134, + 449, + 149 + ], + "score": 1.0, + "content": "for a fixed starting state", + "type": "text" + }, + { + "bbox": [ + 450, + 137, + 459, + 146 + ], + "score": 0.84, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 134, + 464, + 149 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 108, + 155, + 507, + 190 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 169, + 170 + ], + "score": 1.0, + "content": "For a trajectory", + "type": "text" + }, + { + "bbox": [ + 169, + 156, + 284, + 169 + ], + "score": 0.87, + "content": "\\left\\{ { \\left( { \\bar { s } } _ { s _ { 1 } , 1 } , { \\bar { a } } _ { 1 } \\right) } , { \\left( { \\bar { s } } _ { s _ { 2 } , 2 } , { \\bar { a } } _ { 2 } \\right) } , \\cdot \\cdot \\cdot \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 155, + 295, + 170 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 295, + 156, + 308, + 167 + ], + "score": 0.85, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 155, + 325, + 170 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 325, + 156, + 420, + 169 + ], + "score": 0.92, + "content": "\\{ ( s _ { 1 } , a _ { 1 } ) , ( s _ { 2 } , a _ { 2 } ) , \\cdot \\cdot \\cdot \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 155, + 506, + 170 + ], + "score": 1.0, + "content": "be the corresponding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 157, + 180 + ], + "score": 1.0, + "content": "trajectory in", + "type": "text" + }, + { + "bbox": [ + 158, + 168, + 170, + 177 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 167, + 214, + 180 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + }, + { + "bbox": [ + 214, + 169, + 226, + 177 + ], + "score": 0.84, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 167, + 355, + 180 + ], + "score": 1.0, + "content": "has a unique fixed starting state", + "type": "text" + }, + { + "bbox": [ + 355, + 170, + 365, + 179 + ], + "score": 0.87, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 167, + 441, + 180 + ], + "score": 1.0, + "content": ", which means that", + "type": "text" + }, + { + "bbox": [ + 441, + 169, + 490, + 180 + ], + "score": 0.9, + "content": "s _ { t H + 1 } = s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 178, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 119, + 191 + ], + "score": 1.0, + "content": "all", + "type": "text" + }, + { + "bbox": [ + 119, + 179, + 142, + 190 + ], + "score": 0.89, + "content": "t \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 179, + 290, + 191 + ], + "score": 1.0, + "content": ". Denote the corresponding policy of", + "type": "text" + }, + { + "bbox": [ + 290, + 178, + 300, + 189 + ], + "score": 0.87, + "content": "\\bar { \\pi } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 179, + 312, + 191 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 313, + 178, + 323, + 189 + ], + "score": 0.87, + "content": "\\pi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 179, + 480, + 191 + ], + "score": 1.0, + "content": "(may be non-stationary), then we have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 194, + 514, + 245 + ], + "lines": [ + { + "bbox": [ + 111, + 194, + 514, + 245 + ], + "spans": [ + { + "bbox": [ + 111, + 194, + 514, + 245 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\bar { \\gamma } ^ { \\pi ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\mathbb { E } \\left[ \\bar { r } ( \\bar { s } _ { s _ { 1 } , 1 } , \\bar { a } _ { 1 } ) + \\gamma \\bar { r } ( \\bar { s } _ { s _ { 2 } , 2 } , \\bar { a } _ { 2 } ) + \\cdot \\cdot \\cdot + \\gamma ^ { H - 1 } \\bar { r } ( \\bar { s } _ { s _ { H - 1 } , H - 1 } , \\bar { a } _ { H - 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H - 1 } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } \\mathbb { E } \\left[ r _ { 1 } ( s _ { 1 } , a _ { 1 } ) + r _ { 2 } ( s _ { 2 } , a _ { 2 } ) + \\cdot \\cdot \\cdot + r _ { H - 1 } ( s _ { H - 1 } , a _ { H - 1 } ) + \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi + H } ( \\bar { s } _ { s _ { 1 } , 1 } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "52704ae56896c370e8ec3e398577b25911ebb487e2237bb93613270924e4d24c.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 111, + 194, + 514, + 211.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 211.0, + 514, + 228.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 111, + 228.0, + 514, + 245.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 249, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 100, + 245, + 510, + 270 + ], + "spans": [ + { + "bbox": [ + 100, + 245, + 217, + 270 + ], + "score": 1.0, + "content": "Then for a stationary policy", + "type": "text" + }, + { + "bbox": [ + 217, + 254, + 224, + 262 + ], + "score": 0.8, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 245, + 295, + 270 + ], + "score": 1.0, + "content": ", we can conclude", + "type": "text" + }, + { + "bbox": [ + 296, + 249, + 404, + 267 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\pi } ( s _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 245, + 510, + 270 + ], + "score": 1.0, + "content": ". Since the optimal policy", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 266, + 321, + 286 + ], + "spans": [ + { + "bbox": [ + 107, + 270, + 118, + 279 + ], + "score": 0.85, + "content": "\\bar { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 266, + 207, + 286 + ], + "score": 1.0, + "content": "is stationary, we have", + "type": "text" + }, + { + "bbox": [ + 208, + 266, + 315, + 284 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\ast } ( s _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 266, + 321, + 286 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 287, + 322, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 286, + 322, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 164, + 301 + ], + "score": 1.0, + "content": "By definition,", + "type": "text" + }, + { + "bbox": [ + 164, + 289, + 171, + 298 + ], + "score": 0.79, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 286, + 181, + 301 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 182, + 290, + 187, + 298 + ], + "score": 0.73, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 286, + 270, + 301 + ], + "score": 1.0, + "content": "-optimal at time step", + "type": "text" + }, + { + "bbox": [ + 270, + 289, + 275, + 298 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 286, + 322, + 301 + ], + "score": 1.0, + "content": "means that", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 303, + 364, + 320 + ], + "lines": [ + { + "bbox": [ + 246, + 303, + 364, + 320 + ], + "spans": [ + { + "bbox": [ + 246, + 303, + 364, + 320 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon . } \\end{array}", + "type": "interline_equation", + "image_path": "f086319de63ecc59068ffa365e9e62e72d387018083ec5144b0aa425136c465f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 246, + 303, + 364, + 320 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 164, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 322, + 165, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 165, + 336 + ], + "score": 1.0, + "content": "It follows that", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 336, + 432, + 353 + ], + "lines": [ + { + "bbox": [ + 177, + 336, + 432, + 353 + ], + "spans": [ + { + "bbox": [ + 177, + 336, + 432, + 353 + ], + "score": 0.88, + "content": "\\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon ,", + "type": "interline_equation", + "image_path": "5cce7a222551bc6274bbc52ffdb272aeca3dced98abbbc8f80d2fab7858deadf.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 177, + 336, + 432, + 353 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 132, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 133, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 133, + 370 + ], + "score": 1.0, + "content": "hence", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 371, + 504, + 387 + ], + "lines": [ + { + "bbox": [ + 111, + 371, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 111, + 371, + 504, + 387 + ], + "score": 0.86, + "content": "^ { H } V ^ { \\pi ^ { \\varepsilon } } ( s _ { 1 } ) \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) + \\gamma ^ { H } ( \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) ) - \\epsilon \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon .", + "type": "interline_equation", + "image_path": "a3a5a1e7d68659f678d45d1177dbe36c95439ce8dceba0366c8e7ffef18cce5d.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 111, + 371, + 504, + 387 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 183, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 183, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 183, + 403 + ], + "score": 1.0, + "content": "Therefore we have", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 404, + 423, + 431 + ], + "lines": [ + { + "bbox": [ + 187, + 404, + 423, + 431 + ], + "spans": [ + { + "bbox": [ + 187, + 404, + 423, + 431 + ], + "score": 0.93, + "content": "{ V ^ { \\pi } } ^ { t } ( s _ { 1 } ) \\ge \\frac { 1 - \\gamma ^ { H } } { \\gamma ^ { H } } \\bar { V } ^ { * } ( \\bar { s } _ { s 1 , 1 } ) - \\epsilon / \\gamma ^ { H } = V ^ { * } ( s _ { 1 } ) - \\epsilon / \\gamma ^ { H } ,", + "type": "interline_equation", + "image_path": "1f1b44222a0c1087458d2ccf43ee9da0e84754c0602268ed6cbc91b7c6c5c549.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 187, + 404, + 423, + 431 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 306, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 307, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 178, + 451 + ], + "score": 1.0, + "content": "which means that", + "type": "text" + }, + { + "bbox": [ + 178, + 436, + 189, + 446 + ], + "score": 0.88, + "content": "\\pi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 433, + 212, + 451 + ], + "score": 1.0, + "content": "is an", + "type": "text" + }, + { + "bbox": [ + 212, + 435, + 242, + 448 + ], + "score": 0.91, + "content": "( \\epsilon / \\gamma ^ { H } )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 433, + 307, + 451 + ], + "score": 1.0, + "content": "-optimal policy.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 106, + 462, + 372, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 373, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 373, + 480 + ], + "score": 1.0, + "content": "D A HARD INSTANCE FOR DELAYED Q-LEARNING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 108, + 488, + 456, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 456, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 456, + 504 + ], + "score": 1.0, + "content": "In this section, we prove Theorem 2 regarding the performance of Delayed Q-learning.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 506, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 353, + 516 + ], + "score": 1.0, + "content": "Theorem 2. There exists a family of MDPs with constant", + "type": "text" + }, + { + "bbox": [ + 354, + 503, + 362, + 513 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 501, + 383, + 516 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 504, + 392, + 513 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 501, + 505, + 516 + ], + "score": 1.0, + "content": ", in which with probability", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 514, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 129, + 528 + ], + "score": 0.71, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 515, + 169, + 532 + ], + "score": 1.0, + "content": ", Delayed", + "type": "text" + }, + { + "bbox": [ + 170, + 518, + 178, + 529 + ], + "score": 0.71, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 515, + 393, + 532 + ], + "score": 1.0, + "content": "-learning incurs sample complexity of exploration of", + "type": "text" + }, + { + "bbox": [ + 393, + 514, + 441, + 533 + ], + "score": 0.94, + "content": "\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 516, + 505, + 530 + ], + "score": 1.0, + "content": ", assuming that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 531, + 169, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 166, + 545 + ], + "score": 0.91, + "content": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 531, + 169, + 544 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 243, + 559, + 367, + 657 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 243, + 559, + 367, + 657 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 243, + 559, + 367, + 657 + ], + "spans": [ + { + "bbox": [ + 243, + 559, + 367, + 657 + ], + "score": 0.96, + "type": "image", + "image_path": "d5ea7466679255528080a8e623d46ab5efe02c60983c67939e2c04d8974b3167.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 559, + 367, + 608.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 243, + 608.0, + 367, + 657.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 667, + 506, + 690 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "Figure 1: The MDP family. Actions are denoted by arrows. Actions with red color have reward 1,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 678, + 205, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 205, + 691 + ], + "score": 1.0, + "content": "and reward 0 otherwise.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 175, + 713 + ], + "score": 1.0, + "content": "Proof. For each", + "type": "text" + }, + { + "bbox": [ + 175, + 698, + 227, + 711 + ], + "score": 0.92, + "content": "0 < \\epsilon < \\frac { 1 } { 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 696, + 484, + 713 + ], + "score": 1.0, + "content": ", consider the following MDP (see also Fig. 1): state space is", + "type": "text" + }, + { + "bbox": [ + 484, + 699, + 505, + 709 + ], + "score": 0.86, + "content": "{ \\boldsymbol { s } } =", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 140, + 721 + ], + "score": 0.92, + "content": "\\{ a , b , c \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 709, + 213, + 723 + ], + "score": 1.0, + "content": "while action set is", + "type": "text" + }, + { + "bbox": [ + 214, + 710, + 261, + 721 + ], + "score": 0.91, + "content": "\\overset { \\vartriangle } { \\mathcal { A } } = \\{ \\ v { x } , \\ v { y } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 709, + 369, + 723 + ], + "score": 1.0, + "content": "; transition probabilities are", + "type": "text" + }, + { + "bbox": [ + 370, + 710, + 450, + 721 + ], + "score": 0.73, + "content": "P ( b | a , y ) = 1 - 1 0 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 709, + 455, + 723 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 455, + 710, + 505, + 722 + ], + "score": 0.61, + "content": "P ( c | a , y ) =", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 720, + 446, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 125, + 733 + ], + "score": 1.0, + "content": "10\u000f,", + "type": "text" + }, + { + "bbox": [ + 126, + 721, + 182, + 732 + ], + "score": 0.91, + "content": "P ( b | a , x ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 720, + 186, + 733 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 186, + 721, + 289, + 732 + ], + "score": 0.88, + "content": "P ( a | b , \\cdot ) = P ( a | c , \\cdot ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 720, + 396, + 733 + ], + "score": 1.0, + "content": ". Rewards are all 1, except", + "type": "text" + }, + { + "bbox": [ + 396, + 721, + 442, + 732 + ], + "score": 0.92, + "content": "R ( c , \\cdot ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 720, + 446, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 294, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 131, + 81, + 508, + 148 + ], + "lines": [ + { + "bbox": [ + 131, + 79, + 228, + 95 + ], + "spans": [ + { + "bbox": [ + 131, + 79, + 141, + 95 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 81, + 224, + 94 + ], + "score": 0.81, + "content": "\\bar { S } = S \\times H , \\bar { A } = A", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 79, + 228, + 95 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 97, + 211, + 109 + ], + "spans": [ + { + "bbox": [ + 131, + 97, + 142, + 109 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 97, + 207, + 109 + ], + "score": 0.73, + "content": "\\gamma = ( 1 - 1 / H )", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 97, + 211, + 109 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 111, + 507, + 125 + ], + "spans": [ + { + "bbox": [ + 131, + 111, + 184, + 125 + ], + "score": 1.0, + "content": "• for a state", + "type": "text" + }, + { + "bbox": [ + 184, + 115, + 190, + 122 + ], + "score": 0.76, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 111, + 219, + 125 + ], + "score": 1.0, + "content": "at step", + "type": "text" + }, + { + "bbox": [ + 219, + 113, + 226, + 122 + ], + "score": 0.78, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 111, + 242, + 125 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 242, + 113, + 259, + 124 + ], + "score": 0.88, + "content": "\\bar { s } _ { s , h }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 111, + 429, + 125 + ], + "score": 1.0, + "content": "be the corresponding state. For any action", + "type": "text" + }, + { + "bbox": [ + 429, + 115, + 436, + 122 + ], + "score": 0.75, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 111, + 494, + 125 + ], + "score": 1.0, + "content": "and next state", + "type": "text" + }, + { + "bbox": [ + 494, + 113, + 502, + 122 + ], + "score": 0.85, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 111, + 507, + 125 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 140, + 122, + 509, + 138 + ], + "spans": [ + { + "bbox": [ + 140, + 122, + 170, + 138 + ], + "score": 1.0, + "content": "define", + "type": "text" + }, + { + "bbox": [ + 170, + 123, + 289, + 136 + ], + "score": 0.9, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 122, + 308, + 138 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 309, + 124, + 433, + 136 + ], + "score": 0.91, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s , h )", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 122, + 472, + 138 + ], + "score": 1.0, + "content": ". And for", + "type": "text" + }, + { + "bbox": [ + 473, + 124, + 502, + 135 + ], + "score": 0.89, + "content": "h = H", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 122, + 509, + 138 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 134, + 464, + 149 + ], + "spans": [ + { + "bbox": [ + 140, + 134, + 156, + 149 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 156, + 136, + 214, + 147 + ], + "score": 0.92, + "content": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 134, + 232, + 149 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 232, + 136, + 352, + 147 + ], + "score": 0.89, + "content": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 134, + 449, + 149 + ], + "score": 1.0, + "content": "for a fixed starting state", + "type": "text" + }, + { + "bbox": [ + 450, + 137, + 459, + 146 + ], + "score": 0.84, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 134, + 464, + 149 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + } + ], + "index": 2, + "bbox_fs": [ + 131, + 79, + 509, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 155, + 507, + 190 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 506, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 169, + 170 + ], + "score": 1.0, + "content": "For a trajectory", + "type": "text" + }, + { + "bbox": [ + 169, + 156, + 284, + 169 + ], + "score": 0.87, + "content": "\\left\\{ { \\left( { \\bar { s } } _ { s _ { 1 } , 1 } , { \\bar { a } } _ { 1 } \\right) } , { \\left( { \\bar { s } } _ { s _ { 2 } , 2 } , { \\bar { a } } _ { 2 } \\right) } , \\cdot \\cdot \\cdot \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 155, + 295, + 170 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 295, + 156, + 308, + 167 + ], + "score": 0.85, + "content": "\\bar { \\mathcal { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 155, + 325, + 170 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 325, + 156, + 420, + 169 + ], + "score": 0.92, + "content": "\\{ ( s _ { 1 } , a _ { 1 } ) , ( s _ { 2 } , a _ { 2 } ) , \\cdot \\cdot \\cdot \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 155, + 506, + 170 + ], + "score": 1.0, + "content": "be the corresponding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 157, + 180 + ], + "score": 1.0, + "content": "trajectory in", + "type": "text" + }, + { + "bbox": [ + 158, + 168, + 170, + 177 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 167, + 214, + 180 + ], + "score": 1.0, + "content": ". Note that", + "type": "text" + }, + { + "bbox": [ + 214, + 169, + 226, + 177 + ], + "score": 0.84, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 167, + 355, + 180 + ], + "score": 1.0, + "content": "has a unique fixed starting state", + "type": "text" + }, + { + "bbox": [ + 355, + 170, + 365, + 179 + ], + "score": 0.87, + "content": "s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 167, + 441, + 180 + ], + "score": 1.0, + "content": ", which means that", + "type": "text" + }, + { + "bbox": [ + 441, + 169, + 490, + 180 + ], + "score": 0.9, + "content": "s _ { t H + 1 } = s _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 178, + 480, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 179, + 119, + 191 + ], + "score": 1.0, + "content": "all", + "type": "text" + }, + { + "bbox": [ + 119, + 179, + 142, + 190 + ], + "score": 0.89, + "content": "t \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 179, + 290, + 191 + ], + "score": 1.0, + "content": ". Denote the corresponding policy of", + "type": "text" + }, + { + "bbox": [ + 290, + 178, + 300, + 189 + ], + "score": 0.87, + "content": "\\bar { \\pi } ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 179, + 312, + 191 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 313, + 178, + 323, + 189 + ], + "score": 0.87, + "content": "\\pi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 179, + 480, + 191 + ], + "score": 1.0, + "content": "(may be non-stationary), then we have", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 155, + 506, + 191 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 194, + 514, + 245 + ], + "lines": [ + { + "bbox": [ + 111, + 194, + 514, + 245 + ], + "spans": [ + { + "bbox": [ + 111, + 194, + 514, + 245 + ], + "score": 0.9, + "content": "\\begin{array} { r l } & { \\bar { \\gamma } ^ { \\pi ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\mathbb { E } \\left[ \\bar { r } ( \\bar { s } _ { s _ { 1 } , 1 } , \\bar { a } _ { 1 } ) + \\gamma \\bar { r } ( \\bar { s } _ { s _ { 2 } , 2 } , \\bar { a } _ { 2 } ) + \\cdot \\cdot \\cdot + \\gamma ^ { H - 1 } \\bar { r } ( \\bar { s } _ { s _ { H - 1 } , H - 1 } , \\bar { a } _ { H - 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H - 1 } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } \\mathbb { E } \\left[ r _ { 1 } ( s _ { 1 } , a _ { 1 } ) + r _ { 2 } ( s _ { 2 } , a _ { 2 } ) + \\cdot \\cdot \\cdot + r _ { H - 1 } ( s _ { H - 1 } , a _ { H - 1 } ) + \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi + H } ( \\bar { s } _ { s _ { 1 } , 1 } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "52704ae56896c370e8ec3e398577b25911ebb487e2237bb93613270924e4d24c.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 111, + 194, + 514, + 211.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 111, + 211.0, + 514, + 228.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 111, + 228.0, + 514, + 245.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 249, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 100, + 245, + 510, + 270 + ], + "spans": [ + { + "bbox": [ + 100, + 245, + 217, + 270 + ], + "score": 1.0, + "content": "Then for a stationary policy", + "type": "text" + }, + { + "bbox": [ + 217, + 254, + 224, + 262 + ], + "score": 0.8, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 245, + 295, + 270 + ], + "score": 1.0, + "content": ", we can conclude", + "type": "text" + }, + { + "bbox": [ + 296, + 249, + 404, + 267 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\pi } ( s _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 245, + 510, + 270 + ], + "score": 1.0, + "content": ". Since the optimal policy", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 266, + 321, + 286 + ], + "spans": [ + { + "bbox": [ + 107, + 270, + 118, + 279 + ], + "score": 0.85, + "content": "\\bar { \\pi } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 266, + 207, + 286 + ], + "score": 1.0, + "content": "is stationary, we have", + "type": "text" + }, + { + "bbox": [ + 208, + 266, + 315, + 284 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\ast } ( s _ { 1 } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 266, + 321, + 286 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 100, + 245, + 510, + 286 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 287, + 322, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 286, + 322, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 164, + 301 + ], + "score": 1.0, + "content": "By definition,", + "type": "text" + }, + { + "bbox": [ + 164, + 289, + 171, + 298 + ], + "score": 0.79, + "content": "\\bar { \\pi }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 286, + 181, + 301 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 182, + 290, + 187, + 298 + ], + "score": 0.73, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 286, + 270, + 301 + ], + "score": 1.0, + "content": "-optimal at time step", + "type": "text" + }, + { + "bbox": [ + 270, + 289, + 275, + 298 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 286, + 322, + 301 + ], + "score": 1.0, + "content": "means that", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 286, + 322, + 301 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 303, + 364, + 320 + ], + "lines": [ + { + "bbox": [ + 246, + 303, + 364, + 320 + ], + "spans": [ + { + "bbox": [ + 246, + 303, + 364, + 320 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon . } \\end{array}", + "type": "interline_equation", + "image_path": "f086319de63ecc59068ffa365e9e62e72d387018083ec5144b0aa425136c465f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 246, + 303, + 364, + 320 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 164, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 322, + 165, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 165, + 336 + ], + "score": 1.0, + "content": "It follows that", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 322, + 165, + 336 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 336, + 432, + 353 + ], + "lines": [ + { + "bbox": [ + 177, + 336, + 432, + 353 + ], + "spans": [ + { + "bbox": [ + 177, + 336, + 432, + 353 + ], + "score": 0.88, + "content": "\\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon ,", + "type": "interline_equation", + "image_path": "5cce7a222551bc6274bbc52ffdb272aeca3dced98abbbc8f80d2fab7858deadf.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 177, + 336, + 432, + 353 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 132, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 133, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 133, + 370 + ], + "score": 1.0, + "content": "hence", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 356, + 133, + 370 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 371, + 504, + 387 + ], + "lines": [ + { + "bbox": [ + 111, + 371, + 504, + 387 + ], + "spans": [ + { + "bbox": [ + 111, + 371, + 504, + 387 + ], + "score": 0.86, + "content": "^ { H } V ^ { \\pi ^ { \\varepsilon } } ( s _ { 1 } ) \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) + \\gamma ^ { H } ( \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) ) - \\epsilon \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon .", + "type": "interline_equation", + "image_path": "a3a5a1e7d68659f678d45d1177dbe36c95439ce8dceba0366c8e7ffef18cce5d.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 111, + 371, + 504, + 387 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 390, + 183, + 402 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 183, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 183, + 403 + ], + "score": 1.0, + "content": "Therefore we have", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 389, + 183, + 403 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 404, + 423, + 431 + ], + "lines": [ + { + "bbox": [ + 187, + 404, + 423, + 431 + ], + "spans": [ + { + "bbox": [ + 187, + 404, + 423, + 431 + ], + "score": 0.93, + "content": "{ V ^ { \\pi } } ^ { t } ( s _ { 1 } ) \\ge \\frac { 1 - \\gamma ^ { H } } { \\gamma ^ { H } } \\bar { V } ^ { * } ( \\bar { s } _ { s 1 , 1 } ) - \\epsilon / \\gamma ^ { H } = V ^ { * } ( s _ { 1 } ) - \\epsilon / \\gamma ^ { H } ,", + "type": "interline_equation", + "image_path": "1f1b44222a0c1087458d2ccf43ee9da0e84754c0602268ed6cbc91b7c6c5c549.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 187, + 404, + 423, + 431 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 306, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 307, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 178, + 451 + ], + "score": 1.0, + "content": "which means that", + "type": "text" + }, + { + "bbox": [ + 178, + 436, + 189, + 446 + ], + "score": 0.88, + "content": "\\pi ^ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 433, + 212, + 451 + ], + "score": 1.0, + "content": "is an", + "type": "text" + }, + { + "bbox": [ + 212, + 435, + 242, + 448 + ], + "score": 0.91, + "content": "( \\epsilon / \\gamma ^ { H } )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 433, + 307, + 451 + ], + "score": 1.0, + "content": "-optimal policy.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 433, + 307, + 451 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 462, + 372, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 373, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 373, + 480 + ], + "score": 1.0, + "content": "D A HARD INSTANCE FOR DELAYED Q-LEARNING", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 108, + 488, + 456, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 456, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 456, + 504 + ], + "score": 1.0, + "content": "In this section, we prove Theorem 2 regarding the performance of Delayed Q-learning.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 485, + 456, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 502, + 506, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 353, + 516 + ], + "score": 1.0, + "content": "Theorem 2. There exists a family of MDPs with constant", + "type": "text" + }, + { + "bbox": [ + 354, + 503, + 362, + 513 + ], + "score": 0.78, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 501, + 383, + 516 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 383, + 504, + 392, + 513 + ], + "score": 0.71, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 501, + 505, + 516 + ], + "score": 1.0, + "content": ", in which with probability", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 514, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 129, + 528 + ], + "score": 0.71, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 515, + 169, + 532 + ], + "score": 1.0, + "content": ", Delayed", + "type": "text" + }, + { + "bbox": [ + 170, + 518, + 178, + 529 + ], + "score": 0.71, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 515, + 393, + 532 + ], + "score": 1.0, + "content": "-learning incurs sample complexity of exploration of", + "type": "text" + }, + { + "bbox": [ + 393, + 514, + 441, + 533 + ], + "score": 0.94, + "content": "\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 516, + 505, + 530 + ], + "score": 1.0, + "content": ", assuming that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 531, + 169, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 166, + 545 + ], + "score": 0.91, + "content": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 531, + 169, + 544 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 501, + 505, + 545 + ] + }, + { + "type": "image", + "bbox": [ + 243, + 559, + 367, + 657 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 243, + 559, + 367, + 657 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 243, + 559, + 367, + 657 + ], + "spans": [ + { + "bbox": [ + 243, + 559, + 367, + 657 + ], + "score": 0.96, + "type": "image", + "image_path": "d5ea7466679255528080a8e623d46ab5efe02c60983c67939e2c04d8974b3167.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 559, + 367, + 608.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 243, + 608.0, + 367, + 657.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 667, + 506, + 690 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "Figure 1: The MDP family. Actions are denoted by arrows. Actions with red color have reward 1,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 678, + 205, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 205, + 691 + ], + "score": 1.0, + "content": "and reward 0 otherwise.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 105, + 696, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 175, + 713 + ], + "score": 1.0, + "content": "Proof. For each", + "type": "text" + }, + { + "bbox": [ + 175, + 698, + 227, + 711 + ], + "score": 0.92, + "content": "0 < \\epsilon < \\frac { 1 } { 1 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 696, + 484, + 713 + ], + "score": 1.0, + "content": ", consider the following MDP (see also Fig. 1): state space is", + "type": "text" + }, + { + "bbox": [ + 484, + 699, + 505, + 709 + ], + "score": 0.86, + "content": "{ \\boldsymbol { s } } =", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 710, + 140, + 721 + ], + "score": 0.92, + "content": "\\{ a , b , c \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 709, + 213, + 723 + ], + "score": 1.0, + "content": "while action set is", + "type": "text" + }, + { + "bbox": [ + 214, + 710, + 261, + 721 + ], + "score": 0.91, + "content": "\\overset { \\vartriangle } { \\mathcal { A } } = \\{ \\ v { x } , \\ v { y } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 709, + 369, + 723 + ], + "score": 1.0, + "content": "; transition probabilities are", + "type": "text" + }, + { + "bbox": [ + 370, + 710, + 450, + 721 + ], + "score": 0.73, + "content": "P ( b | a , y ) = 1 - 1 0 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 709, + 455, + 723 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 455, + 710, + 505, + 722 + ], + "score": 0.61, + "content": "P ( c | a , y ) =", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 720, + 446, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 125, + 733 + ], + "score": 1.0, + "content": "10\u000f,", + "type": "text" + }, + { + "bbox": [ + 126, + 721, + 182, + 732 + ], + "score": 0.91, + "content": "P ( b | a , x ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 720, + 186, + 733 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 186, + 721, + 289, + 732 + ], + "score": 0.88, + "content": "P ( a | b , \\cdot ) = P ( a | c , \\cdot ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 720, + 396, + 733 + ], + "score": 1.0, + "content": ". Rewards are all 1, except", + "type": "text" + }, + { + "bbox": [ + 396, + 721, + 442, + 732 + ], + "score": 0.92, + "content": "R ( c , \\cdot ) = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 720, + 446, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 696, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 409, + 97 + ], + "score": 1.0, + "content": "Assume that Delayed Q-learning is called for this MDP starting from state", + "type": "text" + }, + { + "bbox": [ + 410, + 85, + 416, + 93 + ], + "score": 0.72, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 81, + 477, + 97 + ], + "score": 1.0, + "content": ", with discount", + "type": "text" + }, + { + "bbox": [ + 477, + 81, + 504, + 96 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\gamma > \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 95, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 193, + 109 + ], + "score": 1.0, + "content": "and precision set as", + "type": "text" + }, + { + "bbox": [ + 194, + 99, + 199, + 106 + ], + "score": 0.49, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 96, + 257, + 109 + ], + "score": 1.0, + "content": ". Denote the", + "type": "text" + }, + { + "bbox": [ + 257, + 97, + 267, + 108 + ], + "score": 0.84, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 96, + 432, + 109 + ], + "score": 1.0, + "content": "value maintained by the algorithm by", + "type": "text" + }, + { + "bbox": [ + 432, + 95, + 441, + 108 + ], + "score": 0.86, + "content": "\\hat { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 96, + 505, + 109 + ], + "score": 1.0, + "content": ". Without loss", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 108, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 104, + 108, + 371, + 122 + ], + "score": 1.0, + "content": "of generality, assume that the initial tie-breaking favors action", + "type": "text" + }, + { + "bbox": [ + 371, + 111, + 378, + 121 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 108, + 453, + 122 + ], + "score": 1.0, + "content": "when comparing", + "type": "text" + }, + { + "bbox": [ + 453, + 108, + 485, + 121 + ], + "score": 0.93, + "content": "\\hat { Q } ( a , x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 108, + 506, + 122 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 107, + 120, + 138, + 133 + ], + "score": 0.91, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 120, + 225, + 135 + ], + "score": 1.0, + "content": ". In that case, unless", + "type": "text" + }, + { + "bbox": [ + 225, + 120, + 257, + 133 + ], + "score": 0.95, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 120, + 426, + 135 + ], + "score": 1.0, + "content": "is updated, the agent will always choose", + "type": "text" + }, + { + "bbox": [ + 427, + 123, + 434, + 133 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 120, + 467, + 135 + ], + "score": 1.0, + "content": "in state", + "type": "text" + }, + { + "bbox": [ + 468, + 124, + 474, + 132 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 120, + 506, + 135 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 107, + 133, + 232, + 145 + ], + "score": 0.91, + "content": "Q ( a , x ) - Q ( a , y ) = 1 0 \\epsilon \\gamma > \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 132, + 332, + 146 + ], + "score": 1.0, + "content": "for any policy, choosing", + "type": "text" + }, + { + "bbox": [ + 333, + 135, + 339, + 144 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 132, + 371, + 146 + ], + "score": 1.0, + "content": "at state", + "type": "text" + }, + { + "bbox": [ + 371, + 135, + 378, + 143 + ], + "score": 0.66, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "implies that the timestep is not", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 107, + 146, + 112, + 154 + ], + "score": 0.59, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "-optimal. In other words, sample complexity for exploration is at least the number of times the agent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 274, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 129, + 169 + ], + "score": 1.0, + "content": "visits", + "type": "text" + }, + { + "bbox": [ + 130, + 158, + 136, + 166 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 155, + 238, + 169 + ], + "score": 1.0, + "content": "before the first update of", + "type": "text" + }, + { + "bbox": [ + 238, + 154, + 269, + 168 + ], + "score": 0.93, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 155, + 274, + 169 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 506, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 259, + 189 + ], + "score": 1.0, + "content": "In the Delayed Q-learning algorithm,", + "type": "text" + }, + { + "bbox": [ + 259, + 173, + 286, + 187 + ], + "score": 0.93, + "content": "\\hat { Q } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 173, + 356, + 189 + ], + "score": 1.0, + "content": "are initialized to", + "type": "text" + }, + { + "bbox": [ + 356, + 175, + 397, + 187 + ], + "score": 0.92, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 173, + 447, + 189 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 447, + 173, + 479, + 187 + ], + "score": 0.94, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 173, + 506, + 189 + ], + "score": 1.0, + "content": "could", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 202, + 201 + ], + "score": 1.0, + "content": "only be updated if max", + "type": "text" + }, + { + "bbox": [ + 202, + 187, + 230, + 200 + ], + "score": 0.88, + "content": "\\hat { Q } ( c , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 186, + 385, + 201 + ], + "score": 1.0, + "content": "is updated (and becomes smaller than", + "type": "text" + }, + { + "bbox": [ + 386, + 188, + 429, + 200 + ], + "score": 0.9, + "content": "1 / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 186, + 505, + 201 + ], + "score": 1.0, + "content": ". According to the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 368, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 243, + 214 + ], + "score": 1.0, + "content": "algorithm, this can only happen if", + "type": "text" + }, + { + "bbox": [ + 244, + 202, + 249, + 210 + ], + "score": 0.74, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 199, + 288, + 214 + ], + "score": 1.0, + "content": "is visited", + "type": "text" + }, + { + "bbox": [ + 289, + 199, + 339, + 213 + ], + "score": 0.93, + "content": "\\begin{array} { r } { m = \\Omega \\left( \\frac { 1 } { \\epsilon ^ { 2 } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 199, + 368, + 214 + ], + "score": 1.0, + "content": "times.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 505, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 259, + 229 + ], + "score": 1.0, + "content": "However, each time the agent visits", + "type": "text" + }, + { + "bbox": [ + 259, + 220, + 266, + 227 + ], + "score": 0.74, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 217, + 345, + 229 + ], + "score": 1.0, + "content": ", there is less than", + "type": "text" + }, + { + "bbox": [ + 345, + 218, + 361, + 227 + ], + "score": 0.61, + "content": "1 0 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 217, + 475, + 229 + ], + "score": 1.0, + "content": "probability of transiting to", + "type": "text" + }, + { + "bbox": [ + 476, + 219, + 482, + 227 + ], + "score": 0.57, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 217, + 505, + 229 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 173, + 240 + ], + "score": 0.91, + "content": "t _ { 0 } = m / ( 1 0 \\epsilon C )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 227, + 205, + 242 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 206, + 227, + 270, + 241 + ], + "score": 0.92, + "content": "\\begin{array} { r } { C ^ { ' } = 3 \\ln \\frac { 1 } { \\delta } + 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 227, + 275, + 242 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 276, + 228, + 282, + 238 + ], + "score": 0.55, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 227, + 363, + 242 + ], + "score": 1.0, + "content": "is chosen such that", + "type": "text" + }, + { + "bbox": [ + 364, + 228, + 396, + 239 + ], + "score": 0.91, + "content": "C \\leq m", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 227, + 447, + 242 + ], + "score": 1.0, + "content": ". In the first", + "type": "text" + }, + { + "bbox": [ + 447, + 228, + 461, + 239 + ], + "score": 0.89, + "content": "2 t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 227, + 506, + 242 + ], + "score": 1.0, + "content": "timesteps,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 107, + 241, + 113, + 249 + ], + "score": 0.64, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 239, + 174, + 251 + ], + "score": 1.0, + "content": "will be visited", + "type": "text" + }, + { + "bbox": [ + 174, + 240, + 184, + 250 + ], + "score": 0.87, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 239, + 372, + 251 + ], + "score": 1.0, + "content": "times. By Chernoff’s bound, with probability", + "type": "text" + }, + { + "bbox": [ + 372, + 240, + 395, + 250 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 239, + 420, + 251 + ], + "score": 1.0, + "content": ", state", + "type": "text" + }, + { + "bbox": [ + 420, + 241, + 426, + 249 + ], + "score": 0.73, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "will be visited less", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 126, + 264 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 127, + 253, + 137, + 262 + ], + "score": 0.73, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 251, + 219, + 264 + ], + "score": 1.0, + "content": "times. In that case,", + "type": "text" + }, + { + "bbox": [ + 219, + 250, + 250, + 263 + ], + "score": 0.94, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 251, + 378, + 264 + ], + "score": 1.0, + "content": "will not be updated in the first", + "type": "text" + }, + { + "bbox": [ + 379, + 252, + 393, + 263 + ], + "score": 0.89, + "content": "2 t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 251, + 505, + 264 + ], + "score": 1.0, + "content": "timesteps. Therefore, with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 357, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 153, + 276 + ], + "score": 1.0, + "content": "probability", + "type": "text" + }, + { + "bbox": [ + 153, + 263, + 176, + 273 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 262, + 357, + 276 + ], + "score": 1.0, + "content": ", sample complexity of exploration is at least", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 279, + 354, + 307 + ], + "lines": [ + { + "bbox": [ + 257, + 279, + 354, + 307 + ], + "spans": [ + { + "bbox": [ + 257, + 279, + 354, + 307 + ], + "score": 0.94, + "content": "t _ { 0 } = \\Omega \\left( { \\frac { 1 } { \\epsilon ^ { 3 } \\left( \\ln { 1 / \\delta } \\right) } } \\right) .", + "type": "interline_equation", + "image_path": "446057ad76737ed78dbeb6c14c391ec04ae8a21d0dea27c95bf135aaaf7cc43f.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 257, + 279, + 354, + 307 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 380, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 378, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 132, + 328 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 132, + 312, + 192, + 325 + ], + "score": 0.92, + "content": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 308, + 270, + 328 + ], + "score": 1.0, + "content": ", it can be seen that", + "type": "text" + }, + { + "bbox": [ + 270, + 311, + 378, + 326 + ], + "score": 0.93, + "content": "\\begin{array} { r } { C = 3 \\ln \\frac { 1 } { \\delta } + 1 < \\frac { 4 } { \\epsilon ^ { 2 } } < m } \\end{array}", + "type": "inline_equation" + } + ], + "index": 16 + } + ], + "index": 16 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 494, + 313, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 496, + 315, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 496, + 315, + 505, + 324 + ], + "score": 0.998, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 409, + 97 + ], + "score": 1.0, + "content": "Assume that Delayed Q-learning is called for this MDP starting from state", + "type": "text" + }, + { + "bbox": [ + 410, + 85, + 416, + 93 + ], + "score": 0.72, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 81, + 477, + 97 + ], + "score": 1.0, + "content": ", with discount", + "type": "text" + }, + { + "bbox": [ + 477, + 81, + 504, + 96 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\gamma > \\frac { 1 } { 2 } } \\end{array}", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 95, + 505, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 193, + 109 + ], + "score": 1.0, + "content": "and precision set as", + "type": "text" + }, + { + "bbox": [ + 194, + 99, + 199, + 106 + ], + "score": 0.49, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 96, + 257, + 109 + ], + "score": 1.0, + "content": ". Denote the", + "type": "text" + }, + { + "bbox": [ + 257, + 97, + 267, + 108 + ], + "score": 0.84, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 96, + 432, + 109 + ], + "score": 1.0, + "content": "value maintained by the algorithm by", + "type": "text" + }, + { + "bbox": [ + 432, + 95, + 441, + 108 + ], + "score": 0.86, + "content": "\\hat { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 96, + 505, + 109 + ], + "score": 1.0, + "content": ". Without loss", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 108, + 506, + 122 + ], + "spans": [ + { + "bbox": [ + 104, + 108, + 371, + 122 + ], + "score": 1.0, + "content": "of generality, assume that the initial tie-breaking favors action", + "type": "text" + }, + { + "bbox": [ + 371, + 111, + 378, + 121 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 108, + 453, + 122 + ], + "score": 1.0, + "content": "when comparing", + "type": "text" + }, + { + "bbox": [ + 453, + 108, + 485, + 121 + ], + "score": 0.93, + "content": "\\hat { Q } ( a , x )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 108, + 506, + 122 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 107, + 120, + 138, + 133 + ], + "score": 0.91, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 120, + 225, + 135 + ], + "score": 1.0, + "content": ". In that case, unless", + "type": "text" + }, + { + "bbox": [ + 225, + 120, + 257, + 133 + ], + "score": 0.95, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 120, + 426, + 135 + ], + "score": 1.0, + "content": "is updated, the agent will always choose", + "type": "text" + }, + { + "bbox": [ + 427, + 123, + 434, + 133 + ], + "score": 0.79, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 120, + 467, + 135 + ], + "score": 1.0, + "content": "in state", + "type": "text" + }, + { + "bbox": [ + 468, + 124, + 474, + 132 + ], + "score": 0.71, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 120, + 506, + 135 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 107, + 133, + 232, + 145 + ], + "score": 0.91, + "content": "Q ( a , x ) - Q ( a , y ) = 1 0 \\epsilon \\gamma > \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 132, + 332, + 146 + ], + "score": 1.0, + "content": "for any policy, choosing", + "type": "text" + }, + { + "bbox": [ + 333, + 135, + 339, + 144 + ], + "score": 0.76, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 132, + 371, + 146 + ], + "score": 1.0, + "content": "at state", + "type": "text" + }, + { + "bbox": [ + 371, + 135, + 378, + 143 + ], + "score": 0.66, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "implies that the timestep is not", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 107, + 146, + 112, + 154 + ], + "score": 0.59, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "-optimal. In other words, sample complexity for exploration is at least the number of times the agent", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 274, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 129, + 169 + ], + "score": 1.0, + "content": "visits", + "type": "text" + }, + { + "bbox": [ + 130, + 158, + 136, + 166 + ], + "score": 0.73, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 155, + 238, + 169 + ], + "score": 1.0, + "content": "before the first update of", + "type": "text" + }, + { + "bbox": [ + 238, + 154, + 269, + 168 + ], + "score": 0.93, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 155, + 274, + 169 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 104, + 81, + 506, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 506, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 259, + 189 + ], + "score": 1.0, + "content": "In the Delayed Q-learning algorithm,", + "type": "text" + }, + { + "bbox": [ + 259, + 173, + 286, + 187 + ], + "score": 0.93, + "content": "\\hat { Q } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 173, + 356, + 189 + ], + "score": 1.0, + "content": "are initialized to", + "type": "text" + }, + { + "bbox": [ + 356, + 175, + 397, + 187 + ], + "score": 0.92, + "content": "1 / ( 1 - \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 173, + 447, + 189 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 447, + 173, + 479, + 187 + ], + "score": 0.94, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 173, + 506, + 189 + ], + "score": 1.0, + "content": "could", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 186, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 202, + 201 + ], + "score": 1.0, + "content": "only be updated if max", + "type": "text" + }, + { + "bbox": [ + 202, + 187, + 230, + 200 + ], + "score": 0.88, + "content": "\\hat { Q } ( c , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 186, + 385, + 201 + ], + "score": 1.0, + "content": "is updated (and becomes smaller than", + "type": "text" + }, + { + "bbox": [ + 386, + 188, + 429, + 200 + ], + "score": 0.9, + "content": "1 / ( 1 - \\gamma ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 186, + 505, + 201 + ], + "score": 1.0, + "content": ". According to the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 368, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 243, + 214 + ], + "score": 1.0, + "content": "algorithm, this can only happen if", + "type": "text" + }, + { + "bbox": [ + 244, + 202, + 249, + 210 + ], + "score": 0.74, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 199, + 288, + 214 + ], + "score": 1.0, + "content": "is visited", + "type": "text" + }, + { + "bbox": [ + 289, + 199, + 339, + 213 + ], + "score": 0.93, + "content": "\\begin{array} { r } { m = \\Omega \\left( \\frac { 1 } { \\epsilon ^ { 2 } } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 199, + 368, + 214 + ], + "score": 1.0, + "content": "times.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 173, + 506, + 214 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 505, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 259, + 229 + ], + "score": 1.0, + "content": "However, each time the agent visits", + "type": "text" + }, + { + "bbox": [ + 259, + 220, + 266, + 227 + ], + "score": 0.74, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 217, + 345, + 229 + ], + "score": 1.0, + "content": ", there is less than", + "type": "text" + }, + { + "bbox": [ + 345, + 218, + 361, + 227 + ], + "score": 0.61, + "content": "1 0 \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 217, + 475, + 229 + ], + "score": 1.0, + "content": "probability of transiting to", + "type": "text" + }, + { + "bbox": [ + 476, + 219, + 482, + 227 + ], + "score": 0.57, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 217, + 505, + 229 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 173, + 240 + ], + "score": 0.91, + "content": "t _ { 0 } = m / ( 1 0 \\epsilon C )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 227, + 205, + 242 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 206, + 227, + 270, + 241 + ], + "score": 0.92, + "content": "\\begin{array} { r } { C ^ { ' } = 3 \\ln \\frac { 1 } { \\delta } + 1 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 227, + 275, + 242 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 276, + 228, + 282, + 238 + ], + "score": 0.55, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 227, + 363, + 242 + ], + "score": 1.0, + "content": "is chosen such that", + "type": "text" + }, + { + "bbox": [ + 364, + 228, + 396, + 239 + ], + "score": 0.91, + "content": "C \\leq m", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 227, + 447, + 242 + ], + "score": 1.0, + "content": ". In the first", + "type": "text" + }, + { + "bbox": [ + 447, + 228, + 461, + 239 + ], + "score": 0.89, + "content": "2 t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 227, + 506, + 242 + ], + "score": 1.0, + "content": "timesteps,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 239, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 107, + 241, + 113, + 249 + ], + "score": 0.64, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 239, + 174, + 251 + ], + "score": 1.0, + "content": "will be visited", + "type": "text" + }, + { + "bbox": [ + 174, + 240, + 184, + 250 + ], + "score": 0.87, + "content": "t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 239, + 372, + 251 + ], + "score": 1.0, + "content": "times. By Chernoff’s bound, with probability", + "type": "text" + }, + { + "bbox": [ + 372, + 240, + 395, + 250 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 239, + 420, + 251 + ], + "score": 1.0, + "content": ", state", + "type": "text" + }, + { + "bbox": [ + 420, + 241, + 426, + 249 + ], + "score": 0.73, + "content": "c", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 239, + 505, + 251 + ], + "score": 1.0, + "content": "will be visited less", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 126, + 264 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 127, + 253, + 137, + 262 + ], + "score": 0.73, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 251, + 219, + 264 + ], + "score": 1.0, + "content": "times. In that case,", + "type": "text" + }, + { + "bbox": [ + 219, + 250, + 250, + 263 + ], + "score": 0.94, + "content": "\\hat { Q } ( a , y )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 251, + 378, + 264 + ], + "score": 1.0, + "content": "will not be updated in the first", + "type": "text" + }, + { + "bbox": [ + 379, + 252, + 393, + 263 + ], + "score": 0.89, + "content": "2 t _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 251, + 505, + 264 + ], + "score": 1.0, + "content": "timesteps. Therefore, with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 357, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 153, + 276 + ], + "score": 1.0, + "content": "probability", + "type": "text" + }, + { + "bbox": [ + 153, + 263, + 176, + 273 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 262, + 357, + 276 + ], + "score": 1.0, + "content": ", sample complexity of exploration is at least", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 217, + 506, + 276 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 257, + 279, + 354, + 307 + ], + "lines": [ + { + "bbox": [ + 257, + 279, + 354, + 307 + ], + "spans": [ + { + "bbox": [ + 257, + 279, + 354, + 307 + ], + "score": 0.94, + "content": "t _ { 0 } = \\Omega \\left( { \\frac { 1 } { \\epsilon ^ { 3 } \\left( \\ln { 1 / \\delta } \\right) } } \\right) .", + "type": "interline_equation", + "image_path": "446057ad76737ed78dbeb6c14c391ec04ae8a21d0dea27c95bf135aaaf7cc43f.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 257, + 279, + 354, + 307 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 312, + 380, + 326 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 378, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 132, + 328 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 132, + 312, + 192, + 325 + ], + "score": 0.92, + "content": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 308, + 270, + 328 + ], + "score": 1.0, + "content": ", it can be seen that", + "type": "text" + }, + { + "bbox": [ + 270, + 311, + 378, + 326 + ], + "score": 0.93, + "content": "\\begin{array} { r } { C = 3 \\ln \\frac { 1 } { \\delta } + 1 < \\frac { 4 } { \\epsilon ^ { 2 } } < m } \\end{array}", + "type": "inline_equation" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 308, + 378, + 328 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/BkglSTNFDB/BkglSTNFDB_model.json b/parse/train/BkglSTNFDB/BkglSTNFDB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6261e34c6bea6e8b40b411592e29f8fca563307e --- /dev/null +++ b/parse/train/BkglSTNFDB/BkglSTNFDB_model.json @@ -0,0 +1,32281 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 993, + 1304, + 993, + 1304, + 1321, + 398, + 1321 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1466, + 1404, + 1466, + 1404, + 1712, + 297, + 1712 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1727, + 1405, + 1727, + 1405, + 1972, + 298, + 1972 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 314, + 545, + 1029, + 545, + 1029, + 668, + 314, + 668 + ], + "score": 0.964 + }, + { + "category_id": 0, + "poly": [ + 300, + 219, + 1400, + 219, + 1400, + 323, + 300, + 323 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 313, + 715, + 1024, + 715, + 1024, + 838, + 313, + 838 + ], + "score": 0.942 + }, + { + "category_id": 2, + "poly": [ + 332, + 2006, + 730, + 2006, + 730, + 2034, + 332, + 2034 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 302, + 1392, + 574, + 1392, + 574, + 1427, + 302, + 1427 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 313, + 374, + 946, + 374, + 946, + 499, + 313, + 499 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.87 + }, + { + "category_id": 0, + "poly": [ + 773, + 919, + 927, + 919, + 927, + 952, + 773, + 952 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 856, + 2088, + 856, + 2112, + 840, + 2112 + ], + "score": 0.737 + }, + { + "category_id": 0, + "poly": [ + 315, + 717, + 461, + 717, + 461, + 746, + 315, + 746 + ], + "score": 0.162 + }, + { + "category_id": 13, + "poly": [ + 759, + 1177, + 899, + 1177, + 899, + 1222, + 759, + 1222 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 575, + 1218, + 715, + 1218, + 715, + 1263, + 575, + 1263 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\tilde { O } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 738, + 1820, + 764, + 1820, + 764, + 1850, + 738, + 1850 + ], + "score": 0.82, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 1141, + 1260, + 1163, + 1260, + 1163, + 1288, + 1141, + 1288 + ], + "score": 0.8, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 822, + 1564, + 838, + 1564, + 838, + 1587, + 822, + 1587 + ], + "score": 0.75, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 1007, + 1266, + 1022, + 1266, + 1022, + 1287, + 1007, + 1287 + ], + "score": 0.73, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1260, + 1238, + 1260, + 1238, + 1287, + 1213, + 1287 + ], + "score": 0.7, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 615, + 1565, + 632, + 1565, + 632, + 1586, + 615, + 1586 + ], + "score": 0.7, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 298, + 1565, + 314, + 1565, + 314, + 1586, + 298, + 1586 + ], + "score": 0.52, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 219.0, + 1405.0, + 219.0, + 1405.0, + 272.0, + 294.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 275.0, + 1142.0, + 275.0, + 1142.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1998.0, + 733.0, + 1998.0, + 733.0, + 2040.0, + 329.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1388.0, + 579.0, + 1388.0, + 579.0, + 1435.0, + 294.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 915.0, + 934.0, + 915.0, + 934.0, + 958.0, + 768.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 859.0, + 2087.0, + 859.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 709.0, + 465.0, + 709.0, + 465.0, + 754.0, + 310.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 993.0, + 1308.0, + 993.0, + 1308.0, + 1029.0, + 393.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1022.0, + 1305.0, + 1022.0, + 1305.0, + 1061.0, + 393.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1055.0, + 1305.0, + 1055.0, + 1305.0, + 1090.0, + 394.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1085.0, + 1305.0, + 1085.0, + 1305.0, + 1120.0, + 394.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1116.0, + 1304.0, + 1116.0, + 1304.0, + 1148.0, + 395.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1146.0, + 1305.0, + 1146.0, + 1305.0, + 1184.0, + 393.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1173.0, + 758.0, + 1173.0, + 758.0, + 1231.0, + 388.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1173.0, + 1311.0, + 1173.0, + 1311.0, + 1231.0, + 900.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1202.0, + 574.0, + 1202.0, + 574.0, + 1279.0, + 380.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1202.0, + 1318.0, + 1202.0, + 1318.0, + 1279.0, + 716.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1258.0, + 1006.0, + 1258.0, + 1006.0, + 1295.0, + 393.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 1258.0, + 1140.0, + 1258.0, + 1140.0, + 1295.0, + 1023.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1258.0, + 1212.0, + 1258.0, + 1212.0, + 1295.0, + 1164.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1258.0, + 1305.0, + 1258.0, + 1305.0, + 1295.0, + 1239.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1290.0, + 619.0, + 1290.0, + 619.0, + 1323.0, + 394.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1501.0, + 295.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1498.0, + 1404.0, + 1498.0, + 1404.0, + 1531.0, + 295.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1528.0, + 1406.0, + 1528.0, + 1406.0, + 1563.0, + 292.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1559.0, + 297.0, + 1559.0, + 297.0, + 1593.0, + 294.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1559.0, + 614.0, + 1559.0, + 614.0, + 1593.0, + 315.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1559.0, + 821.0, + 1559.0, + 821.0, + 1593.0, + 633.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1559.0, + 1407.0, + 1559.0, + 1407.0, + 1593.0, + 839.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1624.0, + 294.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1406.0, + 1620.0, + 1406.0, + 1654.0, + 295.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1650.0, + 1407.0, + 1650.0, + 1407.0, + 1686.0, + 292.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1682.0, + 1044.0, + 1682.0, + 1044.0, + 1715.0, + 295.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1724.0, + 1406.0, + 1724.0, + 1406.0, + 1764.0, + 292.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1758.0, + 1405.0, + 1758.0, + 1405.0, + 1792.0, + 295.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1790.0, + 1405.0, + 1790.0, + 1405.0, + 1823.0, + 293.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1817.0, + 737.0, + 1817.0, + 737.0, + 1853.0, + 293.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1817.0, + 1405.0, + 1817.0, + 1405.0, + 1853.0, + 765.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1884.0, + 293.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1876.0, + 1410.0, + 1876.0, + 1410.0, + 1915.0, + 292.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1908.0, + 1409.0, + 1908.0, + 1409.0, + 1945.0, + 293.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1937.0, + 411.0, + 1937.0, + 411.0, + 1976.0, + 293.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 545.0, + 477.0, + 545.0, + 477.0, + 580.0, + 312.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 574.0, + 1031.0, + 574.0, + 1031.0, + 610.0, + 311.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 604.0, + 521.0, + 604.0, + 521.0, + 642.0, + 310.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 637.0, + 586.0, + 637.0, + 586.0, + 672.0, + 313.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 710.0, + 465.0, + 710.0, + 465.0, + 751.0, + 310.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 745.0, + 1023.0, + 745.0, + 1023.0, + 778.0, + 312.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 772.0, + 798.0, + 772.0, + 798.0, + 811.0, + 311.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 807.0, + 670.0, + 807.0, + 670.0, + 841.0, + 311.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 373.0, + 675.0, + 373.0, + 675.0, + 411.0, + 312.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 404.0, + 890.0, + 404.0, + 890.0, + 443.0, + 310.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 436.0, + 545.0, + 436.0, + 545.0, + 471.0, + 311.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 467.0, + 948.0, + 467.0, + 948.0, + 503.0, + 314.0, + 503.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1490, + 1404, + 1490, + 1404, + 1645, + 298, + 1645 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1113, + 1404, + 1113, + 1404, + 1328, + 298, + 1328 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 935, + 1404, + 935, + 1404, + 1098, + 298, + 1098 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 428, + 1405, + 428, + 1405, + 614, + 298, + 614 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 415, + 298, + 415 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1343, + 1403, + 1343, + 1403, + 1476, + 299, + 1476 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 766, + 1404, + 766, + 1404, + 920, + 299, + 920 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 629, + 1404, + 629, + 1404, + 752, + 299, + 752 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1659, + 1403, + 1659, + 1403, + 1813, + 299, + 1813 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1402, + 1971, + 1402, + 2035, + 297, + 2035 + ], + "score": 0.942 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 816, + 76, + 816, + 104, + 300, + 104 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 300, + 1886, + 551, + 1886, + 551, + 1921, + 300, + 1921 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.7 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.176 + }, + { + "category_id": 13, + "poly": [ + 606, + 2002, + 711, + 2002, + 711, + 2036, + 606, + 2036 + ], + "score": 0.94, + "latex": "p ( s ^ { \\prime } | s , a )" + }, + { + "category_id": 13, + "poly": [ + 599, + 969, + 739, + 969, + 739, + 1012, + 599, + 1012 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 868, + 1376, + 1009, + 1376, + 1009, + 1419, + 868, + 1419 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 644, + 1234, + 739, + 1234, + 739, + 1270, + 644, + 1270 + ], + "score": 0.92, + "latex": "\\Omega ( 1 / \\epsilon ^ { 3 } )" + }, + { + "category_id": 13, + "poly": [ + 1191, + 720, + 1306, + 720, + 1306, + 754, + 1191, + 754 + ], + "score": 0.92, + "latex": "1 / ( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 1013, + 1973, + 1166, + 1973, + 1166, + 2005, + 1013, + 2005 + ], + "score": 0.91, + "latex": "\\langle S , \\mathcal { A } , p , r , \\gamma \\rangle" + }, + { + "category_id": 13, + "poly": [ + 989, + 2005, + 1190, + 2005, + 1190, + 2036, + 989, + 2036 + ], + "score": 0.9, + "latex": "r : S \\times \\mathcal { A } [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1114, + 1160, + 1114, + 1160, + 1148, + 1117, + 1148 + ], + "score": 0.86, + "latex": "1 / \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1253, + 1974, + 1275, + 1974, + 1275, + 2000, + 1253, + 2000 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 371, + 2004, + 395, + 2004, + 395, + 2030, + 371, + 2030 + ], + "score": 0.8, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 1258, + 664, + 1278, + 664, + 1278, + 692, + 1258, + 692 + ], + "score": 0.8, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 323, + 1445, + 344, + 1445, + 344, + 1473, + 323, + 1473 + ], + "score": 0.72, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 395, + 1446, + 419, + 1446, + 419, + 1472, + 395, + 1472 + ], + "score": 0.68, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 298, + 1451, + 312, + 1451, + 312, + 1472, + 298, + 1472 + ], + "score": 0.41, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1883.0, + 556.0, + 1883.0, + 556.0, + 1927.0, + 292.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1490.0, + 1405.0, + 1490.0, + 1405.0, + 1528.0, + 296.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1520.0, + 1402.0, + 1520.0, + 1402.0, + 1557.0, + 294.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1550.0, + 1406.0, + 1550.0, + 1406.0, + 1590.0, + 293.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1583.0, + 1404.0, + 1583.0, + 1404.0, + 1617.0, + 296.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1611.0, + 1195.0, + 1611.0, + 1195.0, + 1650.0, + 292.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1113.0, + 1116.0, + 1113.0, + 1116.0, + 1152.0, + 293.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1113.0, + 1405.0, + 1113.0, + 1405.0, + 1152.0, + 1161.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1143.0, + 1404.0, + 1143.0, + 1404.0, + 1180.0, + 293.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1177.0, + 1404.0, + 1177.0, + 1404.0, + 1208.0, + 296.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1203.0, + 1404.0, + 1203.0, + 1404.0, + 1241.0, + 293.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1236.0, + 643.0, + 1236.0, + 643.0, + 1271.0, + 294.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1236.0, + 1405.0, + 1236.0, + 1405.0, + 1271.0, + 740.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1266.0, + 1404.0, + 1266.0, + 1404.0, + 1300.0, + 294.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 1240.0, + 1297.0, + 1240.0, + 1331.0, + 293.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 935.0, + 1404.0, + 935.0, + 1404.0, + 972.0, + 294.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 964.0, + 598.0, + 964.0, + 598.0, + 1019.0, + 287.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 964.0, + 1411.0, + 964.0, + 1411.0, + 1019.0, + 740.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1006.0, + 1405.0, + 1006.0, + 1405.0, + 1042.0, + 293.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1038.0, + 1406.0, + 1038.0, + 1406.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1067.0, + 698.0, + 1067.0, + 698.0, + 1100.0, + 296.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 428.0, + 1405.0, + 428.0, + 1405.0, + 464.0, + 295.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 460.0, + 1406.0, + 460.0, + 1406.0, + 494.0, + 291.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 488.0, + 1405.0, + 488.0, + 1405.0, + 525.0, + 293.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 521.0, + 1405.0, + 521.0, + 1405.0, + 557.0, + 295.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 546.0, + 1410.0, + 546.0, + 1410.0, + 591.0, + 291.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 580.0, + 1162.0, + 580.0, + 1162.0, + 619.0, + 293.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1403.0, + 229.0, + 1403.0, + 263.0, + 295.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 1408.0, + 258.0, + 1408.0, + 298.0, + 293.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1405.0, + 293.0, + 1405.0, + 325.0, + 296.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 1406.0, + 322.0, + 1406.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 383.0, + 566.0, + 383.0, + 566.0, + 419.0, + 294.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1343.0, + 1404.0, + 1343.0, + 1404.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1375.0, + 867.0, + 1375.0, + 867.0, + 1418.0, + 294.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1375.0, + 1410.0, + 1375.0, + 1410.0, + 1418.0, + 1010.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1413.0, + 1403.0, + 1413.0, + 1403.0, + 1448.0, + 294.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1443.0, + 297.0, + 1443.0, + 297.0, + 1481.0, + 294.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1443.0, + 322.0, + 1443.0, + 322.0, + 1481.0, + 313.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 1443.0, + 394.0, + 1443.0, + 394.0, + 1481.0, + 345.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1443.0, + 711.0, + 1443.0, + 711.0, + 1481.0, + 420.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 764.0, + 1409.0, + 764.0, + 1409.0, + 805.0, + 294.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1406.0, + 798.0, + 1406.0, + 831.0, + 295.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 826.0, + 1406.0, + 826.0, + 1406.0, + 864.0, + 294.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 1406.0, + 858.0, + 1406.0, + 895.0, + 294.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 888.0, + 417.0, + 888.0, + 417.0, + 923.0, + 291.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 629.0, + 1409.0, + 629.0, + 1409.0, + 665.0, + 294.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 660.0, + 1257.0, + 660.0, + 1257.0, + 693.0, + 295.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 660.0, + 1404.0, + 660.0, + 1404.0, + 693.0, + 1279.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 691.0, + 1403.0, + 691.0, + 1403.0, + 723.0, + 294.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 718.0, + 1190.0, + 718.0, + 1190.0, + 757.0, + 293.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 718.0, + 1387.0, + 718.0, + 1387.0, + 757.0, + 1307.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1658.0, + 1405.0, + 1658.0, + 1405.0, + 1696.0, + 294.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1691.0, + 1405.0, + 1691.0, + 1405.0, + 1724.0, + 297.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1720.0, + 1403.0, + 1720.0, + 1403.0, + 1753.0, + 295.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1752.0, + 1405.0, + 1752.0, + 1405.0, + 1789.0, + 294.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1784.0, + 443.0, + 1784.0, + 443.0, + 1814.0, + 295.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1012.0, + 1971.0, + 1012.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1971.0, + 1252.0, + 1971.0, + 1252.0, + 2007.0, + 1167.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2007.0, + 1276.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 370.0, + 2000.0, + 370.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 2000.0, + 605.0, + 2000.0, + 605.0, + 2038.0, + 396.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 2000.0, + 988.0, + 2000.0, + 988.0, + 2038.0, + 712.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2038.0, + 1191.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 413, + 1405, + 413, + 1405, + 601, + 297, + 601 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 712, + 1406, + 712, + 1406, + 806, + 297, + 806 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 298, + 1922, + 1404, + 1922, + 1404, + 2043, + 298, + 2043 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 297, + 894, + 1403, + 894, + 1403, + 989, + 297, + 989 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 299, + 307, + 1401, + 307, + 1401, + 402, + 299, + 402 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 300, + 614, + 1399, + 614, + 1399, + 708, + 300, + 708 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 293, + 228, + 1402, + 228, + 1402, + 294, + 293, + 294 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 298, + 827, + 1401, + 827, + 1401, + 889, + 298, + 889 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 299, + 1195, + 574, + 1195, + 574, + 1232, + 299, + 1232 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 299, + 1007, + 1011, + 1007, + 1011, + 1040, + 299, + 1040 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 816, + 74, + 816, + 105, + 299, + 105 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 286, + 1263, + 1322, + 1263, + 1322, + 1298, + 286, + 1298 + ], + "score": 0.913 + }, + { + "category_id": 8, + "poly": [ + 387, + 1045, + 1286, + 1045, + 1286, + 1117, + 387, + 1117 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 288, + 1121, + 1370, + 1121, + 1370, + 1157, + 288, + 1157 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 300, + 1396, + 777, + 1396, + 777, + 1429, + 300, + 1429 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 298, + 1330, + 518, + 1330, + 518, + 1363, + 298, + 1363 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2113, + 841, + 2113 + ], + "score": 0.652 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 860, + 2087, + 860, + 2113, + 841, + 2113 + ], + "score": 0.581 + }, + { + "category_id": 1, + "poly": [ + 307, + 1435, + 1402, + 1435, + 1402, + 1893, + 307, + 1893 + ], + "score": 0.47 + }, + { + "category_id": 5, + "poly": [ + 307, + 1435, + 1402, + 1435, + 1402, + 1893, + 307, + 1893 + ], + "score": 0.446, + "html": "
Parameters: ∈, γ, δ E ,H← ln1/(1-2)∈1). In 1/q Define (k)= ln(SA(k + 1)(k + 2)/δ),αk = H· H+1 for t=1,2,.. do
5: Take action at ← arg maxa' Q(st, a') Receive reward rt and transit to St+1 N(st,at)← N(st,at)+1
k←N(st,at),bk←1√ H(k) k c2 is a constant and can be set to 4√2
V(St+1) ← maXa∈A Q(st+1,a)
Q(st,at) ←(1-ak)Q(st,at)+αk [r(st,at)+bk +γV(st+1)] 10:
Q(st,at) ← min(Q(st,at),Q(st,at))
" + }, + { + "category_id": 13, + "poly": [ + 470, + 1996, + 618, + 1996, + 618, + 2043, + 470, + 2043 + ], + "score": 0.94, + "latex": "\\frac { \\ln { 1 } / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln { 1 } / \\gamma }" + }, + { + "category_id": 13, + "poly": [ + 648, + 1924, + 941, + 1924, + 941, + 1968, + 648, + 1968 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { R = \\lceil \\ln \\frac { 3 } { \\epsilon ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 799, + 1995, + 1010, + 1995, + 1010, + 2042, + 799, + 2042 + ], + "score": 0.93, + "latex": "H \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma }" + }, + { + "category_id": 13, + "poly": [ + 360, + 1923, + 481, + 1923, + 481, + 1958, + 360, + 1958 + ], + "score": 0.93, + "latex": "c _ { 2 } = 4 \\sqrt { 2 }" + }, + { + "category_id": 13, + "poly": [ + 774, + 340, + 853, + 340, + 853, + 369, + 774, + 369 + ], + "score": 0.93, + "latex": "s _ { t } \\in S" + }, + { + "category_id": 13, + "poly": [ + 1079, + 230, + 1171, + 230, + 1171, + 264, + 1079, + 264 + ], + "score": 0.93, + "latex": "S = | S |" + }, + { + "category_id": 13, + "poly": [ + 533, + 231, + 653, + 231, + 653, + 262, + 533, + 262 + ], + "score": 0.93, + "latex": "0 \\leq \\gamma < 1" + }, + { + "category_id": 13, + "poly": [ + 1222, + 230, + 1320, + 230, + 1320, + 264, + 1222, + 264 + ], + "score": 0.92, + "latex": "A = | { \\mathcal { A } } |" + }, + { + "category_id": 13, + "poly": [ + 1005, + 340, + 1088, + 340, + 1088, + 369, + 1005, + 369 + ], + "score": 0.91, + "latex": "a _ { t } \\in \\mathcal A" + }, + { + "category_id": 13, + "poly": [ + 1158, + 477, + 1402, + 477, + 1402, + 509, + 1158, + 509 + ], + "score": 0.91, + "latex": "V ^ { * } ( s ) = \\operatorname* { s u p } _ { \\pi } V ^ { \\pi } ( s )" + }, + { + "category_id": 13, + "poly": [ + 725, + 447, + 808, + 447, + 808, + 479, + 725, + 479 + ], + "score": 0.91, + "latex": "V ^ { \\pi _ { t } } ( s )" + }, + { + "category_id": 13, + "poly": [ + 295, + 536, + 824, + 536, + 824, + 573, + 295, + 573 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\mathbb { E } [ \\sum _ { i = 2 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = \\stackrel { \\cdot } { s } , a _ { 1 } = \\stackrel { \\cdot } { a } ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1117, + 1963, + 1393, + 1963, + 1393, + 1999, + 1117, + 1999 + ], + "score": 0.91, + "latex": "\\alpha _ { k } = ( H + 1 ) / ( H + k )" + }, + { + "category_id": 13, + "poly": [ + 663, + 374, + 716, + 374, + 716, + 402, + 663, + 402 + ], + "score": 0.9, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 773, + 564, + 773, + 564, + 807, + 298, + 807 + ], + "score": 0.89, + "latex": "\\dot { V } ^ { \\pi _ { t } } \\left( s _ { t } \\right) \\dot { < } V ^ { * } \\left( s _ { t } \\right) - \\epsilon ." + }, + { + "category_id": 13, + "poly": [ + 1209, + 540, + 1307, + 540, + 1307, + 570, + 1209, + 570 + ], + "score": 0.89, + "latex": "Q ^ { * } ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 397, + 1043, + 1286, + 1043, + 1286, + 1120, + 397, + 1120 + ], + "score": 0.89, + "latex": "\\left\\{ \\begin{array} { l l } { V ^ { \\pi _ { t } } ( s ) = Q ^ { \\pi _ { t } } \\left( s , \\pi _ { t } ( s ) \\right) } \\\\ { Q ^ { \\pi _ { t } } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { \\pi _ { t + 1 } } \\right) ( s , a ) , } \\end{array} \\right. \\quad \\left\\{ \\begin{array} { l l } { V ^ { * } ( s ) = Q ^ { * } \\left( s , \\pi ^ { * } ( s ) \\right) } \\\\ { Q ^ { * } ( s , a ) : = \\left( r _ { t } + \\gamma \\mathbb { P } V ^ { * } \\right) ( s , a ) , } \\end{array} \\right." + }, + { + "category_id": 13, + "poly": [ + 711, + 955, + 1009, + 955, + 1009, + 989, + 711, + 989 + ], + "score": 0.89, + "latex": "( S , A , 1 / \\epsilon , 1 / \\delta , 1 / ( 1 - \\gamma ) )" + }, + { + "category_id": 13, + "poly": [ + 1347, + 1689, + 1402, + 1689, + 1402, + 1725, + 1347, + 1725 + ], + "score": 0.88, + "latex": "4 { \\sqrt { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 933, + 1123, + 1371, + 1123, + 1371, + 1158, + 933, + 1158 + ], + "score": 0.87, + "latex": "\\left[ \\mathbb { P } V ^ { \\pi _ { t } } \\right] ( s , a ) : = \\mathbb { E } _ { s ^ { \\prime } \\sim p ( \\cdot \\vert s , a ) } V ^ { \\pi _ { t + 1 } } ( s ^ { \\prime } ) ." + }, + { + "category_id": 13, + "poly": [ + 451, + 1466, + 722, + 1466, + 722, + 1511, + 451, + 1511 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { Q ( s , a ) , \\hat { Q } ( s , a ) \\gets \\frac { 1 } { 1 - \\gamma } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1153, + 509, + 1407, + 509, + 1407, + 540, + 1153, + 540 + ], + "score": 0.86, + "latex": "Q ^ { \\pi _ { t } } ( s , a ) = r ( s , a ) +" + }, + { + "category_id": 13, + "poly": [ + 888, + 421, + 918, + 421, + 918, + 447, + 888, + 447 + ], + "score": 0.86, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 533, + 314, + 561, + 314, + 561, + 339, + 533, + 339 + ], + "score": 0.86, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 447, + 476, + 1002, + 476, + 1002, + 512, + 447, + 512 + ], + "score": 0.85, + "latex": "\\begin{array} { r } { V ^ { \\pi _ { t } } ( s ) = \\mathbb { E } \\bigl [ \\sum _ { i = 1 } ^ { \\infty } \\gamma ^ { i - 1 } r ( s _ { i } , \\pi _ { t + i - 1 } ( s _ { i } ) ) | s _ { 1 } = s \\bigr ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1326, + 343, + 1351, + 343, + 1351, + 369, + 1326, + 369 + ], + "score": 0.85, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 956, + 1358, + 956, + 1358, + 985, + 1295, + 985 + ], + "score": 0.85, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 633, + 572, + 665, + 572, + 665, + 596, + 633, + 596 + ], + "score": 0.84, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 842, + 748, + 871, + 748, + 871, + 774, + 842, + 774 + ], + "score": 0.84, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 563, + 1622, + 589, + 1622, + 589, + 1647, + 563, + 1647 + ], + "score": 0.82, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 452, + 1316, + 452, + 1316, + 476, + 1288, + 476 + ], + "score": 0.82, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 745, + 1624, + 798, + 1624, + 798, + 1650, + 745, + 1650 + ], + "score": 0.8, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1181, + 1928, + 1214, + 1928, + 1214, + 1955, + 1181, + 1955 + ], + "score": 0.8, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1333, + 749, + 1358, + 749, + 1358, + 774, + 1333, + 774 + ], + "score": 0.8, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 298, + 2002, + 326, + 2002, + 326, + 2030, + 298, + 2030 + ], + "score": 0.79, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 426, + 1515, + 930, + 1515, + 930, + 1555, + 426, + 1555 + ], + "score": 0.78, + "latex": "\\begin{array} { r } { \\iota ( k ) = \\ln ( S A ( k + 1 ) ( k + 2 ) / \\delta ) , \\alpha _ { k } = \\frac { H + 1 } { H + k } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 451, + 1964, + 752, + 1964, + 752, + 1998, + 451, + 1998 + ], + "score": 0.75, + "latex": "{ \\cal { \\left( M = 0 \\left( \\ln 1 / ( 1 - \\gamma ) \\epsilon \\right) \\right) } }" + }, + { + "category_id": 13, + "poly": [ + 520, + 1582, + 814, + 1582, + 814, + 1619, + 520, + 1619 + ], + "score": 0.73, + "latex": "a _ { t } \\gets \\arg \\operatorname* { m a x } _ { a ^ { \\prime } } \\hat { Q } ( s _ { t } , a ^ { \\prime } )" + }, + { + "category_id": 14, + "poly": [ + 432, + 1462, + 1332, + 1462, + 1332, + 1562, + 432, + 1562 + ], + "score": 0.69, + "latex": "\\begin{array} { r l } & { \\textnormal { \\texttt { e } } Q ( s , a ) , \\hat { Q } ( s , a ) \\gets \\frac { 1 } { 1 - \\gamma } , N ( s , a ) \\gets 0 , \\epsilon _ { 1 } \\gets \\frac { \\epsilon } { 2 4 R M \\ln \\frac { 1 } { 1 - \\gamma } } , H \\gets \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } . } \\\\ & { \\cdot ( k ) = \\ln ( S A ( k + 1 ) ( k + 2 ) / \\delta ) , \\alpha _ { k } = \\frac { H + 1 } { H + k } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 957, + 748, + 971, + 748, + 971, + 771, + 957, + 771 + ], + "score": 0.68, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 621, + 894, + 684, + 894, + 684, + 924, + 621, + 924 + ], + "score": 0.67, + "latex": "\\mathcal { A L G }" + }, + { + "category_id": 13, + "poly": [ + 628, + 927, + 653, + 927, + 653, + 957, + 628, + 957 + ], + "score": 0.67, + "latex": "i f ," + }, + { + "category_id": 13, + "poly": [ + 386, + 1553, + 548, + 1553, + 548, + 1583, + 386, + 1583 + ], + "score": 0.66, + "latex": "t = 1 , 2 , \\dots \\mathbf { d o }" + }, + { + "category_id": 13, + "poly": [ + 819, + 926, + 834, + 926, + 834, + 952, + 819, + 952 + ], + "score": 0.65, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 980, + 1700, + 1007, + 1700, + 1007, + 1724, + 980, + 1724 + ], + "score": 0.65, + "latex": "c _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 542, + 1196, + 542, + 1196, + 569, + 1168, + 569 + ], + "score": 0.64, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 607, + 449, + 621, + 449, + 621, + 474, + 607, + 474 + ], + "score": 0.62, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1138, + 925, + 1200, + 925, + 1200, + 955, + 1138, + 955 + ], + "score": 0.61, + "latex": "\\mathcal { A L G }" + }, + { + "category_id": 13, + "poly": [ + 1057, + 712, + 1119, + 712, + 1119, + 743, + 1057, + 743 + ], + "score": 0.56, + "latex": "\\mathcal { A L G }" + }, + { + "category_id": 14, + "poly": [ + 382, + 1640, + 1142, + 1640, + 1142, + 1863, + 382, + 1863 + ], + "score": 0.48, + "latex": "\\begin{array} { r l r } & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1 } & \\\\ & { k \\gets N ( s _ { t } , a _ { t } ) , b _ { k } \\gets \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H ( k ) } { k } } } & \\\\ & { \\hat { V } ( s _ { t + 1 } ) \\gets \\operatorname* { m a x } _ { a \\in A } \\hat { Q } ( s _ { t + 1 } , a ) } & \\\\ & { Q ( s _ { t } , a _ { t } ) \\gets ( 1 - \\alpha _ { k } ) Q ( s _ { t } , a _ { t } ) + \\alpha _ { k } \\left[ r ( s _ { t } , a _ { t } ) + b _ { k } + \\gamma \\hat { V } ( s _ { t + 1 } ) \\right] } \\\\ & { \\hat { Q } ( s _ { t } , a _ { t } ) \\gets \\operatorname* { m i n } ( \\hat { Q } ( s _ { t } , a _ { t } ) , Q ( s _ { t } , a _ { t } ) ) } & \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1040, + 750, + 1055, + 750, + 1055, + 771, + 1040, + 771 + ], + "score": 0.47, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 709, + 1582, + 812, + 1582, + 812, + 1619, + 709, + 1619 + ], + "score": 0.42, + "latex": "\\hat { Q } ( s _ { t } , a ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 385, + 1647, + 695, + 1647, + 695, + 1682, + 385, + 1682 + ], + "score": 0.34, + "latex": "N ( s _ { t } , a _ { t } ) \\gets N ( s _ { t } , a _ { t } ) + 1" + }, + { + "category_id": 13, + "poly": [ + 457, + 1964, + 551, + 1964, + 551, + 1994, + 457, + 1994 + ], + "score": 0.28, + "latex": "M = \\mathcal { O }" + }, + { + "category_id": 13, + "poly": [ + 976, + 678, + 1001, + 678, + 1001, + 705, + 976, + 705 + ], + "score": 0.28, + "latex": "\\&" + }, + { + "category_id": 13, + "poly": [ + 898, + 539, + 923, + 539, + 923, + 569, + 898, + 569 + ], + "score": 0.26, + "latex": "\\mathrm { Q }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1191.0, + 575.0, + 1191.0, + 575.0, + 1238.0, + 291.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1394.0, + 781.0, + 1394.0, + 781.0, + 1436.0, + 294.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1329.0, + 521.0, + 1329.0, + 521.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 837.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 411.0, + 887.0, + 411.0, + 887.0, + 453.0, + 292.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 411.0, + 1406.0, + 411.0, + 1406.0, + 453.0, + 919.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 443.0, + 606.0, + 443.0, + 606.0, + 483.0, + 294.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 443.0, + 724.0, + 443.0, + 724.0, + 483.0, + 622.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 443.0, + 1287.0, + 443.0, + 1287.0, + 483.0, + 809.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 443.0, + 1407.0, + 443.0, + 1407.0, + 483.0, + 1317.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 463.0, + 446.0, + 463.0, + 446.0, + 524.0, + 286.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 463.0, + 1157.0, + 463.0, + 1157.0, + 524.0, + 1003.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 463.0, + 1415.0, + 463.0, + 1415.0, + 524.0, + 1403.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 503.0, + 1152.0, + 503.0, + 1152.0, + 544.0, + 290.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1408.0, + 503.0, + 1411.0, + 503.0, + 1411.0, + 544.0, + 1408.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 520.0, + 294.0, + 520.0, + 294.0, + 588.0, + 283.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 520.0, + 897.0, + 520.0, + 897.0, + 588.0, + 825.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 520.0, + 1167.0, + 520.0, + 1167.0, + 588.0, + 924.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 520.0, + 1208.0, + 520.0, + 1208.0, + 588.0, + 1197.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 520.0, + 1416.0, + 520.0, + 1416.0, + 588.0, + 1308.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 567.0, + 632.0, + 567.0, + 632.0, + 603.0, + 295.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 567.0, + 677.0, + 567.0, + 677.0, + 603.0, + 666.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 1056.0, + 710.0, + 1056.0, + 748.0, + 294.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 710.0, + 1406.0, + 710.0, + 1406.0, + 748.0, + 1120.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 742.0, + 841.0, + 742.0, + 841.0, + 781.0, + 292.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 742.0, + 956.0, + 742.0, + 956.0, + 781.0, + 872.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 742.0, + 1039.0, + 742.0, + 1039.0, + 781.0, + 972.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 742.0, + 1332.0, + 742.0, + 1332.0, + 781.0, + 1056.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 742.0, + 1409.0, + 742.0, + 1409.0, + 781.0, + 1359.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 769.0, + 297.0, + 769.0, + 297.0, + 811.0, + 294.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 769.0, + 575.0, + 769.0, + 575.0, + 811.0, + 565.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1918.0, + 359.0, + 1918.0, + 359.0, + 1974.0, + 291.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1918.0, + 647.0, + 1918.0, + 647.0, + 1974.0, + 482.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1918.0, + 1180.0, + 1918.0, + 1180.0, + 1974.0, + 942.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1918.0, + 1410.0, + 1918.0, + 1410.0, + 1974.0, + 1215.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1959.0, + 450.0, + 1959.0, + 450.0, + 1999.0, + 293.0, + 1999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1959.0, + 1116.0, + 1959.0, + 1116.0, + 1999.0, + 753.0, + 1999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1959.0, + 1404.0, + 1959.0, + 1404.0, + 1999.0, + 1394.0, + 1999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1984.0, + 469.0, + 1984.0, + 469.0, + 2050.0, + 327.0, + 2050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1984.0, + 798.0, + 1984.0, + 798.0, + 2050.0, + 619.0, + 2050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 1984.0, + 1018.0, + 1984.0, + 1018.0, + 2050.0, + 1011.0, + 2050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 892.0, + 620.0, + 892.0, + 620.0, + 931.0, + 294.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 892.0, + 1404.0, + 892.0, + 1404.0, + 931.0, + 685.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 923.0, + 627.0, + 923.0, + 627.0, + 960.0, + 292.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 923.0, + 818.0, + 923.0, + 818.0, + 960.0, + 654.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 923.0, + 1137.0, + 923.0, + 1137.0, + 960.0, + 835.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 923.0, + 1404.0, + 923.0, + 1404.0, + 960.0, + 1201.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 955.0, + 710.0, + 955.0, + 710.0, + 993.0, + 292.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 955.0, + 1294.0, + 955.0, + 1294.0, + 993.0, + 1010.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 955.0, + 1370.0, + 955.0, + 1370.0, + 993.0, + 1359.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 306.0, + 532.0, + 306.0, + 532.0, + 345.0, + 293.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 306.0, + 1406.0, + 306.0, + 1406.0, + 345.0, + 562.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 337.0, + 773.0, + 337.0, + 773.0, + 372.0, + 295.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 337.0, + 1004.0, + 337.0, + 1004.0, + 372.0, + 854.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 337.0, + 1325.0, + 337.0, + 1325.0, + 372.0, + 1089.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 337.0, + 1403.0, + 337.0, + 1403.0, + 372.0, + 1352.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 365.0, + 662.0, + 365.0, + 662.0, + 408.0, + 293.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 365.0, + 730.0, + 365.0, + 730.0, + 408.0, + 717.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 611.0, + 1404.0, + 611.0, + 1404.0, + 653.0, + 294.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 645.0, + 1403.0, + 645.0, + 1403.0, + 679.0, + 296.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 677.0, + 975.0, + 677.0, + 975.0, + 711.0, + 296.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 677.0, + 1187.0, + 677.0, + 1187.0, + 711.0, + 1002.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 532.0, + 229.0, + 532.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 229.0, + 1078.0, + 229.0, + 1078.0, + 265.0, + 654.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 229.0, + 1221.0, + 229.0, + 1221.0, + 265.0, + 1172.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 265.0, + 1321.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 966.0, + 258.0, + 966.0, + 297.0, + 293.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 823.0, + 1405.0, + 823.0, + 1405.0, + 866.0, + 293.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 855.0, + 942.0, + 855.0, + 942.0, + 891.0, + 296.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1003.0, + 1014.0, + 1003.0, + 1014.0, + 1047.0, + 293.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1260.0, + 1324.0, + 1260.0, + 1324.0, + 1303.0, + 292.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1112.0, + 932.0, + 1112.0, + 932.0, + 1165.0, + 291.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1112.0, + 1377.0, + 1112.0, + 1377.0, + 1165.0, + 1372.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1431.0, + 561.0, + 1431.0, + 561.0, + 1467.0, + 343.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1463.0, + 431.0, + 1463.0, + 431.0, + 1510.0, + 339.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1508.0, + 425.0, + 1508.0, + 425.0, + 1558.0, + 341.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1549.0, + 385.0, + 1549.0, + 385.0, + 1584.0, + 342.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1549.0, + 552.0, + 1549.0, + 552.0, + 1584.0, + 549.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1584.0, + 342.0, + 1584.0, + 342.0, + 1620.0, + 307.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1583.0, + 519.0, + 1583.0, + 519.0, + 1622.0, + 381.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1680.0, + 381.0, + 1680.0, + 381.0, + 1740.0, + 377.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1691.0, + 1346.0, + 1691.0, + 1346.0, + 1725.0, + 1143.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1777.0, + 340.0, + 1777.0, + 340.0, + 1811.0, + 298.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1851.0, + 439.0, + 1851.0, + 439.0, + 1888.0, + 340.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.25, + 1452.5, + 1338.25, + 1452.5, + 1338.25, + 1504.0, + 1080.25, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.75, + 1480.5, + 1093.75, + 1480.5, + 1093.75, + 1517.5, + 952.75, + 1517.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.75, + 1613.0, + 801.75, + 1613.0, + 801.75, + 1652.0, + 381.75, + 1652.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1498, + 1404, + 1498, + 1404, + 1806, + 297, + 1806 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1130, + 1404, + 1130, + 1404, + 1346, + 298, + 1346 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 929, + 1406, + 929, + 1406, + 1115, + 297, + 1115 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1819, + 1404, + 1819, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1405, + 228, + 1405, + 426, + 297, + 426 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1360, + 1404, + 1360, + 1404, + 1485, + 297, + 1485 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 577, + 1409, + 577, + 1409, + 670, + 298, + 670 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 738, + 678, + 959, + 678, + 959, + 768, + 738, + 768 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 853, + 1405, + 853, + 1405, + 917, + 297, + 917 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 298, + 531, + 1117, + 531, + 1117, + 566, + 298, + 566 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 297, + 788, + 1042, + 788, + 1042, + 826, + 297, + 826 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 297, + 470, + 851, + 470, + 851, + 504, + 297, + 504 + ], + "score": 0.867 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.784 + }, + { + "category_id": 13, + "poly": [ + 346, + 358, + 441, + 358, + 441, + 395, + 346, + 395 + ], + "score": 0.93, + "latex": "\\hat { Q } _ { t } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 394, + 265, + 480, + 265, + 480, + 298, + 394, + 298 + ], + "score": 0.93, + "latex": "Q ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1275, + 1160, + 1361, + 1160, + 1361, + 1195, + 1275, + 1195 + ], + "score": 0.92, + "latex": "o ( T ^ { - 1 } )" + }, + { + "category_id": 14, + "poly": [ + 738, + 674, + 959, + 674, + 959, + 768, + 738, + 768 + ], + "score": 0.92, + "latex": "\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln { 1 / \\delta } } { \\epsilon ^ { 2 } \\left( 1 - \\gamma \\right) ^ { 7 } } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 764, + 1422, + 1028, + 1422, + 1028, + 1455, + 764, + 1455 + ], + "score": 0.92, + "latex": "V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 745, + 296, + 852, + 296, + 852, + 329, + 745, + 329 + ], + "score": 0.92, + "latex": "\\tau ( s , a , k )" + }, + { + "category_id": 13, + "poly": [ + 891, + 1881, + 1144, + 1881, + 1144, + 1914, + 891, + 1914 + ], + "score": 0.92, + "latex": "V ^ { \\ast } ( s _ { t ^ { \\prime } } ) - Q ^ { \\ast } ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } )" + }, + { + "category_id": 13, + "poly": [ + 1223, + 295, + 1403, + 295, + 1403, + 328, + 1223, + 328 + ], + "score": 0.91, + "latex": "( s _ { t } , a _ { t } ) = ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 467, + 1973, + 698, + 1973, + 698, + 2006, + 467, + 2006 + ], + "score": 0.91, + "latex": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 789, + 362, + 851, + 362, + 851, + 395, + 789, + 395 + ], + "score": 0.91, + "latex": "( s , a )" + }, + { + "category_id": 13, + "poly": [ + 580, + 1914, + 681, + 1914, + 681, + 1944, + 580, + 1944 + ], + "score": 0.91, + "latex": "[ t , t + R ]" + }, + { + "category_id": 13, + "poly": [ + 296, + 296, + 358, + 296, + 358, + 328, + 296, + 328 + ], + "score": 0.91, + "latex": "( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1210, + 992, + 1397, + 992, + 1397, + 1026, + 1210, + 1026 + ], + "score": 0.91, + "latex": "( V ^ { * } - V ^ { \\pi } ) ( s _ { 1 } ) )" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1561, + 1167, + 1561, + 1167, + 1590, + 1100, + 1590 + ], + "score": 0.91, + "latex": "h + 1" + }, + { + "category_id": 13, + "poly": [ + 559, + 1882, + 812, + 1882, + 812, + 1915, + 559, + 1915 + ], + "score": 0.91, + "latex": "V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 775, + 791, + 942, + 791, + 942, + 825, + 775, + 825 + ], + "score": 0.9, + "latex": "1 / \\epsilon , 1 / ( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 1117, + 327, + 1291, + 327, + 1291, + 360, + 1117, + 360 + ], + "score": 0.89, + "latex": "\\tau ( s , a , k ) = \\infty" + }, + { + "category_id": 13, + "poly": [ + 659, + 357, + 685, + 357, + 685, + 395, + 659, + 395 + ], + "score": 0.87, + "latex": "\\hat { Q }" + }, + { + "category_id": 13, + "poly": [ + 1063, + 1657, + 1089, + 1657, + 1089, + 1682, + 1063, + 1682 + ], + "score": 0.86, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 846, + 260, + 932, + 260, + 932, + 298, + 846, + 298 + ], + "score": 0.86, + "latex": "\\hat { Q } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 897, + 614, + 925, + 614, + 925, + 640, + 897, + 640 + ], + "score": 0.86, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1153, + 1197, + 1181, + 1197, + 1181, + 1223, + 1153, + 1223 + ], + "score": 0.85, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 944, + 264, + 1040, + 264, + 1040, + 299, + 944, + 299 + ], + "score": 0.85, + "latex": "N _ { t } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 487, + 1854, + 515, + 1854, + 515, + 1881, + 487, + 1881 + ], + "score": 0.85, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 824, + 1166, + 853, + 1166, + 853, + 1192, + 824, + 1192 + ], + "score": 0.85, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 588, + 1625, + 614, + 1625, + 614, + 1652, + 588, + 1652 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1165, + 1245, + 1165, + 1245, + 1193, + 1217, + 1193 + ], + "score": 0.85, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 581, + 362, + 607, + 362, + 607, + 395, + 581, + 395 + ], + "score": 0.85, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 1064, + 1227, + 1093, + 1227, + 1093, + 1253, + 1064, + 1253 + ], + "score": 0.85, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 803, + 1625, + 830, + 1625, + 830, + 1652, + 803, + 1652 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 675, + 1425, + 701, + 1425, + 701, + 1452, + 675, + 1452 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1375, + 369, + 1400, + 369, + 1400, + 393, + 1375, + 393 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1307, + 328, + 1403, + 328, + 1403, + 360, + 1307, + 360 + ], + "score": 0.84, + "latex": "Q _ { t } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1119, + 996, + 1148, + 996, + 1148, + 1023, + 1119, + 1023 + ], + "score": 0.83, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 358, + 1911, + 379, + 1911, + 379, + 1939, + 358, + 1939 + ], + "score": 0.83, + "latex": "t ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1686, + 324, + 1686, + 324, + 1713, + 298, + 1713 + ], + "score": 0.83, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1140, + 614, + 1167, + 614, + 1167, + 640, + 1140, + 640 + ], + "score": 0.82, + "latex": "s _ { t . }" + }, + { + "category_id": 13, + "poly": [ + 380, + 328, + 399, + 328, + 399, + 354, + 380, + 354 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 997, + 1562, + 1016, + 1562, + 1016, + 1588, + 997, + 1588 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 619, + 1225, + 643, + 1225, + 643, + 1250, + 619, + 1250 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 933, + 1856, + 957, + 1856, + 957, + 1881, + 933, + 1881 + ], + "score": 0.81, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1562, + 316, + 1562, + 316, + 1588, + 297, + 1588 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 684, + 1854, + 697, + 1854, + 697, + 1878, + 684, + 1878 + ], + "score": 0.8, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 420, + 1562, + 441, + 1562, + 441, + 1588, + 420, + 1588 + ], + "score": 0.79, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 371, + 787, + 398, + 787, + 398, + 820, + 371, + 820 + ], + "score": 0.79, + "latex": "\\tilde { \\mathcal { O } }" + }, + { + "category_id": 13, + "poly": [ + 349, + 994, + 373, + 994, + 373, + 1020, + 349, + 1020 + ], + "score": 0.79, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 927, + 1915, + 951, + 1915, + 951, + 1939, + 927, + 1939 + ], + "score": 0.78, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 540, + 581, + 606, + 581, + 606, + 609, + 540, + 609 + ], + "score": 0.77, + "latex": "\\epsilon > 0" + }, + { + "category_id": 13, + "poly": [ + 995, + 791, + 1037, + 791, + 1037, + 820, + 995, + 820 + ], + "score": 0.77, + "latex": "S A" + }, + { + "category_id": 13, + "poly": [ + 921, + 1194, + 946, + 1194, + 946, + 1220, + 921, + 1220 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1042, + 580, + 1107, + 580, + 1107, + 609, + 1042, + 609 + ], + "score": 0.76, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 353, + 1623, + 368, + 1623, + 368, + 1649, + 353, + 1649 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 727, + 1857, + 740, + 1857, + 740, + 1878, + 727, + 1878 + ], + "score": 0.75, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1100, + 299, + 1114, + 299, + 1114, + 324, + 1100, + 324 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1624, + 1147, + 1624, + 1147, + 1648, + 1133, + 1648 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 438, + 1715, + 452, + 1715, + 452, + 1741, + 438, + 1741 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 547, + 1425, + 561, + 1425, + 561, + 1449, + 547, + 1449 + ], + "score": 0.72, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 298, + 1428, + 312, + 1428, + 312, + 1449, + 298, + 1449 + ], + "score": 0.67, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 683, + 1685, + 697, + 1685, + 697, + 1709, + 683, + 1709 + ], + "score": 0.64, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 616, + 579, + 844, + 579, + 844, + 613, + 616, + 613 + ], + "score": 0.6, + "latex": "\\delta > 0 , 1 / 2 < \\gamma < 1" + }, + { + "category_id": 13, + "poly": [ + 998, + 617, + 1012, + 617, + 1012, + 637, + 998, + 637 + ], + "score": 0.55, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 719, + 298, + 734, + 298, + 734, + 325, + 719, + 325 + ], + "score": 0.48, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 467.0, + 854.0, + 467.0, + 854.0, + 509.0, + 292.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1496.0, + 1405.0, + 1496.0, + 1405.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1527.0, + 1404.0, + 1527.0, + 1404.0, + 1564.0, + 292.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1558.0, + 296.0, + 1558.0, + 296.0, + 1596.0, + 292.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1558.0, + 419.0, + 1558.0, + 419.0, + 1596.0, + 317.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1558.0, + 996.0, + 1558.0, + 996.0, + 1596.0, + 442.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1558.0, + 1099.0, + 1558.0, + 1099.0, + 1596.0, + 1017.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1558.0, + 1404.0, + 1558.0, + 1404.0, + 1596.0, + 1168.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1625.0, + 294.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1619.0, + 352.0, + 1619.0, + 352.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1619.0, + 587.0, + 1619.0, + 587.0, + 1656.0, + 369.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1619.0, + 802.0, + 1619.0, + 802.0, + 1656.0, + 615.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1619.0, + 1132.0, + 1619.0, + 1132.0, + 1656.0, + 831.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1619.0, + 1406.0, + 1619.0, + 1406.0, + 1656.0, + 1148.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1652.0, + 1062.0, + 1652.0, + 1062.0, + 1686.0, + 292.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1652.0, + 1406.0, + 1652.0, + 1406.0, + 1686.0, + 1090.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1680.0, + 297.0, + 1680.0, + 297.0, + 1717.0, + 291.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1680.0, + 682.0, + 1680.0, + 682.0, + 1717.0, + 325.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1680.0, + 1407.0, + 1680.0, + 1407.0, + 1717.0, + 698.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 437.0, + 1712.0, + 437.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1712.0, + 1407.0, + 1712.0, + 1407.0, + 1748.0, + 453.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1406.0, + 1743.0, + 1406.0, + 1778.0, + 294.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1770.0, + 599.0, + 1770.0, + 599.0, + 1812.0, + 294.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 1408.0, + 1131.0, + 1408.0, + 1165.0, + 294.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1158.0, + 823.0, + 1158.0, + 823.0, + 1198.0, + 289.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1158.0, + 1216.0, + 1158.0, + 1216.0, + 1198.0, + 854.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 1158.0, + 1274.0, + 1158.0, + 1274.0, + 1198.0, + 1246.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 1158.0, + 1406.0, + 1158.0, + 1406.0, + 1198.0, + 1362.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1191.0, + 920.0, + 1191.0, + 920.0, + 1228.0, + 293.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1191.0, + 1152.0, + 1191.0, + 1152.0, + 1228.0, + 947.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1191.0, + 1405.0, + 1191.0, + 1405.0, + 1228.0, + 1182.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 618.0, + 1223.0, + 618.0, + 1258.0, + 294.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1223.0, + 1063.0, + 1223.0, + 1063.0, + 1258.0, + 644.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1223.0, + 1406.0, + 1223.0, + 1406.0, + 1258.0, + 1094.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1249.0, + 1405.0, + 1249.0, + 1405.0, + 1290.0, + 292.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1282.0, + 1405.0, + 1282.0, + 1405.0, + 1319.0, + 293.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1313.0, + 679.0, + 1313.0, + 679.0, + 1352.0, + 294.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 929.0, + 1404.0, + 929.0, + 1404.0, + 966.0, + 294.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 961.0, + 1404.0, + 961.0, + 1404.0, + 996.0, + 294.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 985.0, + 348.0, + 985.0, + 348.0, + 1031.0, + 291.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 985.0, + 1118.0, + 985.0, + 1118.0, + 1031.0, + 374.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 985.0, + 1209.0, + 985.0, + 1209.0, + 1031.0, + 1149.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 985.0, + 1408.0, + 985.0, + 1408.0, + 1031.0, + 1398.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1022.0, + 1406.0, + 1022.0, + 1406.0, + 1059.0, + 292.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1054.0, + 1406.0, + 1054.0, + 1406.0, + 1088.0, + 292.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1086.0, + 379.0, + 1086.0, + 379.0, + 1119.0, + 294.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1854.0, + 295.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 486.0, + 1850.0, + 486.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1850.0, + 683.0, + 1850.0, + 683.0, + 1885.0, + 516.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1850.0, + 726.0, + 1850.0, + 726.0, + 1885.0, + 698.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1850.0, + 932.0, + 1850.0, + 932.0, + 1885.0, + 741.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1850.0, + 1407.0, + 1850.0, + 1407.0, + 1885.0, + 958.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1878.0, + 558.0, + 1878.0, + 558.0, + 1918.0, + 292.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 1878.0, + 890.0, + 1878.0, + 890.0, + 1918.0, + 813.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1878.0, + 1406.0, + 1878.0, + 1406.0, + 1918.0, + 1145.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 357.0, + 1911.0, + 357.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1911.0, + 579.0, + 1911.0, + 579.0, + 1946.0, + 380.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1911.0, + 926.0, + 1911.0, + 926.0, + 1946.0, + 682.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1946.0, + 952.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 466.0, + 1969.0, + 466.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 1969.0, + 1407.0, + 1969.0, + 1407.0, + 2008.0, + 699.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1250.0, + 2000.0, + 1250.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 1406.0, + 227.0, + 1406.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 263.0, + 393.0, + 263.0, + 393.0, + 300.0, + 295.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 263.0, + 845.0, + 263.0, + 845.0, + 300.0, + 481.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 263.0, + 943.0, + 263.0, + 943.0, + 300.0, + 933.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 263.0, + 1406.0, + 263.0, + 1406.0, + 300.0, + 1041.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 295.0, + 718.0, + 295.0, + 718.0, + 332.0, + 359.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 295.0, + 744.0, + 295.0, + 744.0, + 332.0, + 735.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 295.0, + 1099.0, + 295.0, + 1099.0, + 332.0, + 853.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 295.0, + 1222.0, + 295.0, + 1222.0, + 332.0, + 1115.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 322.0, + 379.0, + 322.0, + 379.0, + 363.0, + 292.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 322.0, + 1116.0, + 322.0, + 1116.0, + 363.0, + 400.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 322.0, + 1306.0, + 322.0, + 1306.0, + 363.0, + 1292.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 322.0, + 1407.0, + 322.0, + 1407.0, + 363.0, + 1404.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 356.0, + 345.0, + 356.0, + 345.0, + 402.0, + 292.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 356.0, + 580.0, + 356.0, + 580.0, + 402.0, + 442.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 356.0, + 658.0, + 356.0, + 658.0, + 402.0, + 608.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 356.0, + 788.0, + 356.0, + 788.0, + 402.0, + 686.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 356.0, + 1374.0, + 356.0, + 1374.0, + 402.0, + 852.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 356.0, + 1407.0, + 356.0, + 1407.0, + 402.0, + 1401.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 387.0, + 442.0, + 387.0, + 442.0, + 432.0, + 293.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1358.0, + 1405.0, + 1358.0, + 1405.0, + 1396.0, + 292.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1430.0, + 292.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 297.0, + 1421.0, + 297.0, + 1458.0, + 294.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1421.0, + 546.0, + 1421.0, + 546.0, + 1458.0, + 313.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1421.0, + 674.0, + 1421.0, + 674.0, + 1458.0, + 562.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1421.0, + 763.0, + 1421.0, + 763.0, + 1458.0, + 702.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1421.0, + 1406.0, + 1421.0, + 1406.0, + 1458.0, + 1029.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1451.0, + 1038.0, + 1451.0, + 1038.0, + 1489.0, + 292.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 574.0, + 539.0, + 574.0, + 539.0, + 616.0, + 294.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 574.0, + 615.0, + 574.0, + 615.0, + 616.0, + 607.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 574.0, + 1041.0, + 574.0, + 1041.0, + 616.0, + 845.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 574.0, + 1409.0, + 574.0, + 1409.0, + 616.0, + 1108.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 610.0, + 896.0, + 610.0, + 896.0, + 644.0, + 296.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 610.0, + 997.0, + 610.0, + 997.0, + 644.0, + 926.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 610.0, + 1139.0, + 610.0, + 1139.0, + 644.0, + 1013.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 610.0, + 1406.0, + 610.0, + 1406.0, + 644.0, + 1168.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 640.0, + 360.0, + 640.0, + 360.0, + 673.0, + 292.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 853.0, + 1408.0, + 853.0, + 1408.0, + 889.0, + 295.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 884.0, + 1408.0, + 884.0, + 1408.0, + 920.0, + 295.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 528.0, + 1119.0, + 528.0, + 1119.0, + 572.0, + 295.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 785.0, + 370.0, + 785.0, + 370.0, + 830.0, + 295.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 785.0, + 774.0, + 785.0, + 774.0, + 830.0, + 399.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 785.0, + 994.0, + 785.0, + 994.0, + 830.0, + 943.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 785.0, + 1047.0, + 785.0, + 1047.0, + 830.0, + 1038.0, + 830.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 464, + 1406, + 464, + 1406, + 564, + 298, + 564 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1405, + 228, + 1405, + 324, + 297, + 324 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 402, + 1718, + 1297, + 1718, + 1297, + 1909, + 402, + 1909 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 350, + 579, + 1349, + 579, + 1349, + 1028, + 350, + 1028 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 300, + 1616, + 1404, + 1616, + 1404, + 1703, + 300, + 1703 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 565, + 1207, + 1131, + 1207, + 1131, + 1303, + 565, + 1303 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 1114, + 1399, + 1114, + 1399, + 1191, + 297, + 1191 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 297, + 1516, + 1404, + 1516, + 1404, + 1581, + 297, + 1581 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 297, + 1378, + 1088, + 1378, + 1088, + 1437, + 297, + 1437 + ], + "score": 0.933 + }, + { + "category_id": 8, + "poly": [ + 764, + 1456, + 934, + 1456, + 934, + 1504, + 764, + 1504 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 295, + 1046, + 1298, + 1046, + 1298, + 1095, + 295, + 1095 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 295, + 1332, + 1202, + 1332, + 1202, + 1369, + 295, + 1369 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 817, + 74, + 817, + 106, + 298, + 106 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 297, + 360, + 886, + 360, + 886, + 393, + 297, + 393 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 295, + 416, + 1344, + 416, + 1344, + 453, + 295, + 453 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1364, + 1467, + 1401, + 1467, + 1401, + 1499, + 1364, + 1499 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 289, + 1924, + 1285, + 1924, + 1285, + 1974, + 289, + 1974 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1364, + 942, + 1402, + 942, + 1402, + 973, + 1364, + 973 + ], + "score": 0.866 + }, + { + "category_id": 9, + "poly": [ + 1364, + 1239, + 1401, + 1239, + 1401, + 1270, + 1364, + 1270 + ], + "score": 0.864 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1999, + 1404, + 1999, + 1404, + 2028, + 1374, + 2028 + ], + "score": 0.779 + }, + { + "category_id": 2, + "poly": [ + 840, + 2087, + 860, + 2087, + 860, + 2113, + 840, + 2113 + ], + "score": 0.742 + }, + { + "category_id": 1, + "poly": [ + 285, + 1986, + 1324, + 1986, + 1324, + 2044, + 285, + 2044 + ], + "score": 0.529 + }, + { + "category_id": 8, + "poly": [ + 285, + 1986, + 1324, + 1986, + 1324, + 2044, + 285, + 2044 + ], + "score": 0.502 + }, + { + "category_id": 2, + "poly": [ + 840, + 2087, + 860, + 2087, + 860, + 2113, + 840, + 2113 + ], + "score": 0.142 + }, + { + "category_id": 13, + "poly": [ + 703, + 527, + 933, + 527, + 933, + 561, + 703, + 561 + ], + "score": 0.94, + "latex": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 346, + 582, + 1348, + 582, + 1348, + 1034, + 346, + 1034 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { \\quad V ^ { * } ( s _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + Q ^ { * } ( s _ { t } , a _ { t } ) - V ^ { \\pi } ( s _ { t } ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\mathbb { P } ( V ^ { * } - V ^ { \\pi } ) ( s _ { t } , \\pi _ { t } ( s _ { t } ) ) } \\\\ & { = V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) + \\gamma \\sum _ { s + 1 } ^ { \\infty } p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\cdot [ V ^ { * } ( s _ { t + 1 } ) - Q ^ { * } ( s _ { t + 1 } , a _ { t + 1 } ) ] + } \\\\ & { \\quad \\Big . \\Big . \\Big . \\Big . \\Big . \\Big . } \\\\ & { \\quad \\quad \\Big . \\gamma _ { s + 1 , s ^ { * } + 2 } \\mathrm { ~ } p ( s _ { t + 2 } | s _ { t + 1 } , \\pi _ { t + 1 } ( s _ { t + 1 } ) ) \\cdot p ( s _ { t + 1 } | s _ { t } , \\pi _ { t } ( s _ { t } ) ) \\big [ V ^ { * } ( s _ { t + 2 } ) - Q ^ { * } ( s _ { t + 2 } , a _ { t + 2 } ) \\Big ] } \\\\ & { \\quad \\quad \\cdot \\dots } \\\\ & { \\le \\epsilon _ { 2 } + \\displaystyle \\sum _ { s ^ { \\prime } = s ^ { \\prime } } p ( t r a _ { t } ) \\cdot [ \\frac { n - 1 } { \\lambda _ { 2 } } \\gamma ^ { 2 } \\Delta _ { t + 2 } ] , } \\\\ & { \\quad \\quad \\pi _ { s \\lambda ^ { ( 2 ) } ( L ) } ^ { R , a } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 738, + 1046, + 847, + 1046, + 847, + 1096, + 738, + 1096 + ], + "score": 0.94, + "latex": "\\frac { \\gamma ^ { R } } { 1 - \\gamma } \\leq \\epsilon _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 863, + 1924, + 1125, + 1924, + 1125, + 1975, + 863, + 1975 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { \\infty } \\frac { \\gamma ^ { i - 1 } } { i } = \\frac { 1 } { \\gamma } \\ln \\frac { 1 } { 1 - \\gamma } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1124, + 1115, + 1292, + 1115, + 1292, + 1157, + 1124, + 1157 + ], + "score": 0.93, + "latex": "\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }" + }, + { + "category_id": 13, + "poly": [ + 1269, + 1616, + 1404, + 1616, + 1404, + 1660, + 1269, + 1660 + ], + "score": 0.93, + "latex": "E [ X _ { t } ^ { ( i ) } ] \\ \\leq" + }, + { + "category_id": 14, + "poly": [ + 564, + 1206, + 1134, + 1206, + 1134, + 1303, + 564, + 1303 + ], + "score": 0.93, + "latex": "V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E _ { t r a j } \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] ." + }, + { + "category_id": 13, + "poly": [ + 297, + 525, + 602, + 525, + 602, + 568, + 297, + 568 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { R : = \\lceil \\ln \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } / ( 1 - \\gamma ) \\rceil } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1656, + 458, + 1656, + 458, + 1702, + 298, + 1702 + ], + "score": 0.93, + "latex": "E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ]" + }, + { + "category_id": 13, + "poly": [ + 502, + 1617, + 557, + 1617, + 557, + 1659, + 502, + 1659 + ], + "score": 0.93, + "latex": "X _ { t } ^ { ( i ) }" + }, + { + "category_id": 14, + "poly": [ + 401, + 1716, + 1296, + 1716, + 1296, + 1911, + 401, + 1911 + ], + "score": 0.93, + "latex": "\\begin{array} { r l } & { E \\left[ \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] = \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( i ) } ] \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } E [ X _ { t } ^ { ( 2 ^ { \\lfloor \\log _ { 2 } i \\rfloor } ) } ] } \\\\ & { \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } 2 ^ { - \\lfloor \\log _ { 2 } i \\rfloor - 2 } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq \\displaystyle \\sum _ { i = 1 } ^ { R } \\frac { \\gamma ^ { i - 1 } } { i } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } \\leq 2 \\epsilon _ { 2 } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 586, + 1153, + 753, + 1153, + 753, + 1192, + 586, + 1192 + ], + "score": 0.93, + "latex": "\\left( \\Delta _ { t ^ { \\prime } } \\right) _ { t \\leq t ^ { \\prime } < t + R }" + }, + { + "category_id": 13, + "poly": [ + 500, + 1376, + 783, + 1376, + 783, + 1438, + 500, + 1438 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\xi _ { i } : = \\frac { 1 } { 2 ^ { i + 2 } } \\epsilon _ { 2 } \\left( \\ln \\frac { 1 } { 1 - \\gamma } \\right) ^ { - 1 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1093, + 1334, + 1195, + 1334, + 1195, + 1368, + 1093, + 1368 + ], + "score": 0.92, + "latex": "\\epsilon _ { 2 } = \\epsilon / 3" + }, + { + "category_id": 13, + "poly": [ + 1235, + 499, + 1316, + 499, + 1316, + 528, + 1235, + 528 + ], + "score": 0.91, + "latex": "\\epsilon _ { 2 } > 0" + }, + { + "category_id": 13, + "poly": [ + 732, + 1667, + 881, + 1667, + 881, + 1702, + 732, + 1702 + ], + "score": 0.91, + "latex": "1 / 2 < \\gamma < 1" + }, + { + "category_id": 13, + "poly": [ + 1291, + 1516, + 1404, + 1516, + 1404, + 1551, + 1291, + 1551 + ], + "score": 0.91, + "latex": "V ^ { \\ast } ( s _ { t } ) -" + }, + { + "category_id": 13, + "poly": [ + 298, + 1548, + 438, + 1548, + 438, + 1581, + 298, + 1581 + ], + "score": 0.91, + "latex": "V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1346, + 1110, + 1400, + 1110, + 1400, + 1151, + 1346, + 1151 + ], + "score": 0.91, + "latex": "X _ { t } ^ { ( i ) }" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1933, + 1273, + 1933, + 1273, + 1969, + 1176, + 1969 + ], + "score": 0.91, + "latex": "\\gamma > 1 / 2" + }, + { + "category_id": 14, + "poly": [ + 763, + 1452, + 936, + 1452, + 936, + 1502, + 763, + 1502 + ], + "score": 0.9, + "latex": "E [ X _ { t } ^ { ( 2 ^ { i } ) } ] \\leq \\xi _ { i } ." + }, + { + "category_id": 13, + "poly": [ + 880, + 1392, + 1077, + 1392, + 1077, + 1428, + 880, + 1428 + ], + "score": 0.88, + "latex": "0 \\leq i \\leq \\lfloor \\log _ { 2 } R \\rfloor" + }, + { + "category_id": 13, + "poly": [ + 972, + 528, + 1007, + 528, + 1007, + 558, + 972, + 558 + ], + "score": 0.88, + "latex": "\\Delta _ { t }" + }, + { + "category_id": 13, + "poly": [ + 663, + 1987, + 1322, + 1987, + 1322, + 2045, + 663, + 2045 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { V ^ { \\ast } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) \\leq \\epsilon _ { 2 } + E \\left[ \\sum _ { i = 1 } ^ { R } \\gamma ^ { i - 1 } X _ { t } ^ { ( i ) } \\right] \\leq 3 \\epsilon _ { 2 } = \\epsilon . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 840, + 1124, + 866, + 1124, + 866, + 1149, + 840, + 1149 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 959, + 1523, + 988, + 1523, + 988, + 1548, + 959, + 1548 + ], + "score": 0.85, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 628, + 502, + 656, + 502, + 656, + 528, + 628, + 528 + ], + "score": 0.85, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1268, + 1059, + 1290, + 1059, + 1290, + 1084, + 1268, + 1084 + ], + "score": 0.84, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 425, + 471, + 451, + 471, + 451, + 497, + 425, + 497 + ], + "score": 0.83, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1524, + 1239, + 1524, + 1239, + 1548, + 1213, + 1548 + ], + "score": 0.83, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1150, + 473, + 1175, + 473, + 1175, + 497, + 1150, + 497 + ], + "score": 0.83, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 665, + 1119, + 689, + 1119, + 689, + 1145, + 665, + 1145 + ], + "score": 0.82, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 837, + 468, + 861, + 468, + 861, + 494, + 837, + 494 + ], + "score": 0.8, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 1009, + 1338, + 1023, + 1338, + 1023, + 1362, + 1009, + 1362 + ], + "score": 0.77, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 734, + 1341, + 748, + 1341, + 748, + 1362, + 734, + 1362 + ], + "score": 0.76, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 962, + 500, + 975, + 500, + 975, + 525, + 962, + 525 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 371, + 1156, + 384, + 1156, + 384, + 1181, + 371, + 1181 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1094, + 1628, + 1108, + 1628, + 1108, + 1653, + 1094, + 1653 + ], + "score": 0.73, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1014, + 503, + 1029, + 503, + 1029, + 524, + 1014, + 524 + ], + "score": 0.72, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1321, + 423, + 1334, + 423, + 1334, + 447, + 1321, + 447 + ], + "score": 0.71, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 551, + 1520, + 566, + 1520, + 566, + 1545, + 551, + 1545 + ], + "score": 0.7, + "latex": "^ { l }" + }, + { + "category_id": 13, + "poly": [ + 1017, + 1524, + 1031, + 1524, + 1031, + 1545, + 1017, + 1545 + ], + "score": 0.68, + "latex": "\\epsilon" + }, + { + "category_id": 14, + "poly": [ + 354, + 903, + 816, + 903, + 816, + 1028, + 354, + 1028 + ], + "score": 0.68, + "latex": "\\leq \\epsilon _ { 2 } + \\sum _ { \\stackrel { t r a j \\in \\epsilon } { \\mathrm { T R A J } ( R ) } } p ( t r a j ) \\cdot \\left[ \\sum _ { j = 0 } ^ { R - 1 } \\gamma ^ { j } \\Delta _ { t + j } \\right] ," + }, + { + "category_id": 13, + "poly": [ + 711, + 368, + 725, + 368, + 725, + 389, + 711, + 389 + ], + "score": 0.67, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 568, + 468, + 608, + 468, + 608, + 498, + 568, + 498 + ], + "score": 0.66, + "latex": "( R )" + }, + { + "category_id": 13, + "poly": [ + 1046, + 427, + 1059, + 427, + 1059, + 447, + 1046, + 447 + ], + "score": 0.65, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 298, + 268, + 312, + 268, + 312, + 289, + 298, + 289 + ], + "score": 0.64, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 822, + 1521, + 836, + 1521, + 836, + 1546, + 822, + 1546 + ], + "score": 0.51, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 360.0, + 710.0, + 360.0, + 710.0, + 396.0, + 295.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 360.0, + 887.0, + 360.0, + 887.0, + 396.0, + 726.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1380.0, + 2004.0, + 1402.0, + 2004.0, + 1402.0, + 2027.0, + 1380.0, + 2027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 839.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 839.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 461.0, + 424.0, + 461.0, + 424.0, + 504.0, + 292.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 461.0, + 567.0, + 461.0, + 567.0, + 504.0, + 452.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 461.0, + 836.0, + 461.0, + 836.0, + 504.0, + 609.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 461.0, + 1149.0, + 461.0, + 1149.0, + 504.0, + 862.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 461.0, + 1406.0, + 461.0, + 1406.0, + 504.0, + 1176.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 496.0, + 627.0, + 496.0, + 627.0, + 531.0, + 293.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 496.0, + 961.0, + 496.0, + 961.0, + 531.0, + 657.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 496.0, + 1013.0, + 496.0, + 1013.0, + 531.0, + 976.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 496.0, + 1234.0, + 496.0, + 1234.0, + 531.0, + 1030.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 496.0, + 1404.0, + 496.0, + 1404.0, + 531.0, + 1317.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 522.0, + 296.0, + 522.0, + 296.0, + 573.0, + 292.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 522.0, + 702.0, + 522.0, + 702.0, + 573.0, + 603.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 522.0, + 971.0, + 522.0, + 971.0, + 573.0, + 934.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 522.0, + 1129.0, + 522.0, + 1129.0, + 573.0, + 1008.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 296.0, + 313.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 465.0, + 291.0, + 465.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1609.0, + 501.0, + 1609.0, + 501.0, + 1667.0, + 288.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1609.0, + 1093.0, + 1609.0, + 1093.0, + 1667.0, + 558.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1609.0, + 1268.0, + 1609.0, + 1268.0, + 1667.0, + 1109.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1405.0, + 1609.0, + 1409.0, + 1609.0, + 1409.0, + 1667.0, + 1405.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 275.0, + 1635.0, + 297.0, + 1635.0, + 297.0, + 1719.0, + 275.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1635.0, + 731.0, + 1635.0, + 731.0, + 1719.0, + 459.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1635.0, + 910.0, + 1635.0, + 910.0, + 1719.0, + 882.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1107.0, + 664.0, + 1107.0, + 664.0, + 1163.0, + 288.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1107.0, + 839.0, + 1107.0, + 839.0, + 1163.0, + 690.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1107.0, + 1123.0, + 1107.0, + 1123.0, + 1163.0, + 867.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1107.0, + 1345.0, + 1107.0, + 1345.0, + 1163.0, + 1293.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1107.0, + 1405.0, + 1107.0, + 1405.0, + 1163.0, + 1401.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1148.0, + 370.0, + 1148.0, + 370.0, + 1197.0, + 292.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1148.0, + 585.0, + 1148.0, + 585.0, + 1197.0, + 385.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1148.0, + 1118.0, + 1148.0, + 1118.0, + 1197.0, + 754.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1510.0, + 550.0, + 1510.0, + 550.0, + 1556.0, + 292.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 1510.0, + 821.0, + 1510.0, + 821.0, + 1556.0, + 567.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1510.0, + 958.0, + 1510.0, + 958.0, + 1556.0, + 837.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1510.0, + 1016.0, + 1510.0, + 1016.0, + 1556.0, + 989.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1510.0, + 1212.0, + 1510.0, + 1212.0, + 1556.0, + 1032.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1510.0, + 1290.0, + 1510.0, + 1290.0, + 1556.0, + 1240.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1539.0, + 297.0, + 1539.0, + 297.0, + 1588.0, + 293.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1539.0, + 454.0, + 1539.0, + 454.0, + 1588.0, + 439.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1373.0, + 499.0, + 1373.0, + 499.0, + 1440.0, + 286.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1373.0, + 879.0, + 1373.0, + 879.0, + 1440.0, + 784.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1373.0, + 1096.0, + 1373.0, + 1096.0, + 1440.0, + 1078.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1026.0, + 737.0, + 1026.0, + 737.0, + 1115.0, + 292.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1026.0, + 1267.0, + 1026.0, + 1267.0, + 1115.0, + 848.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1026.0, + 1317.0, + 1026.0, + 1317.0, + 1115.0, + 1291.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1329.0, + 733.0, + 1329.0, + 733.0, + 1374.0, + 292.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1329.0, + 1008.0, + 1329.0, + 1008.0, + 1374.0, + 749.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1329.0, + 1092.0, + 1329.0, + 1092.0, + 1374.0, + 1024.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1329.0, + 1207.0, + 1329.0, + 1207.0, + 1374.0, + 1196.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 412.0, + 1045.0, + 412.0, + 1045.0, + 459.0, + 292.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 412.0, + 1320.0, + 412.0, + 1320.0, + 459.0, + 1060.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 412.0, + 1347.0, + 412.0, + 1347.0, + 459.0, + 1335.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1914.0, + 862.0, + 1914.0, + 862.0, + 1988.0, + 285.0, + 1988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 1914.0, + 1175.0, + 1914.0, + 1175.0, + 1988.0, + 1126.0, + 1988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1914.0, + 1293.0, + 1914.0, + 1293.0, + 1988.0, + 1274.0, + 1988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1979.0, + 662.0, + 1979.0, + 662.0, + 2048.0, + 288.0, + 2048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1323.0, + 1979.0, + 1334.0, + 1979.0, + 1334.0, + 2048.0, + 1323.0, + 2048.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 855, + 1406, + 855, + 1406, + 1014, + 296, + 1014 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 296, + 1107, + 1406, + 1107, + 1406, + 1237, + 296, + 1237 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 520, + 1340, + 1177, + 1340, + 1177, + 1425, + 520, + 1425 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 291, + 449, + 1408, + 449, + 1408, + 519, + 291, + 519 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 385, + 537, + 1310, + 537, + 1310, + 630, + 385, + 630 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 295, + 272, + 1408, + 272, + 1408, + 373, + 295, + 373 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 691, + 775, + 1007, + 775, + 1007, + 840, + 691, + 840 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 356, + 1716, + 1339, + 1716, + 1339, + 1812, + 356, + 1812 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 293, + 1241, + 1403, + 1241, + 1403, + 1318, + 293, + 1318 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 299, + 1444, + 690, + 1444, + 690, + 1480, + 299, + 1480 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 298, + 1050, + 536, + 1050, + 536, + 1083, + 298, + 1083 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 293, + 1546, + 1402, + 1546, + 1402, + 1616, + 293, + 1616 + ], + "score": 0.919 + }, + { + "category_id": 1, + "poly": [ + 298, + 1847, + 805, + 1847, + 805, + 1892, + 298, + 1892 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 293, + 720, + 1404, + 720, + 1404, + 758, + 293, + 758 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 295, + 378, + 973, + 378, + 973, + 413, + 295, + 413 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 290, + 1914, + 1300, + 1914, + 1300, + 1951, + 290, + 1951 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 297, + 1503, + 1223, + 1503, + 1223, + 1538, + 297, + 1538 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 297, + 227, + 939, + 227, + 939, + 264, + 297, + 264 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 293, + 1624, + 1349, + 1624, + 1349, + 1662, + 293, + 1662 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1365, + 793, + 1400, + 793, + 1400, + 824, + 1365, + 824 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1364, + 1366, + 1401, + 1366, + 1401, + 1397, + 1364, + 1397 + ], + "score": 0.846 + }, + { + "category_id": 2, + "poly": [ + 839, + 2088, + 859, + 2088, + 859, + 2113, + 839, + 2113 + ], + "score": 0.815 + }, + { + "category_id": 2, + "poly": [ + 1374, + 648, + 1403, + 648, + 1403, + 678, + 1374, + 678 + ], + "score": 0.812 + }, + { + "category_id": 2, + "poly": [ + 296, + 1973, + 1411, + 1973, + 1411, + 2036, + 296, + 2036 + ], + "score": 0.812 + }, + { + "category_id": 14, + "poly": [ + 690, + 771, + 1010, + 771, + 1010, + 840, + 690, + 840 + ], + "score": 0.93, + "latex": "\\eta _ { j } \\mathrm { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } ." + }, + { + "category_id": 13, + "poly": [ + 914, + 723, + 1110, + 723, + 1110, + 757, + 914, + 757 + ], + "score": 0.93, + "latex": "0 \\leq i < \\lfloor \\log _ { 2 } R \\rfloor" + }, + { + "category_id": 14, + "poly": [ + 387, + 533, + 1309, + 533, + 1309, + 632, + 387, + 632 + ], + "score": 0.93, + "latex": "E \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } \\right] = \\int _ { 0 } ^ { 1 / ( 1 - \\gamma ) } \\operatorname* { P r } \\left[ { X _ { t } ^ { ( 2 ^ { i } ) } } > x \\right] d x \\le \\eta _ { 1 } + \\sum _ { j = 2 } ^ { M } \\eta _ { j } \\operatorname* { P r } [ { X _ { t } ^ { ( 2 ^ { i } ) } } > \\eta _ { j - 1 } ] \\le \\xi _ { i } ." + }, + { + "category_id": 13, + "poly": [ + 663, + 919, + 953, + 919, + 953, + 971, + 663, + 971 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1199, + 1972, + 1404, + 1972, + 1404, + 2008, + 1199, + 2008 + ], + "score": 0.93, + "latex": "2 ^ { M / 2 - 1 } > ( M + 1 )" + }, + { + "category_id": 13, + "poly": [ + 661, + 485, + 794, + 485, + 794, + 517, + 661, + 517 + ], + "score": 0.92, + "latex": "1 \\leq j \\leq M" + }, + { + "category_id": 13, + "poly": [ + 297, + 967, + 454, + 967, + 454, + 1014, + 297, + 1014 + ], + "score": 0.92, + "latex": "X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 }" + }, + { + "category_id": 14, + "poly": [ + 355, + 1716, + 1342, + 1716, + 1342, + 1810, + 355, + 1810 + ], + "score": 0.92, + "latex": "\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 1001, + 857, + 1080, + 857, + 1080, + 890, + 1001, + 890 + ], + "score": 0.92, + "latex": "( t , i , j )" + }, + { + "category_id": 13, + "poly": [ + 536, + 1170, + 720, + 1170, + 720, + 1208, + 536, + 1208 + ], + "score": 0.92, + "latex": "( \\hat { Q } _ { t } - Q ^ { * } ) ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 929, + 1973, + 1154, + 1973, + 1154, + 2009, + 929, + 2009 + ], + "score": 0.92, + "latex": "\\xi _ { i } \\cdot 2 ^ { M / 2 } \\geq 1 / ( 1 - \\gamma )" + }, + { + "category_id": 14, + "poly": [ + 519, + 1333, + 1176, + 1333, + 1176, + 1430, + 519, + 1430 + ], + "score": 0.92, + "latex": "\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 735, + 1626, + 825, + 1626, + 825, + 1661, + 735, + 1661 + ], + "score": 0.92, + "latex": "( w _ { t } ) _ { t \\geq 1 }" + }, + { + "category_id": 13, + "poly": [ + 458, + 1580, + 619, + 1580, + 619, + 1618, + 458, + 1618 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\sum _ { t \\geq 1 } w _ { t } \\leq C } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 936, + 888, + 993, + 888, + 993, + 921, + 936, + 921 + ], + "score": 0.92, + "latex": "( i , j )" + }, + { + "category_id": 13, + "poly": [ + 591, + 1548, + 681, + 1548, + 681, + 1582, + 591, + 1582 + ], + "score": 0.91, + "latex": "( w _ { t } ) _ { t \\geq 1 }" + }, + { + "category_id": 13, + "poly": [ + 543, + 1626, + 618, + 1626, + 618, + 1660, + 543, + 1660 + ], + "score": 0.91, + "latex": "( C , w )" + }, + { + "category_id": 13, + "poly": [ + 1163, + 724, + 1295, + 724, + 1295, + 756, + 1163, + 756 + ], + "score": 0.91, + "latex": "2 \\leq j \\leq M" + }, + { + "category_id": 13, + "poly": [ + 712, + 1240, + 764, + 1240, + 764, + 1282, + 712, + 1282 + ], + "score": 0.91, + "latex": "{ B } _ { \\eta } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 957, + 1247, + 1272, + 1247, + 1272, + 1288, + 957, + 1288 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 992, + 452, + 1203, + 452, + 1203, + 487, + 992, + 487 + ], + "score": 0.91, + "latex": "\\eta _ { M } > 1 / ( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 541, + 283, + 703, + 283, + 703, + 318, + 541, + 318 + ], + "score": 0.91, + "latex": "L = \\lfloor \\log _ { 2 } R \\rfloor" + }, + { + "category_id": 13, + "poly": [ + 371, + 1849, + 636, + 1849, + 636, + 1892, + 371, + 1892 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1626, + 1112, + 1626, + 1112, + 1660, + 1020, + 1660 + ], + "score": 0.9, + "latex": "1 - \\delta / 2" + }, + { + "category_id": 13, + "poly": [ + 726, + 1284, + 816, + 1284, + 816, + 1318, + 726, + 1318 + ], + "score": 0.9, + "latex": "1 - \\delta / 2" + }, + { + "category_id": 13, + "poly": [ + 849, + 1548, + 924, + 1548, + 924, + 1582, + 849, + 1582 + ], + "score": 0.9, + "latex": "( C , w )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1285, + 389, + 1285, + 389, + 1317, + 297, + 1317 + ], + "score": 0.89, + "latex": "\\eta > 2 \\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 598, + 1249, + 667, + 1249, + 667, + 1281, + 598, + 1281 + ], + "score": 0.89, + "latex": "\\eta > 0" + }, + { + "category_id": 13, + "poly": [ + 348, + 1976, + 519, + 1976, + 519, + 2008, + 348, + 2008 + ], + "score": 0.88, + "latex": "\\eta _ { M } > 1 / ( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 334, + 1580, + 396, + 1580, + 396, + 1611, + 334, + 1611 + ], + "score": 0.88, + "latex": "t \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 761, + 272, + 1158, + 272, + 1158, + 329, + 761, + 329 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { M = \\operatorname* { m a x } \\left\\{ \\lceil 2 \\log _ { 2 } \\frac { 1 } { \\xi _ { L } ( 1 - \\gamma ) } \\rceil , 1 0 \\right\\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 371, + 1446, + 411, + 1446, + 411, + 1480, + 371, + 1480 + ], + "score": 0.87, + "latex": "I [ \\cdot ]" + }, + { + "category_id": 13, + "poly": [ + 1222, + 277, + 1397, + 277, + 1397, + 321, + 1222, + 321 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\eta _ { j } = \\frac { \\xi _ { i } } { M } \\cdot 2 ^ { j - 1 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 652, + 729, + 678, + 729, + 678, + 754, + 652, + 754 + ], + "score": 0.85, + "latex": "\\epsilon _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 760, + 455, + 793, + 455, + 793, + 481, + 760, + 481 + ], + "score": 0.83, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 577, + 232, + 612, + 232, + 612, + 262, + 577, + 262 + ], + "score": 0.82, + "latex": "i , t" + }, + { + "category_id": 13, + "poly": [ + 1081, + 1549, + 1187, + 1549, + 1187, + 1580, + 1081, + 1580 + ], + "score": 0.82, + "latex": "C , w > 0" + }, + { + "category_id": 13, + "poly": [ + 468, + 2007, + 498, + 2007, + 498, + 2030, + 468, + 2030 + ], + "score": 0.8, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1549, + 1361, + 1549, + 1361, + 1581, + 1200, + 1581 + ], + "score": 0.79, + "latex": "i f 0 \\le w _ { t } \\le w" + }, + { + "category_id": 13, + "poly": [ + 568, + 728, + 581, + 728, + 581, + 751, + 568, + 751 + ], + "score": 0.79, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 389, + 325, + 823, + 325, + 823, + 373, + 389, + 373 + ], + "score": 0.75, + "latex": "\\begin{array} { r } { \\because j \\le M , \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] \\le \\frac { \\xi _ { i } } { M } , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 539, + 933, + 553, + 933, + 553, + 957, + 539, + 957 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1110, + 982, + 1123, + 982, + 1123, + 1006, + 1110, + 1006 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 378, + 334, + 510, + 334, + 510, + 370, + 378, + 370 + ], + "score": 0.63, + "latex": "2 \\leq j \\leq M" + }, + { + "category_id": 13, + "poly": [ + 533, + 1251, + 547, + 1251, + 547, + 1276, + 533, + 1276 + ], + "score": 0.25, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1046.0, + 538.0, + 1046.0, + 538.0, + 1087.0, + 293.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 651.0, + 1404.0, + 651.0, + 1404.0, + 682.0, + 1376.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1965.0, + 347.0, + 1965.0, + 347.0, + 2013.0, + 331.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1965.0, + 928.0, + 1965.0, + 928.0, + 2013.0, + 520.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1965.0, + 1198.0, + 1965.0, + 1198.0, + 2013.0, + 1155.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1405.0, + 1965.0, + 1409.0, + 1965.0, + 1409.0, + 2013.0, + 1405.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2003.0, + 467.0, + 2003.0, + 467.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 2003.0, + 511.0, + 2003.0, + 511.0, + 2036.0, + 499.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 855.0, + 1000.0, + 855.0, + 1000.0, + 893.0, + 294.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 855.0, + 1408.0, + 855.0, + 1408.0, + 893.0, + 1081.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 886.0, + 935.0, + 886.0, + 935.0, + 924.0, + 294.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 886.0, + 1404.0, + 886.0, + 1404.0, + 924.0, + 994.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 919.0, + 538.0, + 919.0, + 538.0, + 976.0, + 283.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 919.0, + 662.0, + 919.0, + 662.0, + 976.0, + 554.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 919.0, + 1405.0, + 919.0, + 1405.0, + 976.0, + 954.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 971.0, + 1109.0, + 971.0, + 1109.0, + 1018.0, + 455.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 971.0, + 1138.0, + 971.0, + 1138.0, + 1018.0, + 1124.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 966.0, + 378.0, + 966.0, + 378.0, + 1018.0, + 288.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1105.0, + 1409.0, + 1105.0, + 1409.0, + 1146.0, + 293.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1140.0, + 1404.0, + 1140.0, + 1404.0, + 1173.0, + 295.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1172.0, + 535.0, + 1172.0, + 535.0, + 1209.0, + 294.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1172.0, + 1406.0, + 1172.0, + 1406.0, + 1209.0, + 721.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 680.0, + 1204.0, + 680.0, + 1238.0, + 294.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 451.0, + 759.0, + 451.0, + 759.0, + 488.0, + 296.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 451.0, + 991.0, + 451.0, + 991.0, + 488.0, + 794.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 451.0, + 1407.0, + 451.0, + 1407.0, + 488.0, + 1204.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 483.0, + 660.0, + 483.0, + 660.0, + 518.0, + 293.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 483.0, + 806.0, + 483.0, + 806.0, + 518.0, + 795.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 265.0, + 540.0, + 265.0, + 540.0, + 335.0, + 288.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 265.0, + 760.0, + 265.0, + 760.0, + 335.0, + 704.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 265.0, + 1221.0, + 265.0, + 1221.0, + 335.0, + 1159.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 265.0, + 1410.0, + 265.0, + 1410.0, + 335.0, + 1398.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 320.0, + 377.0, + 320.0, + 377.0, + 380.0, + 290.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 320.0, + 835.0, + 320.0, + 835.0, + 380.0, + 824.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1231.0, + 532.0, + 1231.0, + 532.0, + 1293.0, + 289.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1231.0, + 597.0, + 1231.0, + 597.0, + 1293.0, + 548.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1231.0, + 711.0, + 1231.0, + 711.0, + 1293.0, + 668.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1231.0, + 956.0, + 1231.0, + 956.0, + 1293.0, + 765.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1231.0, + 1411.0, + 1231.0, + 1411.0, + 1293.0, + 1273.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1281.0, + 296.0, + 1281.0, + 296.0, + 1321.0, + 292.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1281.0, + 725.0, + 1281.0, + 725.0, + 1321.0, + 390.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1281.0, + 829.0, + 1281.0, + 829.0, + 1321.0, + 817.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1445.0, + 370.0, + 1445.0, + 370.0, + 1482.0, + 297.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1445.0, + 692.0, + 1445.0, + 692.0, + 1482.0, + 412.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1541.0, + 590.0, + 1541.0, + 590.0, + 1588.0, + 291.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1541.0, + 848.0, + 1541.0, + 848.0, + 1588.0, + 682.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1541.0, + 1080.0, + 1541.0, + 1080.0, + 1588.0, + 925.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1541.0, + 1199.0, + 1541.0, + 1199.0, + 1588.0, + 1188.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 1541.0, + 1407.0, + 1541.0, + 1407.0, + 1588.0, + 1362.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1576.0, + 333.0, + 1576.0, + 333.0, + 1618.0, + 293.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1576.0, + 457.0, + 1576.0, + 457.0, + 1618.0, + 397.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1576.0, + 631.0, + 1576.0, + 631.0, + 1618.0, + 620.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1824.0, + 370.0, + 1824.0, + 370.0, + 1911.0, + 284.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1824.0, + 821.0, + 1824.0, + 821.0, + 1911.0, + 637.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 717.0, + 567.0, + 717.0, + 567.0, + 761.0, + 293.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 717.0, + 651.0, + 717.0, + 651.0, + 761.0, + 582.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 717.0, + 913.0, + 717.0, + 913.0, + 761.0, + 679.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 717.0, + 1162.0, + 717.0, + 1162.0, + 761.0, + 1111.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 717.0, + 1406.0, + 717.0, + 1406.0, + 761.0, + 1296.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 378.0, + 971.0, + 378.0, + 971.0, + 414.0, + 298.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1912.0, + 1301.0, + 1912.0, + 1301.0, + 1955.0, + 292.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1502.0, + 1224.0, + 1502.0, + 1224.0, + 1543.0, + 292.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 221.0, + 576.0, + 221.0, + 576.0, + 272.0, + 292.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 221.0, + 942.0, + 221.0, + 942.0, + 272.0, + 613.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1620.0, + 542.0, + 1620.0, + 542.0, + 1666.0, + 291.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1620.0, + 734.0, + 1620.0, + 734.0, + 1666.0, + 619.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1620.0, + 1019.0, + 1620.0, + 1019.0, + 1666.0, + 826.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1620.0, + 1352.0, + 1620.0, + 1352.0, + 1666.0, + 1113.0, + 1666.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 294, + 1128, + 1407, + 1128, + 1407, + 1315, + 294, + 1315 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 295, + 348, + 1407, + 348, + 1407, + 516, + 295, + 516 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 296, + 615, + 1404, + 615, + 1404, + 733, + 296, + 733 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 295, + 1444, + 1407, + 1444, + 1407, + 1657, + 295, + 1657 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 317, + 1669, + 1403, + 1669, + 1403, + 2033, + 317, + 2033 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 295, + 228, + 1405, + 228, + 1405, + 299, + 295, + 299 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 602, + 1323, + 1095, + 1323, + 1095, + 1390, + 602, + 1390 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 408, + 521, + 1286, + 521, + 1286, + 607, + 408, + 607 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 582, + 306, + 1119, + 306, + 1119, + 346, + 582, + 346 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 298, + 1397, + 1077, + 1397, + 1077, + 1434, + 298, + 1434 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 296, + 821, + 985, + 821, + 985, + 857, + 296, + 857 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 299, + 765, + 651, + 765, + 651, + 797, + 299, + 797 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 297, + 1061, + 811, + 1061, + 811, + 1114, + 297, + 1114 + ], + "score": 0.919 + }, + { + "category_id": 9, + "poly": [ + 1365, + 997, + 1401, + 997, + 1401, + 1028, + 1365, + 1028 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 298, + 888, + 634, + 888, + 634, + 921, + 298, + 921 + ], + "score": 0.873 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1339, + 1401, + 1339, + 1401, + 1369, + 1365, + 1369 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 841, + 2086, + 858, + 2086, + 858, + 2112, + 841, + 2112 + ], + "score": 0.773 + }, + { + "category_id": 9, + "poly": [ + 1174, + 1975, + 1403, + 1975, + 1403, + 2009, + 1174, + 2009 + ], + "score": 0.588 + }, + { + "category_id": 8, + "poly": [ + 696, + 977, + 1000, + 977, + 1000, + 1053, + 696, + 1053 + ], + "score": 0.549 + }, + { + "category_id": 1, + "poly": [ + 298, + 931, + 1249, + 931, + 1249, + 1053, + 298, + 1053 + ], + "score": 0.228 + }, + { + "category_id": 8, + "poly": [ + 298, + 931, + 1249, + 931, + 1249, + 1053, + 298, + 1053 + ], + "score": 0.114 + }, + { + "category_id": 13, + "poly": [ + 521, + 647, + 644, + 647, + 644, + 701, + 521, + 701 + ], + "score": 0.93, + "latex": "\\tilde { \\mathcal { O } } \\left( \\sqrt { | J | } \\right)" + }, + { + "category_id": 13, + "poly": [ + 885, + 1232, + 1222, + 1232, + 1222, + 1289, + 885, + 1289 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { T = \\mathcal { O } \\left( \\frac { C } { 2 ^ { i } } \\cdot \\frac { M \\eta _ { j } } { \\xi _ { i } } \\ln ( R M L ) \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 342, + 1128, + 596, + 1128, + 596, + 1173, + 342, + 1173 + ], + "score": 0.93, + "latex": "A _ { t } = I [ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } ]" + }, + { + "category_id": 14, + "poly": [ + 604, + 1318, + 1094, + 1318, + 1094, + 1392, + 604, + 1392 + ], + "score": 0.93, + "latex": "\\operatorname* { P r } \\left[ X _ { k + t R } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] = \\mathbb { E } [ A _ { k + t R } ] > \\frac { \\xi _ { i } } { M \\eta _ { j } } ," + }, + { + "category_id": 14, + "poly": [ + 408, + 518, + 1291, + 518, + 1291, + 611, + 408, + 611 + ], + "score": 0.93, + "latex": "| J | \\eta ( 1 - \\gamma ) ^ { - 1 } \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { S A | J | \\ell ( | J | ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln | J | } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 767, + 261, + 872, + 261, + 872, + 298, + 767, + 298 + ], + "score": 0.93, + "latex": "\\hat { Q } _ { t } \\geq Q ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 394, + 1397, + 611, + 1397, + 611, + 1436, + 394, + 1436 + ], + "score": 0.93, + "latex": "\\textstyle \\sum _ { t } A _ { k + t R } \\geq C / 2 ^ { i }" + }, + { + "category_id": 14, + "poly": [ + 315, + 1678, + 1424, + 1678, + 1424, + 2039, + 315, + 2039 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } { { \\sum _ { t = 1 } ^ { \\ell - 1 } I [ V ^ { * } ( s _ { t } ) - V ^ { \\pi _ { t } } ( s _ { t } ) > \\epsilon ] } } \\\\ & { \\leq \\sum _ { t = 1 } ^ { \\infty } \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } I [ \\eta _ { j } \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { M } ] = \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\sum _ { t = 1 } ^ { \\infty } I [ \\operatorname* { P r } [ X _ { t } ^ { ( 2 ^ { i } ) } > \\eta _ { j - 1 } ] > \\frac { \\xi _ { i } } { \\eta _ { j } M } ] } \\\\ & { \\leq \\sum _ { i = 0 } ^ { L } \\sum _ { j = 2 } ^ { M } \\frac { S A M R \\ln 1 / \\delta \\ln S A } { \\eta _ { j } \\xi _ { i } \\cdot 2 ^ { i } ( 1 - \\gamma ) ^ { 5 } } \\tilde { P } \\leq \\sum _ { i = 0 } ^ { L } \\frac { S A \\cdot 2 ^ { i + 4 } \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\xi _ { i } \\mathrm { ~ a n d } \\eta _ { j } ) } \\\\ & { \\leq \\frac { S A R \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\tilde { P } \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\tilde { P } . } & { \\mathrm { ( B y ~ d e f i n i t i o n ~ o f } \\hbar ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1010, + 1139, + 1106, + 1139, + 1106, + 1173, + 1010, + 1173 + ], + "score": 0.92, + "latex": "\\{ \\mathcal { F } _ { t } \\} _ { t \\ge 1 }" + }, + { + "category_id": 13, + "poly": [ + 810, + 1571, + 969, + 1571, + 969, + 1605, + 810, + 1605 + ], + "score": 0.92, + "latex": "1 - \\delta / ( 2 M L )" + }, + { + "category_id": 13, + "poly": [ + 1136, + 448, + 1330, + 448, + 1330, + 486, + 1136, + 486 + ], + "score": 0.92, + "latex": "( \\hat { Q } - Q ^ { * } ) ( s _ { t } , a _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1233, + 410, + 1397, + 410, + 1397, + 445, + 1233, + 445 + ], + "score": 0.92, + "latex": "| J | \\bar { \\eta ( 1 - \\gamma ) } ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 929, + 1509, + 1180, + 1509, + 1180, + 1545, + 929, + 1545 + ], + "score": 0.92, + "latex": "\\textstyle \\sum _ { t = 0 } ^ { \\infty } A _ { k + t R } \\leq C / 2 ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 829, + 1447, + 1135, + 1447, + 1135, + 1479, + 829, + 1479 + ], + "score": 0.92, + "latex": "[ k + t R , k + t R + R - 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1313, + 350, + 1395, + 350, + 1395, + 384, + 1313, + 384 + ], + "score": 0.91, + "latex": "( | J | , 1 )" + }, + { + "category_id": 13, + "poly": [ + 635, + 383, + 817, + 383, + 817, + 414, + 635, + 414 + ], + "score": 0.91, + "latex": "w _ { t } ~ = ~ I \\left[ t \\in J \\right]" + }, + { + "category_id": 13, + "poly": [ + 1309, + 657, + 1344, + 657, + 1344, + 691, + 1309, + 691 + ], + "score": 0.91, + "latex": "| J |" + }, + { + "category_id": 13, + "poly": [ + 563, + 1478, + 896, + 1478, + 896, + 1511, + 563, + 1511 + ], + "score": 0.91, + "latex": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > ^ { \\mathsf { ^ { * } } } \\eta _ { j - 1 }" + }, + { + "category_id": 13, + "poly": [ + 348, + 1511, + 648, + 1511, + 648, + 1544, + 348, + 1544 + ], + "score": 0.9, + "latex": "[ k + t R , k + t R + R - 1 ]" + }, + { + "category_id": 13, + "poly": [ + 551, + 1448, + 630, + 1448, + 630, + 1479, + 551, + 1479 + ], + "score": 0.9, + "latex": "A _ { k + t R }" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1170, + 1397, + 1170, + 1397, + 1201, + 1266, + 1201 + ], + "score": 0.9, + "latex": "0 \\leq k < R" + }, + { + "category_id": 13, + "poly": [ + 298, + 1242, + 427, + 1242, + 427, + 1274, + 298, + 1274 + ], + "score": 0.9, + "latex": "0 \\leq k < R" + }, + { + "category_id": 13, + "poly": [ + 298, + 1572, + 425, + 1572, + 425, + 1602, + 298, + 1602 + ], + "score": 0.9, + "latex": "0 \\leq k < R" + }, + { + "category_id": 14, + "poly": [ + 696, + 976, + 1001, + 976, + 1001, + 1054, + 696, + 1054 + ], + "score": 0.89, + "latex": "C = \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta _ { j - 1 } ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\tilde { P } ." + }, + { + "category_id": 13, + "poly": [ + 885, + 1398, + 1067, + 1398, + 1067, + 1433, + 885, + 1433 + ], + "score": 0.89, + "latex": "1 - \\delta / ( 2 M R L )" + }, + { + "category_id": 13, + "poly": [ + 673, + 616, + 708, + 616, + 708, + 649, + 673, + 649 + ], + "score": 0.89, + "latex": "| J |" + }, + { + "category_id": 13, + "poly": [ + 923, + 261, + 1204, + 261, + 1204, + 299, + 923, + 299 + ], + "score": 0.89, + "latex": "a _ { t } = \\arg \\operatorname* { m a x } _ { a } \\hat { Q } _ { t } ( s _ { t } , a )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1200, + 690, + 1200, + 690, + 1235, + 298, + 1235 + ], + "score": 0.88, + "latex": "\\{ \\dot { A } _ { k + t R } - E [ A _ { k + t R } \\ | \\ \\mathcal { F } _ { k + t R } ] \\} _ { t \\geq 0 }" + }, + { + "category_id": 13, + "poly": [ + 884, + 1171, + 917, + 1171, + 917, + 1201, + 884, + 1201 + ], + "score": 0.88, + "latex": "A _ { t }" + }, + { + "category_id": 13, + "poly": [ + 537, + 1170, + 798, + 1170, + 798, + 1203, + 537, + 1203 + ], + "score": 0.88, + "latex": "\\bar { \\{ ( s _ { \\tau } , a _ { \\tau } ) : 1 \\leq \\tau \\leq t \\} }" + }, + { + "category_id": 13, + "poly": [ + 1219, + 616, + 1255, + 616, + 1255, + 650, + 1219, + 650 + ], + "score": 0.88, + "latex": "| J |" + }, + { + "category_id": 13, + "poly": [ + 297, + 410, + 657, + 410, + 657, + 450, + 297, + 450 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\sum _ { t \\geq 1 } \\dot { w } _ { t } \\left[ \\bar { V } ^ { \\ast } ( s _ { t } ) - Q ^ { \\ast } ( s _ { t } , \\dot { a } _ { t } ) \\right] , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 580, + 701, + 625, + 701, + 625, + 733, + 580, + 733 + ], + "score": 0.87, + "latex": "1 / \\eta" + }, + { + "category_id": 13, + "poly": [ + 358, + 1066, + 384, + 1066, + 384, + 1098, + 358, + 1098 + ], + "score": 0.86, + "latex": "\\tilde { P }" + }, + { + "category_id": 13, + "poly": [ + 1288, + 1888, + 1312, + 1888, + 1312, + 1918, + 1288, + 1918 + ], + "score": 0.86, + "latex": "\\xi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 532, + 934, + 1164, + 934, + 1164, + 971, + 532, + 971 + ], + "score": 0.85, + "latex": "2 \\leq j \\leq M , \\sum _ { t = 1 } ^ { \\infty } I \\left[ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta _ { j - 1 } \\right] \\leq C" + }, + { + "category_id": 13, + "poly": [ + 296, + 1602, + 664, + 1602, + 664, + 1658, + 296, + 1658 + ], + "score": 0.85, + "latex": "\\mathrm { P r } \\left[ X _ { t } ^ { ( 2 ^ { i } ) } \\geq \\eta _ { j - 1 } \\right] > \\xi _ { i } / ( M \\eta _ { j } )" + }, + { + "category_id": 13, + "poly": [ + 562, + 349, + 1104, + 349, + 1104, + 384, + 562, + 384 + ], + "score": 0.85, + "latex": "J = \\{ t : \\ V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\eta ( 1 - \\gamma ) ^ { - 1 } \\}" + }, + { + "category_id": 14, + "poly": [ + 576, + 304, + 1113, + 304, + 1113, + 345, + 576, + 345 + ], + "score": 0.83, + "latex": "V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\leq \\hat { Q } _ { t } ( s _ { t } , a _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) ." + }, + { + "category_id": 13, + "poly": [ + 296, + 1476, + 323, + 1476, + 323, + 1506, + 296, + 1506 + ], + "score": 0.83, + "latex": "2 ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 815, + 1543, + 839, + 1543, + 839, + 1569, + 815, + 1569 + ], + "score": 0.83, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1302, + 1479, + 1327, + 1479, + 1327, + 1505, + 1302, + 1505 + ], + "score": 0.81, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1363, + 1890, + 1390, + 1890, + 1390, + 1921, + 1363, + 1921 + ], + "score": 0.81, + "latex": "\\eta _ { j }" + }, + { + "category_id": 13, + "poly": [ + 948, + 1172, + 1015, + 1172, + 1015, + 1203, + 948, + 1203 + ], + "score": 0.8, + "latex": "\\mathcal { F } _ { t + R }" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1546, + 1115, + 1546, + 1115, + 1573, + 1099, + 1573 + ], + "score": 0.8, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1572, + 1212, + 1572, + 1212, + 1600, + 1168, + 1600 + ], + "score": 0.78, + "latex": "R T" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1546, + 1042, + 1546, + 1042, + 1569, + 1028, + 1569 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1367, + 1979, + 1391, + 1979, + 1391, + 2004, + 1367, + 2004 + ], + "score": 0.71, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 677, + 1059, + 800, + 1059, + 800, + 1115, + 677, + 1115 + ], + "score": 0.69, + "latex": "\\begin{array} { r } { \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 763.0, + 654.0, + 763.0, + 654.0, + 799.0, + 292.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 839.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1120.0, + 341.0, + 1120.0, + 341.0, + 1186.0, + 287.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1120.0, + 1009.0, + 1120.0, + 1009.0, + 1186.0, + 597.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1120.0, + 1413.0, + 1120.0, + 1413.0, + 1186.0, + 1107.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1165.0, + 536.0, + 1165.0, + 536.0, + 1207.0, + 292.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1165.0, + 883.0, + 1165.0, + 883.0, + 1207.0, + 799.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1165.0, + 947.0, + 1165.0, + 947.0, + 1207.0, + 918.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1165.0, + 1265.0, + 1165.0, + 1265.0, + 1207.0, + 1016.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1165.0, + 1409.0, + 1165.0, + 1409.0, + 1207.0, + 1398.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1195.0, + 297.0, + 1195.0, + 297.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 1195.0, + 1408.0, + 1195.0, + 1408.0, + 1239.0, + 691.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1231.0, + 297.0, + 1231.0, + 297.0, + 1286.0, + 292.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 1231.0, + 884.0, + 1231.0, + 884.0, + 1286.0, + 428.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1280.0, + 643.0, + 1280.0, + 643.0, + 1316.0, + 293.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1228.5, + 1410.0, + 1228.5, + 1410.0, + 1285.0, + 1024.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 346.0, + 561.0, + 346.0, + 561.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 346.0, + 1312.0, + 346.0, + 1312.0, + 387.0, + 1105.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 346.0, + 1409.0, + 346.0, + 1409.0, + 387.0, + 1396.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 377.0, + 634.0, + 377.0, + 634.0, + 416.0, + 292.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 377.0, + 1408.0, + 377.0, + 1408.0, + 416.0, + 818.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 407.0, + 296.0, + 407.0, + 296.0, + 450.0, + 290.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 407.0, + 1232.0, + 407.0, + 1232.0, + 450.0, + 658.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 407.0, + 1408.0, + 407.0, + 1408.0, + 450.0, + 1398.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 450.0, + 1135.0, + 450.0, + 1135.0, + 487.0, + 294.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 450.0, + 1405.0, + 450.0, + 1405.0, + 487.0, + 1331.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 485.0, + 380.0, + 485.0, + 380.0, + 519.0, + 290.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 614.0, + 672.0, + 614.0, + 672.0, + 652.0, + 295.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 614.0, + 1218.0, + 614.0, + 1218.0, + 652.0, + 709.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 614.0, + 1404.0, + 614.0, + 1404.0, + 652.0, + 1256.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 648.0, + 520.0, + 648.0, + 520.0, + 700.0, + 291.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 648.0, + 1308.0, + 648.0, + 1308.0, + 700.0, + 645.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 648.0, + 1407.0, + 648.0, + 1407.0, + 700.0, + 1345.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 699.0, + 579.0, + 699.0, + 579.0, + 738.0, + 296.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 699.0, + 635.0, + 699.0, + 635.0, + 738.0, + 626.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1443.0, + 550.0, + 1443.0, + 550.0, + 1484.0, + 293.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1443.0, + 828.0, + 1443.0, + 828.0, + 1484.0, + 631.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1443.0, + 1408.0, + 1443.0, + 1408.0, + 1484.0, + 1136.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1471.0, + 295.0, + 1471.0, + 295.0, + 1516.0, + 291.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1471.0, + 562.0, + 1471.0, + 562.0, + 1516.0, + 324.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1471.0, + 1301.0, + 1471.0, + 1301.0, + 1516.0, + 897.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1471.0, + 1412.0, + 1471.0, + 1412.0, + 1516.0, + 1328.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1493.0, + 347.0, + 1493.0, + 347.0, + 1577.0, + 284.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1493.0, + 814.0, + 1493.0, + 814.0, + 1577.0, + 649.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1493.0, + 928.0, + 1493.0, + 928.0, + 1577.0, + 840.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1493.0, + 1417.0, + 1493.0, + 1417.0, + 1577.0, + 1181.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1568.0, + 297.0, + 1568.0, + 297.0, + 1607.0, + 293.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1568.0, + 809.0, + 1568.0, + 809.0, + 1607.0, + 426.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1568.0, + 1167.0, + 1568.0, + 1167.0, + 1607.0, + 970.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 1568.0, + 1407.0, + 1568.0, + 1407.0, + 1607.0, + 1213.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1598.0, + 295.0, + 1598.0, + 295.0, + 1659.0, + 289.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1598.0, + 1283.0, + 1598.0, + 1283.0, + 1659.0, + 665.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 766.0, + 262.0, + 766.0, + 304.0, + 294.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 262.0, + 922.0, + 262.0, + 922.0, + 304.0, + 873.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 262.0, + 1216.0, + 262.0, + 1216.0, + 304.0, + 1205.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1392.0, + 393.0, + 1392.0, + 393.0, + 1441.0, + 291.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 1392.0, + 884.0, + 1392.0, + 884.0, + 1441.0, + 612.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1392.0, + 1080.0, + 1392.0, + 1080.0, + 1441.0, + 1068.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 820.0, + 988.0, + 820.0, + 988.0, + 859.0, + 294.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1051.0, + 357.0, + 1051.0, + 357.0, + 1124.0, + 283.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1051.0, + 676.0, + 1051.0, + 676.0, + 1124.0, + 385.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1051.0, + 818.0, + 1051.0, + 818.0, + 1124.0, + 801.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 887.0, + 633.0, + 887.0, + 633.0, + 926.0, + 296.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 926.0, + 531.0, + 926.0, + 531.0, + 977.0, + 293.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 926.0, + 1251.0, + 926.0, + 1251.0, + 977.0, + 1165.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.25, + 966.5, + 1008.25, + 966.5, + 1008.25, + 1026.5, + 752.25, + 1026.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1330, + 1406, + 1330, + 1406, + 1539, + 296, + 1539 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 296, + 226, + 1405, + 226, + 1405, + 378, + 296, + 378 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 947, + 1406, + 947, + 1406, + 1135, + 296, + 1135 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1677, + 1405, + 1677, + 1405, + 1838, + 297, + 1838 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 296, + 787, + 1406, + 787, + 1406, + 936, + 296, + 936 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1852, + 1404, + 1852, + 1404, + 1954, + 298, + 1954 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 296, + 1142, + 1405, + 1142, + 1405, + 1260, + 296, + 1260 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 298, + 631, + 1405, + 631, + 1405, + 695, + 298, + 695 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 588, + 1968, + 1105, + 1968, + 1105, + 2047, + 588, + 2047 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 741, + 391, + 956, + 391, + 956, + 469, + 741, + 469 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 482, + 870, + 482, + 870, + 524, + 297, + 524 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 297, + 1283, + 1202, + 1283, + 1202, + 1319, + 297, + 1319 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 298, + 561, + 531, + 561, + 531, + 598, + 298, + 598 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 295, + 1630, + 1345, + 1630, + 1345, + 1665, + 295, + 1665 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 817, + 74, + 817, + 105, + 298, + 105 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 300, + 1575, + 755, + 1575, + 755, + 1606, + 300, + 1606 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 732, + 836, + 732, + 836, + 763, + 299, + 763 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 1374, + 486, + 1403, + 486, + 1403, + 515, + 1374, + 515 + ], + "score": 0.839 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.825 + }, + { + "category_id": 13, + "poly": [ + 325, + 980, + 482, + 980, + 482, + 1034, + 325, + 1034 + ], + "score": 0.95, + "latex": "\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma ) ^ { 8 } } \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 906, + 1393, + 1063, + 1393, + 1063, + 1448, + 906, + 1448 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\tilde { \\mathcal O } \\left( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 6 } } \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1093, + 1174, + 1227, + 1174, + 1227, + 1229, + 1093, + 1229 + ], + "score": 0.94, + "latex": "\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)" + }, + { + "category_id": 14, + "poly": [ + 588, + 1968, + 1109, + 1968, + 1109, + 2045, + 588, + 2045 + ], + "score": 0.94, + "latex": "\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right) ." + }, + { + "category_id": 13, + "poly": [ + 847, + 1852, + 936, + 1852, + 936, + 1890, + 847, + 1890 + ], + "score": 0.93, + "latex": "\\tilde { \\mathcal { O } } ( { \\sqrt { T } } )" + }, + { + "category_id": 13, + "poly": [ + 961, + 1030, + 1046, + 1030, + 1046, + 1075, + 961, + 1075 + ], + "score": 0.93, + "latex": "\\frac { 1 } { \\epsilon ^ { 2 } ( 1 - \\gamma ) }" + }, + { + "category_id": 13, + "poly": [ + 1036, + 1887, + 1262, + 1887, + 1262, + 1925, + 1036, + 1925 + ], + "score": 0.93, + "latex": "\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )" + }, + { + "category_id": 13, + "poly": [ + 1034, + 261, + 1240, + 261, + 1240, + 308, + 1034, + 308 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\epsilon _ { 1 } = \\frac { \\epsilon } { 2 4 R M \\ln \\frac { 1 } { 1 - \\gamma } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 457, + 821, + 685, + 821, + 685, + 874, + 457, + 874 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\Omega \\left( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln { 1 / \\delta } \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 809, + 1504, + 904, + 1504, + 904, + 1539, + 809, + 1539 + ], + "score": 0.93, + "latex": "\\Omega ( S ^ { 2 } A )" + }, + { + "category_id": 14, + "poly": [ + 741, + 391, + 955, + 391, + 955, + 468, + 741, + 468 + ], + "score": 0.93, + "latex": "\\tilde { \\mathcal { O } } \\left( \\frac { S A \\ln 1 / \\delta } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 860, + 870, + 968, + 870, + 968, + 905, + 860, + 905 + ], + "score": 0.93, + "latex": "1 / ( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 1084, + 1476, + 1170, + 1476, + 1170, + 1509, + 1084, + 1509 + ], + "score": 0.92, + "latex": "O ( S A )" + }, + { + "category_id": 13, + "poly": [ + 887, + 262, + 973, + 262, + 973, + 298, + 887, + 298 + ], + "score": 0.92, + "latex": "\\epsilon _ { 2 } = \\textstyle { \\frac { \\epsilon } { 3 } }" + }, + { + "category_id": 13, + "poly": [ + 419, + 303, + 741, + 303, + 741, + 345, + 419, + 345 + ], + "score": 0.92, + "latex": "\\ln 1 / \\epsilon _ { 1 } = \\mathrm { p o l y } ( \\ln \\textstyle \\frac { 1 } { \\epsilon } , \\ln \\frac { 1 } { 1 - \\gamma } )" + }, + { + "category_id": 13, + "poly": [ + 490, + 1445, + 576, + 1445, + 576, + 1479, + 490, + 1479 + ], + "score": 0.91, + "latex": "( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1709, + 392, + 1709, + 392, + 1749, + 298, + 1749 + ], + "score": 0.91, + "latex": "\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 739, + 1709, + 847, + 1709, + 847, + 1749, + 739, + 1749 + ], + "score": 0.9, + "latex": "\\bar { \\mathcal { O } } \\left( T ^ { 1 / 2 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 644, + 344, + 871, + 344, + 871, + 379, + 644, + 379 + ], + "score": 0.9, + "latex": "\\left( V ^ { * } - \\dot { V } ^ { \\pi } \\right) \\left( s _ { t } \\right) > \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 297, + 1224, + 464, + 1224, + 464, + 1261, + 297, + 1261 + ], + "score": 0.89, + "latex": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 1253, + 869, + 1403, + 869, + 1403, + 905, + 1253, + 905 + ], + "score": 0.88, + "latex": "\\mathrm { { } } ^ { \\mathrm { { 3 } } A , 1 / ( 1 - \\gamma ) }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1070, + 339, + 1070, + 339, + 1105, + 296, + 1105 + ], + "score": 0.87, + "latex": "1 / \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 345, + 903, + 386, + 903, + 386, + 935, + 345, + 935 + ], + "score": 0.87, + "latex": "1 / \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 806, + 226, + 831, + 226, + 831, + 258, + 806, + 258 + ], + "score": 0.87, + "latex": "\\tilde { P }" + }, + { + "category_id": 13, + "poly": [ + 889, + 1892, + 952, + 1892, + 952, + 1921, + 889, + 1921 + ], + "score": 0.86, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 733, + 265, + 760, + 265, + 760, + 292, + 733, + 292 + ], + "score": 0.85, + "latex": "\\epsilon _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 327, + 1926, + 354, + 1926, + 354, + 1953, + 327, + 1953 + ], + "score": 0.85, + "latex": "\\epsilon _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 813, + 267, + 840, + 267, + 840, + 292, + 813, + 292 + ], + "score": 0.85, + "latex": "\\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1862, + 1282, + 1862, + 1282, + 1887, + 1255, + 1887 + ], + "score": 0.82, + "latex": "\\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 580, + 481, + 860, + 481, + 860, + 527, + 580, + 527 + ], + "score": 0.79, + "latex": "\\begin{array} { r } { \\operatorname { p o l y } ( \\ln { \\frac { 1 } { \\epsilon } } , \\ln { \\frac { 1 } { 1 - \\gamma } } , \\ln S A ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1288, + 308, + 1352, + 308, + 1352, + 337, + 1288, + 337 + ], + "score": 0.79, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 984, + 1145, + 1006, + 1145, + 1006, + 1171, + 984, + 1171 + ], + "score": 0.77, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1066, + 1145, + 1089, + 1145, + 1089, + 1171, + 1066, + 1171 + ], + "score": 0.71, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 295, + 1183, + 360, + 1183, + 360, + 1214, + 295, + 1214 + ], + "score": 0.68, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 473, + 1185, + 496, + 1185, + 496, + 1215, + 473, + 1215 + ], + "score": 0.66, + "latex": "Q" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 558.0, + 535.0, + 558.0, + 535.0, + 605.0, + 291.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1574.0, + 757.0, + 1574.0, + 757.0, + 1609.0, + 295.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 838.0, + 731.0, + 838.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 490.0, + 1403.0, + 490.0, + 1403.0, + 516.0, + 1378.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1332.0, + 1404.0, + 1332.0, + 1404.0, + 1366.0, + 295.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1362.0, + 1406.0, + 1362.0, + 1406.0, + 1396.0, + 294.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1402.0, + 905.0, + 1402.0, + 905.0, + 1438.0, + 291.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1404.0, + 1404.0, + 1404.0, + 1404.0, + 1438.0, + 1064.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1444.0, + 489.0, + 1444.0, + 489.0, + 1482.0, + 294.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1444.0, + 1407.0, + 1444.0, + 1407.0, + 1482.0, + 577.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1476.0, + 1083.0, + 1476.0, + 1083.0, + 1509.0, + 294.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1476.0, + 1404.0, + 1476.0, + 1404.0, + 1509.0, + 1171.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1506.0, + 808.0, + 1506.0, + 808.0, + 1540.0, + 296.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1506.0, + 1331.0, + 1506.0, + 1331.0, + 1540.0, + 905.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 223.0, + 805.0, + 223.0, + 805.0, + 267.0, + 291.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 223.0, + 1406.0, + 223.0, + 1406.0, + 267.0, + 832.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 292.0, + 1241.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 299.0, + 418.0, + 299.0, + 418.0, + 353.0, + 290.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 299.0, + 1287.0, + 299.0, + 1287.0, + 353.0, + 742.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 299.0, + 1408.0, + 299.0, + 1408.0, + 353.0, + 1353.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 340.0, + 643.0, + 340.0, + 643.0, + 381.0, + 293.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 340.0, + 902.0, + 340.0, + 902.0, + 381.0, + 872.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 255.0, + 1208.0, + 255.0, + 1208.0, + 304.5, + 289.0, + 304.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.25, + 285.0, + 1229.25, + 285.0, + 1229.25, + 307.0, + 1191.25, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 945.0, + 1406.0, + 945.0, + 1406.0, + 986.0, + 291.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 978.0, + 324.0, + 978.0, + 324.0, + 1026.0, + 290.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 987.0, + 1408.0, + 987.0, + 1408.0, + 1026.0, + 484.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1026.0, + 960.0, + 1026.0, + 960.0, + 1086.0, + 286.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1026.0, + 1412.0, + 1026.0, + 1412.0, + 1086.0, + 1047.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1071.0, + 1406.0, + 1071.0, + 1406.0, + 1107.0, + 340.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1100.0, + 1025.0, + 1100.0, + 1025.0, + 1141.0, + 294.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1676.0, + 1405.0, + 1676.0, + 1405.0, + 1717.0, + 292.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1704.0, + 297.0, + 1704.0, + 297.0, + 1753.0, + 292.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1704.0, + 738.0, + 1704.0, + 738.0, + 1753.0, + 393.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1704.0, + 1408.0, + 1704.0, + 1408.0, + 1753.0, + 848.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1779.0, + 295.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1771.0, + 1408.0, + 1771.0, + 1408.0, + 1815.0, + 291.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 445.0, + 1804.0, + 445.0, + 1843.0, + 294.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 788.0, + 1404.0, + 788.0, + 1404.0, + 824.0, + 295.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 825.0, + 456.0, + 825.0, + 456.0, + 867.0, + 295.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 871.0, + 859.0, + 871.0, + 859.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 871.0, + 1252.0, + 871.0, + 1252.0, + 907.0, + 969.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 899.0, + 344.0, + 899.0, + 344.0, + 938.0, + 293.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 899.0, + 400.0, + 899.0, + 400.0, + 938.0, + 387.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.75, + 815.5, + 1411.75, + 815.5, + 1411.75, + 876.5, + 482.75, + 876.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1852.0, + 846.0, + 1852.0, + 846.0, + 1892.0, + 294.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1852.0, + 1254.0, + 1852.0, + 1254.0, + 1892.0, + 937.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1852.0, + 1404.0, + 1852.0, + 1404.0, + 1892.0, + 1283.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1890.0, + 888.0, + 1890.0, + 888.0, + 1926.0, + 294.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1890.0, + 1035.0, + 1890.0, + 1035.0, + 1926.0, + 953.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1890.0, + 1405.0, + 1890.0, + 1405.0, + 1926.0, + 1263.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1922.0, + 326.0, + 1922.0, + 326.0, + 1957.0, + 296.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1922.0, + 713.0, + 1922.0, + 713.0, + 1957.0, + 355.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1138.0, + 983.0, + 1138.0, + 983.0, + 1179.0, + 294.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1138.0, + 1065.0, + 1138.0, + 1065.0, + 1179.0, + 1007.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1138.0, + 1403.0, + 1138.0, + 1403.0, + 1179.0, + 1090.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1175.0, + 294.0, + 1175.0, + 294.0, + 1221.0, + 291.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1175.0, + 472.0, + 1175.0, + 472.0, + 1221.0, + 361.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1175.0, + 1092.0, + 1175.0, + 1092.0, + 1221.0, + 497.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1183.0, + 1403.0, + 1183.0, + 1403.0, + 1219.0, + 1233.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1223.0, + 296.0, + 1223.0, + 296.0, + 1259.0, + 292.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1223.0, + 473.0, + 1223.0, + 473.0, + 1259.0, + 465.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 1173.0, + 1203.0, + 1173.0, + 1203.0, + 1202.5, + 1140.0, + 1202.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 630.0, + 1406.0, + 630.0, + 1406.0, + 669.0, + 293.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 875.0, + 662.0, + 875.0, + 698.0, + 293.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 477.0, + 579.0, + 477.0, + 579.0, + 530.0, + 292.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 477.0, + 873.0, + 477.0, + 873.0, + 530.0, + 861.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1281.0, + 1204.0, + 1281.0, + 1204.0, + 1322.0, + 293.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1627.0, + 1348.0, + 1627.0, + 1348.0, + 1672.0, + 292.0, + 1672.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 405, + 1405, + 405, + 1405, + 612, + 298, + 612 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 723, + 1401, + 723, + 1401, + 818, + 299, + 818 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1402, + 229, + 1402, + 294, + 296, + 294 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 300, + 1744, + 1405, + 1744, + 1405, + 1839, + 300, + 1839 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 297, + 1221, + 1402, + 1221, + 1402, + 1315, + 297, + 1315 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 300, + 335, + 544, + 335, + 544, + 372, + 300, + 372 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 295, + 1333, + 1401, + 1333, + 1401, + 1397, + 295, + 1397 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 293, + 1663, + 1402, + 1663, + 1402, + 1728, + 293, + 1728 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 295, + 1416, + 1402, + 1416, + 1402, + 1481, + 295, + 1481 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 299, + 914, + 1401, + 914, + 1401, + 977, + 299, + 977 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 299, + 1077, + 1406, + 1077, + 1406, + 1201, + 299, + 1201 + ], + "score": 0.881 + }, + { + "category_id": 1, + "poly": [ + 297, + 1581, + 1404, + 1581, + 1404, + 1645, + 297, + 1645 + ], + "score": 0.878 + }, + { + "category_id": 1, + "poly": [ + 296, + 995, + 1402, + 995, + 1402, + 1059, + 296, + 1059 + ], + "score": 0.87 + }, + { + "category_id": 0, + "poly": [ + 300, + 656, + 674, + 656, + 674, + 691, + 300, + 691 + ], + "score": 0.868 + }, + { + "category_id": 1, + "poly": [ + 296, + 1498, + 1401, + 1498, + 1401, + 1564, + 296, + 1564 + ], + "score": 0.864 + }, + { + "category_id": 0, + "poly": [ + 299, + 862, + 489, + 862, + 489, + 897, + 299, + 897 + ], + "score": 0.824 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1397, + 1971, + 1397, + 2035, + 297, + 2035 + ], + "score": 0.811 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.777 + }, + { + "category_id": 1, + "poly": [ + 297, + 1858, + 1402, + 1858, + 1402, + 1953, + 297, + 1953 + ], + "score": 0.374 + }, + { + "category_id": 13, + "poly": [ + 1065, + 467, + 1206, + 467, + 1206, + 513, + 1065, + 513 + ], + "score": 0.94, + "latex": "\\tilde { \\cal O } \\big ( { \\textstyle \\frac { S A } { \\epsilon ^ { 4 } ( 1 - \\gamma _ { _ { - } } ) ^ { 8 } } } \\big )" + }, + { + "category_id": 13, + "poly": [ + 1059, + 538, + 1199, + 538, + 1199, + 582, + 1059, + 582 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\tilde { \\mathcal { O } } \\big ( \\frac { S A } { \\epsilon ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 774, + 578, + 889, + 578, + 889, + 613, + 774, + 613 + ], + "score": 0.93, + "latex": "1 / ( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 615, + 267, + 642, + 267, + 642, + 292, + 615, + 292 + ], + "score": 0.84, + "latex": "\\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1242, + 236, + 1268, + 236, + 1268, + 261, + 1242, + 261 + ], + "score": 0.84, + "latex": "\\epsilon _ { 1 }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 332.0, + 550.0, + 332.0, + 550.0, + 380.0, + 291.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 657.0, + 676.0, + 657.0, + 676.0, + 697.0, + 294.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 865.0, + 490.0, + 865.0, + 490.0, + 898.0, + 297.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 408.0, + 1405.0, + 408.0, + 1405.0, + 438.0, + 297.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 437.0, + 1064.0, + 437.0, + 1064.0, + 530.0, + 283.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 437.0, + 1406.0, + 437.0, + 1406.0, + 530.0, + 1207.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 506.0, + 1406.0, + 506.0, + 1406.0, + 543.0, + 295.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 523.0, + 1058.0, + 523.0, + 1058.0, + 596.0, + 285.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 523.0, + 1414.0, + 523.0, + 1414.0, + 596.0, + 1200.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 576.0, + 773.0, + 576.0, + 773.0, + 616.0, + 292.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 576.0, + 1165.0, + 576.0, + 1165.0, + 616.0, + 890.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 723.0, + 1405.0, + 723.0, + 1405.0, + 761.0, + 293.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 752.0, + 1406.0, + 752.0, + 1406.0, + 790.0, + 293.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 784.0, + 1181.0, + 784.0, + 1181.0, + 823.0, + 293.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 1241.0, + 227.0, + 1241.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 1269.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 259.0, + 614.0, + 259.0, + 614.0, + 299.0, + 295.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 259.0, + 653.0, + 259.0, + 653.0, + 299.0, + 643.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1782.0, + 296.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1776.0, + 1404.0, + 1776.0, + 1404.0, + 1813.0, + 322.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1809.0, + 596.0, + 1809.0, + 596.0, + 1838.0, + 323.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1221.0, + 1406.0, + 1221.0, + 1406.0, + 1258.0, + 296.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1249.0, + 1408.0, + 1249.0, + 1408.0, + 1291.0, + 321.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1284.0, + 533.0, + 1284.0, + 533.0, + 1314.0, + 324.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1373.0, + 293.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1365.0, + 661.0, + 1365.0, + 661.0, + 1397.0, + 322.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1659.0, + 1404.0, + 1659.0, + 1404.0, + 1701.0, + 295.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1693.0, + 992.0, + 1693.0, + 992.0, + 1728.0, + 320.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1416.0, + 1404.0, + 1416.0, + 1404.0, + 1452.0, + 297.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1447.0, + 1184.0, + 1447.0, + 1184.0, + 1482.0, + 322.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 909.0, + 1405.0, + 909.0, + 1405.0, + 953.0, + 293.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 944.0, + 1137.0, + 944.0, + 1137.0, + 979.0, + 323.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1076.0, + 1406.0, + 1076.0, + 1406.0, + 1113.0, + 294.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1109.0, + 1406.0, + 1109.0, + 1406.0, + 1142.0, + 319.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1140.0, + 1406.0, + 1140.0, + 1406.0, + 1173.0, + 322.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1171.0, + 640.0, + 1171.0, + 640.0, + 1203.0, + 322.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1580.0, + 1408.0, + 1580.0, + 1408.0, + 1618.0, + 295.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1612.0, + 843.0, + 1612.0, + 843.0, + 1644.0, + 326.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1407.0, + 991.0, + 1407.0, + 1033.0, + 294.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1027.0, + 947.0, + 1027.0, + 947.0, + 1062.0, + 323.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1499.0, + 1405.0, + 1499.0, + 1405.0, + 1535.0, + 296.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1530.0, + 1196.0, + 1530.0, + 1196.0, + 1562.0, + 324.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1968.0, + 1403.0, + 1968.0, + 1403.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 1377.0, + 2002.0, + 1377.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1857.0, + 1403.0, + 1857.0, + 1403.0, + 1896.0, + 296.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1890.0, + 1403.0, + 1890.0, + 1403.0, + 1924.0, + 324.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1919.0, + 1236.0, + 1919.0, + 1236.0, + 1955.0, + 321.0, + 1955.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 390, + 1268, + 1310, + 1268, + 1310, + 1643, + 390, + 1643 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 360, + 1777, + 1342, + 1777, + 1342, + 2035, + 360, + 2035 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 520, + 996, + 1177, + 996, + 1177, + 1082, + 520, + 1082 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 298, + 683, + 1405, + 683, + 1405, + 780, + 298, + 780 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 292, + 899, + 1404, + 899, + 1404, + 977, + 292, + 977 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 298, + 342, + 1404, + 342, + 1404, + 467, + 298, + 467 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 298, + 1098, + 690, + 1098, + 690, + 1133, + 298, + 1133 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 301, + 228, + 1401, + 228, + 1401, + 324, + 301, + 324 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 297, + 1714, + 1355, + 1714, + 1355, + 1762, + 297, + 1762 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 298, + 568, + 1406, + 568, + 1406, + 665, + 298, + 665 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 817, + 75, + 817, + 105, + 299, + 105 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 296, + 1210, + 1196, + 1210, + 1196, + 1252, + 296, + 1252 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 296, + 1163, + 1198, + 1163, + 1198, + 1200, + 296, + 1200 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 294, + 486, + 1399, + 486, + 1399, + 551, + 294, + 551 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 865, + 2087, + 865, + 2114, + 835, + 2114 + ], + "score": 0.854 + }, + { + "category_id": 9, + "poly": [ + 1364, + 1022, + 1401, + 1022, + 1401, + 1053, + 1364, + 1053 + ], + "score": 0.839 + }, + { + "category_id": 0, + "poly": [ + 301, + 835, + 646, + 835, + 646, + 872, + 301, + 872 + ], + "score": 0.562 + }, + { + "category_id": 1, + "poly": [ + 301, + 835, + 646, + 835, + 646, + 872, + 301, + 872 + ], + "score": 0.395 + }, + { + "category_id": 14, + "poly": [ + 386, + 1272, + 1313, + 1272, + 1313, + 1645, + 386, + 1645 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } \\leq \\displaystyle \\sum _ { t \\in I } \\left( V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) \\right) \\leq \\displaystyle \\sum _ { t \\in I } \\left[ \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { 1 } { ( 1 - \\gamma ) ^ { 5 / 2 } } \\sqrt { S A | I | \\ell | ( I | ) } + \\frac { S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln | I | \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln \\frac { S A | I | } { \\delta } } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\\\ & { \\qquad \\leq \\displaystyle \\frac { | I | \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\frac { \\sqrt { S A | I | \\ln S A | I | } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\frac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) \\right) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 356, + 1775, + 1344, + 1775, + 1344, + 2038, + 356, + 2038 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { \\frac { \\eta | I | } { 1 - \\gamma } = \\cfrac { k ^ { 2 } S A \\ln S A } { ( 1 - \\gamma ) ^ { 4 } \\eta } \\leq 2 \\cfrac { ( \\eta - \\epsilon _ { 1 } ) | I | } { 1 - \\gamma } } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { \\sqrt { S A | I | \\ln \\left( S A | I | \\right) } } { ( 1 - \\gamma ) ^ { 5 / 2 } } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) } \\\\ & { \\qquad \\leq C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln \\cfrac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\cfrac { S A k } { \\eta ( 1 - \\gamma ) ^ { 4 } } \\sqrt { \\ln S A \\cdot ( \\ln S A + \\ln | I | ) } + \\cfrac { S A \\ln | I | } { ( 1 - \\gamma ) ^ { 3 } } \\right) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 447, + 1714, + 687, + 1714, + 687, + 1763, + 447, + 1763 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { | I | = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 957, + 905, + 1272, + 905, + 1272, + 946, + 957, + 946 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 519, + 990, + 1176, + 990, + 1176, + 1086, + 519, + 1086 + ], + "score": 0.92, + "latex": "\\sum _ { t = 1 } ^ { t = \\infty } I \\left[ B _ { \\eta } ^ { ( t ) } \\right] \\leq \\frac { S A \\ln S A \\ln 1 / \\delta } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\cdot p o l y l o g \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 341, + 1211, + 777, + 1211, + 777, + 1254, + 341, + 1254 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { I = \\{ t \\colon V ^ { * } ( s _ { t } ) - Q ^ { * } ( s _ { t } , a _ { t } ) > \\frac { \\eta } { 1 - \\gamma } \\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 599, + 907, + 667, + 907, + 667, + 939, + 599, + 939 + ], + "score": 0.92, + "latex": "\\eta > 0" + }, + { + "category_id": 13, + "poly": [ + 712, + 899, + 764, + 899, + 764, + 940, + 712, + 940 + ], + "score": 0.92, + "latex": "B _ { \\eta } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 453, + 1168, + 522, + 1168, + 522, + 1199, + 453, + 1199 + ], + "score": 0.91, + "latex": "\\eta > 1" + }, + { + "category_id": 13, + "poly": [ + 1121, + 1168, + 1189, + 1168, + 1189, + 1199, + 1121, + 1199 + ], + "score": 0.91, + "latex": "\\eta \\leq 1" + }, + { + "category_id": 13, + "poly": [ + 726, + 943, + 817, + 943, + 817, + 977, + 726, + 977 + ], + "score": 0.9, + "latex": "1 - \\delta / 2" + }, + { + "category_id": 13, + "poly": [ + 298, + 944, + 388, + 944, + 388, + 976, + 298, + 976 + ], + "score": 0.9, + "latex": "\\eta > 2 \\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 802, + 1722, + 871, + 1722, + 871, + 1750, + 802, + 1750 + ], + "score": 0.9, + "latex": "k > 1" + }, + { + "category_id": 13, + "poly": [ + 1312, + 1723, + 1345, + 1723, + 1345, + 1753, + 1312, + 1753 + ], + "score": 0.88, + "latex": "C _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1099, + 411, + 1099, + 411, + 1133, + 371, + 1133 + ], + "score": 0.87, + "latex": "I [ \\cdot ]" + }, + { + "category_id": 13, + "poly": [ + 1124, + 1214, + 1187, + 1214, + 1187, + 1243, + 1124, + 1243 + ], + "score": 0.84, + "latex": "1 - \\delta" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 833.0, + 649.0, + 833.0, + 649.0, + 877.0, + 295.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 685.0, + 1402.0, + 685.0, + 1402.0, + 719.0, + 295.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 712.0, + 1405.0, + 712.0, + 1405.0, + 754.0, + 321.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 746.0, + 706.0, + 746.0, + 706.0, + 781.0, + 322.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 892.0, + 598.0, + 892.0, + 598.0, + 951.0, + 290.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 892.0, + 711.0, + 892.0, + 711.0, + 951.0, + 668.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 892.0, + 956.0, + 892.0, + 956.0, + 951.0, + 765.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 892.0, + 1411.0, + 892.0, + 1411.0, + 951.0, + 1273.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 941.0, + 297.0, + 941.0, + 297.0, + 981.0, + 292.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 941.0, + 725.0, + 941.0, + 725.0, + 981.0, + 389.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 941.0, + 826.0, + 941.0, + 826.0, + 981.0, + 818.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 345.0, + 1405.0, + 345.0, + 1405.0, + 378.0, + 296.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 374.0, + 1404.0, + 374.0, + 1404.0, + 407.0, + 322.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 403.0, + 1405.0, + 403.0, + 1405.0, + 443.0, + 317.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 436.0, + 563.0, + 436.0, + 563.0, + 469.0, + 321.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1098.0, + 370.0, + 1098.0, + 370.0, + 1134.0, + 296.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1098.0, + 693.0, + 1098.0, + 693.0, + 1134.0, + 412.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 1406.0, + 227.0, + 1406.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 264.0, + 1404.0, + 264.0, + 1404.0, + 294.0, + 325.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 288.0, + 1039.0, + 288.0, + 1039.0, + 329.0, + 320.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 446.0, + 1712.0, + 446.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1712.0, + 801.0, + 1712.0, + 801.0, + 1764.0, + 688.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1712.0, + 1311.0, + 1712.0, + 1311.0, + 1764.0, + 872.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1712.0, + 1357.0, + 1712.0, + 1357.0, + 1764.0, + 1346.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 571.0, + 1408.0, + 571.0, + 1408.0, + 605.0, + 296.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 603.0, + 1404.0, + 603.0, + 1404.0, + 637.0, + 324.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 631.0, + 730.0, + 631.0, + 730.0, + 666.0, + 323.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1205.0, + 340.0, + 1205.0, + 340.0, + 1258.0, + 292.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1205.0, + 1123.0, + 1205.0, + 1123.0, + 1258.0, + 778.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1205.0, + 1202.0, + 1205.0, + 1202.0, + 1258.0, + 1188.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1163.0, + 452.0, + 1163.0, + 452.0, + 1203.0, + 294.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1163.0, + 1120.0, + 1163.0, + 1120.0, + 1203.0, + 523.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1163.0, + 1201.0, + 1163.0, + 1201.0, + 1203.0, + 1190.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 489.0, + 1403.0, + 489.0, + 1403.0, + 521.0, + 298.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 519.0, + 1392.0, + 519.0, + 1392.0, + 552.0, + 322.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 833.0, + 649.0, + 833.0, + 649.0, + 877.0, + 295.0, + 877.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 417, + 272, + 1279, + 272, + 1279, + 603, + 417, + 603 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 581, + 788, + 1114, + 788, + 1114, + 871, + 581, + 871 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 356, + 1483, + 1341, + 1483, + 1341, + 1579, + 356, + 1579 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 534, + 934, + 1159, + 934, + 1159, + 1004, + 534, + 1004 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 696, + 1818, + 1002, + 1818, + 1002, + 1864, + 696, + 1864 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 699, + 1713, + 998, + 1713, + 998, + 1758, + 699, + 1758 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 744, + 688, + 951, + 688, + 951, + 728, + 744, + 728 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 297, + 1017, + 603, + 1017, + 603, + 1049, + 297, + 1049 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 294, + 619, + 843, + 619, + 843, + 672, + 294, + 672 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 297, + 884, + 906, + 884, + 906, + 919, + 297, + 919 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 297, + 740, + 561, + 740, + 561, + 774, + 297, + 774 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 1769, + 770, + 1769, + 770, + 1802, + 297, + 1802 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 296, + 1890, + 1167, + 1890, + 1167, + 1927, + 296, + 1927 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 298, + 1612, + 804, + 1612, + 804, + 1657, + 298, + 1657 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 817, + 73, + 817, + 106, + 297, + 106 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 287, + 1393, + 1347, + 1393, + 1347, + 1432, + 287, + 1432 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 411, + 229, + 411, + 262, + 296, + 262 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 300, + 1325, + 648, + 1325, + 648, + 1362, + 300, + 1362 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 296, + 1663, + 1059, + 1663, + 1059, + 1697, + 296, + 1697 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 293, + 1959, + 992, + 1959, + 992, + 1995, + 293, + 1995 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1351, + 950, + 1401, + 950, + 1401, + 982, + 1351, + 982 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1165, + 1401, + 1165, + 1401, + 1197, + 1351, + 1197 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1087, + 1401, + 1087, + 1401, + 1119, + 1351, + 1119 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 862, + 2087, + 862, + 2114, + 835, + 2114 + ], + "score": 0.849 + }, + { + "category_id": 9, + "poly": [ + 1364, + 693, + 1402, + 693, + 1402, + 724, + 1364, + 724 + ], + "score": 0.847 + }, + { + "category_id": 2, + "poly": [ + 1373, + 1892, + 1404, + 1892, + 1404, + 1923, + 1373, + 1923 + ], + "score": 0.845 + }, + { + "category_id": 1, + "poly": [ + 298, + 2001, + 849, + 2001, + 849, + 2036, + 298, + 2036 + ], + "score": 0.843 + }, + { + "category_id": 8, + "poly": [ + 585, + 1066, + 863, + 1066, + 863, + 1140, + 585, + 1140 + ], + "score": 0.836 + }, + { + "category_id": 8, + "poly": [ + 624, + 1144, + 1111, + 1144, + 1111, + 1221, + 624, + 1221 + ], + "score": 0.82 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1249, + 1404, + 1249, + 1404, + 1279, + 1374, + 1279 + ], + "score": 0.817 + }, + { + "category_id": 8, + "poly": [ + 585, + 1065, + 1113, + 1065, + 1113, + 1221, + 585, + 1221 + ], + "score": 0.118 + }, + { + "category_id": 14, + "poly": [ + 418, + 271, + 1281, + 271, + 1281, + 604, + 418, + 604 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { k ^ { 2 } \\ln ( S A ) \\le C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( k \\left( \\ln S A + \\ln | I | \\right) + \\eta ( 1 - \\gamma ) \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( \\ln S A + 2 \\ln | I | \\right) } \\\\ & { \\qquad \\le k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\cdot \\left( 3 \\ln S A + 4 \\ln k + 6 \\ln \\frac { 1 } { \\eta ( 1 - \\gamma ) } \\right) } \\\\ & { \\qquad \\le 6 k C _ { 1 } \\sqrt { \\ln \\displaystyle \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\left( \\ln S A + \\ln \\epsilon k \\right) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 342, + 618, + 765, + 618, + 765, + 675, + 342, + 675 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { C ^ { \\prime } = \\operatorname* { m a x } \\lbrace 2 , 6 C _ { 1 } \\sqrt { \\ln \\frac { 1 } { \\delta } } \\ln ^ { 2 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } \\rbrace } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 584, + 1063, + 1115, + 1063, + 1115, + 1223, + 584, + 1223 + ], + "score": 0.93, + "latex": "\\begin{array} { r l r } { { \\vert I \\vert = \\frac { S A k ^ { 2 } } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 3 } } \\ln S A } } \\\\ & { \\leq \\frac { S A \\ln S A } { \\eta ^ { 2 } ( 1 - \\gamma ) ^ { 5 } } \\cdot \\ln \\frac { 1 } { \\delta } \\cdot \\mathcal { O } ( \\ln ^ { 8 } \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } ) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 356, + 1484, + 1340, + 1484, + 1340, + 1579, + 356, + 1579 + ], + "score": 0.93, + "latex": "\\sum _ { t \\ge 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\le \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A \\ln C } { ( 1 - \\gamma ) ^ { 3 } } \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 586, + 1769, + 662, + 1769, + 662, + 1802, + 586, + 1802 + ], + "score": 0.92, + "latex": "p ^ { \\prime } \\leq p" + }, + { + "category_id": 13, + "poly": [ + 372, + 1613, + 636, + 1613, + 636, + 1658, + 372, + 1658 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\ell ( C ) = \\iota ( C ) \\ln { \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 324, + 742, + 496, + 742, + 496, + 773, + 324, + 773 + ], + "score": 0.91, + "latex": "k \\geq 1 0 C ^ { \\prime } \\ln C ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 543, + 1396, + 619, + 1396, + 619, + 1430, + 543, + 1430 + ], + "score": 0.91, + "latex": "( C , w )" + }, + { + "category_id": 13, + "poly": [ + 735, + 1396, + 825, + 1396, + 825, + 1431, + 735, + 1431 + ], + "score": 0.91, + "latex": "( w _ { t } ) _ { t \\geq 1 }" + }, + { + "category_id": 14, + "poly": [ + 696, + 1816, + 1002, + 1816, + 1002, + 1861, + 696, + 1861 + ], + "score": 0.91, + "latex": "{ \\hat { Q } } _ { p + 1 } ( s , a ) \\geq Q _ { p ^ { \\prime } + 1 } ( s , a ) ." + }, + { + "category_id": 14, + "poly": [ + 699, + 1711, + 998, + 1711, + 998, + 1756, + 699, + 1756 + ], + "score": 0.91, + "latex": "\\hat { Q } _ { p + 1 } ( s , a ) \\leq Q _ { p + 1 } ( s , a ) ." + }, + { + "category_id": 13, + "poly": [ + 818, + 885, + 903, + 885, + 903, + 917, + 818, + 917 + ], + "score": 0.91, + "latex": "C ^ { \\prime } \\geq 2" + }, + { + "category_id": 14, + "poly": [ + 537, + 930, + 1161, + 930, + 1161, + 1004, + 537, + 1004 + ], + "score": 0.9, + "latex": "k \\leq 1 0 C ^ { \\prime } \\ln C ^ { \\prime } \\leq 3 6 0 C _ { 1 } ^ { 2 } \\operatorname* { m a x } \\{ \\ln ^ { 4 } { \\frac { 1 } { \\epsilon _ { 1 } ( 1 - \\gamma ) } } , 2 0 \\ln 2 \\} ." + }, + { + "category_id": 14, + "poly": [ + 582, + 786, + 1114, + 786, + 1114, + 875, + 582, + 875 + ], + "score": 0.9, + "latex": "\\begin{array} { c } { k - C ^ { \\prime } \\left( 2 + \\ln k \\right) \\geq 8 C ^ { \\prime } \\ln C ^ { \\prime } - ( 2 + \\ln 1 0 ) C ^ { \\prime } } \\\\ { \\geq 4 C ^ { \\prime } \\left( 2 \\ln C ^ { \\prime } - 4 \\right) \\geq 0 , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 745, + 686, + 954, + 686, + 954, + 728, + 745, + 728 + ], + "score": 0.89, + "latex": "k \\leq C ^ { \\prime } ( 2 + \\ln k ) ." + }, + { + "category_id": 13, + "poly": [ + 1020, + 1396, + 1111, + 1396, + 1111, + 1430, + 1020, + 1430 + ], + "score": 0.89, + "latex": "1 - \\delta / 2" + }, + { + "category_id": 13, + "poly": [ + 805, + 2002, + 835, + 2002, + 835, + 2036, + 805, + 2036 + ], + "score": 0.86, + "latex": "\\alpha _ { t } ^ { i }" + }, + { + "category_id": 13, + "poly": [ + 427, + 1776, + 444, + 1776, + 444, + 1802, + 427, + 1802 + ], + "score": 0.54, + "latex": "p" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1323.0, + 649.0, + 1323.0, + 649.0, + 1367.0, + 293.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 1898.0, + 1402.0, + 1898.0, + 1402.0, + 1921.0, + 1379.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 1254.0, + 1403.0, + 1254.0, + 1403.0, + 1282.0, + 1378.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1012.0, + 605.0, + 1012.0, + 605.0, + 1054.0, + 294.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 612.0, + 341.0, + 612.0, + 341.0, + 686.0, + 280.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 612.0, + 854.0, + 612.0, + 854.0, + 686.0, + 766.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 883.0, + 817.0, + 883.0, + 817.0, + 923.0, + 294.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 883.0, + 908.0, + 883.0, + 908.0, + 923.0, + 904.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 738.0, + 323.0, + 738.0, + 323.0, + 778.0, + 294.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 738.0, + 563.0, + 738.0, + 563.0, + 778.0, + 497.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1766.0, + 426.0, + 1766.0, + 426.0, + 1805.0, + 292.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1766.0, + 585.0, + 1766.0, + 585.0, + 1805.0, + 445.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1766.0, + 774.0, + 1766.0, + 774.0, + 1805.0, + 663.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1889.0, + 1169.0, + 1889.0, + 1169.0, + 1930.0, + 295.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1603.0, + 371.0, + 1603.0, + 371.0, + 1672.0, + 284.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1603.0, + 819.0, + 1603.0, + 819.0, + 1672.0, + 637.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1393.0, + 542.0, + 1393.0, + 542.0, + 1434.0, + 294.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1393.0, + 734.0, + 1393.0, + 734.0, + 1434.0, + 620.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1393.0, + 1019.0, + 1393.0, + 1019.0, + 1434.0, + 826.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1393.0, + 1350.0, + 1393.0, + 1350.0, + 1434.0, + 1112.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 412.0, + 229.0, + 412.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1659.0, + 1061.0, + 1659.0, + 1061.0, + 1704.0, + 293.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1956.0, + 994.0, + 1956.0, + 994.0, + 2000.0, + 293.0, + 2000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1998.0, + 804.0, + 1998.0, + 804.0, + 2042.0, + 295.0, + 2042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1998.0, + 855.0, + 1998.0, + 855.0, + 2042.0, + 836.0, + 2042.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 587, + 1518, + 1108, + 1518, + 1108, + 1609, + 587, + 1609 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 497, + 546, + 1200, + 546, + 1200, + 638, + 497, + 638 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 610, + 723, + 1087, + 723, + 1087, + 813, + 610, + 813 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 297, + 863, + 1404, + 863, + 1404, + 927, + 297, + 927 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 433, + 1086, + 1265, + 1086, + 1265, + 1166, + 433, + 1166 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 344, + 939, + 1354, + 939, + 1354, + 1032, + 344, + 1032 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 473, + 1270, + 1224, + 1270, + 1224, + 1356, + 473, + 1356 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 390, + 1808, + 1307, + 1808, + 1307, + 1894, + 390, + 1894 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 297, + 1759, + 690, + 1759, + 690, + 1794, + 297, + 1794 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 294, + 1615, + 1270, + 1615, + 1270, + 1649, + 294, + 1649 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 295, + 1043, + 1226, + 1043, + 1226, + 1078, + 295, + 1078 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 299, + 1903, + 697, + 1903, + 697, + 1938, + 299, + 1938 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 294, + 1219, + 1126, + 1219, + 1126, + 1257, + 294, + 1257 + ], + "score": 0.927 + }, + { + "category_id": 8, + "poly": [ + 298, + 1952, + 1536, + 1952, + 1536, + 2039, + 298, + 2039 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 297, + 648, + 1167, + 648, + 1167, + 683, + 297, + 683 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 296, + 817, + 850, + 817, + 850, + 851, + 296, + 851 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 296, + 500, + 505, + 500, + 505, + 534, + 296, + 534 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 817, + 73, + 817, + 106, + 297, + 106 + ], + "score": 0.913 + }, + { + "category_id": 1, + "poly": [ + 297, + 1485, + 504, + 1485, + 504, + 1518, + 297, + 1518 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1295, + 1401, + 1295, + 1401, + 1328, + 1351, + 1328 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1369, + 1402, + 1369, + 1402, + 1401, + 1351, + 1401 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 297, + 696, + 490, + 696, + 490, + 728, + 297, + 728 + ], + "score": 0.872 + }, + { + "category_id": 1, + "poly": [ + 299, + 1417, + 1073, + 1417, + 1073, + 1458, + 299, + 1458 + ], + "score": 0.866 + }, + { + "category_id": 8, + "poly": [ + 506, + 1664, + 1189, + 1664, + 1189, + 1750, + 506, + 1750 + ], + "score": 0.842 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1174, + 1403, + 1174, + 1403, + 1202, + 1374, + 1202 + ], + "score": 0.813 + }, + { + "category_id": 1, + "poly": [ + 360, + 296, + 1041, + 296, + 1041, + 343, + 360, + 343 + ], + "score": 0.796 + }, + { + "category_id": 1, + "poly": [ + 360, + 357, + 819, + 357, + 819, + 397, + 360, + 397 + ], + "score": 0.739 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2114, + 835, + 2114 + ], + "score": 0.72 + }, + { + "category_id": 1, + "poly": [ + 361, + 222, + 982, + 222, + 982, + 282, + 361, + 282 + ], + "score": 0.669 + }, + { + "category_id": 8, + "poly": [ + 361, + 412, + 1400, + 412, + 1400, + 473, + 361, + 473 + ], + "score": 0.49 + }, + { + "category_id": 1, + "poly": [ + 361, + 412, + 1400, + 412, + 1400, + 473, + 361, + 473 + ], + "score": 0.446 + }, + { + "category_id": 8, + "poly": [ + 476, + 1363, + 729, + 1363, + 729, + 1406, + 476, + 1406 + ], + "score": 0.446 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 865, + 2087, + 865, + 2114, + 835, + 2114 + ], + "score": 0.313 + }, + { + "category_id": 8, + "poly": [ + 506, + 1664, + 1190, + 1664, + 1190, + 1750, + 506, + 1750 + ], + "score": 0.197 + }, + { + "category_id": 14, + "poly": [ + 609, + 721, + 1085, + 721, + 1085, + 815, + 609, + 815 + ], + "score": 0.95, + "latex": "\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( t ) } { i } } \\leq 2 \\sqrt { \\frac { \\iota ( t ) } { t } } ," + }, + { + "category_id": 14, + "poly": [ + 508, + 1657, + 1192, + 1657, + 1192, + 1752, + 508, + 1752 + ], + "score": 0.94, + "latex": "Q _ { p } ( s , a ) = \\alpha _ { t } ^ { 0 } \\frac { 1 } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + b _ { i } + \\gamma \\hat { V } _ { t _ { i } } ( s _ { t _ { i } + 1 } ) \\right] ." + }, + { + "category_id": 13, + "poly": [ + 466, + 1043, + 618, + 1043, + 618, + 1078, + 466, + 1078 + ], + "score": 0.94, + "latex": "f ( t ) = \\iota ( t ) / t" + }, + { + "category_id": 14, + "poly": [ + 499, + 542, + 1200, + 542, + 1200, + 641, + 499, + 641 + ], + "score": 0.94, + "latex": "\\alpha _ { t } = \\frac { H + 1 } { H + t } , \\quad \\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) ." + }, + { + "category_id": 14, + "poly": [ + 433, + 1086, + 1264, + 1086, + 1264, + 1165, + 433, + 1165 + ], + "score": 0.93, + "latex": "\\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t ) } { t } } \\geq \\sqrt { \\frac { \\iota ( t ) } { t } } ." + }, + { + "category_id": 14, + "poly": [ + 591, + 1512, + 1108, + 1512, + 1108, + 1611, + 591, + 1611 + ], + "score": 0.93, + "latex": "\\alpha _ { t } ^ { 0 } = \\prod _ { j = 1 } ^ { t } ( 1 - \\alpha _ { j } ) , \\quad \\alpha _ { t } ^ { i } = \\alpha _ { i } \\prod _ { j = i + 1 } ^ { t } ( 1 - \\alpha _ { j } ) ." + }, + { + "category_id": 14, + "poly": [ + 341, + 938, + 1353, + 938, + 1353, + 1034, + 341, + 1034 + ], + "score": 0.93, + "latex": "\\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } = \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sum _ { i = 1 } ^ { t - 1 } \\alpha _ { t - 1 } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\geq \\alpha _ { t } \\sqrt { \\frac { \\iota ( t ) } { t } } + ( 1 - \\alpha _ { t } ) \\sqrt { \\frac { \\iota ( t - 1 ) } { t - 1 } } ." + }, + { + "category_id": 14, + "poly": [ + 391, + 1802, + 1308, + 1802, + 1308, + 1896, + 391, + 1896 + ], + "score": 0.93, + "latex": "Q ^ { * } ( s , a ) = r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) = \\alpha _ { t } ^ { 0 } Q ^ { * } ( s , a ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ r ( s , a ) + \\gamma \\mathbb { P } V ^ { * } ( s , a ) \\right] ." + }, + { + "category_id": 13, + "poly": [ + 357, + 893, + 575, + 893, + 575, + 931, + 357, + 931 + ], + "score": 0.93, + "latex": "\\alpha _ { t } ^ { i } = ( 1 - \\alpha _ { t } ) \\dot { \\alpha _ { t - 1 } ^ { i } }" + }, + { + "category_id": 13, + "poly": [ + 979, + 1046, + 1117, + 1046, + 1117, + 1077, + 979, + 1077 + ], + "score": 0.92, + "latex": "t \\geq 1 , c \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 882, + 1223, + 950, + 1223, + 950, + 1255, + 882, + 1255 + ], + "score": 0.92, + "latex": "p \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1615, + 1261, + 1615, + 1261, + 1649, + 1176, + 1649 + ], + "score": 0.92, + "latex": "Q ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1003, + 1223, + 1065, + 1223, + 1065, + 1256, + 1003, + 1256 + ], + "score": 0.91, + "latex": "( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1080, + 864, + 1234, + 864, + 1234, + 898, + 1080, + 898 + ], + "score": 0.91, + "latex": "t = 1 , \\alpha _ { t } ^ { t } = 1" + }, + { + "category_id": 14, + "poly": [ + 473, + 1265, + 1222, + 1265, + 1222, + 1407, + 473, + 1407 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { 0 \\le ( Q _ { p } - Q ^ { * } ) ( s , a ) \\le \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } _ { t _ { i } } - V ^ { * } ) ( s _ { t _ { i } + 1 } ) + \\beta _ { t } , } \\\\ & { 0 \\le ( { \\hat { Q } } _ { p } - Q ^ { * } ) ( s , a ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 704, + 1222, + 796, + 1222, + 796, + 1256, + 704, + 1256 + ], + "score": 0.91, + "latex": "1 - \\delta / 2" + }, + { + "category_id": 13, + "poly": [ + 619, + 895, + 777, + 895, + 777, + 925, + 619, + 925 + ], + "score": 0.9, + "latex": "1 \\leq i \\leq t - 1" + }, + { + "category_id": 13, + "poly": [ + 396, + 359, + 634, + 359, + 634, + 397, + 396, + 397 + ], + "score": 0.9, + "latex": "\\textstyle \\sum _ { t = i } ^ { \\infty } \\alpha _ { t } ^ { i } = 1 + 1 / H" + }, + { + "category_id": 14, + "poly": [ + 310, + 1946, + 1428, + 1946, + 1428, + 2040, + 310, + 2040 + ], + "score": 0.9, + "latex": "Q _ { p } - Q ^ { * } ) ( s , a ) = \\alpha _ { t } ^ { 0 } ( \\frac { 1 } { 1 - \\gamma } - Q ^ { * } ( s , a ) ) + \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ b _ { i } + \\gamma \\left( V _ { t _ { i } } - V ^ { * } \\right) \\left( s _ { t _ { i } + 1 } \\right) + \\gamma \\left( V ^ { * } \\left( s _ { t _ { i } + 1 } \\right) - \\mathbb { P } V ^ { * } \\right) \\right] ," + }, + { + "category_id": 13, + "poly": [ + 737, + 1416, + 1064, + 1416, + 1064, + 1457, + 737, + 1457 + ], + "score": 0.89, + "latex": "\\beta _ { t } = c _ { 3 } \\sqrt { H \\iota ( t ) / ( ( 1 - \\gamma ) ^ { 2 } t ) }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1420, + 684, + 1420, + 684, + 1456, + 371, + 1456 + ], + "score": 0.89, + "latex": "t = N _ { p } ( s , a ) , t _ { i } = \\tau ( s , a , i )" + }, + { + "category_id": 13, + "poly": [ + 397, + 223, + 722, + 223, + 722, + 282, + 397, + 282 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\sqrt { \\frac { 1 } { t } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { 1 } { i } } \\leq 2 \\sqrt { \\frac { 1 } { t } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1291, + 865, + 1355, + 865, + 1355, + 895, + 1291, + 895 + ], + "score": 0.88, + "latex": "t \\geq 2" + }, + { + "category_id": 13, + "poly": [ + 1254, + 431, + 1393, + 431, + 1393, + 463, + 1254, + 463 + ], + "score": 0.88, + "latex": "t \\geq 1 , c \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 656, + 298, + 861, + 298, + 861, + 341, + 656, + 341 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\sum _ { i = 1 } ^ { t } ( \\alpha _ { t } ^ { i } ) ^ { 2 } \\le \\frac { 2 H } { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 396, + 300, + 602, + 300, + 602, + 341, + 396, + 341 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { \\operatorname* { m a x } _ { i \\in [ t ] } \\alpha _ { t } ^ { i } \\leq \\frac { 2 H } { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 749, + 362, + 812, + 362, + 812, + 394, + 749, + 394 + ], + "score": 0.83, + "latex": "i \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 787, + 868, + 801, + 868, + 801, + 892, + 787, + 892 + ], + "score": 0.7, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 396, + 413, + 792, + 413, + 792, + 473, + 396, + 473 + ], + "score": 0.7, + "latex": "\\begin{array} { r } { \\sqrt { \\frac { \\iota ( t ) } { t } } \\leq \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\sqrt { \\frac { \\iota ( i ) } { i } } \\leq 2 \\sqrt { \\frac { \\iota ( t ) } { t } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 969, + 304, + 1034, + 304, + 1034, + 337, + 969, + 337 + ], + "score": 0.68, + "latex": "t \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 866, + 427, + 1141, + 427, + 1141, + 466, + 866, + 466 + ], + "score": 0.63, + "latex": "\\iota ( t ) = \\ln ( c ( t + 1 ) ( t + 2 ) )" + }, + { + "category_id": 13, + "poly": [ + 831, + 235, + 971, + 235, + 971, + 272, + 831, + 272 + ], + "score": 0.58, + "latex": "t \\geq 1 , c > 0" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1177.0, + 1403.0, + 1177.0, + 1403.0, + 1205.0, + 1377.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 858.0, + 786.0, + 858.0, + 786.0, + 901.0, + 292.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 858.0, + 1079.0, + 858.0, + 1079.0, + 901.0, + 802.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 858.0, + 1290.0, + 858.0, + 1290.0, + 901.0, + 1235.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 858.0, + 1405.0, + 858.0, + 1405.0, + 901.0, + 1356.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 891.0, + 356.0, + 891.0, + 356.0, + 931.0, + 294.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 891.0, + 618.0, + 891.0, + 618.0, + 931.0, + 576.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 891.0, + 952.0, + 891.0, + 952.0, + 931.0, + 778.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1756.0, + 693.0, + 1756.0, + 693.0, + 1799.0, + 294.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1612.0, + 1175.0, + 1612.0, + 1175.0, + 1653.0, + 292.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1612.0, + 1273.0, + 1612.0, + 1273.0, + 1653.0, + 1262.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1042.0, + 465.0, + 1042.0, + 465.0, + 1082.0, + 295.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1042.0, + 978.0, + 1042.0, + 978.0, + 1082.0, + 619.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1042.0, + 1227.0, + 1042.0, + 1227.0, + 1082.0, + 1118.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1899.0, + 697.0, + 1899.0, + 697.0, + 1943.0, + 295.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1218.0, + 703.0, + 1218.0, + 703.0, + 1262.0, + 295.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1218.0, + 881.0, + 1218.0, + 881.0, + 1262.0, + 797.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1218.0, + 1002.0, + 1218.0, + 1002.0, + 1262.0, + 951.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1218.0, + 1129.0, + 1218.0, + 1129.0, + 1262.0, + 1066.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 639.0, + 1170.0, + 639.0, + 1170.0, + 694.0, + 291.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 854.0, + 812.0, + 854.0, + 857.0, + 294.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 499.0, + 507.0, + 499.0, + 507.0, + 536.0, + 296.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1483.0, + 508.0, + 1483.0, + 508.0, + 1520.0, + 293.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 694.0, + 493.0, + 694.0, + 493.0, + 731.0, + 296.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1409.0, + 370.0, + 1409.0, + 370.0, + 1464.0, + 293.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1409.0, + 736.0, + 1409.0, + 736.0, + 1464.0, + 685.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 1409.0, + 1075.0, + 1409.0, + 1075.0, + 1464.0, + 1065.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 291.0, + 395.0, + 291.0, + 395.0, + 349.0, + 351.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 291.0, + 655.0, + 291.0, + 655.0, + 349.0, + 603.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 291.0, + 968.0, + 291.0, + 968.0, + 349.0, + 862.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 291.0, + 1050.0, + 291.0, + 1050.0, + 349.0, + 1035.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 350.0, + 395.0, + 350.0, + 395.0, + 405.0, + 354.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 350.0, + 748.0, + 350.0, + 748.0, + 405.0, + 635.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 350.0, + 826.0, + 350.0, + 826.0, + 405.0, + 813.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 217.0, + 396.0, + 217.0, + 396.0, + 286.0, + 355.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 217.0, + 830.0, + 217.0, + 830.0, + 286.0, + 723.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 217.0, + 990.0, + 217.0, + 990.0, + 286.0, + 972.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 404.0, + 395.0, + 404.0, + 395.0, + 487.0, + 347.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 404.0, + 865.0, + 404.0, + 865.0, + 487.0, + 793.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 404.0, + 1253.0, + 404.0, + 1253.0, + 487.0, + 1142.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 404.0, + 1417.0, + 404.0, + 1417.0, + 487.0, + 1394.0, + 487.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 386, + 1411, + 386, + 1411, + 484, + 295, + 484 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 456, + 1675, + 1242, + 1675, + 1242, + 1861, + 456, + 1861 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 417, + 1229, + 1406, + 1229, + 1406, + 1546, + 417, + 1546 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 574, + 860, + 1127, + 860, + 1127, + 945, + 574, + 945 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 576, + 324, + 1123, + 324, + 1123, + 380, + 576, + 380 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 417, + 669, + 1281, + 669, + 1281, + 773, + 417, + 773 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1406, + 228, + 1406, + 322, + 297, + 322 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 297, + 953, + 1401, + 953, + 1401, + 1018, + 297, + 1018 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 295, + 1602, + 1403, + 1602, + 1403, + 1668, + 295, + 1668 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 581, + 1096, + 1117, + 1096, + 1117, + 1186, + 581, + 1186 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 296, + 1915, + 1405, + 1915, + 1405, + 1992, + 296, + 1992 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 509, + 491, + 1185, + 491, + 1185, + 584, + 509, + 584 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 300, + 781, + 1408, + 781, + 1408, + 851, + 300, + 851 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 297, + 1190, + 574, + 1190, + 574, + 1224, + 297, + 1224 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 296, + 1866, + 1181, + 1866, + 1181, + 1902, + 296, + 1902 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 294, + 1552, + 955, + 1552, + 955, + 1592, + 294, + 1592 + ], + "score": 0.909 + }, + { + "category_id": 1, + "poly": [ + 297, + 587, + 1172, + 587, + 1172, + 623, + 297, + 623 + ], + "score": 0.907 + }, + { + "category_id": 9, + "poly": [ + 1351, + 520, + 1401, + 520, + 1401, + 552, + 1351, + 552 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1351, + 708, + 1400, + 708, + 1400, + 740, + 1351, + 740 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 296, + 1037, + 1163, + 1037, + 1163, + 1088, + 296, + 1088 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 1374, + 2004, + 1402, + 2004, + 1402, + 2031, + 1374, + 2031 + ], + "score": 0.745 + }, + { + "category_id": 14, + "poly": [ + 454, + 1671, + 1242, + 1671, + 1242, + 1865, + 454, + 1865 + ], + "score": 0.94, + "latex": "\\begin{array} { c l } { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\geq - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ { \\geq \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } b _ { i } - \\gamma \\displaystyle \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| \\geq 0 . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 509, + 488, + 1190, + 488, + 1190, + 585, + 509, + 585 + ], + "score": 0.93, + "latex": "\\mathrm { P r } \\left[ \\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| > \\eta \\right] \\leq 2 \\exp \\left\\{ - \\frac { \\eta ^ { 2 } } { 8 \\left( 1 - \\gamma \\right) ^ { - 2 } \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } } \\right\\} ." + }, + { + "category_id": 13, + "poly": [ + 1106, + 1913, + 1205, + 1913, + 1205, + 1953, + 1106, + 1953 + ], + "score": 0.93, + "latex": "\\hat { Q } _ { p } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 352, + 1954, + 458, + 1954, + 458, + 1991, + 352, + 1991 + ], + "score": 0.93, + "latex": "Q _ { p ^ { \\prime } } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 297, + 261, + 454, + 261, + 454, + 295, + 297, + 295 + ], + "score": 0.93, + "latex": "t _ { i } = \\tau ( s , a , i )" + }, + { + "category_id": 13, + "poly": [ + 1032, + 954, + 1183, + 954, + 1183, + 989, + 1032, + 989 + ], + "score": 0.93, + "latex": "t = N _ { p } ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 572, + 853, + 1124, + 853, + 1124, + 948, + 572, + 948 + ], + "score": 0.93, + "latex": "1 - \\sum _ { s ^ { \\prime } \\in S , a ^ { \\prime } \\in A } \\sum _ { k = 1 } ^ { \\infty } \\frac { \\delta } { 2 S A ( k + 1 ) ( k + 2 ) } = 1 - \\frac { \\delta } { 2 } ." + }, + { + "category_id": 13, + "poly": [ + 1252, + 231, + 1396, + 231, + 1396, + 265, + 1252, + 265 + ], + "score": 0.93, + "latex": "t = N _ { p } ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 413, + 670, + 1282, + 670, + 1282, + 775, + 413, + 775 + ], + "score": 0.93, + "latex": "\\left| \\sum _ { i = 1 } ^ { k } \\Delta _ { i } \\right| \\leq \\frac { 2 \\sqrt { 2 } } { 1 - \\gamma } \\cdot \\sqrt { \\sum _ { i = 1 } ^ { k } ( \\alpha _ { k } ^ { i } ) ^ { 2 } \\cdot \\ln \\frac { 2 ( k + 1 ) ( k + 2 ) S A } { \\delta } } \\leq \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } ." + }, + { + "category_id": 13, + "poly": [ + 485, + 782, + 751, + 782, + 751, + 824, + 485, + 824 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\iota ( k ) = \\ln { \\frac { ( k + 1 ) ( k + 2 ) S A } { \\delta } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1107, + 1635, + 1182, + 1635, + 1182, + 1666, + 1107, + 1666 + ], + "score": 0.93, + "latex": "p ^ { \\prime } = 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 417, + 463, + 417, + 463, + 450, + 297, + 450 + ], + "score": 0.92, + "latex": "\\mathbb { E } \\left[ \\Delta _ { i } | F _ { i } \\right] = 0" + }, + { + "category_id": 13, + "poly": [ + 298, + 1635, + 406, + 1635, + 406, + 1669, + 298, + 1669 + ], + "score": 0.92, + "latex": "Q _ { p } \\geq Q ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 486, + 1634, + 548, + 1634, + 548, + 1668, + 486, + 1668 + ], + "score": 0.92, + "latex": "( s , a )" + }, + { + "category_id": 13, + "poly": [ + 599, + 1635, + 674, + 1635, + 674, + 1667, + 599, + 1667 + ], + "score": 0.92, + "latex": "p \\leq p ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 809, + 1953, + 1181, + 1953, + 1181, + 1991, + 809, + 1991 + ], + "score": 0.92, + "latex": "\\hat { Q } _ { p } ( s , a ) \\geq Q _ { p ^ { \\prime } } ( s , a ) \\geq Q ^ { * } ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 574, + 321, + 1124, + 321, + 1124, + 381, + 574, + 381 + ], + "score": 0.92, + "latex": "\\Delta _ { i } = \\left( \\alpha _ { k } ^ { i } \\cdot I [ t _ { i } < \\infty ] \\cdot \\left( \\mathbb { P } V ^ { * } - \\mathbb { \\hat { P } } _ { t _ { i } } V ^ { * } \\right) ( s , a ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 567, + 1954, + 642, + 1954, + 642, + 1988, + 567, + 1988 + ], + "score": 0.92, + "latex": "p ^ { \\prime } \\leq p" + }, + { + "category_id": 13, + "poly": [ + 783, + 417, + 841, + 417, + 841, + 450, + 783, + 450 + ], + "score": 0.92, + "latex": "F _ { i + 1 }" + }, + { + "category_id": 13, + "poly": [ + 427, + 985, + 517, + 985, + 517, + 1019, + 427, + 1019 + ], + "score": 0.91, + "latex": "1 - \\delta / 2" + }, + { + "category_id": 13, + "poly": [ + 664, + 1867, + 738, + 1867, + 738, + 1899, + 664, + 1899 + ], + "score": 0.91, + "latex": "p ^ { \\prime } + 1" + }, + { + "category_id": 13, + "poly": [ + 781, + 1918, + 872, + 1918, + 872, + 1952, + 781, + 1952 + ], + "score": 0.91, + "latex": "1 - \\delta / 2" + }, + { + "category_id": 13, + "poly": [ + 856, + 589, + 1158, + 589, + 1158, + 622, + 856, + 622 + ], + "score": 0.9, + "latex": "1 - \\delta / \\left[ S A ( k + 1 ) ( k + 2 ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 1125, + 231, + 1196, + 231, + 1196, + 262, + 1125, + 262 + ], + "score": 0.9, + "latex": "p \\in \\mathbb N" + }, + { + "category_id": 13, + "poly": [ + 873, + 263, + 937, + 263, + 937, + 292, + 873, + 292 + ], + "score": 0.9, + "latex": "t \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 1257, + 414, + 1397, + 414, + 1397, + 457, + 1257, + 457 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { | \\Delta _ { i } | \\le \\frac { 2 } { 1 - \\gamma } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 517, + 263, + 582, + 263, + 582, + 290, + 517, + 290 + ], + "score": 0.9, + "latex": "t = 0" + }, + { + "category_id": 13, + "poly": [ + 1329, + 789, + 1397, + 789, + 1397, + 817, + 1329, + 817 + ], + "score": 0.89, + "latex": "k > 0" + }, + { + "category_id": 13, + "poly": [ + 420, + 453, + 454, + 453, + 454, + 482, + 420, + 482 + ], + "score": 0.89, + "latex": "\\Delta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 548, + 418, + 583, + 418, + 583, + 447, + 548, + 447 + ], + "score": 0.89, + "latex": "\\Delta _ { i }" + }, + { + "category_id": 13, + "poly": [ + 925, + 232, + 993, + 232, + 993, + 260, + 925, + 260 + ], + "score": 0.89, + "latex": "s \\in S" + }, + { + "category_id": 13, + "poly": [ + 358, + 784, + 474, + 784, + 474, + 819, + 358, + 819 + ], + "score": 0.89, + "latex": "c _ { 2 } = 4 \\sqrt { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1003, + 232, + 1076, + 232, + 1076, + 260, + 1003, + 260 + ], + "score": 0.88, + "latex": "a \\in A" + }, + { + "category_id": 13, + "poly": [ + 937, + 388, + 1149, + 388, + 1149, + 419, + 937, + 419 + ], + "score": 0.88, + "latex": "( s _ { 1 } , a _ { 1 } , . . . , s _ { t _ { i } } , a _ { t _ { i } } )" + }, + { + "category_id": 14, + "poly": [ + 415, + 1228, + 1285, + 1228, + 1285, + 1418, + 415, + 1418 + ], + "score": 0.88, + "latex": "\\begin{array} { r l } & { ( Q _ { p } - Q ^ { * } ) ( s , a ) \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\gamma \\left| \\displaystyle \\sum _ { i = 1 } ^ { t } \\Delta _ { i } \\right| + \\displaystyle \\sum _ { i = 1 } ^ { t } \\alpha _ { t } ^ { i } \\left[ \\gamma ( \\hat { V } _ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + b _ { i } \\right] } \\\\ & { \\qquad \\leq \\displaystyle \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\displaystyle \\frac { 3 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( t ) } { t } } + \\displaystyle \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( \\hat { V } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 580, + 1093, + 1115, + 1093, + 1115, + 1187, + 580, + 1187 + ], + "score": 0.86, + "latex": "\\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } \\leq \\sum _ { i = 1 } ^ { k } \\alpha _ { k } ^ { i } b _ { i } \\leq \\frac { 2 c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } ." + }, + { + "category_id": 13, + "poly": [ + 346, + 388, + 376, + 388, + 376, + 416, + 346, + 416 + ], + "score": 0.86, + "latex": "F _ { i }" + }, + { + "category_id": 13, + "poly": [ + 407, + 1553, + 729, + 1553, + 729, + 1593, + 407, + 1593 + ], + "score": 0.85, + "latex": "\\beta _ { t } = c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { H \\iota ( t ) / t }" + }, + { + "category_id": 13, + "poly": [ + 996, + 417, + 1242, + 417, + 1242, + 455, + 996, + 455 + ], + "score": 0.85, + "latex": "\\textstyle 0 \\leq V ^ { \\ast } ( s , { \\bar { a } } ) \\leq { \\frac { 1 } { 1 - \\gamma } }" + }, + { + "category_id": 13, + "poly": [ + 1380, + 265, + 1398, + 265, + 1398, + 289, + 1380, + 289 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 469, + 391, + 489, + 391, + 489, + 414, + 469, + 414 + ], + "score": 0.81, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1043, + 790, + 1061, + 790, + 1061, + 816, + 1043, + 816 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1155, + 1874, + 1171, + 1874, + 1171, + 1899, + 1155, + 1899 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1317, + 962, + 1334, + 962, + 1334, + 987, + 1317, + 987 + ], + "score": 0.77, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 951, + 1922, + 1024, + 1922, + 1024, + 1951, + 951, + 1951 + ], + "score": 0.77, + "latex": "p , s , a" + }, + { + "category_id": 13, + "poly": [ + 479, + 822, + 551, + 822, + 551, + 847, + 479, + 847 + ], + "score": 0.77, + "latex": "a \\in A" + }, + { + "category_id": 13, + "poly": [ + 444, + 596, + 461, + 596, + 461, + 621, + 444, + 621 + ], + "score": 0.76, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 400, + 821, + 469, + 821, + 469, + 848, + 400, + 848 + ], + "score": 0.74, + "latex": "s \\in S" + }, + { + "category_id": 13, + "poly": [ + 740, + 1552, + 948, + 1552, + 948, + 1590, + 740, + 1590 + ], + "score": 0.73, + "latex": "c _ { 3 } = 3 c _ { 2 } = 1 2 { \\sqrt { 2 } } ." + }, + { + "category_id": 13, + "poly": [ + 801, + 237, + 819, + 237, + 819, + 258, + 801, + 258 + ], + "score": 0.73, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 937, + 1033, + 1157, + 1033, + 1157, + 1089, + 937, + 1089 + ], + "score": 0.68, + "latex": "\\begin{array} { r } { ( b _ { k } = \\frac { c _ { 2 } } { 1 - \\gamma } \\sqrt { \\frac { H \\iota ( k ) } { k } } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 708, + 238, + 725, + 238, + 725, + 262, + 708, + 262 + ], + "score": 0.6, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 736, + 238, + 753, + 238, + 753, + 259, + 736, + 259 + ], + "score": 0.47, + "latex": "s" + }, + { + "category_id": 14, + "poly": [ + 608, + 1453, + 1121, + 1453, + 1121, + 1546, + 608, + 1546 + ], + "score": 0.38, + "latex": "\\leq \\frac { \\alpha _ { t } ^ { 0 } } { 1 - \\gamma } + \\sum _ { i = 1 } ^ { t } \\gamma \\alpha _ { t } ^ { i } ( { \\hat { V } } ^ { t _ { i } } - V ^ { * } ) ( x _ { t _ { i } + 1 } ) + \\beta _ { t } ." + }, + { + "category_id": 13, + "poly": [ + 399, + 820, + 551, + 820, + 551, + 849, + 399, + 849 + ], + "score": 0.33, + "latex": "s \\in S , a \\in A" + }, + { + "category_id": 13, + "poly": [ + 707, + 237, + 753, + 237, + 753, + 261, + 707, + 261 + ], + "score": 0.31, + "latex": "p , s" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1381.0, + 2009.0, + 1401.0, + 2009.0, + 1401.0, + 2029.0, + 1381.0, + 2029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 345.0, + 381.0, + 345.0, + 423.0, + 292.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 381.0, + 468.0, + 381.0, + 468.0, + 423.0, + 377.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 381.0, + 936.0, + 381.0, + 936.0, + 423.0, + 490.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 381.0, + 1408.0, + 381.0, + 1408.0, + 423.0, + 1150.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 401.0, + 296.0, + 401.0, + 296.0, + 465.0, + 287.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 401.0, + 547.0, + 401.0, + 547.0, + 465.0, + 464.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 401.0, + 782.0, + 401.0, + 782.0, + 465.0, + 584.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 401.0, + 995.0, + 401.0, + 995.0, + 465.0, + 842.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 401.0, + 1256.0, + 401.0, + 1256.0, + 465.0, + 1243.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 401.0, + 1410.0, + 401.0, + 1410.0, + 465.0, + 1398.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 446.0, + 419.0, + 446.0, + 419.0, + 488.0, + 293.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 446.0, + 1271.0, + 446.0, + 1271.0, + 488.0, + 455.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 225.0, + 706.0, + 225.0, + 706.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 225.0, + 800.0, + 225.0, + 800.0, + 267.0, + 754.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 225.0, + 924.0, + 225.0, + 924.0, + 267.0, + 820.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 225.0, + 1002.0, + 225.0, + 1002.0, + 267.0, + 994.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 225.0, + 1124.0, + 225.0, + 1124.0, + 267.0, + 1077.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 225.0, + 1251.0, + 225.0, + 1251.0, + 267.0, + 1197.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 225.0, + 1408.0, + 225.0, + 1408.0, + 267.0, + 1397.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 259.0, + 296.0, + 259.0, + 296.0, + 296.0, + 291.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 259.0, + 516.0, + 259.0, + 516.0, + 296.0, + 455.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 259.0, + 872.0, + 259.0, + 872.0, + 296.0, + 583.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 259.0, + 1379.0, + 259.0, + 1379.0, + 296.0, + 938.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 259.0, + 1411.0, + 259.0, + 1411.0, + 296.0, + 1399.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 288.0, + 381.0, + 288.0, + 381.0, + 325.0, + 292.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 949.0, + 1031.0, + 949.0, + 1031.0, + 989.0, + 293.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 949.0, + 1316.0, + 949.0, + 1316.0, + 989.0, + 1184.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 949.0, + 1403.0, + 949.0, + 1403.0, + 989.0, + 1335.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 986.0, + 426.0, + 986.0, + 426.0, + 1019.0, + 293.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 986.0, + 613.0, + 986.0, + 613.0, + 1019.0, + 518.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1600.0, + 1406.0, + 1600.0, + 1406.0, + 1638.0, + 293.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1632.0, + 297.0, + 1632.0, + 297.0, + 1671.0, + 294.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1632.0, + 485.0, + 1632.0, + 485.0, + 1671.0, + 407.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1632.0, + 598.0, + 1632.0, + 598.0, + 1671.0, + 549.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1632.0, + 1106.0, + 1632.0, + 1106.0, + 1671.0, + 675.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1632.0, + 1260.0, + 1632.0, + 1260.0, + 1671.0, + 1183.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 780.0, + 1913.0, + 780.0, + 1955.0, + 294.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1913.0, + 950.0, + 1913.0, + 950.0, + 1955.0, + 873.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1913.0, + 1105.0, + 1913.0, + 1105.0, + 1955.0, + 1025.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1955.0, + 1206.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1952.0, + 351.0, + 1952.0, + 351.0, + 1995.0, + 295.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1952.0, + 566.0, + 1952.0, + 566.0, + 1995.0, + 459.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1952.0, + 808.0, + 1952.0, + 808.0, + 1995.0, + 643.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1952.0, + 1398.0, + 1952.0, + 1398.0, + 1995.0, + 1182.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 769.0, + 357.0, + 769.0, + 357.0, + 831.0, + 289.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 769.0, + 484.0, + 769.0, + 484.0, + 831.0, + 475.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 769.0, + 1042.0, + 769.0, + 1042.0, + 831.0, + 752.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 769.0, + 1328.0, + 769.0, + 1328.0, + 831.0, + 1062.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 769.0, + 1414.0, + 769.0, + 1414.0, + 831.0, + 1398.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 815.0, + 398.0, + 815.0, + 398.0, + 855.0, + 295.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 815.0, + 912.0, + 815.0, + 912.0, + 855.0, + 552.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1187.0, + 576.0, + 1187.0, + 576.0, + 1229.0, + 293.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1862.0, + 663.0, + 1862.0, + 663.0, + 1905.0, + 293.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1862.0, + 1154.0, + 1862.0, + 1154.0, + 1905.0, + 739.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1862.0, + 1182.0, + 1862.0, + 1182.0, + 1905.0, + 1172.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1548.0, + 406.0, + 1548.0, + 406.0, + 1596.0, + 294.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1548.0, + 739.0, + 1548.0, + 739.0, + 1596.0, + 730.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1548.0, + 955.0, + 1548.0, + 955.0, + 1596.0, + 949.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 584.0, + 443.0, + 584.0, + 443.0, + 627.0, + 295.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 584.0, + 855.0, + 584.0, + 855.0, + 627.0, + 462.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 584.0, + 1171.0, + 584.0, + 1171.0, + 627.0, + 1159.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1045.0, + 936.0, + 1045.0, + 936.0, + 1083.0, + 293.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1029.0, + 1164.0, + 1029.0, + 1164.0, + 1075.0, + 1158.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1062.0, + 1056.0, + 1062.0, + 1056.0, + 1084.5, + 1004.0, + 1084.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 444, + 376, + 1252, + 376, + 1252, + 629, + 444, + 629 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 578, + 882, + 1120, + 882, + 1120, + 1081, + 578, + 1081 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 551, + 1350, + 1145, + 1350, + 1145, + 1592, + 551, + 1592 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 648, + 1663, + 1048, + 1663, + 1048, + 1813, + 648, + 1813 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 716, + 716, + 983, + 716, + 983, + 804, + 716, + 804 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 633, + 1896, + 1064, + 1896, + 1064, + 1983, + 633, + 1983 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 291, + 227, + 1403, + 227, + 1403, + 298, + 291, + 298 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 1224, + 1406, + 1224, + 1406, + 1295, + 297, + 1295 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 297, + 639, + 1403, + 639, + 1403, + 708, + 297, + 708 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 295, + 1831, + 1400, + 1831, + 1400, + 1895, + 295, + 1895 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 295, + 831, + 1063, + 831, + 1063, + 870, + 295, + 870 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 296, + 329, + 870, + 329, + 870, + 364, + 296, + 364 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 297, + 1308, + 804, + 1308, + 804, + 1340, + 297, + 1340 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 296, + 1642, + 377, + 1642, + 377, + 1674, + 296, + 1674 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 817, + 73, + 817, + 106, + 298, + 106 + ], + "score": 0.921 + }, + { + "category_id": 8, + "poly": [ + 704, + 1174, + 1085, + 1174, + 1085, + 1215, + 704, + 1215 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1351, + 739, + 1401, + 739, + 1401, + 771, + 1351, + 771 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1181, + 1400, + 1181, + 1400, + 1212, + 1351, + 1212 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1351, + 563, + 1401, + 563, + 1401, + 595, + 1351, + 595 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1499, + 1401, + 1499, + 1401, + 1531, + 1351, + 1531 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1019, + 1401, + 1019, + 1401, + 1051, + 1351, + 1051 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1597, + 1401, + 1597, + 1401, + 1629, + 1351, + 1629 + ], + "score": 0.885 + }, + { + "category_id": 8, + "poly": [ + 702, + 1090, + 1113, + 1090, + 1113, + 1165, + 702, + 1165 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1103, + 1400, + 1103, + 1400, + 1135, + 1351, + 1135 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 328, + 2002, + 1096, + 2002, + 1096, + 2036, + 328, + 2036 + ], + "score": 0.731 + }, + { + "category_id": 2, + "poly": [ + 328, + 2002, + 1096, + 2002, + 1096, + 2036, + 328, + 2036 + ], + "score": 0.159 + }, + { + "category_id": 13, + "poly": [ + 457, + 262, + 532, + 262, + 532, + 295, + 457, + 295 + ], + "score": 0.95, + "latex": "( C , w )" + }, + { + "category_id": 13, + "poly": [ + 424, + 331, + 602, + 331, + 602, + 365, + 424, + 365 + ], + "score": 0.95, + "latex": "n _ { t } = N _ { t } ( s _ { t } , a _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 577, + 881, + 1119, + 881, + 1119, + 1217, + 577, + 1217 + ], + "score": 0.94, + "latex": "\\begin{array} { l } { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } \\beta _ { n _ { t } } = \\sum _ { s , o } ^ { u ( s , o ) } w _ { \\tau ( s , a , i ) } \\beta _ { i } } \\\\ { \\displaystyle \\qquad \\leq \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sum _ { i = 1 } ^ { C _ { s , a } / w } \\sqrt { \\frac { H \\iota ( i ) } { i } } w } \\\\ { \\displaystyle \\qquad \\leq 2 \\sum _ { s , a } ( 1 - \\gamma ) ^ { - 1 } c _ { 3 } \\sqrt { \\iota ( C ) H C _ { s , a } w } } \\\\ { \\displaystyle \\qquad \\leq 2 c _ { 3 } ( 1 - \\gamma ) ^ { - 1 } \\sqrt { w S A H C \\iota ( C ) } . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 632, + 1889, + 1066, + 1889, + 1066, + 1985, + 632, + 1985 + ], + "score": 0.94, + "latex": "w _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq w \\sum _ { j = n _ { t ^ { \\prime } } } ^ { \\infty } \\alpha _ { j } ^ { n _ { t ^ { \\prime } } } = ( 1 + 1 / H ) w ." + }, + { + "category_id": 14, + "poly": [ + 439, + 374, + 1253, + 374, + 1253, + 630, + 439, + 630 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } { { \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } ( Q _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq \\sum _ { t \\geq 1 } w _ { t } [ \\cfrac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } + \\beta _ { n _ { t } } + \\gamma \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } ( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } ) ( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } ) ] } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 649, + 1662, + 1049, + 1662, + 1049, + 1810, + 649, + 1810 + ], + "score": 0.94, + "latex": "w _ { t ^ { \\prime } + 1 } ^ { \\prime } = \\left( \\sum _ { \\stackrel { t = t ^ { \\prime } + 1 } { ( s _ { t } , a _ { t } ) = ( s _ { t } ^ { \\prime } , a _ { t } ^ { \\prime } ) } } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) ." + }, + { + "category_id": 14, + "poly": [ + 713, + 713, + 984, + 713, + 984, + 806, + 713, + 806 + ], + "score": 0.93, + "latex": "\\sum _ { t \\geq 1 } w _ { t } \\frac { \\alpha _ { n _ { t } } ^ { 0 } } { 1 - \\gamma } \\leq \\frac { S A w } { 1 - \\gamma } ." + }, + { + "category_id": 14, + "poly": [ + 550, + 1348, + 1150, + 1348, + 1150, + 1593, + 550, + 1593 + ], + "score": 0.93, + "latex": "\\begin{array} { c } { { \\displaystyle \\sum _ { t \\ge 1 } w _ { t } \\sum _ { i = 1 } ^ { n _ { t } } \\alpha _ { n _ { t } } ^ { i } \\left( \\hat { V } _ { \\tau ( s _ { t } , a _ { t } , i ) } - V ^ { * } \\right) \\left( s _ { \\tau ( s _ { t } , a _ { t } , i ) + 1 } \\right) } } \\\\ { { \\displaystyle \\le \\sum _ { t ^ { \\prime } \\ge 1 } \\left( \\hat { V } _ { t ^ { \\prime } } - V ^ { * } \\right) \\left( s _ { t ^ { \\prime } + 1 } \\right) \\left( \\sum _ { t = t ^ { \\prime } + 1 } ^ { \\infty } \\alpha _ { n _ { t } } ^ { n _ { t ^ { \\prime } } } w _ { t } \\right) . } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 606, + 834, + 874, + 834, + 874, + 869, + 606, + 869 + ], + "score": 0.93, + "latex": "u ( s , a ) = \\operatorname* { s u p } _ { t } N _ { t } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 823, + 640, + 1007, + 640, + 1007, + 679, + 823, + 679 + ], + "score": 0.93, + "latex": "\\alpha _ { n _ { t } } ^ { 0 } = \\mathbb { I } [ n _ { t } = 0 ]" + }, + { + "category_id": 13, + "poly": [ + 460, + 1834, + 520, + 1834, + 520, + 1870, + 460, + 1870 + ], + "score": 0.92, + "latex": "w _ { t + 1 } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 261, + 1330, + 261, + 1330, + 300, + 1168, + 300 + ], + "score": 0.92, + "latex": "\\textstyle \\sum _ { t \\geq 1 } w _ { t } \\leq C" + }, + { + "category_id": 13, + "poly": [ + 828, + 263, + 969, + 263, + 969, + 293, + 828, + 293 + ], + "score": 0.92, + "latex": "0 \\leq w _ { t } \\leq w" + }, + { + "category_id": 13, + "poly": [ + 975, + 231, + 1051, + 231, + 1051, + 263, + 975, + 263 + ], + "score": 0.92, + "latex": "( C , w )" + }, + { + "category_id": 13, + "poly": [ + 686, + 262, + 793, + 262, + 793, + 294, + 686, + 294 + ], + "score": 0.92, + "latex": "C , w > 0" + }, + { + "category_id": 13, + "poly": [ + 572, + 1832, + 742, + 1832, + 742, + 1871, + 572, + 1871 + ], + "score": 0.92, + "latex": "\\textstyle ( C , ( 1 + { \\frac { 1 } { H } } ) w )" + }, + { + "category_id": 13, + "poly": [ + 1312, + 230, + 1403, + 230, + 1403, + 265, + 1312, + 265 + ], + "score": 0.91, + "latex": "( w _ { t } ) _ { t \\geq 1 }" + }, + { + "category_id": 13, + "poly": [ + 379, + 1225, + 706, + 1225, + 706, + 1265, + 379, + 1265 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { C _ { s , a } = \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } ) = ( s , a ) } w _ { t } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1260, + 377, + 1260, + 377, + 1295, + 298, + 1295 + ], + "score": 0.9, + "latex": "\\iota ( x ) / x" + }, + { + "category_id": 13, + "poly": [ + 1046, + 263, + 1110, + 263, + 1110, + 292, + 1046, + 292 + ], + "score": 0.89, + "latex": "t \\geq 1" + }, + { + "category_id": 13, + "poly": [ + 651, + 2006, + 709, + 2006, + 709, + 2035, + 651, + 2035 + ], + "score": 0.89, + "latex": "( s , a )" + }, + { + "category_id": 13, + "poly": [ + 333, + 2003, + 416, + 2003, + 416, + 2035, + 333, + 2035 + ], + "score": 0.89, + "latex": "^ 2 u ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1865, + 369, + 1865, + 369, + 1897, + 298, + 1897 + ], + "score": 0.88, + "latex": "t ^ { \\prime } \\geq 0" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 2001.0, + 650.0, + 2001.0, + 650.0, + 2038.0, + 417.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 2001.0, + 1097.0, + 2001.0, + 1097.0, + 2038.0, + 710.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 221.0, + 974.0, + 221.0, + 974.0, + 272.0, + 292.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 221.0, + 1311.0, + 221.0, + 1311.0, + 272.0, + 1052.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 456.0, + 258.0, + 456.0, + 300.0, + 293.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 258.0, + 685.0, + 258.0, + 685.0, + 300.0, + 533.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 258.0, + 827.0, + 258.0, + 827.0, + 300.0, + 794.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 258.0, + 1045.0, + 258.0, + 1045.0, + 300.0, + 970.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 258.0, + 1167.0, + 258.0, + 1167.0, + 300.0, + 1111.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 258.0, + 1342.0, + 258.0, + 1342.0, + 300.0, + 1331.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1213.0, + 378.0, + 1213.0, + 378.0, + 1274.0, + 289.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1213.0, + 1409.0, + 1213.0, + 1409.0, + 1274.0, + 707.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1255.0, + 1251.0, + 1255.0, + 1251.0, + 1300.0, + 378.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 635.0, + 822.0, + 635.0, + 822.0, + 684.0, + 291.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 635.0, + 1409.0, + 635.0, + 1409.0, + 684.0, + 1008.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 668.0, + 473.0, + 668.0, + 473.0, + 711.0, + 293.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1825.0, + 459.0, + 1825.0, + 459.0, + 1876.0, + 289.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1825.0, + 571.0, + 1825.0, + 571.0, + 1876.0, + 521.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1825.0, + 1407.0, + 1825.0, + 1407.0, + 1876.0, + 743.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1857.0, + 297.0, + 1857.0, + 297.0, + 1904.0, + 290.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1857.0, + 385.0, + 1857.0, + 385.0, + 1904.0, + 370.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 831.0, + 605.0, + 831.0, + 605.0, + 872.0, + 294.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 831.0, + 1064.0, + 831.0, + 1064.0, + 872.0, + 875.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 329.0, + 423.0, + 329.0, + 423.0, + 369.0, + 295.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 329.0, + 869.0, + 329.0, + 869.0, + 369.0, + 603.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1305.0, + 805.0, + 1305.0, + 805.0, + 1343.0, + 294.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1638.0, + 379.0, + 1638.0, + 379.0, + 1679.0, + 294.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 2001.0, + 650.0, + 2001.0, + 650.0, + 2038.0, + 417.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 2001.0, + 1097.0, + 2001.0, + 1097.0, + 2038.0, + 710.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1267, + 1410, + 1267, + 1410, + 1440, + 296, + 1440 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 301, + 750, + 1398, + 750, + 1398, + 853, + 301, + 853 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 349, + 1454, + 1353, + 1454, + 1353, + 1714, + 349, + 1714 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 1760, + 1406, + 1760, + 1406, + 1849, + 297, + 1849 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 330, + 1087, + 1323, + 1087, + 1323, + 1256, + 330, + 1256 + ], + "score": 0.967 + }, + { + "category_id": 8, + "poly": [ + 410, + 871, + 1285, + 871, + 1285, + 974, + 410, + 974 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 480, + 308, + 1219, + 308, + 1219, + 467, + 480, + 467 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 290, + 225, + 1405, + 225, + 1405, + 297, + 290, + 297 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 317, + 1896, + 1334, + 1896, + 1334, + 1989, + 317, + 1989 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 297, + 1001, + 804, + 1001, + 804, + 1036, + 297, + 1036 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 481, + 474, + 1186, + 474, + 1186, + 558, + 481, + 558 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 817, + 73, + 817, + 106, + 297, + 106 + ], + "score": 0.92 + }, + { + "category_id": 8, + "poly": [ + 480, + 655, + 1168, + 655, + 1168, + 737, + 480, + 737 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1351, + 584, + 1400, + 584, + 1400, + 615, + 1351, + 615 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1351, + 674, + 1401, + 674, + 1401, + 706, + 1351, + 706 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1351, + 494, + 1401, + 494, + 1401, + 525, + 1351, + 525 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1350, + 402, + 1401, + 402, + 1401, + 434, + 1350, + 434 + ], + "score": 0.875 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1192, + 1401, + 1192, + 1401, + 1224, + 1352, + 1224 + ], + "score": 0.869 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 864, + 2087, + 864, + 2113, + 835, + 2113 + ], + "score": 0.863 + }, + { + "category_id": 8, + "poly": [ + 481, + 566, + 1100, + 566, + 1100, + 646, + 481, + 646 + ], + "score": 0.853 + }, + { + "category_id": 9, + "poly": [ + 1353, + 1926, + 1400, + 1926, + 1400, + 1956, + 1353, + 1956 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 1374, + 2004, + 1402, + 2004, + 1402, + 2032, + 1374, + 2032 + ], + "score": 0.804 + }, + { + "category_id": 14, + "poly": [ + 471, + 308, + 1219, + 308, + 1219, + 743, + 471, + 743 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { \\quad \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { = \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t } - \\hat { V } _ { t + 1 } \\right) \\left( s _ { t + 1 } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\underset { t \\geq 1 } { \\sum \\sigma } u _ { t + 1 } ^ { \\prime } \\left( 2 \\alpha _ { n _ { t } + 1 } \\frac { 1 } { 1 - \\gamma } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { V } _ { t + 1 } - V ^ { * } \\right) \\left( s _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\\\ & { \\quad \\leq \\underset { t \\geq 1 } { \\sum \\sigma } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) \\left( s _ { t + 1 } , a _ { t + 1 } \\right) + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln { C } \\right) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 349, + 1450, + 1353, + 1450, + 1353, + 1714, + 349, + 1714 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ^ { ( H ) } \\left( \\hat { Q } _ { t } - Q ^ { * } \\right) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w ^ { ( H ) } S A H C \\iota ( C ) } } { ( 1 - \\gamma ) ^ { 2 } } + \\mathcal { O } \\left( \\frac { w ^ { ( H ) } S A H } { ( 1 - \\gamma ) ^ { 2 } } \\ln C \\right) + \\gamma ^ { H } \\frac { C } { 1 - \\gamma } . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1009, + 1369, + 1128, + 1369, + 1128, + 1412, + 1009, + 1412 + ], + "score": 0.94, + "latex": "\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }" + }, + { + "category_id": 13, + "poly": [ + 379, + 1761, + 778, + 1761, + 778, + 1810, + 379, + 1810 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { H = \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { \\ln 1 / \\gamma } \\leq \\frac { \\ln 1 / ( ( 1 - \\gamma ) \\epsilon _ { 1 } ) } { 1 - \\gamma } , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 315, + 1895, + 1339, + 1895, + 1339, + 1989, + 315, + 1989 + ], + "score": 0.93, + "latex": "\\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) \\leq \\frac { C \\epsilon _ { 1 } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { \\sqrt { w S A C \\ell ( C ) } } { ( 1 - \\gamma ) ^ { 2 . 5 } } + \\frac { w S A } { ( 1 - \\gamma ) ^ { 3 } } \\ln C \\ln \\frac { 1 } { ( 1 - \\gamma ) \\epsilon _ { 1 } } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 803, + 1331, + 921, + 1331, + 921, + 1373, + 803, + 1373 + ], + "score": 0.93, + "latex": "\\{ w _ { t } ^ { ( k ) } \\} _ { t \\ge 1 }" + }, + { + "category_id": 13, + "poly": [ + 597, + 228, + 943, + 228, + 943, + 269, + 597, + 269 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\sum _ { t ^ { \\prime } \\geq 1 } w _ { t ^ { \\prime } + 1 } ^ { \\prime } \\leq \\sum _ { t \\geq 1 } w _ { t } \\leq C . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 573, + 1372, + 751, + 1372, + 751, + 1411, + 573, + 1411 + ], + "score": 0.93, + "latex": "w \\cdot \\left( 1 + 1 / H \\right) ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 606, + 1330, + 724, + 1330, + 724, + 1372, + 606, + 1372 + ], + "score": 0.93, + "latex": "\\{ \\bar { w _ { t } ^ { ( 0 ) } } \\} _ { t \\ge 1 }" + }, + { + "category_id": 13, + "poly": [ + 347, + 1807, + 515, + 1807, + 515, + 1851, + 347, + 1851 + ], + "score": 0.92, + "latex": "\\gamma ^ { H } \\frac { C } { 1 - \\gamma } \\leq C \\epsilon _ { 1 }" + }, + { + "category_id": 14, + "poly": [ + 412, + 866, + 1286, + 866, + 1286, + 975, + 412, + 975 + ], + "score": 0.92, + "latex": "\\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\alpha _ { n _ { t } + 1 } \\leq \\sum _ { s , a } \\sum _ { n = 1 } ^ { C _ { s , a } ^ { \\prime } / w ^ { \\prime } } w ^ { \\prime } \\frac { H } { n } \\leq \\sum _ { s , a } H w ^ { \\prime } \\ln ( C _ { s , a } ^ { \\prime } / w ) \\leq 2 S A H w \\ln C ." + }, + { + "category_id": 14, + "poly": [ + 326, + 1083, + 1326, + 1083, + 1326, + 1260, + 326, + 1260 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { \\displaystyle \\sum _ { t \\geq 1 } w _ { t } ( \\hat { Q } _ { t } - Q ^ { * } ) ( s _ { t } , a _ { t } ) } \\\\ & { \\leq 2 c _ { 3 } \\frac { \\sqrt { w S A H C \\iota ( C ) } } { 1 - \\gamma } + \\mathcal { O } \\left( \\frac { w S A H } { 1 - \\gamma } \\ln C \\right) + \\gamma \\displaystyle \\sum _ { t \\geq 1 } w _ { t + 1 } ^ { \\prime } \\left( \\hat { Q } _ { t + 1 } - Q ^ { * } \\right) ( s _ { t + 1 } , a _ { t + 1 } ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 951, + 1765, + 1344, + 1765, + 1344, + 1804, + 951, + 1804 + ], + "score": 0.92, + "latex": "w ^ { ( H ) } = ( 1 + 1 / H ) ^ { H } w ^ { ( 0 ) } \\leq e w ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 1179, + 1373, + 1285, + 1373, + 1285, + 1411, + 1179, + 1411 + ], + "score": 0.91, + "latex": "( C , w ^ { ( k ) } )" + }, + { + "category_id": 13, + "poly": [ + 1315, + 782, + 1404, + 782, + 1404, + 820, + 1315, + 820 + ], + "score": 0.91, + "latex": "C _ { s , a } ^ { \\prime } =" + }, + { + "category_id": 13, + "poly": [ + 336, + 224, + 492, + 224, + 492, + 269, + 336, + 269 + ], + "score": 0.91, + "latex": "\\textstyle \\sum _ { j = 0 } ^ { i } \\alpha _ { i } ^ { j } = 1" + }, + { + "category_id": 13, + "poly": [ + 401, + 782, + 771, + 782, + 771, + 816, + 401, + 816 + ], + "score": 0.91, + "latex": "\\dot { \\alpha } _ { t } = ( H + 1 ) / ( H + t ) \\overset { \\cdot } { \\leq } H / t" + }, + { + "category_id": 13, + "poly": [ + 481, + 1373, + 534, + 1373, + 534, + 1406, + 481, + 1406 + ], + "score": 0.9, + "latex": "w ^ { ( k ) }" + }, + { + "category_id": 13, + "poly": [ + 629, + 1807, + 980, + 1807, + 980, + 1845, + 629, + 1845 + ], + "score": 0.9, + "latex": "\\ell ( C ) = \\iota ( C ) \\ln ( ( 1 - \\gamma ) ^ { - 1 } \\epsilon _ { 1 } ^ { - 1 } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 816, + 537, + 816, + 537, + 857, + 297, + 857 + ], + "score": 0.85, + "latex": "\\scriptstyle \\sum _ { t \\geq 1 , ( s _ { t } , a _ { t } = s , a } w _ { t + 1 } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 550, + 817, + 761, + 817, + 761, + 851, + 550, + 851 + ], + "score": 0.83, + "latex": "w ^ { \\prime } = w ( 1 + 1 / H )" + }, + { + "category_id": 13, + "poly": [ + 598, + 1411, + 626, + 1411, + 626, + 1435, + 598, + 1435 + ], + "score": 0.81, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 337, + 1378, + 356, + 1378, + 356, + 1405, + 337, + 1405 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1380.0, + 2009.0, + 1401.0, + 2009.0, + 1401.0, + 2030.0, + 1380.0, + 2030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1265.0, + 1406.0, + 1265.0, + 1406.0, + 1304.0, + 294.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1406.0, + 1300.0, + 1406.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1322.0, + 605.0, + 1322.0, + 605.0, + 1383.0, + 285.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1322.0, + 802.0, + 1322.0, + 802.0, + 1383.0, + 725.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1322.0, + 1413.0, + 1322.0, + 1413.0, + 1383.0, + 922.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1366.0, + 336.0, + 1366.0, + 336.0, + 1419.0, + 289.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1366.0, + 480.0, + 1366.0, + 480.0, + 1419.0, + 357.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1366.0, + 572.0, + 1366.0, + 572.0, + 1419.0, + 535.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1366.0, + 1008.0, + 1366.0, + 1008.0, + 1419.0, + 752.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1366.0, + 1178.0, + 1366.0, + 1178.0, + 1419.0, + 1129.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 1366.0, + 1413.0, + 1366.0, + 1413.0, + 1419.0, + 1286.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1406.0, + 597.0, + 1406.0, + 597.0, + 1441.0, + 294.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 1406.0, + 772.0, + 1406.0, + 772.0, + 1441.0, + 627.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 751.0, + 1402.0, + 751.0, + 1402.0, + 786.0, + 297.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 771.0, + 400.0, + 771.0, + 400.0, + 827.0, + 291.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 771.0, + 1314.0, + 771.0, + 1314.0, + 827.0, + 772.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 813.0, + 296.0, + 813.0, + 296.0, + 861.0, + 291.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 813.0, + 549.0, + 813.0, + 549.0, + 861.0, + 538.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 813.0, + 842.0, + 813.0, + 842.0, + 861.0, + 762.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1748.0, + 378.0, + 1748.0, + 378.0, + 1814.0, + 285.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 1748.0, + 950.0, + 1748.0, + 950.0, + 1814.0, + 779.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1748.0, + 1412.0, + 1748.0, + 1412.0, + 1814.0, + 1345.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1804.0, + 628.0, + 1804.0, + 628.0, + 1849.0, + 516.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1804.0, + 1117.0, + 1804.0, + 1117.0, + 1849.0, + 981.0, + 1849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1796.0, + 442.0, + 1796.0, + 442.0, + 1856.0, + 284.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 206.0, + 335.0, + 206.0, + 335.0, + 283.0, + 282.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 206.0, + 596.0, + 206.0, + 596.0, + 283.0, + 493.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 206.0, + 1416.0, + 206.0, + 1416.0, + 283.0, + 944.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 258.0, + 405.0, + 258.0, + 405.0, + 303.0, + 291.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 996.0, + 807.0, + 996.0, + 807.0, + 1043.0, + 293.0, + 1043.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 648, + 1406, + 648, + 1406, + 826, + 296, + 826 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 293, + 1406, + 293, + 1406, + 419, + 298, + 419 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1176, + 1406, + 1176, + 1406, + 1274, + 297, + 1274 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 841, + 1407, + 841, + 1407, + 993, + 297, + 993 + ], + "score": 0.974 + }, + { + "category_id": 8, + "poly": [ + 534, + 1548, + 1163, + 1548, + 1163, + 1798, + 534, + 1798 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 582, + 1052, + 1113, + 1052, + 1113, + 1150, + 582, + 1150 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 365, + 439, + 1404, + 439, + 1404, + 629, + 365, + 629 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 594, + 1286, + 1101, + 1286, + 1101, + 1365, + 594, + 1365 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 295, + 1812, + 1402, + 1812, + 1402, + 1882, + 295, + 1882 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 294, + 1451, + 1405, + 1451, + 1405, + 1533, + 294, + 1533 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 1373, + 1403, + 1373, + 1403, + 1438, + 298, + 1438 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 297, + 1968, + 1403, + 1968, + 1403, + 2036, + 297, + 2036 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 297, + 1004, + 977, + 1004, + 977, + 1038, + 297, + 1038 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 297, + 1915, + 887, + 1915, + 887, + 1947, + 297, + 1947 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 817, + 74, + 817, + 106, + 298, + 106 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 822, + 226, + 822, + 262, + 300, + 262 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2113, + 835, + 2113 + ], + "score": 0.86 + }, + { + "category_id": 14, + "poly": [ + 531, + 1545, + 1167, + 1545, + 1167, + 1800, + 531, + 1800 + ], + "score": 0.94, + "latex": "\\begin{array} { r l r } { { \\operatorname { R e g r e t } ( T ) \\le T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } ( | k : \\{ X _ { k } \\ge \\epsilon _ { 2 } \\cdot 2 ^ { i - 1 } \\} | ) \\epsilon _ { 2 } \\cdot 2 ^ { i } } } \\\\ & { } & \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( T \\epsilon _ { 2 } + \\sum _ { i = 1 } ^ { M } \\frac { S A \\ln 1 / \\delta } { \\epsilon _ { 2 } \\cdot 2 ^ { i - 2 } } ) } \\\\ & { } & { \\le \\tilde { \\mathcal { O } } ( \\sqrt { S A T } \\ln 1 / \\delta ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 594, + 1285, + 1105, + 1285, + 1105, + 1363, + 594, + 1363 + ], + "score": 0.94, + "latex": "\\mathcal { O } \\left( \\frac { S A \\ln S A \\ln 1 / \\delta } { \\epsilon _ { 2 } ^ { 2 } ( 1 - \\gamma ) ^ { 7 } } \\mathrm { p o l y l o g } \\left( \\frac { 1 } { \\epsilon _ { 1 } } , \\frac { 1 } { 1 - \\gamma } \\right) \\right)" + }, + { + "category_id": 14, + "poly": [ + 583, + 1049, + 1114, + 1049, + 1114, + 1151, + 583, + 1151 + ], + "score": 0.94, + "latex": "\\mathrm { R e g r e t } ( T ) = \\sum _ { k = 1 } ^ { T / H } \\left[ V ^ { * } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } ) \\right] \\propto T ^ { 1 / 2 } ." + }, + { + "category_id": 13, + "poly": [ + 439, + 684, + 736, + 684, + 736, + 732, + 439, + 732 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V _ { 1 } ^ { \\ast } ( s _ { 1 } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 888, + 840, + 1012, + 840, + 1012, + 883, + 888, + 883 + ], + "score": 0.94, + "latex": "\\tilde { \\mathcal O } \\big ( \\frac { 3 S A H ^ { 9 } } { \\epsilon ^ { 2 } } \\big )" + }, + { + "category_id": 13, + "poly": [ + 1008, + 954, + 1102, + 954, + 1102, + 995, + 1008, + 995 + ], + "score": 0.94, + "latex": "\\tilde { \\mathcal { O } } \\left( \\epsilon ^ { - 2 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 1203, + 1451, + 1355, + 1451, + 1355, + 1490, + 1203, + 1490 + ], + "score": 0.94, + "latex": "\\epsilon = \\sqrt { S A / T }" + }, + { + "category_id": 13, + "poly": [ + 975, + 1207, + 1217, + 1207, + 1217, + 1246, + 975, + 1246 + ], + "score": 0.93, + "latex": "\\epsilon _ { 2 } > \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )" + }, + { + "category_id": 13, + "poly": [ + 465, + 881, + 589, + 881, + 589, + 924, + 465, + 924 + ], + "score": 0.93, + "latex": "\\tilde { \\mathcal O } \\big ( { \\textstyle { \\frac { 3 S A H ^ { 8 } } { \\epsilon ^ { 2 } } } } \\big )" + }, + { + "category_id": 13, + "poly": [ + 1174, + 692, + 1361, + 692, + 1361, + 728, + 1174, + 728 + ], + "score": 0.93, + "latex": "\\pi _ { h } ( s ) = \\bar { \\pi } ( \\bar { s } _ { s , h } )" + }, + { + "category_id": 13, + "poly": [ + 830, + 922, + 988, + 922, + 988, + 957, + 830, + 957 + ], + "score": 0.93, + "latex": "0 \\leq t < T / H" + }, + { + "category_id": 13, + "poly": [ + 298, + 1487, + 526, + 1487, + 526, + 1527, + 298, + 1527 + ], + "score": 0.93, + "latex": "\\epsilon _ { 2 } = \\tilde { \\mathcal { O } } ( \\epsilon _ { 1 } / ( 1 - \\gamma ) )" + }, + { + "category_id": 13, + "poly": [ + 401, + 790, + 714, + 790, + 714, + 826, + 401, + 826 + ], + "score": 0.93, + "latex": "\\gamma ^ { H } = ( 1 - \\dot { 1 } / H ) ^ { \\tilde { H } } = \\mathcal { O } ( \\dot { 1 } )" + }, + { + "category_id": 13, + "poly": [ + 738, + 593, + 1068, + 593, + 1068, + 625, + 738, + 625 + ], + "score": 0.92, + "latex": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]" + }, + { + "category_id": 13, + "poly": [ + 765, + 1970, + 1263, + 1970, + 1263, + 2007, + 765, + 2007 + ], + "score": 0.91, + "latex": "\\begin{array} { r c l } { \\mathcal { M } } & { = } & { ( S , A , H , r _ { h } ( s , a ) , p _ { h } ( s ^ { \\prime } \\mathrm { ~ ~ \\chi ~ } | \\mathrm { ~ ~ \\chi ~ } s , a ) ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 583, + 1489, + 815, + 1489, + 815, + 1533, + 583, + 1533 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { M = \\lceil \\log _ { 2 } \\frac { 1 } { \\epsilon _ { 2 } ( 1 - \\gamma ) } \\rceil } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 528, + 593, + 687, + 593, + 687, + 625, + 528, + 625 + ], + "score": 0.91, + "latex": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0" + }, + { + "category_id": 13, + "poly": [ + 1241, + 958, + 1332, + 958, + 1332, + 990, + 1241, + 990 + ], + "score": 0.91, + "latex": "S , A , H" + }, + { + "category_id": 13, + "poly": [ + 692, + 761, + 776, + 761, + 776, + 796, + 692, + 796 + ], + "score": 0.91, + "latex": "( \\epsilon / \\gamma ^ { H } )" + }, + { + "category_id": 13, + "poly": [ + 831, + 650, + 869, + 650, + 869, + 687, + 831, + 687 + ], + "score": 0.91, + "latex": "V _ { h } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 443, + 1453, + 717, + 1453, + 717, + 1489, + 443, + 1489 + ], + "score": 0.91, + "latex": "X _ { k } = V ^ { \\ast } ( s _ { 1 } ) - V _ { 1 } ^ { k } ( s _ { 1 } )" + }, + { + "category_id": 13, + "poly": [ + 898, + 559, + 1296, + 559, + 1296, + 593, + 898, + 593 + ], + "score": 0.91, + "latex": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s ^ { \\prime } \\mid s , h )" + }, + { + "category_id": 13, + "poly": [ + 511, + 559, + 844, + 559, + 844, + 593, + 511, + 593 + ], + "score": 0.91, + "latex": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 868, + 326, + 1293, + 326, + 1293, + 359, + 868, + 359 + ], + "score": 0.91, + "latex": "\\mathcal { M } = ( S , A , H , r _ { h } ( s _ { \\perp } a ) , p _ { h } ( s ^ { \\prime } \\mid s , a ) )" + }, + { + "category_id": 13, + "poly": [ + 296, + 2000, + 616, + 2000, + 616, + 2036, + 296, + 2036 + ], + "score": 0.9, + "latex": "( \\bar { S } , \\bar { A } , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )" + }, + { + "category_id": 13, + "poly": [ + 471, + 1847, + 730, + 1847, + 730, + 1879, + 471, + 1879 + ], + "score": 0.9, + "latex": "( H , \\log T , \\log S , \\log A )" + }, + { + "category_id": 13, + "poly": [ + 709, + 921, + 785, + 921, + 785, + 952, + 709, + 952 + ], + "score": 0.9, + "latex": "\\pi ^ { t H + 1 }" + }, + { + "category_id": 13, + "poly": [ + 341, + 650, + 370, + 650, + 370, + 683, + 341, + 683 + ], + "score": 0.88, + "latex": "\\bar { V } _ { t }" + }, + { + "category_id": 13, + "poly": [ + 393, + 849, + 462, + 849, + 462, + 877, + 393, + 877 + ], + "score": 0.88, + "latex": "\\epsilon > 0" + }, + { + "category_id": 13, + "poly": [ + 397, + 592, + 478, + 592, + 478, + 620, + 397, + 620 + ], + "score": 0.88, + "latex": "h = H" + }, + { + "category_id": 13, + "poly": [ + 684, + 530, + 731, + 530, + 731, + 559, + 684, + 559 + ], + "score": 0.88, + "latex": "\\bar { s } _ { s , h }" + }, + { + "category_id": 13, + "poly": [ + 1012, + 359, + 1402, + 359, + 1402, + 390, + 1012, + 390 + ], + "score": 0.88, + "latex": "\\mathcal { M } = ( S , A , \\gamma , \\bar { r } ( \\bar { s } , \\bar { a } ) , \\bar { p } ( \\bar { s } ^ { \\prime } \\mid \\bar { s } , \\bar { a } ) )" + }, + { + "category_id": 13, + "poly": [ + 982, + 1814, + 1218, + 1814, + 1218, + 1850, + 982, + 1850 + ], + "score": 0.87, + "latex": "( 1 / ( 1 - \\gamma ) , \\log { 1 / \\epsilon _ { 1 } } )" + }, + { + "category_id": 13, + "poly": [ + 644, + 650, + 679, + 650, + 679, + 681, + 644, + 681 + ], + "score": 0.87, + "latex": "\\bar { \\mathcal { M } }" + }, + { + "category_id": 13, + "poly": [ + 833, + 730, + 869, + 730, + 869, + 761, + 833, + 761 + ], + "score": 0.86, + "latex": "\\bar { \\mathcal { M } }" + }, + { + "category_id": 13, + "poly": [ + 697, + 1811, + 723, + 1811, + 723, + 1843, + 697, + 1843 + ], + "score": 0.85, + "latex": "\\tilde { \\mathcal { O } }" + }, + { + "category_id": 13, + "poly": [ + 616, + 1411, + 641, + 1411, + 641, + 1435, + 616, + 1435 + ], + "score": 0.85, + "latex": "\\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 425, + 730, + 460, + 730, + 460, + 760, + 425, + 760 + ], + "score": 0.85, + "latex": "\\bar { \\mathcal { M } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1247, + 324, + 1247, + 324, + 1273, + 298, + 1273 + ], + "score": 0.85, + "latex": "\\epsilon _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 809, + 844, + 841, + 844, + 841, + 876, + 809, + 876 + ], + "score": 0.85, + "latex": "\\bar { M }" + }, + { + "category_id": 13, + "poly": [ + 1345, + 853, + 1373, + 853, + 1373, + 878, + 1345, + 878 + ], + "score": 0.85, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1244, + 1379, + 1270, + 1379, + 1270, + 1404, + 1244, + 1404 + ], + "score": 0.84, + "latex": "\\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1341, + 597, + 1369, + 597, + 1369, + 622, + 1341, + 622 + ], + "score": 0.84, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 397, + 560, + 420, + 560, + 420, + 588, + 397, + 588 + ], + "score": 0.84, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 892, + 1457, + 910, + 1457, + 910, + 1483, + 892, + 1483 + ], + "score": 0.83, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 838, + 1007, + 866, + 1007, + 866, + 1033, + 838, + 1033 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1001, + 764, + 1037, + 764, + 1037, + 791, + 1001, + 791 + ], + "score": 0.83, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 617, + 528, + 636, + 528, + 636, + 555, + 617, + 555 + ], + "score": 0.83, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 1109, + 654, + 1145, + 654, + 1145, + 681, + 1109, + 681 + ], + "score": 0.82, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 427, + 761, + 462, + 761, + 462, + 790, + 427, + 790 + ], + "score": 0.82, + "latex": "\\bar { \\mathcal { M } }" + }, + { + "category_id": 13, + "poly": [ + 1132, + 1457, + 1155, + 1457, + 1155, + 1483, + 1132, + 1483 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1317, + 1968, + 1403, + 1968, + 1403, + 2005, + 1317, + 2005 + ], + "score": 0.8, + "latex": "\\begin{array} { r l } { \\bar { \\mathcal { M } } } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1217, + 323, + 1217, + 323, + 1242, + 298, + 1242 + ], + "score": 0.79, + "latex": "\\epsilon _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1373, + 327, + 1401, + 327, + 1401, + 354, + 1373, + 354 + ], + "score": 0.79, + "latex": "H" + }, + { + "category_id": 13, + "poly": [ + 372, + 766, + 393, + 766, + 393, + 790, + 372, + 790 + ], + "score": 0.79, + "latex": "\\bar { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 747, + 733, + 782, + 733, + 782, + 760, + 747, + 760 + ], + "score": 0.78, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 372, + 735, + 393, + 735, + 393, + 761, + 372, + 761 + ], + "score": 0.78, + "latex": "\\bar { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 1268, + 654, + 1286, + 654, + 1286, + 680, + 1268, + 680 + ], + "score": 0.78, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 477, + 1374, + 535, + 1374, + 535, + 1403, + 477, + 1403 + ], + "score": 0.77, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 1220, + 534, + 1238, + 534, + 1238, + 554, + 1220, + 554 + ], + "score": 0.77, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 766, + 655, + 781, + 655, + 781, + 680, + 766, + 680 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1349, + 654, + 1368, + 654, + 1368, + 680, + 1349, + 680 + ], + "score": 0.76, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 517, + 534, + 533, + 534, + 533, + 554, + 517, + 554 + ], + "score": 0.76, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 948, + 767, + 969, + 767, + 969, + 790, + 948, + 790 + ], + "score": 0.75, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 806, + 887, + 850, + 887, + 850, + 921, + 806, + 921 + ], + "score": 0.72, + "latex": "1 / 3" + }, + { + "category_id": 13, + "poly": [ + 477, + 1816, + 539, + 1816, + 539, + 1844, + 477, + 1844 + ], + "score": 0.71, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 396, + 483, + 575, + 483, + 575, + 520, + 396, + 520 + ], + "score": 0.71, + "latex": "\\gamma = ( 1 - 1 / H )" + }, + { + "category_id": 13, + "poly": [ + 1031, + 894, + 1045, + 894, + 1045, + 915, + 1031, + 915 + ], + "score": 0.71, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1294, + 739, + 1308, + 739, + 1308, + 760, + 1294, + 760 + ], + "score": 0.68, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 371, + 965, + 384, + 965, + 384, + 986, + 371, + 986 + ], + "score": 0.65, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1271, + 922, + 1316, + 922, + 1316, + 957, + 1271, + 957 + ], + "score": 0.44, + "latex": "2 / 3" + }, + { + "category_id": 14, + "poly": [ + 392, + 440, + 626, + 440, + 626, + 477, + 392, + 477 + ], + "score": 0.42, + "latex": "\\bar { S } = S \\times H , \\bar { A } = A ;" + }, + { + "category_id": 14, + "poly": [ + 366, + 438, + 630, + 438, + 630, + 522, + 366, + 522 + ], + "score": 0.37, + "latex": "\\begin{array} { l } { { \\bullet \\ { \\bar { S } } = S \\times H , { \\bar { A } } = A ; } } \\\\ { { \\bullet \\ \\gamma = ( 1 - 1 / H ) ; } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 395, + 440, + 625, + 440, + 625, + 477, + 395, + 477 + ], + "score": 0.3, + "latex": "\\bar { S } = S \\times H , \\bar { A } = A ;" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1915.0, + 887.0, + 1915.0, + 887.0, + 1950.0, + 297.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 224.0, + 827.0, + 224.0, + 827.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 649.0, + 340.0, + 649.0, + 340.0, + 687.0, + 294.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 649.0, + 643.0, + 649.0, + 643.0, + 687.0, + 371.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 649.0, + 765.0, + 649.0, + 765.0, + 687.0, + 680.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 649.0, + 830.0, + 649.0, + 830.0, + 687.0, + 782.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 649.0, + 1108.0, + 649.0, + 1108.0, + 687.0, + 870.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 649.0, + 1267.0, + 649.0, + 1267.0, + 687.0, + 1146.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 649.0, + 1348.0, + 649.0, + 1348.0, + 687.0, + 1287.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 649.0, + 1407.0, + 649.0, + 1407.0, + 687.0, + 1369.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 682.0, + 438.0, + 682.0, + 438.0, + 737.0, + 290.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 682.0, + 1173.0, + 682.0, + 1173.0, + 737.0, + 737.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 682.0, + 1409.0, + 682.0, + 1409.0, + 737.0, + 1362.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 730.0, + 371.0, + 730.0, + 371.0, + 767.0, + 294.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 730.0, + 424.0, + 730.0, + 424.0, + 767.0, + 394.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 730.0, + 746.0, + 730.0, + 746.0, + 767.0, + 461.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 730.0, + 832.0, + 730.0, + 832.0, + 767.0, + 783.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 730.0, + 1293.0, + 730.0, + 1293.0, + 767.0, + 870.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1309.0, + 730.0, + 1406.0, + 730.0, + 1406.0, + 767.0, + 1309.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 759.0, + 371.0, + 759.0, + 371.0, + 798.0, + 294.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 759.0, + 426.0, + 759.0, + 426.0, + 798.0, + 394.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 759.0, + 691.0, + 759.0, + 691.0, + 798.0, + 463.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 759.0, + 947.0, + 759.0, + 947.0, + 798.0, + 777.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 759.0, + 1000.0, + 759.0, + 1000.0, + 798.0, + 970.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 759.0, + 1407.0, + 759.0, + 1407.0, + 798.0, + 1038.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 400.0, + 788.0, + 400.0, + 829.0, + 294.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 788.0, + 871.0, + 788.0, + 871.0, + 829.0, + 715.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 294.0, + 1404.0, + 294.0, + 1404.0, + 328.0, + 292.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 324.0, + 867.0, + 324.0, + 867.0, + 361.0, + 295.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 324.0, + 1372.0, + 324.0, + 1372.0, + 361.0, + 1294.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 355.0, + 1011.0, + 355.0, + 1011.0, + 391.0, + 292.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 355.0, + 1406.0, + 355.0, + 1406.0, + 391.0, + 1403.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 384.0, + 450.0, + 384.0, + 450.0, + 422.0, + 291.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1173.0, + 1404.0, + 1173.0, + 1404.0, + 1215.0, + 292.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 297.0, + 1209.0, + 297.0, + 1247.0, + 294.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1209.0, + 974.0, + 1209.0, + 974.0, + 1247.0, + 324.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1209.0, + 1406.0, + 1209.0, + 1406.0, + 1247.0, + 1218.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 297.0, + 1241.0, + 297.0, + 1278.0, + 294.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1241.0, + 684.0, + 1241.0, + 684.0, + 1278.0, + 325.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 833.0, + 392.0, + 833.0, + 392.0, + 931.0, + 289.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 833.0, + 805.0, + 833.0, + 805.0, + 931.0, + 590.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 833.0, + 887.0, + 833.0, + 887.0, + 931.0, + 851.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 833.0, + 1030.0, + 833.0, + 1030.0, + 931.0, + 1013.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 833.0, + 1344.0, + 833.0, + 1344.0, + 931.0, + 1046.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 833.0, + 1410.0, + 833.0, + 1410.0, + 931.0, + 1374.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 914.0, + 708.0, + 914.0, + 708.0, + 963.0, + 292.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 786.0, + 914.0, + 829.0, + 914.0, + 829.0, + 963.0, + 786.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 914.0, + 1270.0, + 914.0, + 1270.0, + 963.0, + 989.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 914.0, + 1408.0, + 914.0, + 1408.0, + 963.0, + 1317.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 954.0, + 370.0, + 954.0, + 370.0, + 998.0, + 292.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 954.0, + 1007.0, + 954.0, + 1007.0, + 998.0, + 385.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 954.0, + 1240.0, + 954.0, + 1240.0, + 998.0, + 1103.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 954.0, + 1409.0, + 954.0, + 1409.0, + 998.0, + 1333.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 436.0, + 636.0, + 436.0, + 636.0, + 479.0, + 631.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 523.0, + 516.0, + 523.0, + 516.0, + 560.0, + 366.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 523.0, + 616.0, + 523.0, + 616.0, + 560.0, + 534.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 523.0, + 683.0, + 523.0, + 683.0, + 560.0, + 637.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 523.0, + 1219.0, + 523.0, + 1219.0, + 560.0, + 732.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 523.0, + 1405.0, + 523.0, + 1405.0, + 560.0, + 1239.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 550.0, + 396.0, + 550.0, + 396.0, + 601.0, + 390.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 550.0, + 510.0, + 550.0, + 510.0, + 601.0, + 421.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 550.0, + 897.0, + 550.0, + 897.0, + 601.0, + 845.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 550.0, + 1409.0, + 550.0, + 1409.0, + 601.0, + 1297.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 587.0, + 396.0, + 587.0, + 396.0, + 629.0, + 392.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 587.0, + 527.0, + 587.0, + 527.0, + 629.0, + 479.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 587.0, + 737.0, + 587.0, + 737.0, + 629.0, + 688.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 587.0, + 1340.0, + 587.0, + 1340.0, + 629.0, + 1069.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 587.0, + 1383.0, + 587.0, + 1383.0, + 629.0, + 1370.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1808.0, + 476.0, + 1808.0, + 476.0, + 1852.0, + 293.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1808.0, + 696.0, + 1808.0, + 696.0, + 1852.0, + 540.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1808.0, + 981.0, + 1808.0, + 981.0, + 1852.0, + 724.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 1808.0, + 1405.0, + 1808.0, + 1405.0, + 1852.0, + 1219.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1843.0, + 470.0, + 1843.0, + 470.0, + 1883.0, + 294.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1843.0, + 742.0, + 1843.0, + 742.0, + 1883.0, + 731.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1449.0, + 442.0, + 1449.0, + 442.0, + 1493.0, + 294.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1449.0, + 891.0, + 1449.0, + 891.0, + 1493.0, + 718.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1449.0, + 1131.0, + 1449.0, + 1131.0, + 1493.0, + 911.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1449.0, + 1202.0, + 1449.0, + 1202.0, + 1493.0, + 1156.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 1449.0, + 1405.0, + 1449.0, + 1405.0, + 1493.0, + 1356.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1485.0, + 297.0, + 1485.0, + 297.0, + 1539.0, + 292.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1485.0, + 582.0, + 1485.0, + 582.0, + 1539.0, + 527.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1485.0, + 1001.0, + 1485.0, + 1001.0, + 1539.0, + 816.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1371.0, + 476.0, + 1371.0, + 476.0, + 1411.0, + 293.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 1371.0, + 1243.0, + 1371.0, + 1243.0, + 1411.0, + 536.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 1371.0, + 1405.0, + 1371.0, + 1405.0, + 1411.0, + 1271.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1401.0, + 615.0, + 1401.0, + 615.0, + 1443.0, + 295.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1401.0, + 654.0, + 1401.0, + 654.0, + 1443.0, + 642.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1965.0, + 764.0, + 1965.0, + 764.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 1965.0, + 1316.0, + 1965.0, + 1316.0, + 2010.0, + 1264.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1965.0, + 1408.0, + 1965.0, + 1408.0, + 2010.0, + 1404.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1998.0, + 295.0, + 1998.0, + 295.0, + 2039.0, + 292.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1998.0, + 775.0, + 1998.0, + 775.0, + 2039.0, + 617.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 997.0, + 837.0, + 997.0, + 837.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 997.0, + 980.0, + 997.0, + 980.0, + 1045.0, + 867.0, + 1045.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1939, + 1404, + 1939, + 1404, + 2037, + 298, + 2037 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 302, + 433, + 1409, + 433, + 1409, + 530, + 302, + 530 + ], + "score": 0.965 + }, + { + "category_id": 3, + "poly": [ + 677, + 1555, + 1022, + 1555, + 1022, + 1825, + 677, + 1825 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 298, + 694, + 1405, + 694, + 1405, + 789, + 298, + 789 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 295, + 542, + 1482, + 542, + 1482, + 682, + 295, + 682 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 687, + 846, + 1011, + 846, + 1011, + 892, + 687, + 892 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 521, + 1126, + 1178, + 1126, + 1178, + 1196, + 521, + 1196 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 297, + 1210, + 852, + 1210, + 852, + 1248, + 297, + 1248 + ], + "score": 0.936 + }, + { + "category_id": 8, + "poly": [ + 494, + 939, + 1205, + 939, + 1205, + 984, + 494, + 984 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 296, + 1397, + 1407, + 1397, + 1407, + 1516, + 296, + 1516 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 366, + 226, + 1412, + 226, + 1412, + 413, + 366, + 413 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 297, + 799, + 897, + 799, + 897, + 833, + 297, + 833 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 297, + 900, + 457, + 900, + 457, + 931, + 297, + 931 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 817, + 74, + 817, + 106, + 298, + 106 + ], + "score": 0.914 + }, + { + "category_id": 1, + "poly": [ + 296, + 994, + 368, + 994, + 368, + 1024, + 296, + 1024 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 298, + 1085, + 509, + 1085, + 509, + 1117, + 298, + 1117 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 835, + 2087, + 863, + 2087, + 863, + 2113, + 835, + 2113 + ], + "score": 0.867 + }, + { + "category_id": 4, + "poly": [ + 292, + 1854, + 1406, + 1854, + 1406, + 1918, + 292, + 1918 + ], + "score": 0.833 + }, + { + "category_id": 0, + "poly": [ + 297, + 1286, + 1035, + 1286, + 1035, + 1326, + 297, + 1326 + ], + "score": 0.827 + }, + { + "category_id": 1, + "poly": [ + 302, + 1356, + 1268, + 1356, + 1268, + 1392, + 302, + 1392 + ], + "score": 0.773 + }, + { + "category_id": 8, + "poly": [ + 291, + 1032, + 1401, + 1032, + 1401, + 1077, + 291, + 1077 + ], + "score": 0.74 + }, + { + "category_id": 1, + "poly": [ + 291, + 1032, + 1401, + 1032, + 1401, + 1077, + 291, + 1077 + ], + "score": 0.135 + }, + { + "category_id": 1, + "poly": [ + 292, + 1854, + 1406, + 1854, + 1406, + 1918, + 292, + 1918 + ], + "score": 0.117 + }, + { + "category_id": 13, + "poly": [ + 1093, + 1428, + 1227, + 1428, + 1227, + 1483, + 1093, + 1483 + ], + "score": 0.94, + "latex": "\\Omega \\left( \\frac { \\epsilon ^ { - 3 } } { \\ln ( 1 / \\delta ) } \\right)" + }, + { + "category_id": 13, + "poly": [ + 578, + 739, + 875, + 739, + 875, + 790, + 578, + 790 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\bar { V } ^ { \\ast } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\ast } ( s _ { 1 } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 823, + 694, + 1123, + 694, + 1123, + 742, + 823, + 742 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\frac { \\gamma ^ { H } } { 1 - \\gamma ^ { H } } V ^ { \\pi } ( s _ { 1 } ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 521, + 1123, + 1175, + 1123, + 1175, + 1199, + 521, + 1199 + ], + "score": 0.93, + "latex": "{ V ^ { \\pi } } ^ { t } ( s _ { 1 } ) \\ge \\frac { 1 - \\gamma ^ { H } } { \\gamma ^ { H } } \\bar { V } ^ { * } ( \\bar { s } _ { s 1 , 1 } ) - \\epsilon / \\gamma ^ { H } = V ^ { * } ( s _ { 1 } ) - \\epsilon / \\gamma ^ { H } ," + }, + { + "category_id": 13, + "poly": [ + 1102, + 2004, + 1229, + 2004, + 1229, + 2036, + 1102, + 2036 + ], + "score": 0.92, + "latex": "R ( c , \\cdot ) = 0" + }, + { + "category_id": 13, + "poly": [ + 298, + 1973, + 391, + 1973, + 391, + 2005, + 298, + 2005 + ], + "score": 0.92, + "latex": "\\{ a , b , c \\}" + }, + { + "category_id": 13, + "poly": [ + 488, + 1939, + 632, + 1939, + 632, + 1976, + 488, + 1976 + ], + "score": 0.92, + "latex": "0 < \\epsilon < \\frac { 1 } { 1 0 }" + }, + { + "category_id": 13, + "poly": [ + 904, + 436, + 1167, + 436, + 1167, + 470, + 904, + 470 + ], + "score": 0.92, + "latex": "\\{ ( s _ { 1 } , a _ { 1 } ) , ( s _ { 2 } , a _ { 2 } ) , \\cdot \\cdot \\cdot \\}" + }, + { + "category_id": 13, + "poly": [ + 436, + 379, + 595, + 379, + 595, + 410, + 436, + 410 + ], + "score": 0.92, + "latex": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = 0" + }, + { + "category_id": 13, + "poly": [ + 859, + 345, + 1204, + 345, + 1204, + 380, + 859, + 380 + ], + "score": 0.91, + "latex": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , h + 1 } \\mid \\bar { s } _ { s , h } , a ) = p _ { h } ( s , h )" + }, + { + "category_id": 13, + "poly": [ + 590, + 1210, + 674, + 1210, + 674, + 1247, + 590, + 1247 + ], + "score": 0.91, + "latex": "( \\epsilon / \\gamma ^ { H } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1478, + 463, + 1478, + 463, + 1516, + 297, + 1516 + ], + "score": 0.91, + "latex": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }" + }, + { + "category_id": 14, + "poly": [ + 685, + 844, + 1012, + 844, + 1012, + 890, + 685, + 890 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\bar { V } ^ { \\bar { \\pi } ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 595, + 1973, + 727, + 1973, + 727, + 2005, + 595, + 2005 + ], + "score": 0.91, + "latex": "\\overset { \\vartriangle } { \\mathcal { A } } = \\{ \\ v { x } , \\ v { y } \\}" + }, + { + "category_id": 13, + "poly": [ + 350, + 2003, + 508, + 2003, + 508, + 2036, + 350, + 2036 + ], + "score": 0.91, + "latex": "P ( b | a , x ) = 1" + }, + { + "category_id": 13, + "poly": [ + 1226, + 472, + 1362, + 472, + 1362, + 500, + 1226, + 500 + ], + "score": 0.9, + "latex": "s _ { t H + 1 } = s _ { 1 }" + }, + { + "category_id": 14, + "poly": [ + 310, + 540, + 1428, + 540, + 1428, + 683, + 310, + 683 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { \\bar { \\gamma } ^ { \\pi ^ { t } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\mathbb { E } \\left[ \\bar { r } ( \\bar { s } _ { s _ { 1 } , 1 } , \\bar { a } _ { 1 } ) + \\gamma \\bar { r } ( \\bar { s } _ { s _ { 2 } , 2 } , \\bar { a } _ { 2 } ) + \\cdot \\cdot \\cdot + \\gamma ^ { H - 1 } \\bar { r } ( \\bar { s } _ { s _ { H - 1 } , H - 1 } , \\bar { a } _ { H - 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H - 1 } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } \\mathbb { E } \\left[ r _ { 1 } ( s _ { 1 } , a _ { 1 } ) + r _ { 2 } ( s _ { 2 } , a _ { 2 } ) + \\cdot \\cdot \\cdot + r _ { H - 1 } ( s _ { H - 1 } , a _ { H - 1 } ) + \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { H + 1 } , 1 } ) \\right] } \\\\ & { \\phantom { \\frac { 1 } { 1 } } = \\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi + H } ( \\bar { s } _ { s _ { 1 } , 1 } ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 474, + 344, + 805, + 344, + 805, + 378, + 474, + 378 + ], + "score": 0.9, + "latex": "\\bar { r } ( \\bar { s } _ { s , h } , a ) = \\gamma ^ { H - h + 1 } r _ { h } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1314, + 346, + 1397, + 346, + 1397, + 375, + 1314, + 375 + ], + "score": 0.89, + "latex": "h = H" + }, + { + "category_id": 13, + "poly": [ + 333, + 499, + 395, + 499, + 395, + 528, + 333, + 528 + ], + "score": 0.89, + "latex": "t \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 646, + 378, + 978, + 378, + 978, + 410, + 646, + 410 + ], + "score": 0.89, + "latex": "\\bar { p } ( \\bar { s } _ { s ^ { \\prime } , 1 } \\mid \\bar { s } _ { s , h } , a ) = I [ s ^ { \\prime } = s _ { 1 } ]" + }, + { + "category_id": 13, + "poly": [ + 674, + 316, + 721, + 316, + 721, + 345, + 674, + 345 + ], + "score": 0.88, + "latex": "\\bar { s } _ { s , h }" + }, + { + "category_id": 13, + "poly": [ + 519, + 2004, + 804, + 2004, + 804, + 2036, + 519, + 2036 + ], + "score": 0.88, + "latex": "P ( a | b , \\cdot ) = P ( a | c , \\cdot ) = 1" + }, + { + "category_id": 13, + "poly": [ + 497, + 1213, + 527, + 1213, + 527, + 1241, + 497, + 1241 + ], + "score": 0.88, + "latex": "\\pi ^ { t }" + }, + { + "category_id": 14, + "poly": [ + 494, + 936, + 1200, + 936, + 1200, + 983, + 494, + 983 + ], + "score": 0.88, + "latex": "\\gamma ^ { H } V ^ { \\pi ^ { t } } ( s _ { 1 } ) + \\gamma ^ { H } \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) = \\bar { V } ^ { \\bar { \\pi } } ( \\bar { s } _ { s _ { 1 } , 1 } ) \\geq \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon ," + }, + { + "category_id": 13, + "poly": [ + 807, + 497, + 836, + 497, + 836, + 525, + 807, + 525 + ], + "score": 0.87, + "latex": "\\bar { \\pi } ^ { t }" + }, + { + "category_id": 13, + "poly": [ + 472, + 436, + 791, + 436, + 791, + 470, + 472, + 470 + ], + "score": 0.87, + "latex": "\\left\\{ { \\left( { \\bar { s } } _ { s _ { 1 } , 1 } , { \\bar { a } } _ { 1 } \\right) } , { \\left( { \\bar { s } } _ { s _ { 2 } , 2 } , { \\bar { a } } _ { 2 } \\right) } , \\cdot \\cdot \\cdot \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 988, + 474, + 1016, + 474, + 1016, + 498, + 988, + 498 + ], + "score": 0.87, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 870, + 497, + 899, + 497, + 899, + 525, + 870, + 525 + ], + "score": 0.87, + "latex": "\\pi ^ { t }" + }, + { + "category_id": 14, + "poly": [ + 310, + 1032, + 1402, + 1032, + 1402, + 1076, + 310, + 1076 + ], + "score": 0.86, + "latex": "^ { H } V ^ { \\pi ^ { \\varepsilon } } ( s _ { 1 } ) \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) + \\gamma ^ { H } ( \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\bar { V } ^ { \\pi _ { t + H } } ( \\bar { s } _ { s _ { 1 } , 1 } ) ) - \\epsilon \\geq ( 1 - \\gamma ^ { H } ) \\bar { V } ^ { * } ( \\bar { s } _ { s _ { 1 } , 1 } ) - \\epsilon ." + }, + { + "category_id": 13, + "poly": [ + 1346, + 1944, + 1404, + 1944, + 1404, + 1972, + 1346, + 1972 + ], + "score": 0.86, + "latex": "{ \\boldsymbol { s } } =" + }, + { + "category_id": 13, + "poly": [ + 1374, + 314, + 1396, + 314, + 1396, + 340, + 1374, + 340 + ], + "score": 0.85, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 822, + 434, + 858, + 434, + 858, + 466, + 822, + 466 + ], + "score": 0.85, + "latex": "\\bar { \\mathcal { M } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 751, + 329, + 751, + 329, + 777, + 298, + 777 + ], + "score": 0.85, + "latex": "\\bar { \\pi } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1250, + 383, + 1277, + 383, + 1277, + 407, + 1250, + 407 + ], + "score": 0.84, + "latex": "s _ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 596, + 470, + 630, + 470, + 630, + 494, + 596, + 494 + ], + "score": 0.84, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 439, + 469, + 474, + 469, + 474, + 494, + 439, + 494 + ], + "score": 0.83, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 395, + 226, + 624, + 226, + 624, + 262, + 395, + 262 + ], + "score": 0.81, + "latex": "\\bar { S } = S \\times H , \\bar { A } = A" + }, + { + "category_id": 13, + "poly": [ + 604, + 707, + 623, + 707, + 623, + 730, + 604, + 730 + ], + "score": 0.8, + "latex": "\\bar { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 457, + 804, + 476, + 804, + 476, + 828, + 457, + 828 + ], + "score": 0.79, + "latex": "\\bar { \\pi }" + }, + { + "category_id": 13, + "poly": [ + 984, + 1399, + 1006, + 1399, + 1006, + 1426, + 984, + 1426 + ], + "score": 0.78, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 610, + 314, + 629, + 314, + 629, + 341, + 610, + 341 + ], + "score": 0.78, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 513, + 320, + 529, + 320, + 529, + 340, + 513, + 340 + ], + "score": 0.76, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1194, + 320, + 1212, + 320, + 1212, + 340, + 1194, + 340 + ], + "score": 0.75, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 752, + 804, + 766, + 804, + 766, + 828, + 752, + 828 + ], + "score": 0.74, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 396, + 270, + 575, + 270, + 575, + 305, + 396, + 305 + ], + "score": 0.73, + "latex": "\\gamma = ( 1 - 1 / H )" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1973, + 1250, + 1973, + 1250, + 2005, + 1028, + 2005 + ], + "score": 0.73, + "latex": "P ( b | a , y ) = 1 - 1 0 \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 506, + 807, + 520, + 807, + 520, + 828, + 506, + 828 + ], + "score": 0.73, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1066, + 1400, + 1089, + 1400, + 1089, + 1426, + 1066, + 1426 + ], + "score": 0.71, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 473, + 1440, + 496, + 1440, + 496, + 1470, + 473, + 1470 + ], + "score": 0.71, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 296, + 1437, + 360, + 1437, + 360, + 1469, + 296, + 1469 + ], + "score": 0.71, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 1265, + 1973, + 1405, + 1973, + 1405, + 2006, + 1265, + 2006 + ], + "score": 0.61, + "latex": "P ( c | a , y ) =" + }, + { + "category_id": 13, + "poly": [ + 971, + 1598, + 985, + 1598, + 985, + 1621, + 971, + 1621 + ], + "score": 0.44, + "latex": "b" + }, + { + "category_id": 13, + "poly": [ + 788, + 1563, + 837, + 1563, + 837, + 1586, + 788, + 1586 + ], + "score": 0.27, + "latex": "x , y" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1553.0, + 787.0, + 1553.0, + 787.0, + 1591.0, + 780.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1553.0, + 847.0, + 1553.0, + 847.0, + 1591.0, + 838.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1592.0, + 970.0, + 1592.0, + 970.0, + 1624.0, + 966.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1592.0, + 994.0, + 1592.0, + 994.0, + 1624.0, + 986.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1627.0, + 826.0, + 1627.0, + 826.0, + 1656.0, + 795.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1645.0, + 970.0, + 1645.0, + 970.0, + 1670.0, + 847.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1677.0, + 736.0, + 1677.0, + 736.0, + 1709.0, + 702.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1684.0, + 842.0, + 1684.0, + 842.0, + 1694.0, + 835.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1689.0, + 825.0, + 1689.0, + 825.0, + 1724.0, + 794.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1756.0, + 942.0, + 1756.0, + 942.0, + 1775.0, + 891.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1756.0, + 995.0, + 1756.0, + 995.0, + 1784.0, + 966.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1757.0, + 895.0, + 1757.0, + 895.0, + 1765.0, + 887.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1789.0, + 849.0, + 1789.0, + 849.0, + 1833.0, + 784.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1740.5, + 876.0, + 1740.5, + 876.0, + 1745.5, + 871.0, + 1745.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 817.0, + 72.0, + 817.0, + 109.0, + 295.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1854.0, + 1408.0, + 1854.0, + 1408.0, + 1890.0, + 295.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1886.0, + 570.0, + 1886.0, + 570.0, + 1922.0, + 294.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1281.0, + 1038.0, + 1281.0, + 1038.0, + 1335.0, + 292.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1936.0, + 487.0, + 1936.0, + 487.0, + 1982.0, + 292.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1936.0, + 1345.0, + 1936.0, + 1345.0, + 1982.0, + 633.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1405.0, + 1936.0, + 1408.0, + 1936.0, + 1408.0, + 1982.0, + 1405.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1971.0, + 594.0, + 1971.0, + 594.0, + 2009.0, + 392.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1971.0, + 1027.0, + 1971.0, + 1027.0, + 2009.0, + 728.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1971.0, + 1264.0, + 1971.0, + 1264.0, + 2009.0, + 1251.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 349.0, + 2000.0, + 349.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 2000.0, + 518.0, + 2000.0, + 518.0, + 2038.0, + 509.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 2000.0, + 1101.0, + 2000.0, + 1101.0, + 2038.0, + 805.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 2000.0, + 1240.0, + 2000.0, + 1240.0, + 2038.0, + 1230.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 432.0, + 471.0, + 432.0, + 471.0, + 474.0, + 293.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 432.0, + 821.0, + 432.0, + 821.0, + 474.0, + 792.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 432.0, + 903.0, + 432.0, + 903.0, + 474.0, + 859.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 432.0, + 1406.0, + 432.0, + 1406.0, + 474.0, + 1168.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 464.0, + 438.0, + 464.0, + 438.0, + 502.0, + 295.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 464.0, + 595.0, + 464.0, + 595.0, + 502.0, + 475.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 464.0, + 987.0, + 464.0, + 987.0, + 502.0, + 631.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 464.0, + 1225.0, + 464.0, + 1225.0, + 502.0, + 1017.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 464.0, + 1405.0, + 464.0, + 1405.0, + 502.0, + 1363.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 498.0, + 332.0, + 498.0, + 332.0, + 532.0, + 296.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 498.0, + 806.0, + 498.0, + 806.0, + 532.0, + 396.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 498.0, + 869.0, + 498.0, + 869.0, + 532.0, + 837.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 498.0, + 1336.0, + 498.0, + 1336.0, + 532.0, + 900.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 682.0, + 603.0, + 682.0, + 603.0, + 751.0, + 280.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 682.0, + 822.0, + 682.0, + 822.0, + 751.0, + 624.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 682.0, + 1417.0, + 682.0, + 1417.0, + 751.0, + 1124.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 740.0, + 297.0, + 740.0, + 297.0, + 796.0, + 287.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 740.0, + 577.0, + 740.0, + 577.0, + 796.0, + 330.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 740.0, + 893.0, + 740.0, + 893.0, + 796.0, + 876.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 737.0, + 783.0, + 737.0, + 783.0, + 769.0, + 731.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1203.0, + 496.0, + 1203.0, + 496.0, + 1254.0, + 292.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1203.0, + 589.0, + 1203.0, + 589.0, + 1254.0, + 528.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1203.0, + 853.0, + 1203.0, + 853.0, + 1254.0, + 675.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1392.0, + 983.0, + 1392.0, + 983.0, + 1435.0, + 294.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1392.0, + 1065.0, + 1392.0, + 1065.0, + 1435.0, + 1007.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1392.0, + 1404.0, + 1392.0, + 1404.0, + 1435.0, + 1090.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1432.0, + 295.0, + 1432.0, + 295.0, + 1478.0, + 291.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1432.0, + 472.0, + 1432.0, + 472.0, + 1478.0, + 361.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1432.0, + 1092.0, + 1432.0, + 1092.0, + 1478.0, + 497.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 1436.0, + 1405.0, + 1436.0, + 1405.0, + 1473.0, + 1230.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1477.0, + 296.0, + 1477.0, + 296.0, + 1512.0, + 292.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1477.0, + 472.0, + 1477.0, + 472.0, + 1512.0, + 464.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1426.5, + 1201.0, + 1426.5, + 1201.0, + 1456.5, + 1145.0, + 1456.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 222.0, + 394.0, + 222.0, + 394.0, + 265.0, + 365.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 222.0, + 636.0, + 222.0, + 636.0, + 265.0, + 625.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 270.0, + 395.0, + 270.0, + 395.0, + 305.0, + 365.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 270.0, + 588.0, + 270.0, + 588.0, + 305.0, + 576.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 309.0, + 512.0, + 309.0, + 512.0, + 348.0, + 366.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 309.0, + 609.0, + 309.0, + 609.0, + 348.0, + 530.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 309.0, + 673.0, + 309.0, + 673.0, + 348.0, + 630.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 309.0, + 1193.0, + 309.0, + 1193.0, + 348.0, + 722.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 309.0, + 1373.0, + 309.0, + 1373.0, + 348.0, + 1213.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 309.0, + 1409.0, + 309.0, + 1409.0, + 348.0, + 1397.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 339.0, + 473.0, + 339.0, + 473.0, + 385.0, + 390.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 339.0, + 858.0, + 339.0, + 858.0, + 385.0, + 806.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 339.0, + 1313.0, + 339.0, + 1313.0, + 385.0, + 1205.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 339.0, + 1415.0, + 339.0, + 1415.0, + 385.0, + 1398.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 373.0, + 435.0, + 373.0, + 435.0, + 414.0, + 391.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 373.0, + 645.0, + 373.0, + 645.0, + 414.0, + 596.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 373.0, + 1249.0, + 373.0, + 1249.0, + 414.0, + 979.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 373.0, + 1291.0, + 373.0, + 1291.0, + 414.0, + 1278.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 797.0, + 456.0, + 797.0, + 456.0, + 837.0, + 296.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 797.0, + 505.0, + 797.0, + 505.0, + 837.0, + 477.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 797.0, + 751.0, + 797.0, + 751.0, + 837.0, + 521.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 797.0, + 896.0, + 797.0, + 896.0, + 837.0, + 767.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 896.0, + 461.0, + 896.0, + 461.0, + 934.0, + 293.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 371.0, + 991.0, + 371.0, + 1030.0, + 294.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1081.0, + 511.0, + 1081.0, + 511.0, + 1120.0, + 294.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1349.0, + 1267.0, + 1349.0, + 1267.0, + 1401.0, + 292.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1026.0, + 309.0, + 1026.0, + 309.0, + 1083.0, + 292.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1026.0, + 1408.0, + 1026.0, + 1408.0, + 1083.0, + 1403.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1854.0, + 1408.0, + 1854.0, + 1408.0, + 1890.0, + 295.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1886.0, + 570.0, + 1886.0, + 570.0, + 1922.0, + 294.0, + 1922.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 226, + 1405, + 226, + 1405, + 470, + 297, + 470 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 296, + 603, + 1405, + 603, + 1405, + 763, + 296, + 763 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 485, + 1406, + 485, + 1406, + 593, + 298, + 593 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 715, + 778, + 983, + 778, + 983, + 854, + 715, + 854 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 295, + 867, + 1058, + 867, + 1058, + 907, + 295, + 907 + ], + "score": 0.929 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.843 + }, + { + "category_id": 2, + "poly": [ + 1374, + 871, + 1403, + 871, + 1403, + 900, + 1374, + 900 + ], + "score": 0.822 + }, + { + "category_id": 13, + "poly": [ + 627, + 335, + 714, + 335, + 714, + 372, + 627, + 372 + ], + "score": 0.95, + "latex": "\\hat { Q } ( a , y )" + }, + { + "category_id": 14, + "poly": [ + 714, + 776, + 984, + 776, + 984, + 854, + 714, + 854 + ], + "score": 0.94, + "latex": "t _ { 0 } = \\Omega \\left( { \\frac { 1 } { \\epsilon ^ { 3 } \\left( \\ln { 1 / \\delta } \\right) } } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 1243, + 483, + 1331, + 483, + 1331, + 522, + 1243, + 522 + ], + "score": 0.94, + "latex": "\\hat { Q } ( a , y )" + }, + { + "category_id": 13, + "poly": [ + 610, + 696, + 697, + 696, + 697, + 733, + 610, + 733 + ], + "score": 0.94, + "latex": "\\hat { Q } ( a , y )" + }, + { + "category_id": 13, + "poly": [ + 1260, + 300, + 1349, + 300, + 1349, + 338, + 1260, + 338 + ], + "score": 0.93, + "latex": "\\hat { Q } ( a , x )" + }, + { + "category_id": 13, + "poly": [ + 663, + 430, + 749, + 430, + 749, + 469, + 663, + 469 + ], + "score": 0.93, + "latex": "\\hat { Q } ( a , y )" + }, + { + "category_id": 13, + "poly": [ + 752, + 866, + 1051, + 866, + 1051, + 908, + 752, + 908 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { C = 3 \\ln \\frac { 1 } { \\delta } + 1 < \\frac { 4 } { \\epsilon ^ { 2 } } < m } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 803, + 555, + 942, + 555, + 942, + 594, + 803, + 594 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { m = \\Omega \\left( \\frac { 1 } { \\epsilon ^ { 2 } } \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 722, + 483, + 796, + 483, + 796, + 522, + 722, + 522 + ], + "score": 0.93, + "latex": "\\hat { Q } ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 990, + 487, + 1104, + 487, + 1104, + 522, + 990, + 522 + ], + "score": 0.92, + "latex": "1 / ( 1 - \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 573, + 633, + 752, + 633, + 752, + 670, + 573, + 670 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { C ^ { ' } = 3 \\ln \\frac { 1 } { \\delta } + 1 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1327, + 227, + 1401, + 227, + 1401, + 267, + 1327, + 267 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\gamma > \\frac { 1 } { 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 369, + 868, + 535, + 868, + 535, + 904, + 369, + 904 + ], + "score": 0.92, + "latex": "\\ln ( 1 / \\delta ) < \\epsilon ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 371, + 645, + 371, + 645, + 404, + 298, + 404 + ], + "score": 0.91, + "latex": "Q ( a , x ) - Q ( a , y ) = 1 0 \\epsilon \\gamma > \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 298, + 335, + 384, + 335, + 384, + 372, + 298, + 372 + ], + "score": 0.91, + "latex": "\\hat { Q } ( a , y )" + }, + { + "category_id": 13, + "poly": [ + 297, + 634, + 483, + 634, + 483, + 668, + 297, + 668 + ], + "score": 0.91, + "latex": "t _ { 0 } = m / ( 1 0 \\epsilon C )" + }, + { + "category_id": 13, + "poly": [ + 1012, + 636, + 1101, + 636, + 1101, + 665, + 1012, + 665 + ], + "score": 0.91, + "latex": "C \\leq m" + }, + { + "category_id": 13, + "poly": [ + 1073, + 524, + 1194, + 524, + 1194, + 558, + 1073, + 558 + ], + "score": 0.9, + "latex": "1 / ( 1 - \\gamma ) )" + }, + { + "category_id": 13, + "poly": [ + 1243, + 636, + 1282, + 636, + 1282, + 666, + 1243, + 666 + ], + "score": 0.89, + "latex": "2 t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1053, + 701, + 1092, + 701, + 1092, + 731, + 1053, + 731 + ], + "score": 0.89, + "latex": "2 t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 563, + 520, + 641, + 520, + 641, + 558, + 563, + 558 + ], + "score": 0.88, + "latex": "\\hat { Q } ( c , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1035, + 667, + 1099, + 667, + 1099, + 695, + 1035, + 695 + ], + "score": 0.87, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 426, + 731, + 490, + 731, + 490, + 760, + 426, + 760 + ], + "score": 0.87, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 486, + 668, + 512, + 668, + 512, + 696, + 486, + 696 + ], + "score": 0.87, + "latex": "t _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1201, + 264, + 1227, + 264, + 1227, + 302, + 1201, + 302 + ], + "score": 0.86, + "latex": "\\hat { Q }" + }, + { + "category_id": 13, + "poly": [ + 716, + 270, + 743, + 270, + 743, + 302, + 716, + 302 + ], + "score": 0.84, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 1187, + 344, + 1207, + 344, + 1207, + 372, + 1187, + 372 + ], + "score": 0.79, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 925, + 376, + 944, + 376, + 944, + 402, + 925, + 402 + ], + "score": 0.76, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1032, + 310, + 1052, + 310, + 1052, + 337, + 1032, + 337 + ], + "score": 0.76, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 678, + 563, + 694, + 563, + 694, + 585, + 678, + 585 + ], + "score": 0.74, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 722, + 612, + 739, + 612, + 739, + 632, + 722, + 632 + ], + "score": 0.74, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 1169, + 671, + 1185, + 671, + 1185, + 693, + 1169, + 693 + ], + "score": 0.73, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 362, + 441, + 380, + 441, + 380, + 463, + 362, + 463 + ], + "score": 0.73, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 353, + 705, + 382, + 705, + 382, + 729, + 353, + 729 + ], + "score": 0.73, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 1139, + 237, + 1157, + 237, + 1157, + 259, + 1139, + 259 + ], + "score": 0.72, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 1300, + 346, + 1318, + 346, + 1318, + 367, + 1300, + 367 + ], + "score": 0.71, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 1033, + 376, + 1051, + 376, + 1051, + 398, + 1033, + 398 + ], + "score": 0.66, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 298, + 672, + 316, + 672, + 316, + 693, + 298, + 693 + ], + "score": 0.64, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 960, + 606, + 1003, + 606, + 1003, + 633, + 960, + 633 + ], + "score": 0.61, + "latex": "1 0 \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 298, + 408, + 312, + 408, + 312, + 428, + 298, + 428 + ], + "score": 0.59, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1323, + 611, + 1339, + 611, + 1339, + 632, + 1323, + 632 + ], + "score": 0.57, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 767, + 635, + 784, + 635, + 784, + 662, + 767, + 662 + ], + "score": 0.55, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 539, + 276, + 555, + 276, + 555, + 297, + 539, + 297 + ], + "score": 0.49, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2122.0, + 831.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 875.0, + 1403.0, + 875.0, + 1403.0, + 901.0, + 1378.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 226.0, + 1138.0, + 226.0, + 1138.0, + 270.0, + 292.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 226.0, + 1326.0, + 226.0, + 1326.0, + 270.0, + 1158.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 226.0, + 1406.0, + 226.0, + 1406.0, + 270.0, + 1402.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 269.0, + 538.0, + 269.0, + 538.0, + 303.0, + 294.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 269.0, + 715.0, + 269.0, + 715.0, + 303.0, + 556.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 269.0, + 1200.0, + 269.0, + 1200.0, + 303.0, + 744.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 269.0, + 1405.0, + 269.0, + 1405.0, + 303.0, + 1228.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 300.0, + 1031.0, + 300.0, + 1031.0, + 340.0, + 291.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 300.0, + 1259.0, + 300.0, + 1259.0, + 340.0, + 1053.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 300.0, + 1406.0, + 300.0, + 1406.0, + 340.0, + 1350.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 335.0, + 297.0, + 335.0, + 297.0, + 377.0, + 294.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 335.0, + 626.0, + 335.0, + 626.0, + 377.0, + 385.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 335.0, + 1186.0, + 335.0, + 1186.0, + 377.0, + 715.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 335.0, + 1299.0, + 335.0, + 1299.0, + 377.0, + 1208.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 335.0, + 1406.0, + 335.0, + 1406.0, + 377.0, + 1319.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 368.0, + 297.0, + 368.0, + 297.0, + 407.0, + 294.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 368.0, + 924.0, + 368.0, + 924.0, + 407.0, + 646.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 368.0, + 1032.0, + 368.0, + 1032.0, + 407.0, + 945.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 368.0, + 1407.0, + 368.0, + 1407.0, + 407.0, + 1052.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 402.0, + 1405.0, + 402.0, + 1405.0, + 435.0, + 313.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 433.0, + 361.0, + 433.0, + 361.0, + 472.0, + 292.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 433.0, + 662.0, + 433.0, + 662.0, + 472.0, + 381.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 433.0, + 762.0, + 433.0, + 762.0, + 472.0, + 750.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 604.0, + 721.0, + 604.0, + 721.0, + 638.0, + 295.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 604.0, + 959.0, + 604.0, + 959.0, + 638.0, + 740.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 604.0, + 1322.0, + 604.0, + 1322.0, + 638.0, + 1004.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 604.0, + 1405.0, + 604.0, + 1405.0, + 638.0, + 1340.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 632.0, + 296.0, + 632.0, + 296.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 632.0, + 572.0, + 632.0, + 572.0, + 674.0, + 484.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 632.0, + 766.0, + 632.0, + 766.0, + 674.0, + 753.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 632.0, + 1011.0, + 632.0, + 1011.0, + 674.0, + 785.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 632.0, + 1242.0, + 632.0, + 1242.0, + 674.0, + 1102.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 632.0, + 1408.0, + 632.0, + 1408.0, + 674.0, + 1283.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 665.0, + 297.0, + 665.0, + 297.0, + 699.0, + 294.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 665.0, + 485.0, + 665.0, + 485.0, + 699.0, + 317.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 665.0, + 1034.0, + 665.0, + 1034.0, + 699.0, + 513.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 665.0, + 1168.0, + 665.0, + 1168.0, + 699.0, + 1100.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 665.0, + 1405.0, + 665.0, + 1405.0, + 699.0, + 1186.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 698.0, + 352.0, + 698.0, + 352.0, + 736.0, + 295.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 698.0, + 609.0, + 698.0, + 609.0, + 736.0, + 383.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 698.0, + 1052.0, + 698.0, + 1052.0, + 736.0, + 698.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 698.0, + 1405.0, + 698.0, + 1405.0, + 736.0, + 1093.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 730.0, + 425.0, + 730.0, + 425.0, + 767.0, + 294.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 730.0, + 992.0, + 730.0, + 992.0, + 767.0, + 491.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 482.0, + 721.0, + 482.0, + 721.0, + 526.0, + 292.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 482.0, + 989.0, + 482.0, + 989.0, + 526.0, + 797.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 482.0, + 1242.0, + 482.0, + 1242.0, + 526.0, + 1105.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 482.0, + 1406.0, + 482.0, + 1406.0, + 526.0, + 1332.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 562.0, + 519.0, + 562.0, + 561.0, + 293.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 519.0, + 1072.0, + 519.0, + 1072.0, + 561.0, + 642.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 519.0, + 1404.0, + 519.0, + 1404.0, + 561.0, + 1195.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 555.0, + 677.0, + 555.0, + 677.0, + 595.0, + 295.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 555.0, + 802.0, + 555.0, + 802.0, + 595.0, + 695.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 555.0, + 1023.0, + 555.0, + 1023.0, + 595.0, + 943.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 857.0, + 368.0, + 857.0, + 368.0, + 913.0, + 292.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 857.0, + 751.0, + 857.0, + 751.0, + 913.0, + 536.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 857.0, + 1064.0, + 857.0, + 1064.0, + 913.0, + 1052.0, + 913.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/LAKplpLMbP8/LAKplpLMbP8.md b/parse/train/LAKplpLMbP8/LAKplpLMbP8.md new file mode 100644 index 0000000000000000000000000000000000000000..b6199876b438099afe82fd7180249a7cb6ab5375 --- /dev/null +++ b/parse/train/LAKplpLMbP8/LAKplpLMbP8.md @@ -0,0 +1,387 @@ +# Repulsive Deep Ensembles are Bayesian + +Francesco D’Angelo ETH Zürich Zürich, Switzerland dngfra@gmail.com + +Vincent Fortuin +ETH Zürich +Zürich, Switzerland +fortuin@inf.ethz.ch + +# Abstract + +Deep ensembles have recently gained popularity in the deep learning community for their conceptual simplicity and efficiency. However, maintaining functional diversity between ensemble members that are independently trained with gradient descent is challenging. This can lead to pathologies when adding more ensemble members, such as a saturation of the ensemble performance, which converges to the performance of a single model. Moreover, this does not only affect the quality of its predictions, but even more so the uncertainty estimates of the ensemble, and thus its performance on out-of-distribution data. We hypothesize that this limitation can be overcome by discouraging different ensemble members from collapsing to the same function. To this end, we introduce a kernelized repulsive term in the update rule of the deep ensembles. We show that this simple modification not only enforces and maintains diversity among the members but, even more importantly, transforms the maximum a posteriori inference into proper Bayesian inference. Namely, we show that the training dynamics of our proposed repulsive ensembles follow a Wasserstein gradient flow of the KL divergence to the true posterior. We study repulsive terms in weight and function space and empirically compare their performance to standard ensembles and Bayesian baselines on synthetic and real-world prediction tasks. + +# 1 Introduction + +There have been many recent advances on the theoretical properties of sampling algorithms for approximate Bayesian inference, which changed our interpretation and understanding of them. Particularly worth mentioning is the work of Jordan et al. [38], who reinterpret Markov Chain Monte Carlo (MCMC) as a gradient flow of the KL divergence over the Wasserstein space of probability measures. This new formulation allowed for a deeper understanding of approximate inference methods but also inspired the inception of new and more efficient inference strategies. Following this direction, Liu and Wang $\pmb { \mathbb { B } } \mathbf { \mathbb { 1 } }$ recently proposed the Stein Variational Gradient Descent (SVGD) method to perform approximate Wasserstein gradient descent. Conceptually, this method, which belongs to the family of particle-optimization variational inference (POVI), introduces a repulsive force through a kernel acting in the parameter space to evolve a set of samples towards high-density regions of the target distribution without collapsing to a point estimate. + +Another method which has achieved great success recently are ensembles of neural networks (socalled deep ensembles), which work well both in terms of predictive performance $\pm 2 \sqrt { 8 0 }$ as well as uncertainty estimation $\begin{array} { r l } { { \| 6 5 \| } } \end{array}$ , and have also been proposed as a way to perform approximate inference in Bayesian neural networks [82, 36]. That being said, while they might allow for the averaging of predictions over several hypotheses, they do not offer any guarantees for the diversity between those hypotheses nor do they provably converge to the true Bayesian posterior under any meaningful limit. In this work, we show how the introduction of a repulsive term between the members in the ensemble, inspired by SVGD, not only naïvely guarantees the diversity among the members, avoiding their collapse in parameter space, but also allows for a reformulation of the method as a gradient flow of the KL divergence in the Wasserstein space of distributions. It thus allows to endow deep ensembles with convergence guarantees to the true Bayesian posterior. + +![](images/c87b35667fe31cb24e44c323fe8bffeb4df13996d7df41a00bfd504161843062.jpg) +Figure 1: BNN 1D regression. The function-space methods (SVGD and WGD) approach the HMC posterior more closely, while the standard deep ensembles and weight-space methods fail to properly account for the uncertainty, especially the in-between uncertainty. + +An additional problem is that BNN inference in weight space can lead to degenerate solutions, due to the overparametrization of these models. That is, several samples could have very different weights but map to the same function, thus giving a false sense of diversity in the ensemble. This property, that we will refer to as non-identifiability of neural networks (see Appendix $\boxed { \mathrm { A } }$ , can lead to redundancies in the posterior distribution. It implies that methods like MCMC sampling, deep ensembles, and SVGD waste computation in local modes that account for equivalent functions. Predictive distributions approximated using samples from these modes do not improve over a simple point estimate and lead to a poor uncertainty estimation. Following this idea, Wang et al. $\mathbb { \left[ \left] \right. \right]} 2 $ introduced a new method to extend POVI methods to function space, overcoming this limitation. Here, we also study an update rule that allows for an approximation of the gradient flow of the KL divergence in function space in our proposed repulsive ensembles. + +We make the following contributions: + +• We derive several different repulsion terms that can be added as regularizers to the gradient updates of deep ensembles to endow them with Bayesian convergence properties. +• We show that these terms approximate Wasserstein gradient flows of the KL divergence and can be used both in weight space and function space. +• We compare these proposed methods theoretically to standard deep ensembles and SVGD and highlight their different guarantees. +• We assess all these methods on synthetic and real-world deep learning tasks and show that our proposed repulsive ensembles can achieve competitive performance and improved uncertainty estimation. + +# 2 Repulsive Deep Ensembles + +In supervised deep learning, we typically consider a likelihood function $p ( \pmb { y } | f ( \pmb { x } ; \mathbf { w } ) )$ (e.g., Gaussian for regression or Categorical for classification) parameterized by a neural network $f ( \pmb { x } ; \mathbf { w } )$ and training data $\textit { D } = \{ ( \boldsymbol { \mathbf { { x } } } _ { i } , \boldsymbol { \mathbf { { y } } } _ { i } ) \} _ { i = 1 } ^ { n }$ with $\textbf { \textit { x } } \in \textbf { \textit { X } }$ and $\textbf { \textit { y } } \in \textbf { \textit { y } }$ . In Bayesian neural networks (BNNs), we are interested in the posterior distribution of all likely networks given by $\begin{array} { r } { p ( \mathbf { w } | \mathcal { D } ) \propto \prod _ { i = 1 } ^ { n } p ( \pmb { y } _ { i } | f ( \pmb { x } _ { i } ; \mathbf { w } ) ) p ( \mathbf { w } ) } \end{array}$ , where $p ( \mathbf { w } )$ is the prior distribution over weights. Crucially, when making a prediction on a test point $\mathbf { \boldsymbol { x } } ^ { * }$ , in the Bayesian approach we do not only use a single parameter $\widehat { \bf w }$ to predict ${ \pmb y } ^ { * } = f ( { \pmb x } ^ { * } ; \widehat { \mathbf { w } } )$ , but we marginalize over the whole posterior, thus taking all possible explanations of the data into account: + +$$ +p ( \pmb { y } ^ { * } | \pmb { x } ^ { * } , \mathcal { D } ) = \int p ( \pmb { y } ^ { * } | f ( \pmb { x } ^ { * } ; \mathbf { w } ) ) p ( \mathbf { w } | \mathcal { D } ) \mathrm { d } \mathbf { w } +$$ + +While approximating the posterior of Bayesian neural networks (or sampling from it) is a challenging task, performing maximum a posteriori (MAP) estimation, which corresponds to finding the mode of the posterior, is usually simple. Ensembles of neural networks use the non-convexity of the MAP optimization problem to create a collection of $K$ independent—and possibly different— solutions. Considering $n$ weight configurations of a neural network $\{ { \bf w } _ { i } \} _ { i = 1 } ^ { n }$ with ${ \bf w } _ { i } \in \mathbb { R } ^ { d }$ , the dynamics of the ensemble under the gradient of the posterior lead to the following update rule at iteration $t$ : + +$$ +\begin{array} { r l } & { \mathbf { w } _ { i } ^ { t + 1 } \mathbf { w } _ { i } ^ { t } + \epsilon _ { t } \phi ( \mathbf { w } _ { i } ^ { t } ) } \\ { \mathrm { w i t h } \quad } & { \phi ( \mathbf { w } _ { i } ^ { t } ) = \nabla _ { \mathbf { w } _ { i } ^ { t } } \log p ( \mathbf { w } _ { i } ^ { t } | \mathcal { D } ) , } \end{array} +$$ + +with step size $\epsilon _ { t }$ . Ensemble methods have a long history [e.g., 45, 26, 6] and were recently revisited for neural networks $\pm 2 \|$ and coined deep ensembles. The predictions of the different members are combined to create a predictive distribution by using the solutions to compute the Bayesian model average (BMA) in Eq. $\mathbb { D }$ . Recent works $\mathbb { \lVert \rVert }$ have shown that deep ensembles can outperform some of the Bayesian approaches for uncertainty estimation. Even more recently, Wilson and Izmailov $\pmb { \Vert 8 2 \Vert }$ argued that deep ensembles can be considered a compelling approach to Bayesian model averaging. Despite these ideas, the ability of deep ensembles to efficiently average over multiple hypotheses and to explore the functional landscape of the posterior distribution studied in $[ \overline { { 1 8 } } ]$ does not guarantee sampling from the right distribution. Indeed, the additional Langevin noise introduced in $\mathbb { [ \breve { ] { ] \mathrm { ~ Z ~ Z ~ } ] } } }$ , which is not considered in deep ensembles, is crucial to ensure samples from the true Bayesian posterior. + +From a practical standpoint, since the quality of an ensemble hinges on the diversity of its members, many methods were recently proposed to improve this diversity without compromising the individual accuracy. For instance, Wenzel et al. [80] propose hyper-deep ensembles that combine deep networks with different hyperparameters. Similarly, cyclical learning-rate schedules can explore several local minima for the ensemble members $\pmb { \mathbb { B 3 } }$ . Alternatively, Rame and Cord $ { \mathbb { I } } { \mathbb { I } }$ proposed an informationtheoretic framework to avoid redundancy in the members and Oswald et al. $\overline { { \| 6 3 \| } }$ studied possible interactions between members based on weight sharing. However, the absence of a constraint that prevents particles from converging to the same mode limits the possibility of improvement by introducing more ensemble members. This means that any hopes to converge to different modes must exclusively rely on: + +1. the randomness of the initialization +2. the noise in the estimation of the gradients due to minibatching +3. the number of local optima that might be reached during gradient descent. + +Moreover, the recent study of Geiger et al. $\pmb { \left. \pmb { \left. \bar { 2 5 } \right. } \right. }$ showed how the empirical test error of the ensemble converges to the one of a single trained model when the number of parameters goes to infinity, leading to deterioration of the performance. In other words, the bigger the model, the harder it is to maintain diversity in the ensemble and avoid collapse to the same solution. This is intuitively due to the fact that bigger models are less sensitive to the initialization. Namely, in order for them to get stuck in a local minimum, they must have second derivatives that are positive simultaneously in all directions. As the number of hidden units gets larger, this becomes less likely. + +# 2.1 Repulsive force in weight space + +To overcome the aforementioned limitations of standard deep ensembles, we introduce, inspired by SVGD $\mathbb { \left[ \left[ 5 1 \right] \right] }$ , a deep ensemble with members that interact with each other through a repulsive component. Using a kernel function to model this interaction, the single models repel each other based on their position in the weight space, so that two members can never assume the same weights. Considering a stationary kernel $\bar { k } ( \cdot , \cdot ) : \mathbb { R } ^ { d } \times \mathbb { R } ^ { d } \to \mathbb { R }$ acting in the parameter space of the neural network, a repulsive term $\mathcal { R }$ can be parameterized through its gradient: + +$$ +\phi ( \mathbf { w } _ { i } ^ { t } ) = \nabla _ { \mathbf { w } _ { i } ^ { t } } \log p ( \mathbf { w } _ { i } ^ { t } | \mathcal { D } ) - \mathcal { R } \left( \left\{ \nabla _ { \mathbf { w } _ { i } ^ { t } } k ( \mathbf { w } _ { i } ^ { t } , \mathbf { w } _ { j } ^ { t } ) \right\} _ { j = 1 } ^ { n } \right) . +$$ + +To get an intuition for the behavior of this repulsive term and its gradients, we can consider the RBF kernel $\begin{array} { r } { k ( \mathbf { w } _ { i } , \mathbf { w } _ { j } ) = \exp \big ( - \frac { 1 } { h } | | \mathbf { w } _ { i } - \mathbf { w } _ { j } | | ^ { 2 } \big ) } \end{array}$ with lengthscale $h$ and notice how its gradient + +$$ +\nabla _ { \mathbf { w } _ { i } ^ { t } } k ( \mathbf { w } _ { i } ^ { t } , \mathbf { w } _ { j } ^ { t } ) = \frac { 2 } { h } ( \mathbf { w } _ { j } ^ { t } - \mathbf { w } _ { i } ^ { t } ) k ( \mathbf { w } _ { i } ^ { t } , \mathbf { w } _ { j } ^ { t } ) +$$ + +drives $\mathbf { w } _ { i }$ away from its neighboring members ${ \bf w } _ { j }$ , thus creating a repulsive effect. Naturally, not all the choices of $\mathcal { R }$ induce this effect. One of the simplest formulations to obtain it is via a linear combination of the kernel gradients scaled by a positive factor, that is, $\begin{array} { r } { \beta \sum _ { j = 1 } ^ { n } \nabla _ { \mathbf { w } _ { i } ^ { t } } k ( \mathbf { w } _ { i } ^ { t } , \mathbf { w } _ { j } ^ { t } ) } \end{array}$ with $\beta \in \mathbb { R } _ { * } ^ { + }$ . We will see in Section $\begin{array} { l } { 3 } \\ { . } \end{array}$ how the choice of $\beta$ can be justified in order to obtain convergence to the Bayesian posterior together with alternative possible formulations of $\mathcal { R }$ that preserve this convergence. + +# 2.2 Repulsive force in function space + +To overcome the aforementioned overparameterization issue, the update in Eq. $\textcircled { 3 }$ can be formulated in function space instead of weight space. Let $f : \mathbf { w } \mapsto f ( \cdot ; \mathbf { w } )$ be the map that maps a configuration of weights $\mathbf { w } \in \mathbb { R } ^ { d }$ to the corresponding neural network regression function and denote as $f _ { i } : = f ( \cdot ; { \mathbf w } _ { i } )$ the function with a certain configuration of weights $\mathbf { w } _ { i }$ . We can now consider $n$ particles in function space $\{ f _ { i } \} _ { i = 1 } ^ { n }$ with $\pmb { f } \in \mathcal { F }$ and model their interaction with a general positive definite kernel $k ( \cdot , \cdot )$ . We also consider the implicit functional likelihood $p ( \pmb { y } | \pmb { x } , \pmb { f } )$ , determined by the measure $p ( \pmb { y } | \pmb { x } , \mathbf { w } )$ in the weight space, as well as the functional prior $p ( f )$ , which can either be defined separately (e.g., using a GP) or modeled as a push-forward measure of the weight-space prior $p ( \mathbf { w } )$ . Together, they determine the posterior in function space $p ( \pmb { f } | \mathcal { D } )$ . The functional evolution of a particle can then be written as: + +$$ +\begin{array} { r l } & { f _ { i } ^ { t + 1 } \gets f _ { i } ^ { t } + \epsilon _ { t } \phi ( f _ { i } ^ { t } ) } \\ { \mathrm { t h } \quad } & { \phi ( f _ { i } ^ { t } ) = \nabla _ { f _ { i } ^ { t } } \log p ( f _ { i } ^ { t } | \mathcal { D } ) - \mathcal { R } \left( \left\{ \nabla _ { f _ { i } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \right\} _ { j = 1 } ^ { n } \right) . } \end{array} +$$ + +However, computing the update in function space is neither tractable nor practical, which is why two additional considerations are needed. The first one regards the infinite dimensionality of function space, which we circumvent using a canonical projection into a subspace: + +Definition 1 (Canonical projection). For any $A \subset { \mathcal { X } }$ , we define $\pi _ { A } : \mathbb { R } ^ { \mathcal { X } } \to \mathbb { R } ^ { A }$ as the canonical projection onto $A$ , that is, $\pi _ { A } ( f ) = \{ f ( a ) \} _ { a \in A }$ . + +In other words, the kernel will not be evaluated directly in function space, but on the projection $k \big ( \pi _ { B } ( f ) , \pi _ { B } ( f ^ { \prime } ) \big )$ , with $B$ being a subset of the input space given by a batch of training data points. The second consideration is to project this update back into the parameter space and evolve a set of particles there, because ultimately we are interested in representing the functions by parameterized neural networks. For this purpose, we can use the Jacobian of the $i$ -th particle as a projector: + +$$ +\phi ( { \mathbf w } _ { i } ^ { t } ) = \left( \frac { \partial f _ { i } ^ { t } } { \partial { \mathbf w } _ { i } ^ { t } } \right) ^ { \top } \left[ \nabla _ { f _ { i } ^ { t } } \log p ( f _ { i } ^ { t } | { \mathcal D } ) - { \mathcal R } \left( \left\{ \nabla _ { f _ { i } ^ { t } } k ( \pi _ { B } ( f _ { i } ^ { t } ) , \pi _ { B } ( f _ { j } ^ { t } ) ) \right\} _ { j = 1 } ^ { n } \right) \right] . +$$ + +# 2.3 Comparison to Stein variational gradient descent + +Note that our update is reminiscent of SVGD $\mathbb { \left[ 5 1 \right] }$ , which in parameter space can be written as: + +$$ +\phi ( \mathbf { w } _ { i } ^ { t } ) = \sum _ { j = 1 } ^ { n } k ( \mathbf { w } _ { i } ^ { t } , \mathbf { w } _ { j } ^ { t } ) \nabla _ { \mathbf { w } _ { i } ^ { t } } \log p ( \mathbf { w } _ { i } ^ { t } | \mathcal { D } ) + \sum _ { j = 1 } ^ { n } \nabla _ { \mathbf { w } _ { j } ^ { t } } k ( \mathbf { w } _ { j } ^ { t } , \mathbf { w } _ { i } ^ { t } ) . +$$ + +It is important to notice that here, the gradients are averaged across all the particles using the kernel matrix. Interestingly, SVGD can be asymptotically interpreted as gradient flow of the KL divergence under a new metric induced by the Stein operator [16, 50] (see Appendix D for more details). Moving the inference from parameter to function space $\pmb { \mathbb { Z } } 6 \|$ leads to the update rule + +$$ +\phi ( \mathbf { w } _ { i } ^ { t } ) = \left( \frac { \partial f _ { i } ^ { t } } { \partial \mathbf { w } _ { i } ^ { t } } \right) ^ { \top } \left( \frac { 1 } { n } \sum _ { j = 1 } ^ { n } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \nabla _ { f _ { j } ^ { t } } \log p ( f _ { j } ^ { t } | \mathcal { D } ) + \nabla _ { f _ { j } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \right) . +$$ + +This way of averaging gradients using a kernel can be dangerous in high-dimensional settings, where kernel methods often suffer from the curse of dimensionality. Moreover, in Eq. $\textcircled{6}$ , the posterior gradients of the particles are averaged using their similarity in weight space, which can be misleading in multi-modal posteriors. Worse yet, in Eq. $\textcircled{7}$ , the gradients are averaged in function space and are then projected back using exclusively the $i$ -th Jacobian, which can be harmful given that it is not guaranteed that distances between functions evaluated on a subset of their input space resemble their true distance. Our proposed method, on the other hand, does not employ any averaging of the posterior gradients and thus comes closest to the true particle gradients in deep ensembles. + +# 3 Repulsive deep ensembles are Bayesian + +So far, we represented the repulsive force as a general function of the gradients of a kernel. In this section, we show how to determine the explicit form of the repulsive term, such that the resulting update rule is equivalent to the discretization of the gradient flow dynamics of the KL divergence in Wasserstein space. We begin by introducing the concepts of particle approximation and gradient flow. + +# 3.1 Particle approximation + +A particle-based approximation of a target measure depends on a set of weighted samples $\{ ( \bar { x _ { i } } , w _ { i } ) \} _ { i = 1 } ^ { n }$ , for which an empirical measure can be defined as + +$$ +\rho ( x ) = \sum _ { i = 1 } ^ { n } w _ { i } \delta ( x - x _ { i } ) , +$$ + +where $\delta ( \cdot )$ is the Dirac delta function and the weights $w _ { i }$ satisfy $w _ { i } \in [ 0 , 1 ]$ and $\textstyle \sum _ { i = 1 } ^ { n } w _ { i } = 1$ . To approximate a target distribution $\pi ( x )$ using the empirical measure, the particles and their weights need to be selected in a principled manner that minimizes some measure of distance between $\pi ( x )$ and $\rho ( x )$ (e.g., a set of $N$ samples with weights $w _ { i } = 1 / N$ obtained using an MCMC method). + +# 3.2 Gradient flow in parameter space + +Given a smooth function $J : \mathbb { R } ^ { d } \mathbb { R }$ in Euclidean space, we can minimize it by creating a path that follows its negative gradient starting from some initial conditions $x _ { 0 }$ . The curve $x ( t )$ with starting point $x _ { 0 }$ described by that path is called gradient flow. The dynamics and evolution in time of a considered point in the space under this minimization problem can be described as the ODE1 + +$$ +\frac { d \boldsymbol { x } } { d t } = - \boldsymbol { \nabla } J ( \boldsymbol { x } ) . +$$ + +We can extend this concept to the space of probability distributions (Wasserstein gradient flow) [3]. Let us consider the space of probability measures $\mathcal { P } _ { 2 } ( \mathcal { M } )$ , that is, the set of probability measures with finite second moments defined on the manifold $\mathcal { M }$ : + +$$ +\mathcal { P } _ { 2 } ( \mathcal { M } ) = \left\{ \varphi : \mathcal { M } \to [ 0 , \infty ) \bigg | \int _ { \mathcal { M } } \mathrm { d } \varphi = 1 , ~ \int _ { \mathcal { M } } | x | ^ { 2 } \varphi ( x ) \mathrm { d } x < + \infty \right\} . +$$ + +Taking $\Pi ( \mu , \nu )$ as the set of joint probability measures with marginals $\mu , \nu$ , we can define the Wasserstein metric on the space $\mathcal { P } _ { 2 } ( \mathcal { M } )$ as: + +$$ +W _ { 2 } ^ { 2 } ( \mu , \nu ) = \operatorname* { i n f } _ { \pi \in \Pi ( \mu , \nu ) } \int | x - y | ^ { 2 } { \mathrm { d } } \pi ( x , y ) . +$$ + +Considering the optimization problem of a functional $J : \mathcal { P } _ { 2 } ( \mathcal { M } ) \mathbb { R }$ , such as the $\mathrm { K L }$ divergence between the particle approximation in Eq. $\textcircled { 8 }$ and the target posterior $\pi ( x )$ , + +$$ +\operatorname* { i n f } _ { \rho \in \mathcal { P } _ { 2 } ( \mathcal { M } ) } D _ { K L } ( \rho , \pi ) = \int _ { \mathcal { M } } ( \log \rho ( x ) - \log \pi ( x ) ) \rho ( x ) \mathrm { d } x , +$$ + +the evolution in time of the measure $\rho$ under the equivalent of the gradient, the Wasserstein gradient flow, is described by the Liouville equation2 [38, 3, 64]: + +$$ +\begin{array} { l } { \displaystyle \frac { \partial \rho ( \boldsymbol { x } ) } { \partial t } = \nabla \cdot \left( \rho ( \boldsymbol { x } ) \nabla \frac { \delta } { \delta \rho } D _ { K L } ( \rho , \pi ) \right) } \\ { \displaystyle = \nabla \cdot \left( \rho ( \boldsymbol { x } ) \nabla \big ( \log \rho ( \boldsymbol { x } ) - \log \pi ( \boldsymbol { x } ) \big ) \right) , } \end{array} +$$ + +where $\nabla _ { \delta \rho } ^ { \delta } D _ { K L } ( \rho , \pi ) \ = : \ \nabla _ { \mathcal { W } _ { 2 } } D _ { K L } ( \rho , \pi )$ is the Wasserstein gradient and the operator $\frac { \delta } { \delta \rho }$ $\mathcal { P } _ { 2 } ( \mathcal { M } ) \stackrel { \cdot } { } \mathbb { R }$ represents the functional derivative or first variation (see Appendix $\boxed { \mathsf { C } }$ for more details). In the particular case of the KL functional, we can recover the Fokker-Planck equation, + +$$ +\begin{array} { r l r } & { } & { \frac { \partial \rho ( x ) } { \partial t } = \nabla \cdot \left( \rho ( x ) \nabla ( \log \rho ( x ) - \log \pi ( x ) ) \right) } \\ & { } & { = - \nabla \cdot \left( \rho ( x ) \nabla \log \pi ( x ) \right) + \nabla ^ { 2 } \rho ( x ) , } \end{array} +$$ + +that admits as unique stationary distribution the posterior $\pi ( x )$ . The deterministic particle dynamics ODE $\left[ \left[ 2 \right] \right]$ related to Eq. $\mathbb { ( 1 1 ) }$ , namely mean-field Wasserstein dynamics, is then given by: + +$$ +\frac { d x } { d t } = - \nabla \big ( \log \rho ( x ) - \log \pi ( x ) \big ) . +$$ + +Considering a discretization of Eq. (12) for a particle system $\{ x \} _ { i = 1 } ^ { n }$ and small stepsize $\epsilon _ { t }$ , we can rewrite Eq. $\mathbb { \underline { { ( 1 2 ) } } }$ as: + +$$ +x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \epsilon _ { t } \bigl ( \nabla \log \pi ( x _ { i } ^ { t } ) - \nabla \log \rho ( x _ { i } ^ { t } ) \bigr ) . +$$ + +Unfortunately, we do not have access to the analytical form of the gradient $\nabla \log \rho$ , so an approximation is needed. At this point, it is crucial to observe the similarity between the discretization of the Wasserstein gradient flow in Eq. $\textcircled { 1 3 }$ and the repulsive update in Eq. $( 3 )$ to notice how, if the kernelized repulsion is an approximation of the gradient of the empirical particle measure, the update rule minimizes the KL divergence between the particle measure and the target posterior. Different sample-based approximations of the gradient that use a kernel function have been recently studied. The simplest one is given by the kernel density estimation (KDE) (details in Appendix $\dot { \mathrm { ~ E ~ } } )$ $\begin{array} { r } { \tilde { \rho } _ { t } ( x ) = \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \dot { k } ( x , x _ { t } ^ { i } ) } \end{array}$ , where $\boldsymbol { k } ( \cdot , \cdot ) : \mathbb { R } ^ { d } \times \mathbb { R } ^ { d } \dot { \mathbb { R } }$ and the gradient of its log density is given by $\mathbb { \ m }$ : + +$$ +\nabla \log \rho ( x _ { i } ^ { t } ) \approx \frac { \sum _ { j = 1 } ^ { n } \nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } . +$$ + +Using this approximation in Eq. $( 1 3 )$ we obtain: + +$$ +x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \epsilon _ { t } \Bigg ( \nabla \log \pi ( x _ { i } ^ { t } ) - \frac { \sum _ { j = 1 } ^ { n } \nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } \Bigg ) , +$$ + +where, if we substitute the posterior for $\pi$ , we obtain an expression for the repulsive force in Eq. $\textcircled{3}$ . This shows that if the repulsive term in Eq. $( 3 )$ is the normalized sum of the gradients $\mathcal { R } =$ $\begin{array} { r } { \left( \bar { \sum _ { j = 1 } ^ { n } } k ( \mathbf { w } _ { i } ^ { t } , \mathbf { w } _ { j } ^ { t } ) \right) ^ { - 1 } \sum _ { j = 1 } ^ { n } \nabla _ { \mathbf { w } _ { i } ^ { t } } k ( \mathbf { w } _ { i } ^ { t } , \mathbf { w } _ { j } ^ { t } ) } \end{array}$ , we do not only encourage diversity of the ensemble members and thus avoid collapse, but surprisingly—in the asymptotic limit of , where the KDE approximation is exact $\left[ \left[ 6 6 \right] \right]$ —also converge to the true Bayesian posterior! + +Nevertheless, approximating the gradient of the empirical measure with the KDE can lead to suboptimal performance, as already studied by Li and Turner $[ \overline { { | 4 6 | } }$ . They instead introduced a new Stein gradient estimator (SGE) that offers better performance, while maintaining the same computational cost. Even more recently, Shi et al. $\mathbb { \left. \boldsymbol { \mathfrak { G } } \boldsymbol { \mathfrak { A } } \right. }$ introduced a spectral method for gradient estimation (SSGE), that also allows for a simple estimation on out-of-sample points. These two estimators can be used in Eq. $\textcircled { 1 3 }$ , to formulate the following update rules with two alternative repulsive forces. The one using the Stein estimator, that we will call SGE-WGD, is: + +$$ +x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \epsilon _ { t } \bigg ( \nabla \log \pi ( x _ { i } ^ { t } ) + \sum _ { j = 1 } ^ { n } ( K + \eta \mathbb { I } ) _ { i j } ^ { - 1 } \sum _ { k = 1 } ^ { n } \nabla _ { x _ { k } ^ { t } } k ( x _ { k } ^ { t } , x _ { j } ^ { t } ) \bigg ) , +$$ + +where $K$ is the kernel Gram matrix, $\eta$ a small constant, and I the identity matrix. We can notice an important difference between KDE and SGE, in that the former is only considering the interaction of the $i$ -th particle being updated with all the others, while the latter is simultaneously considering also the interactions between the remaining particles. The spectral method, that we will call SSGE-WGD, leads to the following update rule: + +$$ +x _ { t + 1 } ^ { i } = x _ { t } ^ { i } + \epsilon _ { t } \bigg ( \nabla \log \pi ( x _ { t } ^ { i } ) + \sum _ { j = 1 } ^ { J } \frac { 1 } { \lambda _ { j } ^ { 2 } } \sum _ { m = 1 } ^ { n } \sum _ { k = 1 } ^ { n } u _ { j k } \nabla _ { x _ { m } } k ( x _ { m } ^ { t } , x _ { k } ^ { t } ) \cdot \sum _ { l = 1 } ^ { n } u _ { j l } k ( x _ { i } ^ { t } , x _ { l } ^ { t } ) \bigg ) +$$ + +where $\lambda _ { j }$ is the $j$ -th eigenvalue of the kernel matrix and $u _ { j k }$ is the $k$ -th component of the $j$ -th eigenvector. Computationally, both SSGE and SGE have a cost of $\mathcal { O } ( M ^ { 3 } + M ^ { 2 } d )$ , with $M$ being the number of points and $d$ their dimensionality. SSGE has an additional cost for predictions of $\mathcal { O } ( M ( d + J ) )$ , where $J$ is the number of eigenvalues. + +![](images/909c60e208223cd54336af828c72fc5777f5cbee78dd1b5dffc27345c64eca36.jpg) +Figure 2: Single Gaussian. We show samples from SVGD, KDE-WGD, SGE-WGD, and SSGEWGD (from left to right). The upper and right plots show the empirical one-dimensional marginal distributions obtained using KDE on the samples (red) and on the particles (blue). + +# 3.3 Gradient flow in function space + +To theoretically justify the update rule introduced in function space in Eq. $\textcircled{5}$ , we can rewrite the Liouville equation for the gradient flow in Eq. $\textcircled { 1 1 }$ in function space as + +$$ +\begin{array} { l } { \displaystyle \frac { \partial \rho ( { \pmb f } ) } { \partial t } = \nabla \cdot \left( \rho ( { \pmb f } ) \nabla \frac { \delta } { \delta \rho } D _ { K L } ( \rho , \pi ) \right) } \\ { \displaystyle = \nabla \cdot \left( \rho ( { \pmb f } ) \nabla \big ( \log \rho ( { \pmb f } ) - \log \pi ( { \pmb f } ) \big ) \right) . } \end{array} +$$ + +Following this update, the mean field functional dynamics are + +$$ +\frac { d \pmb f } { d t } = - \nabla \big ( \log \rho ( \pmb f ) - \log \pi ( \pmb f ) \big ) . +$$ + +Using the same KDE approximation as above, we can obtain a discretized evolution in function space and with it an explicit form for the repulsive force in Eq. $\textcircled{4}$ as + +$$ +\pmb { f } _ { t + 1 } ^ { i } = \pmb { f } _ { t } ^ { i } + \epsilon _ { t } \bigg ( \nabla _ { \pmb { f } } \log \pi ( \pmb { f } _ { t } ^ { i } ) - \frac { \sum _ { j = 1 } ^ { n } \nabla _ { \pmb { f } _ { i } ^ { t } } k ( \pmb { f } _ { i } ^ { t } , \pmb { f } _ { j } ^ { t } ) } { \sum _ { j = 1 } ^ { n } k ( \pmb { f } _ { i } ^ { t } , \pmb { f } _ { j } ^ { t } ) } \bigg ) . +$$ + +The update rules using the SGE and SSGE approximations follow as for the parametric case. It is important to notice that this update rule requires the function space prior gradient: $\nabla _ { f _ { j } } \log { p ( f _ { j } | x , y ) } =$ $\nabla _ { f _ { j } } \log p ( \pmb { y } | \pmb { x } , \pmb { f _ { j } } ) + \nabla _ { f _ { j } } \log p ( \pmb { f _ { j } } )$ . If one wants to use an implicit prior defined in weight space, an additional estimator is needed due to its analytical intractability. We again adopted the SSGE, introduced by Shi et al. $\mathbb { \lVert \boldsymbol { 6 9 } \rVert }$ , which was already used for a similar purpose in Sun et al. $\pmb { \mathbb { Z } 1 }$ . It is also interesting to note that the update rule in Eq. $\textcircled { 1 2 0 }$ readily allows for the use of alternative priors that have an analytical form, such as Gaussian processes. This is an important feature of our method that allows for an explicit encoding of function space properties that can be useful for example in achieving better out of distribution detection capabilities $\checkmark$ . + +# 3.4 The choice of the kernel + +A repulsive effect can always be created in the ensemble by means of the gradient of any kernel function that is measuring the similarity between two members. Nevertheless, it is important to keep in mind that to ensure the asymptotic convergence to the Bayesian posterior, the repulsive component must be a consistent estimator of the gradient in Eq. $( 1 3 )$ , as shown in Section $\checkmark$ . Therefore, some important constraints over the kernel choice are needed. In particular, the SGE and SSGE need a kernel function belonging to the Stein class (see Shi et al. $\dot { \overline { { \vert 6 9 \vert } } }$ for more details). On the other hand, for the KDE, any symmetric probability density function can be used. On this subject, the work of Aggarwal et al. [1] has shown how the Manhattan distance metric (L1 norm) might be preferable over the Euclidean distance metric (L2 norm) for high-dimensional settings. We performed some additional experiments using the L1 norm (Laplace kernel) for the KDE but we could not observe any substantial difference compared to using the L2 norm. Further investigations regarding this hypothesis are left for future research. + +![](images/094ed7ba5611118de5688c3ddd1545168ced55dff3ababfdac0ee5e88615b714.jpg) +Figure 3: BNN 2D classification. We show the entropy of the predictive posteriors. Again, the function-space methods capture the uncertainty better than the weight-space ones, thus approaching the gold-standard HMC posterior. + +# 4 Experiments + +In this section, we compare the different proposed WGD methods with deep ensembles and SVGD on synthetic sampling, regression, and classification tasks and real-world image classification tasks. We use an RBF kernel (except where otherwise specified) with the popular median heuristic $\mathbb { \left[ \left[ 5 1 \right] \right] }$ to choose the kernel bandwidth. In our experiments, an adaptive bandwidth leads to better performance than fixing and tuning a single constant value for the entire evolution of the particles. We also quantitatively assess the uncertainty estimation of the methods in terms of calibration and OOD detection. In our experiments, we report the test accuracy, negative log-likelihood (NLL), and the expected calibration error (ECE) [58]. To assert the robustness on out-of-distribution (OOD) data, we report the ratio between predictive entropy on OOD and test data points $( H _ { o } / H _ { t } )$ , and the OOD detection area under the ROC curve AUROC(H) $\mathbb { H }$ . Moreover, to assess the diversity of the ensemble generated by the different methods in function space, we measure the functional diversity using the model disagreement (MD) (details in Appendix $\bar { \bigtriangledown } \bar { \bigtriangledown }$ . In particular, we report the ratio between the average model disagreement on the OOD and test data points $( M D _ { o } / M D _ { t } )$ and additionally the OOD detection AUROC(MD) computed using this measure instead of the entropy. + +Sampling from synthetic distributions As a sanity check, we first assessed the ability of our different approximations for Wasserstein gradient descent (using KDE, SGE, and SSGE) to sample from a two-dimensional Gaussian distribution (Figure $^ { 2 ) }$ . We see that our SGE-WGD, SSGE-WGD and the SVGD fit the target almost perfectly. We also tested the different methods in a more complex two-dimensional Funnel distribution $\mathbb { \left[ \left[ 5 9 \right] \right] }$ and present the results in Figure $\mathrm { F . l }$ in the Appendix. There, SGE-WGD and SVGD also perform best. + +BNN 1D regression We then assessed the different methods in fitting a BNN posterior on a synthetically generated one-dimensional regression task. The results are reported in Figure 1 consisting of the mean prediction and $\pm 1 , 2 , 3$ standard deviations of the predictive distribution. We can see that all methods performing inference in the weight space (DE, w-SVGD, WGD) are unable to capture the epistemic uncertainty between the two clusters of training data points. Conversely, the functional methods (f-SVGD, fWGD) are perfectly able to infer the diversity of the hypotheses in this region due to the lack of training evidence. They thereby achieve a predictive posterior that very closely resembles the one obtained with the gold-standard HMC sampling. + +BNN 2D classification Next, we investigated the predictive performance and quality of uncertainty estimation of the methods in a two-dimensional synthetic classification setting. The results are displayed in Figure $\textcircled { 3 }$ We can clearly observe that the weight-space methods are overconfident and do not capture the uncertainty well. Moreover, all the functions seems to collapse to the optimal classifier. These methods thus only account for uncertainty close to the decision boundaries and to the origin region, for which the uncertainty is purely aleatoric. In this setting, f-SVGD suffers from similar issues as the weight space methods, being overconfident away from the training data. Conversely, our fWGD methods are confident (low entropy) around the data but not out-of-distribution, thus representing the epistemic uncertainty better. This suggests that the functional diversity captured by this method naturally leads to a distance-aware uncertainty estimation $\mathbb { H 9 } \mathbb { L O }$ , a property that translates into confident predictions only in the proximity of the training data, allowing for a principled OOD detection. + +Table 1: BNN image classification. AUROC $\mathbf { ( H ) }$ is the AUROC computed using the entropy whereas AUROC(MD) is computed using the model disagreement. $\mathbf { H _ { o } / H _ { t } }$ is the ratio of the entropies on OOD and test points respectively and $\mathbf { M D _ { o } / M D _ { t } }$ is the ratio for model disagreement. We see that the best accuracy is achieved by our WGD methods, while our fWGD methods yield the best OOD detection and funtional diversity. All our proposed methods improve over standard deep ensembles in terms of accuracy and diversity, highlighting the effect of our repulsion. + +
PrroiiiieAUROC(H)AUROC(MD)AccuracyH/HtMD/MDtECENLL
Deep ensemble 40.958±0.0010.975±0.00191.122±0.0136.257±0.0056.394±0.0010.012±0.0010.129±0.001
SVGD [510.960±0.0010.973±0.00191.134±0.0246.315±0.0196.395±0.0180.014±0.0010.127±0.001
f-SVGD [760.956±0.0010.975±0.00189.884±0.0155.652±0.0096.531±0.0050.013±0.0010.150±0.001
hyper-DE [80]0.968±0.0010.981±0.00191.160±0.0076.682±0.0657.059±0.1520.014±0.0010.128±0.001
kde-WGD (ours)0.960±0.0010.970±0.00191.238±0.0196.587±0.0196.379±0.0180.014±0.0010.128±0.001
sge-WGD (ours)0.960±0.0010.970±0.00191.312±0.0166.562±0.0076.363±0.0090.012±0.0010.128±0.001
ssge-WGD (ours)0.968±0.0010.979±0.00191.198±0.0246.522±0.0096.610±0.0120.012±0.0010.130±0.001
kde-fWGD (ours)0.971±0.0010.980±0.00191.260±0.0117.079±0.0166.887±0.0150.015±0.0010.125±0.001
sge-fWGD (ours)0.969±0.0010.978±0.00191.192±0.0137.076±0.0046.900±0.0050.015±0.0010.125±0.001
ssge-fWGD (ours)0.971±0.0010.980±0.00191.240±0.0227.129±0.0066.951±0.0050.016±0.0010.124±0.001
CITIIIODeep ensemble 420.843±0.0040.736±0.00585,552±0.0762.244±0.0061.667±0.0080.049±0.0010.277±0.001
SVGD [510.825±0.0010.710±0.00285.142±0.0172.106±0.0031.567±0.0040.052±0.0010.287±0.001
fSVGD 回0.783±0.0010.712±0.00184.510±0.0311.968±0.0041.624±0.0030.049±0.0010.292±0.001
hyper-DE [80]0.789±0.0010.743±0.00184.743±0.0111.951±0.0101.690±0.0150.046±0.0010.288±0.001
kde-WGD (ours)0.838±0.0010.735±0.00485.904±0.0302.205±0.0031.661±0.0080.053±0.0010.276±0.001
sge-WGD (ours)0.837±0.0030.725±0.00485.792±0.0352.214±0.0101.634±0.0040.051±0.0010.275±0.001
ssge-WGD (ours)0.832±0.0030.731±0.00585.638±0.0382.182±0.0151.655±0.0010.049±0.0010.276±0.001
kde-fWGD (ours)0.791±0.0020.758±0.00284.888±0.0301.970±0.0041.749±0.0050.044±0.0010.282±0.001
sge-fWGD (ours)0.795±0.0010.754±0.00284.766±0.0601.984±0.0031.729±0.0020.047±0.0010.288±0.001
ssge-fWGD (ours)0.792±0.0020.752±0.00284.762±0.0341.970±0.0061.723±0.0050.046±0.0010.286±0.001
+ +FashionMNIST classification Moving on to real-world data, we used an image classification setting using the FashionMNIST dataset $\dot { \left[ \left| 8 3 \right| \right] }$ for training and the MNIST dataset $\bar { \mathbb { E } 3 } \mathbb { I }$ as an out-ofdistribution (OOD) task. The results are reported in Table $^ 1 { } _ { . }$ (top). We can see that all our methods improve upon standard deep ensembles and SVGD, highlighting the effectiveness of our proposed repulsion terms when training neural network ensembles. In particular, the sge-WGD offers the best accuracy, whereas the methods in function space all offer a better OOD detection. This is probably due to the fact that these methods achieve a higher entropy ratio and functional diversity measured via the model disagreement when compared to their weight-space counterparts. Interestingly, they also reach the lowest NLL values. We can also notice how the model disagreement (MD) not only serves its purpose as a metric for the functional heterogeneity of the ensemble but also allows for a better OOD detection in comparison to the entropy. To the best of our knowledge, this insight has not been described before, although it has been used in continual learning $\pmb { \mathbb { B } } \mathbf { \mathbb { 1 } }$ . Interestingly, using this metric, the hyper-deep ensemble $[ \textcircled { 8 0 } ]$ shows OOD detection performance comparable with our repulsive ensemble in function space. + +CIFAR classification Finally, we use a ResNet32 architecture $\mathbb { \left| \left[ 2 9 \right] \right| }$ on CIFAR-10 [41] with the SVHN dataset $\mathbb { \left[ 6 1 \right] }$ as OOD data. The results are reported in Table 1 (bottom). We can see that in this case, the weight-space methods achieve better performance in accuracy and OOD detection using the entropy than the ones in function space. Nevertheless, all our repulsive ensembles improve functional diversity, accuracy, and OOD detection when compared to standard SVGD, whereas the standard deep ensemble achieves the best OOD detection using the entropy. + +# 5 Related Work + +The theoretical and empirical properties of SVGD have been well studied [40, 48, 13] and it can also be seen as a Wasserstein gradient flow of the KL divergence in the Stein geometry [16, 50] (see Appendix D for more details). Interestingly, a gradient flow interpretation is also possible for (stochastic gradient) MCMC-type algorithms $\dot { \left. \overline { { 4 8 } } \right. }$ , which can be unified under a general particle inference framework [10]. Moreover, our Wasserstein gradient descent using the SGE approximation can also be derived using an alternative formulation as a gradient flow with smoothed test functions [48]. A projected version of WGD has been studied in Wang et al. $\mathbb { \left. \overline { { \boldsymbol { \mathscr { Q } } \boldsymbol { \cdot } \boldsymbol { \cdot } } } \right. }$ , which could also be readily applied in our framework. Besides particle methods, Bayesian neural networks $\pm \pm \infty$ have gained popularity recently [79, 22, 19, 36], using modern MCMC [59, 79, 22, 24, 21] and variational inference techniques [5, 72, 17, 34]. On the other hand, ensemble methods have also been extensively studied [42, 18, 82, 23, 80, 32, 85, 78].Moreover, repulsive interactions between the members have also been studied in Wabartha et al. [74]. Moreover, providing Bayesian interpretations for deep ensembles has been previously attempted through the lenses of stationary SGD distributions $[ \sqrt { 5 6 } , \sqrt ] { 8 } ]$ , ensembles of linear models $\dot { \bigtriangledown } 5 7 \big |$ , additional random functions $\boxed { 6 2 } , \boxed { 1 2 } , \boxed { 2 7 }$ , approximate inference $\lVert 8 2 \rVert$ , Stein variational inference [15], and marginal likelihood lower bounds $\bar { \| 5 3 \| }$ , and ensembles have also been shown to provide good approximations to the true BNN posterior in some settings $\left[ \left[ 3 6 \right] \right]$ . Furthermore, variational inference in function space has recently gained attention $\mathbb { \ m }$ and the limitations of the KL divergence have been studied in Burt et al. [7]. + +# 6 Conclusion + +We have presented a simple and principled way to improve upon standard deep ensemble methods. To this end, we have shown that the introduction of a kernelized repulsion between members of the ensemble not only improves the accuracy of the predictions but—even more importantly—can be seen as Wasserstein gradient descent on the KL divergence, thus transforming the MAP inference of deep ensembles into proper Bayesian inference. Moreover, we have shown that incorporating functional repulsion between ensemble members can improve the quality of the estimated uncertainties on simple synthetic examples and OOD detection on real-world data and can approach the true Bayesian posterior more closely. + +In future work, it will be interesting to study the impact of the Jacobian in the fWGD update and its implications on the Liouville equation in more detail, also compared to other neural network Jacobian methods, such as neural tangent kernels $\pmb { \Vert 3 7 } \Vert$ and generalized Gauss-Newton approximations $\boldsymbol { \left[ \left[ 3 5 \right] \right] }$ . Moreover, it would be interesting to derive explicit convergence bounds for our proposed method and compare them to the existing bounds for SVGD [40]. + +# Acknowledgments + +VF would like to acknowledge financial support from the Strategic Focus Area “Personalized Health and Related Technologies” of the ETH Domain through the grant #2017-110 and from the Swiss Data Science Center through a PhD fellowship. We thank Florian Wenzel, Alexander Immer, Andrew Gordon Wilson, Pavel Izmailov, Christian Henning, and Johannes von Oswald for helpful discussions. We also thank Dr. Sheldon Cooper, Dr. Leonard Hofstadter and Penny Hofstadter for their support and inspiration. + +# References + +[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of distance metrics in high dimensional space. In International conference on database theory, pages 420–434. Springer, 2001. +[2] Luigi Ambrosio and Gianluca Crippa. Continuity equations and ode flows with non-smooth velocity. Proceedings of the Royal Society of Edinburgh: Section A Mathematics, 144(6): 1191–1244, 2014. +[3] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré. Gradient flows: in metric spaces and in the space of probability measures. Springer Science & Business Media, 2008. +[4] Vijay Badrinarayanan, Bamdev Mishra, and Roberto Cipolla. Symmetry-invariant optimization in deep networks. arXiv preprint arXiv:1511.01754, 2015. +[5] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. arXiv preprint arXiv:1505.05424, 2015. + +[6] Leo Breiman. Bagging predictors. Machine learning, 24(2):123–140, 1996. + +[7] David R Burt, Sebastian W Ober, Adrià Garriga-Alonso, and Mark van der Wilk. Understanding variational inference in function-space. arXiv preprint arXiv:2011.09421, 2020. + +[8] Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference, converges to limit cycles for deep networks. In 2018 Information Theory and Applications Workshop (ITA), pages 1–10. IEEE, 2018. + +[9] An Mei Chen, Haw-minn Lu, and Robert Hecht-Nielsen. On the geometry of feedforward neural network error surfaces. Neural computation, 5(6):910–927, 1993. + +[10] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liqun Chen. A unified particleoptimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659, 2018. + +[11] Jiefeng Chen, Xi Wu, Yingyu Liang, Somesh Jha, et al. Robust out-of-distribution detection in neural networks. arXiv preprint arXiv:2003.09711, 2020. + +[12] Kamil Ciosek, Vincent Fortuin, Ryota Tomioka, Katja Hofmann, and Richard Turner. Conservative uncertainty estimation by fitting prior networks. In International Conference on Learning Representations, 2019. + +[13] Francesco D’Angelo and Vincent Fortuin. Annealed stein variational gradient descent. arXiv preprint arXiv:2101.09815, 2021. + +[14] Francesco D’Angelo and Christian Henning. Uncertainty-based out-of-distribution detection requires suitable function space priors. arXiv preprint arXiv:2110.06020, 2021. + +[15] Francesco D’Angelo, Vincent Fortuin, and Florian Wenzel. On stein variational neural network ensembles. arXiv preprint arXiv:2106.10760, 2021. + +[16] A Duncan, Nikolas Nuesken, and Lukasz Szpruch. On the geometry of stein variational gradient descent. arXiv preprint arXiv:1912.00894, 2019. + +[17] Michael W Dusenberry, Ghassen Jerfel, Yeming Wen, Yi-an Ma, Jasper Snoek, Katherine Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets with rank-1 factors. arXiv preprint arXiv:2005.07186, 2020. + +[18] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757, 2019. + +[19] Vincent Fortuin. Priors in bayesian deep learning: A review. arXiv preprint arXiv:2105.06868, 2021. + +[20] Vincent Fortuin, Mark Collier, Florian Wenzel, James Allingham, Jeremiah Liu, Dustin Tran, Balaji Lakshminarayanan, Jesse Berent, Rodolphe Jenatton, and Effrosyni Kokiopoulou. Deep classifiers with label noise modeling and distance awareness. arXiv preprint arXiv:2110.02609, 2021. + +[21] Vincent Fortuin, Adrià Garriga-Alonso, Mark van der Wilk, and Laurence Aitchison. Bnnpriors: A library for bayesian neural network inference with different prior distributions. Software Impacts, page 100079, 2021. + +[22] Vincent Fortuin, Adrià Garriga-Alonso, Florian Wenzel, Gunnar Rätsch, Richard Turner, Mark van der Wilk, and Laurence Aitchison. Bayesian neural network priors revisited. arXiv preprint arXiv:2102.06571, 2021. + +[23] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry Vetrov, and Andrew Gordon Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. arXiv preprint arXiv:1802.10026, 2018. + +[24] Adrià Garriga-Alonso and Vincent Fortuin. Exact langevin dynamics with stochastic gradients. arXiv preprint arXiv:2102.01691, 2021. + +[25] Mario Geiger, Arthur Jacot, Stefano Spigler, Franck Gabriel, Levent Sagun, Stéphane d’Ascoli, Giulio Biroli, Clément Hongler, and Matthieu Wyart. Scaling description of generalization with number of parameters in deep learning. Journal of Statistical Mechanics: Theory and Experiment, 2020(2):023401, 2020. +[26] L. K. Hansen and P. Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10):993–1001, October 1990. ISSN 0162-8828. doi: 10.1109/34.58871. URL https://doi.org/10.1109/34.58871. +[27] Bobby He, Balaji Lakshminarayanan, and Yee Whye Teh. Bayesian deep ensembles via the neural tangent kernel. arXiv preprint arXiv:2007.05864, 2020. +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015. +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. +[30] Robert Hecht-Nielsen. On the algebraic structure of feedforward network weight spaces. In Advanced Neural Computers, pages 129–135. Elsevier, 1990. +[31] Christian Henning, Maria R Cervera, Francesco D’Angelo, Johannes von Oswald, Regina Traber, Benjamin Ehret, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Posterior meta-replay for continual learning. arXiv preprint arXiv:2103.01133, 2021. +[32] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger. Snapshot ensembles: Train 1, get m for free. arXiv preprint arXiv:1704.00109, 2017. +[33] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger. Snapshot ensembles: Train 1, get M for free. CoRR, abs/1704.00109, 2017. URL http: //arxiv.org/abs/1704.00109. +[34] Alexander Immer, Matthias Bauer, Vincent Fortuin, Gunnar Rätsch, and Mohammad Emtiyaz Khan. Scalable marginal likelihood estimation for model selection in deep learning. arXiv preprint arXiv:2104.04975, 2021. +[35] Alexander Immer, Maciej Korzepa, and Matthias Bauer. Improving predictions of bayesian neural nets via local linearization. In International Conference on Artificial Intelligence and Statistics, pages 703–711. PMLR, 2021. +[36] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Wilson. What are bayesian neural network posteriors really like? arXiv preprint arXiv:2104.14421, 2021. +[37] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: convergence and generalization in neural networks. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 8580–8589, 2018. +[38] Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker– planck equation. SIAM journal on mathematical analysis, 29(1):1–17, 1998. +[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. +[40] Anna Korba, Adil Salim, Michael Arbel, Giulia Luise, and Arthur Gretton. A non-asymptotic analysis for stein variational gradient descent. Advances in Neural Information Processing Systems, 33, 2020. +[41] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. +[42] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in neural information processing systems, pages 6402–6413, 2017. +[43] Yann LeCun. The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/, 1998. +[44] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington, and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. arXiv preprint arXiv:1711.00165, 2017. +[45] E. Levin, N. Tishby, and S. A. Solla. A statistical approach to learning and generalization in layered neural networks. Proc. of the IEEE – Special issue on Neural Networks, 1990. +[46] Yingzhen Li and Richard E Turner. Gradient estimators for implicit models. arXiv preprint arXiv:1705.07107, 2017. +[47] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-ofdistribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017. +[48] Chang Liu, Jingwei Zhuo, Pengyu Cheng, Ruiyi Zhang, and Jun Zhu. Understanding and accelerating particle-based variational inference. In International Conference on Machine Learning, pages 4082–4092. PMLR, 2019. +[49] Jeremiah Zhe Liu, Zi Lin, Shreyas Padhy, Dustin Tran, Tania Bedrax-Weiss, and Balaji Lakshminarayanan. Simple and principled uncertainty estimation with deterministic deep learning via distance awareness. arXiv preprint arXiv:2006.10108, 2020. +[50] Qiang Liu. Stein variational gradient descent as gradient flow. In Advances in neural information processing systems, pages 3115–3123, 2017. +[51] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. In Advances in neural information processing systems, pages 2378–2386, 2016. +[52] Qiang Liu, Jason Lee, and Michael Jordan. A kernelized stein discrepancy for goodness-of-fit tests. In International conference on machine learning, pages 276–284. PMLR, 2016. +[53] Clare Lyle, Lisa Schut, Robin Ru, Yarin Gal, and Mark van der Wilk. A bayesian perspective on training speed and model selection. Advances in Neural Information Processing Systems, 33, 2020. +[54] David JC MacKay. A practical bayesian framework for backpropagation networks. Neural computation, 4(3):448–472, 1992. +[55] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. +[56] Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017. +[57] Alexander G de G Matthews, Jiri Hron, Richard E Turner, and Zoubin Ghahramani. Samplethen-optimize posterior sampling for bayesian linear models. In NeurIPS Workshop on Advances in Approximate Bayesian Inference, 2017. +[58] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 29, 2015. +[59] Radford M Neal. Bayesian learning for neural networks. 1995. +[60] Radford M Neal. Bayesian learning for neural networks, volume 118. Springer Science & Business Media, 2012. +[61] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. +[62] Ian Osband, Benjamin Van Roy, Daniel J Russo, Zheng Wen, et al. Deep exploration via randomized value functions. J. Mach. Learn. Res., 20(124):1–62, 2019. + +[63] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian Henning, and Benjamin F Grewe. Neural networks with late-phase weights. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum? id=C0qJUx5dxFb. + +[64] Felix Otto. The geometry of dissipative evolution equations: the porous medium equation. 2001. + +[65] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift. In Advances in Neural Information Processing Systems, pages 13991–14002, 2019. + +[66] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065–1076, 1962. + +[67] Alexandre Rame and Matthieu Cord. Dice: Diversity in deep ensembles via conditional redundancy adversarial estimation. arXiv preprint arXiv:2101.05544, 2021. + +[68] Geoffrey Roeder, Luke Metz, and Diederik P Kingma. On linear identifiability of learned representations. arXiv preprint arXiv:2007.00810, 2020. + +[69] Jiaxin Shi, Shengyang Sun, and Jun Zhu. A spectral approach to gradient estimation for implicit distributions. In International Conference on Machine Learning, pages 4644–4653. PMLR, 2018. + +[70] Radhey S Singh. Improvement on some known nonparametric uniformly consistent estimators of derivatives of a density. The Annals of Statistics, pages 394–399, 1977. + +[71] Shengyang Sun, Guodong Zhang, Jiaxin Shi, and Roger Grosse. Functional variational bayesian neural networks. arXiv preprint arXiv:1903.05779, 2019. + +[72] Jakub Swiatkowski, Kevin Roth, Bastiaan S Veeling, Linh Tran, Joshua V Dillon, Stephan Mandt, Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. The k-tied normal distribution: A compact parameterization of gaussian mean field posteriors in bayesian neural networks. arXiv preprint arXiv:2002.02655, 2020. + +[73] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F. Grewe. Continual learning with hypernetworks. In International Conference on Learning Representations, 2020. URL https://arxiv.org/abs/1906.00695. + +[74] Maxime Wabartha, Audrey Durand, Vincent François-Lavet, and Joelle Pineau. Handling black swan events in deep learning with diversely extrapolated neural networks. In IJCAI, pages 2140–2147, 2020. + +[75] Yifei Wang, Peng Chen, and Wuchen Li. Projected wasserstein gradient descent for highdimensional bayesian inference. arXiv preprint arXiv:2102.06350, 2021. + +[76] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019. + +[77] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011. + +[78] Yeming Wen, Dustin Tran, and Jimmy Ba. Batchensemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715, 2020. + +[79] Florian Wenzel, Kevin Roth, Bastiaan S Veeling, Jakub Swiatkowski, Linh Tran, Stephan Mandt, ´ Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. How good is the bayes posterior in deep neural networks really? arXiv preprint arXiv:2002.02405, 2020. + +[80] Florian Wenzel, Jasper Snoek, Dustin Tran, and Rodolphe Jenatton. Hyperparameter ensembles for robustness and uncertainty quantification. In Advances in Neural Information Processing Systems, 2020. + +[81] Christopher KI Williams. Computation with infinite neural networks. Neural Computation, 10 (5):1203–1216, 1998. +[82] Andrew Gordon Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic perspective of generalization. arXiv preprint arXiv:2002.08791, 2020. +[83] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. +[84] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016. +[85] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019. \ No newline at end of file diff --git a/parse/train/LAKplpLMbP8/LAKplpLMbP8_content_list.json b/parse/train/LAKplpLMbP8/LAKplpLMbP8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c99047df4faf9a1a3cf8a0e0abab3c55896dde07 --- /dev/null +++ b/parse/train/LAKplpLMbP8/LAKplpLMbP8_content_list.json @@ -0,0 +1,1690 @@ +[ + { + "type": "text", + "text": "Repulsive Deep Ensembles are Bayesian ", + "text_level": 1, + "bbox": [ + 254, + 122, + 743, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Francesco D’Angelo ETH Zürich Zürich, Switzerland dngfra@gmail.com ", + "bbox": [ + 285, + 200, + 429, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vincent Fortuin \nETH Zürich \nZürich, Switzerland \nfortuin@inf.ethz.ch ", + "bbox": [ + 547, + 202, + 712, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 292, + 535, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep ensembles have recently gained popularity in the deep learning community for their conceptual simplicity and efficiency. However, maintaining functional diversity between ensemble members that are independently trained with gradient descent is challenging. This can lead to pathologies when adding more ensemble members, such as a saturation of the ensemble performance, which converges to the performance of a single model. Moreover, this does not only affect the quality of its predictions, but even more so the uncertainty estimates of the ensemble, and thus its performance on out-of-distribution data. We hypothesize that this limitation can be overcome by discouraging different ensemble members from collapsing to the same function. To this end, we introduce a kernelized repulsive term in the update rule of the deep ensembles. We show that this simple modification not only enforces and maintains diversity among the members but, even more importantly, transforms the maximum a posteriori inference into proper Bayesian inference. Namely, we show that the training dynamics of our proposed repulsive ensembles follow a Wasserstein gradient flow of the KL divergence to the true posterior. We study repulsive terms in weight and function space and empirically compare their performance to standard ensembles and Bayesian baselines on synthetic and real-world prediction tasks. ", + "bbox": [ + 233, + 325, + 766, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction ", + "text_level": 1, + "bbox": [ + 174, + 599, + 310, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There have been many recent advances on the theoretical properties of sampling algorithms for approximate Bayesian inference, which changed our interpretation and understanding of them. Particularly worth mentioning is the work of Jordan et al. [38], who reinterpret Markov Chain Monte Carlo (MCMC) as a gradient flow of the KL divergence over the Wasserstein space of probability measures. This new formulation allowed for a deeper understanding of approximate inference methods but also inspired the inception of new and more efficient inference strategies. Following this direction, Liu and Wang $\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }$ recently proposed the Stein Variational Gradient Descent (SVGD) method to perform approximate Wasserstein gradient descent. Conceptually, this method, which belongs to the family of particle-optimization variational inference (POVI), introduces a repulsive force through a kernel acting in the parameter space to evolve a set of samples towards high-density regions of the target distribution without collapsing to a point estimate. ", + "bbox": [ + 174, + 631, + 825, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another method which has achieved great success recently are ensembles of neural networks (socalled deep ensembles), which work well both in terms of predictive performance $\\pm 2 \\sqrt { 8 0 }$ as well as uncertainty estimation $\\begin{array} { r l } { { \\| 6 5 \\| } } \\end{array}$ , and have also been proposed as a way to perform approximate inference in Bayesian neural networks [82, 36]. That being said, while they might allow for the averaging of predictions over several hypotheses, they do not offer any guarantees for the diversity between those hypotheses nor do they provably converge to the true Bayesian posterior under any meaningful limit. In this work, we show how the introduction of a repulsive term between the members in the ensemble, inspired by SVGD, not only naïvely guarantees the diversity among the members, avoiding their collapse in parameter space, but also allows for a reformulation of the method as a gradient flow of the KL divergence in the Wasserstein space of distributions. It thus allows to endow deep ensembles with convergence guarantees to the true Bayesian posterior. ", + "bbox": [ + 174, + 790, + 825, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c87b35667fe31cb24e44c323fe8bffeb4df13996d7df41a00bfd504161843062.jpg", + "image_caption": [ + "Figure 1: BNN 1D regression. The function-space methods (SVGD and WGD) approach the HMC posterior more closely, while the standard deep ensembles and weight-space methods fail to properly account for the uncertainty, especially the in-between uncertainty. " + ], + "image_footnote": [], + "bbox": [ + 196, + 85, + 805, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 327, + 821, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An additional problem is that BNN inference in weight space can lead to degenerate solutions, due to the overparametrization of these models. That is, several samples could have very different weights but map to the same function, thus giving a false sense of diversity in the ensemble. This property, that we will refer to as non-identifiability of neural networks (see Appendix $\\boxed { \\mathrm { A } }$ , can lead to redundancies in the posterior distribution. It implies that methods like MCMC sampling, deep ensembles, and SVGD waste computation in local modes that account for equivalent functions. Predictive distributions approximated using samples from these modes do not improve over a simple point estimate and lead to a poor uncertainty estimation. Following this idea, Wang et al. $\\mathbb { \\left[ \\left] \\right. \\right]} 2 $ introduced a new method to extend POVI methods to function space, overcoming this limitation. Here, we also study an update rule that allows for an approximation of the gradient flow of the KL divergence in function space in our proposed repulsive ensembles. ", + "bbox": [ + 173, + 375, + 825, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We make the following contributions: ", + "bbox": [ + 176, + 534, + 421, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We derive several different repulsion terms that can be added as regularizers to the gradient updates of deep ensembles to endow them with Bayesian convergence properties. \n• We show that these terms approximate Wasserstein gradient flows of the KL divergence and can be used both in weight space and function space. \n• We compare these proposed methods theoretically to standard deep ensembles and SVGD and highlight their different guarantees. \n• We assess all these methods on synthetic and real-world deep learning tasks and show that our proposed repulsive ensembles can achieve competitive performance and improved uncertainty estimation. ", + "bbox": [ + 217, + 560, + 825, + 705 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Repulsive Deep Ensembles ", + "text_level": 1, + "bbox": [ + 174, + 732, + 428, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In supervised deep learning, we typically consider a likelihood function $p ( \\pmb { y } | f ( \\pmb { x } ; \\mathbf { w } ) )$ (e.g., Gaussian for regression or Categorical for classification) parameterized by a neural network $f ( \\pmb { x } ; \\mathbf { w } )$ and training data $\\textit { D } = \\{ ( \\boldsymbol { \\mathbf { { x } } } _ { i } , \\boldsymbol { \\mathbf { { y } } } _ { i } ) \\} _ { i = 1 } ^ { n }$ with $\\textbf { \\textit { x } } \\in \\textbf { \\textit { X } }$ and $\\textbf { \\textit { y } } \\in \\textbf { \\textit { y } }$ . In Bayesian neural networks (BNNs), we are interested in the posterior distribution of all likely networks given by $\\begin{array} { r } { p ( \\mathbf { w } | \\mathcal { D } ) \\propto \\prod _ { i = 1 } ^ { n } p ( \\pmb { y } _ { i } | f ( \\pmb { x } _ { i } ; \\mathbf { w } ) ) p ( \\mathbf { w } ) } \\end{array}$ , where $p ( \\mathbf { w } )$ is the prior distribution over weights. Crucially, when making a prediction on a test point $\\mathbf { \\boldsymbol { x } } ^ { * }$ , in the Bayesian approach we do not only use a single parameter $\\widehat { \\bf w }$ to predict ${ \\pmb y } ^ { * } = f ( { \\pmb x } ^ { * } ; \\widehat { \\mathbf { w } } )$ , but we marginalize over the whole posterior, thus taking all possible explanations of the data into account: ", + "bbox": [ + 173, + 763, + 826, + 876 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/ea58bf85c2893e1394390263ae7f97e3c6d3270c2715a883b91057d7c53c61ba.jpg", + "text": "$$\np ( \\pmb { y } ^ { * } | \\pmb { x } ^ { * } , \\mathcal { D } ) = \\int p ( \\pmb { y } ^ { * } | f ( \\pmb { x } ^ { * } ; \\mathbf { w } ) ) p ( \\mathbf { w } | \\mathcal { D } ) \\mathrm { d } \\mathbf { w }\n$$", + "text_format": "latex", + "bbox": [ + 343, + 883, + 653, + 916 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While approximating the posterior of Bayesian neural networks (or sampling from it) is a challenging task, performing maximum a posteriori (MAP) estimation, which corresponds to finding the mode of the posterior, is usually simple. Ensembles of neural networks use the non-convexity of the MAP optimization problem to create a collection of $K$ independent—and possibly different— solutions. Considering $n$ weight configurations of a neural network $\\{ { \\bf w } _ { i } \\} _ { i = 1 } ^ { n }$ with ${ \\bf w } _ { i } \\in \\mathbb { R } ^ { d }$ , the dynamics of the ensemble under the gradient of the posterior lead to the following update rule at iteration $t$ : ", + "bbox": [ + 173, + 90, + 826, + 175 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/dec175f126457d7570141af68aa834ee6aa688fa0490be6f7e2b319f4c374697.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbf { w } _ { i } ^ { t + 1 } \\mathbf { w } _ { i } ^ { t } + \\epsilon _ { t } \\phi ( \\mathbf { w } _ { i } ^ { t } ) } \\\\ { \\mathrm { w i t h } \\quad } & { \\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 372, + 179, + 625, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with step size $\\epsilon _ { t }$ . Ensemble methods have a long history [e.g., 45, 26, 6] and were recently revisited for neural networks $\\pm 2 \\|$ and coined deep ensembles. The predictions of the different members are combined to create a predictive distribution by using the solutions to compute the Bayesian model average (BMA) in Eq. $\\mathbb { D }$ . Recent works $\\mathbb { \\lVert \\rVert }$ have shown that deep ensembles can outperform some of the Bayesian approaches for uncertainty estimation. Even more recently, Wilson and Izmailov $\\pmb { \\Vert 8 2 \\Vert }$ argued that deep ensembles can be considered a compelling approach to Bayesian model averaging. Despite these ideas, the ability of deep ensembles to efficiently average over multiple hypotheses and to explore the functional landscape of the posterior distribution studied in $[ \\overline { { 1 8 } } ]$ does not guarantee sampling from the right distribution. Indeed, the additional Langevin noise introduced in $\\mathbb { [ \\breve { ] { ] \\mathrm { ~ Z ~ Z ~ } ] } } }$ , which is not considered in deep ensembles, is crucial to ensure samples from the true Bayesian posterior. ", + "bbox": [ + 173, + 224, + 825, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From a practical standpoint, since the quality of an ensemble hinges on the diversity of its members, many methods were recently proposed to improve this diversity without compromising the individual accuracy. For instance, Wenzel et al. [80] propose hyper-deep ensembles that combine deep networks with different hyperparameters. Similarly, cyclical learning-rate schedules can explore several local minima for the ensemble members $\\pmb { \\mathbb { B 3 } }$ . Alternatively, Rame and Cord $ { \\mathbb { I } } { \\mathbb { I } }$ proposed an informationtheoretic framework to avoid redundancy in the members and Oswald et al. $\\overline { { \\| 6 3 \\| } }$ studied possible interactions between members based on weight sharing. However, the absence of a constraint that prevents particles from converging to the same mode limits the possibility of improvement by introducing more ensemble members. This means that any hopes to converge to different modes must exclusively rely on: ", + "bbox": [ + 174, + 371, + 826, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. the randomness of the initialization \n2. the noise in the estimation of the gradients due to minibatching \n3. the number of local optima that might be reached during gradient descent. ", + "bbox": [ + 210, + 520, + 717, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, the recent study of Geiger et al. $\\pmb { \\left. \\pmb { \\left. \\bar { 2 5 } \\right. } \\right. }$ showed how the empirical test error of the ensemble converges to the one of a single trained model when the number of parameters goes to infinity, leading to deterioration of the performance. In other words, the bigger the model, the harder it is to maintain diversity in the ensemble and avoid collapse to the same solution. This is intuitively due to the fact that bigger models are less sensitive to the initialization. Namely, in order for them to get stuck in a local minimum, they must have second derivatives that are positive simultaneously in all directions. As the number of hidden units gets larger, this becomes less likely. ", + "bbox": [ + 174, + 582, + 826, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Repulsive force in weight space ", + "text_level": 1, + "bbox": [ + 176, + 695, + 429, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the aforementioned limitations of standard deep ensembles, we introduce, inspired by SVGD $\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }$ , a deep ensemble with members that interact with each other through a repulsive component. Using a kernel function to model this interaction, the single models repel each other based on their position in the weight space, so that two members can never assume the same weights. Considering a stationary kernel $\\bar { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\to \\mathbb { R }$ acting in the parameter space of the neural network, a repulsive term $\\mathcal { R }$ can be parameterized through its gradient: ", + "bbox": [ + 173, + 720, + 826, + 805 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/52631e0ed6e1a61e9e013d38c7585cc4709cd85b719f28114ba1570f9f671c2a.jpg", + "text": "$$\n\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 299, + 810, + 699, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To get an intuition for the behavior of this repulsive term and its gradients, we can consider the RBF kernel $\\begin{array} { r } { k ( \\mathbf { w } _ { i } , \\mathbf { w } _ { j } ) = \\exp \\big ( - \\frac { 1 } { h } | | \\mathbf { w } _ { i } - \\mathbf { w } _ { j } | | ^ { 2 } \\big ) } \\end{array}$ with lengthscale $h$ and notice how its gradient ", + "bbox": [ + 173, + 849, + 825, + 881 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/f23cff6e6b3dc3c8e61b222583d646eef0fa3e0251332d45f77faee6e2130a87.jpg", + "text": "$$\n\\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) = \\frac { 2 } { h } ( \\mathbf { w } _ { j } ^ { t } - \\mathbf { w } _ { i } ^ { t } ) k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } )\n$$", + "text_format": "latex", + "bbox": [ + 357, + 886, + 638, + 916 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "drives $\\mathbf { w } _ { i }$ away from its neighboring members ${ \\bf w } _ { j }$ , thus creating a repulsive effect. Naturally, not all the choices of $\\mathcal { R }$ induce this effect. One of the simplest formulations to obtain it is via a linear combination of the kernel gradients scaled by a positive factor, that is, $\\begin{array} { r } { \\beta \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}$ with $\\beta \\in \\mathbb { R } _ { * } ^ { + }$ . We will see in Section $\\begin{array} { l } { 3 } \\\\ { . } \\end{array}$ how the choice of $\\beta$ can be justified in order to obtain convergence to the Bayesian posterior together with alternative possible formulations of $\\mathcal { R }$ that preserve this convergence. ", + "bbox": [ + 173, + 90, + 826, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2 Repulsive force in function space ", + "text_level": 1, + "bbox": [ + 176, + 193, + 441, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome the aforementioned overparameterization issue, the update in Eq. $\\textcircled { 3 }$ can be formulated in function space instead of weight space. Let $f : \\mathbf { w } \\mapsto f ( \\cdot ; \\mathbf { w } )$ be the map that maps a configuration of weights $\\mathbf { w } \\in \\mathbb { R } ^ { d }$ to the corresponding neural network regression function and denote as $f _ { i } : = f ( \\cdot ; { \\mathbf w } _ { i } )$ the function with a certain configuration of weights $\\mathbf { w } _ { i }$ . We can now consider $n$ particles in function space $\\{ f _ { i } \\} _ { i = 1 } ^ { n }$ with $\\pmb { f } \\in \\mathcal { F }$ and model their interaction with a general positive definite kernel $k ( \\cdot , \\cdot )$ . We also consider the implicit functional likelihood $p ( \\pmb { y } | \\pmb { x } , \\pmb { f } )$ , determined by the measure $p ( \\pmb { y } | \\pmb { x } , \\mathbf { w } )$ in the weight space, as well as the functional prior $p ( f )$ , which can either be defined separately (e.g., using a GP) or modeled as a push-forward measure of the weight-space prior $p ( \\mathbf { w } )$ . Together, they determine the posterior in function space $p ( \\pmb { f } | \\mathcal { D } )$ . The functional evolution of a particle can then be written as: ", + "bbox": [ + 173, + 218, + 826, + 358 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/6e9ccd562a52a4d7474eb2dbb571b766cb26754c555d1761309dd914491e04b8.jpg", + "text": "$$\n\\begin{array} { r l } & { f _ { i } ^ { t + 1 } \\gets f _ { i } ^ { t } + \\epsilon _ { t } \\phi ( f _ { i } ^ { t } ) } \\\\ { \\mathrm { t h } \\quad } & { \\phi ( f _ { i } ^ { t } ) = \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 297, + 362, + 720, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, computing the update in function space is neither tractable nor practical, which is why two additional considerations are needed. The first one regards the infinite dimensionality of function space, which we circumvent using a canonical projection into a subspace: ", + "bbox": [ + 178, + 422, + 823, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 (Canonical projection). For any $A \\subset { \\mathcal { X } }$ , we define $\\pi _ { A } : \\mathbb { R } ^ { \\mathcal { X } } \\to \\mathbb { R } ^ { A }$ as the canonical projection onto $A$ , that is, $\\pi _ { A } ( f ) = \\{ f ( a ) \\} _ { a \\in A }$ . ", + "bbox": [ + 174, + 468, + 821, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In other words, the kernel will not be evaluated directly in function space, but on the projection $k \\big ( \\pi _ { B } ( f ) , \\pi _ { B } ( f ^ { \\prime } ) \\big )$ , with $B$ being a subset of the input space given by a batch of training data points. The second consideration is to project this update back into the parameter space and evolve a set of particles there, because ultimately we are interested in representing the functions by parameterized neural networks. For this purpose, we can use the Jacobian of the $i$ -th particle as a projector: ", + "bbox": [ + 174, + 507, + 825, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/d7dea1ea5314c734775cfe499bded9ea6e655f30206ef90295c8a3d5d82d37d3.jpg", + "text": "$$\n\\phi ( { \\mathbf w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial { \\mathbf w } _ { i } ^ { t } } \\right) ^ { \\top } \\left[ \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | { \\mathcal D } ) - { \\mathcal R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( \\pi _ { B } ( f _ { i } ^ { t } ) , \\pi _ { B } ( f _ { j } ^ { t } ) ) \\right\\} _ { j = 1 } ^ { n } \\right) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 230, + 593, + 767, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Comparison to Stein variational gradient descent ", + "text_level": 1, + "bbox": [ + 173, + 645, + 555, + 661 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that our update is reminiscent of SVGD $\\mathbb { \\left[ 5 1 \\right] }$ , which in parameter space can be written as: ", + "bbox": [ + 173, + 669, + 795, + 685 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/ed4ba25f1f792a3ef21feee9794a7aaba9ed6181f2313384dd85bba5b6760967.jpg", + "text": "$$\n\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\sum _ { j = 1 } ^ { n } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) + \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { j } ^ { t } } k ( \\mathbf { w } _ { j } ^ { t } , \\mathbf { w } _ { i } ^ { t } ) .\n$$", + "text_format": "latex", + "bbox": [ + 282, + 690, + 715, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to notice that here, the gradients are averaged across all the particles using the kernel matrix. Interestingly, SVGD can be asymptotically interpreted as gradient flow of the KL divergence under a new metric induced by the Stein operator [16, 50] (see Appendix D for more details). Moving the inference from parameter to function space $\\pmb { \\mathbb { Z } } 6 \\|$ leads to the update rule ", + "bbox": [ + 173, + 738, + 826, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/da910223233132ed75eba544433552d88de5b055ed7c20ab997ee811df84045c.jpg", + "text": "$$\n\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial \\mathbf { w } _ { i } ^ { t } } \\right) ^ { \\top } \\left( \\frac { 1 } { n } \\sum _ { j = 1 } ^ { n } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\nabla _ { f _ { j } ^ { t } } \\log p ( f _ { j } ^ { t } | \\mathcal { D } ) + \\nabla _ { f _ { j } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 251, + 799, + 746, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This way of averaging gradients using a kernel can be dangerous in high-dimensional settings, where kernel methods often suffer from the curse of dimensionality. Moreover, in Eq. $\\textcircled{6}$ , the posterior gradients of the particles are averaged using their similarity in weight space, which can be misleading in multi-modal posteriors. Worse yet, in Eq. $\\textcircled{7}$ , the gradients are averaged in function space and are then projected back using exclusively the $i$ -th Jacobian, which can be harmful given that it is not guaranteed that distances between functions evaluated on a subset of their input space resemble their true distance. Our proposed method, on the other hand, does not employ any averaging of the posterior gradients and thus comes closest to the true particle gradients in deep ensembles. ", + "bbox": [ + 174, + 854, + 825, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 90, + 825, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Repulsive deep ensembles are Bayesian ", + "text_level": 1, + "bbox": [ + 174, + 165, + 532, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "So far, we represented the repulsive force as a general function of the gradients of a kernel. In this section, we show how to determine the explicit form of the repulsive term, such that the resulting update rule is equivalent to the discretization of the gradient flow dynamics of the KL divergence in Wasserstein space. We begin by introducing the concepts of particle approximation and gradient flow. ", + "bbox": [ + 173, + 195, + 826, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Particle approximation ", + "text_level": 1, + "bbox": [ + 174, + 267, + 374, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A particle-based approximation of a target measure depends on a set of weighted samples $\\{ ( \\bar { x _ { i } } , w _ { i } ) \\} _ { i = 1 } ^ { n }$ , for which an empirical measure can be defined as ", + "bbox": [ + 173, + 292, + 823, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/b51a8eebc42b5a02f34979b6c46328fac936b6561c35ce4b3a464937651a8dd3.jpg", + "text": "$$\n\\rho ( x ) = \\sum _ { i = 1 } ^ { n } w _ { i } \\delta ( x - x _ { i } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 413, + 325, + 584, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta ( \\cdot )$ is the Dirac delta function and the weights $w _ { i }$ satisfy $w _ { i } \\in [ 0 , 1 ]$ and $\\textstyle \\sum _ { i = 1 } ^ { n } w _ { i } = 1$ . To approximate a target distribution $\\pi ( x )$ using the empirical measure, the particles and their weights need to be selected in a principled manner that minimizes some measure of distance between $\\pi ( x )$ and $\\rho ( x )$ (e.g., a set of $N$ samples with weights $w _ { i } = 1 / N$ obtained using an MCMC method). ", + "bbox": [ + 173, + 369, + 825, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Gradient flow in parameter space ", + "text_level": 1, + "bbox": [ + 176, + 441, + 446, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a smooth function $J : \\mathbb { R } ^ { d } \\mathbb { R }$ in Euclidean space, we can minimize it by creating a path that follows its negative gradient starting from some initial conditions $x _ { 0 }$ . The curve $x ( t )$ with starting point $x _ { 0 }$ described by that path is called gradient flow. The dynamics and evolution in time of a considered point in the space under this minimization problem can be described as the ODE1 ", + "bbox": [ + 173, + 465, + 825, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/4af9c8caebc8dd0f4bb240ad868ba49b8044f29f61003074e861ae4833cdfbee.jpg", + "text": "$$\n\\frac { d \\boldsymbol { x } } { d t } = - \\boldsymbol { \\nabla } J ( \\boldsymbol { x } ) .\n$$", + "text_format": "latex", + "bbox": [ + 442, + 526, + 555, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can extend this concept to the space of probability distributions (Wasserstein gradient flow) [3]. Let us consider the space of probability measures $\\mathcal { P } _ { 2 } ( \\mathcal { M } )$ , that is, the set of probability measures with finite second moments defined on the manifold $\\mathcal { M }$ : ", + "bbox": [ + 173, + 558, + 825, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/c1e5d77c719b4d2fa7a0247805d6e0144779c241f5db03b43a6f1d113276ff19.jpg", + "text": "$$\n\\mathcal { P } _ { 2 } ( \\mathcal { M } ) = \\left\\{ \\varphi : \\mathcal { M } \\to [ 0 , \\infty ) \\bigg | \\int _ { \\mathcal { M } } \\mathrm { d } \\varphi = 1 , ~ \\int _ { \\mathcal { M } } | x | ^ { 2 } \\varphi ( x ) \\mathrm { d } x < + \\infty \\right\\} .\n$$", + "text_format": "latex", + "bbox": [ + 259, + 601, + 736, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking $\\Pi ( \\mu , \\nu )$ as the set of joint probability measures with marginals $\\mu , \\nu$ , we can define the Wasserstein metric on the space $\\mathcal { P } _ { 2 } ( \\mathcal { M } )$ as: ", + "bbox": [ + 178, + 638, + 820, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/22875a9421a3c955c184f36bee03fad9988e4e0d847133375befe3e5951b4f25.jpg", + "text": "$$\nW _ { 2 } ^ { 2 } ( \\mu , \\nu ) = \\operatorname* { i n f } _ { \\pi \\in \\Pi ( \\mu , \\nu ) } \\int | x - y | ^ { 2 } { \\mathrm { d } } \\pi ( x , y ) .\n$$", + "text_format": "latex", + "bbox": [ + 354, + 670, + 642, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Considering the optimization problem of a functional $J : \\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\mathbb { R }$ , such as the $\\mathrm { K L }$ divergence between the particle approximation in Eq. $\\textcircled { 8 }$ and the target posterior $\\pi ( x )$ , ", + "bbox": [ + 173, + 707, + 823, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/202eca2cfd27979881321c7e7947951e3cc456330188130d02fcab3cc4c0abde.jpg", + "text": "$$\n\\operatorname* { i n f } _ { \\rho \\in \\mathcal { P } _ { 2 } ( \\mathcal { M } ) } D _ { K L } ( \\rho , \\pi ) = \\int _ { \\mathcal { M } } ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\rho ( x ) \\mathrm { d } x ,\n$$", + "text_format": "latex", + "bbox": [ + 305, + 739, + 691, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the evolution in time of the measure $\\rho$ under the equivalent of the gradient, the Wasserstein gradient flow, is described by the Liouville equation2 [38, 3, 64]: ", + "bbox": [ + 174, + 775, + 825, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/319c5dba11def8b47c62bbed7e518cec22d2bb6447df21ccd8c8f0ceacda0468.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( \\boldsymbol { x } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\big ( \\log \\rho ( \\boldsymbol { x } ) - \\log \\pi ( \\boldsymbol { x } ) \\big ) \\right) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 339, + 806, + 656, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla _ { \\delta \\rho } ^ { \\delta } D _ { K L } ( \\rho , \\pi ) \\ = : \\ \\nabla _ { \\mathcal { W } _ { 2 } } D _ { K L } ( \\rho , \\pi )$ is the Wasserstein gradient and the operator $\\frac { \\delta } { \\delta \\rho }$ $\\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\stackrel { \\cdot } { } \\mathbb { R }$ represents the functional derivative or first variation (see Appendix $\\boxed { \\mathsf { C } }$ for more details). In the particular case of the KL functional, we can recover the Fokker-Planck equation, ", + "bbox": [ + 173, + 89, + 826, + 136 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/bc5069e97a8be7662b9fea29c7dc30021ca7cea5d088b0d8d45da5c537562e37.jpg", + "text": "$$\n\\begin{array} { r l r } & { } & { \\frac { \\partial \\rho ( x ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( x ) \\nabla ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\right) } \\\\ & { } & { = - \\nabla \\cdot \\left( \\rho ( x ) \\nabla \\log \\pi ( x ) \\right) + \\nabla ^ { 2 } \\rho ( x ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 348, + 137, + 647, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "that admits as unique stationary distribution the posterior $\\pi ( x )$ . The deterministic particle dynamics ODE $\\left[ \\left[ 2 \\right] \\right]$ related to Eq. $\\mathbb { ( 1 1 ) }$ , namely mean-field Wasserstein dynamics, is then given by: ", + "bbox": [ + 173, + 189, + 823, + 217 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/dd2081dc85db6793f99ac9a48cbba2bf3fdd0b388e3458acd6fdadb5a243ac7b.jpg", + "text": "$$\n\\frac { d x } { d t } = - \\nabla \\big ( \\log \\rho ( x ) - \\log \\pi ( x ) \\big ) .\n$$", + "text_format": "latex", + "bbox": [ + 385, + 218, + 612, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering a discretization of Eq. (12) for a particle system $\\{ x \\} _ { i = 1 } ^ { n }$ and small stepsize $\\epsilon _ { t }$ , we can rewrite Eq. $\\mathbb { \\underline { { ( 1 2 ) } } }$ as: ", + "bbox": [ + 171, + 253, + 823, + 284 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/8a84c16f9a2296b238e38136d0db6478f4239c987d86ac3eef254ae1bde2a3a0.jpg", + "text": "$$\nx _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigl ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\nabla \\log \\rho ( x _ { i } ^ { t } ) \\bigr ) .\n$$", + "text_format": "latex", + "bbox": [ + 348, + 281, + 648, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unfortunately, we do not have access to the analytical form of the gradient $\\nabla \\log \\rho$ , so an approximation is needed. At this point, it is crucial to observe the similarity between the discretization of the Wasserstein gradient flow in Eq. $\\textcircled { 1 3 }$ and the repulsive update in Eq. $( 3 )$ to notice how, if the kernelized repulsion is an approximation of the gradient of the empirical particle measure, the update rule minimizes the KL divergence between the particle measure and the target posterior. Different sample-based approximations of the gradient that use a kernel function have been recently studied. The simplest one is given by the kernel density estimation (KDE) (details in Appendix $\\dot { \\mathrm { ~ E ~ } } )$ $\\begin{array} { r } { \\tilde { \\rho } _ { t } ( x ) = \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\dot { k } ( x , x _ { t } ^ { i } ) } \\end{array}$ , where $\\boldsymbol { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\dot { \\mathbb { R } }$ and the gradient of its log density is given by $\\mathbb { \\ m }$ : ", + "bbox": [ + 173, + 299, + 826, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/690e51dceea0054b390381bde15c30e9e687f9c712f52307c2e07985d9200e39.jpg", + "text": "$$\n\\nabla \\log \\rho ( x _ { i } ^ { t } ) \\approx \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } .\n$$", + "text_format": "latex", + "bbox": [ + 377, + 421, + 620, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using this approximation in Eq. $( 1 3 )$ we obtain: ", + "bbox": [ + 174, + 458, + 488, + 473 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/371ffb0715cb9a3b57e74ec4c78b9d2ce1313431a8e5f80ccc30135a0cfa4053.jpg", + "text": "$$\nx _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\Bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } \\Bigg ) ,\n$$", + "text_format": "latex", + "bbox": [ + 315, + 473, + 681, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, if we substitute the posterior for $\\pi$ , we obtain an expression for the repulsive force in Eq. $\\textcircled{3}$ . This shows that if the repulsive term in Eq. $( 3 )$ is the normalized sum of the gradients $\\mathcal { R } =$ $\\begin{array} { r } { \\left( \\bar { \\sum _ { j = 1 } ^ { n } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right) ^ { - 1 } \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}$ , we do not only encourage diversity of the ensemble members and thus avoid collapse, but surprisingly—in the asymptotic limit of , where the KDE approximation is exact $\\left[ \\left[ 6 6 \\right] \\right]$ —also converge to the true Bayesian posterior! ", + "bbox": [ + 173, + 512, + 825, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Nevertheless, approximating the gradient of the empirical measure with the KDE can lead to suboptimal performance, as already studied by Li and Turner $[ \\overline { { | 4 6 | } }$ . They instead introduced a new Stein gradient estimator (SGE) that offers better performance, while maintaining the same computational cost. Even more recently, Shi et al. $\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { A } } \\right. }$ introduced a spectral method for gradient estimation (SSGE), that also allows for a simple estimation on out-of-sample points. These two estimators can be used in Eq. $\\textcircled { 1 3 }$ , to formulate the following update rules with two alternative repulsive forces. The one using the Stein estimator, that we will call SGE-WGD, is: ", + "bbox": [ + 173, + 598, + 826, + 696 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/fc8beab1c532557f6b4c53d364e88ab075d4917c7571f2c182526373978fd6c4.jpg", + "text": "$$\nx _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) + \\sum _ { j = 1 } ^ { n } ( K + \\eta \\mathbb { I } ) _ { i j } ^ { - 1 } \\sum _ { k = 1 } ^ { n } \\nabla _ { x _ { k } ^ { t } } k ( x _ { k } ^ { t } , x _ { j } ^ { t } ) \\bigg ) ,\n$$", + "text_format": "latex", + "bbox": [ + 269, + 696, + 725, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $K$ is the kernel Gram matrix, $\\eta$ a small constant, and I the identity matrix. We can notice an important difference between KDE and SGE, in that the former is only considering the interaction of the $i$ -th particle being updated with all the others, while the latter is simultaneously considering also the interactions between the remaining particles. The spectral method, that we will call SSGE-WGD, leads to the following update rule: ", + "bbox": [ + 174, + 739, + 825, + 809 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/22b05cdc924e14a9f81460eed44ce5ae2949169ff1032670bb23204d0bc3d3f9.jpg", + "text": "$$\nx _ { t + 1 } ^ { i } = x _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { t } ^ { i } ) + \\sum _ { j = 1 } ^ { J } \\frac { 1 } { \\lambda _ { j } ^ { 2 } } \\sum _ { m = 1 } ^ { n } \\sum _ { k = 1 } ^ { n } u _ { j k } \\nabla _ { x _ { m } } k ( x _ { m } ^ { t } , x _ { k } ^ { t } ) \\cdot \\sum _ { l = 1 } ^ { n } u _ { j l } k ( x _ { i } ^ { t } , x _ { l } ^ { t } ) \\bigg )\n$$", + "text_format": "latex", + "bbox": [ + 197, + 809, + 772, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda _ { j }$ is the $j$ -th eigenvalue of the kernel matrix and $u _ { j k }$ is the $k$ -th component of the $j$ -th eigenvector. Computationally, both SSGE and SGE have a cost of $\\mathcal { O } ( M ^ { 3 } + M ^ { 2 } d )$ , with $M$ being the number of points and $d$ their dimensionality. SSGE has an additional cost for predictions of $\\mathcal { O } ( M ( d + J ) )$ , where $J$ is the number of eigenvalues. ", + "bbox": [ + 174, + 854, + 825, + 912 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/909c60e208223cd54336af828c72fc5777f5cbee78dd1b5dffc27345c64eca36.jpg", + "image_caption": [ + "Figure 2: Single Gaussian. We show samples from SVGD, KDE-WGD, SGE-WGD, and SSGEWGD (from left to right). The upper and right plots show the empirical one-dimensional marginal distributions obtained using KDE on the samples (red) and on the particles (blue). " + ], + "image_footnote": [], + "bbox": [ + 181, + 85, + 820, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3 Gradient flow in function space ", + "text_level": 1, + "bbox": [ + 174, + 287, + 429, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To theoretically justify the update rule introduced in function space in Eq. $\\textcircled{5}$ , we can rewrite the Liouville equation for the gradient flow in Eq. $\\textcircled { 1 1 }$ in function space as ", + "bbox": [ + 173, + 314, + 823, + 344 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/5c17155651b4b8a9222af9e3b20c3d19517ca6b37f38c6cba7442ed799860e6c.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( { \\pmb f } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\big ( \\log \\rho ( { \\pmb f } ) - \\log \\pi ( { \\pmb f } ) \\big ) \\right) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 336, + 352, + 663, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following this update, the mean field functional dynamics are ", + "bbox": [ + 173, + 429, + 580, + 444 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/6ec49c1866059fd222d901856c6fd87a5e5117e19a3da56886cb693721fd87d0.jpg", + "text": "$$\n\\frac { d \\pmb f } { d t } = - \\nabla \\big ( \\log \\rho ( \\pmb f ) - \\log \\pi ( \\pmb f ) \\big ) .\n$$", + "text_format": "latex", + "bbox": [ + 382, + 452, + 616, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the same KDE approximation as above, we can obtain a discretized evolution in function space and with it an explicit form for the repulsive force in Eq. $\\textcircled{4}$ as ", + "bbox": [ + 171, + 489, + 825, + 520 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/413dc39054cf2ffe58aac203b1f546f1178b0fd741a1e970d847b042475c73d1.jpg", + "text": "$$\n\\pmb { f } _ { t + 1 } ^ { i } = \\pmb { f } _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla _ { \\pmb { f } } \\log \\pi ( \\pmb { f } _ { t } ^ { i } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { \\pmb { f } _ { i } ^ { t } } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } \\bigg ) .\n$$", + "text_format": "latex", + "bbox": [ + 307, + 526, + 691, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The update rules using the SGE and SSGE approximations follow as for the parametric case. It is important to notice that this update rule requires the function space prior gradient: $\\nabla _ { f _ { j } } \\log { p ( f _ { j } | x , y ) } =$ $\\nabla _ { f _ { j } } \\log p ( \\pmb { y } | \\pmb { x } , \\pmb { f _ { j } } ) + \\nabla _ { f _ { j } } \\log p ( \\pmb { f _ { j } } )$ . If one wants to use an implicit prior defined in weight space, an additional estimator is needed due to its analytical intractability. We again adopted the SSGE, introduced by Shi et al. $\\mathbb { \\lVert \\boldsymbol { 6 9 } \\rVert }$ , which was already used for a similar purpose in Sun et al. $\\pmb { \\mathbb { Z } 1 }$ . It is also interesting to note that the update rule in Eq. $\\textcircled { 1 2 0 }$ readily allows for the use of alternative priors that have an analytical form, such as Gaussian processes. This is an important feature of our method that allows for an explicit encoding of function space properties that can be useful for example in achieving better out of distribution detection capabilities $\\checkmark$ . ", + "bbox": [ + 173, + 574, + 826, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4 The choice of the kernel ", + "text_level": 1, + "bbox": [ + 174, + 718, + 379, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A repulsive effect can always be created in the ensemble by means of the gradient of any kernel function that is measuring the similarity between two members. Nevertheless, it is important to keep in mind that to ensure the asymptotic convergence to the Bayesian posterior, the repulsive component must be a consistent estimator of the gradient in Eq. $( 1 3 )$ , as shown in Section $\\checkmark$ . Therefore, some important constraints over the kernel choice are needed. In particular, the SGE and SSGE need a kernel function belonging to the Stein class (see Shi et al. $\\dot { \\overline { { \\vert 6 9 \\vert } } }$ for more details). On the other hand, for the KDE, any symmetric probability density function can be used. On this subject, the work of Aggarwal et al. [1] has shown how the Manhattan distance metric (L1 norm) might be preferable over the Euclidean distance metric (L2 norm) for high-dimensional settings. We performed some additional experiments using the L1 norm (Laplace kernel) for the KDE but we could not observe any substantial difference compared to using the L2 norm. Further investigations regarding this hypothesis are left for future research. ", + "bbox": [ + 173, + 744, + 825, + 911 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/094ed7ba5611118de5688c3ddd1545168ced55dff3ababfdac0ee5e88615b714.jpg", + "image_caption": [ + "Figure 3: BNN 2D classification. We show the entropy of the predictive posteriors. Again, the function-space methods capture the uncertainty better than the weight-space ones, thus approaching the gold-standard HMC posterior. " + ], + "image_footnote": [], + "bbox": [ + 196, + 93, + 831, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4 Experiments ", + "text_level": 1, + "bbox": [ + 174, + 373, + 312, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we compare the different proposed WGD methods with deep ensembles and SVGD on synthetic sampling, regression, and classification tasks and real-world image classification tasks. We use an RBF kernel (except where otherwise specified) with the popular median heuristic $\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }$ to choose the kernel bandwidth. In our experiments, an adaptive bandwidth leads to better performance than fixing and tuning a single constant value for the entire evolution of the particles. We also quantitatively assess the uncertainty estimation of the methods in terms of calibration and OOD detection. In our experiments, we report the test accuracy, negative log-likelihood (NLL), and the expected calibration error (ECE) [58]. To assert the robustness on out-of-distribution (OOD) data, we report the ratio between predictive entropy on OOD and test data points $( H _ { o } / H _ { t } )$ , and the OOD detection area under the ROC curve AUROC(H) $\\mathbb { H }$ . Moreover, to assess the diversity of the ensemble generated by the different methods in function space, we measure the functional diversity using the model disagreement (MD) (details in Appendix $\\bar { \\bigtriangledown } \\bar { \\bigtriangledown }$ . In particular, we report the ratio between the average model disagreement on the OOD and test data points $( M D _ { o } / M D _ { t } )$ and additionally the OOD detection AUROC(MD) computed using this measure instead of the entropy. ", + "bbox": [ + 174, + 405, + 825, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sampling from synthetic distributions As a sanity check, we first assessed the ability of our different approximations for Wasserstein gradient descent (using KDE, SGE, and SSGE) to sample from a two-dimensional Gaussian distribution (Figure $^ { 2 ) }$ . We see that our SGE-WGD, SSGE-WGD and the SVGD fit the target almost perfectly. We also tested the different methods in a more complex two-dimensional Funnel distribution $\\mathbb { \\left[ \\left[ 5 9 \\right] \\right] }$ and present the results in Figure $\\mathrm { F . l }$ in the Appendix. There, SGE-WGD and SVGD also perform best. ", + "bbox": [ + 173, + 616, + 825, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BNN 1D regression We then assessed the different methods in fitting a BNN posterior on a synthetically generated one-dimensional regression task. The results are reported in Figure 1 consisting of the mean prediction and $\\pm 1 , 2 , 3$ standard deviations of the predictive distribution. We can see that all methods performing inference in the weight space (DE, w-SVGD, WGD) are unable to capture the epistemic uncertainty between the two clusters of training data points. Conversely, the functional methods (f-SVGD, fWGD) are perfectly able to infer the diversity of the hypotheses in this region due to the lack of training evidence. They thereby achieve a predictive posterior that very closely resembles the one obtained with the gold-standard HMC sampling. ", + "bbox": [ + 173, + 714, + 825, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BNN 2D classification Next, we investigated the predictive performance and quality of uncertainty estimation of the methods in a two-dimensional synthetic classification setting. The results are displayed in Figure $\\textcircled { 3 }$ We can clearly observe that the weight-space methods are overconfident and do not capture the uncertainty well. Moreover, all the functions seems to collapse to the optimal classifier. These methods thus only account for uncertainty close to the decision boundaries and to the origin region, for which the uncertainty is purely aleatoric. In this setting, f-SVGD suffers from similar issues as the weight space methods, being overconfident away from the training data. Conversely, our fWGD methods are confident (low entropy) around the data but not out-of-distribution, thus representing the epistemic uncertainty better. This suggests that the functional diversity captured by this method naturally leads to a distance-aware uncertainty estimation $\\mathbb { H 9 } \\mathbb { L O }$ , a property that translates into confident predictions only in the proximity of the training data, allowing for a principled OOD detection. ", + "bbox": [ + 174, + 842, + 825, + 911 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c10e7e6db9d6afe7b283337a217b0daf6bdc8ff99b073aef9c58623cfb518d7c.jpg", + "table_caption": [ + "Table 1: BNN image classification. AUROC $\\mathbf { ( H ) }$ is the AUROC computed using the entropy whereas AUROC(MD) is computed using the model disagreement. $\\mathbf { H _ { o } / H _ { t } }$ is the ratio of the entropies on OOD and test points respectively and $\\mathbf { M D _ { o } / M D _ { t } }$ is the ratio for model disagreement. We see that the best accuracy is achieved by our WGD methods, while our fWGD methods yield the best OOD detection and funtional diversity. All our proposed methods improve over standard deep ensembles in terms of accuracy and diversity, highlighting the effect of our repulsion. " + ], + "table_footnote": [], + "table_body": "
PrroiiiieAUROC(H)AUROC(MD)AccuracyH/HtMD/MDtECENLL
Deep ensemble 40.958±0.0010.975±0.00191.122±0.0136.257±0.0056.394±0.0010.012±0.0010.129±0.001
SVGD [510.960±0.0010.973±0.00191.134±0.0246.315±0.0196.395±0.0180.014±0.0010.127±0.001
f-SVGD [760.956±0.0010.975±0.00189.884±0.0155.652±0.0096.531±0.0050.013±0.0010.150±0.001
hyper-DE [80]0.968±0.0010.981±0.00191.160±0.0076.682±0.0657.059±0.1520.014±0.0010.128±0.001
kde-WGD (ours)0.960±0.0010.970±0.00191.238±0.0196.587±0.0196.379±0.0180.014±0.0010.128±0.001
sge-WGD (ours)0.960±0.0010.970±0.00191.312±0.0166.562±0.0076.363±0.0090.012±0.0010.128±0.001
ssge-WGD (ours)0.968±0.0010.979±0.00191.198±0.0246.522±0.0096.610±0.0120.012±0.0010.130±0.001
kde-fWGD (ours)0.971±0.0010.980±0.00191.260±0.0117.079±0.0166.887±0.0150.015±0.0010.125±0.001
sge-fWGD (ours)0.969±0.0010.978±0.00191.192±0.0137.076±0.0046.900±0.0050.015±0.0010.125±0.001
ssge-fWGD (ours)0.971±0.0010.980±0.00191.240±0.0227.129±0.0066.951±0.0050.016±0.0010.124±0.001
CITIIIODeep ensemble 420.843±0.0040.736±0.00585,552±0.0762.244±0.0061.667±0.0080.049±0.0010.277±0.001
SVGD [510.825±0.0010.710±0.00285.142±0.0172.106±0.0031.567±0.0040.052±0.0010.287±0.001
fSVGD 回0.783±0.0010.712±0.00184.510±0.0311.968±0.0041.624±0.0030.049±0.0010.292±0.001
hyper-DE [80]0.789±0.0010.743±0.00184.743±0.0111.951±0.0101.690±0.0150.046±0.0010.288±0.001
kde-WGD (ours)0.838±0.0010.735±0.00485.904±0.0302.205±0.0031.661±0.0080.053±0.0010.276±0.001
sge-WGD (ours)0.837±0.0030.725±0.00485.792±0.0352.214±0.0101.634±0.0040.051±0.0010.275±0.001
ssge-WGD (ours)0.832±0.0030.731±0.00585.638±0.0382.182±0.0151.655±0.0010.049±0.0010.276±0.001
kde-fWGD (ours)0.791±0.0020.758±0.00284.888±0.0301.970±0.0041.749±0.0050.044±0.0010.282±0.001
sge-fWGD (ours)0.795±0.0010.754±0.00284.766±0.0601.984±0.0031.729±0.0020.047±0.0010.288±0.001
ssge-fWGD (ours)0.792±0.0020.752±0.00284.762±0.0341.970±0.0061.723±0.0050.046±0.0010.286±0.001
", + "bbox": [ + 174, + 180, + 825, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 421, + 825, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "FashionMNIST classification Moving on to real-world data, we used an image classification setting using the FashionMNIST dataset $\\dot { \\left[ \\left| 8 3 \\right| \\right] }$ for training and the MNIST dataset $\\bar { \\mathbb { E } 3 } \\mathbb { I }$ as an out-ofdistribution (OOD) task. The results are reported in Table $^ 1 { } _ { . }$ (top). We can see that all our methods improve upon standard deep ensembles and SVGD, highlighting the effectiveness of our proposed repulsion terms when training neural network ensembles. In particular, the sge-WGD offers the best accuracy, whereas the methods in function space all offer a better OOD detection. This is probably due to the fact that these methods achieve a higher entropy ratio and functional diversity measured via the model disagreement when compared to their weight-space counterparts. Interestingly, they also reach the lowest NLL values. We can also notice how the model disagreement (MD) not only serves its purpose as a metric for the functional heterogeneity of the ensemble but also allows for a better OOD detection in comparison to the entropy. To the best of our knowledge, this insight has not been described before, although it has been used in continual learning $\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }$ . Interestingly, using this metric, the hyper-deep ensemble $[ \\textcircled { 8 0 } ]$ shows OOD detection performance comparable with our repulsive ensemble in function space. ", + "bbox": [ + 173, + 536, + 825, + 729 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "CIFAR classification Finally, we use a ResNet32 architecture $\\mathbb { \\left| \\left[ 2 9 \\right] \\right| }$ on CIFAR-10 [41] with the SVHN dataset $\\mathbb { \\left[ 6 1 \\right] }$ as OOD data. The results are reported in Table 1 (bottom). We can see that in this case, the weight-space methods achieve better performance in accuracy and OOD detection using the entropy than the ones in function space. Nevertheless, all our repulsive ensembles improve functional diversity, accuracy, and OOD detection when compared to standard SVGD, whereas the standard deep ensemble achieves the best OOD detection using the entropy. ", + "bbox": [ + 174, + 746, + 825, + 829 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Related Work ", + "text_level": 1, + "bbox": [ + 174, + 849, + 321, + 867 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The theoretical and empirical properties of SVGD have been well studied [40, 48, 13] and it can also be seen as a Wasserstein gradient flow of the KL divergence in the Stein geometry [16, 50] (see Appendix D for more details). Interestingly, a gradient flow interpretation is also possible for (stochastic gradient) MCMC-type algorithms $\\dot { \\left. \\overline { { 4 8 } } \\right. }$ , which can be unified under a general particle inference framework [10]. Moreover, our Wasserstein gradient descent using the SGE approximation can also be derived using an alternative formulation as a gradient flow with smoothed test functions [48]. A projected version of WGD has been studied in Wang et al. $\\mathbb { \\left. \\overline { { \\boldsymbol { \\mathscr { Q } } \\boldsymbol { \\cdot } \\boldsymbol { \\cdot } } } \\right. }$ , which could also be readily applied in our framework. Besides particle methods, Bayesian neural networks $\\pm \\pm \\infty$ have gained popularity recently [79, 22, 19, 36], using modern MCMC [59, 79, 22, 24, 21] and variational inference techniques [5, 72, 17, 34]. On the other hand, ensemble methods have also been extensively studied [42, 18, 82, 23, 80, 32, 85, 78].Moreover, repulsive interactions between the members have also been studied in Wabartha et al. [74]. Moreover, providing Bayesian interpretations for deep ensembles has been previously attempted through the lenses of stationary SGD distributions $[ \\sqrt { 5 6 } , \\sqrt ] { 8 } ]$ , ensembles of linear models $\\dot { \\bigtriangledown } 5 7 \\big |$ , additional random functions $\\boxed { 6 2 } , \\boxed { 1 2 } , \\boxed { 2 7 }$ , approximate inference $\\lVert 8 2 \\rVert$ , Stein variational inference [15], and marginal likelihood lower bounds $\\bar { \\| 5 3 \\| }$ , and ensembles have also been shown to provide good approximations to the true BNN posterior in some settings $\\left[ \\left[ 3 6 \\right] \\right]$ . Furthermore, variational inference in function space has recently gained attention $\\mathbb { \\ m }$ and the limitations of the KL divergence have been studied in Burt et al. [7]. ", + "bbox": [ + 174, + 882, + 821, + 911 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 90, + 825, + 313 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 Conclusion ", + "text_level": 1, + "bbox": [ + 174, + 330, + 299, + 348 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We have presented a simple and principled way to improve upon standard deep ensemble methods. To this end, we have shown that the introduction of a kernelized repulsion between members of the ensemble not only improves the accuracy of the predictions but—even more importantly—can be seen as Wasserstein gradient descent on the KL divergence, thus transforming the MAP inference of deep ensembles into proper Bayesian inference. Moreover, we have shown that incorporating functional repulsion between ensemble members can improve the quality of the estimated uncertainties on simple synthetic examples and OOD detection on real-world data and can approach the true Bayesian posterior more closely. ", + "bbox": [ + 174, + 361, + 825, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "In future work, it will be interesting to study the impact of the Jacobian in the fWGD update and its implications on the Liouville equation in more detail, also compared to other neural network Jacobian methods, such as neural tangent kernels $\\pmb { \\Vert 3 7 } \\Vert$ and generalized Gauss-Newton approximations $\\boldsymbol { \\left[ \\left[ 3 5 \\right] \\right] }$ . Moreover, it would be interesting to derive explicit convergence bounds for our proposed method and compare them to the existing bounds for SVGD [40]. ", + "bbox": [ + 174, + 478, + 825, + 547 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Acknowledgments ", + "text_level": 1, + "bbox": [ + 176, + 563, + 303, + 578 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "VF would like to acknowledge financial support from the Strategic Focus Area “Personalized Health and Related Technologies” of the ETH Domain through the grant #2017-110 and from the Swiss Data Science Center through a PhD fellowship. We thank Florian Wenzel, Alexander Immer, Andrew Gordon Wilson, Pavel Izmailov, Christian Henning, and Johannes von Oswald for helpful discussions. We also thank Dr. Sheldon Cooper, Dr. Leonard Hofstadter and Penny Hofstadter for their support and inspiration. ", + "bbox": [ + 174, + 588, + 825, + 671 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "References ", + "text_level": 1, + "bbox": [ + 174, + 690, + 266, + 707 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of distance metrics in high dimensional space. In International conference on database theory, pages 420–434. Springer, 2001. \n[2] Luigi Ambrosio and Gianluca Crippa. Continuity equations and ode flows with non-smooth velocity. Proceedings of the Royal Society of Edinburgh: Section A Mathematics, 144(6): 1191–1244, 2014. \n[3] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré. Gradient flows: in metric spaces and in the space of probability measures. Springer Science & Business Media, 2008. \n[4] Vijay Badrinarayanan, Bamdev Mishra, and Roberto Cipolla. Symmetry-invariant optimization in deep networks. arXiv preprint arXiv:1511.01754, 2015. \n[5] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty in neural networks. arXiv preprint arXiv:1505.05424, 2015. ", + "bbox": [ + 179, + 713, + 826, + 911 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "[6] Leo Breiman. Bagging predictors. Machine learning, 24(2):123–140, 1996. ", + "bbox": [ + 179, + 90, + 707, + 107 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[7] David R Burt, Sebastian W Ober, Adrià Garriga-Alonso, and Mark van der Wilk. Understanding variational inference in function-space. arXiv preprint arXiv:2011.09421, 2020. ", + "bbox": [ + 178, + 114, + 823, + 145 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[8] Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference, converges to limit cycles for deep networks. In 2018 Information Theory and Applications Workshop (ITA), pages 1–10. IEEE, 2018. ", + "bbox": [ + 178, + 152, + 823, + 195 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[9] An Mei Chen, Haw-minn Lu, and Robert Hecht-Nielsen. On the geometry of feedforward neural network error surfaces. Neural computation, 5(6):910–927, 1993. ", + "bbox": [ + 178, + 204, + 825, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[10] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liqun Chen. A unified particleoptimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659, 2018. ", + "bbox": [ + 173, + 242, + 825, + 285 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[11] Jiefeng Chen, Xi Wu, Yingyu Liang, Somesh Jha, et al. Robust out-of-distribution detection in neural networks. arXiv preprint arXiv:2003.09711, 2020. ", + "bbox": [ + 173, + 294, + 825, + 323 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[12] Kamil Ciosek, Vincent Fortuin, Ryota Tomioka, Katja Hofmann, and Richard Turner. Conservative uncertainty estimation by fitting prior networks. In International Conference on Learning Representations, 2019. ", + "bbox": [ + 174, + 332, + 823, + 375 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[13] Francesco D’Angelo and Vincent Fortuin. Annealed stein variational gradient descent. arXiv preprint arXiv:2101.09815, 2021. ", + "bbox": [ + 173, + 382, + 821, + 412 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[14] Francesco D’Angelo and Christian Henning. Uncertainty-based out-of-distribution detection requires suitable function space priors. arXiv preprint arXiv:2110.06020, 2021. ", + "bbox": [ + 173, + 421, + 821, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[15] Francesco D’Angelo, Vincent Fortuin, and Florian Wenzel. On stein variational neural network ensembles. arXiv preprint arXiv:2106.10760, 2021. ", + "bbox": [ + 174, + 458, + 823, + 488 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[16] A Duncan, Nikolas Nuesken, and Lukasz Szpruch. On the geometry of stein variational gradient descent. arXiv preprint arXiv:1912.00894, 2019. ", + "bbox": [ + 173, + 496, + 823, + 526 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[17] Michael W Dusenberry, Ghassen Jerfel, Yeming Wen, Yi-an Ma, Jasper Snoek, Katherine Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets with rank-1 factors. arXiv preprint arXiv:2005.07186, 2020. ", + "bbox": [ + 174, + 535, + 823, + 578 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[18] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape perspective. arXiv preprint arXiv:1912.02757, 2019. ", + "bbox": [ + 174, + 587, + 823, + 616 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[19] Vincent Fortuin. Priors in bayesian deep learning: A review. arXiv preprint arXiv:2105.06868, 2021. ", + "bbox": [ + 173, + 625, + 823, + 654 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[20] Vincent Fortuin, Mark Collier, Florian Wenzel, James Allingham, Jeremiah Liu, Dustin Tran, Balaji Lakshminarayanan, Jesse Berent, Rodolphe Jenatton, and Effrosyni Kokiopoulou. Deep classifiers with label noise modeling and distance awareness. arXiv preprint arXiv:2110.02609, 2021. ", + "bbox": [ + 174, + 661, + 826, + 718 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[21] Vincent Fortuin, Adrià Garriga-Alonso, Mark van der Wilk, and Laurence Aitchison. Bnnpriors: A library for bayesian neural network inference with different prior distributions. Software Impacts, page 100079, 2021. ", + "bbox": [ + 173, + 727, + 823, + 771 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[22] Vincent Fortuin, Adrià Garriga-Alonso, Florian Wenzel, Gunnar Rätsch, Richard Turner, Mark van der Wilk, and Laurence Aitchison. Bayesian neural network priors revisited. arXiv preprint arXiv:2102.06571, 2021. ", + "bbox": [ + 174, + 779, + 823, + 821 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[23] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry Vetrov, and Andrew Gordon Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. arXiv preprint arXiv:1802.10026, 2018. ", + "bbox": [ + 174, + 830, + 821, + 875 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[24] Adrià Garriga-Alonso and Vincent Fortuin. Exact langevin dynamics with stochastic gradients. arXiv preprint arXiv:2102.01691, 2021. ", + "bbox": [ + 173, + 882, + 820, + 911 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "[25] Mario Geiger, Arthur Jacot, Stefano Spigler, Franck Gabriel, Levent Sagun, Stéphane d’Ascoli, Giulio Biroli, Clément Hongler, and Matthieu Wyart. Scaling description of generalization with number of parameters in deep learning. Journal of Statistical Mechanics: Theory and Experiment, 2020(2):023401, 2020. \n[26] L. K. Hansen and P. Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell., 12(10):993–1001, October 1990. ISSN 0162-8828. doi: 10.1109/34.58871. URL https://doi.org/10.1109/34.58871. \n[27] Bobby He, Balaji Lakshminarayanan, and Yee Whye Teh. Bayesian deep ensembles via the neural tangent kernel. arXiv preprint arXiv:2007.05864, 2020. \n[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015. \n[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. \n[30] Robert Hecht-Nielsen. On the algebraic structure of feedforward network weight spaces. In Advanced Neural Computers, pages 129–135. Elsevier, 1990. \n[31] Christian Henning, Maria R Cervera, Francesco D’Angelo, Johannes von Oswald, Regina Traber, Benjamin Ehret, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Posterior meta-replay for continual learning. arXiv preprint arXiv:2103.01133, 2021. \n[32] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger. Snapshot ensembles: Train 1, get m for free. arXiv preprint arXiv:1704.00109, 2017. \n[33] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger. Snapshot ensembles: Train 1, get M for free. CoRR, abs/1704.00109, 2017. URL http: //arxiv.org/abs/1704.00109. \n[34] Alexander Immer, Matthias Bauer, Vincent Fortuin, Gunnar Rätsch, and Mohammad Emtiyaz Khan. Scalable marginal likelihood estimation for model selection in deep learning. arXiv preprint arXiv:2104.04975, 2021. \n[35] Alexander Immer, Maciej Korzepa, and Matthias Bauer. Improving predictions of bayesian neural nets via local linearization. In International Conference on Artificial Intelligence and Statistics, pages 703–711. PMLR, 2021. \n[36] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Wilson. What are bayesian neural network posteriors really like? arXiv preprint arXiv:2104.14421, 2021. \n[37] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: convergence and generalization in neural networks. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 8580–8589, 2018. \n[38] Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker– planck equation. SIAM journal on mathematical analysis, 29(1):1–17, 1998. \n[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. \n[40] Anna Korba, Adil Salim, Michael Arbel, Giulia Luise, and Arthur Gretton. A non-asymptotic analysis for stein variational gradient descent. Advances in Neural Information Processing Systems, 33, 2020. \n[41] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009. \n[42] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in neural information processing systems, pages 6402–6413, 2017. \n[43] Yann LeCun. The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/, 1998. \n[44] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington, and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. arXiv preprint arXiv:1711.00165, 2017. \n[45] E. Levin, N. Tishby, and S. A. Solla. A statistical approach to learning and generalization in layered neural networks. Proc. of the IEEE – Special issue on Neural Networks, 1990. \n[46] Yingzhen Li and Richard E Turner. Gradient estimators for implicit models. arXiv preprint arXiv:1705.07107, 2017. \n[47] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-ofdistribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017. \n[48] Chang Liu, Jingwei Zhuo, Pengyu Cheng, Ruiyi Zhang, and Jun Zhu. Understanding and accelerating particle-based variational inference. In International Conference on Machine Learning, pages 4082–4092. PMLR, 2019. \n[49] Jeremiah Zhe Liu, Zi Lin, Shreyas Padhy, Dustin Tran, Tania Bedrax-Weiss, and Balaji Lakshminarayanan. Simple and principled uncertainty estimation with deterministic deep learning via distance awareness. arXiv preprint arXiv:2006.10108, 2020. \n[50] Qiang Liu. Stein variational gradient descent as gradient flow. In Advances in neural information processing systems, pages 3115–3123, 2017. \n[51] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian inference algorithm. In Advances in neural information processing systems, pages 2378–2386, 2016. \n[52] Qiang Liu, Jason Lee, and Michael Jordan. A kernelized stein discrepancy for goodness-of-fit tests. In International conference on machine learning, pages 276–284. PMLR, 2016. \n[53] Clare Lyle, Lisa Schut, Robin Ru, Yarin Gal, and Mark van der Wilk. A bayesian perspective on training speed and model selection. Advances in Neural Information Processing Systems, 33, 2020. \n[54] David JC MacKay. A practical bayesian framework for backpropagation networks. Neural computation, 4(3):448–472, 1992. \n[55] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. \n[56] Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017. \n[57] Alexander G de G Matthews, Jiri Hron, Richard E Turner, and Zoubin Ghahramani. Samplethen-optimize posterior sampling for bayesian linear models. In NeurIPS Workshop on Advances in Approximate Bayesian Inference, 2017. \n[58] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 29, 2015. \n[59] Radford M Neal. Bayesian learning for neural networks. 1995. \n[60] Radford M Neal. Bayesian learning for neural networks, volume 118. Springer Science & Business Media, 2012. \n[61] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. \n[62] Ian Osband, Benjamin Van Roy, Daniel J Russo, Zheng Wen, et al. Deep exploration via randomized value functions. J. Mach. Learn. Res., 20(124):1–62, 2019. ", + "bbox": [ + 168, + 42, + 828, + 915 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 56, + 828, + 919 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "[63] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian Henning, and Benjamin F Grewe. Neural networks with late-phase weights. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum? id=C0qJUx5dxFb. ", + "bbox": [ + 174, + 90, + 826, + 148 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[64] Felix Otto. The geometry of dissipative evolution equations: the porous medium equation. 2001. ", + "bbox": [ + 169, + 155, + 825, + 184 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[65] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift. In Advances in Neural Information Processing Systems, pages 13991–14002, 2019. ", + "bbox": [ + 174, + 193, + 825, + 248 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[66] Emanuel Parzen. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065–1076, 1962. ", + "bbox": [ + 173, + 256, + 823, + 285 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[67] Alexandre Rame and Matthieu Cord. Dice: Diversity in deep ensembles via conditional redundancy adversarial estimation. arXiv preprint arXiv:2101.05544, 2021. ", + "bbox": [ + 173, + 292, + 823, + 321 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[68] Geoffrey Roeder, Luke Metz, and Diederik P Kingma. On linear identifiability of learned representations. arXiv preprint arXiv:2007.00810, 2020. ", + "bbox": [ + 173, + 329, + 823, + 359 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[69] Jiaxin Shi, Shengyang Sun, and Jun Zhu. A spectral approach to gradient estimation for implicit distributions. In International Conference on Machine Learning, pages 4644–4653. PMLR, 2018. ", + "bbox": [ + 173, + 366, + 825, + 409 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[70] Radhey S Singh. Improvement on some known nonparametric uniformly consistent estimators of derivatives of a density. The Annals of Statistics, pages 394–399, 1977. ", + "bbox": [ + 173, + 416, + 823, + 446 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[71] Shengyang Sun, Guodong Zhang, Jiaxin Shi, and Roger Grosse. Functional variational bayesian neural networks. arXiv preprint arXiv:1903.05779, 2019. ", + "bbox": [ + 173, + 454, + 823, + 483 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[72] Jakub Swiatkowski, Kevin Roth, Bastiaan S Veeling, Linh Tran, Joshua V Dillon, Stephan Mandt, Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. The k-tied normal distribution: A compact parameterization of gaussian mean field posteriors in bayesian neural networks. arXiv preprint arXiv:2002.02655, 2020. ", + "bbox": [ + 174, + 491, + 826, + 546 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[73] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F. Grewe. Continual learning with hypernetworks. In International Conference on Learning Representations, 2020. URL https://arxiv.org/abs/1906.00695. ", + "bbox": [ + 178, + 555, + 823, + 598 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[74] Maxime Wabartha, Audrey Durand, Vincent François-Lavet, and Joelle Pineau. Handling black swan events in deep learning with diversely extrapolated neural networks. In IJCAI, pages 2140–2147, 2020. ", + "bbox": [ + 176, + 606, + 823, + 647 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[75] Yifei Wang, Peng Chen, and Wuchen Li. Projected wasserstein gradient descent for highdimensional bayesian inference. arXiv preprint arXiv:2102.06350, 2021. ", + "bbox": [ + 174, + 656, + 823, + 685 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[76] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019. ", + "bbox": [ + 173, + 693, + 825, + 722 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[77] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics. In Proceedings of the 28th international conference on machine learning (ICML-11), pages 681–688, 2011. ", + "bbox": [ + 176, + 729, + 825, + 772 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[78] Yeming Wen, Dustin Tran, and Jimmy Ba. Batchensemble: an alternative approach to efficient ensemble and lifelong learning. arXiv preprint arXiv:2002.06715, 2020. ", + "bbox": [ + 171, + 780, + 825, + 809 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[79] Florian Wenzel, Kevin Roth, Bastiaan S Veeling, Jakub Swiatkowski, Linh Tran, Stephan Mandt, ´ Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. How good is the bayes posterior in deep neural networks really? arXiv preprint arXiv:2002.02405, 2020. ", + "bbox": [ + 173, + 818, + 825, + 861 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[80] Florian Wenzel, Jasper Snoek, Dustin Tran, and Rodolphe Jenatton. Hyperparameter ensembles for robustness and uncertainty quantification. In Advances in Neural Information Processing Systems, 2020. ", + "bbox": [ + 174, + 869, + 825, + 911 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "[81] Christopher KI Williams. Computation with infinite neural networks. Neural Computation, 10 (5):1203–1216, 1998. \n[82] Andrew Gordon Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic perspective of generalization. arXiv preprint arXiv:2002.08791, 2020. \n[83] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. \n[84] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016. \n[85] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclical stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019. ", + "bbox": [ + 173, + 90, + 828, + 271 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/parse/train/LAKplpLMbP8/LAKplpLMbP8_middle.json b/parse/train/LAKplpLMbP8/LAKplpLMbP8_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..a935f021b6d1534c6c6c7cd3c536886861250c0a --- /dev/null +++ b/parse/train/LAKplpLMbP8/LAKplpLMbP8_middle.json @@ -0,0 +1,41100 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 156, + 97, + 455, + 117 + ], + "lines": [ + { + "bbox": [ + 156, + 96, + 455, + 119 + ], + "spans": [ + { + "bbox": [ + 156, + 96, + 455, + 119 + ], + "score": 1.0, + "content": "Repulsive Deep Ensembles are Bayesian", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 175, + 159, + 263, + 204 + ], + "lines": [ + { + "bbox": [ + 174, + 159, + 263, + 172 + ], + "spans": [ + { + "bbox": [ + 174, + 159, + 263, + 172 + ], + "score": 1.0, + "content": "Francesco D’Angelo", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 192, + 169, + 245, + 182 + ], + "spans": [ + { + "bbox": [ + 192, + 169, + 245, + 182 + ], + "score": 1.0, + "content": "ETH Zürich", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 177, + 180, + 260, + 193 + ], + "spans": [ + { + "bbox": [ + 177, + 180, + 260, + 193 + ], + "score": 1.0, + "content": "Zürich, Switzerland", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 192, + 262, + 205 + ], + "spans": [ + { + "bbox": [ + 176, + 192, + 262, + 205 + ], + "score": 1.0, + "content": "dngfra@gmail.com", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "text", + "bbox": [ + 335, + 160, + 436, + 203 + ], + "lines": [ + { + "bbox": [ + 351, + 160, + 420, + 170 + ], + "spans": [ + { + "bbox": [ + 351, + 160, + 420, + 170 + ], + "score": 1.0, + "content": "Vincent Fortuin", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 359, + 169, + 411, + 182 + ], + "spans": [ + { + "bbox": [ + 359, + 169, + 411, + 182 + ], + "score": 1.0, + "content": "ETH Zürich", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 345, + 182, + 426, + 192 + ], + "spans": [ + { + "bbox": [ + 345, + 182, + 426, + 192 + ], + "score": 1.0, + "content": "Zürich, Switzerland", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 335, + 193, + 436, + 203 + ], + "spans": [ + { + "bbox": [ + 335, + 193, + 436, + 203 + ], + "score": 1.0, + "content": "fortuin@inf.ethz.ch", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.0 + }, + { + "type": "title", + "bbox": [ + 283, + 232, + 328, + 244 + ], + "lines": [ + { + "bbox": [ + 282, + 232, + 330, + 245 + ], + "spans": [ + { + "bbox": [ + 282, + 232, + 330, + 245 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 143, + 258, + 469, + 454 + ], + "lines": [ + { + "bbox": [ + 141, + 257, + 469, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 271 + ], + "score": 1.0, + "content": "Deep ensembles have recently gained popularity in the deep learning community", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 267, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 281 + ], + "score": 1.0, + "content": "for their conceptual simplicity and efficiency. However, maintaining functional", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 292 + ], + "score": 1.0, + "content": "diversity between ensemble members that are independently trained with gradient", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "descent is challenging. This can lead to pathologies when adding more ensemble", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 302, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 469, + 313 + ], + "score": 1.0, + "content": "members, such as a saturation of the ensemble performance, which converges to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "score": 1.0, + "content": "the performance of a single model. Moreover, this does not only affect the quality", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "of its predictions, but even more so the uncertainty estimates of the ensemble, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 334, + 469, + 346 + ], + "spans": [ + { + "bbox": [ + 142, + 334, + 469, + 346 + ], + "score": 1.0, + "content": "thus its performance on out-of-distribution data. We hypothesize that this limitation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 344, + 469, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 358 + ], + "score": 1.0, + "content": "can be overcome by discouraging different ensemble members from collapsing to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "score": 1.0, + "content": "the same function. To this end, we introduce a kernelized repulsive term in the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "update rule of the deep ensembles. We show that this simple modification not only", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "score": 1.0, + "content": "enforces and maintains diversity among the members but, even more importantly,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "score": 1.0, + "content": "transforms the maximum a posteriori inference into proper Bayesian inference.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 398, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 398, + 470, + 412 + ], + "score": 1.0, + "content": "Namely, we show that the training dynamics of our proposed repulsive ensembles", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "score": 1.0, + "content": "follow a Wasserstein gradient flow of the KL divergence to the true posterior.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "score": 1.0, + "content": "We study repulsive terms in weight and function space and empirically compare", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 432, + 470, + 445 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 470, + 445 + ], + "score": 1.0, + "content": "their performance to standard ensembles and Bayesian baselines on synthetic and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 443, + 254, + 455 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 254, + 455 + ], + "score": 1.0, + "content": "real-world prediction tasks.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 107, + 475, + 190, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 192, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 192, + 490 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "There have been many recent advances on the theoretical properties of sampling algorithms for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "approximate Bayesian inference, which changed our interpretation and understanding of them.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 521, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 535 + ], + "score": 1.0, + "content": "Particularly worth mentioning is the work of Jordan et al. [38], who reinterpret Markov Chain Monte", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "score": 1.0, + "content": "Carlo (MCMC) as a gradient flow of the KL divergence over the Wasserstein space of probability", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "measures. This new formulation allowed for a deeper understanding of approximate inference", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "methods but also inspired the inception of new and more efficient inference strategies. Following", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 222, + 578 + ], + "score": 1.0, + "content": "this direction, Liu and Wang", + "type": "text" + }, + { + "bbox": [ + 222, + 565, + 240, + 577 + ], + "score": 0.4, + "content": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "recently proposed the Stein Variational Gradient Descent (SVGD)", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "method to perform approximate Wasserstein gradient descent. Conceptually, this method, which", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "belongs to the family of particle-optimization variational inference (POVI), introduces a repulsive", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "force through a kernel acting in the parameter space to evolve a set of samples towards high-density", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 391, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 391, + 622 + ], + "score": 1.0, + "content": "regions of the target distribution without collapsing to a point estimate.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 507, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 507, + 639 + ], + "score": 1.0, + "content": "Another method which has achieved great success recently are ensembles of neural networks (so-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 430, + 649 + ], + "score": 1.0, + "content": "called deep ensembles), which work well both in terms of predictive performance", + "type": "text" + }, + { + "bbox": [ + 430, + 636, + 463, + 648 + ], + "score": 0.54, + "content": "\\pm 2 \\sqrt { 8 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 636, + 506, + 649 + ], + "score": 1.0, + "content": "as well as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 196, + 660 + ], + "score": 1.0, + "content": "uncertainty estimation", + "type": "text" + }, + { + "bbox": [ + 197, + 647, + 214, + 659 + ], + "score": 0.62, + "content": "\\begin{array} { r l } { { \\| 6 5 \\| } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 648, + 506, + 660 + ], + "score": 1.0, + "content": ", and have also been proposed as a way to perform approximate inference", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "in Bayesian neural networks [82, 36]. That being said, while they might allow for the averaging of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 669, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 681 + ], + "score": 1.0, + "content": "predictions over several hypotheses, they do not offer any guarantees for the diversity between those", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "score": 1.0, + "content": "hypotheses nor do they provably converge to the true Bayesian posterior under any meaningful limit.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 506, + 703 + ], + "score": 1.0, + "content": "In this work, we show how the introduction of a repulsive term between the members in the ensemble,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 703, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 714 + ], + "score": 1.0, + "content": "inspired by SVGD, not only naïvely guarantees the diversity among the members, avoiding their", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 731, + 385, + 741 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 387, + 744 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 387, + 744 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 156, + 97, + 455, + 117 + ], + "lines": [ + { + "bbox": [ + 156, + 96, + 455, + 119 + ], + "spans": [ + { + "bbox": [ + 156, + 96, + 455, + 119 + ], + "score": 1.0, + "content": "Repulsive Deep Ensembles are Bayesian", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 175, + 159, + 263, + 204 + ], + "lines": [ + { + "bbox": [ + 174, + 159, + 263, + 172 + ], + "spans": [ + { + "bbox": [ + 174, + 159, + 263, + 172 + ], + "score": 1.0, + "content": "Francesco D’Angelo", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 192, + 169, + 245, + 182 + ], + "spans": [ + { + "bbox": [ + 192, + 169, + 245, + 182 + ], + "score": 1.0, + "content": "ETH Zürich", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 177, + 180, + 260, + 193 + ], + "spans": [ + { + "bbox": [ + 177, + 180, + 260, + 193 + ], + "score": 1.0, + "content": "Zürich, Switzerland", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 192, + 262, + 205 + ], + "spans": [ + { + "bbox": [ + 176, + 192, + 262, + 205 + ], + "score": 1.0, + "content": "dngfra@gmail.com", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.0, + "bbox_fs": [ + 174, + 159, + 263, + 205 + ] + }, + { + "type": "list", + "bbox": [ + 335, + 160, + 436, + 203 + ], + "lines": [ + { + "bbox": [ + 351, + 160, + 420, + 170 + ], + "spans": [ + { + "bbox": [ + 351, + 160, + 420, + 170 + ], + "score": 1.0, + "content": "Vincent Fortuin", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 359, + 169, + 411, + 182 + ], + "spans": [ + { + "bbox": [ + 359, + 169, + 411, + 182 + ], + "score": 1.0, + "content": "ETH Zürich", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 345, + 182, + 426, + 192 + ], + "spans": [ + { + "bbox": [ + 345, + 182, + 426, + 192 + ], + "score": 1.0, + "content": "Zürich, Switzerland", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 335, + 193, + 436, + 203 + ], + "spans": [ + { + "bbox": [ + 335, + 193, + 436, + 203 + ], + "score": 1.0, + "content": "fortuin@inf.ethz.ch", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + } + ], + "index": 5.0, + "bbox_fs": [ + 335, + 160, + 436, + 203 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 232, + 328, + 244 + ], + "lines": [ + { + "bbox": [ + 282, + 232, + 330, + 245 + ], + "spans": [ + { + "bbox": [ + 282, + 232, + 330, + 245 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 143, + 258, + 469, + 454 + ], + "lines": [ + { + "bbox": [ + 141, + 257, + 469, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 271 + ], + "score": 1.0, + "content": "Deep ensembles have recently gained popularity in the deep learning community", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 267, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 281 + ], + "score": 1.0, + "content": "for their conceptual simplicity and efficiency. However, maintaining functional", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 278, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 292 + ], + "score": 1.0, + "content": "diversity between ensemble members that are independently trained with gradient", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "descent is challenging. This can lead to pathologies when adding more ensemble", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 302, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 469, + 313 + ], + "score": 1.0, + "content": "members, such as a saturation of the ensemble performance, which converges to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "score": 1.0, + "content": "the performance of a single model. Moreover, this does not only affect the quality", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "of its predictions, but even more so the uncertainty estimates of the ensemble, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 334, + 469, + 346 + ], + "spans": [ + { + "bbox": [ + 142, + 334, + 469, + 346 + ], + "score": 1.0, + "content": "thus its performance on out-of-distribution data. We hypothesize that this limitation", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 344, + 469, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 469, + 358 + ], + "score": 1.0, + "content": "can be overcome by discouraging different ensemble members from collapsing to", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "spans": [ + { + "bbox": [ + 142, + 356, + 469, + 367 + ], + "score": 1.0, + "content": "the same function. To this end, we introduce a kernelized repulsive term in the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 470, + 379 + ], + "score": 1.0, + "content": "update rule of the deep ensembles. We show that this simple modification not only", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 377, + 470, + 390 + ], + "score": 1.0, + "content": "enforces and maintains diversity among the members but, even more importantly,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 470, + 401 + ], + "score": 1.0, + "content": "transforms the maximum a posteriori inference into proper Bayesian inference.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 398, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 398, + 470, + 412 + ], + "score": 1.0, + "content": "Namely, we show that the training dynamics of our proposed repulsive ensembles", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "score": 1.0, + "content": "follow a Wasserstein gradient flow of the KL divergence to the true posterior.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "score": 1.0, + "content": "We study repulsive terms in weight and function space and empirically compare", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 432, + 470, + 445 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 470, + 445 + ], + "score": 1.0, + "content": "their performance to standard ensembles and Bayesian baselines on synthetic and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 443, + 254, + 455 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 254, + 455 + ], + "score": 1.0, + "content": "real-world prediction tasks.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 18.5, + "bbox_fs": [ + 141, + 257, + 470, + 455 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 475, + 190, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 192, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 192, + 490 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "There have been many recent advances on the theoretical properties of sampling algorithms for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 524 + ], + "score": 1.0, + "content": "approximate Bayesian inference, which changed our interpretation and understanding of them.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 521, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 535 + ], + "score": 1.0, + "content": "Particularly worth mentioning is the work of Jordan et al. [38], who reinterpret Markov Chain Monte", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 547 + ], + "score": 1.0, + "content": "Carlo (MCMC) as a gradient flow of the KL divergence over the Wasserstein space of probability", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 557 + ], + "score": 1.0, + "content": "measures. This new formulation allowed for a deeper understanding of approximate inference", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "methods but also inspired the inception of new and more efficient inference strategies. Following", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 222, + 578 + ], + "score": 1.0, + "content": "this direction, Liu and Wang", + "type": "text" + }, + { + "bbox": [ + 222, + 565, + 240, + 577 + ], + "score": 0.4, + "content": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "recently proposed the Stein Variational Gradient Descent (SVGD)", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "method to perform approximate Wasserstein gradient descent. Conceptually, this method, which", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "belongs to the family of particle-optimization variational inference (POVI), introduces a repulsive", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "force through a kernel acting in the parameter space to evolve a set of samples towards high-density", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 610, + 391, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 391, + 622 + ], + "score": 1.0, + "content": "regions of the target distribution without collapsing to a point estimate.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 500, + 506, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 713 + ], + "lines": [ + { + "bbox": [ + 105, + 625, + 507, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 507, + 639 + ], + "score": 1.0, + "content": "Another method which has achieved great success recently are ensembles of neural networks (so-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 430, + 649 + ], + "score": 1.0, + "content": "called deep ensembles), which work well both in terms of predictive performance", + "type": "text" + }, + { + "bbox": [ + 430, + 636, + 463, + 648 + ], + "score": 0.54, + "content": "\\pm 2 \\sqrt { 8 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 636, + 506, + 649 + ], + "score": 1.0, + "content": "as well as", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 647, + 506, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 196, + 660 + ], + "score": 1.0, + "content": "uncertainty estimation", + "type": "text" + }, + { + "bbox": [ + 197, + 647, + 214, + 659 + ], + "score": 0.62, + "content": "\\begin{array} { r l } { { \\| 6 5 \\| } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 648, + 506, + 660 + ], + "score": 1.0, + "content": ", and have also been proposed as a way to perform approximate inference", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "in Bayesian neural networks [82, 36]. That being said, while they might allow for the averaging of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 669, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 506, + 681 + ], + "score": 1.0, + "content": "predictions over several hypotheses, they do not offer any guarantees for the diversity between those", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 507, + 693 + ], + "score": 1.0, + "content": "hypotheses nor do they provably converge to the true Bayesian posterior under any meaningful limit.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 506, + 703 + ], + "score": 1.0, + "content": "In this work, we show how the introduction of a repulsive term between the members in the ensemble,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 703, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 714 + ], + "score": 1.0, + "content": "inspired by SVGD, not only naïvely guarantees the diversity among the members, avoiding their", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "collapse in parameter space, but also allows for a reformulation of the method as a gradient flow of", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "score": 1.0, + "content": "the KL divergence in the Wasserstein space of distributions. It thus allows to endow deep ensembles", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 281, + 345, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 345, + 294 + ], + "score": 1.0, + "content": "with convergence guarantees to the true Bayesian posterior.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 625, + 507, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 68, + 493, + 199 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 68, + 493, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 68, + 493, + 199 + ], + "spans": [ + { + "bbox": [ + 120, + 68, + 493, + 199 + ], + "score": 0.971, + "type": "image", + "image_path": "c87b35667fe31cb24e44c323fe8bffeb4df13996d7df41a00bfd504161843062.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 68, + 493, + 111.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 111.66666666666666, + 493, + 155.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 155.33333333333331, + 493, + 198.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 203, + 505, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "Figure 1: BNN 1D regression. The function-space methods (SVGD and WGD) approach the HMC", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "posterior more closely, while the standard deep ensembles and weight-space methods fail to properly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 225, + 371, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 371, + 239 + ], + "score": 1.0, + "content": "account for the uncertainty, especially the in-between uncertainty.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 259, + 503, + 293 + ], + "lines": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "collapse in parameter space, but also allows for a reformulation of the method as a gradient flow of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "score": 1.0, + "content": "the KL divergence in the Wasserstein space of distributions. It thus allows to endow deep ensembles", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 281, + 345, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 345, + 294 + ], + "score": 1.0, + "content": "with convergence guarantees to the true Bayesian posterior.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "An additional problem is that BNN inference in weight space can lead to degenerate solutions, due to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "the overparametrization of these models. That is, several samples could have very different weights", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 332 + ], + "score": 1.0, + "content": "but map to the same function, thus giving a false sense of diversity in the ensemble. This property, that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 381, + 343 + ], + "score": 1.0, + "content": "we will refer to as non-identifiability of neural networks (see Appendix", + "type": "text" + }, + { + "bbox": [ + 381, + 330, + 394, + 343 + ], + "score": 0.42, + "content": "\\boxed { \\mathrm { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 330, + 506, + 343 + ], + "score": 1.0, + "content": ", can lead to redundancies in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "the posterior distribution. It implies that methods like MCMC sampling, deep ensembles, and SVGD", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "score": 1.0, + "content": "waste computation in local modes that account for equivalent functions. Predictive distributions", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "approximated using samples from these modes do not improve over a simple point estimate and lead", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 374, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 371, + 386 + ], + "score": 1.0, + "content": "to a poor uncertainty estimation. Following this idea, Wang et al.", + "type": "text" + }, + { + "bbox": [ + 371, + 374, + 389, + 385 + ], + "score": 0.79, + "content": "\\mathbb { \\left[ \\left] \\right. \\right]} 2 ", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 374, + 506, + 386 + ], + "score": 1.0, + "content": "introduced a new method to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "extend POVI methods to function space, overcoming this limitation. Here, we also study an update", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "rule that allows for an approximation of the gradient flow of the KL divergence in function space in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 246, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 246, + 419 + ], + "score": 1.0, + "content": "our proposed repulsive ensembles.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 423, + 258, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 259, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 259, + 436 + ], + "score": 1.0, + "content": "We make the following contributions:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 133, + 444, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 132, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "• We derive several different repulsion terms that can be added as regularizers to the gradient", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 456, + 469, + 468 + ], + "spans": [ + { + "bbox": [ + 142, + 456, + 469, + 468 + ], + "score": 1.0, + "content": "updates of deep ensembles to endow them with Bayesian convergence properties.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 132, + 470, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 505, + 484 + ], + "score": 1.0, + "content": "• We show that these terms approximate Wasserstein gradient flows of the KL divergence and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 483, + 356, + 495 + ], + "spans": [ + { + "bbox": [ + 141, + 483, + 356, + 495 + ], + "score": 1.0, + "content": "can be used both in weight space and function space.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 132, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "• We compare these proposed methods theoretically to standard deep ensembles and SVGD", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 509, + 303, + 522 + ], + "spans": [ + { + "bbox": [ + 141, + 509, + 303, + 522 + ], + "score": 1.0, + "content": "and highlight their different guarantees.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 132, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "• We assess all these methods on synthetic and real-world deep learning tasks and show", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 142, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "that our proposed repulsive ensembles can achieve competitive performance and improved", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 548, + 235, + 560 + ], + "spans": [ + { + "bbox": [ + 142, + 548, + 235, + 560 + ], + "score": 1.0, + "content": "uncertainty estimation.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 107, + 580, + 262, + 594 + ], + "lines": [ + { + "bbox": [ + 104, + 578, + 263, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 578, + 263, + 597 + ], + "score": 1.0, + "content": "2 Repulsive Deep Ensembles", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 506, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 399, + 618 + ], + "score": 1.0, + "content": "In supervised deep learning, we typically consider a likelihood function", + "type": "text" + }, + { + "bbox": [ + 400, + 606, + 455, + 618 + ], + "score": 0.92, + "content": "p ( \\pmb { y } | f ( \\pmb { x } ; \\mathbf { w } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 605, + 507, + 618 + ], + "score": 1.0, + "content": "(e.g., Gaus-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 470, + 629 + ], + "score": 1.0, + "content": "sian for regression or Categorical for classification) parameterized by a neural network", + "type": "text" + }, + { + "bbox": [ + 471, + 617, + 505, + 629 + ], + "score": 0.93, + "content": "f ( \\pmb { x } ; \\mathbf { w } )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 626, + 508, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 184, + 642 + ], + "score": 1.0, + "content": "and training data", + "type": "text" + }, + { + "bbox": [ + 184, + 627, + 270, + 640 + ], + "score": 0.92, + "content": "\\textit { D } = \\{ ( \\boldsymbol { \\mathbf { { x } } } _ { i } , \\boldsymbol { \\mathbf { { y } } } _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 626, + 296, + 642 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 296, + 628, + 333, + 638 + ], + "score": 0.91, + "content": "\\textbf { \\textit { x } } \\in \\textbf { \\textit { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 626, + 356, + 642 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 628, + 392, + 639 + ], + "score": 0.91, + "content": "\\textbf { \\textit { y } } \\in \\textbf { \\textit { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 626, + 508, + 642 + ], + "score": 1.0, + "content": ". In Bayesian neural net-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "works (BNNs), we are interested in the posterior distribution of all likely networks given by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 649, + 507, + 663 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 266, + 662 + ], + "score": 0.91, + "content": "\\begin{array} { r } { p ( \\mathbf { w } | \\mathcal { D } ) \\propto \\prod _ { i = 1 } ^ { n } p ( \\pmb { y } _ { i } | f ( \\pmb { x } _ { i } ; \\mathbf { w } ) ) p ( \\mathbf { w } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 649, + 299, + 663 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 299, + 649, + 321, + 661 + ], + "score": 0.91, + "content": "p ( \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 649, + 507, + 663 + ], + "score": 1.0, + "content": "is the prior distribution over weights. Cru-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 303, + 673 + ], + "score": 1.0, + "content": "cially, when making a prediction on a test point", + "type": "text" + }, + { + "bbox": [ + 303, + 662, + 315, + 671 + ], + "score": 0.86, + "content": "\\mathbf { \\boldsymbol { x } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 660, + 506, + 673 + ], + "score": 1.0, + "content": ", in the Bayesian approach we do not only use", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 183, + 684 + ], + "score": 1.0, + "content": "a single parameter", + "type": "text" + }, + { + "bbox": [ + 183, + 672, + 193, + 681 + ], + "score": 0.76, + "content": "\\widehat { \\bf w }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 671, + 236, + 684 + ], + "score": 1.0, + "content": "to predict", + "type": "text" + }, + { + "bbox": [ + 236, + 671, + 298, + 683 + ], + "score": 0.93, + "content": "{ \\pmb y } ^ { * } = f ( { \\pmb x } ^ { * } ; \\widehat { \\mathbf { w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 671, + 505, + 684 + ], + "score": 1.0, + "content": ", but we marginalize over the whole posterior, thus", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 682, + 333, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 333, + 695 + ], + "score": 1.0, + "content": "taking all possible explanations of the data into account:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 700, + 400, + 726 + ], + "lines": [ + { + "bbox": [ + 210, + 700, + 400, + 726 + ], + "spans": [ + { + "bbox": [ + 210, + 700, + 400, + 726 + ], + "score": 0.94, + "content": "p ( \\pmb { y } ^ { * } | \\pmb { x } ^ { * } , \\mathcal { D } ) = \\int p ( \\pmb { y } ^ { * } | f ( \\pmb { x } ^ { * } ; \\mathbf { w } ) ) p ( \\mathbf { w } | \\mathcal { D } ) \\mathrm { d } \\mathbf { w }", + "type": "interline_equation", + "image_path": "ea58bf85c2893e1394390263ae7f97e3c6d3270c2715a883b91057d7c53c61ba.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 210, + 700, + 400, + 726 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 68, + 493, + 199 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 68, + 493, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 68, + 493, + 199 + ], + "spans": [ + { + "bbox": [ + 120, + 68, + 493, + 199 + ], + "score": 0.971, + "type": "image", + "image_path": "c87b35667fe31cb24e44c323fe8bffeb4df13996d7df41a00bfd504161843062.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 68, + 493, + 111.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 111.66666666666666, + 493, + 155.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 155.33333333333331, + 493, + 198.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 203, + 505, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "Figure 1: BNN 1D regression. The function-space methods (SVGD and WGD) approach the HMC", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "posterior more closely, while the standard deep ensembles and weight-space methods fail to properly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 225, + 371, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 371, + 239 + ], + "score": 1.0, + "content": "account for the uncertainty, especially the in-between uncertainty.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 259, + 503, + 293 + ], + "lines": [], + "index": 7, + "bbox_fs": [ + 105, + 259, + 505, + 294 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "An additional problem is that BNN inference in weight space can lead to degenerate solutions, due to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 309, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 321 + ], + "score": 1.0, + "content": "the overparametrization of these models. That is, several samples could have very different weights", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 506, + 332 + ], + "score": 1.0, + "content": "but map to the same function, thus giving a false sense of diversity in the ensemble. This property, that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 381, + 343 + ], + "score": 1.0, + "content": "we will refer to as non-identifiability of neural networks (see Appendix", + "type": "text" + }, + { + "bbox": [ + 381, + 330, + 394, + 343 + ], + "score": 0.42, + "content": "\\boxed { \\mathrm { A } }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 330, + 506, + 343 + ], + "score": 1.0, + "content": ", can lead to redundancies in", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "the posterior distribution. It implies that methods like MCMC sampling, deep ensembles, and SVGD", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "score": 1.0, + "content": "waste computation in local modes that account for equivalent functions. Predictive distributions", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "approximated using samples from these modes do not improve over a simple point estimate and lead", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 374, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 371, + 386 + ], + "score": 1.0, + "content": "to a poor uncertainty estimation. Following this idea, Wang et al.", + "type": "text" + }, + { + "bbox": [ + 371, + 374, + 389, + 385 + ], + "score": 0.79, + "content": "\\mathbb { \\left[ \\left] \\right. \\right]} 2 ", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 374, + 506, + 386 + ], + "score": 1.0, + "content": "introduced a new method to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "extend POVI methods to function space, overcoming this limitation. Here, we also study an update", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "rule that allows for an approximation of the gradient flow of the KL divergence in function space in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 407, + 246, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 246, + 419 + ], + "score": 1.0, + "content": "our proposed repulsive ensembles.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 298, + 506, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 423, + 258, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 422, + 259, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 259, + 436 + ], + "score": 1.0, + "content": "We make the following contributions:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 422, + 259, + 436 + ] + }, + { + "type": "list", + "bbox": [ + 133, + 444, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 132, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 132, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "• We derive several different repulsion terms that can be added as regularizers to the gradient", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 456, + 469, + 468 + ], + "spans": [ + { + "bbox": [ + 142, + 456, + 469, + 468 + ], + "score": 1.0, + "content": "updates of deep ensembles to endow them with Bayesian convergence properties.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 470, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 132, + 470, + 505, + 484 + ], + "score": 1.0, + "content": "• We show that these terms approximate Wasserstein gradient flows of the KL divergence and", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 483, + 356, + 495 + ], + "spans": [ + { + "bbox": [ + 141, + 483, + 356, + 495 + ], + "score": 1.0, + "content": "can be used both in weight space and function space.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 497, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 132, + 497, + 505, + 511 + ], + "score": 1.0, + "content": "• We compare these proposed methods theoretically to standard deep ensembles and SVGD", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 509, + 303, + 522 + ], + "spans": [ + { + "bbox": [ + 141, + 509, + 303, + 522 + ], + "score": 1.0, + "content": "and highlight their different guarantees.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "• We assess all these methods on synthetic and real-world deep learning tasks and show", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 142, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "that our proposed repulsive ensembles can achieve competitive performance and improved", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 548, + 235, + 560 + ], + "spans": [ + { + "bbox": [ + 142, + 548, + 235, + 560 + ], + "score": 1.0, + "content": "uncertainty estimation.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + } + ], + "index": 25, + "bbox_fs": [ + 132, + 443, + 505, + 560 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 580, + 262, + 594 + ], + "lines": [ + { + "bbox": [ + 104, + 578, + 263, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 578, + 263, + 597 + ], + "score": 1.0, + "content": "2 Repulsive Deep Ensembles", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 506, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 399, + 618 + ], + "score": 1.0, + "content": "In supervised deep learning, we typically consider a likelihood function", + "type": "text" + }, + { + "bbox": [ + 400, + 606, + 455, + 618 + ], + "score": 0.92, + "content": "p ( \\pmb { y } | f ( \\pmb { x } ; \\mathbf { w } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 605, + 507, + 618 + ], + "score": 1.0, + "content": "(e.g., Gaus-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 470, + 629 + ], + "score": 1.0, + "content": "sian for regression or Categorical for classification) parameterized by a neural network", + "type": "text" + }, + { + "bbox": [ + 471, + 617, + 505, + 629 + ], + "score": 0.93, + "content": "f ( \\pmb { x } ; \\mathbf { w } )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 626, + 508, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 184, + 642 + ], + "score": 1.0, + "content": "and training data", + "type": "text" + }, + { + "bbox": [ + 184, + 627, + 270, + 640 + ], + "score": 0.92, + "content": "\\textit { D } = \\{ ( \\boldsymbol { \\mathbf { { x } } } _ { i } , \\boldsymbol { \\mathbf { { y } } } _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 626, + 296, + 642 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 296, + 628, + 333, + 638 + ], + "score": 0.91, + "content": "\\textbf { \\textit { x } } \\in \\textbf { \\textit { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 626, + 356, + 642 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 356, + 628, + 392, + 639 + ], + "score": 0.91, + "content": "\\textbf { \\textit { y } } \\in \\textbf { \\textit { y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 626, + 508, + 642 + ], + "score": 1.0, + "content": ". In Bayesian neural net-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "works (BNNs), we are interested in the posterior distribution of all likely networks given by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 107, + 649, + 507, + 663 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 266, + 662 + ], + "score": 0.91, + "content": "\\begin{array} { r } { p ( \\mathbf { w } | \\mathcal { D } ) \\propto \\prod _ { i = 1 } ^ { n } p ( \\pmb { y } _ { i } | f ( \\pmb { x } _ { i } ; \\mathbf { w } ) ) p ( \\mathbf { w } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 649, + 299, + 663 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 299, + 649, + 321, + 661 + ], + "score": 0.91, + "content": "p ( \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 649, + 507, + 663 + ], + "score": 1.0, + "content": "is the prior distribution over weights. Cru-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 303, + 673 + ], + "score": 1.0, + "content": "cially, when making a prediction on a test point", + "type": "text" + }, + { + "bbox": [ + 303, + 662, + 315, + 671 + ], + "score": 0.86, + "content": "\\mathbf { \\boldsymbol { x } } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 660, + 506, + 673 + ], + "score": 1.0, + "content": ", in the Bayesian approach we do not only use", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 183, + 684 + ], + "score": 1.0, + "content": "a single parameter", + "type": "text" + }, + { + "bbox": [ + 183, + 672, + 193, + 681 + ], + "score": 0.76, + "content": "\\widehat { \\bf w }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 671, + 236, + 684 + ], + "score": 1.0, + "content": "to predict", + "type": "text" + }, + { + "bbox": [ + 236, + 671, + 298, + 683 + ], + "score": 0.93, + "content": "{ \\pmb y } ^ { * } = f ( { \\pmb x } ^ { * } ; \\widehat { \\mathbf { w } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 671, + 505, + 684 + ], + "score": 1.0, + "content": ", but we marginalize over the whole posterior, thus", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 682, + 333, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 333, + 695 + ], + "score": 1.0, + "content": "taking all possible explanations of the data into account:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 605, + 508, + 695 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 700, + 400, + 726 + ], + "lines": [ + { + "bbox": [ + 210, + 700, + 400, + 726 + ], + "spans": [ + { + "bbox": [ + 210, + 700, + 400, + 726 + ], + "score": 0.94, + "content": "p ( \\pmb { y } ^ { * } | \\pmb { x } ^ { * } , \\mathcal { D } ) = \\int p ( \\pmb { y } ^ { * } | f ( \\pmb { x } ^ { * } ; \\mathbf { w } ) ) p ( \\mathbf { w } | \\mathcal { D } ) \\mathrm { d } \\mathbf { w }", + "type": "interline_equation", + "image_path": "ea58bf85c2893e1394390263ae7f97e3c6d3270c2715a883b91057d7c53c61ba.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 210, + 700, + 400, + 726 + ], + "spans": [], + "index": 39 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 139 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "While approximating the posterior of Bayesian neural networks (or sampling from it) is a challenging", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "task, performing maximum a posteriori (MAP) estimation, which corresponds to finding the mode of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 106 + ], + "score": 1.0, + "content": "the posterior, is usually simple. Ensembles of neural networks use the non-convexity of the MAP", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 293, + 118 + ], + "score": 1.0, + "content": "optimization problem to create a collection of", + "type": "text" + }, + { + "bbox": [ + 294, + 106, + 304, + 115 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 105, + 507, + 118 + ], + "score": 1.0, + "content": "independent—and possibly different— solutions.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 114, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 158, + 131 + ], + "score": 1.0, + "content": "Considering", + "type": "text" + }, + { + "bbox": [ + 159, + 118, + 166, + 126 + ], + "score": 0.61, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 114, + 339, + 131 + ], + "score": 1.0, + "content": "weight configurations of a neural network", + "type": "text" + }, + { + "bbox": [ + 339, + 116, + 375, + 128 + ], + "score": 0.93, + "content": "\\{ { \\bf w } _ { i } \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 114, + 397, + 131 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 397, + 116, + 434, + 127 + ], + "score": 0.92, + "content": "{ \\bf w } _ { i } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 114, + 507, + 131 + ], + "score": 1.0, + "content": ", the dynamics of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 477, + 140 + ], + "score": 1.0, + "content": "the ensemble under the gradient of the posterior lead to the following update rule at iteration", + "type": "text" + }, + { + "bbox": [ + 477, + 128, + 482, + 137 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 127, + 487, + 140 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 142, + 383, + 177 + ], + "lines": [ + { + "bbox": [ + 228, + 142, + 383, + 177 + ], + "spans": [ + { + "bbox": [ + 228, + 142, + 383, + 177 + ], + "score": 0.83, + "content": "\\begin{array} { r l } & { \\mathbf { w } _ { i } ^ { t + 1 } \\mathbf { w } _ { i } ^ { t } + \\epsilon _ { t } \\phi ( \\mathbf { w } _ { i } ^ { t } ) } \\\\ { \\mathrm { w i t h } \\quad } & { \\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "dec175f126457d7570141af68aa834ee6aa688fa0490be6f7e2b319f4c374697.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 142, + 383, + 159.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 228, + 159.5, + 383, + 177.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 178, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 162, + 193 + ], + "score": 1.0, + "content": "with step size", + "type": "text" + }, + { + "bbox": [ + 163, + 181, + 171, + 191 + ], + "score": 0.82, + "content": "\\epsilon _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 179, + 506, + 193 + ], + "score": 1.0, + "content": ". Ensemble methods have a long history [e.g., 45, 26, 6] and were recently revisited", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 188, + 204 + ], + "score": 1.0, + "content": "for neural networks", + "type": "text" + }, + { + "bbox": [ + 188, + 190, + 206, + 202 + ], + "score": 0.78, + "content": "\\pm 2 \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 189, + 506, + 204 + ], + "score": 1.0, + "content": "and coined deep ensembles. The predictions of the different members are", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 505, + 213 + ], + "score": 1.0, + "content": "combined to create a predictive distribution by using the solutions to compute the Bayesian model", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 211, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 196, + 226 + ], + "score": 1.0, + "content": "average (BMA) in Eq.", + "type": "text" + }, + { + "bbox": [ + 197, + 212, + 209, + 225 + ], + "score": 0.69, + "content": "\\mathbb { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 211, + 270, + 226 + ], + "score": 1.0, + "content": ". Recent works", + "type": "text" + }, + { + "bbox": [ + 270, + 212, + 288, + 223 + ], + "score": 0.75, + "content": "\\mathbb { \\lVert \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 211, + 506, + 226 + ], + "score": 1.0, + "content": "have shown that deep ensembles can outperform some", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 223, + 504, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 486, + 236 + ], + "score": 1.0, + "content": "of the Bayesian approaches for uncertainty estimation. Even more recently, Wilson and Izmailov", + "type": "text" + }, + { + "bbox": [ + 486, + 223, + 504, + 235 + ], + "score": 0.82, + "content": "\\pmb { \\Vert 8 2 \\Vert }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 233, + 507, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 507, + 248 + ], + "score": 1.0, + "content": "argued that deep ensembles can be considered a compelling approach to Bayesian model averaging.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "Despite these ideas, the ability of deep ensembles to efficiently average over multiple hypotheses and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 407, + 268 + ], + "score": 1.0, + "content": "to explore the functional landscape of the posterior distribution studied in", + "type": "text" + }, + { + "bbox": [ + 407, + 255, + 425, + 267 + ], + "score": 0.77, + "content": "[ \\overline { { 1 8 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "does not guarantee", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 457, + 279 + ], + "score": 1.0, + "content": "sampling from the right distribution. Indeed, the additional Langevin noise introduced in", + "type": "text" + }, + { + "bbox": [ + 457, + 267, + 475, + 278 + ], + "score": 0.78, + "content": "\\mathbb { [ \\breve { ] { ] \\mathrm { ~ Z ~ Z ~ } ] } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 267, + 505, + 279 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 277, + 500, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 500, + 290 + ], + "score": 1.0, + "content": "is not considered in deep ensembles, is crucial to ensure samples from the true Bayesian posterior.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 506, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 307 + ], + "score": 1.0, + "content": "From a practical standpoint, since the quality of an ensemble hinges on the diversity of its members,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 306, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 317 + ], + "score": 1.0, + "content": "many methods were recently proposed to improve this diversity without compromising the individual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "score": 1.0, + "content": "accuracy. For instance, Wenzel et al. [80] propose hyper-deep ensembles that combine deep networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "with different hyperparameters. Similarly, cyclical learning-rate schedules can explore several local", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 244, + 350 + ], + "score": 1.0, + "content": "minima for the ensemble members", + "type": "text" + }, + { + "bbox": [ + 245, + 337, + 262, + 349 + ], + "score": 0.31, + "content": "\\pmb { \\mathbb { B 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 338, + 386, + 350 + ], + "score": 1.0, + "content": ". Alternatively, Rame and Cord", + "type": "text" + }, + { + "bbox": [ + 386, + 337, + 404, + 349 + ], + "score": 0.69, + "content": " { \\mathbb { I } } { \\mathbb { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 338, + 506, + 350 + ], + "score": 1.0, + "content": "proposed an information-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 417, + 361 + ], + "score": 1.0, + "content": "theoretic framework to avoid redundancy in the members and Oswald et al.", + "type": "text" + }, + { + "bbox": [ + 418, + 348, + 435, + 360 + ], + "score": 0.56, + "content": "\\overline { { \\| 6 3 \\| } }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "studied possible", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "interactions between members based on weight sharing. However, the absence of a constraint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "that prevents particles from converging to the same mode limits the possibility of improvement by", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "score": 1.0, + "content": "introducing more ensemble members. This means that any hopes to converge to different modes must", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 392, + 187, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 187, + 406 + ], + "score": 1.0, + "content": "exclusively rely on:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 129, + 412, + 439, + 454 + ], + "lines": [ + { + "bbox": [ + 130, + 412, + 285, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 285, + 424 + ], + "score": 1.0, + "content": "1. the randomness of the initialization", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 128, + 426, + 397, + 440 + ], + "spans": [ + { + "bbox": [ + 128, + 426, + 397, + 440 + ], + "score": 1.0, + "content": "2. the noise in the estimation of the gradients due to minibatching", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 128, + 441, + 439, + 455 + ], + "spans": [ + { + "bbox": [ + 128, + 441, + 439, + 455 + ], + "score": 1.0, + "content": "3. the number of local optima that might be reached during gradient descent.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 506, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 276, + 475 + ], + "score": 1.0, + "content": "Moreover, the recent study of Geiger et al.", + "type": "text" + }, + { + "bbox": [ + 277, + 461, + 295, + 473 + ], + "score": 0.57, + "content": "\\pmb { \\left. \\pmb { \\left. \\bar { 2 5 } \\right. } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "showed how the empirical test error of the ensemble", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "converges to the one of a single trained model when the number of parameters goes to infinity, leading", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "to deterioration of the performance. In other words, the bigger the model, the harder it is to maintain", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "diversity in the ensemble and avoid collapse to the same solution. This is intuitively due to the fact", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "that bigger models are less sensitive to the initialization. Namely, in order for them to get stuck in a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 516, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 529 + ], + "score": 1.0, + "content": "local minimum, they must have second derivatives that are positive simultaneously in all directions.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 527, + 374, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 374, + 540 + ], + "score": 1.0, + "content": "As the number of hidden units gets larger, this becomes less likely.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 263, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 264, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 264, + 567 + ], + "score": 1.0, + "content": "2.1 Repulsive force in weight space", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 571, + 506, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "score": 1.0, + "content": "To overcome the aforementioned limitations of standard deep ensembles, we introduce, inspired", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 150, + 595 + ], + "score": 1.0, + "content": "by SVGD", + "type": "text" + }, + { + "bbox": [ + 150, + 582, + 168, + 594 + ], + "score": 0.48, + "content": "\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 582, + 506, + 595 + ], + "score": 1.0, + "content": ", a deep ensemble with members that interact with each other through a repulsive", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "component. Using a kernel function to model this interaction, the single models repel each other", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "based on their position in the weight space, so that two members can never assume the same weights.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 613, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 236, + 629 + ], + "score": 1.0, + "content": "Considering a stationary kernel", + "type": "text" + }, + { + "bbox": [ + 237, + 615, + 329, + 628 + ], + "score": 0.91, + "content": "\\bar { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 613, + 506, + 629 + ], + "score": 1.0, + "content": "acting in the parameter space of the neural", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 626, + 392, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 210, + 639 + ], + "score": 1.0, + "content": "network, a repulsive term", + "type": "text" + }, + { + "bbox": [ + 210, + 627, + 220, + 636 + ], + "score": 0.82, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 626, + 392, + 639 + ], + "score": 1.0, + "content": "can be parameterized through its gradient:", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 642, + 428, + 670 + ], + "lines": [ + { + "bbox": [ + 183, + 642, + 428, + 670 + ], + "spans": [ + { + "bbox": [ + 183, + 642, + 428, + 670 + ], + "score": 0.92, + "content": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) .", + "type": "interline_equation", + "image_path": "52631e0ed6e1a61e9e013d38c7585cc4709cd85b719f28114ba1570f9f671c2a.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 183, + 642, + 428, + 651.3333333333334 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 183, + 651.3333333333334, + 428, + 660.6666666666667 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 183, + 660.6666666666667, + 428, + 670.0000000000001 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 505, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "To get an intuition for the behavior of this repulsive term and its gradients, we can consider the RBF", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 684, + 475, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 134, + 698 + ], + "score": 1.0, + "content": "kernel", + "type": "text" + }, + { + "bbox": [ + 134, + 684, + 286, + 698 + ], + "score": 0.92, + "content": "\\begin{array} { r } { k ( \\mathbf { w } _ { i } , \\mathbf { w } _ { j } ) = \\exp \\big ( - \\frac { 1 } { h } | | \\mathbf { w } _ { i } - \\mathbf { w } _ { j } | | ^ { 2 } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 684, + 356, + 698 + ], + "score": 1.0, + "content": "with lengthscale", + "type": "text" + }, + { + "bbox": [ + 356, + 687, + 362, + 695 + ], + "score": 0.87, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 684, + 475, + 698 + ], + "score": 1.0, + "content": "and notice how its gradient", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 702, + 391, + 726 + ], + "lines": [ + { + "bbox": [ + 219, + 702, + 391, + 726 + ], + "spans": [ + { + "bbox": [ + 219, + 702, + 391, + 726 + ], + "score": 0.94, + "content": "\\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) = \\frac { 2 } { h } ( \\mathbf { w } _ { j } ^ { t } - \\mathbf { w } _ { i } ^ { t } ) k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } )", + "type": "interline_equation", + "image_path": "f23cff6e6b3dc3c8e61b222583d646eef0fa3e0251332d45f77faee6e2130a87.jpg" + } + ] + } + ], + "index": 50, + "virtual_lines": [ + { + "bbox": [ + 219, + 702, + 391, + 726 + ], + "spans": [], + "index": 50 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 139 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 506, + 87 + ], + "score": 1.0, + "content": "While approximating the posterior of Bayesian neural networks (or sampling from it) is a challenging", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "task, performing maximum a posteriori (MAP) estimation, which corresponds to finding the mode of", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 106 + ], + "score": 1.0, + "content": "the posterior, is usually simple. Ensembles of neural networks use the non-convexity of the MAP", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 293, + 118 + ], + "score": 1.0, + "content": "optimization problem to create a collection of", + "type": "text" + }, + { + "bbox": [ + 294, + 106, + 304, + 115 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 105, + 507, + 118 + ], + "score": 1.0, + "content": "independent—and possibly different— solutions.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 114, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 114, + 158, + 131 + ], + "score": 1.0, + "content": "Considering", + "type": "text" + }, + { + "bbox": [ + 159, + 118, + 166, + 126 + ], + "score": 0.61, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 114, + 339, + 131 + ], + "score": 1.0, + "content": "weight configurations of a neural network", + "type": "text" + }, + { + "bbox": [ + 339, + 116, + 375, + 128 + ], + "score": 0.93, + "content": "\\{ { \\bf w } _ { i } \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 114, + 397, + 131 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 397, + 116, + 434, + 127 + ], + "score": 0.92, + "content": "{ \\bf w } _ { i } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 114, + 507, + 131 + ], + "score": 1.0, + "content": ", the dynamics of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 477, + 140 + ], + "score": 1.0, + "content": "the ensemble under the gradient of the posterior lead to the following update rule at iteration", + "type": "text" + }, + { + "bbox": [ + 477, + 128, + 482, + 137 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 127, + 487, + 140 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 71, + 507, + 140 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 142, + 383, + 177 + ], + "lines": [ + { + "bbox": [ + 228, + 142, + 383, + 177 + ], + "spans": [ + { + "bbox": [ + 228, + 142, + 383, + 177 + ], + "score": 0.83, + "content": "\\begin{array} { r l } & { \\mathbf { w } _ { i } ^ { t + 1 } \\mathbf { w } _ { i } ^ { t } + \\epsilon _ { t } \\phi ( \\mathbf { w } _ { i } ^ { t } ) } \\\\ { \\mathrm { w i t h } \\quad } & { \\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "dec175f126457d7570141af68aa834ee6aa688fa0490be6f7e2b319f4c374697.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 142, + 383, + 159.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 228, + 159.5, + 383, + 177.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 178, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 179, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 179, + 162, + 193 + ], + "score": 1.0, + "content": "with step size", + "type": "text" + }, + { + "bbox": [ + 163, + 181, + 171, + 191 + ], + "score": 0.82, + "content": "\\epsilon _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 179, + 506, + 193 + ], + "score": 1.0, + "content": ". Ensemble methods have a long history [e.g., 45, 26, 6] and were recently revisited", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 189, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 188, + 204 + ], + "score": 1.0, + "content": "for neural networks", + "type": "text" + }, + { + "bbox": [ + 188, + 190, + 206, + 202 + ], + "score": 0.78, + "content": "\\pm 2 \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 189, + 506, + 204 + ], + "score": 1.0, + "content": "and coined deep ensembles. The predictions of the different members are", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 505, + 213 + ], + "score": 1.0, + "content": "combined to create a predictive distribution by using the solutions to compute the Bayesian model", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 211, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 196, + 226 + ], + "score": 1.0, + "content": "average (BMA) in Eq.", + "type": "text" + }, + { + "bbox": [ + 197, + 212, + 209, + 225 + ], + "score": 0.69, + "content": "\\mathbb { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 211, + 270, + 226 + ], + "score": 1.0, + "content": ". Recent works", + "type": "text" + }, + { + "bbox": [ + 270, + 212, + 288, + 223 + ], + "score": 0.75, + "content": "\\mathbb { \\lVert \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 211, + 506, + 226 + ], + "score": 1.0, + "content": "have shown that deep ensembles can outperform some", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 223, + 504, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 486, + 236 + ], + "score": 1.0, + "content": "of the Bayesian approaches for uncertainty estimation. Even more recently, Wilson and Izmailov", + "type": "text" + }, + { + "bbox": [ + 486, + 223, + 504, + 235 + ], + "score": 0.82, + "content": "\\pmb { \\Vert 8 2 \\Vert }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 233, + 507, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 507, + 248 + ], + "score": 1.0, + "content": "argued that deep ensembles can be considered a compelling approach to Bayesian model averaging.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "Despite these ideas, the ability of deep ensembles to efficiently average over multiple hypotheses and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 255, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 407, + 268 + ], + "score": 1.0, + "content": "to explore the functional landscape of the posterior distribution studied in", + "type": "text" + }, + { + "bbox": [ + 407, + 255, + 425, + 267 + ], + "score": 0.77, + "content": "[ \\overline { { 1 8 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 255, + 505, + 268 + ], + "score": 1.0, + "content": "does not guarantee", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 267, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 457, + 279 + ], + "score": 1.0, + "content": "sampling from the right distribution. Indeed, the additional Langevin noise introduced in", + "type": "text" + }, + { + "bbox": [ + 457, + 267, + 475, + 278 + ], + "score": 0.78, + "content": "\\mathbb { [ \\breve { ] { ] \\mathrm { ~ Z ~ Z ~ } ] } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 267, + 505, + 279 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 277, + 500, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 500, + 290 + ], + "score": 1.0, + "content": "is not considered in deep ensembles, is crucial to ensure samples from the true Bayesian posterior.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 179, + 507, + 290 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 294, + 506, + 404 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 307 + ], + "score": 1.0, + "content": "From a practical standpoint, since the quality of an ensemble hinges on the diversity of its members,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 306, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 317 + ], + "score": 1.0, + "content": "many methods were recently proposed to improve this diversity without compromising the individual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 506, + 328 + ], + "score": 1.0, + "content": "accuracy. For instance, Wenzel et al. [80] propose hyper-deep ensembles that combine deep networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "with different hyperparameters. Similarly, cyclical learning-rate schedules can explore several local", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 244, + 350 + ], + "score": 1.0, + "content": "minima for the ensemble members", + "type": "text" + }, + { + "bbox": [ + 245, + 337, + 262, + 349 + ], + "score": 0.31, + "content": "\\pmb { \\mathbb { B 3 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 338, + 386, + 350 + ], + "score": 1.0, + "content": ". Alternatively, Rame and Cord", + "type": "text" + }, + { + "bbox": [ + 386, + 337, + 404, + 349 + ], + "score": 0.69, + "content": " { \\mathbb { I } } { \\mathbb { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 338, + 506, + 350 + ], + "score": 1.0, + "content": "proposed an information-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 417, + 361 + ], + "score": 1.0, + "content": "theoretic framework to avoid redundancy in the members and Oswald et al.", + "type": "text" + }, + { + "bbox": [ + 418, + 348, + 435, + 360 + ], + "score": 0.56, + "content": "\\overline { { \\| 6 3 \\| } }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "studied possible", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 505, + 372 + ], + "score": 1.0, + "content": "interactions between members based on weight sharing. However, the absence of a constraint", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "that prevents particles from converging to the same mode limits the possibility of improvement by", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 506, + 394 + ], + "score": 1.0, + "content": "introducing more ensemble members. This means that any hopes to converge to different modes must", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 392, + 187, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 187, + 406 + ], + "score": 1.0, + "content": "exclusively rely on:", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 294, + 506, + 406 + ] + }, + { + "type": "index", + "bbox": [ + 129, + 412, + 439, + 454 + ], + "lines": [ + { + "bbox": [ + 130, + 412, + 285, + 424 + ], + "spans": [ + { + "bbox": [ + 130, + 412, + 285, + 424 + ], + "score": 1.0, + "content": "1. the randomness of the initialization", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 426, + 397, + 440 + ], + "spans": [ + { + "bbox": [ + 128, + 426, + 397, + 440 + ], + "score": 1.0, + "content": "2. the noise in the estimation of the gradients due to minibatching", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 441, + 439, + 455 + ], + "spans": [ + { + "bbox": [ + 128, + 441, + 439, + 455 + ], + "score": 1.0, + "content": "3. the number of local optima that might be reached during gradient descent.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + } + ], + "index": 29, + "bbox_fs": [ + 128, + 412, + 439, + 455 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 506, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 276, + 475 + ], + "score": 1.0, + "content": "Moreover, the recent study of Geiger et al.", + "type": "text" + }, + { + "bbox": [ + 277, + 461, + 295, + 473 + ], + "score": 0.57, + "content": "\\pmb { \\left. \\pmb { \\left. \\bar { 2 5 } \\right. } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "showed how the empirical test error of the ensemble", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "converges to the one of a single trained model when the number of parameters goes to infinity, leading", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "to deterioration of the performance. In other words, the bigger the model, the harder it is to maintain", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 505, + 507 + ], + "score": 1.0, + "content": "diversity in the ensemble and avoid collapse to the same solution. This is intuitively due to the fact", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "that bigger models are less sensitive to the initialization. Namely, in order for them to get stuck in a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 516, + 507, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 529 + ], + "score": 1.0, + "content": "local minimum, they must have second derivatives that are positive simultaneously in all directions.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 527, + 374, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 374, + 540 + ], + "score": 1.0, + "content": "As the number of hidden units gets larger, this becomes less likely.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 461, + 507, + 540 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 551, + 263, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 549, + 264, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 264, + 567 + ], + "score": 1.0, + "content": "2.1 Repulsive force in weight space", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 571, + 506, + 638 + ], + "lines": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 506, + 584 + ], + "score": 1.0, + "content": "To overcome the aforementioned limitations of standard deep ensembles, we introduce, inspired", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 150, + 595 + ], + "score": 1.0, + "content": "by SVGD", + "type": "text" + }, + { + "bbox": [ + 150, + 582, + 168, + 594 + ], + "score": 0.48, + "content": "\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 582, + 506, + 595 + ], + "score": 1.0, + "content": ", a deep ensemble with members that interact with each other through a repulsive", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "component. Using a kernel function to model this interaction, the single models repel each other", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "based on their position in the weight space, so that two members can never assume the same weights.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 613, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 236, + 629 + ], + "score": 1.0, + "content": "Considering a stationary kernel", + "type": "text" + }, + { + "bbox": [ + 237, + 615, + 329, + 628 + ], + "score": 0.91, + "content": "\\bar { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\to \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 613, + 506, + 629 + ], + "score": 1.0, + "content": "acting in the parameter space of the neural", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 626, + 392, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 210, + 639 + ], + "score": 1.0, + "content": "network, a repulsive term", + "type": "text" + }, + { + "bbox": [ + 210, + 627, + 220, + 636 + ], + "score": 0.82, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 626, + 392, + 639 + ], + "score": 1.0, + "content": "can be parameterized through its gradient:", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 571, + 506, + 639 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 642, + 428, + 670 + ], + "lines": [ + { + "bbox": [ + 183, + 642, + 428, + 670 + ], + "spans": [ + { + "bbox": [ + 183, + 642, + 428, + 670 + ], + "score": 0.92, + "content": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) .", + "type": "interline_equation", + "image_path": "52631e0ed6e1a61e9e013d38c7585cc4709cd85b719f28114ba1570f9f671c2a.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 183, + 642, + 428, + 651.3333333333334 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 183, + 651.3333333333334, + 428, + 660.6666666666667 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 183, + 660.6666666666667, + 428, + 670.0000000000001 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 673, + 505, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "To get an intuition for the behavior of this repulsive term and its gradients, we can consider the RBF", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 684, + 475, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 134, + 698 + ], + "score": 1.0, + "content": "kernel", + "type": "text" + }, + { + "bbox": [ + 134, + 684, + 286, + 698 + ], + "score": 0.92, + "content": "\\begin{array} { r } { k ( \\mathbf { w } _ { i } , \\mathbf { w } _ { j } ) = \\exp \\big ( - \\frac { 1 } { h } | | \\mathbf { w } _ { i } - \\mathbf { w } _ { j } | | ^ { 2 } \\big ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 684, + 356, + 698 + ], + "score": 1.0, + "content": "with lengthscale", + "type": "text" + }, + { + "bbox": [ + 356, + 687, + 362, + 695 + ], + "score": 0.87, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 684, + 475, + 698 + ], + "score": 1.0, + "content": "and notice how its gradient", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 674, + 505, + 698 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 702, + 391, + 726 + ], + "lines": [ + { + "bbox": [ + 219, + 702, + 391, + 726 + ], + "spans": [ + { + "bbox": [ + 219, + 702, + 391, + 726 + ], + "score": 0.94, + "content": "\\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) = \\frac { 2 } { h } ( \\mathbf { w } _ { j } ^ { t } - \\mathbf { w } _ { i } ^ { t } ) k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } )", + "type": "interline_equation", + "image_path": "f23cff6e6b3dc3c8e61b222583d646eef0fa3e0251332d45f77faee6e2130a87.jpg" + } + ] + } + ], + "index": 50, + "virtual_lines": [ + { + "bbox": [ + 219, + 702, + 391, + 726 + ], + "spans": [], + "index": 50 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 133, + 86 + ], + "score": 1.0, + "content": "drives", + "type": "text" + }, + { + "bbox": [ + 133, + 74, + 146, + 84 + ], + "score": 0.86, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 72, + 298, + 86 + ], + "score": 1.0, + "content": "away from its neighboring members", + "type": "text" + }, + { + "bbox": [ + 298, + 74, + 312, + 85 + ], + "score": 0.88, + "content": "{ \\bf w } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 72, + 506, + 86 + ], + "score": 1.0, + "content": ", thus creating a repulsive effect. Naturally, not", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 178, + 95 + ], + "score": 1.0, + "content": "all the choices of", + "type": "text" + }, + { + "bbox": [ + 178, + 84, + 188, + 93 + ], + "score": 0.83, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "induce this effect. One of the simplest formulations to obtain it is via a linear", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 101, + 90, + 509, + 113 + ], + "spans": [ + { + "bbox": [ + 101, + 90, + 387, + 113 + ], + "score": 1.0, + "content": "combination of the kernel gradients scaled by a positive factor, that is,", + "type": "text" + }, + { + "bbox": [ + 388, + 94, + 484, + 108 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\beta \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 90, + 509, + 113 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 106, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 107, + 140, + 119 + ], + "score": 0.91, + "content": "\\beta \\in \\mathbb { R } _ { * } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 106, + 232, + 121 + ], + "score": 1.0, + "content": ". We will see in Section", + "type": "text" + }, + { + "bbox": [ + 233, + 106, + 242, + 121 + ], + "score": 0.79, + "content": "\\begin{array} { l } { 3 } \\\\ { . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 106, + 313, + 121 + ], + "score": 1.0, + "content": "how the choice of", + "type": "text" + }, + { + "bbox": [ + 313, + 108, + 321, + 119 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 106, + 505, + 121 + ], + "score": 1.0, + "content": "can be justified in order to obtain convergence", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 118, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 420, + 131 + ], + "score": 1.0, + "content": "to the Bayesian posterior together with alternative possible formulations of", + "type": "text" + }, + { + "bbox": [ + 420, + 119, + 430, + 128 + ], + "score": 0.84, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 118, + 506, + 131 + ], + "score": 1.0, + "content": "that preserve this", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 131, + 162, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 162, + 143 + ], + "score": 1.0, + "content": "convergence.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 153, + 270, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 271, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 271, + 168 + ], + "score": 1.0, + "content": "2.2 Repulsive force in function space", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 506, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 410, + 186 + ], + "score": 1.0, + "content": "To overcome the aforementioned overparameterization issue, the update in Eq.", + "type": "text" + }, + { + "bbox": [ + 410, + 173, + 423, + 186 + ], + "score": 0.83, + "content": "\\textcircled { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "can be formulated in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 277, + 197 + ], + "score": 1.0, + "content": "function space instead of weight space. Let", + "type": "text" + }, + { + "bbox": [ + 277, + 185, + 348, + 197 + ], + "score": 0.93, + "content": "f : \\mathbf { w } \\mapsto f ( \\cdot ; \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 185, + 505, + 197 + ], + "score": 1.0, + "content": "be the map that maps a configuration of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 138, + 209 + ], + "score": 1.0, + "content": "weights", + "type": "text" + }, + { + "bbox": [ + 138, + 196, + 172, + 207 + ], + "score": 0.9, + "content": "\\mathbf { w } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 196, + 444, + 209 + ], + "score": 1.0, + "content": "to the corresponding neural network regression function and denote as", + "type": "text" + }, + { + "bbox": [ + 445, + 196, + 505, + 209 + ], + "score": 0.92, + "content": "f _ { i } : = f ( \\cdot ; { \\mathbf w } _ { i } )", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 311, + 220 + ], + "score": 1.0, + "content": "the function with a certain configuration of weights", + "type": "text" + }, + { + "bbox": [ + 311, + 209, + 324, + 219 + ], + "score": 0.87, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 207, + 415, + 220 + ], + "score": 1.0, + "content": ". We can now consider", + "type": "text" + }, + { + "bbox": [ + 415, + 210, + 423, + 218 + ], + "score": 0.75, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 207, + 506, + 220 + ], + "score": 1.0, + "content": "particles in function", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 216, + 508, + 233 + ], + "spans": [ + { + "bbox": [ + 104, + 216, + 131, + 233 + ], + "score": 1.0, + "content": "space", + "type": "text" + }, + { + "bbox": [ + 132, + 218, + 164, + 231 + ], + "score": 0.92, + "content": "\\{ f _ { i } \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 216, + 186, + 233 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 187, + 219, + 215, + 230 + ], + "score": 0.92, + "content": "\\pmb { f } \\in \\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 216, + 478, + 233 + ], + "score": 1.0, + "content": "and model their interaction with a general positive definite kernel", + "type": "text" + }, + { + "bbox": [ + 479, + 219, + 503, + 230 + ], + "score": 0.91, + "content": "k ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 216, + 508, + 233 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 228, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 309, + 242 + ], + "score": 1.0, + "content": "We also consider the implicit functional likelihood", + "type": "text" + }, + { + "bbox": [ + 309, + 230, + 349, + 241 + ], + "score": 0.92, + "content": "p ( \\pmb { y } | \\pmb { x } , \\pmb { f } )", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 228, + 462, + 242 + ], + "score": 1.0, + "content": ", determined by the measure", + "type": "text" + }, + { + "bbox": [ + 463, + 230, + 505, + 241 + ], + "score": 0.91, + "content": "p ( \\pmb { y } | \\pmb { x } , \\mathbf { w } )", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 238, + 507, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 306, + 254 + ], + "score": 1.0, + "content": "in the weight space, as well as the functional prior", + "type": "text" + }, + { + "bbox": [ + 307, + 241, + 327, + 252 + ], + "score": 0.91, + "content": "p ( f )", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 238, + 507, + 254 + ], + "score": 1.0, + "content": ", which can either be defined separately (e.g.,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 419, + 266 + ], + "score": 1.0, + "content": "using a GP) or modeled as a push-forward measure of the weight-space prior", + "type": "text" + }, + { + "bbox": [ + 419, + 251, + 441, + 263 + ], + "score": 0.91, + "content": "p ( \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 250, + 506, + 266 + ], + "score": 1.0, + "content": ". Together, they", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 271, + 275 + ], + "score": 1.0, + "content": "determine the posterior in function space", + "type": "text" + }, + { + "bbox": [ + 271, + 262, + 302, + 274 + ], + "score": 0.92, + "content": "p ( \\pmb { f } | \\mathcal { D } )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 261, + 506, + 275 + ], + "score": 1.0, + "content": ". The functional evolution of a particle can then be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 151, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 151, + 285 + ], + "score": 1.0, + "content": "written as:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 287, + 441, + 332 + ], + "lines": [ + { + "bbox": [ + 182, + 287, + 441, + 332 + ], + "spans": [ + { + "bbox": [ + 182, + 287, + 441, + 332 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { f _ { i } ^ { t + 1 } \\gets f _ { i } ^ { t } + \\epsilon _ { t } \\phi ( f _ { i } ^ { t } ) } \\\\ { \\mathrm { t h } \\quad } & { \\phi ( f _ { i } ^ { t } ) = \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "6e9ccd562a52a4d7474eb2dbb571b766cb26754c555d1761309dd914491e04b8.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 182, + 287, + 441, + 302.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 182, + 302.0, + 441, + 317.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 182, + 317.0, + 441, + 332.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 335, + 504, + 369 + ], + "lines": [ + { + "bbox": [ + 107, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 107, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "However, computing the update in function space is neither tractable nor practical, which is why two", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 108, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 108, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "additional considerations are needed. The first one regards the infinite dimensionality of function", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 357, + 403, + 370 + ], + "spans": [ + { + "bbox": [ + 107, + 357, + 403, + 370 + ], + "score": 1.0, + "content": "space, which we circumvent using a canonical projection into a subspace:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 503, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 292, + 385 + ], + "score": 1.0, + "content": "Definition 1 (Canonical projection). For any", + "type": "text" + }, + { + "bbox": [ + 293, + 371, + 323, + 382 + ], + "score": 0.88, + "content": "A \\subset { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 369, + 368, + 385 + ], + "score": 1.0, + "content": ", we define", + "type": "text" + }, + { + "bbox": [ + 368, + 370, + 434, + 383 + ], + "score": 0.91, + "content": "\\pi _ { A } : \\mathbb { R } ^ { \\mathcal { X } } \\to \\mathbb { R } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 369, + 506, + 385 + ], + "score": 1.0, + "content": "as the canonical", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 381, + 302, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 169, + 397 + ], + "score": 1.0, + "content": "projection onto", + "type": "text" + }, + { + "bbox": [ + 170, + 383, + 178, + 393 + ], + "score": 0.69, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 381, + 212, + 397 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + }, + { + "bbox": [ + 212, + 383, + 297, + 395 + ], + "score": 0.93, + "content": "\\pi _ { A } ( f ) = \\{ f ( a ) \\} _ { a \\in A }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 381, + 302, + 397 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "In other words, the kernel will not be evaluated directly in function space, but on the projection", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 107, + 413, + 182, + 426 + ], + "score": 0.91, + "content": "k \\big ( \\pi _ { B } ( f ) , \\pi _ { B } ( f ^ { \\prime } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 413, + 205, + 426 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 206, + 414, + 215, + 424 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "being a subset of the input space given by a batch of training data points.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "score": 1.0, + "content": "The second consideration is to project this update back into the parameter space and evolve a set of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "particles there, because ultimately we are interested in representing the functions by parameterized", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 445, + 477, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 369, + 459 + ], + "score": 1.0, + "content": "neural networks. For this purpose, we can use the Jacobian of the", + "type": "text" + }, + { + "bbox": [ + 369, + 447, + 374, + 456 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 445, + 477, + 459 + ], + "score": 1.0, + "content": "-th particle as a projector:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 470, + 470, + 500 + ], + "lines": [ + { + "bbox": [ + 141, + 470, + 470, + 500 + ], + "spans": [ + { + "bbox": [ + 141, + 470, + 470, + 500 + ], + "score": 0.93, + "content": "\\phi ( { \\mathbf w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial { \\mathbf w } _ { i } ^ { t } } \\right) ^ { \\top } \\left[ \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | { \\mathcal D } ) - { \\mathcal R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( \\pi _ { B } ( f _ { i } ^ { t } ) , \\pi _ { B } ( f _ { j } ^ { t } ) ) \\right\\} _ { j = 1 } ^ { n } \\right) \\right] .", + "type": "interline_equation", + "image_path": "d7dea1ea5314c734775cfe499bded9ea6e655f30206ef90295c8a3d5d82d37d3.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 141, + 470, + 470, + 480.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 141, + 480.0, + 470, + 490.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 141, + 490.0, + 470, + 500.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 511, + 340, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 341, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 341, + 525 + ], + "score": 1.0, + "content": "2.3 Comparison to Stein variational gradient descent", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 487, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 530, + 488, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 289, + 545 + ], + "score": 1.0, + "content": "Note that our update is reminiscent of SVGD", + "type": "text" + }, + { + "bbox": [ + 290, + 533, + 306, + 542 + ], + "score": 0.34, + "content": "\\mathbb { \\left[ 5 1 \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 530, + 488, + 545 + ], + "score": 1.0, + "content": ", which in parameter space can be written as:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 547, + 438, + 582 + ], + "lines": [ + { + "bbox": [ + 173, + 547, + 438, + 582 + ], + "spans": [ + { + "bbox": [ + 173, + 547, + 438, + 582 + ], + "score": 0.94, + "content": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\sum _ { j = 1 } ^ { n } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) + \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { j } ^ { t } } k ( \\mathbf { w } _ { j } ^ { t } , \\mathbf { w } _ { i } ^ { t } ) .", + "type": "interline_equation", + "image_path": "ed4ba25f1f792a3ef21feee9794a7aaba9ed6181f2313384dd85bba5b6760967.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 173, + 547, + 438, + 558.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 173, + 558.6666666666666, + 438, + 570.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 173, + 570.3333333333333, + 438, + 581.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 585, + 506, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "It is important to notice that here, the gradients are averaged across all the particles using the kernel", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "matrix. Interestingly, SVGD can be asymptotically interpreted as gradient flow of the KL divergence", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "under a new metric induced by the Stein operator [16, 50] (see Appendix D for more details). Moving", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 618, + 410, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 295, + 630 + ], + "score": 1.0, + "content": "the inference from parameter to function space", + "type": "text" + }, + { + "bbox": [ + 296, + 618, + 313, + 629 + ], + "score": 0.43, + "content": "\\pmb { \\mathbb { Z } } 6 \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 618, + 410, + 630 + ], + "score": 1.0, + "content": "leads to the update rule", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 633, + 457, + 668 + ], + "lines": [ + { + "bbox": [ + 154, + 633, + 457, + 668 + ], + "spans": [ + { + "bbox": [ + 154, + 633, + 457, + 668 + ], + "score": 0.94, + "content": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial \\mathbf { w } _ { i } ^ { t } } \\right) ^ { \\top } \\left( \\frac { 1 } { n } \\sum _ { j = 1 } ^ { n } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\nabla _ { f _ { j } ^ { t } } \\log p ( f _ { j } ^ { t } | \\mathcal { D } ) + \\nabla _ { f _ { j } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right) .", + "type": "interline_equation", + "image_path": "da910223233132ed75eba544433552d88de5b055ed7c20ab997ee811df84045c.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 154, + 633, + 457, + 644.6666666666666 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 154, + 644.6666666666666, + 457, + 656.3333333333333 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 154, + 656.3333333333333, + 457, + 667.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "This way of averaging gradients using a kernel can be dangerous in high-dimensional settings, where", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 434, + 702 + ], + "score": 1.0, + "content": "kernel methods often suffer from the curse of dimensionality. Moreover, in Eq.", + "type": "text" + }, + { + "bbox": [ + 434, + 689, + 447, + 702 + ], + "score": 0.79, + "content": "\\textcircled{6}", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 687, + 505, + 702 + ], + "score": 1.0, + "content": ", the posterior", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "gradients of the particles are averaged using their similarity in weight space, which can be misleading", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 289, + 724 + ], + "score": 1.0, + "content": "in multi-modal posteriors. Worse yet, in Eq.", + "type": "text" + }, + { + "bbox": [ + 290, + 711, + 302, + 723 + ], + "score": 0.8, + "content": "\\textcircled{7}", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 711, + 505, + 724 + ], + "score": 1.0, + "content": ", the gradients are averaged in function space and", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 133, + 86 + ], + "score": 1.0, + "content": "drives", + "type": "text" + }, + { + "bbox": [ + 133, + 74, + 146, + 84 + ], + "score": 0.86, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 72, + 298, + 86 + ], + "score": 1.0, + "content": "away from its neighboring members", + "type": "text" + }, + { + "bbox": [ + 298, + 74, + 312, + 85 + ], + "score": 0.88, + "content": "{ \\bf w } _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 72, + 506, + 86 + ], + "score": 1.0, + "content": ", thus creating a repulsive effect. Naturally, not", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 178, + 95 + ], + "score": 1.0, + "content": "all the choices of", + "type": "text" + }, + { + "bbox": [ + 178, + 84, + 188, + 93 + ], + "score": 0.83, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "induce this effect. One of the simplest formulations to obtain it is via a linear", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 101, + 90, + 509, + 113 + ], + "spans": [ + { + "bbox": [ + 101, + 90, + 387, + 113 + ], + "score": 1.0, + "content": "combination of the kernel gradients scaled by a positive factor, that is,", + "type": "text" + }, + { + "bbox": [ + 388, + 94, + 484, + 108 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\beta \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 90, + 509, + 113 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 106, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 107, + 107, + 140, + 119 + ], + "score": 0.91, + "content": "\\beta \\in \\mathbb { R } _ { * } ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 106, + 232, + 121 + ], + "score": 1.0, + "content": ". We will see in Section", + "type": "text" + }, + { + "bbox": [ + 233, + 106, + 242, + 121 + ], + "score": 0.79, + "content": "\\begin{array} { l } { 3 } \\\\ { . } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 106, + 313, + 121 + ], + "score": 1.0, + "content": "how the choice of", + "type": "text" + }, + { + "bbox": [ + 313, + 108, + 321, + 119 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 106, + 505, + 121 + ], + "score": 1.0, + "content": "can be justified in order to obtain convergence", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 118, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 420, + 131 + ], + "score": 1.0, + "content": "to the Bayesian posterior together with alternative possible formulations of", + "type": "text" + }, + { + "bbox": [ + 420, + 119, + 430, + 128 + ], + "score": 0.84, + "content": "\\mathcal { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 118, + 506, + 131 + ], + "score": 1.0, + "content": "that preserve this", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 131, + 162, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 162, + 143 + ], + "score": 1.0, + "content": "convergence.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 101, + 72, + 509, + 143 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 153, + 270, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 271, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 271, + 168 + ], + "score": 1.0, + "content": "2.2 Repulsive force in function space", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 506, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 410, + 186 + ], + "score": 1.0, + "content": "To overcome the aforementioned overparameterization issue, the update in Eq.", + "type": "text" + }, + { + "bbox": [ + 410, + 173, + 423, + 186 + ], + "score": 0.83, + "content": "\\textcircled { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 174, + 505, + 186 + ], + "score": 1.0, + "content": "can be formulated in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 277, + 197 + ], + "score": 1.0, + "content": "function space instead of weight space. Let", + "type": "text" + }, + { + "bbox": [ + 277, + 185, + 348, + 197 + ], + "score": 0.93, + "content": "f : \\mathbf { w } \\mapsto f ( \\cdot ; \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 185, + 505, + 197 + ], + "score": 1.0, + "content": "be the map that maps a configuration of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 138, + 209 + ], + "score": 1.0, + "content": "weights", + "type": "text" + }, + { + "bbox": [ + 138, + 196, + 172, + 207 + ], + "score": 0.9, + "content": "\\mathbf { w } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 196, + 444, + 209 + ], + "score": 1.0, + "content": "to the corresponding neural network regression function and denote as", + "type": "text" + }, + { + "bbox": [ + 445, + 196, + 505, + 209 + ], + "score": 0.92, + "content": "f _ { i } : = f ( \\cdot ; { \\mathbf w } _ { i } )", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 311, + 220 + ], + "score": 1.0, + "content": "the function with a certain configuration of weights", + "type": "text" + }, + { + "bbox": [ + 311, + 209, + 324, + 219 + ], + "score": 0.87, + "content": "\\mathbf { w } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 207, + 415, + 220 + ], + "score": 1.0, + "content": ". We can now consider", + "type": "text" + }, + { + "bbox": [ + 415, + 210, + 423, + 218 + ], + "score": 0.75, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 207, + 506, + 220 + ], + "score": 1.0, + "content": "particles in function", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 216, + 508, + 233 + ], + "spans": [ + { + "bbox": [ + 104, + 216, + 131, + 233 + ], + "score": 1.0, + "content": "space", + "type": "text" + }, + { + "bbox": [ + 132, + 218, + 164, + 231 + ], + "score": 0.92, + "content": "\\{ f _ { i } \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 216, + 186, + 233 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 187, + 219, + 215, + 230 + ], + "score": 0.92, + "content": "\\pmb { f } \\in \\mathcal { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 216, + 478, + 233 + ], + "score": 1.0, + "content": "and model their interaction with a general positive definite kernel", + "type": "text" + }, + { + "bbox": [ + 479, + 219, + 503, + 230 + ], + "score": 0.91, + "content": "k ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 216, + 508, + 233 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 228, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 309, + 242 + ], + "score": 1.0, + "content": "We also consider the implicit functional likelihood", + "type": "text" + }, + { + "bbox": [ + 309, + 230, + 349, + 241 + ], + "score": 0.92, + "content": "p ( \\pmb { y } | \\pmb { x } , \\pmb { f } )", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 228, + 462, + 242 + ], + "score": 1.0, + "content": ", determined by the measure", + "type": "text" + }, + { + "bbox": [ + 463, + 230, + 505, + 241 + ], + "score": 0.91, + "content": "p ( \\pmb { y } | \\pmb { x } , \\mathbf { w } )", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 238, + 507, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 306, + 254 + ], + "score": 1.0, + "content": "in the weight space, as well as the functional prior", + "type": "text" + }, + { + "bbox": [ + 307, + 241, + 327, + 252 + ], + "score": 0.91, + "content": "p ( f )", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 238, + 507, + 254 + ], + "score": 1.0, + "content": ", which can either be defined separately (e.g.,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 419, + 266 + ], + "score": 1.0, + "content": "using a GP) or modeled as a push-forward measure of the weight-space prior", + "type": "text" + }, + { + "bbox": [ + 419, + 251, + 441, + 263 + ], + "score": 0.91, + "content": "p ( \\mathbf { w } )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 250, + 506, + 266 + ], + "score": 1.0, + "content": ". Together, they", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 271, + 275 + ], + "score": 1.0, + "content": "determine the posterior in function space", + "type": "text" + }, + { + "bbox": [ + 271, + 262, + 302, + 274 + ], + "score": 0.92, + "content": "p ( \\pmb { f } | \\mathcal { D } )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 261, + 506, + 275 + ], + "score": 1.0, + "content": ". The functional evolution of a particle can then be", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 151, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 151, + 285 + ], + "score": 1.0, + "content": "written as:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 173, + 508, + 285 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 182, + 287, + 441, + 332 + ], + "lines": [ + { + "bbox": [ + 182, + 287, + 441, + 332 + ], + "spans": [ + { + "bbox": [ + 182, + 287, + 441, + 332 + ], + "score": 0.92, + "content": "\\begin{array} { r l } & { f _ { i } ^ { t + 1 } \\gets f _ { i } ^ { t } + \\epsilon _ { t } \\phi ( f _ { i } ^ { t } ) } \\\\ { \\mathrm { t h } \\quad } & { \\phi ( f _ { i } ^ { t } ) = \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "6e9ccd562a52a4d7474eb2dbb571b766cb26754c555d1761309dd914491e04b8.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 182, + 287, + 441, + 302.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 182, + 302.0, + 441, + 317.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 182, + 317.0, + 441, + 332.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 335, + 504, + 369 + ], + "lines": [ + { + "bbox": [ + 107, + 335, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 107, + 335, + 505, + 348 + ], + "score": 1.0, + "content": "However, computing the update in function space is neither tractable nor practical, which is why two", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 108, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 108, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "additional considerations are needed. The first one regards the infinite dimensionality of function", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 357, + 403, + 370 + ], + "spans": [ + { + "bbox": [ + 107, + 357, + 403, + 370 + ], + "score": 1.0, + "content": "space, which we circumvent using a canonical projection into a subspace:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 107, + 335, + 505, + 370 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 503, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 292, + 385 + ], + "score": 1.0, + "content": "Definition 1 (Canonical projection). For any", + "type": "text" + }, + { + "bbox": [ + 293, + 371, + 323, + 382 + ], + "score": 0.88, + "content": "A \\subset { \\mathcal { X } }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 369, + 368, + 385 + ], + "score": 1.0, + "content": ", we define", + "type": "text" + }, + { + "bbox": [ + 368, + 370, + 434, + 383 + ], + "score": 0.91, + "content": "\\pi _ { A } : \\mathbb { R } ^ { \\mathcal { X } } \\to \\mathbb { R } ^ { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 369, + 506, + 385 + ], + "score": 1.0, + "content": "as the canonical", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 381, + 302, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 169, + 397 + ], + "score": 1.0, + "content": "projection onto", + "type": "text" + }, + { + "bbox": [ + 170, + 383, + 178, + 393 + ], + "score": 0.69, + "content": "A", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 381, + 212, + 397 + ], + "score": 1.0, + "content": ", that is,", + "type": "text" + }, + { + "bbox": [ + 212, + 383, + 297, + 395 + ], + "score": 0.93, + "content": "\\pi _ { A } ( f ) = \\{ f ( a ) \\} _ { a \\in A }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 381, + 302, + 397 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 369, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 505, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "In other words, the kernel will not be evaluated directly in function space, but on the projection", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 107, + 413, + 182, + 426 + ], + "score": 0.91, + "content": "k \\big ( \\pi _ { B } ( f ) , \\pi _ { B } ( f ^ { \\prime } ) \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 413, + 205, + 426 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 206, + 414, + 215, + 424 + ], + "score": 0.81, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "being a subset of the input space given by a batch of training data points.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 438 + ], + "score": 1.0, + "content": "The second consideration is to project this update back into the parameter space and evolve a set of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "particles there, because ultimately we are interested in representing the functions by parameterized", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 445, + 477, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 369, + 459 + ], + "score": 1.0, + "content": "neural networks. For this purpose, we can use the Jacobian of the", + "type": "text" + }, + { + "bbox": [ + 369, + 447, + 374, + 456 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 445, + 477, + 459 + ], + "score": 1.0, + "content": "-th particle as a projector:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 402, + 506, + 459 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 141, + 470, + 470, + 500 + ], + "lines": [ + { + "bbox": [ + 141, + 470, + 470, + 500 + ], + "spans": [ + { + "bbox": [ + 141, + 470, + 470, + 500 + ], + "score": 0.93, + "content": "\\phi ( { \\mathbf w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial { \\mathbf w } _ { i } ^ { t } } \\right) ^ { \\top } \\left[ \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | { \\mathcal D } ) - { \\mathcal R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( \\pi _ { B } ( f _ { i } ^ { t } ) , \\pi _ { B } ( f _ { j } ^ { t } ) ) \\right\\} _ { j = 1 } ^ { n } \\right) \\right] .", + "type": "interline_equation", + "image_path": "d7dea1ea5314c734775cfe499bded9ea6e655f30206ef90295c8a3d5d82d37d3.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 141, + 470, + 470, + 480.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 141, + 480.0, + 470, + 490.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 141, + 490.0, + 470, + 500.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "title", + "bbox": [ + 106, + 511, + 340, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 341, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 341, + 525 + ], + "score": 1.0, + "content": "2.3 Comparison to Stein variational gradient descent", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 530, + 487, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 530, + 488, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 289, + 545 + ], + "score": 1.0, + "content": "Note that our update is reminiscent of SVGD", + "type": "text" + }, + { + "bbox": [ + 290, + 533, + 306, + 542 + ], + "score": 0.34, + "content": "\\mathbb { \\left[ 5 1 \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 530, + 488, + 545 + ], + "score": 1.0, + "content": ", which in parameter space can be written as:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 530, + 488, + 545 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 173, + 547, + 438, + 582 + ], + "lines": [ + { + "bbox": [ + 173, + 547, + 438, + 582 + ], + "spans": [ + { + "bbox": [ + 173, + 547, + 438, + 582 + ], + "score": 0.94, + "content": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\sum _ { j = 1 } ^ { n } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) + \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { j } ^ { t } } k ( \\mathbf { w } _ { j } ^ { t } , \\mathbf { w } _ { i } ^ { t } ) .", + "type": "interline_equation", + "image_path": "ed4ba25f1f792a3ef21feee9794a7aaba9ed6181f2313384dd85bba5b6760967.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 173, + 547, + 438, + 558.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 173, + 558.6666666666666, + 438, + 570.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 173, + 570.3333333333333, + 438, + 581.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 585, + 506, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "It is important to notice that here, the gradients are averaged across all the particles using the kernel", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "matrix. Interestingly, SVGD can be asymptotically interpreted as gradient flow of the KL divergence", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "under a new metric induced by the Stein operator [16, 50] (see Appendix D for more details). Moving", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 618, + 410, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 295, + 630 + ], + "score": 1.0, + "content": "the inference from parameter to function space", + "type": "text" + }, + { + "bbox": [ + 296, + 618, + 313, + 629 + ], + "score": 0.43, + "content": "\\pmb { \\mathbb { Z } } 6 \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 618, + 410, + 630 + ], + "score": 1.0, + "content": "leads to the update rule", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 585, + 506, + 630 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 154, + 633, + 457, + 668 + ], + "lines": [ + { + "bbox": [ + 154, + 633, + 457, + 668 + ], + "spans": [ + { + "bbox": [ + 154, + 633, + 457, + 668 + ], + "score": 0.94, + "content": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial \\mathbf { w } _ { i } ^ { t } } \\right) ^ { \\top } \\left( \\frac { 1 } { n } \\sum _ { j = 1 } ^ { n } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\nabla _ { f _ { j } ^ { t } } \\log p ( f _ { j } ^ { t } | \\mathcal { D } ) + \\nabla _ { f _ { j } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right) .", + "type": "interline_equation", + "image_path": "da910223233132ed75eba544433552d88de5b055ed7c20ab997ee811df84045c.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 154, + 633, + 457, + 644.6666666666666 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 154, + 644.6666666666666, + 457, + 656.3333333333333 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 154, + 656.3333333333333, + 457, + 667.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "This way of averaging gradients using a kernel can be dangerous in high-dimensional settings, where", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 434, + 702 + ], + "score": 1.0, + "content": "kernel methods often suffer from the curse of dimensionality. Moreover, in Eq.", + "type": "text" + }, + { + "bbox": [ + 434, + 689, + 447, + 702 + ], + "score": 0.79, + "content": "\\textcircled{6}", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 687, + 505, + 702 + ], + "score": 1.0, + "content": ", the posterior", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "gradients of the particles are averaged using their similarity in weight space, which can be misleading", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 289, + 724 + ], + "score": 1.0, + "content": "in multi-modal posteriors. Worse yet, in Eq.", + "type": "text" + }, + { + "bbox": [ + 290, + 711, + 302, + 723 + ], + "score": 0.8, + "content": "\\textcircled{7}", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 711, + 505, + 724 + ], + "score": 1.0, + "content": ", the gradients are averaged in function space and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 293, + 85 + ], + "score": 1.0, + "content": "are then projected back using exclusively the", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 294, + 74, + 298, + 83 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 299, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "-th Jacobian, which can be harmful given that it is", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "not guaranteed that distances between functions evaluated on a subset of their input space resemble", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "their true distance. Our proposed method, on the other hand, does not employ any averaging of the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 469, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 469, + 118 + ], + "score": 1.0, + "content": "posterior gradients and thus comes closest to the true particle gradients in deep ensembles.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 677, + 505, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 293, + 85 + ], + "score": 1.0, + "content": "are then projected back using exclusively the", + "type": "text" + }, + { + "bbox": [ + 294, + 74, + 298, + 83 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "-th Jacobian, which can be harmful given that it is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "not guaranteed that distances between functions evaluated on a subset of their input space resemble", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "their true distance. Our proposed method, on the other hand, does not employ any averaging of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 469, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 469, + 118 + ], + "score": 1.0, + "content": "posterior gradients and thus comes closest to the true particle gradients in deep ensembles.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 107, + 131, + 326, + 145 + ], + "lines": [ + { + "bbox": [ + 104, + 130, + 327, + 149 + ], + "spans": [ + { + "bbox": [ + 104, + 130, + 327, + 149 + ], + "score": 1.0, + "content": "3 Repulsive deep ensembles are Bayesian", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 506, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 169 + ], + "score": 1.0, + "content": "So far, we represented the repulsive force as a general function of the gradients of a kernel. In this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 166, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 505, + 179 + ], + "score": 1.0, + "content": "section, we show how to determine the explicit form of the repulsive term, such that the resulting", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "update rule is equivalent to the discretization of the gradient flow dynamics of the KL divergence in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 507, + 201 + ], + "score": 1.0, + "content": "Wasserstein space. We begin by introducing the concepts of particle approximation and gradient flow.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 212, + 229, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 230, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 230, + 227 + ], + "score": 1.0, + "content": "3.1 Particle approximation", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "score": 1.0, + "content": "A particle-based approximation of a target measure depends on a set of weighted samples", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 242, + 367, + 257 + ], + "spans": [ + { + "bbox": [ + 107, + 243, + 162, + 256 + ], + "score": 0.92, + "content": "\\{ ( \\bar { x _ { i } } , w _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 242, + 367, + 257 + ], + "score": 1.0, + "content": ", for which an empirical measure can be defined as", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 258, + 358, + 290 + ], + "lines": [ + { + "bbox": [ + 253, + 258, + 358, + 290 + ], + "spans": [ + { + "bbox": [ + 253, + 258, + 358, + 290 + ], + "score": 0.95, + "content": "\\rho ( x ) = \\sum _ { i = 1 } ^ { n } w _ { i } \\delta ( x - x _ { i } ) ,", + "type": "interline_equation", + "image_path": "b51a8eebc42b5a02f34979b6c46328fac936b6561c35ce4b3a464937651a8dd3.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 253, + 258, + 358, + 274.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 253, + 274.0, + 358, + 290.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 102, + 287, + 509, + 311 + ], + "spans": [ + { + "bbox": [ + 102, + 287, + 133, + 311 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 293, + 150, + 306 + ], + "score": 0.91, + "content": "\\delta ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 287, + 326, + 311 + ], + "score": 1.0, + "content": "is the Dirac delta function and the weights", + "type": "text" + }, + { + "bbox": [ + 326, + 295, + 338, + 304 + ], + "score": 0.86, + "content": "w _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 287, + 368, + 311 + ], + "score": 1.0, + "content": "satisfy", + "type": "text" + }, + { + "bbox": [ + 369, + 293, + 413, + 305 + ], + "score": 0.92, + "content": "w _ { i } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 287, + 431, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 293, + 487, + 306 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { i = 1 } ^ { n } w _ { i } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 287, + 509, + 311 + ], + "score": 1.0, + "content": ". To", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 304, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 241, + 316 + ], + "score": 1.0, + "content": "approximate a target distribution", + "type": "text" + }, + { + "bbox": [ + 241, + 305, + 261, + 316 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 304, + 506, + 316 + ], + "score": 1.0, + "content": "using the empirical measure, the particles and their weights", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 484, + 328 + ], + "score": 1.0, + "content": "need to be selected in a principled manner that minimizes some measure of distance between", + "type": "text" + }, + { + "bbox": [ + 484, + 316, + 505, + 327 + ], + "score": 0.91, + "content": "\\pi ( x )", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 326, + 488, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 123, + 339 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 326, + 143, + 338 + ], + "score": 0.92, + "content": "\\rho ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 326, + 198, + 339 + ], + "score": 1.0, + "content": "(e.g., a set of", + "type": "text" + }, + { + "bbox": [ + 199, + 327, + 208, + 336 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 326, + 298, + 339 + ], + "score": 1.0, + "content": "samples with weights", + "type": "text" + }, + { + "bbox": [ + 299, + 326, + 342, + 338 + ], + "score": 0.92, + "content": "w _ { i } = 1 / N", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 326, + 488, + 339 + ], + "score": 1.0, + "content": "obtained using an MCMC method).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 350, + 273, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 275, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 275, + 364 + ], + "score": 1.0, + "content": "3.2 Gradient flow in parameter space", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 207, + 383 + ], + "score": 1.0, + "content": "Given a smooth function", + "type": "text" + }, + { + "bbox": [ + 207, + 369, + 257, + 380 + ], + "score": 0.9, + "content": "J : \\mathbb { R } ^ { d } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 369, + 506, + 383 + ], + "score": 1.0, + "content": "in Euclidean space, we can minimize it by creating a path that", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 373, + 394 + ], + "score": 1.0, + "content": "follows its negative gradient starting from some initial conditions", + "type": "text" + }, + { + "bbox": [ + 373, + 383, + 384, + 393 + ], + "score": 0.86, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 380, + 432, + 394 + ], + "score": 1.0, + "content": ". The curve", + "type": "text" + }, + { + "bbox": [ + 432, + 381, + 450, + 393 + ], + "score": 0.91, + "content": "x ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "with starting", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 392, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 130, + 404 + ], + "score": 1.0, + "content": "point", + "type": "text" + }, + { + "bbox": [ + 130, + 393, + 142, + 403 + ], + "score": 0.85, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 392, + 506, + 404 + ], + "score": 1.0, + "content": "described by that path is called gradient flow. The dynamics and evolution in time of a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 403, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 481, + 415 + ], + "score": 1.0, + "content": "considered point in the space under this minimization problem can be described as the ODE1", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 417, + 340, + 441 + ], + "lines": [ + { + "bbox": [ + 271, + 417, + 340, + 441 + ], + "spans": [ + { + "bbox": [ + 271, + 417, + 340, + 441 + ], + "score": 0.93, + "content": "\\frac { d \\boldsymbol { x } } { d t } = - \\boldsymbol { \\nabla } J ( \\boldsymbol { x } ) .", + "type": "interline_equation", + "image_path": "4af9c8caebc8dd0f4bb240ad868ba49b8044f29f61003074e861ae4833cdfbee.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 271, + 417, + 340, + 441 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "score": 1.0, + "content": "We can extend this concept to the space of probability distributions (Wasserstein gradient flow) [3].", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 311, + 466 + ], + "score": 1.0, + "content": "Let us consider the space of probability measures", + "type": "text" + }, + { + "bbox": [ + 311, + 453, + 343, + 466 + ], + "score": 0.92, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 453, + 505, + 466 + ], + "score": 1.0, + "content": ", that is, the set of probability measures", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 464, + 333, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 316, + 476 + ], + "score": 1.0, + "content": "with finite second moments defined on the manifold", + "type": "text" + }, + { + "bbox": [ + 316, + 466, + 329, + 474 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 464, + 333, + 476 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 476, + 451, + 504 + ], + "lines": [ + { + "bbox": [ + 159, + 476, + 451, + 504 + ], + "spans": [ + { + "bbox": [ + 159, + 476, + 451, + 504 + ], + "score": 0.96, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } ) = \\left\\{ \\varphi : \\mathcal { M } \\to [ 0 , \\infty ) \\bigg | \\int _ { \\mathcal { M } } \\mathrm { d } \\varphi = 1 , ~ \\int _ { \\mathcal { M } } | x | ^ { 2 } \\varphi ( x ) \\mathrm { d } x < + \\infty \\right\\} .", + "type": "interline_equation", + "image_path": "c1e5d77c719b4d2fa7a0247805d6e0144779c241f5db03b43a6f1d113276ff19.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 159, + 476, + 451, + 485.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 159, + 485.3333333333333, + 451, + 494.66666666666663 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 159, + 494.66666666666663, + 451, + 503.99999999999994 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 506, + 502, + 529 + ], + "lines": [ + { + "bbox": [ + 107, + 505, + 504, + 519 + ], + "spans": [ + { + "bbox": [ + 107, + 505, + 137, + 519 + ], + "score": 1.0, + "content": "Taking", + "type": "text" + }, + { + "bbox": [ + 138, + 506, + 170, + 518 + ], + "score": 0.92, + "content": "\\Pi ( \\mu , \\nu )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 505, + 405, + 519 + ], + "score": 1.0, + "content": "as the set of joint probability measures with marginals", + "type": "text" + }, + { + "bbox": [ + 406, + 508, + 423, + 518 + ], + "score": 0.87, + "content": "\\mu , \\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 505, + 504, + 519 + ], + "score": 1.0, + "content": ", we can define the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 516, + 282, + 530 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 235, + 530 + ], + "score": 1.0, + "content": "Wasserstein metric on the space", + "type": "text" + }, + { + "bbox": [ + 236, + 517, + 268, + 529 + ], + "score": 0.93, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 516, + 282, + 530 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 531, + 393, + 557 + ], + "lines": [ + { + "bbox": [ + 217, + 531, + 393, + 557 + ], + "spans": [ + { + "bbox": [ + 217, + 531, + 393, + 557 + ], + "score": 0.94, + "content": "W _ { 2 } ^ { 2 } ( \\mu , \\nu ) = \\operatorname* { i n f } _ { \\pi \\in \\Pi ( \\mu , \\nu ) } \\int | x - y | ^ { 2 } { \\mathrm { d } } \\pi ( x , y ) .", + "type": "interline_equation", + "image_path": "22875a9421a3c955c184f36bee03fad9988e4e0d847133375befe3e5951b4f25.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 217, + 531, + 393, + 557 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 323, + 574 + ], + "score": 1.0, + "content": "Considering the optimization problem of a functional", + "type": "text" + }, + { + "bbox": [ + 324, + 560, + 393, + 573 + ], + "score": 0.92, + "content": "J : \\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 559, + 443, + 574 + ], + "score": 1.0, + "content": ", such as the", + "type": "text" + }, + { + "bbox": [ + 443, + 561, + 458, + 571 + ], + "score": 0.39, + "content": "\\mathrm { K L }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 559, + 505, + 574 + ], + "score": 1.0, + "content": "divergence", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 571, + 409, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 276, + 584 + ], + "score": 1.0, + "content": "between the particle approximation in Eq.", + "type": "text" + }, + { + "bbox": [ + 276, + 571, + 289, + 585 + ], + "score": 0.83, + "content": "\\textcircled { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 571, + 385, + 584 + ], + "score": 1.0, + "content": "and the target posterior", + "type": "text" + }, + { + "bbox": [ + 385, + 572, + 405, + 583 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 571, + 409, + 584 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 586, + 423, + 612 + ], + "lines": [ + { + "bbox": [ + 187, + 586, + 423, + 612 + ], + "spans": [ + { + "bbox": [ + 187, + 586, + 423, + 612 + ], + "score": 0.93, + "content": "\\operatorname* { i n f } _ { \\rho \\in \\mathcal { P } _ { 2 } ( \\mathcal { M } ) } D _ { K L } ( \\rho , \\pi ) = \\int _ { \\mathcal { M } } ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\rho ( x ) \\mathrm { d } x ,", + "type": "interline_equation", + "image_path": "202eca2cfd27979881321c7e7947951e3cc456330188130d02fcab3cc4c0abde.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 187, + 586, + 423, + 612 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 252, + 626 + ], + "score": 1.0, + "content": "the evolution in time of the measure", + "type": "text" + }, + { + "bbox": [ + 252, + 616, + 259, + 626 + ], + "score": 0.8, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "under the equivalent of the gradient, the Wasserstein gradient", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 623, + 331, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 331, + 638 + ], + "score": 1.0, + "content": "flow, is described by the Liouville equation2 [38, 3, 64]:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 639, + 402, + 695 + ], + "lines": [ + { + "bbox": [ + 208, + 639, + 402, + 695 + ], + "spans": [ + { + "bbox": [ + 208, + 639, + 402, + 695 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( \\boldsymbol { x } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\big ( \\log \\rho ( \\boldsymbol { x } ) - \\log \\pi ( \\boldsymbol { x } ) \\big ) \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "319c5dba11def8b47c62bbed7e518cec22d2bb6447df21ccd8c8f0ceacda0468.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 639, + 402, + 653.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 208, + 653.0, + 402, + 667.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 208, + 667.0, + 402, + 681.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 208, + 681.0, + 402, + 695.0 + ], + "spans": [], + "index": 41 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 118, + 700, + 381, + 722 + ], + "lines": [ + { + "bbox": [ + 119, + 698, + 381, + 713 + ], + "spans": [ + { + "bbox": [ + 119, + 698, + 381, + 713 + ], + "score": 1.0, + "content": "1 Together with the initial condition, this is know as the Cauchy problem.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 709, + 266, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 266, + 724 + ], + "score": 1.0, + "content": "2 Also referred to as continuity equation.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 117 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 72, + 506, + 118 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 131, + 326, + 145 + ], + "lines": [ + { + "bbox": [ + 104, + 130, + 327, + 149 + ], + "spans": [ + { + "bbox": [ + 104, + 130, + 327, + 149 + ], + "score": 1.0, + "content": "3 Repulsive deep ensembles are Bayesian", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 506, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 169 + ], + "score": 1.0, + "content": "So far, we represented the repulsive force as a general function of the gradients of a kernel. In this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 166, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 505, + 179 + ], + "score": 1.0, + "content": "section, we show how to determine the explicit form of the repulsive term, such that the resulting", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "update rule is equivalent to the discretization of the gradient flow dynamics of the KL divergence in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 507, + 201 + ], + "score": 1.0, + "content": "Wasserstein space. We begin by introducing the concepts of particle approximation and gradient flow.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 155, + 507, + 201 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 212, + 229, + 224 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 230, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 230, + 227 + ], + "score": 1.0, + "content": "3.1 Particle approximation", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 505, + 246 + ], + "score": 1.0, + "content": "A particle-based approximation of a target measure depends on a set of weighted samples", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 242, + 367, + 257 + ], + "spans": [ + { + "bbox": [ + 107, + 243, + 162, + 256 + ], + "score": 0.92, + "content": "\\{ ( \\bar { x _ { i } } , w _ { i } ) \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 242, + 367, + 257 + ], + "score": 1.0, + "content": ", for which an empirical measure can be defined as", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 232, + 505, + 257 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 258, + 358, + 290 + ], + "lines": [ + { + "bbox": [ + 253, + 258, + 358, + 290 + ], + "spans": [ + { + "bbox": [ + 253, + 258, + 358, + 290 + ], + "score": 0.95, + "content": "\\rho ( x ) = \\sum _ { i = 1 } ^ { n } w _ { i } \\delta ( x - x _ { i } ) ,", + "type": "interline_equation", + "image_path": "b51a8eebc42b5a02f34979b6c46328fac936b6561c35ce4b3a464937651a8dd3.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 253, + 258, + 358, + 274.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 253, + 274.0, + 358, + 290.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 293, + 505, + 338 + ], + "lines": [ + { + "bbox": [ + 102, + 287, + 509, + 311 + ], + "spans": [ + { + "bbox": [ + 102, + 287, + 133, + 311 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 293, + 150, + 306 + ], + "score": 0.91, + "content": "\\delta ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 287, + 326, + 311 + ], + "score": 1.0, + "content": "is the Dirac delta function and the weights", + "type": "text" + }, + { + "bbox": [ + 326, + 295, + 338, + 304 + ], + "score": 0.86, + "content": "w _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 287, + 368, + 311 + ], + "score": 1.0, + "content": "satisfy", + "type": "text" + }, + { + "bbox": [ + 369, + 293, + 413, + 305 + ], + "score": 0.92, + "content": "w _ { i } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 287, + 431, + 311 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 432, + 293, + 487, + 306 + ], + "score": 0.91, + "content": "\\textstyle \\sum _ { i = 1 } ^ { n } w _ { i } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 287, + 509, + 311 + ], + "score": 1.0, + "content": ". To", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 304, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 241, + 316 + ], + "score": 1.0, + "content": "approximate a target distribution", + "type": "text" + }, + { + "bbox": [ + 241, + 305, + 261, + 316 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 304, + 506, + 316 + ], + "score": 1.0, + "content": "using the empirical measure, the particles and their weights", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 484, + 328 + ], + "score": 1.0, + "content": "need to be selected in a principled manner that minimizes some measure of distance between", + "type": "text" + }, + { + "bbox": [ + 484, + 316, + 505, + 327 + ], + "score": 0.91, + "content": "\\pi ( x )", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 326, + 488, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 123, + 339 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 326, + 143, + 338 + ], + "score": 0.92, + "content": "\\rho ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 326, + 198, + 339 + ], + "score": 1.0, + "content": "(e.g., a set of", + "type": "text" + }, + { + "bbox": [ + 199, + 327, + 208, + 336 + ], + "score": 0.85, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 326, + 298, + 339 + ], + "score": 1.0, + "content": "samples with weights", + "type": "text" + }, + { + "bbox": [ + 299, + 326, + 342, + 338 + ], + "score": 0.92, + "content": "w _ { i } = 1 / N", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 326, + 488, + 339 + ], + "score": 1.0, + "content": "obtained using an MCMC method).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 102, + 287, + 509, + 339 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 350, + 273, + 362 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 275, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 275, + 364 + ], + "score": 1.0, + "content": "3.2 Gradient flow in parameter space", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 369, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 207, + 383 + ], + "score": 1.0, + "content": "Given a smooth function", + "type": "text" + }, + { + "bbox": [ + 207, + 369, + 257, + 380 + ], + "score": 0.9, + "content": "J : \\mathbb { R } ^ { d } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 369, + 506, + 383 + ], + "score": 1.0, + "content": "in Euclidean space, we can minimize it by creating a path that", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 373, + 394 + ], + "score": 1.0, + "content": "follows its negative gradient starting from some initial conditions", + "type": "text" + }, + { + "bbox": [ + 373, + 383, + 384, + 393 + ], + "score": 0.86, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 380, + 432, + 394 + ], + "score": 1.0, + "content": ". The curve", + "type": "text" + }, + { + "bbox": [ + 432, + 381, + 450, + 393 + ], + "score": 0.91, + "content": "x ( t )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "with starting", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 392, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 130, + 404 + ], + "score": 1.0, + "content": "point", + "type": "text" + }, + { + "bbox": [ + 130, + 393, + 142, + 403 + ], + "score": 0.85, + "content": "x _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 392, + 506, + 404 + ], + "score": 1.0, + "content": "described by that path is called gradient flow. The dynamics and evolution in time of a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 403, + 481, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 481, + 415 + ], + "score": 1.0, + "content": "considered point in the space under this minimization problem can be described as the ODE1", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 369, + 506, + 415 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 271, + 417, + 340, + 441 + ], + "lines": [ + { + "bbox": [ + 271, + 417, + 340, + 441 + ], + "spans": [ + { + "bbox": [ + 271, + 417, + 340, + 441 + ], + "score": 0.93, + "content": "\\frac { d \\boldsymbol { x } } { d t } = - \\boldsymbol { \\nabla } J ( \\boldsymbol { x } ) .", + "type": "interline_equation", + "image_path": "4af9c8caebc8dd0f4bb240ad868ba49b8044f29f61003074e861ae4833cdfbee.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 271, + 417, + 340, + 441 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 442, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 504, + 455 + ], + "score": 1.0, + "content": "We can extend this concept to the space of probability distributions (Wasserstein gradient flow) [3].", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 311, + 466 + ], + "score": 1.0, + "content": "Let us consider the space of probability measures", + "type": "text" + }, + { + "bbox": [ + 311, + 453, + 343, + 466 + ], + "score": 0.92, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 453, + 505, + 466 + ], + "score": 1.0, + "content": ", that is, the set of probability measures", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 464, + 333, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 316, + 476 + ], + "score": 1.0, + "content": "with finite second moments defined on the manifold", + "type": "text" + }, + { + "bbox": [ + 316, + 466, + 329, + 474 + ], + "score": 0.83, + "content": "\\mathcal { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 464, + 333, + 476 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 442, + 505, + 476 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 476, + 451, + 504 + ], + "lines": [ + { + "bbox": [ + 159, + 476, + 451, + 504 + ], + "spans": [ + { + "bbox": [ + 159, + 476, + 451, + 504 + ], + "score": 0.96, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } ) = \\left\\{ \\varphi : \\mathcal { M } \\to [ 0 , \\infty ) \\bigg | \\int _ { \\mathcal { M } } \\mathrm { d } \\varphi = 1 , ~ \\int _ { \\mathcal { M } } | x | ^ { 2 } \\varphi ( x ) \\mathrm { d } x < + \\infty \\right\\} .", + "type": "interline_equation", + "image_path": "c1e5d77c719b4d2fa7a0247805d6e0144779c241f5db03b43a6f1d113276ff19.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 159, + 476, + 451, + 485.3333333333333 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 159, + 485.3333333333333, + 451, + 494.66666666666663 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 159, + 494.66666666666663, + 451, + 503.99999999999994 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 109, + 506, + 502, + 529 + ], + "lines": [ + { + "bbox": [ + 107, + 505, + 504, + 519 + ], + "spans": [ + { + "bbox": [ + 107, + 505, + 137, + 519 + ], + "score": 1.0, + "content": "Taking", + "type": "text" + }, + { + "bbox": [ + 138, + 506, + 170, + 518 + ], + "score": 0.92, + "content": "\\Pi ( \\mu , \\nu )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 505, + 405, + 519 + ], + "score": 1.0, + "content": "as the set of joint probability measures with marginals", + "type": "text" + }, + { + "bbox": [ + 406, + 508, + 423, + 518 + ], + "score": 0.87, + "content": "\\mu , \\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 505, + 504, + 519 + ], + "score": 1.0, + "content": ", we can define the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 107, + 516, + 282, + 530 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 235, + 530 + ], + "score": 1.0, + "content": "Wasserstein metric on the space", + "type": "text" + }, + { + "bbox": [ + 236, + 517, + 268, + 529 + ], + "score": 0.93, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 516, + 282, + 530 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 107, + 505, + 504, + 530 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 531, + 393, + 557 + ], + "lines": [ + { + "bbox": [ + 217, + 531, + 393, + 557 + ], + "spans": [ + { + "bbox": [ + 217, + 531, + 393, + 557 + ], + "score": 0.94, + "content": "W _ { 2 } ^ { 2 } ( \\mu , \\nu ) = \\operatorname* { i n f } _ { \\pi \\in \\Pi ( \\mu , \\nu ) } \\int | x - y | ^ { 2 } { \\mathrm { d } } \\pi ( x , y ) .", + "type": "interline_equation", + "image_path": "22875a9421a3c955c184f36bee03fad9988e4e0d847133375befe3e5951b4f25.jpg" + } + ] + } + ], + "index": 32, + "virtual_lines": [ + { + "bbox": [ + 217, + 531, + 393, + 557 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 504, + 583 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 323, + 574 + ], + "score": 1.0, + "content": "Considering the optimization problem of a functional", + "type": "text" + }, + { + "bbox": [ + 324, + 560, + 393, + 573 + ], + "score": 0.92, + "content": "J : \\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 559, + 443, + 574 + ], + "score": 1.0, + "content": ", such as the", + "type": "text" + }, + { + "bbox": [ + 443, + 561, + 458, + 571 + ], + "score": 0.39, + "content": "\\mathrm { K L }", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 559, + 505, + 574 + ], + "score": 1.0, + "content": "divergence", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 571, + 409, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 276, + 584 + ], + "score": 1.0, + "content": "between the particle approximation in Eq.", + "type": "text" + }, + { + "bbox": [ + 276, + 571, + 289, + 585 + ], + "score": 0.83, + "content": "\\textcircled { 8 }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 571, + 385, + 584 + ], + "score": 1.0, + "content": "and the target posterior", + "type": "text" + }, + { + "bbox": [ + 385, + 572, + 405, + 583 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 571, + 409, + 584 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 559, + 505, + 585 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 586, + 423, + 612 + ], + "lines": [ + { + "bbox": [ + 187, + 586, + 423, + 612 + ], + "spans": [ + { + "bbox": [ + 187, + 586, + 423, + 612 + ], + "score": 0.93, + "content": "\\operatorname* { i n f } _ { \\rho \\in \\mathcal { P } _ { 2 } ( \\mathcal { M } ) } D _ { K L } ( \\rho , \\pi ) = \\int _ { \\mathcal { M } } ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\rho ( x ) \\mathrm { d } x ,", + "type": "interline_equation", + "image_path": "202eca2cfd27979881321c7e7947951e3cc456330188130d02fcab3cc4c0abde.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 187, + 586, + 423, + 612 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 252, + 626 + ], + "score": 1.0, + "content": "the evolution in time of the measure", + "type": "text" + }, + { + "bbox": [ + 252, + 616, + 259, + 626 + ], + "score": 0.8, + "content": "\\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "under the equivalent of the gradient, the Wasserstein gradient", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 623, + 331, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 331, + 638 + ], + "score": 1.0, + "content": "flow, is described by the Liouville equation2 [38, 3, 64]:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 613, + 505, + 638 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 639, + 402, + 695 + ], + "lines": [ + { + "bbox": [ + 208, + 639, + 402, + 695 + ], + "spans": [ + { + "bbox": [ + 208, + 639, + 402, + 695 + ], + "score": 0.94, + "content": "\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( \\boldsymbol { x } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\big ( \\log \\rho ( \\boldsymbol { x } ) - \\log \\pi ( \\boldsymbol { x } ) \\big ) \\right) , } \\end{array}", + "type": "interline_equation", + "image_path": "319c5dba11def8b47c62bbed7e518cec22d2bb6447df21ccd8c8f0ceacda0468.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 639, + 402, + 653.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 208, + 653.0, + 402, + 667.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 208, + 667.0, + 402, + 681.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 208, + 681.0, + 402, + 695.0 + ], + "spans": [], + "index": 41 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 71, + 506, + 108 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 496, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 135, + 86 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 72, + 289, + 86 + ], + "score": 0.93, + "content": "\\nabla _ { \\delta \\rho } ^ { \\delta } D _ { K L } ( \\rho , \\pi ) \\ = : \\ \\nabla _ { \\mathcal { W } _ { 2 } } D _ { K L } ( \\rho , \\pi )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 70, + 484, + 86 + ], + "score": 1.0, + "content": "is the Wasserstein gradient and the operator", + "type": "text" + }, + { + "bbox": [ + 485, + 72, + 496, + 87 + ], + "score": 0.7, + "content": "\\frac { \\delta } { \\delta \\rho }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 85, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 107, + 86, + 167, + 98 + ], + "score": 0.91, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\stackrel { \\cdot } { } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 86, + 453, + 98 + ], + "score": 1.0, + "content": "represents the functional derivative or first variation (see Appendix", + "type": "text" + }, + { + "bbox": [ + 454, + 85, + 465, + 99 + ], + "score": 0.87, + "content": "\\boxed { \\mathsf { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 86, + 506, + 98 + ], + "score": 1.0, + "content": "for more", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 96, + 491, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 491, + 110 + ], + "score": 1.0, + "content": "details). In the particular case of the KL functional, we can recover the Fokker-Planck equation,", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 109, + 396, + 150 + ], + "lines": [ + { + "bbox": [ + 213, + 109, + 396, + 150 + ], + "spans": [ + { + "bbox": [ + 213, + 109, + 396, + 150 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } & { } & { \\frac { \\partial \\rho ( x ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( x ) \\nabla ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\right) } \\\\ & { } & { = - \\nabla \\cdot \\left( \\rho ( x ) \\nabla \\log \\pi ( x ) \\right) + \\nabla ^ { 2 } \\rho ( x ) , } \\end{array}", + "type": "interline_equation", + "image_path": "bc5069e97a8be7662b9fea29c7dc30021ca7cea5d088b0d8d45da5c537562e37.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 213, + 109, + 396, + 122.66666666666667 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 213, + 122.66666666666667, + 396, + 136.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 213, + 136.33333333333334, + 396, + 150.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 504, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 335, + 162 + ], + "score": 1.0, + "content": "that admits as unique stationary distribution the posterior", + "type": "text" + }, + { + "bbox": [ + 335, + 150, + 356, + 162 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 149, + 505, + 162 + ], + "score": 1.0, + "content": ". The deterministic particle dynamics", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 459, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 129, + 173 + ], + "score": 1.0, + "content": "ODE", + "type": "text" + }, + { + "bbox": [ + 130, + 160, + 142, + 172 + ], + "score": 0.61, + "content": "\\left[ \\left[ 2 \\right] \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 159, + 199, + 173 + ], + "score": 1.0, + "content": "related to Eq.", + "type": "text" + }, + { + "bbox": [ + 200, + 160, + 217, + 173 + ], + "score": 0.58, + "content": "\\mathbb { ( 1 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 159, + 459, + 173 + ], + "score": 1.0, + "content": ", namely mean-field Wasserstein dynamics, is then given by:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 173, + 375, + 196 + ], + "lines": [ + { + "bbox": [ + 236, + 173, + 375, + 196 + ], + "spans": [ + { + "bbox": [ + 236, + 173, + 375, + 196 + ], + "score": 0.94, + "content": "\\frac { d x } { d t } = - \\nabla \\big ( \\log \\rho ( x ) - \\log \\pi ( x ) \\big ) .", + "type": "interline_equation", + "image_path": "dd2081dc85db6793f99ac9a48cbba2bf3fdd0b388e3458acd6fdadb5a243ac7b.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 236, + 173, + 375, + 196 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 201, + 504, + 225 + ], + "lines": [ + { + "bbox": [ + 106, + 200, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 354, + 216 + ], + "score": 1.0, + "content": "Considering a discretization of Eq. (12) for a particle system", + "type": "text" + }, + { + "bbox": [ + 354, + 202, + 384, + 215 + ], + "score": 0.93, + "content": "\\{ x \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 200, + 461, + 216 + ], + "score": 1.0, + "content": "and small stepsize", + "type": "text" + }, + { + "bbox": [ + 461, + 205, + 470, + 213 + ], + "score": 0.83, + "content": "\\epsilon _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 200, + 506, + 216 + ], + "score": 1.0, + "content": ", we can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 209, + 186, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 209, + 153, + 227 + ], + "score": 1.0, + "content": "rewrite Eq.", + "type": "text" + }, + { + "bbox": [ + 153, + 213, + 171, + 226 + ], + "score": 0.7, + "content": "\\mathbb { \\underline { { ( 1 2 ) } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 209, + 186, + 227 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 223, + 397, + 238 + ], + "lines": [ + { + "bbox": [ + 213, + 223, + 397, + 238 + ], + "spans": [ + { + "bbox": [ + 213, + 223, + 397, + 238 + ], + "score": 0.91, + "content": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigl ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\nabla \\log \\rho ( x _ { i } ^ { t } ) \\bigr ) .", + "type": "interline_equation", + "image_path": "8a84c16f9a2296b238e38136d0db6478f4239c987d86ac3eef254ae1bde2a3a0.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 213, + 223, + 397, + 238 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 506, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 414, + 250 + ], + "score": 1.0, + "content": "Unfortunately, we do not have access to the analytical form of the gradient", + "type": "text" + }, + { + "bbox": [ + 414, + 237, + 444, + 248 + ], + "score": 0.9, + "content": "\\nabla \\log \\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 236, + 507, + 250 + ], + "score": 1.0, + "content": ", so an approx-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 249, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 259 + ], + "score": 1.0, + "content": "imation is needed. At this point, it is crucial to observe the similarity between the discretization", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 270, + 272 + ], + "score": 1.0, + "content": "of the Wasserstein gradient flow in Eq.", + "type": "text" + }, + { + "bbox": [ + 270, + 258, + 288, + 272 + ], + "score": 0.89, + "content": "\\textcircled { 1 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 258, + 420, + 272 + ], + "score": 1.0, + "content": "and the repulsive update in Eq.", + "type": "text" + }, + { + "bbox": [ + 420, + 258, + 434, + 271 + ], + "score": 0.87, + "content": "( 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 258, + 506, + 272 + ], + "score": 1.0, + "content": "to notice how, if", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "the kernelized repulsion is an approximation of the gradient of the empirical particle measure, the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 280, + 507, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 507, + 293 + ], + "score": 1.0, + "content": "update rule minimizes the KL divergence between the particle measure and the target posterior.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "Different sample-based approximations of the gradient that use a kernel function have been recently", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 492, + 315 + ], + "score": 1.0, + "content": "studied. The simplest one is given by the kernel density estimation (KDE) (details in Appendix", + "type": "text" + }, + { + "bbox": [ + 492, + 302, + 505, + 315 + ], + "score": 0.68, + "content": "\\dot { \\mathrm { ~ E ~ } } )", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 311, + 507, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 210, + 326 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\tilde { \\rho } _ { t } ( x ) = \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\dot { k } ( x , x _ { t } ^ { i } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 311, + 241, + 328 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 241, + 313, + 333, + 325 + ], + "score": 0.91, + "content": "\\boldsymbol { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\dot { \\mathbb { R } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 311, + 507, + 328 + ], + "score": 1.0, + "content": "and the gradient of its log density is given", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 323, + 141, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 119, + 338 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 325, + 137, + 336 + ], + "score": 0.75, + "content": "\\mathbb { \\ m }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 324, + 141, + 337 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 231, + 334, + 380, + 365 + ], + "lines": [ + { + "bbox": [ + 231, + 334, + 380, + 365 + ], + "spans": [ + { + "bbox": [ + 231, + 334, + 380, + 365 + ], + "score": 0.95, + "content": "\\nabla \\log \\rho ( x _ { i } ^ { t } ) \\approx \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } .", + "type": "interline_equation", + "image_path": "690e51dceea0054b390381bde15c30e9e687f9c712f52307c2e07985d9200e39.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 231, + 334, + 380, + 349.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 231, + 349.5, + 380, + 365.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 299, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 300, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 236, + 377 + ], + "score": 1.0, + "content": "Using this approximation in Eq.", + "type": "text" + }, + { + "bbox": [ + 237, + 363, + 254, + 376 + ], + "score": 0.82, + "content": "( 1 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 362, + 300, + 377 + ], + "score": 1.0, + "content": "we obtain:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 375, + 417, + 406 + ], + "lines": [ + { + "bbox": [ + 193, + 375, + 417, + 406 + ], + "spans": [ + { + "bbox": [ + 193, + 375, + 417, + 406 + ], + "score": 0.93, + "content": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\Bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } \\Bigg ) ,", + "type": "interline_equation", + "image_path": "371ffb0715cb9a3b57e74ec4c78b9d2ce1313431a8e5f80ccc30135a0cfa4053.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 375, + 417, + 390.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 193, + 390.5, + 417, + 406.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 278, + 419 + ], + "score": 1.0, + "content": "where, if we substitute the posterior for", + "type": "text" + }, + { + "bbox": [ + 279, + 409, + 286, + 416 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 405, + 505, + 419 + ], + "score": 1.0, + "content": ", we obtain an expression for the repulsive force in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 415, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 123, + 432 + ], + "score": 1.0, + "content": "Eq.", + "type": "text" + }, + { + "bbox": [ + 123, + 417, + 135, + 430 + ], + "score": 0.84, + "content": "\\textcircled{3}", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 415, + 314, + 432 + ], + "score": 1.0, + "content": ". This shows that if the repulsive term in Eq.", + "type": "text" + }, + { + "bbox": [ + 314, + 417, + 327, + 430 + ], + "score": 0.88, + "content": "( 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 415, + 483, + 432 + ], + "score": 1.0, + "content": "is the normalized sum of the gradients", + "type": "text" + }, + { + "bbox": [ + 484, + 417, + 506, + 428 + ], + "score": 0.85, + "content": "\\mathcal { R } =", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 425, + 510, + 455 + ], + "spans": [ + { + "bbox": [ + 107, + 428, + 288, + 450 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\left( \\bar { \\sum _ { j = 1 } ^ { n } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right) ^ { - 1 } \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 425, + 510, + 455 + ], + "score": 1.0, + "content": ", we do not only encourage diversity of the ensemble", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 447, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 426, + 460 + ], + "score": 1.0, + "content": "members and thus avoid collapse, but surprisingly—in the asymptotic limit of", + "type": "text" + }, + { + "bbox": [ + 459, + 447, + 504, + 460 + ], + "score": 1.0, + "content": ", where the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 458, + 429, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 223, + 472 + ], + "score": 1.0, + "content": "KDE approximation is exact", + "type": "text" + }, + { + "bbox": [ + 223, + 459, + 240, + 470 + ], + "score": 0.77, + "content": "\\left[ \\left[ 6 6 \\right] \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 458, + 429, + 472 + ], + "score": 1.0, + "content": "—also converge to the true Bayesian posterior!", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 506, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "score": 1.0, + "content": "Nevertheless, approximating the gradient of the empirical measure with the KDE can lead to subop-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 333, + 498 + ], + "score": 1.0, + "content": "timal performance, as already studied by Li and Turner", + "type": "text" + }, + { + "bbox": [ + 333, + 486, + 351, + 497 + ], + "score": 0.84, + "content": "[ \\overline { { | 4 6 | } }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 486, + 506, + 498 + ], + "score": 1.0, + "content": ". They instead introduced a new Stein", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "gradient estimator (SGE) that offers better performance, while maintaining the same computational", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 245, + 520 + ], + "score": 1.0, + "content": "cost. Even more recently, Shi et al.", + "type": "text" + }, + { + "bbox": [ + 245, + 508, + 263, + 519 + ], + "score": 0.85, + "content": "\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { A } } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "introduced a spectral method for gradient estimation (SSGE),", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "that also allows for a simple estimation on out-of-sample points. These two estimators can be used in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 529, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 122, + 544 + ], + "score": 1.0, + "content": "Eq.", + "type": "text" + }, + { + "bbox": [ + 123, + 529, + 140, + 543 + ], + "score": 0.77, + "content": "\\textcircled { 1 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 529, + 506, + 544 + ], + "score": 1.0, + "content": ", to formulate the following update rules with two alternative repulsive forces. The one using", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 541, + 315, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 315, + 552 + ], + "score": 1.0, + "content": "the Stein estimator, that we will call SGE-WGD, is:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 552, + 444, + 585 + ], + "lines": [ + { + "bbox": [ + 165, + 552, + 444, + 585 + ], + "spans": [ + { + "bbox": [ + 165, + 552, + 444, + 585 + ], + "score": 0.94, + "content": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) + \\sum _ { j = 1 } ^ { n } ( K + \\eta \\mathbb { I } ) _ { i j } ^ { - 1 } \\sum _ { k = 1 } ^ { n } \\nabla _ { x _ { k } ^ { t } } k ( x _ { k } ^ { t } , x _ { j } ^ { t } ) \\bigg ) ,", + "type": "interline_equation", + "image_path": "fc8beab1c532557f6b4c53d364e88ab075d4917c7571f2c182526373978fd6c4.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 165, + 552, + 444, + 563.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 165, + 563.0, + 444, + 574.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 165, + 574.0, + 444, + 585.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 133, + 598 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 586, + 144, + 596 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 585, + 253, + 598 + ], + "score": 1.0, + "content": "is the kernel Gram matrix,", + "type": "text" + }, + { + "bbox": [ + 253, + 588, + 260, + 597 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "a small constant, and I the identity matrix. We can notice an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "important difference between KDE and SGE, in that the former is only considering the interaction of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 121, + 620 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 609, + 126, + 618 + ], + "score": 0.81, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "-th particle being updated with all the others, while the latter is simultaneously considering also", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "the interactions between the remaining particles. The spectral method, that we will call SSGE-WGD,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 630, + 245, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 245, + 641 + ], + "score": 1.0, + "content": "leads to the following update rule:", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + }, + { + "type": "interline_equation", + "bbox": [ + 121, + 641, + 473, + 676 + ], + "lines": [ + { + "bbox": [ + 121, + 641, + 473, + 676 + ], + "spans": [ + { + "bbox": [ + 121, + 641, + 473, + 676 + ], + "score": 0.94, + "content": "x _ { t + 1 } ^ { i } = x _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { t } ^ { i } ) + \\sum _ { j = 1 } ^ { J } \\frac { 1 } { \\lambda _ { j } ^ { 2 } } \\sum _ { m = 1 } ^ { n } \\sum _ { k = 1 } ^ { n } u _ { j k } \\nabla _ { x _ { m } } k ( x _ { m } ^ { t } , x _ { k } ^ { t } ) \\cdot \\sum _ { l = 1 } ^ { n } u _ { j l } k ( x _ { i } ^ { t } , x _ { l } ^ { t } ) \\bigg )", + "type": "interline_equation", + "image_path": "22b05cdc924e14a9f81460eed44ce5ae2949169ff1032670bb23204d0bc3d3f9.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 121, + 641, + 473, + 652.6666666666666 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 121, + 652.6666666666666, + 473, + 664.3333333333333 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 121, + 664.3333333333333, + 473, + 675.9999999999999 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 134, + 690 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 677, + 146, + 689 + ], + "score": 0.89, + "content": "\\lambda _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 676, + 174, + 690 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 174, + 678, + 180, + 689 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 676, + 345, + 690 + ], + "score": 1.0, + "content": "-th eigenvalue of the kernel matrix and", + "type": "text" + }, + { + "bbox": [ + 346, + 679, + 361, + 689 + ], + "score": 0.83, + "content": "u _ { j k }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 676, + 389, + 690 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 390, + 677, + 396, + 687 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 676, + 486, + 690 + ], + "score": 1.0, + "content": "-th component of the", + "type": "text" + }, + { + "bbox": [ + 487, + 678, + 492, + 689 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "-th", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 378, + 703 + ], + "score": 1.0, + "content": "eigenvector. Computationally, both SSGE and SGE have a cost of", + "type": "text" + }, + { + "bbox": [ + 378, + 688, + 442, + 701 + ], + "score": 0.91, + "content": "\\mathcal { O } ( M ^ { 3 } + M ^ { 2 } d )", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 687, + 466, + 703 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 467, + 690, + 478, + 699 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "being", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 213, + 712 + ], + "score": 1.0, + "content": "the number of points and", + "type": "text" + }, + { + "bbox": [ + 214, + 700, + 220, + 710 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "their dimensionality. SSGE has an additional cost for predictions of", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 107, + 711, + 324, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 711, + 165, + 723 + ], + "score": 0.92, + "content": "\\mathcal { O } ( M ( d + J ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 711, + 196, + 723 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 197, + 711, + 204, + 721 + ], + "score": 0.83, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 711, + 324, + 723 + ], + "score": 1.0, + "content": "is the number of eigenvalues.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 71, + 506, + 108 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 496, + 87 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 135, + 86 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 72, + 289, + 86 + ], + "score": 0.93, + "content": "\\nabla _ { \\delta \\rho } ^ { \\delta } D _ { K L } ( \\rho , \\pi ) \\ = : \\ \\nabla _ { \\mathcal { W } _ { 2 } } D _ { K L } ( \\rho , \\pi )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 70, + 484, + 86 + ], + "score": 1.0, + "content": "is the Wasserstein gradient and the operator", + "type": "text" + }, + { + "bbox": [ + 485, + 72, + 496, + 87 + ], + "score": 0.7, + "content": "\\frac { \\delta } { \\delta \\rho }", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 85, + 506, + 99 + ], + "spans": [ + { + "bbox": [ + 107, + 86, + 167, + 98 + ], + "score": 0.91, + "content": "\\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\stackrel { \\cdot } { } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 86, + 453, + 98 + ], + "score": 1.0, + "content": "represents the functional derivative or first variation (see Appendix", + "type": "text" + }, + { + "bbox": [ + 454, + 85, + 465, + 99 + ], + "score": 0.87, + "content": "\\boxed { \\mathsf { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 86, + 506, + 98 + ], + "score": 1.0, + "content": "for more", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 96, + 491, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 491, + 110 + ], + "score": 1.0, + "content": "details). In the particular case of the KL functional, we can recover the Fokker-Planck equation,", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 104, + 70, + 506, + 110 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 109, + 396, + 150 + ], + "lines": [ + { + "bbox": [ + 213, + 109, + 396, + 150 + ], + "spans": [ + { + "bbox": [ + 213, + 109, + 396, + 150 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } & { } & { \\frac { \\partial \\rho ( x ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( x ) \\nabla ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\right) } \\\\ & { } & { = - \\nabla \\cdot \\left( \\rho ( x ) \\nabla \\log \\pi ( x ) \\right) + \\nabla ^ { 2 } \\rho ( x ) , } \\end{array}", + "type": "interline_equation", + "image_path": "bc5069e97a8be7662b9fea29c7dc30021ca7cea5d088b0d8d45da5c537562e37.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 213, + 109, + 396, + 122.66666666666667 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 213, + 122.66666666666667, + 396, + 136.33333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 213, + 136.33333333333334, + 396, + 150.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 150, + 504, + 172 + ], + "lines": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 335, + 162 + ], + "score": 1.0, + "content": "that admits as unique stationary distribution the posterior", + "type": "text" + }, + { + "bbox": [ + 335, + 150, + 356, + 162 + ], + "score": 0.92, + "content": "\\pi ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 149, + 505, + 162 + ], + "score": 1.0, + "content": ". The deterministic particle dynamics", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 459, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 129, + 173 + ], + "score": 1.0, + "content": "ODE", + "type": "text" + }, + { + "bbox": [ + 130, + 160, + 142, + 172 + ], + "score": 0.61, + "content": "\\left[ \\left[ 2 \\right] \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 159, + 199, + 173 + ], + "score": 1.0, + "content": "related to Eq.", + "type": "text" + }, + { + "bbox": [ + 200, + 160, + 217, + 173 + ], + "score": 0.58, + "content": "\\mathbb { ( 1 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 159, + 459, + 173 + ], + "score": 1.0, + "content": ", namely mean-field Wasserstein dynamics, is then given by:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 149, + 505, + 173 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 173, + 375, + 196 + ], + "lines": [ + { + "bbox": [ + 236, + 173, + 375, + 196 + ], + "spans": [ + { + "bbox": [ + 236, + 173, + 375, + 196 + ], + "score": 0.94, + "content": "\\frac { d x } { d t } = - \\nabla \\big ( \\log \\rho ( x ) - \\log \\pi ( x ) \\big ) .", + "type": "interline_equation", + "image_path": "dd2081dc85db6793f99ac9a48cbba2bf3fdd0b388e3458acd6fdadb5a243ac7b.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 236, + 173, + 375, + 196 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 201, + 504, + 225 + ], + "lines": [ + { + "bbox": [ + 106, + 200, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 354, + 216 + ], + "score": 1.0, + "content": "Considering a discretization of Eq. (12) for a particle system", + "type": "text" + }, + { + "bbox": [ + 354, + 202, + 384, + 215 + ], + "score": 0.93, + "content": "\\{ x \\} _ { i = 1 } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 200, + 461, + 216 + ], + "score": 1.0, + "content": "and small stepsize", + "type": "text" + }, + { + "bbox": [ + 461, + 205, + 470, + 213 + ], + "score": 0.83, + "content": "\\epsilon _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 200, + 506, + 216 + ], + "score": 1.0, + "content": ", we can", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 209, + 186, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 209, + 153, + 227 + ], + "score": 1.0, + "content": "rewrite Eq.", + "type": "text" + }, + { + "bbox": [ + 153, + 213, + 171, + 226 + ], + "score": 0.7, + "content": "\\mathbb { \\underline { { ( 1 2 ) } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 209, + 186, + 227 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 200, + 506, + 227 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 213, + 223, + 397, + 238 + ], + "lines": [ + { + "bbox": [ + 213, + 223, + 397, + 238 + ], + "spans": [ + { + "bbox": [ + 213, + 223, + 397, + 238 + ], + "score": 0.91, + "content": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigl ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\nabla \\log \\rho ( x _ { i } ^ { t } ) \\bigr ) .", + "type": "interline_equation", + "image_path": "8a84c16f9a2296b238e38136d0db6478f4239c987d86ac3eef254ae1bde2a3a0.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 213, + 223, + 397, + 238 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 506, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 507, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 414, + 250 + ], + "score": 1.0, + "content": "Unfortunately, we do not have access to the analytical form of the gradient", + "type": "text" + }, + { + "bbox": [ + 414, + 237, + 444, + 248 + ], + "score": 0.9, + "content": "\\nabla \\log \\rho", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 236, + 507, + 250 + ], + "score": 1.0, + "content": ", so an approx-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 249, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 259 + ], + "score": 1.0, + "content": "imation is needed. At this point, it is crucial to observe the similarity between the discretization", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 270, + 272 + ], + "score": 1.0, + "content": "of the Wasserstein gradient flow in Eq.", + "type": "text" + }, + { + "bbox": [ + 270, + 258, + 288, + 272 + ], + "score": 0.89, + "content": "\\textcircled { 1 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 258, + 420, + 272 + ], + "score": 1.0, + "content": "and the repulsive update in Eq.", + "type": "text" + }, + { + "bbox": [ + 420, + 258, + 434, + 271 + ], + "score": 0.87, + "content": "( 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 258, + 506, + 272 + ], + "score": 1.0, + "content": "to notice how, if", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "the kernelized repulsion is an approximation of the gradient of the empirical particle measure, the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 280, + 507, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 507, + 293 + ], + "score": 1.0, + "content": "update rule minimizes the KL divergence between the particle measure and the target posterior.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "Different sample-based approximations of the gradient that use a kernel function have been recently", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 302, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 492, + 315 + ], + "score": 1.0, + "content": "studied. The simplest one is given by the kernel density estimation (KDE) (details in Appendix", + "type": "text" + }, + { + "bbox": [ + 492, + 302, + 505, + 315 + ], + "score": 0.68, + "content": "\\dot { \\mathrm { ~ E ~ } } )", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 311, + 507, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 210, + 326 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\tilde { \\rho } _ { t } ( x ) = \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\dot { k } ( x , x _ { t } ^ { i } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 311, + 241, + 328 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 241, + 313, + 333, + 325 + ], + "score": 0.91, + "content": "\\boldsymbol { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\dot { \\mathbb { R } }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 311, + 507, + 328 + ], + "score": 1.0, + "content": "and the gradient of its log density is given", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 323, + 141, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 119, + 338 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 119, + 325, + 137, + 336 + ], + "score": 0.75, + "content": "\\mathbb { \\ m }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 324, + 141, + 337 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 236, + 507, + 338 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 231, + 334, + 380, + 365 + ], + "lines": [ + { + "bbox": [ + 231, + 334, + 380, + 365 + ], + "spans": [ + { + "bbox": [ + 231, + 334, + 380, + 365 + ], + "score": 0.95, + "content": "\\nabla \\log \\rho ( x _ { i } ^ { t } ) \\approx \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } .", + "type": "interline_equation", + "image_path": "690e51dceea0054b390381bde15c30e9e687f9c712f52307c2e07985d9200e39.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 231, + 334, + 380, + 349.5 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 231, + 349.5, + 380, + 365.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 299, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 300, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 236, + 377 + ], + "score": 1.0, + "content": "Using this approximation in Eq.", + "type": "text" + }, + { + "bbox": [ + 237, + 363, + 254, + 376 + ], + "score": 0.82, + "content": "( 1 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 362, + 300, + 377 + ], + "score": 1.0, + "content": "we obtain:", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 362, + 300, + 377 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 193, + 375, + 417, + 406 + ], + "lines": [ + { + "bbox": [ + 193, + 375, + 417, + 406 + ], + "spans": [ + { + "bbox": [ + 193, + 375, + 417, + 406 + ], + "score": 0.93, + "content": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\Bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } \\Bigg ) ,", + "type": "interline_equation", + "image_path": "371ffb0715cb9a3b57e74ec4c78b9d2ce1313431a8e5f80ccc30135a0cfa4053.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 193, + 375, + 417, + 390.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 193, + 390.5, + 417, + 406.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 406, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 405, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 278, + 419 + ], + "score": 1.0, + "content": "where, if we substitute the posterior for", + "type": "text" + }, + { + "bbox": [ + 279, + 409, + 286, + 416 + ], + "score": 0.77, + "content": "\\pi", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 405, + 505, + 419 + ], + "score": 1.0, + "content": ", we obtain an expression for the repulsive force in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 415, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 123, + 432 + ], + "score": 1.0, + "content": "Eq.", + "type": "text" + }, + { + "bbox": [ + 123, + 417, + 135, + 430 + ], + "score": 0.84, + "content": "\\textcircled{3}", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 415, + 314, + 432 + ], + "score": 1.0, + "content": ". This shows that if the repulsive term in Eq.", + "type": "text" + }, + { + "bbox": [ + 314, + 417, + 327, + 430 + ], + "score": 0.88, + "content": "( 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 415, + 483, + 432 + ], + "score": 1.0, + "content": "is the normalized sum of the gradients", + "type": "text" + }, + { + "bbox": [ + 484, + 417, + 506, + 428 + ], + "score": 0.85, + "content": "\\mathcal { R } =", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 425, + 510, + 455 + ], + "spans": [ + { + "bbox": [ + 107, + 428, + 288, + 450 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\left( \\bar { \\sum _ { j = 1 } ^ { n } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right) ^ { - 1 } \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 425, + 510, + 455 + ], + "score": 1.0, + "content": ", we do not only encourage diversity of the ensemble", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 447, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 426, + 460 + ], + "score": 1.0, + "content": "members and thus avoid collapse, but surprisingly—in the asymptotic limit of", + "type": "text" + }, + { + "bbox": [ + 459, + 447, + 504, + 460 + ], + "score": 1.0, + "content": ", where the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 458, + 429, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 223, + 472 + ], + "score": 1.0, + "content": "KDE approximation is exact", + "type": "text" + }, + { + "bbox": [ + 223, + 459, + 240, + 470 + ], + "score": 0.77, + "content": "\\left[ \\left[ 6 6 \\right] \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 458, + 429, + 472 + ], + "score": 1.0, + "content": "—also converge to the true Bayesian posterior!", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 405, + 510, + 472 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 474, + 506, + 552 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 507, + 488 + ], + "score": 1.0, + "content": "Nevertheless, approximating the gradient of the empirical measure with the KDE can lead to subop-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 333, + 498 + ], + "score": 1.0, + "content": "timal performance, as already studied by Li and Turner", + "type": "text" + }, + { + "bbox": [ + 333, + 486, + 351, + 497 + ], + "score": 0.84, + "content": "[ \\overline { { | 4 6 | } }", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 486, + 506, + 498 + ], + "score": 1.0, + "content": ". They instead introduced a new Stein", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "gradient estimator (SGE) that offers better performance, while maintaining the same computational", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 245, + 520 + ], + "score": 1.0, + "content": "cost. Even more recently, Shi et al.", + "type": "text" + }, + { + "bbox": [ + 245, + 508, + 263, + 519 + ], + "score": 0.85, + "content": "\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { A } } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "introduced a spectral method for gradient estimation (SSGE),", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "that also allows for a simple estimation on out-of-sample points. These two estimators can be used in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 529, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 122, + 544 + ], + "score": 1.0, + "content": "Eq.", + "type": "text" + }, + { + "bbox": [ + 123, + 529, + 140, + 543 + ], + "score": 0.77, + "content": "\\textcircled { 1 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 529, + 506, + 544 + ], + "score": 1.0, + "content": ", to formulate the following update rules with two alternative repulsive forces. The one using", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 541, + 315, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 315, + 552 + ], + "score": 1.0, + "content": "the Stein estimator, that we will call SGE-WGD, is:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 474, + 507, + 552 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 552, + 444, + 585 + ], + "lines": [ + { + "bbox": [ + 165, + 552, + 444, + 585 + ], + "spans": [ + { + "bbox": [ + 165, + 552, + 444, + 585 + ], + "score": 0.94, + "content": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) + \\sum _ { j = 1 } ^ { n } ( K + \\eta \\mathbb { I } ) _ { i j } ^ { - 1 } \\sum _ { k = 1 } ^ { n } \\nabla _ { x _ { k } ^ { t } } k ( x _ { k } ^ { t } , x _ { j } ^ { t } ) \\bigg ) ,", + "type": "interline_equation", + "image_path": "fc8beab1c532557f6b4c53d364e88ab075d4917c7571f2c182526373978fd6c4.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 165, + 552, + 444, + 563.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 165, + 563.0, + 444, + 574.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 165, + 574.0, + 444, + 585.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 586, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 133, + 598 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 586, + 144, + 596 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 585, + 253, + 598 + ], + "score": 1.0, + "content": "is the kernel Gram matrix,", + "type": "text" + }, + { + "bbox": [ + 253, + 588, + 260, + 597 + ], + "score": 0.82, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "a small constant, and I the identity matrix. We can notice an", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "important difference between KDE and SGE, in that the former is only considering the interaction of", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 121, + 620 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 609, + 126, + 618 + ], + "score": 0.81, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "-th particle being updated with all the others, while the latter is simultaneously considering also", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 631 + ], + "score": 1.0, + "content": "the interactions between the remaining particles. The spectral method, that we will call SSGE-WGD,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 630, + 245, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 245, + 641 + ], + "score": 1.0, + "content": "leads to the following update rule:", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 585, + 506, + 641 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 121, + 641, + 473, + 676 + ], + "lines": [ + { + "bbox": [ + 121, + 641, + 473, + 676 + ], + "spans": [ + { + "bbox": [ + 121, + 641, + 473, + 676 + ], + "score": 0.94, + "content": "x _ { t + 1 } ^ { i } = x _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { t } ^ { i } ) + \\sum _ { j = 1 } ^ { J } \\frac { 1 } { \\lambda _ { j } ^ { 2 } } \\sum _ { m = 1 } ^ { n } \\sum _ { k = 1 } ^ { n } u _ { j k } \\nabla _ { x _ { m } } k ( x _ { m } ^ { t } , x _ { k } ^ { t } ) \\cdot \\sum _ { l = 1 } ^ { n } u _ { j l } k ( x _ { i } ^ { t } , x _ { l } ^ { t } ) \\bigg )", + "type": "interline_equation", + "image_path": "22b05cdc924e14a9f81460eed44ce5ae2949169ff1032670bb23204d0bc3d3f9.jpg" + } + ] + } + ], + "index": 47, + "virtual_lines": [ + { + "bbox": [ + 121, + 641, + 473, + 652.6666666666666 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 121, + 652.6666666666666, + 473, + 664.3333333333333 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 121, + 664.3333333333333, + 473, + 675.9999999999999 + ], + "spans": [], + "index": 48 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 134, + 690 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 677, + 146, + 689 + ], + "score": 0.89, + "content": "\\lambda _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 676, + 174, + 690 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 174, + 678, + 180, + 689 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 676, + 345, + 690 + ], + "score": 1.0, + "content": "-th eigenvalue of the kernel matrix and", + "type": "text" + }, + { + "bbox": [ + 346, + 679, + 361, + 689 + ], + "score": 0.83, + "content": "u _ { j k }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 676, + 389, + 690 + ], + "score": 1.0, + "content": "is the", + "type": "text" + }, + { + "bbox": [ + 390, + 677, + 396, + 687 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 676, + 486, + 690 + ], + "score": 1.0, + "content": "-th component of the", + "type": "text" + }, + { + "bbox": [ + 487, + 678, + 492, + 689 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "-th", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 378, + 703 + ], + "score": 1.0, + "content": "eigenvector. Computationally, both SSGE and SGE have a cost of", + "type": "text" + }, + { + "bbox": [ + 378, + 688, + 442, + 701 + ], + "score": 0.91, + "content": "\\mathcal { O } ( M ^ { 3 } + M ^ { 2 } d )", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 687, + 466, + 703 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 467, + 690, + 478, + 699 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "being", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 213, + 712 + ], + "score": 1.0, + "content": "the number of points and", + "type": "text" + }, + { + "bbox": [ + 214, + 700, + 220, + 710 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "their dimensionality. SSGE has an additional cost for predictions of", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 107, + 711, + 324, + 723 + ], + "spans": [ + { + "bbox": [ + 107, + 711, + 165, + 723 + ], + "score": 0.92, + "content": "\\mathcal { O } ( M ( d + J ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 711, + 196, + 723 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 197, + 711, + 204, + 721 + ], + "score": 0.83, + "content": "J", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 711, + 324, + 723 + ], + "score": 1.0, + "content": "is the number of eigenvalues.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50.5, + "bbox_fs": [ + 105, + 676, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 68, + 502, + 168 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 68, + 502, + 168 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 68, + 502, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 68, + 502, + 168 + ], + "score": 0.966, + "type": "image", + "image_path": "909c60e208223cd54336af828c72fc5777f5cbee78dd1b5dffc27345c64eca36.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 68, + 502, + 101.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 101.33333333333334, + 502, + 134.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 134.66666666666669, + 502, + 168.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 172, + 506, + 206 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 172, + 507, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 507, + 185 + ], + "score": 1.0, + "content": "Figure 2: Single Gaussian. We show samples from SVGD, KDE-WGD, SGE-WGD, and SSGE-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "WGD (from left to right). The upper and right plots show the empirical one-dimensional marginal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 194, + 435, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 435, + 207 + ], + "score": 1.0, + "content": "distributions obtained using KDE on the samples (red) and on the particles (blue).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 228, + 263, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 265, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 265, + 243 + ], + "score": 1.0, + "content": "3.3 Gradient flow in function space", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 249, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 410, + 262 + ], + "score": 1.0, + "content": "To theoretically justify the update rule introduced in function space in Eq.", + "type": "text" + }, + { + "bbox": [ + 410, + 249, + 423, + 262 + ], + "score": 0.54, + "content": "\\textcircled{5}", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 248, + 505, + 262 + ], + "score": 1.0, + "content": ", we can rewrite the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 260, + 394, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 292, + 273 + ], + "score": 1.0, + "content": "Liouville equation for the gradient flow in Eq.", + "type": "text" + }, + { + "bbox": [ + 293, + 260, + 310, + 273 + ], + "score": 0.79, + "content": "\\textcircled { 1 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 261, + 394, + 273 + ], + "score": 1.0, + "content": "in function space as", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 279, + 406, + 335 + ], + "lines": [ + { + "bbox": [ + 206, + 279, + 406, + 335 + ], + "spans": [ + { + "bbox": [ + 206, + 279, + 406, + 335 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( { \\pmb f } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\big ( \\log \\rho ( { \\pmb f } ) - \\log \\pi ( { \\pmb f } ) \\big ) \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "5c17155651b4b8a9222af9e3b20c3d19517ca6b37f38c6cba7442ed799860e6c.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 279, + 406, + 293.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 206, + 293.0, + 406, + 307.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 206, + 307.0, + 406, + 321.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 206, + 321.0, + 406, + 335.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 355, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 355, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 355, + 353 + ], + "score": 1.0, + "content": "Following this update, the mean field functional dynamics are", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 358, + 377, + 383 + ], + "lines": [ + { + "bbox": [ + 234, + 358, + 377, + 383 + ], + "spans": [ + { + "bbox": [ + 234, + 358, + 377, + 383 + ], + "score": 0.94, + "content": "\\frac { d \\pmb f } { d t } = - \\nabla \\big ( \\log \\rho ( \\pmb f ) - \\log \\pi ( \\pmb f ) \\big ) .", + "type": "interline_equation", + "image_path": "6ec49c1866059fd222d901856c6fd87a5e5117e19a3da56886cb693721fd87d0.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 234, + 358, + 377, + 383 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 388, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 402 + ], + "score": 1.0, + "content": "Using the same KDE approximation as above, we can obtain a discretized evolution in function space", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 398, + 361, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 334, + 414 + ], + "score": 1.0, + "content": "and with it an explicit form for the repulsive force in Eq.", + "type": "text" + }, + { + "bbox": [ + 334, + 399, + 347, + 413 + ], + "score": 0.86, + "content": "\\textcircled{4}", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 398, + 361, + 414 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 417, + 423, + 450 + ], + "lines": [ + { + "bbox": [ + 188, + 417, + 423, + 450 + ], + "spans": [ + { + "bbox": [ + 188, + 417, + 423, + 450 + ], + "score": 0.93, + "content": "\\pmb { f } _ { t + 1 } ^ { i } = \\pmb { f } _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla _ { \\pmb { f } } \\log \\pi ( \\pmb { f } _ { t } ^ { i } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { \\pmb { f } _ { i } ^ { t } } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } \\bigg ) .", + "type": "interline_equation", + "image_path": "413dc39054cf2ffe58aac203b1f546f1178b0fd741a1e970d847b042475c73d1.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 417, + 423, + 433.5 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 188, + 433.5, + 423, + 450.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 506, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 456, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 506, + 468 + ], + "score": 1.0, + "content": "The update rules using the SGE and SSGE approximations follow as for the parametric case. It is im-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 416, + 479 + ], + "score": 1.0, + "content": "portant to notice that this update rule requires the function space prior gradient:", + "type": "text" + }, + { + "bbox": [ + 417, + 466, + 506, + 479 + ], + "score": 0.91, + "content": "\\nabla _ { f _ { j } } \\log { p ( f _ { j } | x , y ) } =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 476, + 507, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 252, + 491 + ], + "score": 0.9, + "content": "\\nabla _ { f _ { j } } \\log p ( \\pmb { y } | \\pmb { x } , \\pmb { f _ { j } } ) + \\nabla _ { f _ { j } } \\log p ( \\pmb { f _ { j } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 476, + 507, + 493 + ], + "score": 1.0, + "content": ". If one wants to use an implicit prior defined in weight space,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "score": 1.0, + "content": "an additional estimator is needed due to its analytical intractability. We again adopted the SSGE,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 204, + 513 + ], + "score": 1.0, + "content": "introduced by Shi et al.", + "type": "text" + }, + { + "bbox": [ + 204, + 500, + 222, + 511 + ], + "score": 0.85, + "content": "\\mathbb { \\lVert \\boldsymbol { 6 9 } \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 499, + 464, + 513 + ], + "score": 1.0, + "content": ", which was already used for a similar purpose in Sun et al.", + "type": "text" + }, + { + "bbox": [ + 465, + 500, + 482, + 511 + ], + "score": 0.77, + "content": "\\pmb { \\mathbb { Z } 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 499, + 506, + 513 + ], + "score": 1.0, + "content": ". It is", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 303, + 524 + ], + "score": 1.0, + "content": "also interesting to note that the update rule in Eq.", + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 321, + 524 + ], + "score": 0.88, + "content": "\\textcircled { 1 2 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "readily allows for the use of alternative priors", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "that have an analytical form, such as Gaussian processes. This is an important feature of our method", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "that allows for an explicit encoding of function space properties that can be useful for example in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 543, + 353, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 334, + 555 + ], + "score": 1.0, + "content": "achieving better out of distribution detection capabilities", + "type": "text" + }, + { + "bbox": [ + 334, + 543, + 352, + 555 + ], + "score": 0.49, + "content": "\\checkmark", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 543, + 353, + 555 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 232, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 233, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 233, + 583 + ], + "score": 1.0, + "content": "3.4 The choice of the kernel", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "A repulsive effect can always be created in the ensemble by means of the gradient of any kernel", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "score": 1.0, + "content": "function that is measuring the similarity between two members. Nevertheless, it is important to keep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "in mind that to ensure the asymptotic convergence to the Bayesian posterior, the repulsive component", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 320, + 637 + ], + "score": 1.0, + "content": "must be a consistent estimator of the gradient in Eq.", + "type": "text" + }, + { + "bbox": [ + 320, + 623, + 338, + 636 + ], + "score": 0.52, + "content": "( 1 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 623, + 424, + 637 + ], + "score": 1.0, + "content": ", as shown in Section", + "type": "text" + }, + { + "bbox": [ + 424, + 623, + 432, + 636 + ], + "score": 0.68, + "content": "\\checkmark", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 623, + 506, + 637 + ], + "score": 1.0, + "content": ". Therefore, some", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 634, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 646 + ], + "score": 1.0, + "content": "important constraints over the kernel choice are needed. In particular, the SGE and SSGE need a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 338, + 657 + ], + "score": 1.0, + "content": "kernel function belonging to the Stein class (see Shi et al.", + "type": "text" + }, + { + "bbox": [ + 338, + 645, + 356, + 657 + ], + "score": 0.64, + "content": "\\dot { \\overline { { \\vert 6 9 \\vert } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "for more details). On the other hand,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "for the KDE, any symmetric probability density function can be used. On this subject, the work of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "Aggarwal et al. [1] has shown how the Manhattan distance metric (L1 norm) might be preferable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "over the Euclidean distance metric (L2 norm) for high-dimensional settings. We performed some", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "additional experiments using the L1 norm (Laplace kernel) for the KDE but we could not observe", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "any substantial difference compared to using the L2 norm. Further investigations regarding this", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 711, + 260, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 260, + 722 + ], + "score": 1.0, + "content": "hypothesis are left for future research.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 68, + 502, + 168 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 68, + 502, + 168 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 68, + 502, + 168 + ], + "spans": [ + { + "bbox": [ + 111, + 68, + 502, + 168 + ], + "score": 0.966, + "type": "image", + "image_path": "909c60e208223cd54336af828c72fc5777f5cbee78dd1b5dffc27345c64eca36.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 68, + 502, + 101.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 101.33333333333334, + 502, + 134.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 134.66666666666669, + 502, + 168.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 172, + 506, + 206 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 172, + 507, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 507, + 185 + ], + "score": 1.0, + "content": "Figure 2: Single Gaussian. We show samples from SVGD, KDE-WGD, SGE-WGD, and SSGE-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "WGD (from left to right). The upper and right plots show the empirical one-dimensional marginal", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 194, + 435, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 435, + 207 + ], + "score": 1.0, + "content": "distributions obtained using KDE on the samples (red) and on the particles (blue).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 228, + 263, + 240 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 265, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 265, + 243 + ], + "score": 1.0, + "content": "3.3 Gradient flow in function space", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 249, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 410, + 262 + ], + "score": 1.0, + "content": "To theoretically justify the update rule introduced in function space in Eq.", + "type": "text" + }, + { + "bbox": [ + 410, + 249, + 423, + 262 + ], + "score": 0.54, + "content": "\\textcircled{5}", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 248, + 505, + 262 + ], + "score": 1.0, + "content": ", we can rewrite the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 260, + 394, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 292, + 273 + ], + "score": 1.0, + "content": "Liouville equation for the gradient flow in Eq.", + "type": "text" + }, + { + "bbox": [ + 293, + 260, + 310, + 273 + ], + "score": 0.79, + "content": "\\textcircled { 1 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 261, + 394, + 273 + ], + "score": 1.0, + "content": "in function space as", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 248, + 505, + 273 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 279, + 406, + 335 + ], + "lines": [ + { + "bbox": [ + 206, + 279, + 406, + 335 + ], + "spans": [ + { + "bbox": [ + 206, + 279, + 406, + 335 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( { \\pmb f } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\big ( \\log \\rho ( { \\pmb f } ) - \\log \\pi ( { \\pmb f } ) \\big ) \\right) . } \\end{array}", + "type": "interline_equation", + "image_path": "5c17155651b4b8a9222af9e3b20c3d19517ca6b37f38c6cba7442ed799860e6c.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 206, + 279, + 406, + 293.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 206, + 293.0, + 406, + 307.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 206, + 307.0, + 406, + 321.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 206, + 321.0, + 406, + 335.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 340, + 355, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 355, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 355, + 353 + ], + "score": 1.0, + "content": "Following this update, the mean field functional dynamics are", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 340, + 355, + 353 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 358, + 377, + 383 + ], + "lines": [ + { + "bbox": [ + 234, + 358, + 377, + 383 + ], + "spans": [ + { + "bbox": [ + 234, + 358, + 377, + 383 + ], + "score": 0.94, + "content": "\\frac { d \\pmb f } { d t } = - \\nabla \\big ( \\log \\rho ( \\pmb f ) - \\log \\pi ( \\pmb f ) \\big ) .", + "type": "interline_equation", + "image_path": "6ec49c1866059fd222d901856c6fd87a5e5117e19a3da56886cb693721fd87d0.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 234, + 358, + 377, + 383 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 388, + 505, + 412 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 402 + ], + "score": 1.0, + "content": "Using the same KDE approximation as above, we can obtain a discretized evolution in function space", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 398, + 361, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 334, + 414 + ], + "score": 1.0, + "content": "and with it an explicit form for the repulsive force in Eq.", + "type": "text" + }, + { + "bbox": [ + 334, + 399, + 347, + 413 + ], + "score": 0.86, + "content": "\\textcircled{4}", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 398, + 361, + 414 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 387, + 505, + 414 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 188, + 417, + 423, + 450 + ], + "lines": [ + { + "bbox": [ + 188, + 417, + 423, + 450 + ], + "spans": [ + { + "bbox": [ + 188, + 417, + 423, + 450 + ], + "score": 0.93, + "content": "\\pmb { f } _ { t + 1 } ^ { i } = \\pmb { f } _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla _ { \\pmb { f } } \\log \\pi ( \\pmb { f } _ { t } ^ { i } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { \\pmb { f } _ { i } ^ { t } } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } \\bigg ) .", + "type": "interline_equation", + "image_path": "413dc39054cf2ffe58aac203b1f546f1178b0fd741a1e970d847b042475c73d1.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 417, + 423, + 433.5 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 188, + 433.5, + 423, + 450.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 455, + 506, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 456, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 506, + 468 + ], + "score": 1.0, + "content": "The update rules using the SGE and SSGE approximations follow as for the parametric case. It is im-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 416, + 479 + ], + "score": 1.0, + "content": "portant to notice that this update rule requires the function space prior gradient:", + "type": "text" + }, + { + "bbox": [ + 417, + 466, + 506, + 479 + ], + "score": 0.91, + "content": "\\nabla _ { f _ { j } } \\log { p ( f _ { j } | x , y ) } =", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 476, + 507, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 252, + 491 + ], + "score": 0.9, + "content": "\\nabla _ { f _ { j } } \\log p ( \\pmb { y } | \\pmb { x } , \\pmb { f _ { j } } ) + \\nabla _ { f _ { j } } \\log p ( \\pmb { f _ { j } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 476, + 507, + 493 + ], + "score": 1.0, + "content": ". If one wants to use an implicit prior defined in weight space,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "score": 1.0, + "content": "an additional estimator is needed due to its analytical intractability. We again adopted the SSGE,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 204, + 513 + ], + "score": 1.0, + "content": "introduced by Shi et al.", + "type": "text" + }, + { + "bbox": [ + 204, + 500, + 222, + 511 + ], + "score": 0.85, + "content": "\\mathbb { \\lVert \\boldsymbol { 6 9 } \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 499, + 464, + 513 + ], + "score": 1.0, + "content": ", which was already used for a similar purpose in Sun et al.", + "type": "text" + }, + { + "bbox": [ + 465, + 500, + 482, + 511 + ], + "score": 0.77, + "content": "\\pmb { \\mathbb { Z } 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 499, + 506, + 513 + ], + "score": 1.0, + "content": ". It is", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 303, + 524 + ], + "score": 1.0, + "content": "also interesting to note that the update rule in Eq.", + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 321, + 524 + ], + "score": 0.88, + "content": "\\textcircled { 1 2 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "readily allows for the use of alternative priors", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "that have an analytical form, such as Gaussian processes. This is an important feature of our method", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "that allows for an explicit encoding of function space properties that can be useful for example in", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 543, + 353, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 334, + 555 + ], + "score": 1.0, + "content": "achieving better out of distribution detection capabilities", + "type": "text" + }, + { + "bbox": [ + 334, + 543, + 352, + 555 + ], + "score": 0.49, + "content": "\\checkmark", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 543, + 353, + 555 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 456, + 507, + 555 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 569, + 232, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 233, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 233, + 583 + ], + "score": 1.0, + "content": "3.4 The choice of the kernel", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "A repulsive effect can always be created in the ensemble by means of the gradient of any kernel", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 615 + ], + "score": 1.0, + "content": "function that is measuring the similarity between two members. Nevertheless, it is important to keep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "in mind that to ensure the asymptotic convergence to the Bayesian posterior, the repulsive component", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 623, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 320, + 637 + ], + "score": 1.0, + "content": "must be a consistent estimator of the gradient in Eq.", + "type": "text" + }, + { + "bbox": [ + 320, + 623, + 338, + 636 + ], + "score": 0.52, + "content": "( 1 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 623, + 424, + 637 + ], + "score": 1.0, + "content": ", as shown in Section", + "type": "text" + }, + { + "bbox": [ + 424, + 623, + 432, + 636 + ], + "score": 0.68, + "content": "\\checkmark", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 623, + 506, + 637 + ], + "score": 1.0, + "content": ". Therefore, some", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 634, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 646 + ], + "score": 1.0, + "content": "important constraints over the kernel choice are needed. In particular, the SGE and SSGE need a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 338, + 657 + ], + "score": 1.0, + "content": "kernel function belonging to the Stein class (see Shi et al.", + "type": "text" + }, + { + "bbox": [ + 338, + 645, + 356, + 657 + ], + "score": 0.64, + "content": "\\dot { \\overline { { \\vert 6 9 \\vert } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "for more details). On the other hand,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "for the KDE, any symmetric probability density function can be used. On this subject, the work of", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "Aggarwal et al. [1] has shown how the Manhattan distance metric (L1 norm) might be preferable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "over the Euclidean distance metric (L2 norm) for high-dimensional settings. We performed some", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "additional experiments using the L1 norm (Laplace kernel) for the KDE but we could not observe", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "any substantial difference compared to using the L2 norm. Further investigations regarding this", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 711, + 260, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 260, + 722 + ], + "score": 1.0, + "content": "hypothesis are left for future research.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 590, + 506, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 74, + 509, + 232 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 74, + 509, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 74, + 509, + 232 + ], + "spans": [ + { + "bbox": [ + 120, + 74, + 509, + 232 + ], + "score": 0.974, + "type": "image", + "image_path": "094ed7ba5611118de5688c3ddd1545168ced55dff3ababfdac0ee5e88615b714.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 74, + 509, + 126.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 126.66666666666666, + 509, + 179.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 179.33333333333331, + 509, + 231.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 242, + 507, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "Figure 3: BNN 2D classification. We show the entropy of the predictive posteriors. Again, the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "function-space methods capture the uncertainty better than the weight-space ones, thus approaching", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 264, + 244, + 276 + ], + "spans": [ + { + "bbox": [ + 107, + 264, + 244, + 276 + ], + "score": 1.0, + "content": "the gold-standard HMC posterior.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 296, + 191, + 310 + ], + "lines": [ + { + "bbox": [ + 104, + 294, + 193, + 313 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 193, + 313 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 504, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 504, + 334 + ], + "score": 1.0, + "content": "In this section, we compare the different proposed WGD methods with deep ensembles and SVGD", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 333, + 507, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 507, + 345 + ], + "score": 1.0, + "content": "on synthetic sampling, regression, and classification tasks and real-world image classification tasks.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 343, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 475, + 356 + ], + "score": 1.0, + "content": "We use an RBF kernel (except where otherwise specified) with the popular median heuristic", + "type": "text" + }, + { + "bbox": [ + 476, + 343, + 493, + 355 + ], + "score": 0.64, + "content": "\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 343, + 504, + 356 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "choose the kernel bandwidth. In our experiments, an adaptive bandwidth leads to better performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "than fixing and tuning a single constant value for the entire evolution of the particles. We also", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 377, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 388 + ], + "score": 1.0, + "content": "quantitatively assess the uncertainty estimation of the methods in terms of calibration and OOD", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "detection. In our experiments, we report the test accuracy, negative log-likelihood (NLL), and the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "score": 1.0, + "content": "expected calibration error (ECE) [58]. To assert the robustness on out-of-distribution (OOD) data,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 409, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 408, + 421 + ], + "score": 1.0, + "content": "we report the ratio between predictive entropy on OOD and test data points", + "type": "text" + }, + { + "bbox": [ + 408, + 409, + 445, + 421 + ], + "score": 0.91, + "content": "( H _ { o } / H _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 409, + 506, + 421 + ], + "score": 1.0, + "content": ", and the OOD", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 420, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 312, + 433 + ], + "score": 1.0, + "content": "detection area under the ROC curve AUROC(H)", + "type": "text" + }, + { + "bbox": [ + 312, + 420, + 330, + 432 + ], + "score": 0.47, + "content": "\\mathbb { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 420, + 505, + 433 + ], + "score": 1.0, + "content": ". Moreover, to assess the diversity of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "ensemble generated by the different methods in function space, we measure the functional diversity", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 329, + 455 + ], + "score": 1.0, + "content": "using the model disagreement (MD) (details in Appendix", + "type": "text" + }, + { + "bbox": [ + 329, + 442, + 341, + 455 + ], + "score": 0.43, + "content": "\\bar { \\bigtriangledown } \\bar { \\bigtriangledown }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 442, + 505, + 455 + ], + "score": 1.0, + "content": ". In particular, we report the ratio between", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 365, + 466 + ], + "score": 1.0, + "content": "the average model disagreement on the OOD and test data points", + "type": "text" + }, + { + "bbox": [ + 366, + 453, + 423, + 465 + ], + "score": 0.91, + "content": "( M D _ { o } / M D _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "and additionally the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 463, + 438, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 438, + 477 + ], + "score": 1.0, + "content": "OOD detection AUROC(MD) computed using this measure instead of the entropy.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 488, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "Sampling from synthetic distributions As a sanity check, we first assessed the ability of our", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "different approximations for Wasserstein gradient descent (using KDE, SGE, and SSGE) to sample", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 321, + 522 + ], + "score": 1.0, + "content": "from a two-dimensional Gaussian distribution (Figure", + "type": "text" + }, + { + "bbox": [ + 321, + 509, + 332, + 522 + ], + "score": 0.66, + "content": "^ { 2 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 509, + 505, + 522 + ], + "score": 1.0, + "content": ". We see that our SGE-WGD, SSGE-WGD", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "score": 1.0, + "content": "and the SVGD fit the target almost perfectly. We also tested the different methods in a more complex", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 530, + 507, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 251, + 544 + ], + "score": 1.0, + "content": "two-dimensional Funnel distribution", + "type": "text" + }, + { + "bbox": [ + 251, + 531, + 269, + 543 + ], + "score": 0.82, + "content": "\\mathbb { \\left[ \\left[ 5 9 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 530, + 395, + 544 + ], + "score": 1.0, + "content": "and present the results in Figure", + "type": "text" + }, + { + "bbox": [ + 395, + 531, + 411, + 544 + ], + "score": 0.36, + "content": "\\mathrm { F . l }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 530, + 507, + 544 + ], + "score": 1.0, + "content": "in the Appendix. There,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 541, + 276, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 276, + 555 + ], + "score": 1.0, + "content": "SGE-WGD and SVGD also perform best.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "BNN 1D regression We then assessed the different methods in fitting a BNN posterior on a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 577, + 502, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 502, + 591 + ], + "score": 1.0, + "content": "synthetically generated one-dimensional regression task. The results are reported in Figure 1", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 257, + 601 + ], + "score": 1.0, + "content": "consisting of the mean prediction and", + "type": "text" + }, + { + "bbox": [ + 258, + 588, + 291, + 600 + ], + "score": 0.66, + "content": "\\pm 1 , 2 , 3", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "standard deviations of the predictive distribution. We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 612 + ], + "score": 1.0, + "content": "can see that all methods performing inference in the weight space (DE, w-SVGD, WGD) are unable", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "to capture the epistemic uncertainty between the two clusters of training data points. Conversely, the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "functional methods (f-SVGD, fWGD) are perfectly able to infer the diversity of the hypotheses in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "score": 1.0, + "content": "this region due to the lack of training evidence. They thereby achieve a predictive posterior that very", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 642, + 407, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 407, + 657 + ], + "score": 1.0, + "content": "closely resembles the one obtained with the gold-standard HMC sampling.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "BNN 2D classification Next, we investigated the predictive performance and quality of uncertainty", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "estimation of the methods in a two-dimensional synthetic classification setting. The results are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 688, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 182, + 703 + ], + "score": 1.0, + "content": "displayed in Figure", + "type": "text" + }, + { + "bbox": [ + 182, + 688, + 193, + 702 + ], + "score": 0.31, + "content": "\\textcircled { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 688, + 506, + 703 + ], + "score": 1.0, + "content": "We can clearly observe that the weight-space methods are overconfident and do", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "not capture the uncertainty well. Moreover, all the functions seems to collapse to the optimal classifier.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "These methods thus only account for uncertainty close to the decision boundaries and to the origin", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 120, + 74, + 509, + 232 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 74, + 509, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 74, + 509, + 232 + ], + "spans": [ + { + "bbox": [ + 120, + 74, + 509, + 232 + ], + "score": 0.974, + "type": "image", + "image_path": "094ed7ba5611118de5688c3ddd1545168ced55dff3ababfdac0ee5e88615b714.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 120, + 74, + 509, + 126.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 120, + 126.66666666666666, + 509, + 179.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 120, + 179.33333333333331, + 509, + 231.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 242, + 507, + 275 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 255 + ], + "score": 1.0, + "content": "Figure 3: BNN 2D classification. We show the entropy of the predictive posteriors. Again, the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 267 + ], + "score": 1.0, + "content": "function-space methods capture the uncertainty better than the weight-space ones, thus approaching", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 264, + 244, + 276 + ], + "spans": [ + { + "bbox": [ + 107, + 264, + 244, + 276 + ], + "score": 1.0, + "content": "the gold-standard HMC posterior.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 296, + 191, + 310 + ], + "lines": [ + { + "bbox": [ + 104, + 294, + 193, + 313 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 193, + 313 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 321, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 504, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 504, + 334 + ], + "score": 1.0, + "content": "In this section, we compare the different proposed WGD methods with deep ensembles and SVGD", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 333, + 507, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 507, + 345 + ], + "score": 1.0, + "content": "on synthetic sampling, regression, and classification tasks and real-world image classification tasks.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 343, + 504, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 475, + 356 + ], + "score": 1.0, + "content": "We use an RBF kernel (except where otherwise specified) with the popular median heuristic", + "type": "text" + }, + { + "bbox": [ + 476, + 343, + 493, + 355 + ], + "score": 0.64, + "content": "\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 343, + 504, + 356 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "choose the kernel bandwidth. In our experiments, an adaptive bandwidth leads to better performance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "than fixing and tuning a single constant value for the entire evolution of the particles. We also", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 377, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 388 + ], + "score": 1.0, + "content": "quantitatively assess the uncertainty estimation of the methods in terms of calibration and OOD", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "detection. In our experiments, we report the test accuracy, negative log-likelihood (NLL), and the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 506, + 411 + ], + "score": 1.0, + "content": "expected calibration error (ECE) [58]. To assert the robustness on out-of-distribution (OOD) data,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 409, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 408, + 421 + ], + "score": 1.0, + "content": "we report the ratio between predictive entropy on OOD and test data points", + "type": "text" + }, + { + "bbox": [ + 408, + 409, + 445, + 421 + ], + "score": 0.91, + "content": "( H _ { o } / H _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 409, + 506, + 421 + ], + "score": 1.0, + "content": ", and the OOD", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 420, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 312, + 433 + ], + "score": 1.0, + "content": "detection area under the ROC curve AUROC(H)", + "type": "text" + }, + { + "bbox": [ + 312, + 420, + 330, + 432 + ], + "score": 0.47, + "content": "\\mathbb { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 420, + 505, + 433 + ], + "score": 1.0, + "content": ". Moreover, to assess the diversity of the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "ensemble generated by the different methods in function space, we measure the functional diversity", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 329, + 455 + ], + "score": 1.0, + "content": "using the model disagreement (MD) (details in Appendix", + "type": "text" + }, + { + "bbox": [ + 329, + 442, + 341, + 455 + ], + "score": 0.43, + "content": "\\bar { \\bigtriangledown } \\bar { \\bigtriangledown }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 442, + 505, + 455 + ], + "score": 1.0, + "content": ". In particular, we report the ratio between", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 365, + 466 + ], + "score": 1.0, + "content": "the average model disagreement on the OOD and test data points", + "type": "text" + }, + { + "bbox": [ + 366, + 453, + 423, + 465 + ], + "score": 0.91, + "content": "( M D _ { o } / M D _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "and additionally the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 463, + 438, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 438, + 477 + ], + "score": 1.0, + "content": "OOD detection AUROC(MD) computed using this measure instead of the entropy.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 322, + 507, + 477 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 488, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 506, + 500 + ], + "score": 1.0, + "content": "Sampling from synthetic distributions As a sanity check, we first assessed the ability of our", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "different approximations for Wasserstein gradient descent (using KDE, SGE, and SSGE) to sample", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 321, + 522 + ], + "score": 1.0, + "content": "from a two-dimensional Gaussian distribution (Figure", + "type": "text" + }, + { + "bbox": [ + 321, + 509, + 332, + 522 + ], + "score": 0.66, + "content": "^ { 2 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 509, + 505, + 522 + ], + "score": 1.0, + "content": ". We see that our SGE-WGD, SSGE-WGD", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 533 + ], + "score": 1.0, + "content": "and the SVGD fit the target almost perfectly. We also tested the different methods in a more complex", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 530, + 507, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 251, + 544 + ], + "score": 1.0, + "content": "two-dimensional Funnel distribution", + "type": "text" + }, + { + "bbox": [ + 251, + 531, + 269, + 543 + ], + "score": 0.82, + "content": "\\mathbb { \\left[ \\left[ 5 9 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 530, + 395, + 544 + ], + "score": 1.0, + "content": "and present the results in Figure", + "type": "text" + }, + { + "bbox": [ + 395, + 531, + 411, + 544 + ], + "score": 0.36, + "content": "\\mathrm { F . l }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 530, + 507, + 544 + ], + "score": 1.0, + "content": "in the Appendix. There,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 541, + 276, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 276, + 555 + ], + "score": 1.0, + "content": "SGE-WGD and SVGD also perform best.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 487, + 507, + 555 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 566, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 506, + 579 + ], + "score": 1.0, + "content": "BNN 1D regression We then assessed the different methods in fitting a BNN posterior on a", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 577, + 502, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 502, + 591 + ], + "score": 1.0, + "content": "synthetically generated one-dimensional regression task. The results are reported in Figure 1", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 257, + 601 + ], + "score": 1.0, + "content": "consisting of the mean prediction and", + "type": "text" + }, + { + "bbox": [ + 258, + 588, + 291, + 600 + ], + "score": 0.66, + "content": "\\pm 1 , 2 , 3", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 589, + 505, + 601 + ], + "score": 1.0, + "content": "standard deviations of the predictive distribution. We", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 600, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 505, + 612 + ], + "score": 1.0, + "content": "can see that all methods performing inference in the weight space (DE, w-SVGD, WGD) are unable", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 623 + ], + "score": 1.0, + "content": "to capture the epistemic uncertainty between the two clusters of training data points. Conversely, the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "functional methods (f-SVGD, fWGD) are perfectly able to infer the diversity of the hypotheses in", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 645 + ], + "score": 1.0, + "content": "this region due to the lack of training evidence. They thereby achieve a predictive posterior that very", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 642, + 407, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 407, + 657 + ], + "score": 1.0, + "content": "closely resembles the one obtained with the gold-standard HMC sampling.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 566, + 506, + 657 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "BNN 2D classification Next, we investigated the predictive performance and quality of uncertainty", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "estimation of the methods in a two-dimensional synthetic classification setting. The results are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 688, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 182, + 703 + ], + "score": 1.0, + "content": "displayed in Figure", + "type": "text" + }, + { + "bbox": [ + 182, + 688, + 193, + 702 + ], + "score": 0.31, + "content": "\\textcircled { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 688, + 506, + 703 + ], + "score": 1.0, + "content": "We can clearly observe that the weight-space methods are overconfident and do", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "not capture the uncertainty well. Moreover, all the functions seems to collapse to the optimal classifier.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "These methods thus only account for uncertainty close to the decision boundaries and to the origin", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "region, for which the uncertainty is purely aleatoric. In this setting, f-SVGD suffers from similar", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "score": 1.0, + "content": "issues as the weight space methods, being overconfident away from the training data. Conversely,", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "our fWGD methods are confident (low entropy) around the data but not out-of-distribution, thus", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "representing the epistemic uncertainty better. This suggests that the functional diversity captured", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 406, + 392 + ], + "score": 1.0, + "content": "by this method naturally leads to a distance-aware uncertainty estimation", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 407, + 378, + 439, + 390 + ], + "score": 0.56, + "content": "\\mathbb { H 9 } \\mathbb { L O }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 440, + 378, + 505, + 392 + ], + "score": 1.0, + "content": ", a property that", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 389, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 104, + 389, + 506, + 402 + ], + "score": 1.0, + "content": "translates into confident predictions only in the proximity of the training data, allowing for a principled", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 400, + 172, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 172, + 412 + ], + "score": 1.0, + "content": "OOD detection.", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 666, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 143, + 505, + 312 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 69, + 505, + 136 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 69, + 505, + 83 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 288, + 83 + ], + "score": 1.0, + "content": "Table 1: BNN image classification. AUROC", + "type": "text" + }, + { + "bbox": [ + 288, + 70, + 302, + 81 + ], + "score": 0.29, + "content": "\\mathbf { ( H ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 69, + 505, + 83 + ], + "score": 1.0, + "content": "is the AUROC computed using the entropy whereas", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 346, + 95 + ], + "score": 1.0, + "content": "AUROC(MD) is computed using the model disagreement.", + "type": "text" + }, + { + "bbox": [ + 347, + 81, + 380, + 93 + ], + "score": 0.93, + "content": "\\mathbf { H _ { o } / H _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "is the ratio of the entropies on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 92, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 257, + 104 + ], + "score": 1.0, + "content": "OOD and test points respectively and", + "type": "text" + }, + { + "bbox": [ + 257, + 92, + 312, + 104 + ], + "score": 0.9, + "content": "\\mathbf { M D _ { o } / M D _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "is the ratio for model disagreement. We see that", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "score": 1.0, + "content": "the best accuracy is achieved by our WGD methods, while our fWGD methods yield the best OOD", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "score": 1.0, + "content": "detection and funtional diversity. All our proposed methods improve over standard deep ensembles in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 395, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 395, + 139 + ], + "score": 1.0, + "content": "terms of accuracy and diversity, highlighting the effect of our repulsion.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 143, + 505, + 312 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 143, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 505, + 312 + ], + "score": 0.985, + "html": "
PrroiiiieAUROC(H)AUROC(MD)AccuracyH/HtMD/MDtECENLL
Deep ensemble 40.958±0.0010.975±0.00191.122±0.0136.257±0.0056.394±0.0010.012±0.0010.129±0.001
SVGD [510.960±0.0010.973±0.00191.134±0.0246.315±0.0196.395±0.0180.014±0.0010.127±0.001
f-SVGD [760.956±0.0010.975±0.00189.884±0.0155.652±0.0096.531±0.0050.013±0.0010.150±0.001
hyper-DE [80]0.968±0.0010.981±0.00191.160±0.0076.682±0.0657.059±0.1520.014±0.0010.128±0.001
kde-WGD (ours)0.960±0.0010.970±0.00191.238±0.0196.587±0.0196.379±0.0180.014±0.0010.128±0.001
sge-WGD (ours)0.960±0.0010.970±0.00191.312±0.0166.562±0.0076.363±0.0090.012±0.0010.128±0.001
ssge-WGD (ours)0.968±0.0010.979±0.00191.198±0.0246.522±0.0096.610±0.0120.012±0.0010.130±0.001
kde-fWGD (ours)0.971±0.0010.980±0.00191.260±0.0117.079±0.0166.887±0.0150.015±0.0010.125±0.001
sge-fWGD (ours)0.969±0.0010.978±0.00191.192±0.0137.076±0.0046.900±0.0050.015±0.0010.125±0.001
ssge-fWGD (ours)0.971±0.0010.980±0.00191.240±0.0227.129±0.0066.951±0.0050.016±0.0010.124±0.001
CITIIIODeep ensemble 420.843±0.0040.736±0.00585,552±0.0762.244±0.0061.667±0.0080.049±0.0010.277±0.001
SVGD [510.825±0.0010.710±0.00285.142±0.0172.106±0.0031.567±0.0040.052±0.0010.287±0.001
fSVGD 回0.783±0.0010.712±0.00184.510±0.0311.968±0.0041.624±0.0030.049±0.0010.292±0.001
hyper-DE [80]0.789±0.0010.743±0.00184.743±0.0111.951±0.0101.690±0.0150.046±0.0010.288±0.001
kde-WGD (ours)0.838±0.0010.735±0.00485.904±0.0302.205±0.0031.661±0.0080.053±0.0010.276±0.001
sge-WGD (ours)0.837±0.0030.725±0.00485.792±0.0352.214±0.0101.634±0.0040.051±0.0010.275±0.001
ssge-WGD (ours)0.832±0.0030.731±0.00585.638±0.0382.182±0.0151.655±0.0010.049±0.0010.276±0.001
kde-fWGD (ours)0.791±0.0020.758±0.00284.888±0.0301.970±0.0041.749±0.0050.044±0.0010.282±0.001
sge-fWGD (ours)0.795±0.0010.754±0.00284.766±0.0601.984±0.0031.729±0.0020.047±0.0010.288±0.001
ssge-fWGD (ours)0.792±0.0020.752±0.00284.762±0.0341.970±0.0061.723±0.0050.046±0.0010.286±0.001
", + "type": "table", + "image_path": "c10e7e6db9d6afe7b283337a217b0daf6bdc8ff99b073aef9c58623cfb518d7c.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 107, + 143, + 505, + 199.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 199.33333333333334, + 505, + 255.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 107, + 255.66666666666669, + 505, + 312.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "region, for which the uncertainty is purely aleatoric. In this setting, f-SVGD suffers from similar", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 359 + ], + "score": 1.0, + "content": "issues as the weight space methods, being overconfident away from the training data. Conversely,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "our fWGD methods are confident (low entropy) around the data but not out-of-distribution, thus", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "representing the epistemic uncertainty better. This suggests that the functional diversity captured", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 378, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 406, + 392 + ], + "score": 1.0, + "content": "by this method naturally leads to a distance-aware uncertainty estimation", + "type": "text" + }, + { + "bbox": [ + 407, + 378, + 439, + 390 + ], + "score": 0.56, + "content": "\\mathbb { H 9 } \\mathbb { L O }", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 378, + 505, + 392 + ], + "score": 1.0, + "content": ", a property that", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 389, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 104, + 389, + 506, + 402 + ], + "score": 1.0, + "content": "translates into confident predictions only in the proximity of the training data, allowing for a principled", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 400, + 172, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 172, + 412 + ], + "score": 1.0, + "content": "OOD detection.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "FashionMNIST classification Moving on to real-world data, we used an image classification", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 271, + 448 + ], + "score": 1.0, + "content": "setting using the FashionMNIST dataset", + "type": "text" + }, + { + "bbox": [ + 271, + 435, + 289, + 447 + ], + "score": 0.78, + "content": "\\dot { \\left[ \\left| 8 3 \\right| \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 435, + 434, + 448 + ], + "score": 1.0, + "content": "for training and the MNIST dataset", + "type": "text" + }, + { + "bbox": [ + 434, + 435, + 452, + 447 + ], + "score": 0.78, + "content": "\\bar { \\mathbb { E } 3 } \\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "as an out-of-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 340, + 459 + ], + "score": 1.0, + "content": "distribution (OOD) task. The results are reported in Table", + "type": "text" + }, + { + "bbox": [ + 340, + 447, + 350, + 460 + ], + "score": 0.55, + "content": "^ 1 { } _ { . }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "(top). We can see that all our methods", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "improve upon standard deep ensembles and SVGD, highlighting the effectiveness of our proposed", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "repulsion terms when training neural network ensembles. In particular, the sge-WGD offers the best", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "accuracy, whereas the methods in function space all offer a better OOD detection. This is probably", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "due to the fact that these methods achieve a higher entropy ratio and functional diversity measured", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "via the model disagreement when compared to their weight-space counterparts. Interestingly, they", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "also reach the lowest NLL values. We can also notice how the model disagreement (MD) not only", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "serves its purpose as a metric for the functional heterogeneity of the ensemble but also allows for a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "better OOD detection in comparison to the entropy. To the best of our knowledge, this insight has", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 543, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 104, + 543, + 403, + 559 + ], + "score": 1.0, + "content": "not been described before, although it has been used in continual learning", + "type": "text" + }, + { + "bbox": [ + 404, + 545, + 421, + 556 + ], + "score": 0.27, + "content": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 543, + 505, + 559 + ], + "score": 1.0, + "content": ". Interestingly, using", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 257, + 569 + ], + "score": 1.0, + "content": "this metric, the hyper-deep ensemble", + "type": "text" + }, + { + "bbox": [ + 257, + 555, + 275, + 567 + ], + "score": 0.44, + "content": "[ \\textcircled { 8 0 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "shows OOD detection performance comparable with our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 568, + 258, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 258, + 579 + ], + "score": 1.0, + "content": "repulsive ensemble in function space.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 371, + 604 + ], + "score": 1.0, + "content": "CIFAR classification Finally, we use a ResNet32 architecture", + "type": "text" + }, + { + "bbox": [ + 371, + 591, + 389, + 603 + ], + "score": 0.4, + "content": "\\mathbb { \\left| \\left[ 2 9 \\right] \\right| }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "on CIFAR-10 [41] with the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 165, + 615 + ], + "score": 1.0, + "content": "SVHN dataset", + "type": "text" + }, + { + "bbox": [ + 165, + 602, + 183, + 614 + ], + "score": 0.46, + "content": "\\mathbb { \\left[ 6 1 \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "as OOD data. The results are reported in Table 1 (bottom). We can see that in this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 614, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 625 + ], + "score": 1.0, + "content": "case, the weight-space methods achieve better performance in accuracy and OOD detection using the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "entropy than the ones in function space. Nevertheless, all our repulsive ensembles improve functional", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "score": 1.0, + "content": "diversity, accuracy, and OOD detection when compared to standard SVGD, whereas the standard", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 645, + 374, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 374, + 659 + ], + "score": 1.0, + "content": "deep ensemble achieves the best OOD detection using the entropy.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 107, + 673, + 197, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 673, + 198, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 198, + 689 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "The theoretical and empirical properties of SVGD have been well studied [40, 48, 13] and it can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "also be seen as a Wasserstein gradient flow of the KL divergence in the Stein geometry [16, 50]", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 143, + 505, + 312 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 69, + 505, + 136 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 69, + 505, + 83 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 288, + 83 + ], + "score": 1.0, + "content": "Table 1: BNN image classification. AUROC", + "type": "text" + }, + { + "bbox": [ + 288, + 70, + 302, + 81 + ], + "score": 0.29, + "content": "\\mathbf { ( H ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 69, + 505, + 83 + ], + "score": 1.0, + "content": "is the AUROC computed using the entropy whereas", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 346, + 95 + ], + "score": 1.0, + "content": "AUROC(MD) is computed using the model disagreement.", + "type": "text" + }, + { + "bbox": [ + 347, + 81, + 380, + 93 + ], + "score": 0.93, + "content": "\\mathbf { H _ { o } / H _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "is the ratio of the entropies on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 92, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 257, + 104 + ], + "score": 1.0, + "content": "OOD and test points respectively and", + "type": "text" + }, + { + "bbox": [ + 257, + 92, + 312, + 104 + ], + "score": 0.9, + "content": "\\mathbf { M D _ { o } / M D _ { t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "is the ratio for model disagreement. We see that", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 116 + ], + "score": 1.0, + "content": "the best accuracy is achieved by our WGD methods, while our fWGD methods yield the best OOD", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 506, + 128 + ], + "score": 1.0, + "content": "detection and funtional diversity. All our proposed methods improve over standard deep ensembles in", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 124, + 395, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 395, + 139 + ], + "score": 1.0, + "content": "terms of accuracy and diversity, highlighting the effect of our repulsion.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 143, + 505, + 312 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 143, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 505, + 312 + ], + "score": 0.985, + "html": "
PrroiiiieAUROC(H)AUROC(MD)AccuracyH/HtMD/MDtECENLL
Deep ensemble 40.958±0.0010.975±0.00191.122±0.0136.257±0.0056.394±0.0010.012±0.0010.129±0.001
SVGD [510.960±0.0010.973±0.00191.134±0.0246.315±0.0196.395±0.0180.014±0.0010.127±0.001
f-SVGD [760.956±0.0010.975±0.00189.884±0.0155.652±0.0096.531±0.0050.013±0.0010.150±0.001
hyper-DE [80]0.968±0.0010.981±0.00191.160±0.0076.682±0.0657.059±0.1520.014±0.0010.128±0.001
kde-WGD (ours)0.960±0.0010.970±0.00191.238±0.0196.587±0.0196.379±0.0180.014±0.0010.128±0.001
sge-WGD (ours)0.960±0.0010.970±0.00191.312±0.0166.562±0.0076.363±0.0090.012±0.0010.128±0.001
ssge-WGD (ours)0.968±0.0010.979±0.00191.198±0.0246.522±0.0096.610±0.0120.012±0.0010.130±0.001
kde-fWGD (ours)0.971±0.0010.980±0.00191.260±0.0117.079±0.0166.887±0.0150.015±0.0010.125±0.001
sge-fWGD (ours)0.969±0.0010.978±0.00191.192±0.0137.076±0.0046.900±0.0050.015±0.0010.125±0.001
ssge-fWGD (ours)0.971±0.0010.980±0.00191.240±0.0227.129±0.0066.951±0.0050.016±0.0010.124±0.001
CITIIIODeep ensemble 420.843±0.0040.736±0.00585,552±0.0762.244±0.0061.667±0.0080.049±0.0010.277±0.001
SVGD [510.825±0.0010.710±0.00285.142±0.0172.106±0.0031.567±0.0040.052±0.0010.287±0.001
fSVGD 回0.783±0.0010.712±0.00184.510±0.0311.968±0.0041.624±0.0030.049±0.0010.292±0.001
hyper-DE [80]0.789±0.0010.743±0.00184.743±0.0111.951±0.0101.690±0.0150.046±0.0010.288±0.001
kde-WGD (ours)0.838±0.0010.735±0.00485.904±0.0302.205±0.0031.661±0.0080.053±0.0010.276±0.001
sge-WGD (ours)0.837±0.0030.725±0.00485.792±0.0352.214±0.0101.634±0.0040.051±0.0010.275±0.001
ssge-WGD (ours)0.832±0.0030.731±0.00585.638±0.0382.182±0.0151.655±0.0010.049±0.0010.276±0.001
kde-fWGD (ours)0.791±0.0020.758±0.00284.888±0.0301.970±0.0041.749±0.0050.044±0.0010.282±0.001
sge-fWGD (ours)0.795±0.0010.754±0.00284.766±0.0601.984±0.0031.729±0.0020.047±0.0010.288±0.001
ssge-fWGD (ours)0.792±0.0020.752±0.00284.762±0.0341.970±0.0061.723±0.0050.046±0.0010.286±0.001
", + "type": "table", + "image_path": "c10e7e6db9d6afe7b283337a217b0daf6bdc8ff99b073aef9c58623cfb518d7c.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 107, + 143, + 505, + 199.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 199.33333333333334, + 505, + 255.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 107, + 255.66666666666669, + 505, + 312.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 505, + 411 + ], + "lines": [], + "index": 12, + "bbox_fs": [ + 104, + 334, + 506, + 412 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "FashionMNIST classification Moving on to real-world data, we used an image classification", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 271, + 448 + ], + "score": 1.0, + "content": "setting using the FashionMNIST dataset", + "type": "text" + }, + { + "bbox": [ + 271, + 435, + 289, + 447 + ], + "score": 0.78, + "content": "\\dot { \\left[ \\left| 8 3 \\right| \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 435, + 434, + 448 + ], + "score": 1.0, + "content": "for training and the MNIST dataset", + "type": "text" + }, + { + "bbox": [ + 434, + 435, + 452, + 447 + ], + "score": 0.78, + "content": "\\bar { \\mathbb { E } 3 } \\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "as an out-of-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 340, + 459 + ], + "score": 1.0, + "content": "distribution (OOD) task. The results are reported in Table", + "type": "text" + }, + { + "bbox": [ + 340, + 447, + 350, + 460 + ], + "score": 0.55, + "content": "^ 1 { } _ { . }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "(top). We can see that all our methods", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "improve upon standard deep ensembles and SVGD, highlighting the effectiveness of our proposed", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "repulsion terms when training neural network ensembles. In particular, the sge-WGD offers the best", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "accuracy, whereas the methods in function space all offer a better OOD detection. This is probably", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "due to the fact that these methods achieve a higher entropy ratio and functional diversity measured", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 502, + 505, + 514 + ], + "score": 1.0, + "content": "via the model disagreement when compared to their weight-space counterparts. Interestingly, they", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 525 + ], + "score": 1.0, + "content": "also reach the lowest NLL values. We can also notice how the model disagreement (MD) not only", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "serves its purpose as a metric for the functional heterogeneity of the ensemble but also allows for a", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "better OOD detection in comparison to the entropy. To the best of our knowledge, this insight has", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 543, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 104, + 543, + 403, + 559 + ], + "score": 1.0, + "content": "not been described before, although it has been used in continual learning", + "type": "text" + }, + { + "bbox": [ + 404, + 545, + 421, + 556 + ], + "score": 0.27, + "content": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 543, + 505, + 559 + ], + "score": 1.0, + "content": ". Interestingly, using", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 257, + 569 + ], + "score": 1.0, + "content": "this metric, the hyper-deep ensemble", + "type": "text" + }, + { + "bbox": [ + 257, + 555, + 275, + 567 + ], + "score": 0.44, + "content": "[ \\textcircled { 8 0 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "shows OOD detection performance comparable with our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 568, + 258, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 258, + 579 + ], + "score": 1.0, + "content": "repulsive ensemble in function space.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 424, + 506, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 371, + 604 + ], + "score": 1.0, + "content": "CIFAR classification Finally, we use a ResNet32 architecture", + "type": "text" + }, + { + "bbox": [ + 371, + 591, + 389, + 603 + ], + "score": 0.4, + "content": "\\mathbb { \\left| \\left[ 2 9 \\right] \\right| }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "on CIFAR-10 [41] with the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 165, + 615 + ], + "score": 1.0, + "content": "SVHN dataset", + "type": "text" + }, + { + "bbox": [ + 165, + 602, + 183, + 614 + ], + "score": 0.46, + "content": "\\mathbb { \\left[ 6 1 \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "as OOD data. The results are reported in Table 1 (bottom). We can see that in this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 614, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 625 + ], + "score": 1.0, + "content": "case, the weight-space methods achieve better performance in accuracy and OOD detection using the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "entropy than the ones in function space. Nevertheless, all our repulsive ensembles improve functional", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 506, + 648 + ], + "score": 1.0, + "content": "diversity, accuracy, and OOD detection when compared to standard SVGD, whereas the standard", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 645, + 374, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 374, + 659 + ], + "score": 1.0, + "content": "deep ensemble achieves the best OOD detection using the entropy.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 591, + 506, + 659 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 673, + 197, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 673, + 198, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 198, + 689 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 503, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "The theoretical and empirical properties of SVGD have been well studied [40, 48, 13] and it can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "also be seen as a Wasserstein gradient flow of the KL divergence in the Stein geometry [16, 50]", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "(see Appendix D for more details). Interestingly, a gradient flow interpretation is also possible for", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 294, + 96 + ], + "score": 1.0, + "content": "(stochastic gradient) MCMC-type algorithms", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 294, + 83, + 312, + 95 + ], + "score": 0.35, + "content": "\\dot { \\left. \\overline { { 4 8 } } \\right. }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 312, + 83, + 505, + 96 + ], + "score": 1.0, + "content": ", which can be unified under a general particle", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "inference framework [10]. Moreover, our Wasserstein gradient descent using the SGE approximation", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "can also be derived using an alternative formulation as a gradient flow with smoothed test functions", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 372, + 128 + ], + "score": 1.0, + "content": "[48]. A projected version of WGD has been studied in Wang et al.", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 372, + 116, + 390, + 128 + ], + "score": 0.71, + "content": "\\mathbb { \\left. \\overline { { \\boldsymbol { \\mathscr { Q } } \\boldsymbol { \\cdot } \\boldsymbol { \\cdot } } } \\right. }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 390, + 116, + 505, + 128 + ], + "score": 1.0, + "content": ", which could also be readily", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 421, + 140 + ], + "score": 1.0, + "content": "applied in our framework. Besides particle methods, Bayesian neural networks", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 421, + 127, + 454, + 138 + ], + "score": 0.35, + "content": "\\pm \\pm \\infty", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 455, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "have gained", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "popularity recently [79, 22, 19, 36], using modern MCMC [59, 79, 22, 24, 21] and variational", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "inference techniques [5, 72, 17, 34]. On the other hand, ensemble methods have also been extensively", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 504, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 504, + 172 + ], + "score": 1.0, + "content": "studied [42, 18, 82, 23, 80, 32, 85, 78].Moreover, repulsive interactions between the members have", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "score": 1.0, + "content": "also been studied in Wabartha et al. [74]. Moreover, providing Bayesian interpretations for deep", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 475, + 194 + ], + "score": 1.0, + "content": "ensembles has been previously attempted through the lenses of stationary SGD distributions", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 475, + 181, + 502, + 193 + ], + "score": 0.47, + "content": "[ \\sqrt { 5 6 } , \\sqrt ] { 8 } ]", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 503, + 181, + 505, + 194 + ], + "score": 1.0, + "content": ",", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 220, + 206 + ], + "score": 1.0, + "content": "ensembles of linear models", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 221, + 192, + 238, + 204 + ], + "score": 0.39, + "content": "\\dot { \\bigtriangledown } 5 7 \\big |", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 239, + 191, + 359, + 206 + ], + "score": 1.0, + "content": ", additional random functions", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 360, + 192, + 407, + 204 + ], + "score": 0.64, + "content": "\\boxed { 6 2 } , \\boxed { 1 2 } , \\boxed { 2 7 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 408, + 191, + 506, + 206 + ], + "score": 1.0, + "content": ", approximate inference", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 107, + 203, + 124, + 215 + ], + "score": 0.48, + "content": "\\lVert 8 2 \\rVert", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 125, + 203, + 420, + 216 + ], + "score": 1.0, + "content": ", Stein variational inference [15], and marginal likelihood lower bounds", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 420, + 203, + 438, + 215 + ], + "score": 0.8, + "content": "\\bar { \\| 5 3 \\| }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 438, + 203, + 505, + 216 + ], + "score": 1.0, + "content": ", and ensembles", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "have also been shown to provide good approximations to the true BNN posterior in some settings", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 225, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 124, + 237 + ], + "score": 0.31, + "content": "\\left[ \\left[ 3 6 \\right] \\right]", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 124, + 225, + 454, + 237 + ], + "score": 1.0, + "content": ". Furthermore, variational inference in function space has recently gained attention", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 455, + 225, + 473, + 237 + ], + "score": 0.83, + "content": "\\mathbb { \\ m }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 473, + 225, + 506, + 237 + ], + "score": 1.0, + "content": "and the", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 236, + 378, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 378, + 248 + ], + "score": 1.0, + "content": "limitations of the KL divergence have been studied in Burt et al. [7].", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 698, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "(see Appendix D for more details). Interestingly, a gradient flow interpretation is also possible for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 294, + 96 + ], + "score": 1.0, + "content": "(stochastic gradient) MCMC-type algorithms", + "type": "text" + }, + { + "bbox": [ + 294, + 83, + 312, + 95 + ], + "score": 0.35, + "content": "\\dot { \\left. \\overline { { 4 8 } } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 83, + 505, + 96 + ], + "score": 1.0, + "content": ", which can be unified under a general particle", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 107 + ], + "score": 1.0, + "content": "inference framework [10]. Moreover, our Wasserstein gradient descent using the SGE approximation", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "can also be derived using an alternative formulation as a gradient flow with smoothed test functions", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 372, + 128 + ], + "score": 1.0, + "content": "[48]. A projected version of WGD has been studied in Wang et al.", + "type": "text" + }, + { + "bbox": [ + 372, + 116, + 390, + 128 + ], + "score": 0.71, + "content": "\\mathbb { \\left. \\overline { { \\boldsymbol { \\mathscr { Q } } \\boldsymbol { \\cdot } \\boldsymbol { \\cdot } } } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 116, + 505, + 128 + ], + "score": 1.0, + "content": ", which could also be readily", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 421, + 140 + ], + "score": 1.0, + "content": "applied in our framework. Besides particle methods, Bayesian neural networks", + "type": "text" + }, + { + "bbox": [ + 421, + 127, + 454, + 138 + ], + "score": 0.35, + "content": "\\pm \\pm \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "have gained", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "popularity recently [79, 22, 19, 36], using modern MCMC [59, 79, 22, 24, 21] and variational", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "inference techniques [5, 72, 17, 34]. On the other hand, ensemble methods have also been extensively", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 504, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 504, + 172 + ], + "score": 1.0, + "content": "studied [42, 18, 82, 23, 80, 32, 85, 78].Moreover, repulsive interactions between the members have", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 183 + ], + "score": 1.0, + "content": "also been studied in Wabartha et al. [74]. Moreover, providing Bayesian interpretations for deep", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 475, + 194 + ], + "score": 1.0, + "content": "ensembles has been previously attempted through the lenses of stationary SGD distributions", + "type": "text" + }, + { + "bbox": [ + 475, + 181, + 502, + 193 + ], + "score": 0.47, + "content": "[ \\sqrt { 5 6 } , \\sqrt ] { 8 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 181, + 505, + 194 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 220, + 206 + ], + "score": 1.0, + "content": "ensembles of linear models", + "type": "text" + }, + { + "bbox": [ + 221, + 192, + 238, + 204 + ], + "score": 0.39, + "content": "\\dot { \\bigtriangledown } 5 7 \\big |", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 191, + 359, + 206 + ], + "score": 1.0, + "content": ", additional random functions", + "type": "text" + }, + { + "bbox": [ + 360, + 192, + 407, + 204 + ], + "score": 0.64, + "content": "\\boxed { 6 2 } , \\boxed { 1 2 } , \\boxed { 2 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 191, + 506, + 206 + ], + "score": 1.0, + "content": ", approximate inference", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 107, + 203, + 124, + 215 + ], + "score": 0.48, + "content": "\\lVert 8 2 \\rVert", + "type": "inline_equation" + }, + { + "bbox": [ + 125, + 203, + 420, + 216 + ], + "score": 1.0, + "content": ", Stein variational inference [15], and marginal likelihood lower bounds", + "type": "text" + }, + { + "bbox": [ + 420, + 203, + 438, + 215 + ], + "score": 0.8, + "content": "\\bar { \\| 5 3 \\| }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 203, + 505, + 216 + ], + "score": 1.0, + "content": ", and ensembles", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "have also been shown to provide good approximations to the true BNN posterior in some settings", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 225, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 124, + 237 + ], + "score": 0.31, + "content": "\\left[ \\left[ 3 6 \\right] \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 225, + 454, + 237 + ], + "score": 1.0, + "content": ". Furthermore, variational inference in function space has recently gained attention", + "type": "text" + }, + { + "bbox": [ + 455, + 225, + 473, + 237 + ], + "score": 0.83, + "content": "\\mathbb { \\ m }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 225, + 506, + 237 + ], + "score": 1.0, + "content": "and the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 236, + 378, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 378, + 248 + ], + "score": 1.0, + "content": "limitations of the KL divergence have been studied in Burt et al. [7].", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 107, + 262, + 183, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 185, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 185, + 278 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "We have presented a simple and principled way to improve upon standard deep ensemble methods.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 298, + 504, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 504, + 308 + ], + "score": 1.0, + "content": "To this end, we have shown that the introduction of a kernelized repulsion between members of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "score": 1.0, + "content": "ensemble not only improves the accuracy of the predictions but—even more importantly—can be seen", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "as Wasserstein gradient descent on the KL divergence, thus transforming the MAP inference of deep", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "ensembles into proper Bayesian inference. Moreover, we have shown that incorporating functional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "repulsion between ensemble members can improve the quality of the estimated uncertainties on", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "simple synthetic examples and OOD detection on real-world data and can approach the true Bayesian", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 200, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 200, + 376 + ], + "score": 1.0, + "content": "posterior more closely.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "score": 1.0, + "content": "In future work, it will be interesting to study the impact of the Jacobian in the fWGD update and its", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "implications on the Liouville equation in more detail, also compared to other neural network Jacobian", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 270, + 414 + ], + "score": 1.0, + "content": "methods, such as neural tangent kernels", + "type": "text" + }, + { + "bbox": [ + 271, + 400, + 289, + 412 + ], + "score": 0.59, + "content": "\\pmb { \\Vert 3 7 } \\Vert", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 400, + 485, + 414 + ], + "score": 1.0, + "content": "and generalized Gauss-Newton approximations", + "type": "text" + }, + { + "bbox": [ + 485, + 400, + 503, + 412 + ], + "score": 0.26, + "content": "\\boldsymbol { \\left[ \\left[ 3 5 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 400, + 505, + 414 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "Moreover, it would be interesting to derive explicit convergence bounds for our proposed method and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 423, + 318, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 318, + 434 + ], + "score": 1.0, + "content": "compare them to the existing bounds for SVGD [40].", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 446, + 186, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 187, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 187, + 460 + ], + "score": 1.0, + "content": "Acknowledgments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 466, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "VF would like to acknowledge financial support from the Strategic Focus Area “Personalized Health", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 477, + 504, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 504, + 489 + ], + "score": 1.0, + "content": "and Related Technologies” of the ETH Domain through the grant #2017-110 and from the Swiss", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Data Science Center through a PhD fellowship. We thank Florian Wenzel, Alexander Immer, Andrew", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 499, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 506, + 511 + ], + "score": 1.0, + "content": "Gordon Wilson, Pavel Izmailov, Christian Henning, and Johannes von Oswald for helpful discussions.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 509, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 523 + ], + "score": 1.0, + "content": "We also thank Dr. Sheldon Cooper, Dr. Leonard Hofstadter and Penny Hofstadter for their support", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 521, + 170, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 170, + 533 + ], + "score": 1.0, + "content": "and inspiration.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 107, + 547, + 163, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 165, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 165, + 562 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 110, + 565, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 110, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 127, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "distance metrics in high dimensional space. In International conference on database theory,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 588, + 257, + 600 + ], + "spans": [ + { + "bbox": [ + 127, + 588, + 257, + 600 + ], + "score": 1.0, + "content": "pages 420–434. Springer, 2001.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 110, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 110, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "[2] Luigi Ambrosio and Gianluca Crippa. Continuity equations and ode flows with non-smooth", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 127, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "velocity. Proceedings of the Royal Society of Edinburgh: Section A Mathematics, 144(6):", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 627, + 201, + 639 + ], + "spans": [ + { + "bbox": [ + 127, + 627, + 201, + 639 + ], + "score": 1.0, + "content": "1191–1244, 2014.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 111, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 111, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "[3] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré. Gradient flows: in metric spaces and in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 655, + 441, + 667 + ], + "spans": [ + { + "bbox": [ + 127, + 655, + 441, + 667 + ], + "score": 1.0, + "content": "the space of probability measures. Springer Science & Business Media, 2008.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 111, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 111, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "[4] Vijay Badrinarayanan, Bamdev Mishra, and Roberto Cipolla. Symmetry-invariant optimization", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 682, + 364, + 695 + ], + "spans": [ + { + "bbox": [ + 127, + 682, + 364, + 695 + ], + "score": 1.0, + "content": "in deep networks. arXiv preprint arXiv:1511.01754, 2015.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 110, + 699, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 110, + 699, + 505, + 714 + ], + "score": 1.0, + "content": "[5] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 126, + 711, + 371, + 723 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 371, + 723 + ], + "score": 1.0, + "content": "in neural networks. arXiv preprint arXiv:1505.05424, 2015.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 505, + 248 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 72, + 506, + 248 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 262, + 183, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 261, + 185, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 185, + 278 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 286, + 505, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "We have presented a simple and principled way to improve upon standard deep ensemble methods.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 298, + 504, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 504, + 308 + ], + "score": 1.0, + "content": "To this end, we have shown that the introduction of a kernelized repulsion between members of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "score": 1.0, + "content": "ensemble not only improves the accuracy of the predictions but—even more importantly—can be seen", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 332 + ], + "score": 1.0, + "content": "as Wasserstein gradient descent on the KL divergence, thus transforming the MAP inference of deep", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "ensembles into proper Bayesian inference. Moreover, we have shown that incorporating functional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "repulsion between ensemble members can improve the quality of the estimated uncertainties on", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "simple synthetic examples and OOD detection on real-world data and can approach the true Bayesian", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 200, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 200, + 376 + ], + "score": 1.0, + "content": "posterior more closely.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 285, + 506, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 379, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 391 + ], + "score": 1.0, + "content": "In future work, it will be interesting to study the impact of the Jacobian in the fWGD update and its", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "implications on the Liouville equation in more detail, also compared to other neural network Jacobian", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 400, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 270, + 414 + ], + "score": 1.0, + "content": "methods, such as neural tangent kernels", + "type": "text" + }, + { + "bbox": [ + 271, + 400, + 289, + 412 + ], + "score": 0.59, + "content": "\\pmb { \\Vert 3 7 } \\Vert", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 400, + 485, + 414 + ], + "score": 1.0, + "content": "and generalized Gauss-Newton approximations", + "type": "text" + }, + { + "bbox": [ + 485, + 400, + 503, + 412 + ], + "score": 0.26, + "content": "\\boldsymbol { \\left[ \\left[ 3 5 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 400, + 505, + 414 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 506, + 425 + ], + "score": 1.0, + "content": "Moreover, it would be interesting to derive explicit convergence bounds for our proposed method and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 423, + 318, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 318, + 434 + ], + "score": 1.0, + "content": "compare them to the existing bounds for SVGD [40].", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 379, + 506, + 434 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 446, + 186, + 458 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 187, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 187, + 460 + ], + "score": 1.0, + "content": "Acknowledgments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 466, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 479 + ], + "score": 1.0, + "content": "VF would like to acknowledge financial support from the Strategic Focus Area “Personalized Health", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 477, + 504, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 504, + 489 + ], + "score": 1.0, + "content": "and Related Technologies” of the ETH Domain through the grant #2017-110 and from the Swiss", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Data Science Center through a PhD fellowship. We thank Florian Wenzel, Alexander Immer, Andrew", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 499, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 506, + 511 + ], + "score": 1.0, + "content": "Gordon Wilson, Pavel Izmailov, Christian Henning, and Johannes von Oswald for helpful discussions.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 509, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 523 + ], + "score": 1.0, + "content": "We also thank Dr. Sheldon Cooper, Dr. Leonard Hofstadter and Penny Hofstadter for their support", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 521, + 170, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 170, + 533 + ], + "score": 1.0, + "content": "and inspiration.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 466, + 506, + 533 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 547, + 163, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 165, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 165, + 562 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "list", + "bbox": [ + 110, + 565, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 110, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 110, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "[1] Charu C Aggarwal, Alexander Hinneburg, and Daniel A Keim. On the surprising behavior of", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 576, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 127, + 576, + 506, + 589 + ], + "score": 1.0, + "content": "distance metrics in high dimensional space. In International conference on database theory,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 588, + 257, + 600 + ], + "spans": [ + { + "bbox": [ + 127, + 588, + 257, + 600 + ], + "score": 1.0, + "content": "pages 420–434. Springer, 2001.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 110, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "[2] Luigi Ambrosio and Gianluca Crippa. Continuity equations and ode flows with non-smooth", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 127, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "velocity. Proceedings of the Royal Society of Edinburgh: Section A Mathematics, 144(6):", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 627, + 201, + 639 + ], + "spans": [ + { + "bbox": [ + 127, + 627, + 201, + 639 + ], + "score": 1.0, + "content": "1191–1244, 2014.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 111, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "[3] Luigi Ambrosio, Nicola Gigli, and Giuseppe Savaré. Gradient flows: in metric spaces and in", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 655, + 441, + 667 + ], + "spans": [ + { + "bbox": [ + 127, + 655, + 441, + 667 + ], + "score": 1.0, + "content": "the space of probability measures. Springer Science & Business Media, 2008.", + "type": "text" + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 111, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "[4] Vijay Badrinarayanan, Bamdev Mishra, and Roberto Cipolla. Symmetry-invariant optimization", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 682, + 364, + 695 + ], + "spans": [ + { + "bbox": [ + 127, + 682, + 364, + 695 + ], + "score": 1.0, + "content": "in deep networks. arXiv preprint arXiv:1511.01754, 2015.", + "type": "text" + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 699, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 110, + 699, + 505, + 714 + ], + "score": 1.0, + "content": "[5] Charles Blundell, Julien Cornebise, Koray Kavukcuoglu, and Daan Wierstra. Weight uncertainty", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 711, + 371, + 723 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 371, + 723 + ], + "score": 1.0, + "content": "in neural networks. arXiv preprint arXiv:1505.05424, 2015.", + "type": "text" + } + ], + "index": 49, + "is_list_end_line": true + } + ], + "index": 43.5, + "bbox_fs": [ + 110, + 565, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 110, + 72, + 433, + 85 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 433, + 86 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 433, + 86 + ], + "score": 1.0, + "content": "[6] Leo Breiman. Bagging predictors. Machine learning, 24(2):123–140, 1996.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 109, + 91, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 110, + 91, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 110, + 91, + 506, + 106 + ], + "score": 1.0, + "content": "[7] David R Burt, Sebastian W Ober, Adrià Garriga-Alonso, and Mark van der Wilk. Understanding", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 102, + 449, + 115 + ], + "spans": [ + { + "bbox": [ + 127, + 102, + 449, + 115 + ], + "score": 1.0, + "content": "variational inference in function-space. arXiv preprint arXiv:2011.09421, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 109, + 121, + 504, + 155 + ], + "lines": [ + { + "bbox": [ + 110, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 110, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "[8] Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "converges to limit cycles for deep networks. In 2018 Information Theory and Applications", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 144, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 144, + 297, + 156 + ], + "score": 1.0, + "content": "Workshop (ITA), pages 1–10. IEEE, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 109, + 162, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 111, + 162, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 111, + 162, + 505, + 175 + ], + "score": 1.0, + "content": "[9] An Mei Chen, Haw-minn Lu, and Robert Hecht-Nielsen. On the geometry of feedforward", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 173, + 420, + 186 + ], + "spans": [ + { + "bbox": [ + 127, + 173, + 420, + 186 + ], + "score": 1.0, + "content": "neural network error surfaces. Neural computation, 5(6):910–927, 1993.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "[10] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liqun Chen. A unified particle-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 127, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "optimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 213, + 154, + 227 + ], + "spans": [ + { + "bbox": [ + 127, + 213, + 154, + 227 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "[11] Jiefeng Chen, Xi Wu, Yingyu Liang, Somesh Jha, et al. Robust out-of-distribution detection in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 244, + 360, + 257 + ], + "spans": [ + { + "bbox": [ + 127, + 244, + 360, + 257 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:2003.09711, 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 504, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "[12] Kamil Ciosek, Vincent Fortuin, Ryota Tomioka, Katja Hofmann, and Richard Turner. Conserva-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 126, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "tive uncertainty estimation by fitting prior networks. In International Conference on Learning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 285, + 222, + 298 + ], + "spans": [ + { + "bbox": [ + 127, + 285, + 222, + 298 + ], + "score": 1.0, + "content": "Representations, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 503, + 327 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "[13] Francesco D’Angelo and Vincent Fortuin. Annealed stein variational gradient descent. arXiv", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 316, + 266, + 327 + ], + "spans": [ + { + "bbox": [ + 126, + 316, + 266, + 327 + ], + "score": 1.0, + "content": "preprint arXiv:2101.09815, 2021.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 503, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "[14] Francesco D’Angelo and Christian Henning. Uncertainty-based out-of-distribution detection", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 345, + 448, + 358 + ], + "spans": [ + { + "bbox": [ + 127, + 345, + 448, + 358 + ], + "score": 1.0, + "content": "requires suitable function space priors. arXiv preprint arXiv:2110.06020, 2021.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 504, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "score": 1.0, + "content": "[15] Francesco D’Angelo, Vincent Fortuin, and Florian Wenzel. On stein variational neural network", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 376, + 338, + 387 + ], + "spans": [ + { + "bbox": [ + 127, + 376, + 338, + 387 + ], + "score": 1.0, + "content": "ensembles. arXiv preprint arXiv:2106.10760, 2021.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 504, + 417 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "[16] A Duncan, Nikolas Nuesken, and Lukasz Szpruch. On the geometry of stein variational gradient", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 405, + 326, + 417 + ], + "spans": [ + { + "bbox": [ + 127, + 405, + 326, + 417 + ], + "score": 1.0, + "content": "descent. arXiv preprint arXiv:1912.00894, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 504, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "score": 1.0, + "content": "[17] Michael W Dusenberry, Ghassen Jerfel, Yeming Wen, Yi-an Ma, Jasper Snoek, Katherine", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 127, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 446, + 371, + 458 + ], + "spans": [ + { + "bbox": [ + 127, + 446, + 371, + 458 + ], + "score": 1.0, + "content": "with rank-1 factors. arXiv preprint arXiv:2005.07186, 2020.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 504, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "score": 1.0, + "content": "[18] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 477, + 341, + 488 + ], + "spans": [ + { + "bbox": [ + 127, + 477, + 341, + 488 + ], + "score": 1.0, + "content": "perspective. arXiv preprint arXiv:1912.02757, 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 504, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 506, + 508 + ], + "score": 1.0, + "content": "[19] Vincent Fortuin. Priors in bayesian deep learning: A review. arXiv preprint arXiv:2105.06868,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 505, + 155, + 519 + ], + "spans": [ + { + "bbox": [ + 126, + 505, + 155, + 519 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 506, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 507, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 507, + 538 + ], + "score": 1.0, + "content": "[20] Vincent Fortuin, Mark Collier, Florian Wenzel, James Allingham, Jeremiah Liu, Dustin Tran,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 534, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 126, + 534, + 506, + 550 + ], + "score": 1.0, + "content": "Balaji Lakshminarayanan, Jesse Berent, Rodolphe Jenatton, and Effrosyni Kokiopoulou. Deep", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 126, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "classifiers with label noise modeling and distance awareness. arXiv preprint arXiv:2110.02609,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 557, + 155, + 570 + ], + "spans": [ + { + "bbox": [ + 126, + 557, + 155, + 570 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "[21] Vincent Fortuin, Adrià Garriga-Alonso, Mark van der Wilk, and Laurence Aitchison. Bnnpriors:", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 126, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 126, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "A library for bayesian neural network inference with different prior distributions. Software", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 128, + 599, + 246, + 612 + ], + "spans": [ + { + "bbox": [ + 128, + 599, + 246, + 612 + ], + "score": 1.0, + "content": "Impacts, page 100079, 2021.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "[22] Vincent Fortuin, Adrià Garriga-Alonso, Florian Wenzel, Gunnar Rätsch, Richard Turner, Mark", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 126, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "van der Wilk, and Laurence Aitchison. Bayesian neural network priors revisited. arXiv preprint", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 640, + 231, + 651 + ], + "spans": [ + { + "bbox": [ + 127, + 640, + 231, + 651 + ], + "score": 1.0, + "content": "arXiv:2102.06571, 2021.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "[23] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry Vetrov, and Andrew Gordon", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 668, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 127, + 668, + 505, + 683 + ], + "score": 1.0, + "content": "Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. arXiv preprint", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 681, + 231, + 693 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 231, + 693 + ], + "score": 1.0, + "content": "arXiv:1802.10026, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 502, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 504, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 504, + 713 + ], + "score": 1.0, + "content": "[24] Adrià Garriga-Alonso and Vincent Fortuin. Exact langevin dynamics with stochastic gradients.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 711, + 291, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 291, + 723 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2102.01691, 2021.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 110, + 72, + 433, + 85 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 433, + 86 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 433, + 86 + ], + "score": 1.0, + "content": "[6] Leo Breiman. Bagging predictors. Machine learning, 24(2):123–140, 1996.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 111, + 72, + 433, + 86 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 91, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 110, + 91, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 110, + 91, + 506, + 106 + ], + "score": 1.0, + "content": "[7] David R Burt, Sebastian W Ober, Adrià Garriga-Alonso, and Mark van der Wilk. Understanding", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 102, + 449, + 115 + ], + "spans": [ + { + "bbox": [ + 127, + 102, + 449, + 115 + ], + "score": 1.0, + "content": "variational inference in function-space. arXiv preprint arXiv:2011.09421, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 110, + 91, + 506, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 121, + 504, + 155 + ], + "lines": [ + { + "bbox": [ + 110, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 110, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "[8] Pratik Chaudhari and Stefano Soatto. Stochastic gradient descent performs variational inference,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "converges to limit cycles for deep networks. In 2018 Information Theory and Applications", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 144, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 144, + 297, + 156 + ], + "score": 1.0, + "content": "Workshop (ITA), pages 1–10. IEEE, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 110, + 121, + 506, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 162, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 111, + 162, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 111, + 162, + 505, + 175 + ], + "score": 1.0, + "content": "[9] An Mei Chen, Haw-minn Lu, and Robert Hecht-Nielsen. On the geometry of feedforward", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 173, + 420, + 186 + ], + "spans": [ + { + "bbox": [ + 127, + 173, + 420, + 186 + ], + "score": 1.0, + "content": "neural network error surfaces. Neural computation, 5(6):910–927, 1993.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 111, + 162, + 505, + 186 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 192, + 505, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "[10] Changyou Chen, Ruiyi Zhang, Wenlin Wang, Bai Li, and Liqun Chen. A unified particle-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 203, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 127, + 203, + 506, + 216 + ], + "score": 1.0, + "content": "optimization framework for scalable bayesian sampling. arXiv preprint arXiv:1805.11659,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 213, + 154, + 227 + ], + "spans": [ + { + "bbox": [ + 127, + 213, + 154, + 227 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 191, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 256 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 246 + ], + "score": 1.0, + "content": "[11] Jiefeng Chen, Xi Wu, Yingyu Liang, Somesh Jha, et al. Robust out-of-distribution detection in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 244, + 360, + 257 + ], + "spans": [ + { + "bbox": [ + 127, + 244, + 360, + 257 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:2003.09711, 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 233, + 505, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 504, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "[12] Kamil Ciosek, Vincent Fortuin, Ryota Tomioka, Katja Hofmann, and Richard Turner. Conserva-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 126, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "tive uncertainty estimation by fitting prior networks. In International Conference on Learning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 285, + 222, + 298 + ], + "spans": [ + { + "bbox": [ + 127, + 285, + 222, + 298 + ], + "score": 1.0, + "content": "Representations, 2019.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 263, + 506, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 503, + 327 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "[13] Francesco D’Angelo and Vincent Fortuin. Annealed stein variational gradient descent. arXiv", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 316, + 266, + 327 + ], + "spans": [ + { + "bbox": [ + 126, + 316, + 266, + 327 + ], + "score": 1.0, + "content": "preprint arXiv:2101.09815, 2021.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 304, + 505, + 327 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 334, + 503, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 347 + ], + "score": 1.0, + "content": "[14] Francesco D’Angelo and Christian Henning. Uncertainty-based out-of-distribution detection", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 345, + 448, + 358 + ], + "spans": [ + { + "bbox": [ + 127, + 345, + 448, + 358 + ], + "score": 1.0, + "content": "requires suitable function space priors. arXiv preprint arXiv:2110.06020, 2021.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 334, + 505, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 504, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "score": 1.0, + "content": "[15] Francesco D’Angelo, Vincent Fortuin, and Florian Wenzel. On stein variational neural network", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 376, + 338, + 387 + ], + "spans": [ + { + "bbox": [ + 127, + 376, + 338, + 387 + ], + "score": 1.0, + "content": "ensembles. arXiv preprint arXiv:2106.10760, 2021.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 363, + 506, + 387 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 504, + 417 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "[16] A Duncan, Nikolas Nuesken, and Lukasz Szpruch. On the geometry of stein variational gradient", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 405, + 326, + 417 + ], + "spans": [ + { + "bbox": [ + 127, + 405, + 326, + 417 + ], + "score": 1.0, + "content": "descent. arXiv preprint arXiv:1912.00894, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 394, + 505, + 417 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 504, + 458 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "score": 1.0, + "content": "[17] Michael W Dusenberry, Ghassen Jerfel, Yeming Wen, Yi-an Ma, Jasper Snoek, Katherine", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 127, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "Heller, Balaji Lakshminarayanan, and Dustin Tran. Efficient and scalable bayesian neural nets", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 446, + 371, + 458 + ], + "spans": [ + { + "bbox": [ + 127, + 446, + 371, + 458 + ], + "score": 1.0, + "content": "with rank-1 factors. arXiv preprint arXiv:2005.07186, 2020.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 423, + 505, + 458 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 504, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 479 + ], + "score": 1.0, + "content": "[18] Stanislav Fort, Huiyi Hu, and Balaji Lakshminarayanan. Deep ensembles: A loss landscape", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 477, + 341, + 488 + ], + "spans": [ + { + "bbox": [ + 127, + 477, + 341, + 488 + ], + "score": 1.0, + "content": "perspective. arXiv preprint arXiv:1912.02757, 2019.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 464, + 505, + 488 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 504, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 494, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 506, + 508 + ], + "score": 1.0, + "content": "[19] Vincent Fortuin. Priors in bayesian deep learning: A review. arXiv preprint arXiv:2105.06868,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 505, + 155, + 519 + ], + "spans": [ + { + "bbox": [ + 126, + 505, + 155, + 519 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 494, + 506, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 506, + 569 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 507, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 507, + 538 + ], + "score": 1.0, + "content": "[20] Vincent Fortuin, Mark Collier, Florian Wenzel, James Allingham, Jeremiah Liu, Dustin Tran,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 534, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 126, + 534, + 506, + 550 + ], + "score": 1.0, + "content": "Balaji Lakshminarayanan, Jesse Berent, Rodolphe Jenatton, and Effrosyni Kokiopoulou. Deep", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 126, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "classifiers with label noise modeling and distance awareness. arXiv preprint arXiv:2110.02609,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 557, + 155, + 570 + ], + "spans": [ + { + "bbox": [ + 126, + 557, + 155, + 570 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 524, + 507, + 570 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 576, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "[21] Vincent Fortuin, Adrià Garriga-Alonso, Mark van der Wilk, and Laurence Aitchison. Bnnpriors:", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 126, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 126, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "A library for bayesian neural network inference with different prior distributions. Software", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 128, + 599, + 246, + 612 + ], + "spans": [ + { + "bbox": [ + 128, + 599, + 246, + 612 + ], + "score": 1.0, + "content": "Impacts, page 100079, 2021.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 576, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "[22] Vincent Fortuin, Adrià Garriga-Alonso, Florian Wenzel, Gunnar Rätsch, Richard Turner, Mark", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 126, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "van der Wilk, and Laurence Aitchison. Bayesian neural network priors revisited. arXiv preprint", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 640, + 231, + 651 + ], + "spans": [ + { + "bbox": [ + 127, + 640, + 231, + 651 + ], + "score": 1.0, + "content": "arXiv:2102.06571, 2021.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 617, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 503, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "[23] Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry Vetrov, and Andrew Gordon", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 668, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 127, + 668, + 505, + 683 + ], + "score": 1.0, + "content": "Wilson. Loss surfaces, mode connectivity, and fast ensembling of dnns. arXiv preprint", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 681, + 231, + 693 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 231, + 693 + ], + "score": 1.0, + "content": "arXiv:1802.10026, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 658, + 505, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 502, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 504, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 504, + 713 + ], + "score": 1.0, + "content": "[24] Adrià Garriga-Alonso and Vincent Fortuin. Exact langevin dynamics with stochastic gradients.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 711, + 291, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 291, + 723 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2102.01691, 2021.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 699, + 504, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 103, + 34, + 507, + 725 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[25] Mario Geiger, Arthur Jacot, Stefano Spigler, Franck Gabriel, Levent Sagun, Stéphane d’Ascoli,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "Giulio Biroli, Clément Hongler, and Matthieu Wyart. Scaling description of generalization", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 95, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 126, + 95, + 506, + 109 + ], + "score": 1.0, + "content": "with number of parameters in deep learning. Journal of Statistical Mechanics: Theory and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 105, + 273, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 105, + 273, + 119 + ], + "score": 1.0, + "content": "Experiment, 2020(2):023401, 2020.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "[26] L. K. Hansen and P. Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 135, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 127, + 135, + 506, + 147 + ], + "score": 1.0, + "content": "Intell., 12(10):993–1001, October 1990. ISSN 0162-8828. doi: 10.1109/34.58871. URL", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 146, + 300, + 160 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 300, + 160 + ], + "score": 1.0, + "content": "https://doi.org/10.1109/34.58871.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "[27] Bobby He, Balaji Lakshminarayanan, and Yee Whye Teh. Bayesian deep ensembles via the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 176, + 381, + 189 + ], + "spans": [ + { + "bbox": [ + 126, + 176, + 381, + 189 + ], + "score": 1.0, + "content": "neural tangent kernel. arXiv preprint arXiv:2007.05864, 2020.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 194, + 507, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 507, + 209 + ], + "score": 1.0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers:", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 127, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 216, + 410, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 410, + 230 + ], + "score": 1.0, + "content": "international conference on computer vision, pages 1026–1034, 2015.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 235, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 506, + 250 + ], + "score": 1.0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 247, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 126, + 247, + 506, + 259 + ], + "score": 1.0, + "content": "recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 257, + 218, + 270 + ], + "spans": [ + { + "bbox": [ + 126, + 257, + 218, + 270 + ], + "score": 1.0, + "content": "pages 770–778, 2016.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "[30] Robert Hecht-Nielsen. On the algebraic structure of feedforward network weight spaces. In", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 126, + 288, + 376, + 302 + ], + "spans": [ + { + "bbox": [ + 126, + 288, + 376, + 302 + ], + "score": 1.0, + "content": "Advanced Neural Computers, pages 129–135. Elsevier, 1990.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 306, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 506, + 318 + ], + "score": 1.0, + "content": "[31] Christian Henning, Maria R Cervera, Francesco D’Angelo, Johannes von Oswald, Regina", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 128, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 128, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "Traber, Benjamin Ehret, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Posterior", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 327, + 433, + 342 + ], + "spans": [ + { + "bbox": [ + 126, + 327, + 433, + 342 + ], + "score": 1.0, + "content": "meta-replay for continual learning. arXiv preprint arXiv:2103.01133, 2021.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 347, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 507, + 361 + ], + "score": 1.0, + "content": "[32] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 358, + 472, + 372 + ], + "spans": [ + { + "bbox": [ + 126, + 358, + 472, + 372 + ], + "score": 1.0, + "content": "Snapshot ensembles: Train 1, get m for free. arXiv preprint arXiv:1704.00109, 2017.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 376, + 507, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 507, + 390 + ], + "score": 1.0, + "content": "[33] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 388, + 507, + 403 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 507, + 403 + ], + "score": 1.0, + "content": "Snapshot ensembles: Train 1, get M for free. CoRR, abs/1704.00109, 2017. URL http:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 398, + 268, + 412 + ], + "spans": [ + { + "bbox": [ + 126, + 398, + 268, + 412 + ], + "score": 1.0, + "content": "//arxiv.org/abs/1704.00109.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "score": 1.0, + "content": "[34] Alexander Immer, Matthias Bauer, Vincent Fortuin, Gunnar Rätsch, and Mohammad Emtiyaz", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 428, + 504, + 441 + ], + "spans": [ + { + "bbox": [ + 127, + 428, + 504, + 441 + ], + "score": 1.0, + "content": "Khan. Scalable marginal likelihood estimation for model selection in deep learning. arXiv", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 439, + 267, + 452 + ], + "spans": [ + { + "bbox": [ + 126, + 439, + 267, + 452 + ], + "score": 1.0, + "content": "preprint arXiv:2104.04975, 2021.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "[35] Alexander Immer, Maciej Korzepa, and Matthias Bauer. Improving predictions of bayesian", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 126, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "neural nets via local linearization. In International Conference on Artificial Intelligence and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 480, + 292, + 494 + ], + "spans": [ + { + "bbox": [ + 126, + 480, + 292, + 494 + ], + "score": 1.0, + "content": "Statistics, pages 703–711. PMLR, 2021.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 500, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 506, + 511 + ], + "score": 1.0, + "content": "[36] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Wilson. What are", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 509, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 126, + 509, + 480, + 523 + ], + "score": 1.0, + "content": "bayesian neural network posteriors really like? arXiv preprint arXiv:2104.14421, 2021.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 543 + ], + "score": 1.0, + "content": "[37] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: convergence and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 540, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 126, + 540, + 506, + 553 + ], + "score": 1.0, + "content": "generalization in neural networks. In Proceedings of the 32nd International Conference on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 550, + 390, + 565 + ], + "spans": [ + { + "bbox": [ + 126, + 550, + 390, + 565 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pages 8580–8589, 2018.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 568, + 507, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 507, + 583 + ], + "score": 1.0, + "content": "[38] Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker–", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 580, + 435, + 594 + ], + "spans": [ + { + "bbox": [ + 126, + 580, + 435, + 594 + ], + "score": 1.0, + "content": "planck equation. SIAM journal on mathematical analysis, 29(1):1–17, 1998.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 611, + 226, + 623 + ], + "spans": [ + { + "bbox": [ + 127, + 611, + 226, + 623 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "score": 1.0, + "content": "[40] Anna Korba, Adil Salim, Michael Arbel, Giulia Luise, and Arthur Gretton. A non-asymptotic", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 126, + 639, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 126, + 639, + 506, + 655 + ], + "score": 1.0, + "content": "analysis for stein variational gradient descent. Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 651, + 205, + 664 + ], + "spans": [ + { + "bbox": [ + 126, + 651, + 205, + 664 + ], + "score": 1.0, + "content": "Systems, 33, 2020.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 669, + 465, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 465, + 683 + ], + "score": 1.0, + "content": "[41] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "[42] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "predictive uncertainty estimation using deep ensembles. In Advances in neural information", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 125, + 711, + 309, + 724 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 309, + 724 + ], + "score": 1.0, + "content": "processing systems, pages 6402–6413, 2017.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 23 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 103, + 34, + 507, + 725 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[25] Mario Geiger, Arthur Jacot, Stefano Spigler, Franck Gabriel, Levent Sagun, Stéphane d’Ascoli,", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 83, + 506, + 97 + ], + "score": 1.0, + "content": "Giulio Biroli, Clément Hongler, and Matthieu Wyart. Scaling description of generalization", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 95, + 506, + 109 + ], + "spans": [ + { + "bbox": [ + 126, + 95, + 506, + 109 + ], + "score": 1.0, + "content": "with number of parameters in deep learning. Journal of Statistical Mechanics: Theory and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 105, + 273, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 105, + 273, + 119 + ], + "score": 1.0, + "content": "Experiment, 2020(2):023401, 2020.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "[26] L. K. Hansen and P. Salamon. Neural network ensembles. IEEE Trans. Pattern Anal. Mach.", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 135, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 127, + 135, + 506, + 147 + ], + "score": 1.0, + "content": "Intell., 12(10):993–1001, October 1990. ISSN 0162-8828. doi: 10.1109/34.58871. URL", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 146, + 300, + 160 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 300, + 160 + ], + "score": 1.0, + "content": "https://doi.org/10.1109/34.58871.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "[27] Bobby He, Balaji Lakshminarayanan, and Yee Whye Teh. Bayesian deep ensembles via the", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 176, + 381, + 189 + ], + "spans": [ + { + "bbox": [ + 126, + 176, + 381, + 189 + ], + "score": 1.0, + "content": "neural tangent kernel. arXiv preprint arXiv:2007.05864, 2020.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 194, + 507, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 507, + 209 + ], + "score": 1.0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers:", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 127, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 216, + 410, + 230 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 410, + 230 + ], + "score": 1.0, + "content": "international conference on computer vision, pages 1026–1034, 2015.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 235, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 104, + 235, + 506, + 250 + ], + "score": 1.0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 247, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 126, + 247, + 506, + 259 + ], + "score": 1.0, + "content": "recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 126, + 257, + 218, + 270 + ], + "spans": [ + { + "bbox": [ + 126, + 257, + 218, + 270 + ], + "score": 1.0, + "content": "pages 770–778, 2016.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 290 + ], + "score": 1.0, + "content": "[30] Robert Hecht-Nielsen. On the algebraic structure of feedforward network weight spaces. In", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 288, + 376, + 302 + ], + "spans": [ + { + "bbox": [ + 126, + 288, + 376, + 302 + ], + "score": 1.0, + "content": "Advanced Neural Computers, pages 129–135. Elsevier, 1990.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 306, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 506, + 318 + ], + "score": 1.0, + "content": "[31] Christian Henning, Maria R Cervera, Francesco D’Angelo, Johannes von Oswald, Regina", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 128, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "Traber, Benjamin Ehret, Seijin Kobayashi, João Sacramento, and Benjamin F Grewe. Posterior", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 327, + 433, + 342 + ], + "spans": [ + { + "bbox": [ + 126, + 327, + 433, + 342 + ], + "score": 1.0, + "content": "meta-replay for continual learning. arXiv preprint arXiv:2103.01133, 2021.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 347, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 507, + 361 + ], + "score": 1.0, + "content": "[32] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E Hopcroft, and Kilian Q Weinberger.", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 358, + 472, + 372 + ], + "spans": [ + { + "bbox": [ + 126, + 358, + 472, + 372 + ], + "score": 1.0, + "content": "Snapshot ensembles: Train 1, get m for free. arXiv preprint arXiv:1704.00109, 2017.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 376, + 507, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 507, + 390 + ], + "score": 1.0, + "content": "[33] Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger.", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 388, + 507, + 403 + ], + "spans": [ + { + "bbox": [ + 127, + 388, + 507, + 403 + ], + "score": 1.0, + "content": "Snapshot ensembles: Train 1, get M for free. CoRR, abs/1704.00109, 2017. URL http:", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 398, + 268, + 412 + ], + "spans": [ + { + "bbox": [ + 126, + 398, + 268, + 412 + ], + "score": 1.0, + "content": "//arxiv.org/abs/1704.00109.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 506, + 432 + ], + "score": 1.0, + "content": "[34] Alexander Immer, Matthias Bauer, Vincent Fortuin, Gunnar Rätsch, and Mohammad Emtiyaz", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 428, + 504, + 441 + ], + "spans": [ + { + "bbox": [ + 127, + 428, + 504, + 441 + ], + "score": 1.0, + "content": "Khan. Scalable marginal likelihood estimation for model selection in deep learning. arXiv", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 439, + 267, + 452 + ], + "spans": [ + { + "bbox": [ + 126, + 439, + 267, + 452 + ], + "score": 1.0, + "content": "preprint arXiv:2104.04975, 2021.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "[35] Alexander Immer, Maciej Korzepa, and Matthias Bauer. Improving predictions of bayesian", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 126, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "neural nets via local linearization. In International Conference on Artificial Intelligence and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 480, + 292, + 494 + ], + "spans": [ + { + "bbox": [ + 126, + 480, + 292, + 494 + ], + "score": 1.0, + "content": "Statistics, pages 703–711. PMLR, 2021.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 500, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 506, + 511 + ], + "score": 1.0, + "content": "[36] Pavel Izmailov, Sharad Vikram, Matthew D Hoffman, and Andrew Gordon Wilson. What are", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 509, + 480, + 523 + ], + "spans": [ + { + "bbox": [ + 126, + 509, + 480, + 523 + ], + "score": 1.0, + "content": "bayesian neural network posteriors really like? arXiv preprint arXiv:2104.14421, 2021.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 528, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 543 + ], + "score": 1.0, + "content": "[37] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: convergence and", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 540, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 126, + 540, + 506, + 553 + ], + "score": 1.0, + "content": "generalization in neural networks. In Proceedings of the 32nd International Conference on", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 550, + 390, + 565 + ], + "spans": [ + { + "bbox": [ + 126, + 550, + 390, + 565 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pages 8580–8589, 2018.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 568, + 507, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 507, + 583 + ], + "score": 1.0, + "content": "[38] Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker–", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 580, + 435, + 594 + ], + "spans": [ + { + "bbox": [ + 126, + 580, + 435, + 594 + ], + "score": 1.0, + "content": "planck equation. SIAM journal on mathematical analysis, 29(1):1–17, 1998.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "[39] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 611, + 226, + 623 + ], + "spans": [ + { + "bbox": [ + 127, + 611, + 226, + 623 + ], + "score": 1.0, + "content": "arXiv:1412.6980, 2014.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 643 + ], + "score": 1.0, + "content": "[40] Anna Korba, Adil Salim, Michael Arbel, Giulia Luise, and Arthur Gretton. A non-asymptotic", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 639, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 126, + 639, + 506, + 655 + ], + "score": 1.0, + "content": "analysis for stein variational gradient descent. Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 651, + 205, + 664 + ], + "spans": [ + { + "bbox": [ + 126, + 651, + 205, + 664 + ], + "score": 1.0, + "content": "Systems, 33, 2020.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 669, + 465, + 683 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 465, + 683 + ], + "score": 1.0, + "content": "[41] Alex Krizhevsky et al. Learning multiple layers of features from tiny images. 2009.", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "[42] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "predictive uncertainty estimation using deep ensembles. In Advances in neural information", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 125, + 711, + 309, + 724 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 309, + 724 + ], + "score": 1.0, + "content": "processing systems, pages 6402–6413, 2017.", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "score": 1.0, + "content": "[43] Yann LeCun. The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/,", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 81, + 154, + 99 + ], + "spans": [ + { + "bbox": [ + 126, + 81, + 154, + 99 + ], + "score": 1.0, + "content": "1998.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 102, + 507, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 507, + 116 + ], + "score": 1.0, + "content": "[44] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington,", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 113, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 126, + 113, + 506, + 127 + ], + "score": 1.0, + "content": "and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 123, + 232, + 137 + ], + "spans": [ + { + "bbox": [ + 126, + 123, + 232, + 137 + ], + "score": 1.0, + "content": "arXiv:1711.00165, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 141, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 156 + ], + "score": 1.0, + "content": "[45] E. Levin, N. Tishby, and S. A. Solla. A statistical approach to learning and generalization in", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 154, + 474, + 167 + ], + "spans": [ + { + "bbox": [ + 126, + 154, + 474, + 167 + ], + "score": 1.0, + "content": "layered neural networks. Proc. of the IEEE – Special issue on Neural Networks, 1990.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "score": 1.0, + "content": "[46] Yingzhen Li and Richard E Turner. Gradient estimators for implicit models. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 183, + 231, + 197 + ], + "spans": [ + { + "bbox": [ + 127, + 183, + 231, + 197 + ], + "score": 1.0, + "content": "arXiv:1705.07107, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 201, + 507, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 507, + 216 + ], + "score": 1.0, + "content": "[47] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 213, + 483, + 225 + ], + "spans": [ + { + "bbox": [ + 127, + 213, + 483, + 225 + ], + "score": 1.0, + "content": "distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 231, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 245 + ], + "score": 1.0, + "content": "[48] Chang Liu, Jingwei Zhuo, Pengyu Cheng, Ruiyi Zhang, and Jun Zhu. Understanding and", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 241, + 507, + 257 + ], + "spans": [ + { + "bbox": [ + 125, + 241, + 507, + 257 + ], + "score": 1.0, + "content": "accelerating particle-based variational inference. In International Conference on Machine", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 253, + 302, + 267 + ], + "spans": [ + { + "bbox": [ + 126, + 253, + 302, + 267 + ], + "score": 1.0, + "content": "Learning, pages 4082–4092. PMLR, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "[49] Jeremiah Zhe Liu, Zi Lin, Shreyas Padhy, Dustin Tran, Tania Bedrax-Weiss, and Balaji Laksh-", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 282, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 506, + 297 + ], + "score": 1.0, + "content": "minarayanan. Simple and principled uncertainty estimation with deterministic deep learning via", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 293, + 372, + 307 + ], + "spans": [ + { + "bbox": [ + 127, + 293, + 372, + 307 + ], + "score": 1.0, + "content": "distance awareness. arXiv preprint arXiv:2006.10108, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 107, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "[50] Qiang Liu. Stein variational gradient descent as gradient flow. In Advances in neural information", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 324, + 310, + 337 + ], + "spans": [ + { + "bbox": [ + 125, + 324, + 310, + 337 + ], + "score": 1.0, + "content": "processing systems, pages 3115–3123, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 341, + 507, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 341, + 507, + 356 + ], + "score": 1.0, + "content": "[51] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 127, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "inference algorithm. In Advances in neural information processing systems, pages 2378–2386,", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 362, + 155, + 378 + ], + "spans": [ + { + "bbox": [ + 127, + 362, + 155, + 378 + ], + "score": 1.0, + "content": "2016.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "[52] Qiang Liu, Jason Lee, and Michael Jordan. A kernelized stein discrepancy for goodness-of-fit", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 393, + 473, + 407 + ], + "spans": [ + { + "bbox": [ + 127, + 393, + 473, + 407 + ], + "score": 1.0, + "content": "tests. In International conference on machine learning, pages 276–284. PMLR, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "score": 1.0, + "content": "[53] Clare Lyle, Lisa Schut, Robin Ru, Yarin Gal, and Mark van der Wilk. A bayesian perspective", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 422, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 422, + 507, + 437 + ], + "score": 1.0, + "content": "on training speed and model selection. Advances in Neural Information Processing Systems, 33,", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 433, + 155, + 448 + ], + "spans": [ + { + "bbox": [ + 126, + 433, + 155, + 448 + ], + "score": 1.0, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "[54] David JC MacKay. A practical bayesian framework for backpropagation networks. Neural", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 464, + 268, + 477 + ], + "spans": [ + { + "bbox": [ + 126, + 464, + 268, + 477 + ], + "score": 1.0, + "content": "computation, 4(3):448–472, 1992.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 482, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 496 + ], + "score": 1.0, + "content": "[55] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 493, + 266, + 506 + ], + "spans": [ + { + "bbox": [ + 125, + 493, + 266, + 506 + ], + "score": 1.0, + "content": "preprint arXiv:1905.00076, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 510, + 507, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 507, + 526 + ], + "score": 1.0, + "content": "[56] Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 523, + 424, + 537 + ], + "spans": [ + { + "bbox": [ + 127, + 523, + 424, + 537 + ], + "score": 1.0, + "content": "approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 541, + 507, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 507, + 555 + ], + "score": 1.0, + "content": "[57] Alexander G de G Matthews, Jiri Hron, Richard E Turner, and Zoubin Ghahramani. Sample-", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 127, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "then-optimize posterior sampling for bayesian linear models. In NeurIPS Workshop on Advances", + "type": "text", + "cross_page": true + } + ], + "index": 34 + }, + { + "bbox": [ + 128, + 564, + 298, + 576 + ], + "spans": [ + { + "bbox": [ + 128, + 564, + 298, + 576 + ], + "score": 1.0, + "content": "in Approximate Bayesian Inference, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "[58] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 591, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 127, + 591, + 505, + 606 + ], + "score": 1.0, + "content": "probabilities using bayesian binning. In Proceedings of the AAAI Conference on Artificial", + "type": "text", + "cross_page": true + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 603, + 253, + 617 + ], + "spans": [ + { + "bbox": [ + 127, + 603, + 253, + 617 + ], + "score": 1.0, + "content": "Intelligence, volume 29, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 622, + 383, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 383, + 635 + ], + "score": 1.0, + "content": "[59] Radford M Neal. Bayesian learning for neural networks. 1995.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "[60] Radford M Neal. Bayesian learning for neural networks, volume 118. Springer Science &", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 652, + 221, + 664 + ], + "spans": [ + { + "bbox": [ + 127, + 652, + 221, + 664 + ], + "score": 1.0, + "content": "Business Media, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "score": 1.0, + "content": "[61] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 682, + 429, + 694 + ], + "spans": [ + { + "bbox": [ + 127, + 682, + 429, + 694 + ], + "score": 1.0, + "content": "Reading digits in natural images with unsupervised feature learning. 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "[62] Ian Osband, Benjamin Van Roy, Daniel J Russo, Zheng Wen, et al. Deep exploration via", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 711, + 416, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 416, + 723 + ], + "score": 1.0, + "content": "randomized value functions. J. Mach. Learn. Res., 20(124):1–62, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_end_line": true + } + ], + "index": 23, + "bbox_fs": [ + 104, + 72, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 45, + 507, + 728 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 85 + ], + "score": 1.0, + "content": "[43] Yann LeCun. The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 81, + 154, + 99 + ], + "spans": [ + { + "bbox": [ + 126, + 81, + 154, + 99 + ], + "score": 1.0, + "content": "1998.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 102, + 507, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 507, + 116 + ], + "score": 1.0, + "content": "[44] Jaehoon Lee, Yasaman Bahri, Roman Novak, Samuel S Schoenholz, Jeffrey Pennington,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 113, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 126, + 113, + 506, + 127 + ], + "score": 1.0, + "content": "and Jascha Sohl-Dickstein. Deep neural networks as gaussian processes. arXiv preprint", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 123, + 232, + 137 + ], + "spans": [ + { + "bbox": [ + 126, + 123, + 232, + 137 + ], + "score": 1.0, + "content": "arXiv:1711.00165, 2017.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 141, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 505, + 156 + ], + "score": 1.0, + "content": "[45] E. Levin, N. Tishby, and S. A. Solla. A statistical approach to learning and generalization in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 154, + 474, + 167 + ], + "spans": [ + { + "bbox": [ + 126, + 154, + 474, + 167 + ], + "score": 1.0, + "content": "layered neural networks. Proc. of the IEEE – Special issue on Neural Networks, 1990.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "score": 1.0, + "content": "[46] Yingzhen Li and Richard E Turner. Gradient estimators for implicit models. arXiv preprint", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 183, + 231, + 197 + ], + "spans": [ + { + "bbox": [ + 127, + 183, + 231, + 197 + ], + "score": 1.0, + "content": "arXiv:1705.07107, 2017.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 201, + 507, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 507, + 216 + ], + "score": 1.0, + "content": "[47] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 213, + 483, + 225 + ], + "spans": [ + { + "bbox": [ + 127, + 213, + 483, + 225 + ], + "score": 1.0, + "content": "distribution image detection in neural networks. arXiv preprint arXiv:1706.02690, 2017.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 231, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 245 + ], + "score": 1.0, + "content": "[48] Chang Liu, Jingwei Zhuo, Pengyu Cheng, Ruiyi Zhang, and Jun Zhu. Understanding and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 125, + 241, + 507, + 257 + ], + "spans": [ + { + "bbox": [ + 125, + 241, + 507, + 257 + ], + "score": 1.0, + "content": "accelerating particle-based variational inference. In International Conference on Machine", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 253, + 302, + 267 + ], + "spans": [ + { + "bbox": [ + 126, + 253, + 302, + 267 + ], + "score": 1.0, + "content": "Learning, pages 4082–4092. PMLR, 2019.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "[49] Jeremiah Zhe Liu, Zi Lin, Shreyas Padhy, Dustin Tran, Tania Bedrax-Weiss, and Balaji Laksh-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 282, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 506, + 297 + ], + "score": 1.0, + "content": "minarayanan. Simple and principled uncertainty estimation with deterministic deep learning via", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 293, + 372, + 307 + ], + "spans": [ + { + "bbox": [ + 127, + 293, + 372, + 307 + ], + "score": 1.0, + "content": "distance awareness. arXiv preprint arXiv:2006.10108, 2020.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 107, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "[50] Qiang Liu. Stein variational gradient descent as gradient flow. In Advances in neural information", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 324, + 310, + 337 + ], + "spans": [ + { + "bbox": [ + 125, + 324, + 310, + 337 + ], + "score": 1.0, + "content": "processing systems, pages 3115–3123, 2017.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 341, + 507, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 341, + 507, + 356 + ], + "score": 1.0, + "content": "[51] Qiang Liu and Dilin Wang. Stein variational gradient descent: A general purpose bayesian", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 127, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "inference algorithm. In Advances in neural information processing systems, pages 2378–2386,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 362, + 155, + 378 + ], + "spans": [ + { + "bbox": [ + 127, + 362, + 155, + 378 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "[52] Qiang Liu, Jason Lee, and Michael Jordan. A kernelized stein discrepancy for goodness-of-fit", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 127, + 393, + 473, + 407 + ], + "spans": [ + { + "bbox": [ + 127, + 393, + 473, + 407 + ], + "score": 1.0, + "content": "tests. In International conference on machine learning, pages 276–284. PMLR, 2016.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 506, + 426 + ], + "score": 1.0, + "content": "[53] Clare Lyle, Lisa Schut, Robin Ru, Yarin Gal, and Mark van der Wilk. A bayesian perspective", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 422, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 127, + 422, + 507, + 437 + ], + "score": 1.0, + "content": "on training speed and model selection. Advances in Neural Information Processing Systems, 33,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 433, + 155, + 448 + ], + "spans": [ + { + "bbox": [ + 126, + 433, + 155, + 448 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "[54] David JC MacKay. A practical bayesian framework for backpropagation networks. Neural", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 464, + 268, + 477 + ], + "spans": [ + { + "bbox": [ + 126, + 464, + 268, + 477 + ], + "score": 1.0, + "content": "computation, 4(3):448–472, 1992.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 482, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 496 + ], + "score": 1.0, + "content": "[55] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 493, + 266, + 506 + ], + "spans": [ + { + "bbox": [ + 125, + 493, + 266, + 506 + ], + "score": 1.0, + "content": "preprint arXiv:1905.00076, 2019.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 510, + 507, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 507, + 526 + ], + "score": 1.0, + "content": "[56] Stephan Mandt, Matthew D Hoffman, and David M Blei. Stochastic gradient descent as", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 523, + 424, + 537 + ], + "spans": [ + { + "bbox": [ + 127, + 523, + 424, + 537 + ], + "score": 1.0, + "content": "approximate bayesian inference. arXiv preprint arXiv:1704.04289, 2017.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 541, + 507, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 507, + 555 + ], + "score": 1.0, + "content": "[57] Alexander G de G Matthews, Jiri Hron, Richard E Turner, and Zoubin Ghahramani. Sample-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 127, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "then-optimize posterior sampling for bayesian linear models. In NeurIPS Workshop on Advances", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 128, + 564, + 298, + 576 + ], + "spans": [ + { + "bbox": [ + 128, + 564, + 298, + 576 + ], + "score": 1.0, + "content": "in Approximate Bayesian Inference, 2017.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 581, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 506, + 595 + ], + "score": 1.0, + "content": "[58] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 591, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 127, + 591, + 505, + 606 + ], + "score": 1.0, + "content": "probabilities using bayesian binning. In Proceedings of the AAAI Conference on Artificial", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 603, + 253, + 617 + ], + "spans": [ + { + "bbox": [ + 127, + 603, + 253, + 617 + ], + "score": 1.0, + "content": "Intelligence, volume 29, 2015.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 622, + 383, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 383, + 635 + ], + "score": 1.0, + "content": "[59] Radford M Neal. Bayesian learning for neural networks. 1995.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "[60] Radford M Neal. Bayesian learning for neural networks, volume 118. Springer Science &", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 652, + 221, + 664 + ], + "spans": [ + { + "bbox": [ + 127, + 652, + 221, + 664 + ], + "score": 1.0, + "content": "Business Media, 2012.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "score": 1.0, + "content": "[61] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 682, + 429, + 694 + ], + "spans": [ + { + "bbox": [ + 127, + 682, + 429, + 694 + ], + "score": 1.0, + "content": "Reading digits in natural images with unsupervised feature learning. 2011.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "[62] Ian Osband, Benjamin Van Roy, Daniel J Russo, Zheng Wen, et al. Deep exploration via", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 127, + 711, + 416, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 416, + 723 + ], + "score": 1.0, + "content": "randomized value functions. J. Mach. Learn. Res., 20(124):1–62, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 22.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 45, + 507, + 728 + ], + "lines": [], + "index": 22.5, + "bbox_fs": [ + 104, + 73, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 506, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[63] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "Henning, and Benjamin F Grewe. Neural networks with late-phase weights. In International", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 127, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2021. URL https://openreview.net/forum?", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 106, + 205, + 118 + ], + "spans": [ + { + "bbox": [ + 127, + 106, + 205, + 118 + ], + "score": 1.0, + "content": "id=C0qJUx5dxFb.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 104, + 123, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "[64] Felix Otto. The geometry of dissipative evolution equations: the porous medium equation.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 131, + 155, + 147 + ], + "spans": [ + { + "bbox": [ + 126, + 131, + 155, + 147 + ], + "score": 1.0, + "content": "2001.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "score": 1.0, + "content": "[65] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 163, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 127, + 163, + 506, + 176 + ], + "score": 1.0, + "content": "Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty?", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 175, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 127, + 175, + 505, + 188 + ], + "score": 1.0, + "content": "evaluating predictive uncertainty under dataset shift. In Advances in Neural Information", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 186, + 321, + 197 + ], + "spans": [ + { + "bbox": [ + 127, + 186, + 321, + 197 + ], + "score": 1.0, + "content": "Processing Systems, pages 13991–14002, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "score": 1.0, + "content": "[66] Emanuel Parzen. On estimation of a probability density function and mode. The annals of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 214, + 325, + 227 + ], + "spans": [ + { + "bbox": [ + 127, + 214, + 325, + 227 + ], + "score": 1.0, + "content": "mathematical statistics, 33(3):1065–1076, 1962.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 107, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "[67] Alexandre Rame and Matthieu Cord. Dice: Diversity in deep ensembles via conditional", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 243, + 433, + 256 + ], + "spans": [ + { + "bbox": [ + 127, + 243, + 433, + 256 + ], + "score": 1.0, + "content": "redundancy adversarial estimation. arXiv preprint arXiv:2101.05544, 2021.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "[68] Geoffrey Roeder, Luke Metz, and Diederik P Kingma. On linear identifiability of learned", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 273, + 357, + 285 + ], + "spans": [ + { + "bbox": [ + 127, + 273, + 357, + 285 + ], + "score": 1.0, + "content": "representations. arXiv preprint arXiv:2007.00810, 2020.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 290, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 304 + ], + "score": 1.0, + "content": "[69] Jiaxin Shi, Shengyang Sun, and Jun Zhu. A spectral approach to gradient estimation for implicit", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 302, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 127, + 302, + 506, + 314 + ], + "score": 1.0, + "content": "distributions. In International Conference on Machine Learning, pages 4644–4653. PMLR,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 312, + 154, + 325 + ], + "spans": [ + { + "bbox": [ + 126, + 312, + 154, + 325 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "[70] Radhey S Singh. Improvement on some known nonparametric uniformly consistent estimators", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 341, + 425, + 354 + ], + "spans": [ + { + "bbox": [ + 127, + 341, + 425, + 354 + ], + "score": 1.0, + "content": "of derivatives of a density. The Annals of Statistics, pages 394–399, 1977.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 374 + ], + "score": 1.0, + "content": "[71] Shengyang Sun, Guodong Zhang, Jiaxin Shi, and Roger Grosse. Functional variational bayesian", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 371, + 361, + 383 + ], + "spans": [ + { + "bbox": [ + 127, + 371, + 361, + 383 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:1903.05779, 2019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 506, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "score": 1.0, + "content": "[72] Jakub Swiatkowski, Kevin Roth, Bastiaan S Veeling, Linh Tran, Joshua V Dillon, Stephan", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 127, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Mandt, Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. The k-tied", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 127, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "normal distribution: A compact parameterization of gaussian mean field posteriors in bayesian", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 423, + 360, + 434 + ], + "spans": [ + { + "bbox": [ + 127, + 423, + 360, + 434 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:2002.02655, 2020.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 109, + 440, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "[73] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F. Grewe. Continual", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 127, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "learning with hypernetworks. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 461, + 321, + 474 + ], + "spans": [ + { + "bbox": [ + 127, + 461, + 321, + 474 + ], + "score": 1.0, + "content": "URL https://arxiv.org/abs/1906.00695.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 480, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "score": 1.0, + "content": "[74] Maxime Wabartha, Audrey Durand, Vincent François-Lavet, and Joelle Pineau. Handling black", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 126, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "swan events in deep learning with diversely extrapolated neural networks. In IJCAI, pages", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 502, + 202, + 513 + ], + "spans": [ + { + "bbox": [ + 127, + 502, + 202, + 513 + ], + "score": 1.0, + "content": "2140–2147, 2020.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 520, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "[75] Yifei Wang, Peng Chen, and Wuchen Li. Projected wasserstein gradient descent for high-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 530, + 423, + 543 + ], + "spans": [ + { + "bbox": [ + 127, + 530, + 423, + 543 + ], + "score": 1.0, + "content": "dimensional bayesian inference. arXiv preprint arXiv:2102.06350, 2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "[76] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 560, + 397, + 572 + ], + "spans": [ + { + "bbox": [ + 127, + 560, + 397, + 572 + ], + "score": 1.0, + "content": "bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 108, + 578, + 505, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "[77] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 126, + 588, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 126, + 588, + 505, + 603 + ], + "score": 1.0, + "content": "In Proceedings of the 28th international conference on machine learning (ICML-11), pages", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 128, + 600, + 193, + 612 + ], + "spans": [ + { + "bbox": [ + 128, + 600, + 193, + 612 + ], + "score": 1.0, + "content": "681–688, 2011.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 105, + 618, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "[78] Yeming Wen, Dustin Tran, and Jimmy Ba. Batchensemble: an alternative approach to efficient", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 629, + 419, + 642 + ], + "spans": [ + { + "bbox": [ + 127, + 629, + 419, + 642 + ], + "score": 1.0, + "content": "ensemble and lifelong learning. arXiv preprint arXiv:2002.06715, 2020.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 648, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "[79] Florian Wenzel, Kevin Roth, Bastiaan S Veeling, Jakub Swiatkowski, Linh Tran, Stephan Mandt, ´", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 127, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. How good is the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 670, + 482, + 683 + ], + "spans": [ + { + "bbox": [ + 127, + 670, + 482, + 683 + ], + "score": 1.0, + "content": "bayes posterior in deep neural networks really? arXiv preprint arXiv:2002.02405, 2020.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "[80] Florian Wenzel, Jasper Snoek, Dustin Tran, and Rodolphe Jenatton. Hyperparameter ensembles", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 126, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "for robustness and uncertainty quantification. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 711, + 189, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 189, + 723 + ], + "score": 1.0, + "content": "Systems, 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 72, + 506, + 118 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[63] Johannes Von Oswald, Seijin Kobayashi, Joao Sacramento, Alexander Meulemans, Christian", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "Henning, and Benjamin F Grewe. Neural networks with late-phase weights. In International", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 127, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "Conference on Learning Representations, 2021. URL https://openreview.net/forum?", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 106, + 205, + 118 + ], + "spans": [ + { + "bbox": [ + 127, + 106, + 205, + 118 + ], + "score": 1.0, + "content": "id=C0qJUx5dxFb.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 72, + 506, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 123, + 505, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "[64] Felix Otto. The geometry of dissipative evolution equations: the porous medium equation.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 131, + 155, + 147 + ], + "spans": [ + { + "bbox": [ + 126, + 131, + 155, + 147 + ], + "score": 1.0, + "content": "2001.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 122, + 506, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 153, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 506, + 165 + ], + "score": 1.0, + "content": "[65] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 163, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 127, + 163, + 506, + 176 + ], + "score": 1.0, + "content": "Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model’s uncertainty?", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 175, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 127, + 175, + 505, + 188 + ], + "score": 1.0, + "content": "evaluating predictive uncertainty under dataset shift. In Advances in Neural Information", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 186, + 321, + 197 + ], + "spans": [ + { + "bbox": [ + 127, + 186, + 321, + 197 + ], + "score": 1.0, + "content": "Processing Systems, pages 13991–14002, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 152, + 506, + 197 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 504, + 226 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 506, + 217 + ], + "score": 1.0, + "content": "[66] Emanuel Parzen. On estimation of a probability density function and mode. The annals of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 214, + 325, + 227 + ], + "spans": [ + { + "bbox": [ + 127, + 214, + 325, + 227 + ], + "score": 1.0, + "content": "mathematical statistics, 33(3):1065–1076, 1962.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 202, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 107, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 107, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "[67] Alexandre Rame and Matthieu Cord. Dice: Diversity in deep ensembles via conditional", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 243, + 433, + 256 + ], + "spans": [ + { + "bbox": [ + 127, + 243, + 433, + 256 + ], + "score": 1.0, + "content": "redundancy adversarial estimation. arXiv preprint arXiv:2101.05544, 2021.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 107, + 232, + 505, + 256 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "[68] Geoffrey Roeder, Luke Metz, and Diederik P Kingma. On linear identifiability of learned", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 273, + 357, + 285 + ], + "spans": [ + { + "bbox": [ + 127, + 273, + 357, + 285 + ], + "score": 1.0, + "content": "representations. arXiv preprint arXiv:2007.00810, 2020.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 261, + 505, + 285 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 290, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 290, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 290, + 505, + 304 + ], + "score": 1.0, + "content": "[69] Jiaxin Shi, Shengyang Sun, and Jun Zhu. A spectral approach to gradient estimation for implicit", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 302, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 127, + 302, + 506, + 314 + ], + "score": 1.0, + "content": "distributions. In International Conference on Machine Learning, pages 4644–4653. PMLR,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 312, + 154, + 325 + ], + "spans": [ + { + "bbox": [ + 126, + 312, + 154, + 325 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 290, + 506, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "[70] Radhey S Singh. Improvement on some known nonparametric uniformly consistent estimators", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 341, + 425, + 354 + ], + "spans": [ + { + "bbox": [ + 127, + 341, + 425, + 354 + ], + "score": 1.0, + "content": "of derivatives of a density. The Annals of Statistics, pages 394–399, 1977.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 331, + 505, + 354 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 374 + ], + "score": 1.0, + "content": "[71] Shengyang Sun, Guodong Zhang, Jiaxin Shi, and Roger Grosse. Functional variational bayesian", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 127, + 371, + 361, + 383 + ], + "spans": [ + { + "bbox": [ + 127, + 371, + 361, + 383 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:1903.05779, 2019.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 106, + 359, + 505, + 383 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 506, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 506, + 401 + ], + "score": 1.0, + "content": "[72] Jakub Swiatkowski, Kevin Roth, Bastiaan S Veeling, Linh Tran, Joshua V Dillon, Stephan", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 127, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Mandt, Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. The k-tied", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 127, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "normal distribution: A compact parameterization of gaussian mean field posteriors in bayesian", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 423, + 360, + 434 + ], + "spans": [ + { + "bbox": [ + 127, + 423, + 360, + 434 + ], + "score": 1.0, + "content": "neural networks. arXiv preprint arXiv:2002.02655, 2020.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 389, + 506, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 440, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "[73] Johannes von Oswald, Christian Henning, João Sacramento, and Benjamin F. Grewe. Continual", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 127, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "learning with hypernetworks. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 461, + 321, + 474 + ], + "spans": [ + { + "bbox": [ + 127, + 461, + 321, + 474 + ], + "score": 1.0, + "content": "URL https://arxiv.org/abs/1906.00695.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 439, + 506, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 480, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 493 + ], + "score": 1.0, + "content": "[74] Maxime Wabartha, Audrey Durand, Vincent François-Lavet, and Joelle Pineau. Handling black", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 126, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "swan events in deep learning with diversely extrapolated neural networks. In IJCAI, pages", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 502, + 202, + 513 + ], + "spans": [ + { + "bbox": [ + 127, + 502, + 202, + 513 + ], + "score": 1.0, + "content": "2140–2147, 2020.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 479, + 506, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 520, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "[75] Yifei Wang, Peng Chen, and Wuchen Li. Projected wasserstein gradient descent for high-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 530, + 423, + 543 + ], + "spans": [ + { + "bbox": [ + 127, + 530, + 423, + 543 + ], + "score": 1.0, + "content": "dimensional bayesian inference. arXiv preprint arXiv:2102.06350, 2021.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 519, + 506, + 543 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 549, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "[76] Ziyu Wang, Tongzheng Ren, Jun Zhu, and Bo Zhang. Function space particle optimization for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 560, + 397, + 572 + ], + "spans": [ + { + "bbox": [ + 127, + 560, + 397, + 572 + ], + "score": 1.0, + "content": "bayesian neural networks. arXiv preprint arXiv:1902.09754, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 549, + 505, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 578, + 505, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "[77] Max Welling and Yee W Teh. Bayesian learning via stochastic gradient langevin dynamics.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 126, + 588, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 126, + 588, + 505, + 603 + ], + "score": 1.0, + "content": "In Proceedings of the 28th international conference on machine learning (ICML-11), pages", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 128, + 600, + 193, + 612 + ], + "spans": [ + { + "bbox": [ + 128, + 600, + 193, + 612 + ], + "score": 1.0, + "content": "681–688, 2011.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 578, + 506, + 612 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 618, + 505, + 641 + ], + "lines": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "[78] Yeming Wen, Dustin Tran, and Jimmy Ba. Batchensemble: an alternative approach to efficient", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 629, + 419, + 642 + ], + "spans": [ + { + "bbox": [ + 127, + 629, + 419, + 642 + ], + "score": 1.0, + "content": "ensemble and lifelong learning. arXiv preprint arXiv:2002.06715, 2020.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 618, + 505, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 648, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "[79] Florian Wenzel, Kevin Roth, Bastiaan S Veeling, Jakub Swiatkowski, Linh Tran, Stephan Mandt, ´", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 127, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 127, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "Jasper Snoek, Tim Salimans, Rodolphe Jenatton, and Sebastian Nowozin. How good is the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 670, + 482, + 683 + ], + "spans": [ + { + "bbox": [ + 127, + 670, + 482, + 683 + ], + "score": 1.0, + "content": "bayes posterior in deep neural networks really? arXiv preprint arXiv:2002.02405, 2020.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 648, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "[80] Florian Wenzel, Jasper Snoek, Dustin Tran, and Rodolphe Jenatton. Hyperparameter ensembles", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 698, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 126, + 698, + 505, + 713 + ], + "score": 1.0, + "content": "for robustness and uncertainty quantification. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 711, + 189, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 189, + 723 + ], + "score": 1.0, + "content": "Systems, 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 687, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 507, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "score": 1.0, + "content": "[81] Christopher KI Williams. Computation with infinite neural networks. Neural Computation, 10", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 84, + 217, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 84, + 217, + 97 + ], + "score": 1.0, + "content": "(5):1203–1216, 1998.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 102, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 504, + 115 + ], + "score": 1.0, + "content": "[82] Andrew Gordon Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 114, + 411, + 127 + ], + "spans": [ + { + "bbox": [ + 126, + 114, + 411, + 127 + ], + "score": 1.0, + "content": "perspective of generalization. arXiv preprint arXiv:2002.08791, 2020.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "[83] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 142, + 470, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 470, + 156 + ], + "score": 1.0, + "content": "benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 161, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 177 + ], + "score": 1.0, + "content": "[84] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 173, + 487, + 187 + ], + "spans": [ + { + "bbox": [ + 127, + 173, + 487, + 187 + ], + "score": 1.0, + "content": "deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "[85] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclical", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 127, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 106, + 72, + 507, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "score": 1.0, + "content": "[81] Christopher KI Williams. Computation with infinite neural networks. Neural Computation, 10", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 84, + 217, + 97 + ], + "spans": [ + { + "bbox": [ + 127, + 84, + 217, + 97 + ], + "score": 1.0, + "content": "(5):1203–1216, 1998.", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 102, + 504, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 504, + 115 + ], + "score": 1.0, + "content": "[82] Andrew Gordon Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 114, + 411, + 127 + ], + "spans": [ + { + "bbox": [ + 126, + 114, + 411, + 127 + ], + "score": 1.0, + "content": "perspective of generalization. arXiv preprint arXiv:2002.08791, 2020.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "[83] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 142, + 470, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 470, + 156 + ], + "score": 1.0, + "content": "benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 161, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 177 + ], + "score": 1.0, + "content": "[84] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 173, + 487, + 187 + ], + "spans": [ + { + "bbox": [ + 127, + 173, + 487, + 187 + ], + "score": 1.0, + "content": "deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "[85] Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen, and Andrew Gordon Wilson. Cyclical", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 127, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "stochastic gradient mcmc for bayesian deep learning. arXiv preprint arXiv:1902.03932, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 72, + 506, + 217 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/LAKplpLMbP8/LAKplpLMbP8_model.json b/parse/train/LAKplpLMbP8/LAKplpLMbP8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..526d16a7ebb52c456111e259c530bb0eab925850 --- /dev/null +++ b/parse/train/LAKplpLMbP8/LAKplpLMbP8_model.json @@ -0,0 +1,19037 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 717, + 1303, + 717, + 1303, + 1262, + 398, + 1262 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1391, + 1404, + 1391, + 1404, + 1725, + 298, + 1725 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1739, + 1404, + 1739, + 1404, + 1983, + 299, + 1983 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 931, + 445, + 1212, + 445, + 1212, + 566, + 931, + 566 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 488, + 444, + 731, + 444, + 731, + 567, + 488, + 567 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 436, + 271, + 1265, + 271, + 1265, + 327, + 436, + 327 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 299, + 1320, + 530, + 1320, + 530, + 1357, + 299, + 1357 + ], + "score": 0.907 + }, + { + "category_id": 0, + "poly": [ + 788, + 645, + 912, + 645, + 912, + 680, + 788, + 680 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 298, + 2033, + 1070, + 2033, + 1070, + 2061, + 298, + 2061 + ], + "score": 0.844 + }, + { + "category_id": 13, + "poly": [ + 548, + 1799, + 597, + 1799, + 597, + 1831, + 548, + 1831 + ], + "score": 0.62, + "latex": "\\begin{array} { r l } { { \\| 6 5 \\| } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1197, + 1767, + 1287, + 1767, + 1287, + 1801, + 1197, + 1801 + ], + "score": 0.54, + "latex": "\\pm 2 \\sqrt { 8 0 }" + }, + { + "category_id": 13, + "poly": [ + 619, + 1571, + 668, + 1571, + 668, + 1604, + 619, + 1604 + ], + "score": 0.4, + "latex": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 268.0, + 1266.0, + 268.0, + 1266.0, + 332.0, + 434.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1316.0, + 536.0, + 1316.0, + 536.0, + 1363.0, + 292.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 645.0, + 918.0, + 645.0, + 918.0, + 683.0, + 784.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2028.0, + 1075.0, + 2028.0, + 1075.0, + 2067.0, + 293.0, + 2067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 715.0, + 1305.0, + 715.0, + 1305.0, + 754.0, + 393.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 744.0, + 1305.0, + 744.0, + 1305.0, + 782.0, + 392.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 774.0, + 1307.0, + 774.0, + 1307.0, + 812.0, + 393.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 808.0, + 1305.0, + 808.0, + 1305.0, + 841.0, + 395.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 839.0, + 1304.0, + 839.0, + 1304.0, + 871.0, + 394.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 868.0, + 1305.0, + 868.0, + 1305.0, + 904.0, + 394.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 898.0, + 1304.0, + 898.0, + 1304.0, + 931.0, + 395.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 929.0, + 1305.0, + 929.0, + 1305.0, + 962.0, + 395.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 957.0, + 1305.0, + 957.0, + 1305.0, + 995.0, + 392.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 991.0, + 1304.0, + 991.0, + 1304.0, + 1020.0, + 396.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1019.0, + 1307.0, + 1019.0, + 1307.0, + 1055.0, + 394.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1049.0, + 1307.0, + 1049.0, + 1307.0, + 1085.0, + 394.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1078.0, + 1308.0, + 1078.0, + 1308.0, + 1116.0, + 393.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1108.0, + 1306.0, + 1108.0, + 1306.0, + 1147.0, + 392.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1139.0, + 1307.0, + 1139.0, + 1307.0, + 1176.0, + 393.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1170.0, + 1306.0, + 1170.0, + 1306.0, + 1207.0, + 393.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1201.0, + 1306.0, + 1201.0, + 1306.0, + 1237.0, + 394.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1232.0, + 706.0, + 1232.0, + 706.0, + 1265.0, + 394.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1425.0, + 294.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1421.0, + 1407.0, + 1421.0, + 1407.0, + 1458.0, + 292.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1449.0, + 1404.0, + 1449.0, + 1404.0, + 1487.0, + 293.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1478.0, + 1404.0, + 1478.0, + 1404.0, + 1520.0, + 293.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1510.0, + 1406.0, + 1510.0, + 1406.0, + 1548.0, + 293.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1541.0, + 1404.0, + 1541.0, + 1404.0, + 1579.0, + 293.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1573.0, + 618.0, + 1573.0, + 618.0, + 1607.0, + 296.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1573.0, + 1405.0, + 1573.0, + 1405.0, + 1607.0, + 669.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1602.0, + 1405.0, + 1602.0, + 1405.0, + 1636.0, + 293.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1404.0, + 1633.0, + 1404.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1663.0, + 1404.0, + 1663.0, + 1404.0, + 1701.0, + 293.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1695.0, + 1088.0, + 1695.0, + 1088.0, + 1729.0, + 294.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1738.0, + 1409.0, + 1738.0, + 1409.0, + 1775.0, + 294.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1768.0, + 1196.0, + 1768.0, + 1196.0, + 1804.0, + 295.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1768.0, + 1406.0, + 1768.0, + 1406.0, + 1804.0, + 1288.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1800.0, + 547.0, + 1800.0, + 547.0, + 1834.0, + 295.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1800.0, + 1406.0, + 1800.0, + 1406.0, + 1834.0, + 598.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1829.0, + 1405.0, + 1829.0, + 1405.0, + 1865.0, + 293.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1860.0, + 1406.0, + 1860.0, + 1406.0, + 1894.0, + 293.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1890.0, + 1409.0, + 1890.0, + 1409.0, + 1927.0, + 294.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1920.0, + 1408.0, + 1920.0, + 1408.0, + 1955.0, + 293.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1953.0, + 1405.0, + 1953.0, + 1405.0, + 1986.0, + 295.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 445.0, + 1169.0, + 445.0, + 1169.0, + 474.0, + 977.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 472.0, + 1144.0, + 472.0, + 1144.0, + 506.0, + 999.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 506.0, + 1186.0, + 506.0, + 1186.0, + 535.0, + 959.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 538.0, + 1213.0, + 538.0, + 1213.0, + 566.0, + 931.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 442.0, + 731.0, + 442.0, + 731.0, + 479.0, + 485.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 471.0, + 681.0, + 471.0, + 681.0, + 506.0, + 536.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 501.0, + 724.0, + 501.0, + 724.0, + 538.0, + 494.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 535.0, + 729.0, + 535.0, + 729.0, + 570.0, + 489.0, + 570.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1683, + 1406, + 1683, + 1406, + 1929, + 297, + 1929 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 827, + 1405, + 827, + 1405, + 1163, + 297, + 1163 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 336, + 189, + 1371, + 189, + 1371, + 554, + 336, + 554 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 371, + 1235, + 1405, + 1235, + 1405, + 1554, + 371, + 1554 + ], + "score": 0.966 + }, + { + "category_id": 4, + "poly": [ + 297, + 566, + 1403, + 566, + 1403, + 658, + 297, + 658 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 300, + 721, + 1398, + 721, + 1398, + 815, + 300, + 815 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 582, + 1948, + 1114, + 1948, + 1114, + 2015, + 582, + 2015 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 299, + 1612, + 728, + 1612, + 728, + 1651, + 299, + 1651 + ], + "score": 0.923 + }, + { + "category_id": 1, + "poly": [ + 300, + 1177, + 717, + 1177, + 717, + 1209, + 300, + 1209 + ], + "score": 0.922 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1967, + 1400, + 1967, + 1400, + 1997, + 1366, + 1997 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.744 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.174 + }, + { + "category_id": 14, + "poly": [ + 585, + 1945, + 1113, + 1945, + 1113, + 2017, + 585, + 2017 + ], + "score": 0.94, + "latex": "p ( \\pmb { y } ^ { * } | \\pmb { x } ^ { * } , \\mathcal { D } ) = \\int p ( \\pmb { y } ^ { * } | f ( \\pmb { x } ^ { * } ; \\mathbf { w } ) ) p ( \\mathbf { w } | \\mathcal { D } ) \\mathrm { d } \\mathbf { w }" + }, + { + "category_id": 13, + "poly": [ + 657, + 1866, + 830, + 1866, + 830, + 1899, + 657, + 1899 + ], + "score": 0.93, + "latex": "{ \\pmb y } ^ { * } = f ( { \\pmb x } ^ { * } ; \\widehat { \\mathbf { w } } )" + }, + { + "category_id": 13, + "poly": [ + 1309, + 1715, + 1403, + 1715, + 1403, + 1748, + 1309, + 1748 + ], + "score": 0.93, + "latex": "f ( \\pmb { x } ; \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 513, + 1744, + 750, + 1744, + 750, + 1778, + 513, + 1778 + ], + "score": 0.92, + "latex": "\\textit { D } = \\{ ( \\boldsymbol { \\mathbf { { x } } } _ { i } , \\boldsymbol { \\mathbf { { y } } } _ { i } ) \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 1112, + 1684, + 1266, + 1684, + 1266, + 1718, + 1112, + 1718 + ], + "score": 0.92, + "latex": "p ( \\pmb { y } | f ( \\pmb { x } ; \\mathbf { w } ) )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1804, + 739, + 1804, + 739, + 1840, + 298, + 1840 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { p ( \\mathbf { w } | \\mathcal { D } ) \\propto \\prod _ { i = 1 } ^ { n } p ( \\pmb { y } _ { i } | f ( \\pmb { x } _ { i } ; \\mathbf { w } ) ) p ( \\mathbf { w } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 990, + 1745, + 1090, + 1745, + 1090, + 1776, + 990, + 1776 + ], + "score": 0.91, + "latex": "\\textbf { \\textit { y } } \\in \\textbf { \\textit { y } }" + }, + { + "category_id": 13, + "poly": [ + 832, + 1805, + 894, + 1805, + 894, + 1838, + 832, + 1838 + ], + "score": 0.91, + "latex": "p ( \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 824, + 1746, + 927, + 1746, + 927, + 1774, + 824, + 1774 + ], + "score": 0.91, + "latex": "\\textbf { \\textit { x } } \\in \\textbf { \\textit { X } }" + }, + { + "category_id": 13, + "poly": [ + 844, + 1839, + 877, + 1839, + 877, + 1864, + 844, + 1864 + ], + "score": 0.86, + "latex": "\\mathbf { \\boldsymbol { x } } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1039, + 1081, + 1039, + 1081, + 1072, + 1032, + 1072 + ], + "score": 0.79, + "latex": "\\mathbb { \\left[ \\left] \\right. \\right]} 2 " + }, + { + "category_id": 13, + "poly": [ + 511, + 1867, + 538, + 1867, + 538, + 1894, + 511, + 1894 + ], + "score": 0.76, + "latex": "\\widehat { \\bf w }" + }, + { + "category_id": 13, + "poly": [ + 1060, + 918, + 1095, + 918, + 1095, + 954, + 1060, + 954 + ], + "score": 0.42, + "latex": "\\boxed { \\mathrm { A } }" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 196.0, + 513.0, + 196.0, + 513.0, + 220.0, + 369.0, + 220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 200.0, + 684.0, + 200.0, + 684.0, + 223.0, + 605.0, + 223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 200.0, + 895.0, + 200.0, + 895.0, + 223.0, + 805.0, + 223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 196.0, + 1305.0, + 196.0, + 1305.0, + 220.0, + 1213.0, + 220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 258.0, + 819.0, + 258.0, + 819.0, + 270.0, + 794.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 259.0, + 1024.0, + 259.0, + 1024.0, + 267.0, + 1000.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 349.0, + 1383.0, + 349.0, + 1383.0, + 574.0, + 314.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 398.0, + 534.0, + 398.0, + 534.0, + 428.0, + 481.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.75, + 196.0, + 1104.75, + 196.0, + 1104.75, + 220.0, + 1005.75, + 220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 565.0, + 1405.0, + 565.0, + 1405.0, + 601.0, + 293.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 596.0, + 1404.0, + 596.0, + 1404.0, + 634.0, + 293.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 625.0, + 1032.0, + 625.0, + 1032.0, + 664.0, + 292.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1606.0, + 733.0, + 1606.0, + 733.0, + 1661.0, + 289.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1682.0, + 1111.0, + 1682.0, + 1111.0, + 1719.0, + 294.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1682.0, + 1409.0, + 1682.0, + 1409.0, + 1719.0, + 1267.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1715.0, + 1308.0, + 1715.0, + 1308.0, + 1748.0, + 295.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1740.0, + 512.0, + 1740.0, + 512.0, + 1784.0, + 291.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1740.0, + 823.0, + 1740.0, + 823.0, + 1784.0, + 751.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1740.0, + 989.0, + 1740.0, + 989.0, + 1784.0, + 928.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1740.0, + 1412.0, + 1740.0, + 1412.0, + 1784.0, + 1091.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1811.0, + 291.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1805.0, + 297.0, + 1805.0, + 297.0, + 1842.0, + 292.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1805.0, + 831.0, + 1805.0, + 831.0, + 1842.0, + 740.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1805.0, + 1409.0, + 1805.0, + 1409.0, + 1842.0, + 895.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 843.0, + 1835.0, + 843.0, + 1872.0, + 294.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1835.0, + 1407.0, + 1835.0, + 1407.0, + 1872.0, + 878.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1866.0, + 510.0, + 1866.0, + 510.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1866.0, + 656.0, + 1866.0, + 656.0, + 1900.0, + 539.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 1866.0, + 1404.0, + 1866.0, + 1404.0, + 1900.0, + 831.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1897.0, + 926.0, + 1897.0, + 926.0, + 1931.0, + 296.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 829.0, + 1406.0, + 829.0, + 1406.0, + 864.0, + 294.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 859.0, + 1403.0, + 859.0, + 1403.0, + 893.0, + 294.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 890.0, + 1406.0, + 890.0, + 1406.0, + 924.0, + 294.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 919.0, + 1059.0, + 919.0, + 1059.0, + 954.0, + 294.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 919.0, + 1406.0, + 919.0, + 1406.0, + 954.0, + 1096.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 949.0, + 1405.0, + 949.0, + 1405.0, + 984.0, + 295.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 981.0, + 1406.0, + 981.0, + 1406.0, + 1016.0, + 295.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1011.0, + 1406.0, + 1011.0, + 1406.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1041.0, + 1031.0, + 1041.0, + 1031.0, + 1074.0, + 292.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1041.0, + 1406.0, + 1041.0, + 1406.0, + 1074.0, + 1082.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1071.0, + 1403.0, + 1071.0, + 1403.0, + 1105.0, + 294.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1100.0, + 1405.0, + 1100.0, + 1405.0, + 1137.0, + 292.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1133.0, + 685.0, + 1133.0, + 685.0, + 1166.0, + 293.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1231.0, + 1405.0, + 1231.0, + 1405.0, + 1272.0, + 368.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1267.0, + 1303.0, + 1267.0, + 1303.0, + 1300.0, + 395.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1306.0, + 1404.0, + 1306.0, + 1404.0, + 1347.0, + 368.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1343.0, + 990.0, + 1343.0, + 990.0, + 1377.0, + 394.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1383.0, + 1404.0, + 1383.0, + 1404.0, + 1421.0, + 368.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1416.0, + 842.0, + 1416.0, + 842.0, + 1452.0, + 394.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1460.0, + 1404.0, + 1460.0, + 1404.0, + 1496.0, + 369.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1493.0, + 1404.0, + 1493.0, + 1404.0, + 1527.0, + 395.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1523.0, + 655.0, + 1523.0, + 655.0, + 1556.0, + 395.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 722.0, + 1404.0, + 722.0, + 1404.0, + 756.0, + 295.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 750.0, + 1404.0, + 750.0, + 1404.0, + 789.0, + 294.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 783.0, + 960.0, + 783.0, + 960.0, + 819.0, + 294.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1174.0, + 720.0, + 1174.0, + 720.0, + 1213.0, + 296.0, + 1213.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 497, + 1405, + 497, + 1405, + 805, + 297, + 805 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 817, + 1406, + 817, + 1406, + 1124, + 298, + 1124 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1588, + 1406, + 1588, + 1406, + 1774, + 297, + 1774 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1282, + 1408, + 1282, + 1408, + 1499, + 298, + 1499 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 202, + 1406, + 202, + 1406, + 387, + 296, + 387 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 361, + 1145, + 1220, + 1145, + 1220, + 1263, + 361, + 1263 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 612, + 1953, + 1088, + 1953, + 1088, + 2017, + 612, + 2017 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 296, + 1872, + 1405, + 1872, + 1405, + 1940, + 296, + 1940 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 511, + 1788, + 1181, + 1788, + 1181, + 1859, + 511, + 1859 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 635, + 399, + 1063, + 399, + 1063, + 486, + 635, + 486 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 300, + 1533, + 732, + 1533, + 732, + 1567, + 300, + 1567 + ], + "score": 0.924 + }, + { + "category_id": 9, + "poly": [ + 1366, + 450, + 1400, + 450, + 1400, + 479, + 1366, + 479 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1808, + 1400, + 1808, + 1400, + 1838, + 1366, + 1838 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.669 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.485 + }, + { + "category_id": 14, + "poly": [ + 611, + 1951, + 1087, + 1951, + 1087, + 2017, + 611, + 2017 + ], + "score": 0.94, + "latex": "\\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) = \\frac { 2 } { h } ( \\mathbf { w } _ { j } ^ { t } - \\mathbf { w } _ { i } ^ { t } ) k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } )" + }, + { + "category_id": 13, + "poly": [ + 944, + 324, + 1043, + 324, + 1043, + 358, + 944, + 358 + ], + "score": 0.93, + "latex": "\\{ { \\bf w } _ { i } \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 14, + "poly": [ + 511, + 1785, + 1189, + 1785, + 1189, + 1862, + 511, + 1862 + ], + "score": 0.92, + "latex": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) ." + }, + { + "category_id": 13, + "poly": [ + 374, + 1902, + 797, + 1902, + 797, + 1940, + 374, + 1940 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { k ( \\mathbf { w } _ { i } , \\mathbf { w } _ { j } ) = \\exp \\big ( - \\frac { 1 } { h } | | \\mathbf { w } _ { i } - \\mathbf { w } _ { j } | | ^ { 2 } \\big ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1104, + 323, + 1206, + 323, + 1206, + 355, + 1104, + 355 + ], + "score": 0.92, + "latex": "{ \\bf w } _ { i } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 659, + 1710, + 914, + 1710, + 914, + 1745, + 659, + 1745 + ], + "score": 0.91, + "latex": "\\bar { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\to \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 991, + 1909, + 1008, + 1909, + 1008, + 1931, + 991, + 1931 + ], + "score": 0.87, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 817, + 296, + 846, + 296, + 846, + 322, + 817, + 322 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 14, + "poly": [ + 635, + 397, + 1064, + 397, + 1064, + 492, + 635, + 492 + ], + "score": 0.83, + "latex": "\\begin{array} { r l } & { \\mathbf { w } _ { i } ^ { t + 1 } \\mathbf { w } _ { i } ^ { t } + \\epsilon _ { t } \\phi ( \\mathbf { w } _ { i } ^ { t } ) } \\\\ { \\mathrm { w i t h } \\quad } & { \\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 586, + 1743, + 613, + 1743, + 613, + 1769, + 586, + 1769 + ], + "score": 0.82, + "latex": "\\mathcal { R }" + }, + { + "category_id": 13, + "poly": [ + 1352, + 621, + 1400, + 621, + 1400, + 653, + 1352, + 653 + ], + "score": 0.82, + "latex": "\\pmb { \\Vert 8 2 \\Vert }" + }, + { + "category_id": 13, + "poly": [ + 453, + 505, + 477, + 505, + 477, + 531, + 453, + 531 + ], + "score": 0.82, + "latex": "\\epsilon _ { t }" + }, + { + "category_id": 13, + "poly": [ + 524, + 529, + 573, + 529, + 573, + 562, + 524, + 562 + ], + "score": 0.78, + "latex": "\\pm 2 \\|" + }, + { + "category_id": 13, + "poly": [ + 1272, + 743, + 1320, + 743, + 1320, + 774, + 1272, + 774 + ], + "score": 0.78, + "latex": "\\mathbb { [ \\breve { ] { ] \\mathrm { ~ Z ~ Z ~ } ] } } }" + }, + { + "category_id": 13, + "poly": [ + 1133, + 711, + 1182, + 711, + 1182, + 743, + 1133, + 743 + ], + "score": 0.77, + "latex": "[ \\overline { { 1 8 } } ]" + }, + { + "category_id": 13, + "poly": [ + 1327, + 358, + 1339, + 358, + 1339, + 382, + 1327, + 382 + ], + "score": 0.77, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 752, + 590, + 800, + 590, + 800, + 622, + 752, + 622 + ], + "score": 0.75, + "latex": "\\mathbb { \\lVert \\rVert }" + }, + { + "category_id": 13, + "poly": [ + 1074, + 938, + 1123, + 938, + 1123, + 970, + 1074, + 970 + ], + "score": 0.69, + "latex": " { \\mathbb { I } } { \\mathbb { I } }" + }, + { + "category_id": 13, + "poly": [ + 548, + 590, + 582, + 590, + 582, + 626, + 548, + 626 + ], + "score": 0.69, + "latex": "\\mathbb { D }" + }, + { + "category_id": 13, + "poly": [ + 442, + 330, + 462, + 330, + 462, + 352, + 442, + 352 + ], + "score": 0.61, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 770, + 1283, + 820, + 1283, + 820, + 1316, + 770, + 1316 + ], + "score": 0.57, + "latex": "\\pmb { \\left. \\pmb { \\left. \\bar { 2 5 } \\right. } \\right. }" + }, + { + "category_id": 13, + "poly": [ + 1162, + 968, + 1211, + 968, + 1211, + 1001, + 1162, + 1001 + ], + "score": 0.56, + "latex": "\\overline { { \\| 6 3 \\| } }" + }, + { + "category_id": 13, + "poly": [ + 419, + 1619, + 468, + 1619, + 468, + 1652, + 419, + 1652 + ], + "score": 0.48, + "latex": "\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }" + }, + { + "category_id": 13, + "poly": [ + 681, + 938, + 730, + 938, + 730, + 971, + 681, + 971 + ], + "score": 0.31, + "latex": "\\pmb { \\mathbb { B 3 } }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1527.0, + 735.0, + 1527.0, + 735.0, + 1576.0, + 292.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 498.0, + 452.0, + 498.0, + 452.0, + 538.0, + 294.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 498.0, + 1407.0, + 498.0, + 1407.0, + 538.0, + 478.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 526.0, + 523.0, + 526.0, + 523.0, + 568.0, + 294.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 526.0, + 1406.0, + 526.0, + 1406.0, + 568.0, + 574.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 562.0, + 1403.0, + 562.0, + 1403.0, + 594.0, + 296.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 587.0, + 547.0, + 587.0, + 547.0, + 629.0, + 292.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 587.0, + 751.0, + 587.0, + 751.0, + 629.0, + 583.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 587.0, + 1407.0, + 587.0, + 1407.0, + 629.0, + 801.0, + 629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 620.0, + 1351.0, + 620.0, + 1351.0, + 656.0, + 295.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 648.0, + 1410.0, + 648.0, + 1410.0, + 691.0, + 292.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 679.0, + 1406.0, + 679.0, + 1406.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 711.0, + 1132.0, + 711.0, + 1132.0, + 747.0, + 295.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 711.0, + 1405.0, + 711.0, + 1405.0, + 747.0, + 1183.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 742.0, + 1271.0, + 742.0, + 1271.0, + 776.0, + 294.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1321.0, + 742.0, + 1405.0, + 742.0, + 1405.0, + 776.0, + 1321.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 771.0, + 1391.0, + 771.0, + 1391.0, + 808.0, + 292.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 817.0, + 1408.0, + 817.0, + 1408.0, + 854.0, + 293.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 850.0, + 1404.0, + 850.0, + 1404.0, + 881.0, + 296.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 878.0, + 1406.0, + 878.0, + 1406.0, + 913.0, + 295.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 908.0, + 1406.0, + 908.0, + 1406.0, + 944.0, + 295.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 939.0, + 680.0, + 939.0, + 680.0, + 974.0, + 293.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 939.0, + 1073.0, + 939.0, + 1073.0, + 974.0, + 731.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 939.0, + 1408.0, + 939.0, + 1408.0, + 974.0, + 1124.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 967.0, + 1161.0, + 967.0, + 1161.0, + 1004.0, + 293.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 967.0, + 1406.0, + 967.0, + 1406.0, + 1004.0, + 1212.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1002.0, + 1404.0, + 1002.0, + 1404.0, + 1034.0, + 296.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 1404.0, + 1030.0, + 1404.0, + 1066.0, + 295.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1061.0, + 1406.0, + 1061.0, + 1406.0, + 1096.0, + 295.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 521.0, + 1089.0, + 521.0, + 1128.0, + 295.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1587.0, + 1406.0, + 1587.0, + 1406.0, + 1624.0, + 292.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1617.0, + 418.0, + 1617.0, + 418.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1617.0, + 1406.0, + 1617.0, + 1406.0, + 1655.0, + 469.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1652.0, + 1406.0, + 1652.0, + 1406.0, + 1684.0, + 295.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1678.0, + 1407.0, + 1678.0, + 1407.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1705.0, + 658.0, + 1705.0, + 658.0, + 1749.0, + 292.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1705.0, + 1408.0, + 1705.0, + 1408.0, + 1749.0, + 915.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 585.0, + 1741.0, + 585.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1741.0, + 1090.0, + 1741.0, + 1090.0, + 1777.0, + 614.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1282.0, + 769.0, + 1282.0, + 769.0, + 1320.0, + 295.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1282.0, + 1405.0, + 1282.0, + 1405.0, + 1320.0, + 821.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1315.0, + 1405.0, + 1315.0, + 1405.0, + 1351.0, + 293.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1347.0, + 1404.0, + 1347.0, + 1404.0, + 1378.0, + 296.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1374.0, + 1405.0, + 1374.0, + 1405.0, + 1410.0, + 293.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1404.0, + 1406.0, + 1404.0, + 1406.0, + 1441.0, + 293.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1436.0, + 1410.0, + 1436.0, + 1410.0, + 1471.0, + 293.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1465.0, + 1041.0, + 1465.0, + 1041.0, + 1502.0, + 292.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 199.0, + 1407.0, + 199.0, + 1407.0, + 242.0, + 291.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 233.0, + 1406.0, + 233.0, + 1406.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 264.0, + 1406.0, + 264.0, + 1406.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 293.0, + 816.0, + 293.0, + 816.0, + 329.0, + 293.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 293.0, + 1409.0, + 293.0, + 1409.0, + 329.0, + 847.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 317.0, + 441.0, + 317.0, + 441.0, + 366.0, + 290.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 317.0, + 943.0, + 317.0, + 943.0, + 366.0, + 463.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 317.0, + 1103.0, + 317.0, + 1103.0, + 366.0, + 1044.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 317.0, + 1411.0, + 317.0, + 1411.0, + 366.0, + 1207.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 354.0, + 1326.0, + 354.0, + 1326.0, + 390.0, + 295.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 354.0, + 1353.0, + 354.0, + 1353.0, + 390.0, + 1340.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1146.0, + 793.0, + 1146.0, + 793.0, + 1178.0, + 363.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1184.0, + 1103.0, + 1184.0, + 1103.0, + 1224.0, + 358.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1226.0, + 1221.0, + 1226.0, + 1221.0, + 1265.0, + 358.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1873.0, + 1405.0, + 1873.0, + 1405.0, + 1906.0, + 296.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1901.0, + 373.0, + 1901.0, + 373.0, + 1941.0, + 294.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1901.0, + 990.0, + 1901.0, + 990.0, + 1941.0, + 798.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1901.0, + 1320.0, + 1901.0, + 1320.0, + 1941.0, + 1009.0, + 1941.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 481, + 1407, + 481, + 1407, + 791, + 296, + 791 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 201, + 1406, + 201, + 1406, + 394, + 296, + 394 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1627, + 1406, + 1627, + 1406, + 1751, + 297, + 1751 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1117, + 1404, + 1117, + 1404, + 1274, + 298, + 1274 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1883, + 1403, + 1883, + 1403, + 2010, + 298, + 2010 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 305, + 931, + 1400, + 931, + 1400, + 1025, + 305, + 1025 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 481, + 1524, + 1214, + 1524, + 1214, + 1614, + 481, + 1614 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 298, + 1031, + 1399, + 1031, + 1399, + 1100, + 298, + 1100 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 425, + 1763, + 1268, + 1763, + 1268, + 1855, + 425, + 1855 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 390, + 1309, + 1307, + 1309, + 1307, + 1391, + 390, + 1391 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 297, + 1421, + 946, + 1421, + 946, + 1456, + 297, + 1456 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 300, + 427, + 751, + 427, + 751, + 461, + 300, + 461 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 295, + 1474, + 1354, + 1474, + 1354, + 1511, + 295, + 1511 + ], + "score": 0.925 + }, + { + "category_id": 8, + "poly": [ + 470, + 803, + 1220, + 803, + 1220, + 919, + 470, + 919 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1551, + 1400, + 1551, + 1400, + 1581, + 1366, + 1581 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 869, + 1400, + 869, + 1400, + 898, + 1366, + 898 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1795, + 1400, + 1795, + 1400, + 1825, + 1366, + 1825 + ], + "score": 0.87 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1338, + 1401, + 1338, + 1401, + 1368, + 1365, + 1368 + ], + "score": 0.842 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2084, + 840, + 2084 + ], + "score": 0.778 + }, + { + "category_id": 8, + "poly": [ + 588, + 802, + 844, + 802, + 844, + 844, + 588, + 844 + ], + "score": 0.264 + }, + { + "category_id": 14, + "poly": [ + 482, + 1521, + 1217, + 1521, + 1217, + 1617, + 482, + 1617 + ], + "score": 0.94, + "latex": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\sum _ { j = 1 } ^ { n } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\nabla _ { \\mathbf { w } _ { i } ^ { t } } \\log p ( \\mathbf { w } _ { i } ^ { t } | \\mathcal { D } ) + \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { j } ^ { t } } k ( \\mathbf { w } _ { j } ^ { t } , \\mathbf { w } _ { i } ^ { t } ) ." + }, + { + "category_id": 14, + "poly": [ + 429, + 1761, + 1270, + 1761, + 1270, + 1858, + 429, + 1858 + ], + "score": 0.94, + "latex": "\\phi ( \\mathbf { w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial \\mathbf { w } _ { i } ^ { t } } \\right) ^ { \\top } \\left( \\frac { 1 } { n } \\sum _ { j = 1 } ^ { n } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\nabla _ { f _ { j } ^ { t } } \\log p ( f _ { j } ^ { t } | \\mathcal { D } ) + \\nabla _ { f _ { j } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right) ." + }, + { + "category_id": 13, + "poly": [ + 771, + 515, + 967, + 515, + 967, + 548, + 771, + 548 + ], + "score": 0.93, + "latex": "f : \\mathbf { w } \\mapsto f ( \\cdot ; \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 590, + 1064, + 826, + 1064, + 826, + 1099, + 590, + 1099 + ], + "score": 0.93, + "latex": "\\pi _ { A } ( f ) = \\{ f ( a ) \\} _ { a \\in A }" + }, + { + "category_id": 14, + "poly": [ + 392, + 1306, + 1307, + 1306, + 1307, + 1391, + 392, + 1391 + ], + "score": 0.93, + "latex": "\\phi ( { \\mathbf w } _ { i } ^ { t } ) = \\left( \\frac { \\partial f _ { i } ^ { t } } { \\partial { \\mathbf w } _ { i } ^ { t } } \\right) ^ { \\top } \\left[ \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | { \\mathcal D } ) - { \\mathcal R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( \\pi _ { B } ( f _ { i } ^ { t } ) , \\pi _ { B } ( f _ { j } ^ { t } ) ) \\right\\} _ { j = 1 } ^ { n } \\right) \\right] ." + }, + { + "category_id": 13, + "poly": [ + 1078, + 262, + 1345, + 262, + 1345, + 302, + 1078, + 302 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\beta \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 367, + 608, + 458, + 608, + 458, + 642, + 367, + 642 + ], + "score": 0.92, + "latex": "\\{ f _ { i } \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 755, + 729, + 841, + 729, + 841, + 763, + 755, + 763 + ], + "score": 0.92, + "latex": "p ( \\pmb { f } | \\mathcal { D } )" + }, + { + "category_id": 13, + "poly": [ + 1237, + 547, + 1403, + 547, + 1403, + 581, + 1237, + 581 + ], + "score": 0.92, + "latex": "f _ { i } : = f ( \\cdot ; { \\mathbf w } _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 860, + 639, + 971, + 639, + 971, + 671, + 860, + 671 + ], + "score": 0.92, + "latex": "p ( \\pmb { y } | \\pmb { x } , \\pmb { f } )" + }, + { + "category_id": 13, + "poly": [ + 520, + 609, + 598, + 609, + 598, + 640, + 520, + 640 + ], + "score": 0.92, + "latex": "\\pmb { f } \\in \\mathcal { F }" + }, + { + "category_id": 14, + "poly": [ + 506, + 798, + 1227, + 798, + 1227, + 923, + 506, + 923 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } & { f _ { i } ^ { t + 1 } \\gets f _ { i } ^ { t } + \\epsilon _ { t } \\phi ( f _ { i } ^ { t } ) } \\\\ { \\mathrm { t h } \\quad } & { \\phi ( f _ { i } ^ { t } ) = \\nabla _ { f _ { i } ^ { t } } \\log p ( f _ { i } ^ { t } | \\mathcal { D } ) - \\mathcal { R } \\left( \\left\\{ \\nabla _ { f _ { i } ^ { t } } k ( f _ { i } ^ { t } , f _ { j } ^ { t } ) \\right\\} _ { j = 1 } ^ { n } \\right) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1331, + 609, + 1398, + 609, + 1398, + 641, + 1331, + 641 + ], + "score": 0.91, + "latex": "k ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1024, + 1030, + 1207, + 1030, + 1207, + 1064, + 1024, + 1064 + ], + "score": 0.91, + "latex": "\\pi _ { A } : \\mathbb { R } ^ { \\mathcal { X } } \\to \\mathbb { R } ^ { A }" + }, + { + "category_id": 13, + "poly": [ + 1165, + 699, + 1226, + 699, + 1226, + 733, + 1165, + 733 + ], + "score": 0.91, + "latex": "p ( \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 298, + 389, + 298, + 389, + 333, + 298, + 333 + ], + "score": 0.91, + "latex": "\\beta \\in \\mathbb { R } _ { * } ^ { + }" + }, + { + "category_id": 13, + "poly": [ + 853, + 671, + 909, + 671, + 909, + 702, + 853, + 702 + ], + "score": 0.91, + "latex": "p ( f )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1149, + 507, + 1149, + 507, + 1186, + 298, + 1186 + ], + "score": 0.91, + "latex": "k \\big ( \\pi _ { B } ( f ) , \\pi _ { B } ( f ^ { \\prime } ) \\big )" + }, + { + "category_id": 13, + "poly": [ + 1287, + 640, + 1403, + 640, + 1403, + 672, + 1287, + 672 + ], + "score": 0.91, + "latex": "p ( \\pmb { y } | \\pmb { x } , \\mathbf { w } )" + }, + { + "category_id": 13, + "poly": [ + 386, + 545, + 479, + 545, + 479, + 576, + 386, + 576 + ], + "score": 0.9, + "latex": "\\mathbf { w } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 829, + 208, + 867, + 208, + 867, + 237, + 829, + 237 + ], + "score": 0.88, + "latex": "{ \\bf w } _ { j }" + }, + { + "category_id": 13, + "poly": [ + 814, + 1033, + 898, + 1033, + 898, + 1062, + 814, + 1062 + ], + "score": 0.88, + "latex": "A \\subset { \\mathcal { X } }" + }, + { + "category_id": 13, + "poly": [ + 866, + 582, + 901, + 582, + 901, + 609, + 866, + 609 + ], + "score": 0.87, + "latex": "\\mathbf { w } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 372, + 208, + 408, + 208, + 408, + 234, + 372, + 234 + ], + "score": 0.86, + "latex": "\\mathbf { w } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 872, + 301, + 893, + 301, + 893, + 332, + 872, + 332 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1169, + 332, + 1196, + 332, + 1196, + 358, + 1169, + 358 + ], + "score": 0.84, + "latex": "\\mathcal { R }" + }, + { + "category_id": 13, + "poly": [ + 1141, + 483, + 1176, + 483, + 1176, + 519, + 1141, + 519 + ], + "score": 0.83, + "latex": "\\textcircled { 3 }" + }, + { + "category_id": 13, + "poly": [ + 497, + 235, + 524, + 235, + 524, + 261, + 497, + 261 + ], + "score": 0.83, + "latex": "\\mathcal { R }" + }, + { + "category_id": 13, + "poly": [ + 573, + 1151, + 599, + 1151, + 599, + 1178, + 573, + 1178 + ], + "score": 0.81, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 806, + 1977, + 841, + 1977, + 841, + 2011, + 806, + 2011 + ], + "score": 0.8, + "latex": "\\textcircled{7}" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1914, + 1242, + 1914, + 1242, + 1950, + 1207, + 1950 + ], + "score": 0.79, + "latex": "\\textcircled{6}" + }, + { + "category_id": 13, + "poly": [ + 648, + 296, + 674, + 296, + 674, + 337, + 648, + 337 + ], + "score": 0.79, + "latex": "\\begin{array} { l } { 3 } \\\\ { . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1027, + 1243, + 1039, + 1243, + 1039, + 1268, + 1027, + 1268 + ], + "score": 0.76, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1155, + 584, + 1175, + 584, + 1175, + 606, + 1155, + 606 + ], + "score": 0.75, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 473, + 1066, + 496, + 1066, + 496, + 1092, + 473, + 1092 + ], + "score": 0.69, + "latex": "A" + }, + { + "category_id": 13, + "poly": [ + 823, + 1719, + 872, + 1719, + 872, + 1749, + 823, + 1749 + ], + "score": 0.43, + "latex": "\\pmb { \\mathbb { Z } } 6 \\|" + }, + { + "category_id": 13, + "poly": [ + 806, + 1481, + 852, + 1481, + 852, + 1508, + 806, + 1508 + ], + "score": 0.34, + "latex": "\\mathbb { \\left[ 5 1 \\right] }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1418.0, + 948.0, + 1418.0, + 948.0, + 1461.0, + 292.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 424.0, + 754.0, + 424.0, + 754.0, + 467.0, + 294.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 863.0, + 2058.0, + 863.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 484.0, + 1140.0, + 484.0, + 1140.0, + 519.0, + 295.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 484.0, + 1405.0, + 484.0, + 1405.0, + 519.0, + 1177.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 516.0, + 770.0, + 516.0, + 770.0, + 548.0, + 296.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 516.0, + 1405.0, + 516.0, + 1405.0, + 548.0, + 968.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 545.0, + 385.0, + 545.0, + 385.0, + 583.0, + 293.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 545.0, + 1236.0, + 545.0, + 1236.0, + 583.0, + 480.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 545.0, + 1407.0, + 545.0, + 1407.0, + 583.0, + 1404.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 576.0, + 865.0, + 576.0, + 865.0, + 613.0, + 294.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 576.0, + 1154.0, + 576.0, + 1154.0, + 613.0, + 902.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 576.0, + 1407.0, + 576.0, + 1407.0, + 613.0, + 1176.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 601.0, + 366.0, + 601.0, + 366.0, + 649.0, + 290.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 601.0, + 519.0, + 601.0, + 519.0, + 649.0, + 459.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 601.0, + 1330.0, + 601.0, + 1330.0, + 649.0, + 599.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 601.0, + 1412.0, + 601.0, + 1412.0, + 649.0, + 1399.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 636.0, + 859.0, + 636.0, + 859.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 636.0, + 1286.0, + 636.0, + 1286.0, + 674.0, + 972.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 636.0, + 1407.0, + 636.0, + 1407.0, + 674.0, + 1404.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 663.0, + 852.0, + 663.0, + 852.0, + 708.0, + 292.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 663.0, + 1409.0, + 663.0, + 1409.0, + 708.0, + 910.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 695.0, + 1164.0, + 695.0, + 1164.0, + 739.0, + 293.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 695.0, + 1407.0, + 695.0, + 1407.0, + 739.0, + 1227.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 727.0, + 754.0, + 727.0, + 754.0, + 765.0, + 293.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 727.0, + 1407.0, + 727.0, + 1407.0, + 765.0, + 842.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 757.0, + 420.0, + 757.0, + 420.0, + 794.0, + 293.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 202.0, + 371.0, + 202.0, + 371.0, + 239.0, + 295.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 202.0, + 828.0, + 202.0, + 828.0, + 239.0, + 409.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 202.0, + 1406.0, + 202.0, + 1406.0, + 239.0, + 868.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 233.0, + 496.0, + 233.0, + 496.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 233.0, + 1404.0, + 233.0, + 1404.0, + 266.0, + 525.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 252.0, + 1077.0, + 252.0, + 1077.0, + 314.0, + 282.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 252.0, + 1415.0, + 252.0, + 1415.0, + 314.0, + 1346.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 297.0, + 297.0, + 297.0, + 297.0, + 337.0, + 294.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 297.0, + 647.0, + 297.0, + 647.0, + 337.0, + 390.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 297.0, + 871.0, + 297.0, + 871.0, + 337.0, + 675.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 297.0, + 1404.0, + 297.0, + 1404.0, + 337.0, + 894.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 329.0, + 1168.0, + 329.0, + 1168.0, + 366.0, + 293.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 329.0, + 1406.0, + 329.0, + 1406.0, + 366.0, + 1197.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 364.0, + 450.0, + 364.0, + 450.0, + 398.0, + 295.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1627.0, + 1404.0, + 1627.0, + 1404.0, + 1660.0, + 292.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1656.0, + 1407.0, + 1656.0, + 1407.0, + 1693.0, + 294.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1684.0, + 1406.0, + 1684.0, + 1406.0, + 1724.0, + 296.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1719.0, + 822.0, + 1719.0, + 822.0, + 1752.0, + 295.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1719.0, + 1139.0, + 1719.0, + 1139.0, + 1752.0, + 873.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1117.0, + 1404.0, + 1117.0, + 1404.0, + 1153.0, + 295.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1149.0, + 297.0, + 1149.0, + 297.0, + 1186.0, + 294.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1149.0, + 572.0, + 1149.0, + 572.0, + 1186.0, + 508.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1149.0, + 1408.0, + 1149.0, + 1408.0, + 1186.0, + 600.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 1408.0, + 1179.0, + 1408.0, + 1217.0, + 294.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1210.0, + 1406.0, + 1210.0, + 1406.0, + 1245.0, + 293.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1238.0, + 1026.0, + 1238.0, + 1026.0, + 1277.0, + 293.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 1238.0, + 1327.0, + 1238.0, + 1327.0, + 1277.0, + 1040.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1919.0, + 292.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1206.0, + 1911.0, + 1206.0, + 1950.0, + 293.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1950.0, + 1243.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1983.0, + 292.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 805.0, + 1975.0, + 805.0, + 2012.0, + 294.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1975.0, + 1404.0, + 1975.0, + 1404.0, + 2012.0, + 842.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 931.0, + 1405.0, + 931.0, + 1405.0, + 968.0, + 299.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 965.0, + 1403.0, + 965.0, + 1403.0, + 995.0, + 301.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 994.0, + 1120.0, + 994.0, + 1120.0, + 1030.0, + 299.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1026.0, + 813.0, + 1026.0, + 813.0, + 1071.0, + 293.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1026.0, + 1023.0, + 1026.0, + 1023.0, + 1071.0, + 899.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1026.0, + 1408.0, + 1026.0, + 1408.0, + 1071.0, + 1208.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1060.0, + 472.0, + 1060.0, + 472.0, + 1104.0, + 292.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1060.0, + 589.0, + 1060.0, + 589.0, + 1104.0, + 497.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1060.0, + 839.0, + 1060.0, + 839.0, + 1104.0, + 827.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1473.0, + 805.0, + 1473.0, + 805.0, + 1515.0, + 292.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1473.0, + 1357.0, + 1473.0, + 1357.0, + 1515.0, + 853.0, + 1515.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 433, + 1406, + 433, + 1406, + 558, + 297, + 558 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1027, + 1405, + 1027, + 1405, + 1152, + 297, + 1152 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 815, + 1405, + 815, + 1405, + 941, + 297, + 941 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 202, + 1405, + 202, + 1405, + 327, + 298, + 327 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1230, + 1405, + 1230, + 1405, + 1322, + 297, + 1322 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 580, + 1777, + 1119, + 1777, + 1119, + 1929, + 580, + 1929 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 701, + 722, + 996, + 722, + 996, + 807, + 701, + 807 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 295, + 647, + 1402, + 647, + 1402, + 709, + 295, + 709 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 296, + 1557, + 1401, + 1557, + 1401, + 1622, + 296, + 1622 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 605, + 1480, + 1093, + 1480, + 1093, + 1548, + 605, + 1548 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 305, + 1407, + 1397, + 1407, + 1397, + 1470, + 305, + 1470 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 444, + 1326, + 1253, + 1326, + 1253, + 1401, + 444, + 1401 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 754, + 1162, + 942, + 1162, + 942, + 1225, + 754, + 1225 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 299, + 1707, + 1405, + 1707, + 1405, + 1770, + 299, + 1770 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 299, + 366, + 906, + 366, + 906, + 405, + 299, + 405 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 520, + 1632, + 1177, + 1632, + 1177, + 1701, + 520, + 1701 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 299, + 590, + 638, + 590, + 638, + 624, + 299, + 624 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 301, + 974, + 759, + 974, + 759, + 1007, + 301, + 1007 + ], + "score": 0.926 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1497, + 1400, + 1497, + 1400, + 1527, + 1352, + 1527 + ], + "score": 0.904 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1180, + 1400, + 1180, + 1400, + 1209, + 1366, + 1209 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1837, + 1401, + 1837, + 1401, + 1868, + 1351, + 1868 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1366, + 745, + 1400, + 745, + 1400, + 775, + 1366, + 775 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 330, + 1946, + 1060, + 1946, + 1060, + 2008, + 330, + 2008 + ], + "score": 0.818 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.763 + }, + { + "category_id": 2, + "poly": [ + 329, + 1946, + 1059, + 1946, + 1059, + 2008, + 329, + 2008 + ], + "score": 0.232 + }, + { + "category_id": 14, + "poly": [ + 443, + 1323, + 1254, + 1323, + 1254, + 1401, + 443, + 1401 + ], + "score": 0.96, + "latex": "\\mathcal { P } _ { 2 } ( \\mathcal { M } ) = \\left\\{ \\varphi : \\mathcal { M } \\to [ 0 , \\infty ) \\bigg | \\int _ { \\mathcal { M } } \\mathrm { d } \\varphi = 1 , ~ \\int _ { \\mathcal { M } } | x | ^ { 2 } \\varphi ( x ) \\mathrm { d } x < + \\infty \\right\\} ." + }, + { + "category_id": 14, + "poly": [ + 703, + 718, + 995, + 718, + 995, + 808, + 703, + 808 + ], + "score": 0.95, + "latex": "\\rho ( x ) = \\sum _ { i = 1 } ^ { n } w _ { i } \\delta ( x - x _ { i } ) ," + }, + { + "category_id": 14, + "poly": [ + 605, + 1476, + 1094, + 1476, + 1094, + 1549, + 605, + 1549 + ], + "score": 0.94, + "latex": "W _ { 2 } ^ { 2 } ( \\mu , \\nu ) = \\operatorname* { i n f } _ { \\pi \\in \\Pi ( \\mu , \\nu ) } \\int | x - y | ^ { 2 } { \\mathrm { d } } \\pi ( x , y ) ." + }, + { + "category_id": 14, + "poly": [ + 579, + 1775, + 1119, + 1775, + 1119, + 1932, + 579, + 1932 + ], + "score": 0.94, + "latex": "\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( \\boldsymbol { x } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( \\boldsymbol { x } ) \\nabla \\big ( \\log \\rho ( \\boldsymbol { x } ) - \\log \\pi ( \\boldsymbol { x } ) \\big ) \\right) , } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 753, + 1160, + 946, + 1160, + 946, + 1226, + 753, + 1226 + ], + "score": 0.93, + "latex": "\\frac { d \\boldsymbol { x } } { d t } = - \\boldsymbol { \\nabla } J ( \\boldsymbol { x } ) ." + }, + { + "category_id": 13, + "poly": [ + 656, + 1438, + 745, + 1438, + 745, + 1471, + 656, + 1471 + ], + "score": 0.93, + "latex": "\\mathcal { P } _ { 2 } ( \\mathcal { M } )" + }, + { + "category_id": 14, + "poly": [ + 520, + 1628, + 1176, + 1628, + 1176, + 1701, + 520, + 1701 + ], + "score": 0.93, + "latex": "\\operatorname* { i n f } _ { \\rho \\in \\mathcal { P } _ { 2 } ( \\mathcal { M } ) } D _ { K L } ( \\rho , \\pi ) = \\int _ { \\mathcal { M } } ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\rho ( x ) \\mathrm { d } x ," + }, + { + "category_id": 13, + "poly": [ + 865, + 1261, + 953, + 1261, + 953, + 1295, + 865, + 1295 + ], + "score": 0.92, + "latex": "\\mathcal { P } _ { 2 } ( \\mathcal { M } )" + }, + { + "category_id": 13, + "poly": [ + 384, + 1407, + 474, + 1407, + 474, + 1440, + 384, + 1440 + ], + "score": 0.92, + "latex": "\\Pi ( \\mu , \\nu )" + }, + { + "category_id": 13, + "poly": [ + 831, + 907, + 950, + 907, + 950, + 941, + 831, + 941 + ], + "score": 0.92, + "latex": "w _ { i } = 1 / N" + }, + { + "category_id": 13, + "poly": [ + 1025, + 816, + 1148, + 816, + 1148, + 849, + 1025, + 849 + ], + "score": 0.92, + "latex": "w _ { i } \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1071, + 1589, + 1126, + 1589, + 1126, + 1622, + 1071, + 1622 + ], + "score": 0.92, + "latex": "\\pi ( x )" + }, + { + "category_id": 13, + "poly": [ + 345, + 908, + 399, + 908, + 399, + 941, + 345, + 941 + ], + "score": 0.92, + "latex": "\\rho ( x )" + }, + { + "category_id": 13, + "poly": [ + 299, + 677, + 452, + 677, + 452, + 712, + 299, + 712 + ], + "score": 0.92, + "latex": "\\{ ( \\bar { x _ { i } } , w _ { i } ) \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 671, + 849, + 727, + 849, + 727, + 880, + 671, + 880 + ], + "score": 0.92, + "latex": "\\pi ( x )" + }, + { + "category_id": 13, + "poly": [ + 900, + 1557, + 1093, + 1557, + 1093, + 1592, + 900, + 1592 + ], + "score": 0.92, + "latex": "J : \\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1060, + 1252, + 1060, + 1252, + 1093, + 1202, + 1093 + ], + "score": 0.91, + "latex": "x ( t )" + }, + { + "category_id": 13, + "poly": [ + 1200, + 815, + 1355, + 815, + 1355, + 851, + 1200, + 851 + ], + "score": 0.91, + "latex": "\\textstyle \\sum _ { i = 1 } ^ { n } w _ { i } = 1" + }, + { + "category_id": 13, + "poly": [ + 373, + 816, + 418, + 816, + 418, + 850, + 373, + 850 + ], + "score": 0.91, + "latex": "\\delta ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 1347, + 878, + 1403, + 878, + 1403, + 911, + 1347, + 911 + ], + "score": 0.91, + "latex": "\\pi ( x )" + }, + { + "category_id": 13, + "poly": [ + 576, + 1027, + 715, + 1027, + 715, + 1057, + 576, + 1057 + ], + "score": 0.9, + "latex": "J : \\mathbb { R } ^ { d } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 1128, + 1413, + 1176, + 1413, + 1176, + 1440, + 1128, + 1440 + ], + "score": 0.87, + "latex": "\\mu , \\nu" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1064, + 1069, + 1064, + 1069, + 1092, + 1038, + 1092 + ], + "score": 0.86, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 908, + 821, + 940, + 821, + 940, + 847, + 908, + 847 + ], + "score": 0.86, + "latex": "w _ { i }" + }, + { + "category_id": 13, + "poly": [ + 553, + 911, + 580, + 911, + 580, + 935, + 553, + 935 + ], + "score": 0.85, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 363, + 1094, + 396, + 1094, + 396, + 1121, + 363, + 1121 + ], + "score": 0.85, + "latex": "x _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 769, + 1587, + 804, + 1587, + 804, + 1625, + 769, + 1625 + ], + "score": 0.83, + "latex": "\\textcircled { 8 }" + }, + { + "category_id": 13, + "poly": [ + 880, + 1295, + 915, + 1295, + 915, + 1319, + 880, + 1319 + ], + "score": 0.83, + "latex": "\\mathcal { M }" + }, + { + "category_id": 13, + "poly": [ + 701, + 1713, + 720, + 1713, + 720, + 1740, + 701, + 1740 + ], + "score": 0.8, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 817, + 206, + 830, + 206, + 830, + 231, + 817, + 231 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1232, + 1559, + 1273, + 1559, + 1273, + 1587, + 1232, + 1587 + ], + "score": 0.39, + "latex": "\\mathrm { K L }" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 363.0, + 909.0, + 363.0, + 909.0, + 414.0, + 289.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 585.0, + 640.0, + 585.0, + 640.0, + 631.0, + 292.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 969.0, + 765.0, + 969.0, + 765.0, + 1013.0, + 292.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1941.0, + 1061.0, + 1941.0, + 1061.0, + 1981.0, + 331.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1974.0, + 739.0, + 1974.0, + 739.0, + 2011.0, + 330.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2092.0, + 838.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1941.0, + 1061.0, + 1941.0, + 1061.0, + 1981.0, + 331.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1972.0, + 739.0, + 1972.0, + 739.0, + 2013.0, + 330.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 433.0, + 1406.0, + 433.0, + 1406.0, + 470.0, + 294.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 463.0, + 1404.0, + 463.0, + 1404.0, + 499.0, + 294.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 497.0, + 1404.0, + 497.0, + 1404.0, + 529.0, + 296.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 524.0, + 1409.0, + 524.0, + 1409.0, + 561.0, + 292.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1025.0, + 575.0, + 1025.0, + 575.0, + 1064.0, + 294.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1025.0, + 1406.0, + 1025.0, + 1406.0, + 1064.0, + 716.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1057.0, + 1037.0, + 1057.0, + 1037.0, + 1095.0, + 292.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1057.0, + 1201.0, + 1057.0, + 1201.0, + 1095.0, + 1070.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1095.0, + 1253.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1091.0, + 362.0, + 1091.0, + 362.0, + 1123.0, + 294.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1091.0, + 1408.0, + 1091.0, + 1408.0, + 1123.0, + 397.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1120.0, + 1338.0, + 1120.0, + 1338.0, + 1153.0, + 292.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 799.0, + 372.0, + 799.0, + 372.0, + 865.0, + 285.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 799.0, + 907.0, + 799.0, + 907.0, + 865.0, + 419.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 799.0, + 1024.0, + 799.0, + 1024.0, + 865.0, + 941.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 799.0, + 1199.0, + 799.0, + 1199.0, + 865.0, + 1149.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 799.0, + 1416.0, + 799.0, + 1416.0, + 865.0, + 1356.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 846.0, + 670.0, + 846.0, + 670.0, + 880.0, + 292.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 846.0, + 1406.0, + 846.0, + 1406.0, + 880.0, + 728.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 877.0, + 1346.0, + 877.0, + 1346.0, + 913.0, + 294.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 906.0, + 344.0, + 906.0, + 344.0, + 942.0, + 294.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 906.0, + 552.0, + 906.0, + 552.0, + 942.0, + 400.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 906.0, + 830.0, + 906.0, + 830.0, + 942.0, + 581.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 906.0, + 1356.0, + 906.0, + 1356.0, + 942.0, + 951.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 202.0, + 816.0, + 202.0, + 816.0, + 238.0, + 292.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 202.0, + 1406.0, + 202.0, + 1406.0, + 238.0, + 831.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 233.0, + 1406.0, + 233.0, + 1406.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 263.0, + 1406.0, + 263.0, + 1406.0, + 300.0, + 293.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 295.0, + 1304.0, + 295.0, + 1304.0, + 330.0, + 292.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1229.0, + 1402.0, + 1229.0, + 1402.0, + 1266.0, + 294.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1260.0, + 864.0, + 1260.0, + 864.0, + 1296.0, + 294.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1260.0, + 1403.0, + 1260.0, + 1403.0, + 1296.0, + 954.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1290.0, + 879.0, + 1290.0, + 879.0, + 1324.0, + 295.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1290.0, + 926.0, + 1290.0, + 926.0, + 1324.0, + 916.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 645.0, + 1404.0, + 645.0, + 1404.0, + 684.0, + 293.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 673.0, + 298.0, + 673.0, + 298.0, + 716.0, + 294.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 673.0, + 1022.0, + 673.0, + 1022.0, + 716.0, + 453.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1553.0, + 899.0, + 1553.0, + 899.0, + 1595.0, + 295.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1553.0, + 1231.0, + 1553.0, + 1231.0, + 1595.0, + 1094.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1553.0, + 1405.0, + 1553.0, + 1405.0, + 1595.0, + 1274.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1588.0, + 768.0, + 1588.0, + 768.0, + 1624.0, + 295.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1588.0, + 1070.0, + 1588.0, + 1070.0, + 1624.0, + 805.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1588.0, + 1138.0, + 1588.0, + 1138.0, + 1624.0, + 1127.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1403.0, + 383.0, + 1403.0, + 383.0, + 1444.0, + 298.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1403.0, + 1127.0, + 1403.0, + 1127.0, + 1444.0, + 475.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1403.0, + 1402.0, + 1403.0, + 1402.0, + 1444.0, + 1177.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1435.0, + 655.0, + 1435.0, + 655.0, + 1473.0, + 299.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1435.0, + 786.0, + 1435.0, + 786.0, + 1473.0, + 746.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1704.0, + 700.0, + 1704.0, + 700.0, + 1741.0, + 294.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1704.0, + 1405.0, + 1704.0, + 1405.0, + 1741.0, + 721.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1732.0, + 920.0, + 1732.0, + 920.0, + 1773.0, + 295.0, + 1773.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 661, + 1407, + 661, + 1407, + 933, + 295, + 933 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 296, + 1319, + 1407, + 1319, + 1407, + 1535, + 296, + 1535 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1130, + 1405, + 1130, + 1405, + 1310, + 297, + 1310 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1882, + 1403, + 1882, + 1403, + 2009, + 298, + 2009 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1629, + 1405, + 1629, + 1405, + 1782, + 298, + 1782 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 199, + 1406, + 199, + 1406, + 301, + 297, + 301 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 594, + 306, + 1104, + 306, + 1104, + 414, + 594, + 414 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 458, + 1539, + 1237, + 1539, + 1237, + 1624, + 458, + 1624 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 335, + 1787, + 1310, + 1787, + 1310, + 1878, + 335, + 1878 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 643, + 931, + 1055, + 931, + 1055, + 1012, + 643, + 1012 + ], + "score": 0.945 + }, + { + "category_id": 8, + "poly": [ + 658, + 486, + 1040, + 486, + 1040, + 547, + 658, + 547 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 534, + 1048, + 1160, + 1048, + 1160, + 1127, + 534, + 1127 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 296, + 417, + 1400, + 417, + 1400, + 480, + 296, + 480 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 292, + 560, + 1402, + 560, + 1402, + 625, + 292, + 625 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 299, + 1010, + 831, + 1010, + 831, + 1044, + 299, + 1044 + ], + "score": 0.929 + }, + { + "category_id": 8, + "poly": [ + 593, + 623, + 1099, + 623, + 1099, + 662, + 593, + 662 + ], + "score": 0.927 + }, + { + "category_id": 9, + "poly": [ + 1351, + 957, + 1400, + 957, + 1400, + 987, + 1351, + 987 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1562, + 1400, + 1562, + 1400, + 1592, + 1352, + 1592 + ], + "score": 0.899 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1071, + 1400, + 1071, + 1400, + 1103, + 1352, + 1103 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1351, + 500, + 1400, + 500, + 1400, + 530, + 1351, + 530 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1815, + 1400, + 1815, + 1400, + 1846, + 1352, + 1846 + ], + "score": 0.862 + }, + { + "category_id": 9, + "poly": [ + 1352, + 627, + 1400, + 627, + 1400, + 657, + 1352, + 657 + ], + "score": 0.855 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2085, + 840, + 2085 + ], + "score": 0.806 + }, + { + "category_id": 14, + "poly": [ + 643, + 928, + 1057, + 928, + 1057, + 1015, + 643, + 1015 + ], + "score": 0.95, + "latex": "\\nabla \\log \\rho ( x _ { i } ^ { t } ) \\approx \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } ." + }, + { + "category_id": 14, + "poly": [ + 657, + 481, + 1043, + 481, + 1043, + 547, + 657, + 547 + ], + "score": 0.94, + "latex": "\\frac { d x } { d t } = - \\nabla \\big ( \\log \\rho ( x ) - \\log \\pi ( x ) \\big ) ." + }, + { + "category_id": 14, + "poly": [ + 461, + 1534, + 1236, + 1534, + 1236, + 1627, + 461, + 1627 + ], + "score": 0.94, + "latex": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) + \\sum _ { j = 1 } ^ { n } ( K + \\eta \\mathbb { I } ) _ { i j } ^ { - 1 } \\sum _ { k = 1 } ^ { n } \\nabla _ { x _ { k } ^ { t } } k ( x _ { k } ^ { t } , x _ { j } ^ { t } ) \\bigg ) ," + }, + { + "category_id": 14, + "poly": [ + 338, + 1783, + 1314, + 1783, + 1314, + 1880, + 338, + 1880 + ], + "score": 0.94, + "latex": "x _ { t + 1 } ^ { i } = x _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla \\log \\pi ( x _ { t } ^ { i } ) + \\sum _ { j = 1 } ^ { J } \\frac { 1 } { \\lambda _ { j } ^ { 2 } } \\sum _ { m = 1 } ^ { n } \\sum _ { k = 1 } ^ { n } u _ { j k } \\nabla _ { x _ { m } } k ( x _ { m } ^ { t } , x _ { k } ^ { t } ) \\cdot \\sum _ { l = 1 } ^ { n } u _ { j l } k ( x _ { i } ^ { t } , x _ { l } ^ { t } ) \\bigg )" + }, + { + "category_id": 13, + "poly": [ + 299, + 1190, + 800, + 1190, + 800, + 1251, + 299, + 1251 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\left( \\bar { \\sum _ { j = 1 } ^ { n } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) \\right) ^ { - 1 } \\sum _ { j = 1 } ^ { n } \\nabla _ { \\mathbf { w } _ { i } ^ { t } } k ( \\mathbf { w } _ { i } ^ { t } , \\mathbf { w } _ { j } ^ { t } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 985, + 563, + 1068, + 563, + 1068, + 598, + 985, + 598 + ], + "score": 0.93, + "latex": "\\{ x \\} _ { i = 1 } ^ { n }" + }, + { + "category_id": 14, + "poly": [ + 537, + 1044, + 1160, + 1044, + 1160, + 1129, + 537, + 1129 + ], + "score": 0.93, + "latex": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\Bigg ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { x _ { i } ^ { t } } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( x _ { i } ^ { t } , x _ { j } ^ { t } ) } \\Bigg ) ," + }, + { + "category_id": 13, + "poly": [ + 377, + 200, + 803, + 200, + 803, + 241, + 377, + 241 + ], + "score": 0.93, + "latex": "\\nabla _ { \\delta \\rho } ^ { \\delta } D _ { K L } ( \\rho , \\pi ) \\ = : \\ \\nabla _ { \\mathcal { W } _ { 2 } } D _ { K L } ( \\rho , \\pi )" + }, + { + "category_id": 14, + "poly": [ + 593, + 304, + 1102, + 304, + 1102, + 417, + 593, + 417 + ], + "score": 0.92, + "latex": "\\begin{array} { r l r } & { } & { \\frac { \\partial \\rho ( x ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( x ) \\nabla ( \\log \\rho ( x ) - \\log \\pi ( x ) ) \\right) } \\\\ & { } & { = - \\nabla \\cdot \\left( \\rho ( x ) \\nabla \\log \\pi ( x ) \\right) + \\nabla ^ { 2 } \\rho ( x ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 933, + 418, + 989, + 418, + 989, + 451, + 933, + 451 + ], + "score": 0.92, + "latex": "\\pi ( x )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1975, + 460, + 1975, + 460, + 2009, + 298, + 2009 + ], + "score": 0.92, + "latex": "\\mathcal { O } ( M ( d + J ) )" + }, + { + "category_id": 13, + "poly": [ + 672, + 870, + 925, + 870, + 925, + 905, + 672, + 905 + ], + "score": 0.91, + "latex": "\\boldsymbol { k } ( \\cdot , \\cdot ) : \\mathbb { R } ^ { d } \\times \\mathbb { R } ^ { d } \\dot { \\mathbb { R } }" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1912, + 1228, + 1912, + 1228, + 1948, + 1051, + 1948 + ], + "score": 0.91, + "latex": "\\mathcal { O } ( M ^ { 3 } + M ^ { 2 } d )" + }, + { + "category_id": 13, + "poly": [ + 298, + 240, + 465, + 240, + 465, + 273, + 298, + 273 + ], + "score": 0.91, + "latex": "\\mathcal { P } _ { 2 } ( \\mathcal { M } ) \\stackrel { \\cdot } { } \\mathbb { R }" + }, + { + "category_id": 14, + "poly": [ + 593, + 621, + 1105, + 621, + 1105, + 663, + 593, + 663 + ], + "score": 0.91, + "latex": "x _ { i } ^ { t + 1 } = x _ { i } ^ { t } + \\epsilon _ { t } \\bigl ( \\nabla \\log \\pi ( x _ { i } ^ { t } ) - \\nabla \\log \\rho ( x _ { i } ^ { t } ) \\bigr ) ." + }, + { + "category_id": 13, + "poly": [ + 297, + 870, + 585, + 870, + 585, + 908, + 297, + 908 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\tilde { \\rho } _ { t } ( x ) = \\frac { 1 } { n } \\sum _ { i = 1 } ^ { n } \\dot { k } ( x , x _ { t } ^ { i } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1151, + 661, + 1236, + 661, + 1236, + 691, + 1151, + 691 + ], + "score": 0.9, + "latex": "\\nabla \\log \\rho" + }, + { + "category_id": 13, + "poly": [ + 375, + 1882, + 406, + 1882, + 406, + 1916, + 375, + 1916 + ], + "score": 0.89, + "latex": "\\lambda _ { j }" + }, + { + "category_id": 13, + "poly": [ + 751, + 719, + 801, + 719, + 801, + 756, + 751, + 756 + ], + "score": 0.89, + "latex": "\\textcircled { 1 3 }" + }, + { + "category_id": 13, + "poly": [ + 874, + 1159, + 910, + 1159, + 910, + 1196, + 874, + 1196 + ], + "score": 0.88, + "latex": "( 3 )" + }, + { + "category_id": 13, + "poly": [ + 1262, + 238, + 1294, + 238, + 1294, + 276, + 1262, + 276 + ], + "score": 0.87, + "latex": "\\boxed { \\mathsf { C } }" + }, + { + "category_id": 13, + "poly": [ + 1169, + 719, + 1206, + 719, + 1206, + 755, + 1169, + 755 + ], + "score": 0.87, + "latex": "( 3 )" + }, + { + "category_id": 13, + "poly": [ + 1185, + 1249, + 1275, + 1249, + 1275, + 1274, + 1185, + 1274 + ], + "score": 0.85, + "latex": "n \\to \\infty" + }, + { + "category_id": 13, + "poly": [ + 683, + 1412, + 732, + 1412, + 732, + 1444, + 683, + 1444 + ], + "score": 0.85, + "latex": "\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { A } } \\right. }" + }, + { + "category_id": 13, + "poly": [ + 1345, + 1161, + 1406, + 1161, + 1406, + 1191, + 1345, + 1191 + ], + "score": 0.85, + "latex": "\\mathcal { R } =" + }, + { + "category_id": 13, + "poly": [ + 927, + 1351, + 975, + 1351, + 975, + 1383, + 927, + 1383 + ], + "score": 0.84, + "latex": "[ \\overline { { | 4 6 | } }" + }, + { + "category_id": 13, + "poly": [ + 343, + 1160, + 377, + 1160, + 377, + 1195, + 343, + 1195 + ], + "score": 0.84, + "latex": "\\textcircled{3}" + }, + { + "category_id": 13, + "poly": [ + 962, + 1887, + 1004, + 1887, + 1004, + 1916, + 962, + 1916 + ], + "score": 0.83, + "latex": "u _ { j k }" + }, + { + "category_id": 13, + "poly": [ + 1283, + 570, + 1307, + 570, + 1307, + 594, + 1283, + 594 + ], + "score": 0.83, + "latex": "\\epsilon _ { t }" + }, + { + "category_id": 13, + "poly": [ + 548, + 1977, + 569, + 1977, + 569, + 2003, + 548, + 2003 + ], + "score": 0.83, + "latex": "J" + }, + { + "category_id": 13, + "poly": [ + 372, + 1630, + 400, + 1630, + 400, + 1656, + 372, + 1656 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 485, + 1884, + 501, + 1884, + 501, + 1914, + 485, + 1914 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1353, + 1885, + 1369, + 1885, + 1369, + 1914, + 1353, + 1914 + ], + "score": 0.82, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1298, + 1918, + 1330, + 1918, + 1330, + 1942, + 1298, + 1942 + ], + "score": 0.82, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 659, + 1011, + 708, + 1011, + 708, + 1047, + 659, + 1047 + ], + "score": 0.82, + "latex": "( 1 3 )" + }, + { + "category_id": 13, + "poly": [ + 1084, + 1883, + 1102, + 1883, + 1102, + 1909, + 1084, + 1909 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 705, + 1635, + 723, + 1635, + 723, + 1661, + 705, + 1661 + ], + "score": 0.82, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 595, + 1947, + 612, + 1947, + 612, + 1973, + 595, + 1973 + ], + "score": 0.82, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 338, + 1692, + 350, + 1692, + 350, + 1717, + 338, + 1717 + ], + "score": 0.81, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 621, + 1276, + 668, + 1276, + 668, + 1308, + 621, + 1308 + ], + "score": 0.77, + "latex": "\\left[ \\left[ 6 6 \\right] \\right]" + }, + { + "category_id": 13, + "poly": [ + 342, + 1472, + 389, + 1472, + 389, + 1509, + 342, + 1509 + ], + "score": 0.77, + "latex": "\\textcircled { 1 3 }" + }, + { + "category_id": 13, + "poly": [ + 775, + 1137, + 795, + 1137, + 795, + 1157, + 775, + 1157 + ], + "score": 0.77, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 333, + 904, + 381, + 904, + 381, + 934, + 333, + 934 + ], + "score": 0.75, + "latex": "\\mathbb { \\ m }" + }, + { + "category_id": 13, + "poly": [ + 1348, + 200, + 1380, + 200, + 1380, + 243, + 1348, + 243 + ], + "score": 0.7, + "latex": "\\frac { \\delta } { \\delta \\rho }" + }, + { + "category_id": 13, + "poly": [ + 427, + 593, + 476, + 593, + 476, + 629, + 427, + 629 + ], + "score": 0.7, + "latex": "\\mathbb { \\underline { { ( 1 2 ) } } }" + }, + { + "category_id": 13, + "poly": [ + 1369, + 841, + 1403, + 841, + 1403, + 877, + 1369, + 877 + ], + "score": 0.68, + "latex": "\\dot { \\mathrm { ~ E ~ } } )" + }, + { + "category_id": 13, + "poly": [ + 362, + 446, + 397, + 446, + 397, + 479, + 362, + 479 + ], + "score": 0.61, + "latex": "\\left[ \\left[ 2 \\right] \\right]" + }, + { + "category_id": 13, + "poly": [ + 556, + 446, + 604, + 446, + 604, + 482, + 556, + 482 + ], + "score": 0.58, + "latex": "\\mathbb { ( 1 1 ) }" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2089.0, + 840.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 656.0, + 1150.0, + 656.0, + 1150.0, + 695.0, + 294.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 656.0, + 1409.0, + 656.0, + 1409.0, + 695.0, + 1237.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 692.0, + 1403.0, + 692.0, + 1403.0, + 721.0, + 297.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 718.0, + 750.0, + 718.0, + 750.0, + 758.0, + 293.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 718.0, + 1168.0, + 718.0, + 1168.0, + 758.0, + 802.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 718.0, + 1408.0, + 718.0, + 1408.0, + 758.0, + 1207.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 749.0, + 1405.0, + 749.0, + 1405.0, + 785.0, + 294.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 780.0, + 1409.0, + 780.0, + 1409.0, + 816.0, + 294.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 809.0, + 1405.0, + 809.0, + 1405.0, + 846.0, + 293.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 841.0, + 1368.0, + 841.0, + 1368.0, + 877.0, + 293.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 866.0, + 671.0, + 866.0, + 671.0, + 912.0, + 586.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 866.0, + 1410.0, + 866.0, + 1410.0, + 912.0, + 926.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 899.0, + 332.0, + 899.0, + 332.0, + 939.0, + 294.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 900.0, + 394.0, + 900.0, + 394.0, + 937.0, + 382.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1318.0, + 1409.0, + 1318.0, + 1409.0, + 1358.0, + 293.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1352.0, + 926.0, + 1352.0, + 926.0, + 1386.0, + 295.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1352.0, + 1407.0, + 1352.0, + 1407.0, + 1386.0, + 976.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1383.0, + 1405.0, + 1383.0, + 1405.0, + 1417.0, + 294.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1412.0, + 682.0, + 1412.0, + 682.0, + 1447.0, + 294.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1412.0, + 1407.0, + 1412.0, + 1407.0, + 1447.0, + 733.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1443.0, + 1407.0, + 1443.0, + 1407.0, + 1478.0, + 294.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1470.0, + 341.0, + 1470.0, + 341.0, + 1512.0, + 294.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1470.0, + 1407.0, + 1470.0, + 1407.0, + 1512.0, + 390.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1505.0, + 875.0, + 1505.0, + 875.0, + 1536.0, + 295.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1126.0, + 774.0, + 1126.0, + 774.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1126.0, + 1405.0, + 1126.0, + 1405.0, + 1164.0, + 796.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1154.0, + 342.0, + 1154.0, + 342.0, + 1200.0, + 293.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1154.0, + 873.0, + 1154.0, + 873.0, + 1200.0, + 378.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1154.0, + 1344.0, + 1154.0, + 1344.0, + 1200.0, + 911.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1181.0, + 1419.0, + 1181.0, + 1419.0, + 1264.0, + 801.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1244.0, + 1184.0, + 1244.0, + 1184.0, + 1279.0, + 295.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1244.0, + 1402.0, + 1244.0, + 1402.0, + 1279.0, + 1276.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1274.0, + 620.0, + 1274.0, + 620.0, + 1312.0, + 295.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1274.0, + 1192.0, + 1274.0, + 1192.0, + 1312.0, + 669.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 374.0, + 1880.0, + 374.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 1880.0, + 484.0, + 1880.0, + 484.0, + 1917.0, + 407.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1880.0, + 961.0, + 1880.0, + 961.0, + 1917.0, + 502.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1880.0, + 1083.0, + 1880.0, + 1083.0, + 1917.0, + 1005.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 1880.0, + 1352.0, + 1880.0, + 1352.0, + 1917.0, + 1103.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1917.0, + 1370.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1050.0, + 1910.0, + 1050.0, + 1953.0, + 292.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1910.0, + 1297.0, + 1910.0, + 1297.0, + 1953.0, + 1229.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1953.0, + 1331.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1947.0, + 594.0, + 1947.0, + 594.0, + 1980.0, + 294.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1947.0, + 1405.0, + 1947.0, + 1405.0, + 1980.0, + 613.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1975.0, + 547.0, + 1975.0, + 547.0, + 2011.0, + 461.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1975.0, + 901.0, + 1975.0, + 901.0, + 2011.0, + 570.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1625.0, + 371.0, + 1625.0, + 371.0, + 1663.0, + 293.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1625.0, + 704.0, + 1625.0, + 704.0, + 1663.0, + 401.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 1625.0, + 1406.0, + 1625.0, + 1406.0, + 1663.0, + 724.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1692.0, + 293.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1690.0, + 337.0, + 1690.0, + 337.0, + 1723.0, + 296.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1690.0, + 1405.0, + 1690.0, + 1405.0, + 1723.0, + 351.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1717.0, + 1408.0, + 1717.0, + 1408.0, + 1754.0, + 293.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1750.0, + 681.0, + 1750.0, + 681.0, + 1783.0, + 296.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 197.0, + 376.0, + 197.0, + 376.0, + 241.0, + 291.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 197.0, + 1347.0, + 197.0, + 1347.0, + 241.0, + 804.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 239.0, + 1261.0, + 239.0, + 1261.0, + 274.0, + 466.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 239.0, + 1406.0, + 239.0, + 1406.0, + 274.0, + 1295.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 269.0, + 1364.0, + 269.0, + 1364.0, + 306.0, + 294.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 214.5, + 1379.0, + 214.5, + 1379.0, + 243.0, + 1351.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 414.0, + 932.0, + 414.0, + 932.0, + 451.0, + 294.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 414.0, + 1404.0, + 414.0, + 1404.0, + 451.0, + 990.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 443.0, + 361.0, + 443.0, + 361.0, + 483.0, + 295.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 443.0, + 555.0, + 443.0, + 555.0, + 483.0, + 398.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 443.0, + 1276.0, + 443.0, + 1276.0, + 483.0, + 605.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 558.0, + 984.0, + 558.0, + 984.0, + 602.0, + 295.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 558.0, + 1282.0, + 558.0, + 1282.0, + 602.0, + 1069.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 558.0, + 1407.0, + 558.0, + 1407.0, + 602.0, + 1308.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 583.0, + 426.0, + 583.0, + 426.0, + 633.0, + 291.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 583.0, + 519.0, + 583.0, + 519.0, + 633.0, + 477.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1008.0, + 658.0, + 1008.0, + 658.0, + 1048.0, + 296.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1008.0, + 834.0, + 1008.0, + 834.0, + 1048.0, + 709.0, + 1048.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1264, + 1406, + 1264, + 1406, + 1544, + 297, + 1544 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1641, + 1405, + 1641, + 1405, + 2007, + 297, + 2007 + ], + "score": 0.984 + }, + { + "category_id": 8, + "poly": [ + 576, + 774, + 1123, + 774, + 1123, + 930, + 576, + 930 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 309, + 191, + 1395, + 191, + 1395, + 469, + 309, + 469 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 523, + 1162, + 1175, + 1162, + 1175, + 1249, + 523, + 1249 + ], + "score": 0.961 + }, + { + "category_id": 8, + "poly": [ + 653, + 999, + 1045, + 999, + 1045, + 1065, + 653, + 1065 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 292, + 1080, + 1403, + 1080, + 1403, + 1147, + 292, + 1147 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 296, + 692, + 1401, + 692, + 1401, + 760, + 296, + 760 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 296, + 945, + 988, + 945, + 988, + 979, + 296, + 979 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 299, + 1583, + 647, + 1583, + 647, + 1616, + 299, + 1616 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 299, + 636, + 733, + 636, + 733, + 669, + 299, + 669 + ], + "score": 0.924 + }, + { + "category_id": 4, + "poly": [ + 297, + 480, + 1406, + 480, + 1406, + 574, + 297, + 574 + ], + "score": 0.923 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1017, + 1400, + 1017, + 1400, + 1048, + 1352, + 1048 + ], + "score": 0.911 + }, + { + "category_id": 9, + "poly": [ + 1351, + 837, + 1400, + 837, + 1400, + 868, + 1351, + 868 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1189, + 1400, + 1189, + 1400, + 1220, + 1352, + 1220 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.717 + }, + { + "category_id": 4, + "poly": [ + 300, + 480, + 1401, + 480, + 1401, + 574, + 300, + 574 + ], + "score": 0.123 + }, + { + "category_id": 2, + "poly": [ + 841, + 2060, + 859, + 2060, + 859, + 2084, + 841, + 2084 + ], + "score": 0.103 + }, + { + "category_id": 14, + "poly": [ + 652, + 996, + 1048, + 996, + 1048, + 1065, + 652, + 1065 + ], + "score": 0.94, + "latex": "\\frac { d \\pmb f } { d t } = - \\nabla \\big ( \\log \\rho ( \\pmb f ) - \\log \\pi ( \\pmb f ) \\big ) ." + }, + { + "category_id": 14, + "poly": [ + 574, + 775, + 1128, + 775, + 1128, + 933, + 574, + 933 + ], + "score": 0.93, + "latex": "\\begin{array} { l } { \\displaystyle \\frac { \\partial \\rho ( { \\pmb f } ) } { \\partial t } = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\frac { \\delta } { \\delta \\rho } D _ { K L } ( \\rho , \\pi ) \\right) } \\\\ { \\displaystyle = \\nabla \\cdot \\left( \\rho ( { \\pmb f } ) \\nabla \\big ( \\log \\rho ( { \\pmb f } ) - \\log \\pi ( { \\pmb f } ) \\big ) \\right) . } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 523, + 1161, + 1177, + 1161, + 1177, + 1250, + 523, + 1250 + ], + "score": 0.93, + "latex": "\\pmb { f } _ { t + 1 } ^ { i } = \\pmb { f } _ { t } ^ { i } + \\epsilon _ { t } \\bigg ( \\nabla _ { \\pmb { f } } \\log \\pi ( \\pmb { f } _ { t } ^ { i } ) - \\frac { \\sum _ { j = 1 } ^ { n } \\nabla _ { \\pmb { f } _ { i } ^ { t } } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } { \\sum _ { j = 1 } ^ { n } k ( \\pmb { f } _ { i } ^ { t } , \\pmb { f } _ { j } ^ { t } ) } \\bigg ) ." + }, + { + "category_id": 13, + "poly": [ + 1159, + 1296, + 1406, + 1296, + 1406, + 1332, + 1159, + 1332 + ], + "score": 0.91, + "latex": "\\nabla _ { f _ { j } } \\log { p ( f _ { j } | x , y ) } =" + }, + { + "category_id": 13, + "poly": [ + 296, + 1328, + 700, + 1328, + 700, + 1366, + 296, + 1366 + ], + "score": 0.9, + "latex": "\\nabla _ { f _ { j } } \\log p ( \\pmb { y } | \\pmb { x } , \\pmb { f _ { j } } ) + \\nabla _ { f _ { j } } \\log p ( \\pmb { f _ { j } } )" + }, + { + "category_id": 13, + "poly": [ + 845, + 1419, + 894, + 1419, + 894, + 1456, + 845, + 1456 + ], + "score": 0.88, + "latex": "\\textcircled { 1 2 0 }" + }, + { + "category_id": 13, + "poly": [ + 930, + 1111, + 966, + 1111, + 966, + 1148, + 930, + 1148 + ], + "score": 0.86, + "latex": "\\textcircled{4}" + }, + { + "category_id": 13, + "poly": [ + 568, + 1389, + 617, + 1389, + 617, + 1422, + 568, + 1422 + ], + "score": 0.85, + "latex": "\\mathbb { \\lVert \\boldsymbol { 6 9 } \\rVert }" + }, + { + "category_id": 13, + "poly": [ + 814, + 724, + 863, + 724, + 863, + 760, + 814, + 760 + ], + "score": 0.79, + "latex": "\\textcircled { 1 1 }" + }, + { + "category_id": 13, + "poly": [ + 1292, + 1389, + 1340, + 1389, + 1340, + 1422, + 1292, + 1422 + ], + "score": 0.77, + "latex": "\\pmb { \\mathbb { Z } 1 }" + }, + { + "category_id": 13, + "poly": [ + 1179, + 1732, + 1201, + 1732, + 1201, + 1768, + 1179, + 1768 + ], + "score": 0.68, + "latex": "\\checkmark" + }, + { + "category_id": 13, + "poly": [ + 941, + 1793, + 990, + 1793, + 990, + 1825, + 941, + 1825 + ], + "score": 0.64, + "latex": "\\dot { \\overline { { \\vert 6 9 \\vert } } }" + }, + { + "category_id": 13, + "poly": [ + 1141, + 693, + 1175, + 693, + 1175, + 730, + 1141, + 730 + ], + "score": 0.54, + "latex": "\\textcircled{5}" + }, + { + "category_id": 13, + "poly": [ + 890, + 1732, + 940, + 1732, + 940, + 1768, + 890, + 1768 + ], + "score": 0.52, + "latex": "( 1 3 )" + }, + { + "category_id": 13, + "poly": [ + 929, + 1511, + 978, + 1511, + 978, + 1543, + 929, + 1543 + ], + "score": 0.49, + "latex": "\\checkmark" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 245.0, + 485.0, + 245.0, + 485.0, + 258.0, + 469.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 243.0, + 530.0, + 243.0, + 530.0, + 259.0, + 489.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 245.0, + 747.0, + 245.0, + 747.0, + 258.0, + 730.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 241.0, + 811.0, + 241.0, + 811.0, + 260.0, + 748.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 245.0, + 1013.0, + 245.0, + 1013.0, + 258.0, + 997.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 241.0, + 1081.0, + 241.0, + 1081.0, + 260.0, + 1015.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 244.0, + 1346.0, + 244.0, + 1346.0, + 259.0, + 1278.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 295.0, + 868.0, + 295.0, + 868.0, + 304.0, + 857.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 323.0, + 867.0, + 323.0, + 867.0, + 332.0, + 857.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 406.0, + 1136.0, + 406.0, + 1136.0, + 414.0, + 1126.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 432.0, + 595.0, + 432.0, + 595.0, + 441.0, + 583.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 456.0, + 395.0, + 456.0, + 395.0, + 465.0, + 385.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 456.0, + 667.0, + 456.0, + 667.0, + 465.0, + 652.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 456.0, + 691.0, + 456.0, + 691.0, + 465.0, + 677.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 456.0, + 711.0, + 456.0, + 711.0, + 465.0, + 698.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.0, + 456.0, + 735.0, + 456.0, + 735.0, + 465.0, + 724.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 456.0, + 958.0, + 456.0, + 958.0, + 465.0, + 948.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 456.0, + 983.0, + 456.0, + 983.0, + 465.0, + 968.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 456.0, + 1052.0, + 456.0, + 1052.0, + 465.0, + 1041.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 456.0, + 1210.0, + 456.0, + 1210.0, + 465.0, + 1195.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 456.0, + 1231.0, + 456.0, + 1231.0, + 465.0, + 1215.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 456.0, + 1255.0, + 456.0, + 1255.0, + 465.0, + 1240.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 456.0, + 1276.0, + 456.0, + 1276.0, + 465.0, + 1265.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 307.5, + 730.0, + 307.5, + 730.0, + 354.5, + 716.0, + 354.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 649.0, + 1581.0, + 649.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 632.0, + 737.0, + 632.0, + 737.0, + 675.0, + 292.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 478.0, + 1409.0, + 478.0, + 1409.0, + 516.0, + 295.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 508.0, + 1404.0, + 508.0, + 1404.0, + 546.0, + 292.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1209.0, + 539.0, + 1209.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 478.0, + 1406.0, + 478.0, + 1406.0, + 516.0, + 295.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 508.0, + 1406.0, + 508.0, + 1406.0, + 546.0, + 294.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 539.0, + 1210.0, + 539.0, + 1210.0, + 576.0, + 295.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1267.0, + 1408.0, + 1267.0, + 1408.0, + 1300.0, + 296.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1296.0, + 1158.0, + 1296.0, + 1158.0, + 1333.0, + 294.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1323.0, + 295.0, + 1323.0, + 295.0, + 1371.0, + 291.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1323.0, + 1411.0, + 1323.0, + 1411.0, + 1371.0, + 701.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1360.0, + 1407.0, + 1360.0, + 1407.0, + 1393.0, + 295.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1387.0, + 567.0, + 1387.0, + 567.0, + 1426.0, + 292.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1387.0, + 1291.0, + 1387.0, + 1291.0, + 1426.0, + 618.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 1387.0, + 1408.0, + 1387.0, + 1408.0, + 1426.0, + 1341.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1418.0, + 844.0, + 1418.0, + 844.0, + 1457.0, + 292.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1418.0, + 1404.0, + 1418.0, + 1404.0, + 1457.0, + 895.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1450.0, + 1406.0, + 1450.0, + 1406.0, + 1487.0, + 294.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1481.0, + 1406.0, + 1481.0, + 1406.0, + 1517.0, + 295.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1510.0, + 928.0, + 1510.0, + 928.0, + 1544.0, + 294.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1510.0, + 982.0, + 1510.0, + 982.0, + 1544.0, + 979.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1641.0, + 1406.0, + 1641.0, + 1406.0, + 1678.0, + 294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1669.0, + 1406.0, + 1669.0, + 1406.0, + 1710.0, + 292.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1701.0, + 1406.0, + 1701.0, + 1406.0, + 1737.0, + 292.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1733.0, + 889.0, + 1733.0, + 889.0, + 1771.0, + 294.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1733.0, + 1178.0, + 1733.0, + 1178.0, + 1771.0, + 941.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1733.0, + 1408.0, + 1733.0, + 1408.0, + 1771.0, + 1202.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1763.0, + 1407.0, + 1763.0, + 1407.0, + 1796.0, + 294.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1792.0, + 940.0, + 1792.0, + 940.0, + 1826.0, + 294.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1792.0, + 1407.0, + 1792.0, + 1407.0, + 1826.0, + 991.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1820.0, + 1408.0, + 1820.0, + 1408.0, + 1859.0, + 292.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1854.0, + 1405.0, + 1854.0, + 1405.0, + 1888.0, + 295.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1920.0, + 294.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1951.0, + 292.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1947.0, + 1403.0, + 1947.0, + 1403.0, + 1981.0, + 295.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1977.0, + 724.0, + 1977.0, + 724.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1077.0, + 1405.0, + 1077.0, + 1405.0, + 1117.0, + 295.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1107.0, + 929.0, + 1107.0, + 929.0, + 1151.0, + 292.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1107.0, + 1005.0, + 1107.0, + 1005.0, + 1151.0, + 967.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 689.0, + 1140.0, + 689.0, + 1140.0, + 730.0, + 294.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 689.0, + 1405.0, + 689.0, + 1405.0, + 730.0, + 1176.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 725.0, + 813.0, + 725.0, + 813.0, + 761.0, + 294.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 725.0, + 1095.0, + 725.0, + 1095.0, + 761.0, + 864.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 945.0, + 988.0, + 945.0, + 988.0, + 982.0, + 296.0, + 982.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 894, + 1404, + 894, + 1404, + 1321, + 298, + 1321 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1356, + 1405, + 1356, + 1405, + 1539, + 297, + 1539 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1574, + 1404, + 1574, + 1404, + 1819, + 297, + 1819 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1854, + 1403, + 1854, + 1403, + 2007, + 299, + 2007 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 336, + 207, + 1416, + 207, + 1416, + 646, + 336, + 646 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 301, + 674, + 1409, + 674, + 1409, + 766, + 301, + 766 + ], + "score": 0.957 + }, + { + "category_id": 0, + "poly": [ + 298, + 824, + 532, + 824, + 532, + 862, + 298, + 862 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2085, + 841, + 2085 + ], + "score": 0.8 + }, + { + "category_id": 13, + "poly": [ + 1136, + 1137, + 1238, + 1137, + 1238, + 1170, + 1136, + 1170 + ], + "score": 0.91, + "latex": "( H _ { o } / H _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1017, + 1260, + 1175, + 1260, + 1175, + 1292, + 1017, + 1292 + ], + "score": 0.91, + "latex": "( M D _ { o } / M D _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 699, + 1477, + 748, + 1477, + 748, + 1509, + 699, + 1509 + ], + "score": 0.82, + "latex": "\\mathbb { \\left[ \\left[ 5 9 \\right] \\right] }" + }, + { + "category_id": 13, + "poly": [ + 717, + 1636, + 809, + 1636, + 809, + 1667, + 717, + 1667 + ], + "score": 0.66, + "latex": "\\pm 1 , 2 , 3" + }, + { + "category_id": 13, + "poly": [ + 894, + 1416, + 924, + 1416, + 924, + 1452, + 894, + 1452 + ], + "score": 0.66, + "latex": "^ { 2 ) }" + }, + { + "category_id": 13, + "poly": [ + 1323, + 955, + 1372, + 955, + 1372, + 988, + 1323, + 988 + ], + "score": 0.64, + "latex": "\\mathbb { \\left[ \\left[ 5 1 \\right] \\right] }" + }, + { + "category_id": 13, + "poly": [ + 869, + 1167, + 919, + 1167, + 919, + 1200, + 869, + 1200 + ], + "score": 0.47, + "latex": "\\mathbb { H }" + }, + { + "category_id": 13, + "poly": [ + 916, + 1228, + 949, + 1228, + 949, + 1264, + 916, + 1264 + ], + "score": 0.43, + "latex": "\\bar { \\bigtriangledown } \\bar { \\bigtriangledown }" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1476, + 1143, + 1476, + 1143, + 1513, + 1099, + 1513 + ], + "score": 0.36, + "latex": "\\mathrm { F . l }" + }, + { + "category_id": 13, + "poly": [ + 508, + 1913, + 537, + 1913, + 537, + 1950, + 508, + 1950 + ], + "score": 0.31, + "latex": "\\textcircled { 3 }" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 210.0, + 511.0, + 210.0, + 511.0, + 239.0, + 363.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 212.0, + 679.0, + 212.0, + 679.0, + 236.0, + 598.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 212.0, + 885.0, + 212.0, + 885.0, + 236.0, + 795.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 212.0, + 1090.0, + 212.0, + 1090.0, + 236.0, + 990.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 214.0, + 1286.0, + 214.0, + 1286.0, + 235.0, + 1195.0, + 235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1374.0, + 223.0, + 1414.0, + 223.0, + 1414.0, + 244.0, + 1374.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 313.0, + 425.0, + 313.0, + 425.0, + 331.0, + 397.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 313.0, + 473.0, + 313.0, + 473.0, + 336.0, + 443.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 312.0, + 626.0, + 312.0, + 626.0, + 334.0, + 597.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 313.0, + 674.0, + 313.0, + 674.0, + 336.0, + 646.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 315.0, + 827.0, + 315.0, + 827.0, + 333.0, + 800.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 313.0, + 876.0, + 313.0, + 876.0, + 336.0, + 848.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 315.0, + 1028.0, + 315.0, + 1028.0, + 331.0, + 1001.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 313.0, + 1075.0, + 313.0, + 1075.0, + 337.0, + 1048.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 313.0, + 1230.0, + 313.0, + 1230.0, + 331.0, + 1202.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 312.0, + 1277.0, + 312.0, + 1277.0, + 336.0, + 1247.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 301.0, + 1418.0, + 301.0, + 1418.0, + 331.0, + 1373.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 339.0, + 435.0, + 339.0, + 435.0, + 364.0, + 410.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 339.0, + 464.0, + 339.0, + 464.0, + 362.0, + 437.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 337.0, + 665.0, + 337.0, + 665.0, + 364.0, + 611.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 337.0, + 838.0, + 337.0, + 838.0, + 365.0, + 812.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 341.0, + 865.0, + 341.0, + 865.0, + 359.0, + 844.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 337.0, + 1064.0, + 337.0, + 1064.0, + 363.0, + 1012.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 336.0, + 1242.0, + 336.0, + 1242.0, + 365.0, + 1212.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 346.0, + 1263.0, + 346.0, + 1263.0, + 354.0, + 1252.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 383.0, + 1418.0, + 383.0, + 1418.0, + 413.0, + 1370.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 428.0, + 462.0, + 428.0, + 462.0, + 452.0, + 406.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 428.0, + 676.0, + 428.0, + 676.0, + 452.0, + 601.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 429.0, + 888.0, + 429.0, + 888.0, + 453.0, + 789.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 428.0, + 1096.0, + 428.0, + 1096.0, + 453.0, + 983.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 428.0, + 1292.0, + 428.0, + 1292.0, + 453.0, + 1189.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 467.0, + 1417.0, + 467.0, + 1417.0, + 494.0, + 1373.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 531.0, + 418.0, + 531.0, + 418.0, + 546.0, + 397.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 532.0, + 472.0, + 532.0, + 472.0, + 550.0, + 447.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 534.0, + 619.0, + 534.0, + 619.0, + 547.0, + 601.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 531.0, + 674.0, + 531.0, + 674.0, + 553.0, + 644.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 532.0, + 875.0, + 532.0, + 875.0, + 555.0, + 851.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 531.0, + 1075.0, + 531.0, + 1075.0, + 555.0, + 1049.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 532.0, + 1274.0, + 532.0, + 1274.0, + 553.0, + 1251.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 550.0, + 465.0, + 550.0, + 465.0, + 580.0, + 406.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 552.0, + 665.0, + 552.0, + 665.0, + 580.0, + 633.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 567.0, + 833.0, + 567.0, + 833.0, + 577.0, + 818.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 556.0, + 865.0, + 556.0, + 865.0, + 574.0, + 838.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 554.0, + 1067.0, + 554.0, + 1067.0, + 576.0, + 1037.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 552.0, + 1241.0, + 552.0, + 1241.0, + 583.0, + 1210.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 548.0, + 1417.0, + 548.0, + 1417.0, + 578.0, + 1371.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1379.0, + 633.0, + 1413.0, + 633.0, + 1413.0, + 650.0, + 1379.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 553.0, + 1261.0, + 553.0, + 1261.0, + 575.5, + 1249.0, + 575.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 551.5, + 648.0, + 551.5, + 648.0, + 580.5, + 602.0, + 580.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 560.5, + 1047.0, + 560.5, + 1047.0, + 575.5, + 1007.0, + 575.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 671.0, + 1405.0, + 671.0, + 1405.0, + 710.0, + 294.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 701.0, + 1405.0, + 701.0, + 1405.0, + 743.0, + 294.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 736.0, + 678.0, + 736.0, + 678.0, + 769.0, + 298.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 819.0, + 537.0, + 819.0, + 537.0, + 870.0, + 289.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2089.0, + 839.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 896.0, + 1401.0, + 896.0, + 1401.0, + 928.0, + 296.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 925.0, + 1409.0, + 925.0, + 1409.0, + 959.0, + 293.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 955.0, + 1322.0, + 955.0, + 1322.0, + 991.0, + 294.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 955.0, + 1402.0, + 955.0, + 1402.0, + 991.0, + 1373.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 986.0, + 1406.0, + 986.0, + 1406.0, + 1021.0, + 294.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1016.0, + 1406.0, + 1016.0, + 1406.0, + 1052.0, + 294.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1048.0, + 1404.0, + 1048.0, + 1404.0, + 1080.0, + 293.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1077.0, + 1406.0, + 1077.0, + 1406.0, + 1113.0, + 294.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1108.0, + 1408.0, + 1108.0, + 1408.0, + 1143.0, + 294.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1137.0, + 1135.0, + 1137.0, + 1135.0, + 1172.0, + 293.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 1137.0, + 1406.0, + 1137.0, + 1406.0, + 1172.0, + 1239.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1167.0, + 868.0, + 1167.0, + 868.0, + 1203.0, + 294.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1167.0, + 1405.0, + 1167.0, + 1405.0, + 1203.0, + 920.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1195.0, + 1404.0, + 1195.0, + 1404.0, + 1234.0, + 293.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1229.0, + 915.0, + 1229.0, + 915.0, + 1265.0, + 296.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1229.0, + 1405.0, + 1229.0, + 1405.0, + 1265.0, + 950.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1260.0, + 1016.0, + 1260.0, + 1016.0, + 1295.0, + 294.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1260.0, + 1404.0, + 1260.0, + 1404.0, + 1295.0, + 1176.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1287.0, + 1218.0, + 1287.0, + 1218.0, + 1327.0, + 293.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1355.0, + 1406.0, + 1355.0, + 1406.0, + 1391.0, + 295.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1386.0, + 1405.0, + 1386.0, + 1405.0, + 1421.0, + 295.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1416.0, + 893.0, + 1416.0, + 893.0, + 1451.0, + 294.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1416.0, + 1405.0, + 1416.0, + 1405.0, + 1451.0, + 925.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1446.0, + 1405.0, + 1446.0, + 1405.0, + 1482.0, + 295.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1474.0, + 698.0, + 1474.0, + 698.0, + 1513.0, + 294.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1474.0, + 1098.0, + 1474.0, + 1098.0, + 1513.0, + 749.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1474.0, + 1409.0, + 1474.0, + 1409.0, + 1513.0, + 1144.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1504.0, + 767.0, + 1504.0, + 767.0, + 1542.0, + 294.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1573.0, + 1407.0, + 1573.0, + 1407.0, + 1611.0, + 292.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1395.0, + 1605.0, + 1395.0, + 1643.0, + 294.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1637.0, + 716.0, + 1637.0, + 716.0, + 1670.0, + 295.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1637.0, + 1405.0, + 1637.0, + 1405.0, + 1670.0, + 810.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1667.0, + 1405.0, + 1667.0, + 1405.0, + 1700.0, + 295.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 1404.0, + 1698.0, + 1404.0, + 1731.0, + 295.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1725.0, + 1404.0, + 1725.0, + 1404.0, + 1761.0, + 292.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1754.0, + 1405.0, + 1754.0, + 1405.0, + 1794.0, + 292.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1785.0, + 1131.0, + 1785.0, + 1131.0, + 1825.0, + 292.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1891.0, + 293.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1919.0, + 295.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 507.0, + 1912.0, + 507.0, + 1953.0, + 293.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1912.0, + 1407.0, + 1912.0, + 1407.0, + 1953.0, + 538.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1946.0, + 1407.0, + 1946.0, + 1407.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1977.0, + 1403.0, + 1977.0, + 1403.0, + 2010.0, + 295.0, + 2010.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 298, + 399, + 1403, + 399, + 1403, + 869, + 298, + 869 + ], + "score": 0.985, + "html": "
PrroiiiieAUROC(H)AUROC(MD)AccuracyH/HtMD/MDtECENLL
Deep ensemble 40.958±0.0010.975±0.00191.122±0.0136.257±0.0056.394±0.0010.012±0.0010.129±0.001
SVGD [510.960±0.0010.973±0.00191.134±0.0246.315±0.0196.395±0.0180.014±0.0010.127±0.001
f-SVGD [760.956±0.0010.975±0.00189.884±0.0155.652±0.0096.531±0.0050.013±0.0010.150±0.001
hyper-DE [80]0.968±0.0010.981±0.00191.160±0.0076.682±0.0657.059±0.1520.014±0.0010.128±0.001
kde-WGD (ours)0.960±0.0010.970±0.00191.238±0.0196.587±0.0196.379±0.0180.014±0.0010.128±0.001
sge-WGD (ours)0.960±0.0010.970±0.00191.312±0.0166.562±0.0076.363±0.0090.012±0.0010.128±0.001
ssge-WGD (ours)0.968±0.0010.979±0.00191.198±0.0246.522±0.0096.610±0.0120.012±0.0010.130±0.001
kde-fWGD (ours)0.971±0.0010.980±0.00191.260±0.0117.079±0.0166.887±0.0150.015±0.0010.125±0.001
sge-fWGD (ours)0.969±0.0010.978±0.00191.192±0.0137.076±0.0046.900±0.0050.015±0.0010.125±0.001
ssge-fWGD (ours)0.971±0.0010.980±0.00191.240±0.0227.129±0.0066.951±0.0050.016±0.0010.124±0.001
CITIIIODeep ensemble 420.843±0.0040.736±0.00585,552±0.0762.244±0.0061.667±0.0080.049±0.0010.277±0.001
SVGD [510.825±0.0010.710±0.00285.142±0.0172.106±0.0031.567±0.0040.052±0.0010.287±0.001
fSVGD 回0.783±0.0010.712±0.00184.510±0.0311.968±0.0041.624±0.0030.049±0.0010.292±0.001
hyper-DE [80]0.789±0.0010.743±0.00184.743±0.0111.951±0.0101.690±0.0150.046±0.0010.288±0.001
kde-WGD (ours)0.838±0.0010.735±0.00485.904±0.0302.205±0.0031.661±0.0080.053±0.0010.276±0.001
sge-WGD (ours)0.837±0.0030.725±0.00485.792±0.0352.214±0.0101.634±0.0040.051±0.0010.275±0.001
ssge-WGD (ours)0.832±0.0030.731±0.00585.638±0.0382.182±0.0151.655±0.0010.049±0.0010.276±0.001
kde-fWGD (ours)0.791±0.0020.758±0.00284.888±0.0301.970±0.0041.749±0.0050.044±0.0010.282±0.001
sge-fWGD (ours)0.795±0.0010.754±0.00284.766±0.0601.984±0.0031.729±0.0020.047±0.0010.288±0.001
ssge-fWGD (ours)0.792±0.0020.752±0.00284.762±0.0341.970±0.0061.723±0.0050.046±0.0010.286±0.001
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1181, + 1405, + 1181, + 1405, + 1606, + 297, + 1606 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1643, + 1404, + 1643, + 1404, + 1827, + 298, + 1827 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 930, + 1405, + 930, + 1405, + 1144, + 297, + 1144 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1943, + 1398, + 1943, + 1398, + 2007, + 299, + 2007 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 298, + 1872, + 548, + 1872, + 548, + 1910, + 298, + 1910 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2084, + 840, + 2084 + ], + "score": 0.798 + }, + { + "category_id": 6, + "poly": [ + 296, + 194, + 1403, + 194, + 1403, + 380, + 296, + 380 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 296, + 194, + 1403, + 194, + 1403, + 380, + 296, + 380 + ], + "score": 0.409 + }, + { + "category_id": 13, + "poly": [ + 964, + 227, + 1057, + 227, + 1057, + 261, + 964, + 261 + ], + "score": 0.93, + "latex": "\\mathbf { H _ { o } / H _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 715, + 258, + 867, + 258, + 867, + 290, + 715, + 290 + ], + "score": 0.9, + "latex": "\\mathbf { M D _ { o } / M D _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1211, + 1257, + 1211, + 1257, + 1243, + 1208, + 1243 + ], + "score": 0.78, + "latex": "\\bar { \\mathbb { E } 3 } \\mathbb { I }" + }, + { + "category_id": 13, + "poly": [ + 754, + 1211, + 803, + 1211, + 803, + 1244, + 754, + 1244 + ], + "score": 0.78, + "latex": "\\dot { \\left[ \\left| 8 3 \\right| \\right] }" + }, + { + "category_id": 13, + "poly": [ + 1131, + 1052, + 1222, + 1052, + 1222, + 1085, + 1131, + 1085 + ], + "score": 0.56, + "latex": "\\mathbb { H 9 } \\mathbb { L O }" + }, + { + "category_id": 13, + "poly": [ + 947, + 1242, + 973, + 1242, + 973, + 1279, + 947, + 1279 + ], + "score": 0.55, + "latex": "^ 1 { } _ { . }" + }, + { + "category_id": 13, + "poly": [ + 461, + 1673, + 509, + 1673, + 509, + 1706, + 461, + 1706 + ], + "score": 0.46, + "latex": "\\mathbb { \\left[ 6 1 \\right] }" + }, + { + "category_id": 13, + "poly": [ + 715, + 1544, + 764, + 1544, + 764, + 1577, + 715, + 1577 + ], + "score": 0.44, + "latex": "[ \\textcircled { 8 0 } ]" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1642, + 1082, + 1642, + 1082, + 1675, + 1032, + 1675 + ], + "score": 0.4, + "latex": "\\mathbb { \\left| \\left[ 2 9 \\right] \\right| }" + }, + { + "category_id": 13, + "poly": [ + 802, + 197, + 840, + 197, + 840, + 227, + 802, + 227 + ], + "score": 0.29, + "latex": "\\mathbf { ( H ) }" + }, + { + "category_id": 13, + "poly": [ + 1123, + 1514, + 1171, + 1514, + 1171, + 1547, + 1123, + 1547 + ], + "score": 0.27, + "latex": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1871.0, + 551.0, + 1871.0, + 551.0, + 1915.0, + 292.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 860.0, + 2060.0, + 860.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 194.0, + 801.0, + 194.0, + 801.0, + 232.0, + 293.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 194.0, + 1405.0, + 194.0, + 1405.0, + 232.0, + 841.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 963.0, + 226.0, + 963.0, + 264.0, + 293.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 264.0, + 1058.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 259.0, + 714.0, + 259.0, + 714.0, + 291.0, + 296.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 259.0, + 1404.0, + 259.0, + 1404.0, + 291.0, + 868.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 1404.0, + 287.0, + 1404.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 314.0, + 1408.0, + 314.0, + 1408.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 347.0, + 1098.0, + 347.0, + 1098.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1180.0, + 1405.0, + 1180.0, + 1405.0, + 1216.0, + 295.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 753.0, + 1211.0, + 753.0, + 1246.0, + 294.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1211.0, + 1207.0, + 1211.0, + 1207.0, + 1246.0, + 804.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1211.0, + 1408.0, + 1211.0, + 1408.0, + 1246.0, + 1258.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1242.0, + 946.0, + 1242.0, + 946.0, + 1277.0, + 295.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1242.0, + 1405.0, + 1242.0, + 1405.0, + 1277.0, + 974.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1272.0, + 1406.0, + 1272.0, + 1406.0, + 1308.0, + 295.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1304.0, + 1406.0, + 1304.0, + 1406.0, + 1337.0, + 294.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1333.0, + 1405.0, + 1333.0, + 1405.0, + 1368.0, + 294.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1363.0, + 1405.0, + 1363.0, + 1405.0, + 1399.0, + 295.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1395.0, + 1403.0, + 1395.0, + 1403.0, + 1430.0, + 295.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1422.0, + 1405.0, + 1422.0, + 1405.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1454.0, + 1407.0, + 1454.0, + 1407.0, + 1489.0, + 294.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1520.0, + 292.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1510.0, + 1122.0, + 1510.0, + 1122.0, + 1554.0, + 291.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1510.0, + 1405.0, + 1510.0, + 1405.0, + 1554.0, + 1172.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1544.0, + 714.0, + 1544.0, + 714.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1544.0, + 1405.0, + 1544.0, + 1405.0, + 1582.0, + 765.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1578.0, + 718.0, + 1578.0, + 718.0, + 1609.0, + 296.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1642.0, + 1031.0, + 1642.0, + 1031.0, + 1678.0, + 296.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 1642.0, + 1405.0, + 1642.0, + 1405.0, + 1678.0, + 1083.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1674.0, + 460.0, + 1674.0, + 460.0, + 1710.0, + 296.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 1674.0, + 1406.0, + 1674.0, + 1406.0, + 1710.0, + 510.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1706.0, + 1404.0, + 1706.0, + 1404.0, + 1738.0, + 296.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1735.0, + 1405.0, + 1735.0, + 1405.0, + 1771.0, + 294.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1763.0, + 1408.0, + 1763.0, + 1408.0, + 1802.0, + 294.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1793.0, + 1039.0, + 1793.0, + 1039.0, + 1833.0, + 295.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 930.0, + 1408.0, + 930.0, + 1408.0, + 966.0, + 292.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 962.0, + 1407.0, + 962.0, + 1407.0, + 998.0, + 292.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 993.0, + 1405.0, + 993.0, + 1405.0, + 1027.0, + 295.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1022.0, + 1406.0, + 1022.0, + 1406.0, + 1057.0, + 292.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1052.0, + 1130.0, + 1052.0, + 1130.0, + 1090.0, + 295.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 1052.0, + 1405.0, + 1052.0, + 1405.0, + 1090.0, + 1223.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1082.0, + 1406.0, + 1082.0, + 1406.0, + 1119.0, + 291.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1112.0, + 479.0, + 1112.0, + 479.0, + 1146.0, + 295.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1404.0, + 1939.0, + 1404.0, + 1981.0, + 293.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2010.0, + 293.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 194.0, + 801.0, + 194.0, + 801.0, + 232.0, + 293.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 194.0, + 1405.0, + 194.0, + 1405.0, + 232.0, + 841.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 963.0, + 226.0, + 963.0, + 264.0, + 293.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 264.0, + 1058.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 259.0, + 714.0, + 259.0, + 714.0, + 291.0, + 296.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 259.0, + 1404.0, + 259.0, + 1404.0, + 291.0, + 868.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 287.0, + 1404.0, + 287.0, + 1404.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 314.0, + 1408.0, + 314.0, + 1408.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 347.0, + 1098.0, + 347.0, + 1098.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 202, + 1404, + 202, + 1404, + 689, + 298, + 689 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 796, + 1404, + 796, + 1404, + 1040, + 298, + 1040 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1054, + 1404, + 1054, + 1404, + 1207, + 299, + 1207 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1296, + 1405, + 1296, + 1405, + 1480, + 298, + 1480 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 308, + 1572, + 1407, + 1572, + 1407, + 2008, + 308, + 2008 + ], + "score": 0.955 + }, + { + "category_id": 0, + "poly": [ + 300, + 1241, + 517, + 1241, + 517, + 1273, + 300, + 1273 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 1522, + 455, + 1522, + 455, + 1557, + 299, + 1557 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 298, + 729, + 509, + 729, + 509, + 767, + 298, + 767 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.838 + }, + { + "category_id": 13, + "poly": [ + 1264, + 626, + 1314, + 626, + 1314, + 659, + 1264, + 659 + ], + "score": 0.83, + "latex": "\\mathbb { \\ m }" + }, + { + "category_id": 13, + "poly": [ + 1169, + 566, + 1218, + 566, + 1218, + 598, + 1169, + 598 + ], + "score": 0.8, + "latex": "\\bar { \\| 5 3 \\| }" + }, + { + "category_id": 13, + "poly": [ + 1035, + 323, + 1084, + 323, + 1084, + 356, + 1035, + 356 + ], + "score": 0.71, + "latex": "\\mathbb { \\left. \\overline { { \\boldsymbol { \\mathscr { Q } } \\boldsymbol { \\cdot } \\boldsymbol { \\cdot } } } \\right. }" + }, + { + "category_id": 13, + "poly": [ + 1000, + 535, + 1133, + 535, + 1133, + 568, + 1000, + 568 + ], + "score": 0.64, + "latex": "\\boxed { 6 2 } , \\boxed { 1 2 } , \\boxed { 2 7 }" + }, + { + "category_id": 13, + "poly": [ + 753, + 1113, + 803, + 1113, + 803, + 1146, + 753, + 1146 + ], + "score": 0.59, + "latex": "\\pmb { \\Vert 3 7 } \\Vert" + }, + { + "category_id": 13, + "poly": [ + 298, + 565, + 347, + 565, + 347, + 598, + 298, + 598 + ], + "score": 0.48, + "latex": "\\lVert 8 2 \\rVert" + }, + { + "category_id": 13, + "poly": [ + 1321, + 504, + 1397, + 504, + 1397, + 538, + 1321, + 538 + ], + "score": 0.47, + "latex": "[ \\sqrt { 5 6 } , \\sqrt ] { 8 } ]" + }, + { + "category_id": 13, + "poly": [ + 614, + 535, + 663, + 535, + 663, + 567, + 614, + 567 + ], + "score": 0.39, + "latex": "\\dot { \\bigtriangledown } 5 7 \\big |" + }, + { + "category_id": 13, + "poly": [ + 819, + 233, + 868, + 233, + 868, + 265, + 819, + 265 + ], + "score": 0.35, + "latex": "\\dot { \\left. \\overline { { 4 8 } } \\right. }" + }, + { + "category_id": 13, + "poly": [ + 1172, + 353, + 1263, + 353, + 1263, + 385, + 1172, + 385 + ], + "score": 0.35, + "latex": "\\pm \\pm \\infty" + }, + { + "category_id": 13, + "poly": [ + 297, + 626, + 346, + 626, + 346, + 659, + 297, + 659 + ], + "score": 0.31, + "latex": "\\left[ \\left[ 3 6 \\right] \\right]" + }, + { + "category_id": 13, + "poly": [ + 1349, + 1113, + 1398, + 1113, + 1398, + 1146, + 1349, + 1146 + ], + "score": 0.26, + "latex": "\\boldsymbol { \\left[ \\left[ 3 5 \\right] \\right] }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1236.0, + 522.0, + 1236.0, + 522.0, + 1280.0, + 295.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1518.0, + 460.0, + 1518.0, + 460.0, + 1563.0, + 295.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 725.0, + 514.0, + 725.0, + 514.0, + 774.0, + 292.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 201.0, + 1408.0, + 201.0, + 1408.0, + 239.0, + 293.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 233.0, + 818.0, + 233.0, + 818.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 233.0, + 1404.0, + 233.0, + 1404.0, + 267.0, + 869.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 294.0, + 1405.0, + 294.0, + 1405.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 323.0, + 1034.0, + 323.0, + 1034.0, + 357.0, + 297.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 323.0, + 1404.0, + 323.0, + 1404.0, + 357.0, + 1085.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 353.0, + 1171.0, + 353.0, + 1171.0, + 389.0, + 292.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 353.0, + 1406.0, + 353.0, + 1406.0, + 389.0, + 1264.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 382.0, + 1407.0, + 382.0, + 1407.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 415.0, + 1404.0, + 415.0, + 1404.0, + 449.0, + 294.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 444.0, + 1402.0, + 444.0, + 1402.0, + 479.0, + 294.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 474.0, + 1403.0, + 474.0, + 1403.0, + 509.0, + 293.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 505.0, + 1320.0, + 505.0, + 1320.0, + 540.0, + 293.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 505.0, + 1405.0, + 505.0, + 1405.0, + 540.0, + 1398.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 532.0, + 613.0, + 532.0, + 613.0, + 573.0, + 292.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 532.0, + 999.0, + 532.0, + 999.0, + 573.0, + 664.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 532.0, + 1406.0, + 532.0, + 1406.0, + 573.0, + 1134.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 565.0, + 1168.0, + 565.0, + 1168.0, + 600.0, + 348.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 565.0, + 1405.0, + 565.0, + 1405.0, + 600.0, + 1219.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 595.0, + 1404.0, + 595.0, + 1404.0, + 633.0, + 295.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 626.0, + 1263.0, + 626.0, + 1263.0, + 661.0, + 347.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 626.0, + 1406.0, + 626.0, + 1406.0, + 661.0, + 1315.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 658.0, + 1051.0, + 658.0, + 1051.0, + 689.0, + 297.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 794.0, + 1406.0, + 794.0, + 1406.0, + 832.0, + 293.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 828.0, + 1402.0, + 828.0, + 1402.0, + 858.0, + 296.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 857.0, + 1406.0, + 857.0, + 1406.0, + 894.0, + 293.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 886.0, + 1405.0, + 886.0, + 1405.0, + 924.0, + 292.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 918.0, + 1404.0, + 918.0, + 1404.0, + 951.0, + 294.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 948.0, + 1405.0, + 948.0, + 1405.0, + 984.0, + 292.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 979.0, + 1404.0, + 979.0, + 1404.0, + 1012.0, + 294.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1010.0, + 556.0, + 1010.0, + 556.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1054.0, + 1403.0, + 1054.0, + 1403.0, + 1087.0, + 295.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1084.0, + 1404.0, + 1084.0, + 1404.0, + 1117.0, + 297.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1113.0, + 752.0, + 1113.0, + 752.0, + 1150.0, + 294.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1113.0, + 1348.0, + 1113.0, + 1348.0, + 1150.0, + 804.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1113.0, + 1405.0, + 1113.0, + 1405.0, + 1150.0, + 1399.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1144.0, + 1406.0, + 1144.0, + 1406.0, + 1181.0, + 295.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1175.0, + 884.0, + 1175.0, + 884.0, + 1208.0, + 294.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1295.0, + 1405.0, + 1295.0, + 1405.0, + 1331.0, + 296.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1327.0, + 1402.0, + 1327.0, + 1402.0, + 1359.0, + 296.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1355.0, + 1403.0, + 1355.0, + 1403.0, + 1392.0, + 293.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1387.0, + 1408.0, + 1387.0, + 1408.0, + 1422.0, + 295.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1416.0, + 1406.0, + 1416.0, + 1406.0, + 1454.0, + 293.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1449.0, + 474.0, + 1449.0, + 474.0, + 1482.0, + 295.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1571.0, + 1408.0, + 1571.0, + 1408.0, + 1608.0, + 307.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1602.0, + 1406.0, + 1602.0, + 1406.0, + 1638.0, + 355.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1634.0, + 714.0, + 1634.0, + 714.0, + 1667.0, + 354.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1679.0, + 1407.0, + 1679.0, + 1407.0, + 1717.0, + 307.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1710.0, + 1408.0, + 1710.0, + 1408.0, + 1745.0, + 353.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1742.0, + 561.0, + 1742.0, + 561.0, + 1775.0, + 355.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1789.0, + 1405.0, + 1789.0, + 1405.0, + 1825.0, + 309.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1821.0, + 1227.0, + 1821.0, + 1227.0, + 1854.0, + 354.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1867.0, + 1405.0, + 1867.0, + 1405.0, + 1903.0, + 309.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1897.0, + 1012.0, + 1897.0, + 1012.0, + 1932.0, + 353.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1985.0, + 307.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1976.0, + 1031.0, + 1976.0, + 1031.0, + 2009.0, + 352.0, + 2009.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 835, + 2061, + 863, + 2061, + 863, + 2086, + 835, + 2086 + ], + "score": 0.801 + }, + { + "category_id": 1, + "poly": [ + 297, + 1376, + 1402, + 1376, + 1402, + 1439, + 297, + 1439 + ], + "score": 0.577 + }, + { + "category_id": 1, + "poly": [ + 297, + 1094, + 1400, + 1094, + 1400, + 1161, + 297, + 1161 + ], + "score": 0.577 + }, + { + "category_id": 1, + "poly": [ + 298, + 1011, + 1400, + 1011, + 1400, + 1077, + 298, + 1077 + ], + "score": 0.535 + }, + { + "category_id": 1, + "poly": [ + 299, + 1178, + 1400, + 1178, + 1400, + 1273, + 299, + 1273 + ], + "score": 0.526 + }, + { + "category_id": 1, + "poly": [ + 298, + 1292, + 1400, + 1292, + 1400, + 1358, + 298, + 1358 + ], + "score": 0.523 + }, + { + "category_id": 1, + "poly": [ + 296, + 928, + 1399, + 928, + 1399, + 994, + 296, + 994 + ], + "score": 0.516 + }, + { + "category_id": 1, + "poly": [ + 297, + 1602, + 1402, + 1602, + 1402, + 1698, + 297, + 1698 + ], + "score": 0.503 + }, + { + "category_id": 1, + "poly": [ + 296, + 844, + 1399, + 844, + 1399, + 911, + 296, + 911 + ], + "score": 0.503 + }, + { + "category_id": 1, + "poly": [ + 299, + 732, + 1401, + 732, + 1401, + 826, + 299, + 826 + ], + "score": 0.479 + }, + { + "category_id": 1, + "poly": [ + 297, + 535, + 1403, + 535, + 1403, + 628, + 297, + 628 + ], + "score": 0.476 + }, + { + "category_id": 1, + "poly": [ + 295, + 648, + 1403, + 648, + 1403, + 713, + 295, + 713 + ], + "score": 0.471 + }, + { + "category_id": 1, + "poly": [ + 303, + 255, + 1401, + 255, + 1401, + 321, + 303, + 321 + ], + "score": 0.46 + }, + { + "category_id": 1, + "poly": [ + 297, + 1944, + 1397, + 1944, + 1397, + 2008, + 297, + 2008 + ], + "score": 0.448 + }, + { + "category_id": 1, + "poly": [ + 298, + 1716, + 1402, + 1716, + 1402, + 1811, + 298, + 1811 + ], + "score": 0.447 + }, + { + "category_id": 1, + "poly": [ + 304, + 338, + 1402, + 338, + 1402, + 433, + 304, + 433 + ], + "score": 0.439 + }, + { + "category_id": 1, + "poly": [ + 303, + 452, + 1403, + 452, + 1403, + 517, + 303, + 517 + ], + "score": 0.437 + }, + { + "category_id": 1, + "poly": [ + 299, + 1830, + 1399, + 1830, + 1399, + 1925, + 299, + 1925 + ], + "score": 0.416 + }, + { + "category_id": 1, + "poly": [ + 308, + 202, + 1204, + 202, + 1204, + 238, + 308, + 238 + ], + "score": 0.392 + }, + { + "category_id": 1, + "poly": [ + 298, + 1458, + 1407, + 1458, + 1407, + 1582, + 298, + 1582 + ], + "score": 0.391 + }, + { + "category_id": 2, + "poly": [ + 835, + 2061, + 863, + 2061, + 863, + 2086, + 835, + 2086 + ], + "score": 0.13 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2098.0, + 832.0, + 2098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 868.0, + 2058.0, + 868.0, + 2098.0, + 832.0, + 2098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1373.0, + 1407.0, + 1373.0, + 1407.0, + 1413.0, + 295.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1404.0, + 432.0, + 1404.0, + 432.0, + 1442.0, + 351.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1095.0, + 1404.0, + 1095.0, + 1404.0, + 1131.0, + 296.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1126.0, + 907.0, + 1126.0, + 907.0, + 1161.0, + 355.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1010.0, + 1406.0, + 1010.0, + 1406.0, + 1048.0, + 293.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1045.0, + 940.0, + 1045.0, + 940.0, + 1075.0, + 354.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1177.0, + 1405.0, + 1177.0, + 1405.0, + 1217.0, + 293.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1211.0, + 1404.0, + 1211.0, + 1404.0, + 1245.0, + 355.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1240.0, + 1033.0, + 1240.0, + 1033.0, + 1274.0, + 355.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1290.0, + 1404.0, + 1290.0, + 1404.0, + 1332.0, + 293.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1325.0, + 949.0, + 1325.0, + 949.0, + 1357.0, + 353.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 930.0, + 1404.0, + 930.0, + 1404.0, + 966.0, + 297.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 960.0, + 1246.0, + 960.0, + 1246.0, + 995.0, + 354.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1602.0, + 1408.0, + 1602.0, + 1408.0, + 1640.0, + 295.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1633.0, + 1404.0, + 1633.0, + 1404.0, + 1670.0, + 352.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1665.0, + 684.0, + 1665.0, + 684.0, + 1700.0, + 356.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 845.0, + 1403.0, + 845.0, + 1403.0, + 881.0, + 296.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 878.0, + 739.0, + 878.0, + 739.0, + 909.0, + 351.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 732.0, + 1407.0, + 732.0, + 1407.0, + 769.0, + 294.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 760.0, + 1405.0, + 760.0, + 1405.0, + 802.0, + 352.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 794.0, + 617.0, + 794.0, + 617.0, + 828.0, + 355.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 533.0, + 1408.0, + 533.0, + 1408.0, + 572.0, + 292.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 565.0, + 1408.0, + 565.0, + 1408.0, + 602.0, + 354.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 592.0, + 430.0, + 592.0, + 430.0, + 633.0, + 353.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 648.0, + 1405.0, + 648.0, + 1405.0, + 684.0, + 294.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 679.0, + 1000.0, + 679.0, + 1000.0, + 714.0, + 353.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 253.0, + 1406.0, + 253.0, + 1406.0, + 295.0, + 307.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 285.0, + 1249.0, + 285.0, + 1249.0, + 320.0, + 354.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1402.0, + 1943.0, + 1402.0, + 1981.0, + 294.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1975.0, + 809.0, + 1975.0, + 809.0, + 2009.0, + 353.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1716.0, + 1407.0, + 1716.0, + 1407.0, + 1751.0, + 294.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1746.0, + 1407.0, + 1746.0, + 1407.0, + 1785.0, + 352.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1779.0, + 643.0, + 1779.0, + 643.0, + 1809.0, + 354.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 337.0, + 1407.0, + 337.0, + 1407.0, + 377.0, + 308.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 368.0, + 1403.0, + 368.0, + 1403.0, + 406.0, + 354.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 400.0, + 827.0, + 400.0, + 827.0, + 434.0, + 355.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 452.0, + 1404.0, + 452.0, + 1404.0, + 488.0, + 309.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 482.0, + 1168.0, + 482.0, + 1168.0, + 517.0, + 353.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1829.0, + 1405.0, + 1829.0, + 1405.0, + 1865.0, + 294.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1858.0, + 1405.0, + 1858.0, + 1405.0, + 1898.0, + 353.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1892.0, + 643.0, + 1892.0, + 643.0, + 1926.0, + 355.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 200.0, + 1205.0, + 200.0, + 1205.0, + 239.0, + 309.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1458.0, + 1410.0, + 1458.0, + 1410.0, + 1497.0, + 293.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1485.0, + 1408.0, + 1485.0, + 1408.0, + 1529.0, + 350.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1519.0, + 1408.0, + 1519.0, + 1408.0, + 1557.0, + 351.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1549.0, + 431.0, + 1549.0, + 431.0, + 1585.0, + 351.0, + 1585.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.778 + }, + { + "category_id": 1, + "poly": [ + 288, + 97, + 1409, + 97, + 1409, + 2016, + 288, + 2016 + ], + "score": 0.578 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.182 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 201.0, + 1408.0, + 201.0, + 1408.0, + 240.0, + 293.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 232.0, + 1408.0, + 232.0, + 1408.0, + 272.0, + 355.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 264.0, + 1408.0, + 264.0, + 1408.0, + 304.0, + 352.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 293.0, + 761.0, + 293.0, + 761.0, + 331.0, + 350.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 348.0, + 1406.0, + 348.0, + 1406.0, + 381.0, + 295.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 377.0, + 1408.0, + 377.0, + 1408.0, + 411.0, + 355.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 407.0, + 834.0, + 407.0, + 834.0, + 447.0, + 352.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 459.0, + 1408.0, + 459.0, + 1408.0, + 499.0, + 293.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 489.0, + 1060.0, + 489.0, + 1060.0, + 526.0, + 350.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 541.0, + 1410.0, + 541.0, + 1410.0, + 581.0, + 295.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 575.0, + 1404.0, + 575.0, + 1404.0, + 609.0, + 355.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 602.0, + 1141.0, + 602.0, + 1141.0, + 640.0, + 350.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 653.0, + 1406.0, + 653.0, + 1406.0, + 695.0, + 291.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 688.0, + 1406.0, + 688.0, + 1406.0, + 722.0, + 352.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 716.0, + 608.0, + 716.0, + 608.0, + 751.0, + 350.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 766.0, + 1408.0, + 766.0, + 1408.0, + 806.0, + 293.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 800.0, + 1046.0, + 800.0, + 1046.0, + 840.0, + 350.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 852.0, + 1406.0, + 852.0, + 1406.0, + 886.0, + 297.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 884.0, + 1406.0, + 884.0, + 1406.0, + 918.0, + 357.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 911.0, + 1205.0, + 911.0, + 1205.0, + 951.0, + 352.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 964.0, + 1410.0, + 964.0, + 1410.0, + 1004.0, + 293.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 996.0, + 1313.0, + 996.0, + 1313.0, + 1035.0, + 352.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1046.0, + 1410.0, + 1046.0, + 1410.0, + 1086.0, + 295.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1078.0, + 1410.0, + 1078.0, + 1410.0, + 1120.0, + 355.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1107.0, + 747.0, + 1107.0, + 747.0, + 1145.0, + 350.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1160.0, + 1408.0, + 1160.0, + 1408.0, + 1200.0, + 293.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1191.0, + 1402.0, + 1191.0, + 1402.0, + 1225.0, + 355.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1221.0, + 742.0, + 1221.0, + 742.0, + 1258.0, + 350.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1273.0, + 1406.0, + 1273.0, + 1406.0, + 1313.0, + 295.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1303.0, + 1408.0, + 1303.0, + 1408.0, + 1343.0, + 352.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1334.0, + 812.0, + 1334.0, + 812.0, + 1374.0, + 352.0, + 1374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1389.0, + 1406.0, + 1389.0, + 1406.0, + 1422.0, + 295.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1416.0, + 1334.0, + 1416.0, + 1334.0, + 1454.0, + 350.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1469.0, + 1406.0, + 1469.0, + 1406.0, + 1509.0, + 293.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1500.0, + 1408.0, + 1500.0, + 1408.0, + 1538.0, + 350.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1530.0, + 1086.0, + 1530.0, + 1086.0, + 1570.0, + 352.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1580.0, + 1410.0, + 1580.0, + 1410.0, + 1620.0, + 295.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1612.0, + 1211.0, + 1612.0, + 1211.0, + 1650.0, + 350.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1664.0, + 1406.0, + 1664.0, + 1406.0, + 1704.0, + 295.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1698.0, + 628.0, + 1698.0, + 628.0, + 1732.0, + 355.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1748.0, + 1408.0, + 1748.0, + 1408.0, + 1788.0, + 293.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1776.0, + 1406.0, + 1776.0, + 1406.0, + 1820.0, + 350.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1810.0, + 572.0, + 1810.0, + 572.0, + 1845.0, + 350.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1860.0, + 1294.0, + 1860.0, + 1294.0, + 1898.0, + 291.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1408.0, + 1912.0, + 1408.0, + 1952.0, + 295.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1944.0, + 1408.0, + 1944.0, + 1408.0, + 1984.0, + 352.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1976.0, + 861.0, + 1976.0, + 861.0, + 2013.0, + 348.0, + 2013.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 293, + 126, + 1410, + 126, + 1410, + 2023, + 293, + 2023 + ], + "score": 0.678 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 205.0, + 1407.0, + 205.0, + 1407.0, + 238.0, + 296.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 225.0, + 430.0, + 225.0, + 430.0, + 275.0, + 351.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 284.0, + 1409.0, + 284.0, + 1409.0, + 324.0, + 294.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 315.0, + 1407.0, + 315.0, + 1407.0, + 355.0, + 351.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 344.0, + 646.0, + 344.0, + 646.0, + 382.0, + 351.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 394.0, + 1405.0, + 394.0, + 1405.0, + 434.0, + 296.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 428.0, + 1318.0, + 428.0, + 1318.0, + 465.0, + 351.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 480.0, + 1407.0, + 480.0, + 1407.0, + 519.0, + 294.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 509.0, + 644.0, + 509.0, + 644.0, + 548.0, + 353.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 561.0, + 1409.0, + 561.0, + 1409.0, + 600.0, + 294.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 594.0, + 1344.0, + 594.0, + 1344.0, + 627.0, + 355.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 644.0, + 1407.0, + 644.0, + 1407.0, + 683.0, + 294.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 671.0, + 1409.0, + 671.0, + 1409.0, + 715.0, + 349.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 704.0, + 841.0, + 704.0, + 841.0, + 744.0, + 351.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 756.0, + 1405.0, + 756.0, + 1405.0, + 790.0, + 296.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 785.0, + 1407.0, + 785.0, + 1407.0, + 825.0, + 353.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 814.0, + 1035.0, + 814.0, + 1035.0, + 854.0, + 353.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 871.0, + 1405.0, + 871.0, + 1405.0, + 904.0, + 298.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 900.0, + 862.0, + 900.0, + 862.0, + 937.0, + 349.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 948.0, + 1409.0, + 948.0, + 1409.0, + 991.0, + 289.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 981.0, + 1407.0, + 981.0, + 1407.0, + 1020.0, + 353.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1008.0, + 433.0, + 1008.0, + 433.0, + 1050.0, + 353.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1062.0, + 1407.0, + 1062.0, + 1407.0, + 1102.0, + 296.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1093.0, + 1314.0, + 1093.0, + 1314.0, + 1133.0, + 353.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1145.0, + 1407.0, + 1145.0, + 1407.0, + 1185.0, + 294.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1174.0, + 1409.0, + 1174.0, + 1409.0, + 1214.0, + 353.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1203.0, + 431.0, + 1203.0, + 431.0, + 1245.0, + 351.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1260.0, + 1405.0, + 1260.0, + 1405.0, + 1293.0, + 296.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1289.0, + 746.0, + 1289.0, + 746.0, + 1326.0, + 351.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1339.0, + 1405.0, + 1339.0, + 1405.0, + 1378.0, + 296.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 1370.0, + 741.0, + 1370.0, + 741.0, + 1407.0, + 348.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1418.0, + 1409.0, + 1418.0, + 1409.0, + 1463.0, + 292.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1453.0, + 1179.0, + 1453.0, + 1179.0, + 1493.0, + 353.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1503.0, + 1409.0, + 1503.0, + 1409.0, + 1543.0, + 296.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1534.0, + 1407.0, + 1534.0, + 1407.0, + 1574.0, + 353.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1568.0, + 828.0, + 1568.0, + 828.0, + 1601.0, + 357.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1615.0, + 1407.0, + 1615.0, + 1407.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1644.0, + 1405.0, + 1644.0, + 1405.0, + 1684.0, + 353.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1676.0, + 704.0, + 1676.0, + 704.0, + 1715.0, + 353.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 1065.0, + 1728.0, + 1065.0, + 1765.0, + 294.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1780.0, + 1409.0, + 1780.0, + 1409.0, + 1819.0, + 296.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1813.0, + 615.0, + 1813.0, + 615.0, + 1846.0, + 355.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1863.0, + 1409.0, + 1863.0, + 1409.0, + 1902.0, + 294.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1896.0, + 1192.0, + 1896.0, + 1192.0, + 1929.0, + 355.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1407.0, + 1944.0, + 1407.0, + 1984.0, + 296.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1977.0, + 1156.0, + 1977.0, + 1156.0, + 2011.0, + 353.0, + 2011.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 298, + 202, + 1406, + 202, + 1406, + 328, + 298, + 328 + ], + "score": 0.643 + }, + { + "category_id": 1, + "poly": [ + 291, + 343, + 1403, + 343, + 1403, + 406, + 291, + 406 + ], + "score": 0.616 + }, + { + "category_id": 1, + "poly": [ + 297, + 808, + 1403, + 808, + 1403, + 902, + 297, + 902 + ], + "score": 0.578 + }, + { + "category_id": 1, + "poly": [ + 296, + 646, + 1402, + 646, + 1402, + 711, + 296, + 711 + ], + "score": 0.566 + }, + { + "category_id": 1, + "poly": [ + 300, + 1607, + 1403, + 1607, + 1403, + 1701, + 300, + 1701 + ], + "score": 0.557 + }, + { + "category_id": 1, + "poly": [ + 297, + 919, + 1400, + 919, + 1400, + 985, + 297, + 985 + ], + "score": 0.555 + }, + { + "category_id": 1, + "poly": [ + 296, + 727, + 1402, + 727, + 1402, + 792, + 296, + 792 + ], + "score": 0.551 + }, + { + "category_id": 1, + "poly": [ + 297, + 1526, + 1403, + 1526, + 1403, + 1590, + 297, + 1590 + ], + "score": 0.539 + }, + { + "category_id": 1, + "poly": [ + 296, + 1000, + 1401, + 1000, + 1401, + 1065, + 296, + 1065 + ], + "score": 0.538 + }, + { + "category_id": 1, + "poly": [ + 303, + 1223, + 1402, + 1223, + 1402, + 1319, + 303, + 1319 + ], + "score": 0.537 + }, + { + "category_id": 1, + "poly": [ + 292, + 1718, + 1403, + 1718, + 1403, + 1783, + 292, + 1783 + ], + "score": 0.533 + }, + { + "category_id": 1, + "poly": [ + 295, + 566, + 1400, + 566, + 1400, + 630, + 295, + 630 + ], + "score": 0.527 + }, + { + "category_id": 1, + "poly": [ + 296, + 1801, + 1403, + 1801, + 1403, + 1897, + 296, + 1897 + ], + "score": 0.521 + }, + { + "category_id": 1, + "poly": [ + 301, + 1334, + 1401, + 1334, + 1401, + 1427, + 301, + 1427 + ], + "score": 0.496 + }, + { + "category_id": 1, + "poly": [ + 298, + 1445, + 1402, + 1445, + 1402, + 1509, + 298, + 1509 + ], + "score": 0.489 + }, + { + "category_id": 1, + "poly": [ + 299, + 425, + 1405, + 425, + 1405, + 548, + 299, + 548 + ], + "score": 0.375 + }, + { + "category_id": 1, + "poly": [ + 299, + 1914, + 1405, + 1914, + 1405, + 2007, + 299, + 2007 + ], + "score": 0.336 + }, + { + "category_id": 1, + "poly": [ + 298, + 1081, + 1406, + 1081, + 1406, + 1205, + 298, + 1205 + ], + "score": 0.305 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2095.0, + 832.0, + 2095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 200.0, + 1407.0, + 200.0, + 1407.0, + 239.0, + 292.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 232.0, + 1408.0, + 232.0, + 1408.0, + 267.0, + 352.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 263.0, + 1407.0, + 263.0, + 1407.0, + 299.0, + 354.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 296.0, + 571.0, + 296.0, + 571.0, + 330.0, + 354.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 340.0, + 1406.0, + 340.0, + 1406.0, + 380.0, + 294.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 366.0, + 432.0, + 366.0, + 432.0, + 409.0, + 351.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 840.0, + 1408.0, + 840.0, + 1408.0, + 874.0, + 355.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 867.0, + 430.0, + 867.0, + 430.0, + 904.0, + 351.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 646.0, + 1404.0, + 646.0, + 1404.0, + 682.0, + 298.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 677.0, + 1204.0, + 677.0, + 1204.0, + 712.0, + 354.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1607.0, + 1407.0, + 1607.0, + 1407.0, + 1644.0, + 294.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1635.0, + 1405.0, + 1635.0, + 1405.0, + 1676.0, + 352.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1669.0, + 538.0, + 1669.0, + 538.0, + 1700.0, + 357.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 920.0, + 1404.0, + 920.0, + 1404.0, + 956.0, + 296.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 949.0, + 1183.0, + 949.0, + 1183.0, + 984.0, + 353.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 726.0, + 1403.0, + 726.0, + 1403.0, + 762.0, + 296.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 760.0, + 992.0, + 760.0, + 992.0, + 793.0, + 353.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1527.0, + 1405.0, + 1527.0, + 1405.0, + 1563.0, + 296.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1557.0, + 1105.0, + 1557.0, + 1105.0, + 1591.0, + 353.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 999.0, + 1405.0, + 999.0, + 1405.0, + 1039.0, + 296.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1031.0, + 1003.0, + 1031.0, + 1003.0, + 1066.0, + 353.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1222.0, + 1406.0, + 1222.0, + 1406.0, + 1259.0, + 294.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1253.0, + 1407.0, + 1253.0, + 1407.0, + 1291.0, + 353.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1283.0, + 894.0, + 1283.0, + 894.0, + 1319.0, + 355.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1718.0, + 1405.0, + 1718.0, + 1405.0, + 1754.0, + 296.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1748.0, + 1165.0, + 1748.0, + 1165.0, + 1784.0, + 353.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 563.0, + 1407.0, + 563.0, + 1407.0, + 604.0, + 294.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 595.0, + 903.0, + 595.0, + 903.0, + 631.0, + 353.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1800.0, + 1408.0, + 1800.0, + 1408.0, + 1838.0, + 294.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1834.0, + 1405.0, + 1834.0, + 1405.0, + 1869.0, + 354.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1862.0, + 1340.0, + 1862.0, + 1340.0, + 1899.0, + 355.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1333.0, + 1406.0, + 1333.0, + 1406.0, + 1371.0, + 293.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1364.0, + 1406.0, + 1364.0, + 1406.0, + 1402.0, + 351.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1395.0, + 563.0, + 1395.0, + 563.0, + 1426.0, + 355.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1444.0, + 1407.0, + 1444.0, + 1407.0, + 1482.0, + 293.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1474.0, + 1175.0, + 1474.0, + 1175.0, + 1510.0, + 353.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 424.0, + 1406.0, + 424.0, + 1406.0, + 460.0, + 295.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 455.0, + 1406.0, + 455.0, + 1406.0, + 491.0, + 355.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 487.0, + 1405.0, + 487.0, + 1405.0, + 523.0, + 355.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 517.0, + 892.0, + 517.0, + 892.0, + 549.0, + 353.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1406.0, + 1910.0, + 1406.0, + 1952.0, + 293.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1983.0, + 352.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1975.0, + 525.0, + 1975.0, + 525.0, + 2010.0, + 355.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1081.0, + 1406.0, + 1081.0, + 1406.0, + 1116.0, + 293.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1113.0, + 1404.0, + 1113.0, + 1404.0, + 1145.0, + 354.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1144.0, + 1406.0, + 1144.0, + 1406.0, + 1180.0, + 354.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1175.0, + 1000.0, + 1175.0, + 1000.0, + 1206.0, + 354.0, + 1206.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 202, + 1410, + 202, + 1410, + 598, + 295, + 598 + ], + "score": 0.793 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.773 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.197 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 202.0, + 1406.0, + 202.0, + 1406.0, + 242.0, + 294.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 235.0, + 603.0, + 235.0, + 603.0, + 271.0, + 355.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 286.0, + 1402.0, + 286.0, + 1402.0, + 322.0, + 296.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 317.0, + 1143.0, + 317.0, + 1143.0, + 353.0, + 351.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 368.0, + 1404.0, + 368.0, + 1404.0, + 404.0, + 296.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 397.0, + 1308.0, + 397.0, + 1308.0, + 436.0, + 353.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 449.0, + 1406.0, + 449.0, + 1406.0, + 492.0, + 293.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 483.0, + 1354.0, + 483.0, + 1354.0, + 520.0, + 353.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 534.0, + 1407.0, + 534.0, + 1407.0, + 574.0, + 296.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 564.0, + 1408.0, + 564.0, + 1408.0, + 603.0, + 354.0, + 603.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/Xh5eMZVONGF/Xh5eMZVONGF.md b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF.md new file mode 100644 index 0000000000000000000000000000000000000000..17a65f63b68aa6afe84121f828168987809b3182 --- /dev/null +++ b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF.md @@ -0,0 +1,501 @@ +# LANGUAGE-AGNOSTIC REPRESENTATION LEARNING OF SOURCE CODE FROM STRUCTURE AND CONTEXT + +Daniel Zugner, Tobias Kirschstein ¨ Technical University of Munich {zuegnerd,kirschto}@in.tum.de + +Michele Catasta +Stanford University +pirroh@cs.stanford.edu + +Jure Leskovec Stanford University jure@cs.stanford.edu + +Stephan Gunnemann ¨ Technical University of Munich guennemann@in.tum.de + +# ABSTRACT + +Source code (Context) and its parsed abstract syntax tree (AST; Structure) are two complementary representations of the same computer program. Traditionally, designers of machine learning models have relied predominantly either on Structure or Context. We propose a new model, which jointly learns on Context and Structure of source code. In contrast to previous approaches, our model uses only language-agnostic features, i.e., source code and features that can be computed directly from the AST. Besides obtaining state-of-the-art on monolingual code summarization on all five programming languages considered in this work, we propose the first multilingual code summarization model. We show that jointly training on non-parallel data from multiple programming languages improves results on all individual languages, where the strongest gains are on low-resource languages. Remarkably, multilingual training only from Context does not lead to the same improvements, highlighting the benefits of combining Structure and Context for representation learning on code. + +# 1 INTRODUCTION + +Machine learning for code is an active and growing area of research which aims at building models that can learn semantically meaningful representations of programs. These embeddings can be used on downstream tasks, such as code generation, bug detection, or code summarization. We focus our work on two complementary data representations of programs: the source code (referred to as Context in this work), and the abstract syntax tree (AST; referred to as Structure). Traditionally, researchers and practitioners have decided to predominantly leverage either Structure or Context in their machine learning models. In this work, we show that jointly learning on Context and Structure improves representation learning on source code (see Fig. 1). + +The source code representation naturally lends itself to models from natural language processing (NLP), e.g., long short-term memory networks (Hochreiter & Schmidhuber, 1997) (LSTM) or Transformers (Vaswani et al., 2017; Radford et al., 2019; Dai et al., 2019; Yang et al., 2019; Shaw et al., 2018). On the other hand, models leveraging the structure representations are typically based on graph neural networks (GNNs) (Kipf & Welling, 2017; Xu et al., 2019; Velickovi ˇ c et al., 2018; ´ You et al., 2019; Hamilton et al., 2017; Li et al., 2015; Klicpera et al., 2020). While the AST representation makes the highly structured nature of source code explicit to the models, since most GNNs use the message-passing framework, their learned representations are inherently local and struggle to leverage long-range interactions. + +Recently, Hellendoorn et al. (2020) have explored models that can leverage several representations, including both Structure and Context. Their Graph Relational Embedding Attention Transformer (GREAT) extends Shaw et al. (2018), which biases the self-attention computation in a localized way given the underlying graph. The language-specific representations used by GREAT include a combination of the data flow graph, control flow graph, syntactic edges (inspired by Allamanis et al. (2018)), etc. which require specialized pipelines and static analysis tools to be obtained. + +![](images/7701ebd563c4ccbab0fb3932736f0a3f5ce8408d760ee36d22e78173cbc12d18.jpg) +Figure 1: Context and Structure both encapsulate valuable information about source code. In this realistic example, token 1 and 4 are distant in the sequence of tokens (Context), but only 5 hops away when traversing the Abstract Syntax Tree (Structure). As such, a method that relies only on the sequence of tokens could neglect the relationship between a method name and its return variable. Conversely, token 1 and 2 showcase the opposite setting. Hence, unifying Structure and Context leads to a more powerful representation of source code. + +We propose the CODE TRANSFORMER1, which combines distances computed on Structure and Context in the self-attention operation. In contrast to the localized treatment via edges described above, we make the full Structure accessible to the model at each layer by computing pairwise distances on the AST, such as shortest path lengths. To this end, we draw inspiration from the XLNet architecture (Yang et al., 2019), which uses relative distances instead of absolute positions in the attention computation. Importantly, all our features are language-agnostic2, i.e., can easily be computed for any programming language based on the source code and AST. + +We use two datasets comprising 5 different programming languages in total, and evaluate the representations learned by our model on the task of code summarization, where the model predicts a method’s name based on its body. Besides setting the state-of-the-art on all five languages for singlelanguage training, we also train the first multilingual model for code summarization. This is enabled by the fact that our model uses only language-agnostic features that can easily be obtained for any programming language. Remarkably, training our model on multiple programming languages substantially improves the performance on all languages. Moreover, multilingual training only from Context does not lead to the same improvements, highlighting the benefits of combining Structure and Context for representation learning on code. + +# 2 RELATED WORK + +Machine Learning for Code. Early research learned language models on raw text data, e.g., (Wang et al., 2016; Raychev et al., 2014; Dam et al., 2016), providing evidence for the naturalness assumption (Hindle et al., 2012). For example, Allamanis et al. (2015) learned distributed representations of variables and methods, finding that they were indeed able to encode common semantic properties from the regularities present in source code. Alon et al. (2019b) also found evidence of semantic arithmetic in their embedding space, dubbed code2vec. These representations—and their variants like (Mou et al., 2016)—can then be used to predict sequences of identifier sub-tokens (Allamanis et al., 2015) or API calls (Acharya et al., 2007; Nguyen et al., 2017). They can be used as advanced auto-completion tools (Hindle et al., 2012; Bhoopchand et al., 2016), including for user-provided tokens like Variable Names (Raychev et al., 2014; Allamanis et al., 2014). These are useful for deobfuscating Android applications (Bichsel et al., 2016) for example. + +Several works leverage structured graphical models for probabilistic models of source code, usually through parse trees (Maddison & Tarlow, 2014; Bielik et al., 2016). Unlike previous works where hand-crafted features were used as node features (Raychev et al., 2014) or as explicit semantic edges (Allamanis et al., 2018), our work does not augment the existing syntactic relationships between the different elements to enhance the predictive capabilities of the model. Other approaches (Alon et al., 2018; Li et al., 2017) also leverage the AST structure, but linearize the graph by first traversing it. + +Learning representations of structured languages. While models of language have dramatically improved in their ability to learn structure (syntax) and semantics from scratch, it can be argued that directly providing the model with the underlying structure of the language can help with generalization (Battaglia et al., 2018), managing long-ranging dependencies (Tai et al., 2015), or representing the compositional aspect of natural language (Socher et al., 2013). Notably, tree structures have shown promising results and inspired new architectures (Shen et al., 2019), including in the domain of source code (Fernandes et al., 2019), where the underlying syntax is directly available. Our work pursues this line of research, showing the benefits of explicitly integrating structural information as an inductive bias. Shiv & Quirk (2019) propose positional encodings for nodes on trees; however, their approach assumes regular trees, which is an unrealistic assumption when working with Abstract Syntax Trees, as an AST node can have arbitrarily many children, e.g., the arguments of a function. + +Graph Neural Networks. GNNs provide a powerful tool for machine learning on graphs, thanks to their ability to recursively incorporate information from neighboring nodes in the network (Battaglia et al., 2018), naturally capturing the graph structure simultaneously with the nodes’ features. (Gori et al., 2005; Scarselli et al., 2008) are able to learn vector representations of nodes and graphs in an end-to-end fashion, encoding structural and feature information in the embedding space. Under this model, GNNs have achieved state-of-the-art performance across a variety of tasks, such as node classification (Kipf & Welling, 2017; Hamilton et al., 2017; Klicpera et al., 2019a), link prediction (Zhang & Chen, 2018; Schlichtkrull et al., 2018), graph clustering (Defferrard et al., 2016; Ying et al., 2018) or graph classification (Ying et al., 2018; Dai et al., 2016; Duvenaud et al., 2015). + +# 3 INTEGRATING STRUCTURE AND CONTEXT IN THE CODE TRANSFORMER + +Self-attention is the core operation powering the Transformer. It enables the model to selectively focus on relevant parts of the input. The matrix form equation for attention with a single head is + +$$ +{ \mathrm { A t t e n t i o n } } ( Q , K , V ) = { \mathrm { s o f t m a x } } \left( { \frac { Q K ^ { T } } { \sqrt { d _ { k } } } } \right) V , +$$ + +where $Q , K \in \mathbb { R } ^ { N \times d _ { k } }$ and $V \in \mathbb { R } ^ { N \times d _ { v } }$ . $N$ is the number of input tokens, $d _ { k }$ the key dimension, and $d _ { v }$ the value dimension (typically we have $d _ { k } = d _ { v }$ ). The attention score of query $Q _ { i }$ and key $K _ { j }$ before softmax is + +$$ +\pmb { A } _ { i j } = \pmb { Q } _ { i } ^ { T } \pmb { K } _ { j } = \pmb { E } _ { i } ^ { T } \pmb { W } _ { q } ^ { T } \pmb { W } _ { k } \pmb { E } _ { j } , +$$ + +where $E _ { i } , E _ { j } \in \mathbb { R } ^ { d }$ are the $d$ -dimensional embeddings of tokens $i$ and $j$ , and $W _ { q }$ , $W _ { k } \in \mathbb { R } ^ { d _ { k } \times d }$ are the query and key projection matrices, respectively. + +Observe that Eq. (2) contains no assumption about potential structure in the input domain: in the attention operation we compute all dot products of query and key vectors equally, effectively viewing them as unordered sets of vectors. This means, however, that the model is oblivious to structured inputs (such as text or graphs) and therefore is unable to distinguish, for example, a variable name occurring as an argument and in the return statement of a method. + +In NLP, it is common to bias Transformers towards sequential inputs by adding positional encodings to the token embeddings. These positional encodings are obtained by applying an encoding function $\phi : \mathbb { R } \to \mathbb { R } ^ { d }$ to each token’s position $p _ { i }$ . These positional encodings make the information about the sequence of tokens available to the model. Eq. (2) becomes: + +$$ +A _ { i j } = ( E _ { i } + \phi ( p _ { i } ) ) ^ { T } W _ { q } ^ { T } W _ { k } ( E _ { j } + \phi ( p _ { j } ) ) , , +$$ + +which factorizes into + +$$ +A _ { i j } = \underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } } _ { ( \mathrm { u } ) { A } _ { i j } ^ { \mathrm { c e } } } { + \underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } { \phi } ( p _ { j } ) } _ { ( \mathrm { b } ) { A } _ { i j } ^ { \mathrm { c p } } } } + \underbrace { { \phi } ( p _ { i } ) ^ { T } W _ { q } ^ { T } W _ { k } { E } _ { j } } _ { ( \mathrm { c } ) { A } _ { i j } ^ { \mathrm { c e } } } { + \underbrace { { { \phi } ( p _ { i } ) } ^ { T } W _ { q } ^ { T } W _ { k } { \phi } ( p _ { j } ) } _ { ( \mathrm { d } ) { A } _ { i j } ^ { \mathrm { p } } } } . +$$ + +We can interpret the terms (a)-(d) as follows. (a) ${ \cal { A } } _ { i j } ^ { \mathrm { c c } }$ is the contribution from the ‘match’ between the content embeddings of tokens $i$ and $j$ ; (b) $A _ { i j } ^ { \mathrm { c p } }$ steers the attention towards certain positions based + +on the content of token $i$ ; (c) $A _ { i j } ^ { \mathrm { p c } }$ biases towards content embeddings based on the position of token $i$ ; (d) $A _ { i j } ^ { \mathrm { p p } }$ controls which positions should attend to which other positions. + +In our model, we adopt the formulation of Dai et al. (2019); Yang et al. (2019). They modify Eq. (4) by replacing the absolute position encodings $\phi ( p _ { i } )$ with relative position encodings $\phi ( r _ { i \to j } )$ : + +$$ +\begin{array} { r } { A _ { i j } ^ { \mathrm { r e l } } = \ E _ { i } ^ { T } W _ { q } ^ { T } W _ { k } E _ { j } + E _ { i } ^ { T } W _ { q } ^ { T } W _ { r } \phi ( r _ { i j } ) + u ^ { T } W _ { k } E _ { j } + v ^ { T } W _ { r } \phi ( r _ { i j } ) , } \end{array} +$$ + +where $r _ { i \to j }$ is the relative distance from token $i$ to token $j$ in the sequence, $\boldsymbol { u } , \boldsymbol { v } \in \mathbb { R } ^ { d _ { k } }$ are learnable bias vectors, and $W _ { r }$ is a key projection matrix for the relative distances. Besides fixing issues with absolute position encodings such as ambiguity when processing two sentences at a time, Eq. (5) enables native application of the powerful self-attention operation on domains such as graphs, where absolute coordinates are not available. We adopt the (non-trainable) sinusoidal encoding function proposed by (Vaswani et al., 2017) for all relations; see Appendix A.1 for details on the distance encoding function. + +# 3.1 INTEGRATING SOURCE CODE AND AST REPRESENTATIONS OF PROGRAMS. + +To enable the model to integrate information both the Context and Structure of programs, we modify Eq. (5) to be able to incorporate multiple different relations. To this end, we use one key projection matrix W (s)r per relation $s$ , and sum their contributions in the raw attention score. This enables the CODE TRANSFORMER to combine information from multiple relations between tokens in the attention computation. Besides the token distance in the Context, we include pairwise relations based on the AST as described in the following. See Fig. 2 for a visualization of the Structure distances we use. + +Shortest path length. We include the number of hops required to reach node $j$ starting from node $i$ and vice versa. Here, we treat the AST as an undirected graph, since otherwise most distances would be undefined: e.g., all other nodes in the AST would be unreachable from the leaves. + +Similar to the distance of two tokens on the source code sequence, the shortestpath length is a global distance. This makes the whole graph structure acces + +![](images/971f2174ff96f3b5b7aea4b5cb4f59f9af4bf791dc6a632edbb218691a5d22cf.jpg) +Figure 2: Structure distances used by our model. + +sible to the model at each layer. In contrast, Hellendoorn et al. (2020) add bias terms to the attention computation only for edges (i.e. shortest-path distance of 1), which is a local operation that only exchanges information between immediate neighbors (similar to message passing in GNNs). The equivalent localized operation on the source code sequence would be to treat the sequence as a chain graph and only compute attention terms for neighboring tokens, which in turn highlights the benefit of non-local attention operations. + +Ancestor distance. Since we treat the ASTs as undirected for the computation of the shortest-path length, we lose the direction information of the edges. To avoid this, we also include the distance on the ordered set of ancestors and descendants of a node in the AST (red arrow in Fig. 2). Again, we include number of (vertical) hops to avoid locality in the attention computation. For example, a node $r _ { i \to j } = 2$ for “grand-children” $j$ of $i$ , and $r _ { j i } = - 2$ in the other direction. + +Sibling distance. The neighbor sets in graphs are typically considered to be unordered, but in an AST, the order of children encodes their order of occurrence in the source code. To avoid information loss when encoding the AST, we further include the distance on the ordered set of siblings $\{ v _ { i } \}$ of a node, where we again avoid locality by encoding the number of hops, i.e. $r _ { v _ { 1 } v _ { 3 } } ~ = ~ 2$ and rv3→v1 = −2. + +Personalized PageRank (Page et al., 1999) (PPR). PPR is a well-studied proximity measure which has been shown to be very effective in learning with graphs (Klicpera et al., 2019a;b; Bojchevski et al., 2020). PPR captures the local graph structure around a pair of nodes $( i , j )$ . E.g., if $i$ has many neighbors, its PPR score for $j$ will be low even when they are only few hops apart, which complements the purely hop-based distances described above. + +![](images/2ae4d77322414ffc49b55c0b5e39c6d3b3d0f5b89a9cfca8c5f81d4390780d66.jpg) +Figure 3: Left: Sequence (Context) and AST (Structure) representation of an input code snippet. Center: The CODE TRANSFORMER jointly leverages the sequence of tokens and the Abstract Syntax Tree to learn expressive representations of source code. In addition to the input token and node embeddings the model uses different distances between the tokens, e.g., shortest paths on the AST or personalized PageRank, to reason about their relative positions. The output embeddings can be used for downstream tasks such as code summarization (right). + +Input embeddings to the model. To combine the Context and Structure information, we assign each token in the sequence to an AST node by selecting the AST node whose range in the source code is the shortest one containing the token. We concatenate the (sub-) token embeddings with the embedding of the token’s assigned AST node type as well as the token type returned by the tokenizer. That is, among all the internal nodes, we use as input only those corresponding to a token in the sequence; however, the remaining internal nodes can used by the model since their presence affects the distances between the remaining AST nodes. See Appendices A.3 and A.4 for details. + +# 3.2 EFFICIENT RELATIVE ATTENTION COMPUTATION. + +Na¨ıvely, we need to compute and materialize a tensor of dimension $N \times N \times d$ to hold all pairwise relative position encodings $\phi ( r _ { i \to j } )$ in Eq. (5) , where $N$ is the input length. This is prohibitive for fast GPU training. While for discrete distance values (e.g., sequence distance or shortest-path length on a graph) we only need to compute unique distance values occurring in the input, this does not generalize to continuous distances such as PPR. Therefore, we propose a constant-time approximation of the relational attention computation by grouping the values into $k \ll N ^ { 2 }$ bins. Since closer samples are typically more relevant for a query sample, we increase the bin widths exponentially with growing distance values. Throughout our experiments we have found the CODE TRANSFORMER to be relatively insensitive to the number of bins; we thus set $k = 3 2$ in our experiments. + +# 4 EXPERIMENTAL SETUP + +Code summarization is one of the most popular tasks in machine learning for code. Given the body of a function, the task is to predict the function’s name. As observed by Alon et al. (2019b) and Allamanis et al. (2016), this is a useful benchmark as method names in open-source projects tend to be precise and descriptive, and functions typically form complete logical units. See Fig. 3 (right) for a visual overview of the task. We use two complementary representations of programs: the source code as a sequence of tokens (Context) and the AST (Structure). As shown in Fig. 3 (left), tokens that are far away on the sequence may be very close on the AST and vice versa. In this task we make use of the CODE TRANSFORMER’s ability jointly leverage both Structure and Context and show that it improves learning. Further, we show the benefit of using only language-agnostic features in our model by training the first multilingual model for code summarization. + +Datasets. To highlight the benefit of only relying on language-agnostic representations such as source code and abstract syntax trees, we evaluate on challenging datasets in four programming languages introduced in the CodeSearchNet (CSN) Challenge (Husain et al., 2019): Python, Javascript, Go, and Ruby. Similar to Java-small, the datasets from CodeSearchNet have been carefully deduplicated by the creators to avoid data leakage from the training set, e.g., via copy-and-paste code. + +We further evaluate on Java-small (Allamanis et al., 2016), a popular and challenging code summarization dataset. It contains 11 open-source Java projects. We use the split as in Alon et al. (2019a), where 9 of these projects are used for training, one for validation, and one for test. The dataset contains roughly 700K samples (function definitions). Moreover, we also experiment with pre-training our model on Java-medium and Java-large (Alon et al., 2019a) before fine-tuning on Java-small, making sure to avoid leakage by removing the test and validation projects of Java-small from the pre-training dataset. See Table 1 for a summary of the datasets we use in this work. + +
Samples per partition
DatasetTrainVal.Test
CSN-Python412,17823,10722,176
CSN-Javascript123,8898,2536,483
CSN-Ruby48,7912,2092,279
CSN-Go317,83214,24214,291
Java-small691,97423,84457,088
+ +Table 1: Dataset statistics. + +Preprocessing. Each token of the source code is split into subtokens respective to code naming conventions, i.e., get TrainingData is converted to [get, training, data]. Following Alon et al. (2019a) we use at most six subtokens for the method names, truncating longer function names if necessary. In addition to the tokenized source code we produce an AST for each method using the open-source AST parser Semantic3. We limit the vocabulary to subtokens with at least 100 occurrences in the training set, and only consider snippets with 512 or fewer tokens (after removing punctuation). We refer the reader to the appendix for further details on the data preprocessing. + +Pointer network. We add a pointer network (Vinyals et al., 2015) (as described in Fernandes et al. (2019)) to the decoders of all Transformer-based models. This enables them to enhance their predictions by pointing at positions in the input sequence. For instance, when predicting the method name get url, the model can point directly to occurrences of the variable url. This often improves results for less frequent tokens, and even enables the model to predict tokens which are not in the vocabulary by pointing at their positions in the input. + +Baselines. We compare with code2seq (Alon et al., 2019a), the Graph Relational Embedding Attention Transformer (GREAT) (Hellendoorn et al., 2020), and the BiLSTM+GNN LSTM+Pointer model presented in Fernandes et al. (2019). Code2seq is a non-Transformer model and state of the art for code summarization using only AST information. GREAT is a recent Transformer model using the framework presented in (Shaw et al., 2018) to bias the attention via edges. In the original formulation, GREAT additionally uses hand-crafted, language-specific edges such as dataflow, ‘computed from’, or ‘next lexical use’ edges, which require specialized preprocessing and static analysis tools. While this approach of leveraging language-specific features can certainly improve results on specific tasks and programming languages, our goal is to have a flexible model that can be used on any programming language. Since the specialized preprocessing used by GREAT is proprietary and not public, we produce the results for GREAT using edges from the AST instead, i.e. it has access to the same information as our proposed model. Note that the preprocessing of Fernandes et al. (2019) is language specific, which is why we only compare with their results on Java-small. + +# 5 RESULTS + +# 5.1 MONOLINGUAL CODE SUMMARIZATION + +CSN dataset. First, we study the performance (measured by F1 score) of our model and the baselines on the traditional setting, where training and evaluation are performed on a single programming language. The results are shown in the upper part of Table 2. The CODE TRANSFORMER (without multi-language training) substantially outperforms all other models on all but one language, highlighting the effectiveness of jointly learning from Structure and Context. The only exception is Ruby, where it performs on par with its Context-only variant. We attribute this to the fact that there are relatively few samples in the Ruby dataset, and that Ruby is an dynamically typed language, which could make the Structure less powerful for learning. Interestingly, the Context-only CODE TRANSFORMER outperforms GREAT on all languages. We attribute this to the fact that GREAT uses the Structure of the programs only in a localized way (see Sec. 3.1). Another noteworthy finding is that code2seq performs comparably to the Transformer-based baselines on Go. We hypothesize that ASTs are more informative on Go since it is a compiled and strongly typed language. + +Table 2: Code summarization results on the CSN dataset (micro F1). + +
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq35.7924.8529.3430.1819.8823.9723.2310.3114.2852.3043.4347.45
GREAT35.0731.5933.2431.2026.8428.8624.6422.2323.3850.0146.5148.20
Ours w/o structure37.3831.9834.4733.1726.7029.5929.8525.8727.7251.7847.5749.59
Ours w/o pointer net37.7431.8534.5533.1228.7030.7523.3225.2124.2354.3150.1252.13
Ours36.4033.6634.9735.0629.6132.1131.4224.4627.5055.1048.0551.34
code2seq (Multilanguage)34.4925.4929.3231.6222.1626.0623.9717.0619.9352.7044.3648.17
GREAT (Multilanguage)36.7531.5433.9433.5827.7830.4130.0524.3326.8952.6548.3050.38
Ours w/o structure (Mult.)38.4830.1433.8035.3827.4130.8932.6126.7629.4055.0347.3450.90
Ours w/o pointer (Mult.)38.9133.1235.7837.2129.7533.0734.5227.3130.5056.0750.7653.28
Ours (Multilanguage)38.8933.8236.1836.9529.9833.1033.9328.9431.2456.0050.4453.07
Ours (Mult.+ Finetune)39.8532.7935.9837.0029.7933.0035.8527.7531.2855.6351.1253.28
Ours (Mult. + LM Pretrain)39.6735.2937.3537.0631.9434.3135.1929.3632.0157.7351.8954.65
+ +Java-small results. In Table 3 we present code summarization results on the Java-small dataset. Among all models equipped with a pointer network, the CODE TRANSFORMER (without pretraining) obtains state-of-the-art on code summarization, outperforming all baselines, including the previous state-of-the-art on Java-small proposed by Fernandes et al. (2019). Further, pre-training on Java-medium and Java-large on the permutation language modeling objective (Yang et al., 2019) substantially improves precision, recall, and F1 score after fine-tuning on Java-small. To avoid leakage, we exclude the projects used in the validation and test splits of Java-small from pre-training. + +Ablation study. We further perform ablations where we remove our model’s access to the Context or Structure, also presented in Table 3. + +With pointer network. We find that both ablations lead to a substantial drop in performance, highlighting the benefit of learning jointly from Structure and Context. Interestingly, the model without access to the Structure performs slightly better than the variant without Context. Note that our model without Structure is related to the XLNet (Yang et al., 2019) model, where we add a pointer network to the decoder and concatenate the token types to their respective input tokens (see Appendix A.4). Without pointer network. We repeat the ablation on the variants without pointer network. Here, the variant without Context performs better than the variant without Structure, + +
ModelPrec.Rec.F1
Without pointer net code2seq Ours w/o structure Ours w/o context51.23 50.70 51.8137.31 45.49 46.0443.18 47.96
Ours With pointer net50.3346.8048.75 48.50
Fernandes et al. (2019)=151.4
GREAT Ours w/o structure53.6046.4149.75
Ours w/o context55.4846.0750.34
54.4545.2949.45
Ours54.8549.8452.22
Ours + Pretrain57.0250.8753.77
+ +Table 3: Results on Java-small and ablation study. + +indicating that the pointer network helps to compensate for the lack of access to Structure. The Structure-only variant (w/o pointer net) of our model even outperforms the full variant in this scenario. Inspection of the results revealed that the Structure-only variant has better performance on longer method names, which have an outsize influence on the micro-F1 score used in this work. + +Ablation of the AST-based distances. In Table 4 we compare the performance of our model when trained with each of the four different AST distances (sibling shortest paths, ancestor shortest paths, shortest paths, personalized PageRank; see Section 3.1). Here, the model is trained on Java-small in the Structure-only setting and without pointer network. For reference, we also show the results of training our model using all four AST distance functions (c.f. Table 3). We find that, while the personalized PageRank distance performs best on its own, each of the individual distances on their own performs substantially worse than their combination, highlighting the usefulness of combining the distances in our model as well as their complementary nature. + +# 5.2 MULTILINGUAL CODE SUMMARIZATION + +Setup. A key contribution of our proposed architecture is that it only uses language-agnostic features, i.e. the source code and features that can be directly computed from the AST. We use this fact to study the first multilanguage code summarization model. We train our model jointly on Python, Javascript, Ruby, and Go. The shared sub-token vocabulary is the union of the individual vocabularies, enabling us to evaluate the multi-language model on the individual languages and compare with the single-language models. As proposed by Conneau & Lample (2019), we add a learned language embedding to each input embedding. + +Table 4: AST distance ablation study. + +
AST distanceF1 score
Sibling shortest paths46.17
Ancestor shortest paths47.89
Shortest paths47.76
Personalized PageRank48.47
All the above (c.f. Table 3)48.75
+ +Results. In the lower part of Table 2 we can see the results of training our CODE TRANSFORMER jointly on all four programming languages. Our multi-lingual variants substantially outperform the mono-lingual models on all languages. The strongest improvement is on Ruby, which is also the programming language with the smallest number of samples in the dataset. Fine-tuning on the individual languages after joint training on code summarization only has a marginal effect on performance, indicating that the multilingual objective is well-aligned with the individual languages. In the last row, we have a variant of our model where we pre-train on the multi-lingual masked language modeling task, followed by finetuning on code summarization on the individual languages. + +Further, we observe that similar to the results on Java-small, removing the pointer network generally leads to weaker performance. One notable exception is Go, where the variant without the pointer network performs better in terms of F1 score. Our investigation revealed that there seems to be some violation of the i.i.d. assumption in the split provided by the creators of the dataset. In Figure 7 we show that in the test partition of the Go dataset, the share of tokens from the labels also occurring in the methods’ bodies – exactly the scenario where the pointer network can improve predictions – is substantially lower compared to the train/validation partitions. + +Remarkably, the multi-language Context-only variant (i.e. without access to the Structure) performs substantially worse than the full multi-language variant. This highlights that Structure is crucial to exploit the commonalities of different programming languages. Also notably, the GREAT baseline’s results also improve substantially when trained in the multi-language setting, though it is still outperformed by our model. However, our results indicate that any representation learning model for code can benefit from multi-language training, especially when evaluating on low-resource languages. + +In Table 16 we present results using the sample-F1 score. At the time of submission, our monolingual model on Python outperforms the state of the art on the ogbg-code $2 ^ { 4 }$ (Hu et al., 2020) leaderboard by $112 \%$ , and our multilanguage variant with LM pretraining outperforms it by $122 \%$ . + +Qualitative analysis of multilingual representations. Learning the CODE TRANSFORMER on multiple programming languages jointly provides us with embeddings in a shared representation space. In Fig. 4 we show a t-SNE (Maaten & Hinton, 2008) visualization of the ca. 40,000 snippets from the validation sets of four programming languages from the CSN dataset. For the embedding of a snippet, we use the representation of the method name in the final layer of the encoder. Note that the true method names are masked, i.e., inaccessible to the model. Further, note that in contrast to the monolingual embeddings learned by Kanade et al. (2020), the embeddings we evaluate are learned on the task of code summarization (though a similar study could be performed by using our model that was trained on the traditional language modeling pretraining task on multiple languages). + +While snippets from the same language tend to be grouped together, there are interesting intersections of the different programming languages. For example, we highlight all methods whose names start with the subtoken parse or main. We see that snippets starting with parse are predominantly in an intersection region of Python and Javascript. From these snippets, we display the cross-language pair with smallest Euclidean embedding distance in Fig. 5. Remarkably, both snippets are effectively the same method in Javascript and Python – it is worth reminding that the model has never seen any parallel data during training. On the other hand, snippets starting with main tend to lie at an intersectional region of Python, Javascript, and Go. In Table 6 in the appendix we show additional cross-lingual pairs with similar embeddings, including a failure case of a main function, where embedding distance is not representative of semantic similarity. We attribute this to the fact that we used the encoder output embedding of the masked method name – the representation used by the decoder to predict the method name – as a snippet’s representation. Thus, snippets with completely different semantics (as is to be expected for very generic method names starting with main) have similar representations because they are predictive of the method name. + +![](images/e875f6f92fb5baf876c7eada99cff2f2f6a1ae4c2907de273c6e3de07dd8a6a8.jpg) +Figure 4: t-SNE visualization of the CODE TRANSFORMER’s learned multilingual representations. + +![](images/b4f73ecaf41d7a935e9b22345a3d4464c58d6b6b17f19a5180425db1c0483a08.jpg) +Figure 5: Example snippet starting with parse (left) and its best embedding match from other languages (right). Both methods parse an input string to convert it into a boolean value. Note that even though they are semantically very similar, their method names are not; nonetheless, their representations in the CODE TRANSFORMER encoder reflect their semantic similarity. + +As another qualitative insight into the representations learned by the CODE TRANSFORMER we have found that the language embeddings of languages with similar roots in language design are close; see Table 5 in the appendix for the pairwise similarity matrix of the learned language embeddings. + +# 6 CONCLUSION + +We present the CODE TRANSFORMER, which learns jointly from Structure and Context of programs while only relying on language-agnostic features. Our model obtains state-of-the-art performance on code summarization on five different programming languages. Besides these results for training on individual languages, the language-agnostic nature of our model allows us to train it jointly on multiple programming languages. The resulting multilingual model substantially outperforms its mono-lingual variant on all programming languages, setting the state of the art on each language. We observe the largest improvement from multilingual training on the language with fewest resources, indicating that multilingual training can improve learning for less widely used programming languages. Remarkably, multilingual training only from Context does not lead to the same improvements, highlighting the benefits of combining Structure and Context. + +# REFERENCES + +Mithun Acharya, Tao Xie, Jian Pei, and Jun Xu. Mining api patterns as partial orders from source code: From usage scenarios to specifications. In Proceedings of the the 6th Joint Meeting of the European Software Engineering Conference and the ACM SIGSOFT Symposium on The Foundations of Software Engineering, ESEC-FSE ’07, pp. 25–34, New York, NY, USA, 2007. Association for Computing Machinery. ISBN 9781595938114. doi: 10.1145/1287624.1287630. URL https://doi.org/10.1145/1287624.1287630. + +Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Learning natural coding conventions. In Proceedings of the 22nd ACM SIGSOFT International Symposium on Foundations of Software Engineering, FSE 2014, pp. 281–293, New York, NY, USA, 2014. Association for Computing Machinery. ISBN 9781450330565. doi: 10.1145/2635868.2635883. URL https://doi.org/10.1145/2635868.2635883. + +Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Suggesting accurate method and class names. In Proceedings of the 2015 10th Joint Meeting on Foundations of Software Engineering, ESEC/FSE 2015, pp. 38–49, New York, NY, USA, 2015. Association for Computing Machinery. ISBN 9781450336758. doi: 10.1145/2786805.2786849. URL https://doi. org/10.1145/2786805.2786849. + +Miltiadis Allamanis, Hao Peng, and Charles Sutton. A convolutional attention network for extreme summarization of source code. In International conference on machine learning, pp. 2091–2100, 2016. + +Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs with graphs. In International Conference on Learning Representations, 2018. URL https: //openreview.net/forum?id $=$ BJOFETxR-. + +Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representation for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450356985. doi: 10.1145/3192366.3192412. URL https://doi.org/10.1145/3192366.3192412. + +Uri Alon, Shaked Brody, Omer Levy, and Eran Yahav. code2seq: Generating sequences from structured representations of code. In International Conference on Learning Representations, 2019a. URL https://openreview.net/forum?id $=$ H1gKYo09tX. + +Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. code2vec: Learning distributed representations of code. Proceedings of the ACM on Programming Languages, 3(POPL):1–29, 2019b. + +Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018. + +Avishkar Bhoopchand, Tim Rocktaschel, Earl Barr, and Sebastian Riedel. Learning python code ¨ suggestion with a sparse pointer network. arXiv preprint arXiv:1611.08307, 2016. + +Benjamin Bichsel, Veselin Raychev, Petar Tsankov, and Martin Vechev. Statistical deobfuscation of android applications. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, CCS ’16, pp. 343–355, New York, NY, USA, 2016. Association for Computing Machinery. ISBN 9781450341394. doi: 10.1145/2976749.2978422. URL https: //doi.org/10.1145/2976749.2978422. + +Pavol Bielik, Veselin Raychev, and Martin Vechev. Phog: probabilistic model for code. In International Conference on Machine Learning, pp. 2933–2942, 2016. + +Aleksandar Bojchevski, Johannes Klicpera, Bryan Perozzi, Amol Kapoor, Martin Blais, Benedek Rozemberczki, Michal Lukasik, and Stephan G ´ unnemann. Scaling graph neural networks with ¨ approximate pagerank. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery amp; Data Mining, KDD ’20, pp. 2464–2473, New York, NY, USA, 2020. + +Association for Computing Machinery. ISBN 9781450379984. doi: 10.1145/3394486.3403296. +URL https://doi.org/10.1145/3394486.3403296. + +Dylan Bourgeois. Learning representations of source code from structure and context. 2019. URL http://infoscience.epfl.ch/record/277163. + +Alexis Conneau and Guillaume Lample. Cross-lingual language model pretraining. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´ R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 7059– 7069. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8928-cross-lingual-language-model-pretraining.pdf. + +Hanjun Dai, Bo Dai, and Le Song. Discriminative embeddings of latent variable models for structured data. In International conference on machine learning, pp. 2702–2711, 2016. + +Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. + +Hoa Khanh Dam, Truyen Tran, and Trang Pham. A deep language model for software code. In FSE 2016: Proceedings of the Foundations Software Engineering International Symposium [The Conference], 2016. + +Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨ graphs with fast localized spectral filtering. In Advances in neural information processing systems, pp. 3844–3852, 2016. + +David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´ Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. In Advances in neural information processing systems, pp. 2224–2232, 2015. + +Patrick Fernandes, Miltiadis Allamanis, and Marc Brockschmidt. Structured neural summarization. In International Conference on Learning Representations, 2019. URL https: //openreview.net/forum?id $=$ H1ersoRqtm. + +Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pp. 729–734. IEEE, 2005. + +William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NIPS, 2017. + +Vincent J. Hellendoorn, Charles Sutton, Rishabh Singh, Petros Maniatis, and David Bieber. Global relational models of source code. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id $=$ B1lnbRNtwr. + +Abram Hindle, Earl T. Barr, Zhendong Su, Mark Gabel, and Premkumar Devanbu. On the naturalness of software. In Proceedings of the 34th International Conference on Software Engineering, ICSE ’12, pp. 837–847. IEEE Press, 2012. ISBN 9781467310673. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. + +Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 22118–22133. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/file/ fb60d411a5c5b72b2e7d3527cfc84fd0-Paper.pdf. + +Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. Codesearchnet challenge: Evaluating the state of semantic code search. arXiv preprint arXiv:1909.09436, 2019. + +Aditya Kanade, Petros Maniatis, Gogul Balakrishnan, and Kensen Shi. Learning and evaluating contextual embedding of source code. In Proceedings of the 37th International Conference on Machine Learning, 2020. + +Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR), 2017. + +Johannes Klicpera, Aleksandar Bojchevski, and Stephan Gunnemann. Predict then propagate: ¨ Graph neural networks meet personalized pagerank. In International Conference on Learning Representations (ICLR), 2019a. + +Johannes Klicpera, Stefan Weißenberger, and Stephan Gunnemann. Diffusion improves graph learn-¨ ing. In Conference on Neural Information Processing Systems (NeurIPS), 2019b. + +Johannes Klicpera, Janek Groß, and Stephan Gunnemann. Directional message passing for molec- ¨ ular graphs. In International Conference on Learning Representations (ICLR), 2020. + +Jian Li, Yue Wang, Michael R Lyu, and Irwin King. Code completion with neural attention and pointer networks. arXiv preprint arXiv:1711.09573, 2017. + +Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. + +Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(Nov):2579–2605, 2008. + +Chris Maddison and Daniel Tarlow. Structured generative models of natural source code. In International Conference on Machine Learning, pp. 649–657, 2014. + +Lili Mou, Ge Li, Lu Zhang, Tao Wang, and Zhi Jin. Convolutional neural networks over tree structures for programming language processing. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16, pp. 1287–1293. AAAI Press, 2016. + +Trong Duc Nguyen, Anh Tuan Nguyen, Hung Dang Phan, and Tien N. Nguyen. Exploring api embedding for api usages and applications. In Proceedings of the 39th International Conference on Software Engineering, ICSE ’17, pp. 438–449. IEEE Press, 2017. ISBN 9781538638682. doi: 10.1109/ICSE.2017.47. URL https://doi.org/10.1109/ICSE.2017.47. + +Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford InfoLab, 1999. + +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019. + +Veselin Raychev, Martin Vechev, and Eran Yahav. Code completion with statistical language models. In Proceedings of the 35th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’14, pp. 419–428, New York, NY, USA, 2014. Association for Computing Machinery. ISBN 9781450327848. doi: 10.1145/2594291.2594321. URL https://doi. org/10.1145/2594291.2594321. + +Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008. + +Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. In Aldo Gangemi, Roberto Navigli, Maria-Esther Vidal, Pascal Hitzler, Raphael Troncy, Laura Hollink, Anna Tordai, and ¨ Mehwish Alam (eds.), The Semantic Web, pp. 593–607, Cham, 2018. Springer International Publishing. ISBN 978-3-319-93417-4. + +Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pp. 464–468, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. doi: 10.18653/v1/N18-2074. URL https://www.aclweb.org/anthology/N18-2074. + +Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. Ordered neurons: Integrating tree structures into recurrent neural networks. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id ${ . } =$ B1l6qiR5F7. + +Vighnesh Shiv and Chris Quirk. Novel positional encodings to enable tree-based transformers. In Advances in Neural Information Processing Systems, pp. 12081–12091, 2019. + +Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pp. 1631–1642, 2013. + +Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075, 2015. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017. + +Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li ´ o, and Yoshua \` Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rJXMpikCZ. + +Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in neural information processing systems, pp. 2692–2700, 2015. + +Song Wang, Devin Chollak, Dana Movshovitz-Attias, and Lin Tan. Bugram: Bug detection with n-gram language models. In Proceedings of the 31st IEEE/ACM International Conference on Automated Software Engineering, ASE 2016, pp. 708–719, New York, NY, USA, 2016. Association for Computing Machinery. ISBN 9781450338455. doi: 10.1145/2970276.2970341. URL https://doi.org/10.1145/2970276.2970341. + +Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations, 2019. URL https: //openreview.net/forum?id $=$ ryGs6iA5Km. + +Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pp. 5754–5764, 2019. + +Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hierarchical graph representation learning with differentiable pooling. In Advances in neural information processing systems, pp. 4800–4810, 2018. + +Jiaxuan You, Rex Ying, and Jure Leskovec. Position-aware graph neural networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 7134– 7143. PMLR, 09–15 Jun 2019. URL http://proceedings.mlr.press/v97/you19b. html. + +Muhan Zhang and Yixin Chen. Link prediction based on graph neural networks. In Advances in Neural Information Processing Systems, pp. 5165–5175, 2018. + +Table 5: Pairwise cosine similarities of the learned language embeddings of the CODE TRANSFORMER. + +
PythonJavascriptGoRuby
Python1.000.430.420.79
Javascript0.431.000.840.39
Go0.430.841.000.38
Ruby0.790.390.381.00
+ +# ACKNOWLEDGEMENTS + +We are grateful to Dylan Bourgeois for having paved the way to this research contribution with his thesis work (Bourgeois, 2019). We further thank Simon Geisler for his helpful suggestions and proofreading the paper, as well as the anonymous reviewers for their constructive feedback and fruitful discussions. + +This research was supported by the TUM International Graduate School of Science and Engineering (IGSSE). Stanford University is supported by DARPA under Nos. N660011924033 (MCS); ARO under Nos. W911NF-16-1- 0342 (MURI), W911NF-16-1-0171 (DURIP); NSF under Nos. OAC1835598 (CINES), OAC-1934578 (HDR), CCF-1918940 (Expeditions), IIS-2030477 (RAPID); Stanford Data Science Initiative, Wu Tsai Neurosciences Institute, Chan Zuckerberg Biohub, Amazon, JPMorgan Chase, Docomo, Hitachi, JD.com, KDDI, NVIDIA, Dell, Toshiba, Intel, and UnitedHealth Group. Jure Leskovec is a Chan Zuckerberg Biohub investigator. + +# A APPENDIX + +# A.1 DISTANCE ENCODING FUNCTION + +For encoding scalar relation values via vectors we employ encoding functions $\phi : \mathbb { R } \to \mathbb { R } ^ { d }$ , where $d$ is the model’s embedding dimension. We choose the popular sinusoidal encoding function presented in Vaswani et al. (2017): + +$$ +\phi ( r _ { i j } ) _ { 2 k } = \sin ( \frac { r _ { i j } } { M ^ { 2 k / d } } ) \qquad \phi ( r _ { i j } ) _ { 2 k + 1 } = \cos ( \frac { r _ { i j } } { M ^ { 2 k / d } } ) , +$$ + +where $1 \leq k < d / 2$ is the position in the encoding vector and $M$ is some constant; we adopt $M = 1 0 , 0 0 0$ as chosen by (Vaswani et al., 2017). Note that the distance encoding functions have no trainable parameters. + +# A.2 MULTILINGUAL REPRESENTATION ANALYSIS + +In Table 5, we show the pairwise cosine similarities of the learned language embeddings of the CODE TRANSFORMER. We can see that the pairs Python-Ruby and Javascript-Go have similar language embeddings. This aligns well with roots of language design and common use cases of the languages. + +Moreover, in Table 6, we show selected snippets starting with is, main, or load (left) and their best embedding matches from other languages (right). + +# A.3 DATA PREPROCESSING + +# A.3.1 TEXTUAL CODE SNIPPET PREPROCESSING + +1. Tokenize code snippets with Pygments language-specific tokenizer. + +2. Remove comments (both multi-line, single-line and doc comments). The comment token types. pygments.token.Comment and pygments.token.Literal.String.Doc that are generated by Pygments are used to identify comments. + +
Published as a conference paper at ICLR 2021
function isEqualArray(a,b){ if (a return true;
if ((a === undefined) undefined)){ return false;func areSameFloat32Array(a,b []float32) bool if len(a) != len(b){ return false
var =a.length; if (i !== b.length){ return false;for i := O;i < len(a);i++{ if a[i]!=b[i]{ return false J
while (i--){ if(a[i] !== b[i]){ return false; }return true
return true;
function main() var rawData $(' .HeaderTexture[data-login- user-email]') .data() ;func TaskSayHello(t *tasking.T){
if (rawData) 1 me={username := t.Flags.String("name") if username user, := user.Current ()
name:rawData.loginUserName, mail: rawData.loginUserEmail };username = user.Name
getCartId(function (cart){ me.cart = cart;} ift.Flags.Bool("verbose")
injectMenu(); refreshUsers(actions.updateUsers); listenOnChanges(onChange);t.Logf("Hello %s, the time now is %s\n", username,time.Now())
listenonorderConfirm(onConfirm); );}else{ t.Logf("Hello %s\n",username)
} else{ caliback('no user');
func Backup(filename string)error{
info,err := os.Stat(filename) if err != nil
if os.IsNotExist(err){
return nil
return err
def _load_rule_file(self, filename):if info.Size() == 0{ return nil
if not (os.path.exists(filename)):}
sys.stderr.write(files,err := filepath.Glob(
"rflint: %s: No such file or "directory\n" % filenamefilename + _BACKUP_SUFFIX
return) if err nil 1
try:return err
basename = os.path.basename(filename)}
(name,ext)= os.path.splitext(basename) imp.load_source(name,filename) numBackup := byte(1)
except Exception as e: sys.stderr.write(if len(files) !=0{
"rflint: %s: exception whilelastFile := files[len(files)-1]
"loading:%s\n" % (filename, str(e))numBackup = lastFile[len(lastFile)-2] + 1
if numBackup >
numBackup = '1'
else{
'1'
numBackup =
return Copy (filename,
fmt.Sprintf("%s+%s~",
string(numBackup)))
+ +Table 6: Selected snippets starting with is, main, or load (left) and their best embedding matches from other languages (right). + +![](images/e37f9cb44d930542340c5ad3d88619f8535c0f5512daf8c0c8d8a050432bff4b.jpg) +Figure 6: Example snippet and its corresponding AST obtained from GitHub Semantic. + +3. Empty lines are removed. + +4. Hard coded strings and numbers are replaced with a special [MASK STRING] and [MASK NUMBER] token. + +5. Indentation style of the code snippet is detected and whitespace characters at the beginning of a line are replaced with a single [INDENT] or [DEDENT] token when indentation changes. + +6. Tokens are further split into sub tokens, e.g., setBottomHeight [‘set’, ‘bottom’, ‘height’]. Throughout our experiments, we use 5 input sub tokens. If a token consists of less than 5 sub tokens, the remaining spaces are filled with a special [PAD] token. + +7. Any remaining tokens that only consist of white spaces are removed. The only white space characters that are kept are line breaks $^ { \bullet } \backslash \boldsymbol { \mathrm { n } } ^ { \prime }$ . + +8. Any code snippets where the Pygments tokenizer cannot parse a token are discarded. + +# A.3.2 STAGE 1 PREPROCESSING (GENERATION OF ASTS) + +1. Stripped code snippets are used to generate language-specific ASTs. For Java, we use the AST parser from the java-parser project. The ASTs contain node types and source ranges. For Python, JavaScript, Ruby and Go, we use semantic. + +2. Snippets that lead to an AST parse error are discarded. + +3. We calculate a mapping between tokens and nodes in the AST. Every token is assigned to the node in the AST with shortest source range that still encompasses the source range of the token. + +To find such a node, we originally intended to make use of the assumption that source ranges of child nodes do not overlap. Then, one could easily find the node with smallest encompassing source range by greedily selecting at every layer in the AST the child that encompasses the token’s source range (there can only be at most one child that fulfills this). However, this assumption does not hold for all ASTs (see Figure 6 for an example). As a heuristic, we greedily select the child node with the shorter source range in case there were multiple child nodes with encompassing source ranges. This approximation seems to be sufficient in our case, and limits runtime as we do not have to consider multiple paths in the AST. It is also sufficient to stop when no child node encompasses the source range of the token, as in ASTs the source ranges of child nodes are always contained in the source ranges of their parent. + +1. Tokens are vocabularized. Any token occurring less than 100 times in the training set is replaced by an $ token. + +2. We calculate multiple pair-wise relations between nodes in the AST: + +• Personalized Page Rank (PPR) We interpret the negative logarithm of PPR as a distance. We use a teleport probability of $\alpha = 0 . 1 5$ and a threshold of $e ^ { - 5 }$ , i.e., anything with $- \log P P R > 5$ is considered unreachable +• Shortest path length between two nodes +• Ancestor shortest paths (bidirectional). That is, the parent has an ancestor shortest path distance of 1 to all its children and the child has a distance of -1 to its parents. We consider nodes that are not ancestors or descendants of a node (i.e. not reachable by following only parent or only child relations) as not connected in the ancestor shortest paths relation. We encode this with a very large value in their distance; we have found a value of $1 , 0 0 0$ to work well in practice. +• Next sibling shortest paths (bidirectional, analogous to the ancestor shortest paths) + +Note that the ancestor shortest paths and next sibling shortest paths are required because treating the AST as a normal graph leads to ambiguity. In a graph, the neighbors of a node have no ordering; however in the AST, the order of the children of a node reflects their order in the code. Therefore, we explicitly include the next sibling shortest paths. The ancestor shortest paths would not be required if we treated the AST as a directed graph; in this case, however, a leaf node could not reach any other node in the AST, and therefore both PPR and shortest path length are not useful in this case. Therefore, we model the AST as undirected and inject the ancestor / child edges to avoid ambiguity. + +3. Distance values are binned into 32 bins using area-based exponential binning with a growth factor of 1.3, i.e., the area of a bin’s rectangle ( $x$ : bin range, $y$ : number of values in bin) will be approximately 1.3 times bigger for the next bin (going away from the bin that contains the zero value). Additionally, for discrete distance measures (such as sequence distance or shortest path length), we hard-code 9 values around 0 to have their own bins. For instance, on the sequence distance the values $- 4 , - 3 , \dots , 4$ have their individual bins, and around those values we employ the exponential binning. + +4. Punctuation tokens (such as points or brackets) are removed from the input sequence, as experiments showed that their presence does not improve performance but slows down training due to bigger input sizes. + +5. Snippets that are longer than MAX NUM TOKENS after punctuation tokens are removed are discarded from the training set. Throughout our experiments, we use MAX NUM TOKENS $= ~ 5 1 2$ . During evaluation on the test set, we use MAX NUM TOKENS $= ~ 1 0 0 0$ . + +# A.4 INPUT EMBEDDINGS TO THE MODEL + +Besides its five subtokens (e.g., [‘get’, ‘data’, ‘[PAD]’, ‘[PAD]’, ‘[PAD]’]), each input token has a token type (coming from the Pygments tokenizer) and an AST node type. The AST node type is the type of the node assigned to each respective token, as described in Section A.3.2. We concatenate the embeddings of the five subtokens, the token type, and the AST node type. Then, we apply a linear layer (without activation function) to project down to the model’s embedding dimension. + +# A.5 INPUT TO THE GREAT BASELINE + +As mentioned in the main text, we also compare with GREAT Hellendoorn et al. (2020). Since their preprocessing pipeline is proprietary and could not be shared with us even after contacting the authors, we provide to GREAT the same AST distances as our model. Since GREAT uses edges instead of distances to encode relations in the Structure, we essentially threshold the ancestor, sibling, and shortest-paths distances and provide the edges where the distances are equal to 1 (including their edge types) to the model. + +(a) CODE TRANSFORMER +Table 7: Code Summarization hyperparameters + +
HyperparameterValue
ActivationGELU
Input Nonlinearitytanh
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
+ +(b) GREAT (Hellendoorn et al., 2020) + +
HyperparameterValue
ActivationGELU
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
+ +# A.6 EXPERIMENTAL SETUP + +Table 7 shows hyperparameters of our models for code summarization. For all our experiments, we use a Transformer Decoder with one layer and teacher forcing to generate 6 output sub tokens. We also employ label smoothing of 0.1. As optimizer, we use Adam with a learning rate of $8 e ^ { - 5 }$ and weight decay of $3 e ^ { - 5 }$ . Batch size during training is 8 with a simulated batch size of 128 achieved by gradient accumulation. + +Apart from comparing the CODE TRANSFORMER to baselines, we performed the following hyperparameter comparisons and ablation studies: + +• CODE TRANSFORMER (structure-only) Using only AST information as input, i.e., masking all tokens that do not correspond to a leaf of the AST, and removing the token distance as a relation to be used by the model. Further, token types are not fed into the model. +• CODE TRANSFORMER (context-only) Here, we do not include any information on the AST (i.e. node types and distances on the AST). This is effectively the XLNet backbone plus encoding of the token type returned by the tokenizer. +• CODE TRANSFORMER (Max-Dist.) Applying a Max Distance Mask of 5 to the shortest paths distance (i.e., model cannot see a node that is more than 5 hops away no matter how small the other distances are). Early results showed that, as expected, results deteriorate substantially when limiting our model’s receptive field. Hence, we do not include these results in this work. +• Using 16 and 64 bins instead of 32 bins. This had no noticeable effect on performance. + +# A.7 CODE SUMMARIZATION EXAMPLES + +In the Tables 8, 9, 10, 11, 12, 13, 14 and 15 we present example functions from the Java-small dataset along with the different models’ predictions for the function name. + +public Summation next() { return parts $[ \dot { 1 } + + ]$ ; } + +Table 8: The CODE TRANSFORMER is the only model to correctly identify the notion of getting the next entry. + +
ModelPrediction
GREAT get x map
code2seq get parts
Ours w/o structureget
CODE TRANSFORMERget next
Ground Truthnext
+ +private Path findCacheFile(Path[] cacheFiles, String fileName) { if (cacheFiles $! =$ null && cacheFiles.length $> 0$ ) { for (Path file : cacheFiles) { if (file.getName().equals(fileName)) { return file; } return null; +} + +
ModelPrediction
GREATget path
code2seqfind file
Ours w/o structureget file
CODE TRANSFORMERfind cache
Ground Truthfind cache file
+ +Table 9: The CODE TRANSFORMER is the only model to both recognize that the task is to find a file as well as the fact that it is about the cache. However, it did not correctly predict the file part of the method name. + +public int compare(Pair p1, Pair p2) { LoggedJob $\ j 1 \ = \ \mathtt { p 1 }$ .first(); LoggedJob $\begin{array} { r } { \dot { \bf { \ j } } 2 { \bf { \nu } } = { \bf { \ p } } 2 } \end{array}$ .first(); return (j1.getSubmitTime() $<$ j2.getSubmitTime()) ? -1 : (j1.getSubmitTime() $= =$ j2.getSubmitTime()) ? 0 : 1; } + +
ModelPrediction
GREATrun
code2seq get submit time
Ours w/o structurecompare
CODE TRANSFORMERcompare
Ground Truthcompare
+ +Table 10: The CODE TRANSFORMER and the its context-only variant are the only models correctly recognizing the ‘compare’ template in the method body. + +public static MNTPROC fromValue(int value) { if (value $< ~ 0$ || value $> =$ values().length) { return null; } return values()[value]; +} + +
ModelPrediction
GREAT get value
code2seq get value
Ours w/o structureto
CODE TRANSFORMERfrom value
Ground Truthfrom value
+ +Table 11: The CODE TRANSFORMER is the only model to recognize that the snippet is similar to a static factory method which is often preceded with from. + +private Iterable listRootBlobs(String aPrefix, + +boolean useFlatBlobListing, +EnumSet listingDetails, +BlobRequestOptions options, +OperationContext opContext) + +throws StorageException, URISyntaxException { + +CloudBlobDirectoryWrapper directory $=$ this.container.getDirectoryReference(aPrefix); return directory.listBlobs(null, useFlatBlobListing, listingDetails, options, opContext); + +} + +
ModelPrediction
GREATlist blobs
code2seqlist blobs
Ours w/o structurelist blobs
CODE TRANSFORMERlist blobs by prefix
Ground Truthlist root blobs
+ +Table 12: All models could correctly identify the listBlobs() call in the return statement. However, the CODE TRANSFORMER additionally comprehended that the specified prefix is quite important. + +private static void dumpOpCounts(EnumMap $<$ FSEditLogOpCodes, Holder $<$ > opCounts) { StringBuilder sb $=$ new StringBuilder(); sb.append("Summary of operations loaded from edit log:\n "); Joiner.on("\n ").withKeyValueSeparator(" $^ { 1 1 } =$ ").appendTo(sb, opCounts); FSImage.LOG.debug(sb.toString()); +} + +Table 13: Only the CODE TRANSFORMER could correctly identify that it is the op counts that should be logged. + +
ModelPrediction
GREATappend
code2seqadd
Ours w/o structurelog
CODE TRANSFORMERlog op counts
Ground Truthdump op counts
+ +static String execCommand(File f, String... cmd) throws IOException { String[] args $=$ new String[cmd.length $^ +$ 1]; System.arraycopy(cmd, 0, args, 0, cmd.length); args[cmd.length] $=$ f.getCanonicalPath(); String output $=$ Shell.execCommand(args); return output; + +Table 14: Only the CODE TRANSFORMER and code2seq could identify that the relevant part of the method is concerned with executing a command instead of returning something. + +
ModelPrediction
GREATget canonical path
code2seqexec
Ours w/o structure get output
CODE TRANSFORMERexec command
Ground Truthexec command
+ +# protected void subView(Class cls) { + +indent(of(ENDTAG)); +sb.setLength(0); +out.print(sb.append('[').append(cls.getName()).append(']').toString()); +out.println(); + +Table 15: Only the CODE TRANSFORMER was able to link the print functionality to the object that should be printed, which can only be inferred from the object’s class in the method parameters. + +
ModelPrediction
GREATprint
code2seqprint
Ours w/o structureprint
CODE TRANSFORMERprint sub view
Ground Truthsub view
+ +![](images/1225dc2212652b0900da96594dc1baedeb731793c2a7f99b2f39f0a306a759b8.jpg) +Figure 7: Share of tokens in the labels also occurring in the bodies of methods. + +Table 16: Code summarization results on the CSN dataset (sample-F1). + +
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq111-11=1111-
GREAT34.9331.1231.6129.6924.2425.5525.6921.4922.1848.3845.9745.71
Ours w/o structure36.8732.1732.9731.3025.0326.6431.4325.3426.6349.7846.7346.69
Ours w/o pointer net38.7731.7233.2732.7025.5027.3332.1230.1729.3653.0948.7049.26
Ours36.6833.8633.8433.3627.5529.0231.5324.7226.4352.0047.3547.93
code2seq (Multilanguage)-====-
GREAT (Multilanguage)35.7330.8131.7431.4926.1727.4129.7224.2025.4350.3247.9447.66
Ours w/o structure (Mult.)36.7829.9231.5832.6026.0227.7431.7126.0727.2451.9147.5848.15
Ours w/o pointer (Mult.)37.1830.5232.0433.9525.9228.1132.7625.0427.0153.5048.5449.35
Ours (Multilanguage)38.1033.3234.1834.2928.6930.0833.3028.3329.2953.8650.4650.61
Ours (Mult. + Finetune)38.2932.4133.6534.4328.2829.9132.8927.1528.4953.8550.85
Ours (Mult.+ LMPretrain)50.81
38.9734.7735.3435.2330.2631.3833.7329.1529.9455.3152.0352.13
+ +# A.8 ESTIMATION OF POINTER NETWORK POTENTIAL + +In Table 2 we observe that the pointer network improves the F1 score for all languages except Go, where counterintuitively it leads to reduced performance as measured by F1 score on the test set (while it improves by about 3 points on validation). To investigate this, in Figure 7 we plot the share of tokens in the labels also occurring in the bodies of methods in the different languages. Intuitively, this gives an indication on how much gain we can expect from using a pointer network. If the share were zero, then no token in the labels ever occur in the bodies of the methods, so the pointer network cannot improve the prediction by pointing at the input. We see that for Go, there is a strong mismatch between the test partition and the train/validation partitions, which much fewer tokens from the labels occurring in the bodies of methods on test compared to train/validation. Thus, we attribute the drop in performance observed by adding a pointer network on Go to this apparent violation of the i.i.d. assumption. + +# A.9 CODE SUMMARIZATION RESULTS ON THE CSN DATASET (SAMPLE-F1) + +In Table 16, we present our results on the CSN dataset as measured by the sample-F1 score. \ No newline at end of file diff --git a/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_content_list.json b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f4b80266dca43fd05f47158e384ef2e7dc5a076 --- /dev/null +++ b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_content_list.json @@ -0,0 +1,2644 @@ +[ + { + "type": "text", + "text": "LANGUAGE-AGNOSTIC REPRESENTATION LEARNING OF SOURCE CODE FROM STRUCTURE AND CONTEXT ", + "text_level": 1, + "bbox": [ + 174, + 99, + 823, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daniel Zugner, Tobias Kirschstein ¨ Technical University of Munich {zuegnerd,kirschto}@in.tum.de ", + "bbox": [ + 184, + 170, + 465, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Michele Catasta \nStanford University \npirroh@cs.stanford.edu ", + "bbox": [ + 596, + 170, + 813, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jure Leskovec Stanford University jure@cs.stanford.edu ", + "bbox": [ + 184, + 234, + 382, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stephan Gunnemann ¨ Technical University of Munich guennemann@in.tum.de ", + "bbox": [ + 596, + 234, + 807, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 313, + 544, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Source code (Context) and its parsed abstract syntax tree (AST; Structure) are two complementary representations of the same computer program. Traditionally, designers of machine learning models have relied predominantly either on Structure or Context. We propose a new model, which jointly learns on Context and Structure of source code. In contrast to previous approaches, our model uses only language-agnostic features, i.e., source code and features that can be computed directly from the AST. Besides obtaining state-of-the-art on monolingual code summarization on all five programming languages considered in this work, we propose the first multilingual code summarization model. We show that jointly training on non-parallel data from multiple programming languages improves results on all individual languages, where the strongest gains are on low-resource languages. Remarkably, multilingual training only from Context does not lead to the same improvements, highlighting the benefits of combining Structure and Context for representation learning on code. ", + "bbox": [ + 233, + 342, + 764, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 559, + 336, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine learning for code is an active and growing area of research which aims at building models that can learn semantically meaningful representations of programs. These embeddings can be used on downstream tasks, such as code generation, bug detection, or code summarization. We focus our work on two complementary data representations of programs: the source code (referred to as Context in this work), and the abstract syntax tree (AST; referred to as Structure). Traditionally, researchers and practitioners have decided to predominantly leverage either Structure or Context in their machine learning models. In this work, we show that jointly learning on Context and Structure improves representation learning on source code (see Fig. 1). ", + "bbox": [ + 174, + 589, + 825, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The source code representation naturally lends itself to models from natural language processing (NLP), e.g., long short-term memory networks (Hochreiter & Schmidhuber, 1997) (LSTM) or Transformers (Vaswani et al., 2017; Radford et al., 2019; Dai et al., 2019; Yang et al., 2019; Shaw et al., 2018). On the other hand, models leveraging the structure representations are typically based on graph neural networks (GNNs) (Kipf & Welling, 2017; Xu et al., 2019; Velickovi ˇ c et al., 2018; ´ You et al., 2019; Hamilton et al., 2017; Li et al., 2015; Klicpera et al., 2020). While the AST representation makes the highly structured nature of source code explicit to the models, since most GNNs use the message-passing framework, their learned representations are inherently local and struggle to leverage long-range interactions. ", + "bbox": [ + 174, + 708, + 825, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Hellendoorn et al. (2020) have explored models that can leverage several representations, including both Structure and Context. Their Graph Relational Embedding Attention Transformer (GREAT) extends Shaw et al. (2018), which biases the self-attention computation in a localized way given the underlying graph. The language-specific representations used by GREAT include a combination of the data flow graph, control flow graph, syntactic edges (inspired by Allamanis et al. (2018)), etc. which require specialized pipelines and static analysis tools to be obtained. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7701ebd563c4ccbab0fb3932736f0a3f5ce8408d760ee36d22e78173cbc12d18.jpg", + "image_caption": [ + "Figure 1: Context and Structure both encapsulate valuable information about source code. In this realistic example, token 1 and 4 are distant in the sequence of tokens (Context), but only 5 hops away when traversing the Abstract Syntax Tree (Structure). As such, a method that relies only on the sequence of tokens could neglect the relationship between a method name and its return variable. Conversely, token 1 and 2 showcase the opposite setting. Hence, unifying Structure and Context leads to a more powerful representation of source code. " + ], + "image_footnote": [], + "bbox": [ + 271, + 101, + 728, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose the CODE TRANSFORMER1, which combines distances computed on Structure and Context in the self-attention operation. In contrast to the localized treatment via edges described above, we make the full Structure accessible to the model at each layer by computing pairwise distances on the AST, such as shortest path lengths. To this end, we draw inspiration from the XLNet architecture (Yang et al., 2019), which uses relative distances instead of absolute positions in the attention computation. Importantly, all our features are language-agnostic2, i.e., can easily be computed for any programming language based on the source code and AST. ", + "bbox": [ + 174, + 358, + 825, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use two datasets comprising 5 different programming languages in total, and evaluate the representations learned by our model on the task of code summarization, where the model predicts a method’s name based on its body. Besides setting the state-of-the-art on all five languages for singlelanguage training, we also train the first multilingual model for code summarization. This is enabled by the fact that our model uses only language-agnostic features that can easily be obtained for any programming language. Remarkably, training our model on multiple programming languages substantially improves the performance on all languages. Moreover, multilingual training only from Context does not lead to the same improvements, highlighting the benefits of combining Structure and Context for representation learning on code. ", + "bbox": [ + 174, + 463, + 825, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 608, + 339, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine Learning for Code. Early research learned language models on raw text data, e.g., (Wang et al., 2016; Raychev et al., 2014; Dam et al., 2016), providing evidence for the naturalness assumption (Hindle et al., 2012). For example, Allamanis et al. (2015) learned distributed representations of variables and methods, finding that they were indeed able to encode common semantic properties from the regularities present in source code. Alon et al. (2019b) also found evidence of semantic arithmetic in their embedding space, dubbed code2vec. These representations—and their variants like (Mou et al., 2016)—can then be used to predict sequences of identifier sub-tokens (Allamanis et al., 2015) or API calls (Acharya et al., 2007; Nguyen et al., 2017). They can be used as advanced auto-completion tools (Hindle et al., 2012; Bhoopchand et al., 2016), including for user-provided tokens like Variable Names (Raychev et al., 2014; Allamanis et al., 2014). These are useful for deobfuscating Android applications (Bichsel et al., 2016) for example. ", + "bbox": [ + 173, + 643, + 825, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several works leverage structured graphical models for probabilistic models of source code, usually through parse trees (Maddison & Tarlow, 2014; Bielik et al., 2016). Unlike previous works where hand-crafted features were used as node features (Raychev et al., 2014) or as explicit semantic edges (Allamanis et al., 2018), our work does not augment the existing syntactic relationships between the different elements to enhance the predictive capabilities of the model. Other approaches (Alon et al., 2018; Li et al., 2017) also leverage the AST structure, but linearize the graph by first traversing it. ", + "bbox": [ + 174, + 804, + 825, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 821, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning representations of structured languages. While models of language have dramatically improved in their ability to learn structure (syntax) and semantics from scratch, it can be argued that directly providing the model with the underlying structure of the language can help with generalization (Battaglia et al., 2018), managing long-ranging dependencies (Tai et al., 2015), or representing the compositional aspect of natural language (Socher et al., 2013). Notably, tree structures have shown promising results and inspired new architectures (Shen et al., 2019), including in the domain of source code (Fernandes et al., 2019), where the underlying syntax is directly available. Our work pursues this line of research, showing the benefits of explicitly integrating structural information as an inductive bias. Shiv & Quirk (2019) propose positional encodings for nodes on trees; however, their approach assumes regular trees, which is an unrealistic assumption when working with Abstract Syntax Trees, as an AST node can have arbitrarily many children, e.g., the arguments of a function. ", + "bbox": [ + 173, + 142, + 825, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Graph Neural Networks. GNNs provide a powerful tool for machine learning on graphs, thanks to their ability to recursively incorporate information from neighboring nodes in the network (Battaglia et al., 2018), naturally capturing the graph structure simultaneously with the nodes’ features. (Gori et al., 2005; Scarselli et al., 2008) are able to learn vector representations of nodes and graphs in an end-to-end fashion, encoding structural and feature information in the embedding space. Under this model, GNNs have achieved state-of-the-art performance across a variety of tasks, such as node classification (Kipf & Welling, 2017; Hamilton et al., 2017; Klicpera et al., 2019a), link prediction (Zhang & Chen, 2018; Schlichtkrull et al., 2018), graph clustering (Defferrard et al., 2016; Ying et al., 2018) or graph classification (Ying et al., 2018; Dai et al., 2016; Duvenaud et al., 2015). ", + "bbox": [ + 173, + 319, + 825, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 INTEGRATING STRUCTURE AND CONTEXT IN THE CODE TRANSFORMER ", + "text_level": 1, + "bbox": [ + 169, + 464, + 805, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-attention is the core operation powering the Transformer. It enables the model to selectively focus on relevant parts of the input. The matrix form equation for attention with a single head is ", + "bbox": [ + 174, + 494, + 823, + 525 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/bf5062f76acf55f459a4ae89adb1abed3d6a40f640c4bd316cd5ed6cd5cde38d.jpg", + "text": "$$\n{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { k } } } } \\right) V ,\n$$", + "text_format": "latex", + "bbox": [ + 341, + 529, + 653, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q , K \\in \\mathbb { R } ^ { N \\times d _ { k } }$ and $V \\in \\mathbb { R } ^ { N \\times d _ { v } }$ . $N$ is the number of input tokens, $d _ { k }$ the key dimension, and $d _ { v }$ the value dimension (typically we have $d _ { k } = d _ { v }$ ). The attention score of query $Q _ { i }$ and key $K _ { j }$ before softmax is ", + "bbox": [ + 174, + 568, + 825, + 611 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/a084e175bb3c6930841d0180a4202d6eff0f84ecb6a8a9cc250c5cd2833a474e.jpg", + "text": "$$\n\\pmb { A } _ { i j } = \\pmb { Q } _ { i } ^ { T } \\pmb { K } _ { j } = \\pmb { E } _ { i } ^ { T } \\pmb { W } _ { q } ^ { T } \\pmb { W } _ { k } \\pmb { E } _ { j } ,\n$$", + "text_format": "latex", + "bbox": [ + 385, + 609, + 611, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E _ { i } , E _ { j } \\in \\mathbb { R } ^ { d }$ are the $d$ -dimensional embeddings of tokens $i$ and $j$ , and $W _ { q }$ , $W _ { k } \\in \\mathbb { R } ^ { d _ { k } \\times d }$ are the query and key projection matrices, respectively. ", + "bbox": [ + 174, + 633, + 823, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observe that Eq. (2) contains no assumption about potential structure in the input domain: in the attention operation we compute all dot products of query and key vectors equally, effectively viewing them as unordered sets of vectors. This means, however, that the model is oblivious to structured inputs (such as text or graphs) and therefore is unable to distinguish, for example, a variable name occurring as an argument and in the return statement of a method. ", + "bbox": [ + 174, + 667, + 825, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In NLP, it is common to bias Transformers towards sequential inputs by adding positional encodings to the token embeddings. These positional encodings are obtained by applying an encoding function $\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }$ to each token’s position $p _ { i }$ . These positional encodings make the information about the sequence of tokens available to the model. Eq. (2) becomes: ", + "bbox": [ + 173, + 744, + 825, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/b96beefee59c632a287223572ec7702368e163a81b493e1a921e5c4d74a409f9.jpg", + "text": "$$\nA _ { i j } = ( E _ { i } + \\phi ( p _ { i } ) ) ^ { T } W _ { q } ^ { T } W _ { k } ( E _ { j } + \\phi ( p _ { j } ) ) , ,\n$$", + "text_format": "latex", + "bbox": [ + 344, + 804, + 650, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which factorizes into ", + "bbox": [ + 173, + 828, + 313, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/df07c855af71840c08fb5fcac8d0a6450e30cf45761ee10da8da265d06d28ade.jpg", + "text": "$$\nA _ { i j } = \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } } _ { ( \\mathrm { u } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { b } ) { A } _ { i j } ^ { \\mathrm { c p } } } } + \\underbrace { { \\phi } ( p _ { i } ) ^ { T } W _ { q } ^ { T } W _ { k } { E } _ { j } } _ { ( \\mathrm { c } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { { \\phi } ( p _ { i } ) } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { d } ) { A } _ { i j } ^ { \\mathrm { p } } } } .\n$$", + "text_format": "latex", + "bbox": [ + 192, + 844, + 787, + 888 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can interpret the terms (a)-(d) as follows. (a) ${ \\cal { A } } _ { i j } ^ { \\mathrm { c c } }$ is the contribution from the ‘match’ between the content embeddings of tokens $i$ and $j$ ; (b) $A _ { i j } ^ { \\mathrm { c p } }$ steers the attention towards certain positions based ", + "bbox": [ + 176, + 892, + 825, + 926 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on the content of token $i$ ; (c) $A _ { i j } ^ { \\mathrm { p c } }$ biases towards content embeddings based on the position of token $i$ ; (d) $A _ { i j } ^ { \\mathrm { p p } }$ controls which positions should attend to which other positions. ", + "bbox": [ + 169, + 102, + 823, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our model, we adopt the formulation of Dai et al. (2019); Yang et al. (2019). They modify Eq. (4) by replacing the absolute position encodings $\\phi ( p _ { i } )$ with relative position encodings $\\phi ( r _ { i \\to j } )$ : ", + "bbox": [ + 169, + 141, + 825, + 170 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/4abc95bdea1392a92734f005afce75f982c5226661a1db2266b0c52e16c233d4.jpg", + "text": "$$\n\\begin{array} { r } { A _ { i j } ^ { \\mathrm { r e l } } = \\ E _ { i } ^ { T } W _ { q } ^ { T } W _ { k } E _ { j } + E _ { i } ^ { T } W _ { q } ^ { T } W _ { r } \\phi ( r _ { i j } ) + u ^ { T } W _ { k } E _ { j } + v ^ { T } W _ { r } \\phi ( r _ { i j } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 236, + 175, + 759, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r _ { i \\to j }$ is the relative distance from token $i$ to token $j$ in the sequence, $\\boldsymbol { u } , \\boldsymbol { v } \\in \\mathbb { R } ^ { d _ { k } }$ are learnable bias vectors, and $W _ { r }$ is a key projection matrix for the relative distances. Besides fixing issues with absolute position encodings such as ambiguity when processing two sentences at a time, Eq. (5) enables native application of the powerful self-attention operation on domains such as graphs, where absolute coordinates are not available. We adopt the (non-trainable) sinusoidal encoding function proposed by (Vaswani et al., 2017) for all relations; see Appendix A.1 for details on the distance encoding function. ", + "bbox": [ + 173, + 202, + 825, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 INTEGRATING SOURCE CODE AND AST REPRESENTATIONS OF PROGRAMS. ", + "text_level": 1, + "bbox": [ + 176, + 315, + 736, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable the model to integrate information both the Context and Structure of programs, we modify Eq. (5) to be able to incorporate multiple different relations. To this end, we use one key projection matrix W (s)r per relation $s$ , and sum their contributions in the raw attention score. This enables the CODE TRANSFORMER to combine information from multiple relations between tokens in the attention computation. Besides the token distance in the Context, we include pairwise relations based on the AST as described in the following. See Fig. 2 for a visualization of the Structure distances we use. ", + "bbox": [ + 173, + 342, + 825, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shortest path length. We include the number of hops required to reach node $j$ starting from node $i$ and vice versa. Here, we treat the AST as an undirected graph, since otherwise most distances would be undefined: e.g., all other nodes in the AST would be unreachable from the leaves. ", + "bbox": [ + 173, + 454, + 436, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to the distance of two tokens on the source code sequence, the shortestpath length is a global distance. This makes the whole graph structure acces", + "bbox": [ + 174, + 571, + 436, + 627 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/971f2174ff96f3b5b7aea4b5cb4f59f9af4bf791dc6a632edbb218691a5d22cf.jpg", + "image_caption": [ + "Figure 2: Structure distances used by our model. " + ], + "image_footnote": [], + "bbox": [ + 454, + 469, + 820, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sible to the model at each layer. In contrast, Hellendoorn et al. (2020) add bias terms to the attention computation only for edges (i.e. shortest-path distance of 1), which is a local operation that only exchanges information between immediate neighbors (similar to message passing in GNNs). The equivalent localized operation on the source code sequence would be to treat the sequence as a chain graph and only compute attention terms for neighboring tokens, which in turn highlights the benefit of non-local attention operations. ", + "bbox": [ + 174, + 627, + 825, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ancestor distance. Since we treat the ASTs as undirected for the computation of the shortest-path length, we lose the direction information of the edges. To avoid this, we also include the distance on the ordered set of ancestors and descendants of a node in the AST (red arrow in Fig. 2). Again, we include number of (vertical) hops to avoid locality in the attention computation. For example, a node $r _ { i \\to j } = 2$ for “grand-children” $j$ of $i$ , and $r _ { j i } = - 2$ in the other direction. ", + "bbox": [ + 173, + 720, + 825, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sibling distance. The neighbor sets in graphs are typically considered to be unordered, but in an AST, the order of children encodes their order of occurrence in the source code. To avoid information loss when encoding the AST, we further include the distance on the ordered set of siblings $\\{ v _ { i } \\}$ of a node, where we again avoid locality by encoding the number of hops, i.e. $r _ { v _ { 1 } v _ { 3 } } ~ = ~ 2$ and rv3→v1 = −2. ", + "bbox": [ + 173, + 801, + 825, + 871 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Personalized PageRank (Page et al., 1999) (PPR). PPR is a well-studied proximity measure which has been shown to be very effective in learning with graphs (Klicpera et al., 2019a;b; Bojchevski et al., 2020). PPR captures the local graph structure around a pair of nodes $( i , j )$ . E.g., if $i$ has many neighbors, its PPR score for $j$ will be low even when they are only few hops apart, which complements the purely hop-based distances described above. ", + "bbox": [ + 174, + 881, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ae4d77322414ffc49b55c0b5e39c6d3b3d0f5b89a9cfca8c5f81d4390780d66.jpg", + "image_caption": [ + "Figure 3: Left: Sequence (Context) and AST (Structure) representation of an input code snippet. Center: The CODE TRANSFORMER jointly leverages the sequence of tokens and the Abstract Syntax Tree to learn expressive representations of source code. In addition to the input token and node embeddings the model uses different distances between the tokens, e.g., shortest paths on the AST or personalized PageRank, to reason about their relative positions. The output embeddings can be used for downstream tasks such as code summarization (right). " + ], + "image_footnote": [], + "bbox": [ + 187, + 101, + 813, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 443, + 823, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input embeddings to the model. To combine the Context and Structure information, we assign each token in the sequence to an AST node by selecting the AST node whose range in the source code is the shortest one containing the token. We concatenate the (sub-) token embeddings with the embedding of the token’s assigned AST node type as well as the token type returned by the tokenizer. That is, among all the internal nodes, we use as input only those corresponding to a token in the sequence; however, the remaining internal nodes can used by the model since their presence affects the distances between the remaining AST nodes. See Appendices A.3 and A.4 for details. ", + "bbox": [ + 174, + 483, + 825, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 EFFICIENT RELATIVE ATTENTION COMPUTATION. ", + "text_level": 1, + "bbox": [ + 174, + 602, + 552, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Na¨ıvely, we need to compute and materialize a tensor of dimension $N \\times N \\times d$ to hold all pairwise relative position encodings $\\phi ( r _ { i \\to j } )$ in Eq. (5) , where $N$ is the input length. This is prohibitive for fast GPU training. While for discrete distance values (e.g., sequence distance or shortest-path length on a graph) we only need to compute unique distance values occurring in the input, this does not generalize to continuous distances such as PPR. Therefore, we propose a constant-time approximation of the relational attention computation by grouping the values into $k \\ll N ^ { 2 }$ bins. Since closer samples are typically more relevant for a query sample, we increase the bin widths exponentially with growing distance values. Throughout our experiments we have found the CODE TRANSFORMER to be relatively insensitive to the number of bins; we thus set $k = 3 2$ in our experiments. ", + "bbox": [ + 174, + 628, + 825, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 779, + 395, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Code summarization is one of the most popular tasks in machine learning for code. Given the body of a function, the task is to predict the function’s name. As observed by Alon et al. (2019b) and Allamanis et al. (2016), this is a useful benchmark as method names in open-source projects tend to be precise and descriptive, and functions typically form complete logical units. See Fig. 3 (right) for a visual overview of the task. We use two complementary representations of programs: the source code as a sequence of tokens (Context) and the AST (Structure). As shown in Fig. 3 (left), tokens that are far away on the sequence may be very close on the AST and vice versa. In this task we make use of the CODE TRANSFORMER’s ability jointly leverage both Structure and Context and show that it improves learning. Further, we show the benefit of using only language-agnostic features in our model by training the first multilingual model for code summarization. ", + "bbox": [ + 174, + 811, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To highlight the benefit of only relying on language-agnostic representations such as source code and abstract syntax trees, we evaluate on challenging datasets in four programming languages introduced in the CodeSearchNet (CSN) Challenge (Husain et al., 2019): Python, Javascript, Go, and Ruby. Similar to Java-small, the datasets from CodeSearchNet have been carefully deduplicated by the creators to avoid data leakage from the training set, e.g., via copy-and-paste code. ", + "bbox": [ + 174, + 142, + 823, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further evaluate on Java-small (Allamanis et al., 2016), a popular and challenging code summarization dataset. It contains 11 open-source Java projects. We use the split as in Alon et al. (2019a), where 9 of these projects are used for training, one for validation, and one for test. The dataset contains roughly 700K samples (function definitions). Moreover, we also experiment with pre-training our model on Java-medium and Java-large (Alon et al., 2019a) before fine-tuning on Java-small, making sure to avoid leakage by removing the test and validation projects of Java-small from the pre-training dataset. See Table 1 for a summary of the datasets we use in this work. ", + "bbox": [ + 174, + 219, + 529, + 377 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b4225af91868e1a8a78db0e52181e077f82a2809a40ec936d5ee0af636f96fb7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Dataset statistics. " + ], + "table_body": "
Samples per partition
DatasetTrainVal.Test
CSN-Python412,17823,10722,176
CSN-Javascript123,8898,2536,483
CSN-Ruby48,7912,2092,279
CSN-Go317,83214,24214,291
Java-small691,97423,84457,088
", + "bbox": [ + 545, + 236, + 825, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 372, + 717, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Preprocessing. Each token of the source code is split into subtokens respective to code naming conventions, i.e., get TrainingData is converted to [get, training, data]. Following Alon et al. (2019a) we use at most six subtokens for the method names, truncating longer function names if necessary. In addition to the tokenized source code we produce an AST for each method using the open-source AST parser Semantic3. We limit the vocabulary to subtokens with at least 100 occurrences in the training set, and only consider snippets with 512 or fewer tokens (after removing punctuation). We refer the reader to the appendix for further details on the data preprocessing. ", + "bbox": [ + 174, + 396, + 825, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pointer network. We add a pointer network (Vinyals et al., 2015) (as described in Fernandes et al. (2019)) to the decoders of all Transformer-based models. This enables them to enhance their predictions by pointing at positions in the input sequence. For instance, when predicting the method name get url, the model can point directly to occurrences of the variable url. This often improves results for less frequent tokens, and even enables the model to predict tokens which are not in the vocabulary by pointing at their positions in the input. ", + "bbox": [ + 174, + 503, + 823, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare with code2seq (Alon et al., 2019a), the Graph Relational Embedding Attention Transformer (GREAT) (Hellendoorn et al., 2020), and the BiLSTM+GNN LSTM+Pointer model presented in Fernandes et al. (2019). Code2seq is a non-Transformer model and state of the art for code summarization using only AST information. GREAT is a recent Transformer model using the framework presented in (Shaw et al., 2018) to bias the attention via edges. In the original formulation, GREAT additionally uses hand-crafted, language-specific edges such as dataflow, ‘computed from’, or ‘next lexical use’ edges, which require specialized preprocessing and static analysis tools. While this approach of leveraging language-specific features can certainly improve results on specific tasks and programming languages, our goal is to have a flexible model that can be used on any programming language. Since the specialized preprocessing used by GREAT is proprietary and not public, we produce the results for GREAT using edges from the AST instead, i.e. it has access to the same information as our proposed model. Note that the preprocessing of Fernandes et al. (2019) is language specific, which is why we only compare with their results on Java-small. ", + "bbox": [ + 174, + 598, + 825, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 799, + 281, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 MONOLINGUAL CODE SUMMARIZATION ", + "text_level": 1, + "bbox": [ + 178, + 832, + 490, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CSN dataset. First, we study the performance (measured by F1 score) of our model and the baselines on the traditional setting, where training and evaluation are performed on a single programming language. The results are shown in the upper part of Table 2. The CODE TRANSFORMER (without multi-language training) substantially outperforms all other models on all but one language, highlighting the effectiveness of jointly learning from Structure and Context. The only exception is Ruby, where it performs on par with its Context-only variant. We attribute this to the fact that there are relatively few samples in the Ruby dataset, and that Ruby is an dynamically typed language, which could make the Structure less powerful for learning. Interestingly, the Context-only CODE TRANSFORMER outperforms GREAT on all languages. We attribute this to the fact that GREAT uses the Structure of the programs only in a localized way (see Sec. 3.1). Another noteworthy finding is that code2seq performs comparably to the Transformer-based baselines on Go. We hypothesize that ASTs are more informative on Go since it is a compiled and strongly typed language. ", + "bbox": [ + 176, + 857, + 823, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6ea8249b1a6b65be5a059f68d87e3dba30b23b79c16802b7d266349476c02871.jpg", + "table_caption": [ + "Table 2: Code summarization results on the CSN dataset (micro F1). " + ], + "table_footnote": [], + "table_body": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq35.7924.8529.3430.1819.8823.9723.2310.3114.2852.3043.4347.45
GREAT35.0731.5933.2431.2026.8428.8624.6422.2323.3850.0146.5148.20
Ours w/o structure37.3831.9834.4733.1726.7029.5929.8525.8727.7251.7847.5749.59
Ours w/o pointer net37.7431.8534.5533.1228.7030.7523.3225.2124.2354.3150.1252.13
Ours36.4033.6634.9735.0629.6132.1131.4224.4627.5055.1048.0551.34
code2seq (Multilanguage)34.4925.4929.3231.6222.1626.0623.9717.0619.9352.7044.3648.17
GREAT (Multilanguage)36.7531.5433.9433.5827.7830.4130.0524.3326.8952.6548.3050.38
Ours w/o structure (Mult.)38.4830.1433.8035.3827.4130.8932.6126.7629.4055.0347.3450.90
Ours w/o pointer (Mult.)38.9133.1235.7837.2129.7533.0734.5227.3130.5056.0750.7653.28
Ours (Multilanguage)38.8933.8236.1836.9529.9833.1033.9328.9431.2456.0050.4453.07
Ours (Mult.+ Finetune)39.8532.7935.9837.0029.7933.0035.8527.7531.2855.6351.1253.28
Ours (Mult. + LM Pretrain)39.6735.2937.3537.0631.9434.3135.1929.3632.0157.7351.8954.65
", + "bbox": [ + 176, + 79, + 821, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 294, + 825, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Java-small results. In Table 3 we present code summarization results on the Java-small dataset. Among all models equipped with a pointer network, the CODE TRANSFORMER (without pretraining) obtains state-of-the-art on code summarization, outperforming all baselines, including the previous state-of-the-art on Java-small proposed by Fernandes et al. (2019). Further, pre-training on Java-medium and Java-large on the permutation language modeling objective (Yang et al., 2019) substantially improves precision, recall, and F1 score after fine-tuning on Java-small. To avoid leakage, we exclude the projects used in the validation and test splits of Java-small from pre-training. ", + "bbox": [ + 174, + 435, + 825, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study. We further perform ablations where we remove our model’s access to the Context or Structure, also presented in Table 3. ", + "bbox": [ + 174, + 547, + 823, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With pointer network. We find that both ablations lead to a substantial drop in performance, highlighting the benefit of learning jointly from Structure and Context. Interestingly, the model without access to the Structure performs slightly better than the variant without Context. Note that our model without Structure is related to the XLNet (Yang et al., 2019) model, where we add a pointer network to the decoder and concatenate the token types to their respective input tokens (see Appendix A.4). Without pointer network. We repeat the ablation on the variants without pointer network. Here, the variant without Context performs better than the variant without Structure, ", + "bbox": [ + 174, + 584, + 482, + 790 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/93c3db634ac78ceb6a3e42a79f6c881109ae26412d90ebf329ef70021d85d788.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Results on Java-small and ablation study. " + ], + "table_body": "
ModelPrec.Rec.F1
Without pointer net code2seq Ours w/o structure Ours w/o context51.23 50.70 51.8137.31 45.49 46.0443.18 47.96
Ours With pointer net50.3346.8048.75 48.50
Fernandes et al. (2019)=151.4
GREAT Ours w/o structure53.6046.4149.75
Ours w/o context55.4846.0750.34
54.4545.2949.45
Ours54.8549.8452.22
Ours + Pretrain57.0250.8753.77
", + "bbox": [ + 514, + 583, + 810, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicating that the pointer network helps to compensate for the lack of access to Structure. The Structure-only variant (w/o pointer net) of our model even outperforms the full variant in this scenario. Inspection of the results revealed that the Structure-only variant has better performance on longer method names, which have an outsize influence on the micro-F1 score used in this work. ", + "bbox": [ + 174, + 791, + 825, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of the AST-based distances. In Table 4 we compare the performance of our model when trained with each of the four different AST distances (sibling shortest paths, ancestor shortest paths, shortest paths, personalized PageRank; see Section 3.1). Here, the model is trained on Java-small in the Structure-only setting and without pointer network. For reference, we also show the results of training our model using all four AST distance functions (c.f. Table 3). We find that, while the personalized PageRank distance performs best on its own, each of the individual distances on their own performs substantially worse than their combination, highlighting the usefulness of combining the distances in our model as well as their complementary nature. ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 MULTILINGUAL CODE SUMMARIZATION ", + "text_level": 1, + "bbox": [ + 176, + 162, + 490, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Setup. A key contribution of our proposed architecture is that it only uses language-agnostic features, i.e. the source code and features that can be directly computed from the AST. We use this fact to study the first multilanguage code summarization model. We train our model jointly on Python, Javascript, Ruby, and Go. The shared sub-token vocabulary is the union of the individual vocabularies, enabling us to evaluate the multi-language model on the individual languages and compare with the single-language models. As proposed by Conneau & Lample (2019), we add a learned language embedding to each input embedding. ", + "bbox": [ + 174, + 188, + 552, + 299 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/44d93ce3802abbb255784a0d5f66ef881ac7c557a4abf8c4d5752cc699deb37d.jpg", + "table_caption": [ + "Table 4: AST distance ablation study. " + ], + "table_footnote": [], + "table_body": "
AST distanceF1 score
Sibling shortest paths46.17
Ancestor shortest paths47.89
Shortest paths47.76
Personalized PageRank48.47
All the above (c.f. Table 3)48.75
", + "bbox": [ + 578, + 184, + 810, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 300, + 825, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. In the lower part of Table 2 we can see the results of training our CODE TRANSFORMER jointly on all four programming languages. Our multi-lingual variants substantially outperform the mono-lingual models on all languages. The strongest improvement is on Ruby, which is also the programming language with the smallest number of samples in the dataset. Fine-tuning on the individual languages after joint training on code summarization only has a marginal effect on performance, indicating that the multilingual objective is well-aligned with the individual languages. In the last row, we have a variant of our model where we pre-train on the multi-lingual masked language modeling task, followed by finetuning on code summarization on the individual languages. ", + "bbox": [ + 173, + 342, + 825, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further, we observe that similar to the results on Java-small, removing the pointer network generally leads to weaker performance. One notable exception is Go, where the variant without the pointer network performs better in terms of F1 score. Our investigation revealed that there seems to be some violation of the i.i.d. assumption in the split provided by the creators of the dataset. In Figure 7 we show that in the test partition of the Go dataset, the share of tokens from the labels also occurring in the methods’ bodies – exactly the scenario where the pointer network can improve predictions – is substantially lower compared to the train/validation partitions. ", + "bbox": [ + 174, + 460, + 825, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remarkably, the multi-language Context-only variant (i.e. without access to the Structure) performs substantially worse than the full multi-language variant. This highlights that Structure is crucial to exploit the commonalities of different programming languages. Also notably, the GREAT baseline’s results also improve substantially when trained in the multi-language setting, though it is still outperformed by our model. However, our results indicate that any representation learning model for code can benefit from multi-language training, especially when evaluating on low-resource languages. ", + "bbox": [ + 174, + 565, + 825, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 16 we present results using the sample-F1 score. At the time of submission, our monolingual model on Python outperforms the state of the art on the ogbg-code $2 ^ { 4 }$ (Hu et al., 2020) leaderboard by $112 \\%$ , and our multilanguage variant with LM pretraining outperforms it by $122 \\%$ . ", + "bbox": [ + 176, + 655, + 823, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative analysis of multilingual representations. Learning the CODE TRANSFORMER on multiple programming languages jointly provides us with embeddings in a shared representation space. In Fig. 4 we show a t-SNE (Maaten & Hinton, 2008) visualization of the ca. 40,000 snippets from the validation sets of four programming languages from the CSN dataset. For the embedding of a snippet, we use the representation of the method name in the final layer of the encoder. Note that the true method names are masked, i.e., inaccessible to the model. Further, note that in contrast to the monolingual embeddings learned by Kanade et al. (2020), the embeddings we evaluate are learned on the task of code summarization (though a similar study could be performed by using our model that was trained on the traditional language modeling pretraining task on multiple languages). ", + "bbox": [ + 174, + 713, + 825, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While snippets from the same language tend to be grouped together, there are interesting intersections of the different programming languages. For example, we highlight all methods whose names start with the subtoken parse or main. We see that snippets starting with parse are predominantly in an intersection region of Python and Javascript. From these snippets, we display the cross-language pair with smallest Euclidean embedding distance in Fig. 5. Remarkably, both snippets are effectively the same method in Javascript and Python – it is worth reminding that the model has never seen any parallel data during training. On the other hand, snippets starting with main tend to lie at an intersectional region of Python, Javascript, and Go. In Table 6 in the appendix we show additional cross-lingual pairs with similar embeddings, including a failure case of a main function, where embedding distance is not representative of semantic similarity. We attribute this to the fact that we used the encoder output embedding of the masked method name – the representation used by the decoder to predict the method name – as a snippet’s representation. Thus, snippets with completely different semantics (as is to be expected for very generic method names starting with main) have similar representations because they are predictive of the method name. ", + "bbox": [ + 176, + 844, + 823, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e875f6f92fb5baf876c7eada99cff2f2f6a1ae4c2907de273c6e3de07dd8a6a8.jpg", + "image_caption": [ + "Figure 4: t-SNE visualization of the CODE TRANSFORMER’s learned multilingual representations. " + ], + "image_footnote": [], + "bbox": [ + 223, + 80, + 774, + 304 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/b4f73ecaf41d7a935e9b22345a3d4464c58d6b6b17f19a5180425db1c0483a08.jpg", + "image_caption": [ + "Figure 5: Example snippet starting with parse (left) and its best embedding match from other languages (right). Both methods parse an input string to convert it into a boolean value. Note that even though they are semantically very similar, their method names are not; nonetheless, their representations in the CODE TRANSFORMER encoder reflect their semantic similarity. " + ], + "image_footnote": [], + "bbox": [ + 204, + 340, + 776, + 440 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 534, + 825, + 672 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As another qualitative insight into the representations learned by the CODE TRANSFORMER we have found that the language embeddings of languages with similar roots in language design are close; see Table 5 in the appendix for the pairwise similarity matrix of the learned language embeddings. ", + "bbox": [ + 176, + 679, + 823, + 722 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 174, + 742, + 318, + 758 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present the CODE TRANSFORMER, which learns jointly from Structure and Context of programs while only relying on language-agnostic features. Our model obtains state-of-the-art performance on code summarization on five different programming languages. Besides these results for training on individual languages, the language-agnostic nature of our model allows us to train it jointly on multiple programming languages. The resulting multilingual model substantially outperforms its mono-lingual variant on all programming languages, setting the state of the art on each language. We observe the largest improvement from multilingual training on the language with fewest resources, indicating that multilingual training can improve learning for less widely used programming languages. Remarkably, multilingual training only from Context does not lead to the same improvements, highlighting the benefits of combining Structure and Context. ", + "bbox": [ + 174, + 773, + 825, + 912 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mithun Acharya, Tao Xie, Jian Pei, and Jun Xu. Mining api patterns as partial orders from source code: From usage scenarios to specifications. In Proceedings of the the 6th Joint Meeting of the European Software Engineering Conference and the ACM SIGSOFT Symposium on The Foundations of Software Engineering, ESEC-FSE ’07, pp. 25–34, New York, NY, USA, 2007. Association for Computing Machinery. ISBN 9781595938114. doi: 10.1145/1287624.1287630. URL https://doi.org/10.1145/1287624.1287630. ", + "bbox": [ + 173, + 126, + 825, + 209 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Learning natural coding conventions. In Proceedings of the 22nd ACM SIGSOFT International Symposium on Foundations of Software Engineering, FSE 2014, pp. 281–293, New York, NY, USA, 2014. Association for Computing Machinery. ISBN 9781450330565. doi: 10.1145/2635868.2635883. URL https://doi.org/10.1145/2635868.2635883. ", + "bbox": [ + 173, + 218, + 825, + 289 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Suggesting accurate method and class names. In Proceedings of the 2015 10th Joint Meeting on Foundations of Software Engineering, ESEC/FSE 2015, pp. 38–49, New York, NY, USA, 2015. Association for Computing Machinery. ISBN 9781450336758. doi: 10.1145/2786805.2786849. URL https://doi. org/10.1145/2786805.2786849. ", + "bbox": [ + 173, + 297, + 825, + 367 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Miltiadis Allamanis, Hao Peng, and Charles Sutton. A convolutional attention network for extreme summarization of source code. In International conference on machine learning, pp. 2091–2100, 2016. ", + "bbox": [ + 173, + 377, + 821, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs with graphs. In International Conference on Learning Representations, 2018. URL https: //openreview.net/forum?id $=$ BJOFETxR-. ", + "bbox": [ + 173, + 428, + 823, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representation for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450356985. doi: 10.1145/3192366.3192412. URL https://doi.org/10.1145/3192366.3192412. ", + "bbox": [ + 174, + 479, + 825, + 550 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Uri Alon, Shaked Brody, Omer Levy, and Eran Yahav. code2seq: Generating sequences from structured representations of code. In International Conference on Learning Representations, 2019a. URL https://openreview.net/forum?id $=$ H1gKYo09tX. ", + "bbox": [ + 174, + 559, + 825, + 602 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. code2vec: Learning distributed representations of code. Proceedings of the ACM on Programming Languages, 3(POPL):1–29, 2019b. ", + "bbox": [ + 173, + 609, + 823, + 640 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018. ", + "bbox": [ + 173, + 647, + 825, + 704 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Avishkar Bhoopchand, Tim Rocktaschel, Earl Barr, and Sebastian Riedel. Learning python code ¨ suggestion with a sparse pointer network. arXiv preprint arXiv:1611.08307, 2016. ", + "bbox": [ + 173, + 713, + 825, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benjamin Bichsel, Veselin Raychev, Petar Tsankov, and Martin Vechev. Statistical deobfuscation of android applications. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, CCS ’16, pp. 343–355, New York, NY, USA, 2016. Association for Computing Machinery. ISBN 9781450341394. doi: 10.1145/2976749.2978422. URL https: //doi.org/10.1145/2976749.2978422. ", + "bbox": [ + 173, + 751, + 825, + 820 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pavol Bielik, Veselin Raychev, and Martin Vechev. Phog: probabilistic model for code. In International Conference on Machine Learning, pp. 2933–2942, 2016. ", + "bbox": [ + 169, + 829, + 823, + 859 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aleksandar Bojchevski, Johannes Klicpera, Bryan Perozzi, Amol Kapoor, Martin Blais, Benedek Rozemberczki, Michal Lukasik, and Stephan G ´ unnemann. Scaling graph neural networks with ¨ approximate pagerank. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery amp; Data Mining, KDD ’20, pp. 2464–2473, New York, NY, USA, 2020. ", + "bbox": [ + 176, + 867, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Association for Computing Machinery. ISBN 9781450379984. doi: 10.1145/3394486.3403296. \nURL https://doi.org/10.1145/3394486.3403296. ", + "bbox": [ + 183, + 103, + 821, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dylan Bourgeois. Learning representations of source code from structure and context. 2019. URL http://infoscience.epfl.ch/record/277163. ", + "bbox": [ + 173, + 142, + 823, + 170 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexis Conneau and Guillaume Lample. Cross-lingual language model pretraining. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´ R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 7059– 7069. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8928-cross-lingual-language-model-pretraining.pdf. ", + "bbox": [ + 173, + 179, + 825, + 250 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hanjun Dai, Bo Dai, and Le Song. Discriminative embeddings of latent variable models for structured data. In International conference on machine learning, pp. 2702–2711, 2016. ", + "bbox": [ + 173, + 258, + 821, + 289 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. ", + "bbox": [ + 173, + 296, + 823, + 339 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hoa Khanh Dam, Truyen Tran, and Trang Pham. A deep language model for software code. In FSE 2016: Proceedings of the Foundations Software Engineering International Symposium [The Conference], 2016. ", + "bbox": [ + 173, + 348, + 825, + 391 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨ graphs with fast localized spectral filtering. In Advances in neural information processing systems, pp. 3844–3852, 2016. ", + "bbox": [ + 171, + 400, + 823, + 443 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´ Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular fingerprints. In Advances in neural information processing systems, pp. 2224–2232, 2015. ", + "bbox": [ + 176, + 452, + 821, + 496 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Patrick Fernandes, Miltiadis Allamanis, and Marc Brockschmidt. Structured neural summarization. In International Conference on Learning Representations, 2019. URL https: //openreview.net/forum?id $=$ H1ersoRqtm. ", + "bbox": [ + 176, + 503, + 820, + 547 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2, pp. 729–734. IEEE, 2005. ", + "bbox": [ + 174, + 556, + 823, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs. In NIPS, 2017. ", + "bbox": [ + 173, + 608, + 821, + 637 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Vincent J. Hellendoorn, Charles Sutton, Rishabh Singh, Petros Maniatis, and David Bieber. Global relational models of source code. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id $=$ B1lnbRNtwr. ", + "bbox": [ + 174, + 645, + 825, + 689 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Abram Hindle, Earl T. Barr, Zhendong Su, Mark Gabel, and Premkumar Devanbu. On the naturalness of software. In Proceedings of the 34th International Conference on Software Engineering, ICSE ’12, pp. 837–847. IEEE Press, 2012. ISBN 9781467310673. ", + "bbox": [ + 174, + 698, + 825, + 741 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 171, + 750, + 821, + 779 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin (eds.), Advances in Neural Information Processing Systems, volume 33, pp. 22118–22133. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/file/ fb60d411a5c5b72b2e7d3527cfc84fd0-Paper.pdf. ", + "bbox": [ + 174, + 787, + 825, + 872 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt. Codesearchnet challenge: Evaluating the state of semantic code search. arXiv preprint arXiv:1909.09436, 2019. ", + "bbox": [ + 174, + 882, + 823, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aditya Kanade, Petros Maniatis, Gogul Balakrishnan, and Kensen Shi. Learning and evaluating contextual embedding of source code. In Proceedings of the 37th International Conference on Machine Learning, 2020. ", + "bbox": [ + 174, + 103, + 821, + 146 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR), 2017. ", + "bbox": [ + 169, + 155, + 821, + 184 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Johannes Klicpera, Aleksandar Bojchevski, and Stephan Gunnemann. Predict then propagate: ¨ Graph neural networks meet personalized pagerank. In International Conference on Learning Representations (ICLR), 2019a. ", + "bbox": [ + 174, + 191, + 823, + 234 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Johannes Klicpera, Stefan Weißenberger, and Stephan Gunnemann. Diffusion improves graph learn-¨ ing. In Conference on Neural Information Processing Systems (NeurIPS), 2019b. ", + "bbox": [ + 169, + 242, + 823, + 272 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Johannes Klicpera, Janek Groß, and Stephan Gunnemann. Directional message passing for molec- ¨ ular graphs. In International Conference on Learning Representations (ICLR), 2020. ", + "bbox": [ + 173, + 280, + 823, + 309 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jian Li, Yue Wang, Michael R Lyu, and Irwin King. Code completion with neural attention and pointer networks. arXiv preprint arXiv:1711.09573, 2017. ", + "bbox": [ + 171, + 318, + 823, + 347 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. arXiv preprint arXiv:1511.05493, 2015. ", + "bbox": [ + 171, + 354, + 823, + 385 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(Nov):2579–2605, 2008. ", + "bbox": [ + 171, + 392, + 823, + 421 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Chris Maddison and Daniel Tarlow. Structured generative models of natural source code. In International Conference on Machine Learning, pp. 649–657, 2014. ", + "bbox": [ + 171, + 430, + 821, + 459 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lili Mou, Ge Li, Lu Zhang, Tao Wang, and Zhi Jin. Convolutional neural networks over tree structures for programming language processing. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, AAAI’16, pp. 1287–1293. AAAI Press, 2016. ", + "bbox": [ + 174, + 467, + 823, + 511 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Trong Duc Nguyen, Anh Tuan Nguyen, Hung Dang Phan, and Tien N. Nguyen. Exploring api embedding for api usages and applications. In Proceedings of the 39th International Conference on Software Engineering, ICSE ’17, pp. 438–449. IEEE Press, 2017. ISBN 9781538638682. doi: 10.1109/ICSE.2017.47. URL https://doi.org/10.1109/ICSE.2017.47. ", + "bbox": [ + 173, + 518, + 825, + 575 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order to the web. Technical report, Stanford InfoLab, 1999. ", + "bbox": [ + 174, + 583, + 823, + 613 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019. ", + "bbox": [ + 173, + 621, + 823, + 650 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Veselin Raychev, Martin Vechev, and Eran Yahav. Code completion with statistical language models. In Proceedings of the 35th ACM SIGPLAN Conference on Programming Language Design and Implementation, PLDI ’14, pp. 419–428, New York, NY, USA, 2014. Association for Computing Machinery. ISBN 9781450327848. doi: 10.1145/2594291.2594321. URL https://doi. org/10.1145/2594291.2594321. ", + "bbox": [ + 173, + 659, + 825, + 729 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008. ", + "bbox": [ + 173, + 737, + 821, + 766 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. In Aldo Gangemi, Roberto Navigli, Maria-Esther Vidal, Pascal Hitzler, Raphael Troncy, Laura Hollink, Anna Tordai, and ¨ Mehwish Alam (eds.), The Semantic Web, pp. 593–607, Cham, 2018. Springer International Publishing. ISBN 978-3-319-93417-4. ", + "bbox": [ + 173, + 775, + 825, + 844 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pp. 464–468, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. doi: 10.18653/v1/N18-2074. URL https://www.aclweb.org/anthology/N18-2074. ", + "bbox": [ + 174, + 853, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. Ordered neurons: Integrating tree structures into recurrent neural networks. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id ${ . } =$ B1l6qiR5F7. ", + "bbox": [ + 178, + 103, + 823, + 146 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Vighnesh Shiv and Chris Quirk. Novel positional encodings to enable tree-based transformers. In Advances in Neural Information Processing Systems, pp. 12081–12091, 2019. ", + "bbox": [ + 171, + 155, + 823, + 184 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pp. 1631–1642, 2013. ", + "bbox": [ + 173, + 193, + 825, + 250 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075, 2015. ", + "bbox": [ + 173, + 257, + 821, + 287 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017. ", + "bbox": [ + 174, + 295, + 825, + 338 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li ´ o, and Yoshua \\` Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rJXMpikCZ. ", + "bbox": [ + 173, + 347, + 825, + 390 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in neural information processing systems, pp. 2692–2700, 2015. ", + "bbox": [ + 173, + 398, + 823, + 428 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Song Wang, Devin Chollak, Dana Movshovitz-Attias, and Lin Tan. Bugram: Bug detection with n-gram language models. In Proceedings of the 31st IEEE/ACM International Conference on Automated Software Engineering, ASE 2016, pp. 708–719, New York, NY, USA, 2016. Association for Computing Machinery. ISBN 9781450338455. doi: 10.1145/2970276.2970341. URL https://doi.org/10.1145/2970276.2970341. ", + "bbox": [ + 173, + 435, + 825, + 507 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations, 2019. URL https: //openreview.net/forum?id $=$ ryGs6iA5Km. ", + "bbox": [ + 174, + 516, + 823, + 559 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pp. 5754–5764, 2019. ", + "bbox": [ + 176, + 568, + 823, + 611 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hierarchical graph representation learning with differentiable pooling. In Advances in neural information processing systems, pp. 4800–4810, 2018. ", + "bbox": [ + 173, + 619, + 825, + 662 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jiaxuan You, Rex Ying, and Jure Leskovec. Position-aware graph neural networks. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 7134– 7143. PMLR, 09–15 Jun 2019. URL http://proceedings.mlr.press/v97/you19b. html. ", + "bbox": [ + 173, + 670, + 825, + 741 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Muhan Zhang and Yixin Chen. Link prediction based on graph neural networks. In Advances in Neural Information Processing Systems, pp. 5165–5175, 2018. ", + "bbox": [ + 176, + 751, + 823, + 779 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/26bec4900b3fdc755545a32837761c235cd99ba3567c283f36c397e15113a344.jpg", + "table_caption": [ + "Table 5: Pairwise cosine similarities of the learned language embeddings of the CODE TRANSFORMER. " + ], + "table_footnote": [], + "table_body": "
PythonJavascriptGoRuby
Python1.000.430.420.79
Javascript0.431.000.840.39
Go0.430.841.000.38
Ruby0.790.390.381.00
", + "bbox": [ + 328, + 101, + 669, + 186 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "ACKNOWLEDGEMENTS ", + "text_level": 1, + "bbox": [ + 176, + 252, + 367, + 267 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We are grateful to Dylan Bourgeois for having paved the way to this research contribution with his thesis work (Bourgeois, 2019). We further thank Simon Geisler for his helpful suggestions and proofreading the paper, as well as the anonymous reviewers for their constructive feedback and fruitful discussions. ", + "bbox": [ + 174, + 282, + 825, + 338 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "This research was supported by the TUM International Graduate School of Science and Engineering (IGSSE). Stanford University is supported by DARPA under Nos. N660011924033 (MCS); ARO under Nos. W911NF-16-1- 0342 (MURI), W911NF-16-1-0171 (DURIP); NSF under Nos. OAC1835598 (CINES), OAC-1934578 (HDR), CCF-1918940 (Expeditions), IIS-2030477 (RAPID); Stanford Data Science Initiative, Wu Tsai Neurosciences Institute, Chan Zuckerberg Biohub, Amazon, JPMorgan Chase, Docomo, Hitachi, JD.com, KDDI, NVIDIA, Dell, Toshiba, Intel, and UnitedHealth Group. Jure Leskovec is a Chan Zuckerberg Biohub investigator. ", + "bbox": [ + 173, + 345, + 825, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 464, + 299, + 479 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.1 DISTANCE ENCODING FUNCTION ", + "text_level": 1, + "bbox": [ + 176, + 496, + 444, + 511 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For encoding scalar relation values via vectors we employ encoding functions $\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }$ , where $d$ is the model’s embedding dimension. We choose the popular sinusoidal encoding function presented in Vaswani et al. (2017): ", + "bbox": [ + 174, + 522, + 825, + 564 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/f46894c730c22bd189c627fee0b9dffb06a45fdfb55c7f527bb4fe0d0c5327e8.jpg", + "text": "$$\n\\phi ( r _ { i j } ) _ { 2 k } = \\sin ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) \\qquad \\phi ( r _ { i j } ) _ { 2 k + 1 } = \\cos ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 258, + 570, + 738, + 598 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where $1 \\leq k < d / 2$ is the position in the encoding vector and $M$ is some constant; we adopt $M = 1 0 , 0 0 0$ as chosen by (Vaswani et al., 2017). Note that the distance encoding functions have no trainable parameters. ", + "bbox": [ + 174, + 606, + 825, + 648 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.2 MULTILINGUAL REPRESENTATION ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 666, + 526, + 680 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In Table 5, we show the pairwise cosine similarities of the learned language embeddings of the CODE TRANSFORMER. We can see that the pairs Python-Ruby and Javascript-Go have similar language embeddings. This aligns well with roots of language design and common use cases of the languages. ", + "bbox": [ + 173, + 690, + 825, + 748 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Moreover, in Table 6, we show selected snippets starting with is, main, or load (left) and their best embedding matches from other languages (right). ", + "bbox": [ + 173, + 755, + 823, + 784 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.3 DATA PREPROCESSING ", + "text_level": 1, + "bbox": [ + 176, + 800, + 375, + 815 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.3.1 TEXTUAL CODE SNIPPET PREPROCESSING ", + "text_level": 1, + "bbox": [ + 176, + 827, + 522, + 842 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "1. Tokenize code snippets with Pygments language-specific tokenizer. ", + "bbox": [ + 210, + 848, + 676, + 863 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "2. Remove comments (both multi-line, single-line and doc comments). The comment token types. pygments.token.Comment and pygments.token.Literal.String.Doc that are generated by Pygments are used to identify comments. ", + "bbox": [ + 214, + 867, + 825, + 924 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/dfeb9bd55674c1d88bd8727a7ae6e6c8b0e9d1b458288922467db71461c08de5.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6: Selected snippets starting with is, main, or load (left) and their best embedding matches from other languages (right). " + ], + "table_body": "
Published as a conference paper at ICLR 2021
function isEqualArray(a,b){ if (a return true;
if ((a === undefined) undefined)){ return false;func areSameFloat32Array(a,b []float32) bool if len(a) != len(b){ return false
var =a.length; if (i !== b.length){ return false;for i := O;i < len(a);i++{ if a[i]!=b[i]{ return false J
while (i--){ if(a[i] !== b[i]){ return false; }return true
return true;
function main() var rawData $(' .HeaderTexture[data-login- user-email]') .data() ;func TaskSayHello(t *tasking.T){
if (rawData) 1 me={username := t.Flags.String("name") if username user, := user.Current ()
name:rawData.loginUserName, mail: rawData.loginUserEmail };username = user.Name
getCartId(function (cart){ me.cart = cart;} ift.Flags.Bool("verbose")
injectMenu(); refreshUsers(actions.updateUsers); listenOnChanges(onChange);t.Logf("Hello %s, the time now is %s\\n", username,time.Now())
listenonorderConfirm(onConfirm); );}else{ t.Logf("Hello %s\\n",username)
} else{ caliback('no user');
func Backup(filename string)error{
info,err := os.Stat(filename) if err != nil
if os.IsNotExist(err){
return nil
return err
def _load_rule_file(self, filename):if info.Size() == 0{ return nil
if not (os.path.exists(filename)):}
sys.stderr.write(files,err := filepath.Glob(
"rflint: %s: No such file or "directory\\n" % filenamefilename + _BACKUP_SUFFIX
return) if err nil 1
try:return err
basename = os.path.basename(filename)}
(name,ext)= os.path.splitext(basename) imp.load_source(name,filename) numBackup := byte(1)
except Exception as e: sys.stderr.write(if len(files) !=0{
"rflint: %s: exception whilelastFile := files[len(files)-1]
"loading:%s\\n" % (filename, str(e))numBackup = lastFile[len(lastFile)-2] + 1
if numBackup >
numBackup = '1'
else{
'1'
numBackup =
return Copy (filename,
fmt.Sprintf("%s+%s~",
string(numBackup)))
", + "bbox": [ + 173, + 15, + 866, + 864 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/e37f9cb44d930542340c5ad3d88619f8535c0f5512daf8c0c8d8a050432bff4b.jpg", + "image_caption": [ + "Figure 6: Example snippet and its corresponding AST obtained from GitHub Semantic. " + ], + "image_footnote": [], + "bbox": [ + 300, + 99, + 686, + 333 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "3. Empty lines are removed. ", + "bbox": [ + 214, + 388, + 405, + 402 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "4. Hard coded strings and numbers are replaced with a special [MASK STRING] and [MASK NUMBER] token. ", + "bbox": [ + 210, + 411, + 825, + 439 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "5. Indentation style of the code snippet is detected and whitespace characters at the beginning of a line are replaced with a single [INDENT] or [DEDENT] token when indentation changes. ", + "bbox": [ + 212, + 446, + 825, + 489 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "6. Tokens are further split into sub tokens, e.g., setBottomHeight [‘set’, ‘bottom’, ‘height’]. Throughout our experiments, we use 5 input sub tokens. If a token consists of less than 5 sub tokens, the remaining spaces are filled with a special [PAD] token. ", + "bbox": [ + 212, + 497, + 825, + 540 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "7. Any remaining tokens that only consist of white spaces are removed. The only white space characters that are kept are line breaks $^ { \\bullet } \\backslash \\boldsymbol { \\mathrm { n } } ^ { \\prime }$ . ", + "bbox": [ + 210, + 547, + 821, + 577 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "8. Any code snippets where the Pygments tokenizer cannot parse a token are discarded. ", + "bbox": [ + 207, + 584, + 803, + 599 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "A.3.2 STAGE 1 PREPROCESSING (GENERATION OF ASTS) ", + "text_level": 1, + "bbox": [ + 174, + 618, + 588, + 633 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "1. Stripped code snippets are used to generate language-specific ASTs. For Java, we use the AST parser from the java-parser project. The ASTs contain node types and source ranges. For Python, JavaScript, Ruby and Go, we use semantic. ", + "bbox": [ + 214, + 645, + 823, + 686 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "2. Snippets that lead to an AST parse error are discarded. ", + "bbox": [ + 212, + 695, + 606, + 709 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "3. We calculate a mapping between tokens and nodes in the AST. Every token is assigned to the node in the AST with shortest source range that still encompasses the source range of the token. ", + "bbox": [ + 210, + 718, + 823, + 758 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "To find such a node, we originally intended to make use of the assumption that source ranges of child nodes do not overlap. Then, one could easily find the node with smallest encompassing source range by greedily selecting at every layer in the AST the child that encompasses the token’s source range (there can only be at most one child that fulfills this). However, this assumption does not hold for all ASTs (see Figure 6 for an example). As a heuristic, we greedily select the child node with the shorter source range in case there were multiple child nodes with encompassing source ranges. This approximation seems to be sufficient in our case, and limits runtime as we do not have to consider multiple paths in the AST. It is also sufficient to stop when no child node encompasses the source range of the token, as in ASTs the source ranges of child nodes are always contained in the source ranges of their parent. ", + "bbox": [ + 232, + 760, + 825, + 911 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "1. Tokens are vocabularized. Any token occurring less than 100 times in the training set is replaced by an $ token. ", + "bbox": [ + 210, + 125, + 825, + 154 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "2. We calculate multiple pair-wise relations between nodes in the AST: ", + "bbox": [ + 212, + 159, + 689, + 172 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "• Personalized Page Rank (PPR) We interpret the negative logarithm of PPR as a distance. We use a teleport probability of $\\alpha = 0 . 1 5$ and a threshold of $e ^ { - 5 }$ , i.e., anything with $- \\log P P R > 5$ is considered unreachable \n• Shortest path length between two nodes \n• Ancestor shortest paths (bidirectional). That is, the parent has an ancestor shortest path distance of 1 to all its children and the child has a distance of -1 to its parents. We consider nodes that are not ancestors or descendants of a node (i.e. not reachable by following only parent or only child relations) as not connected in the ancestor shortest paths relation. We encode this with a very large value in their distance; we have found a value of $1 , 0 0 0$ to work well in practice. \n• Next sibling shortest paths (bidirectional, analogous to the ancestor shortest paths) ", + "bbox": [ + 251, + 178, + 825, + 354 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Note that the ancestor shortest paths and next sibling shortest paths are required because treating the AST as a normal graph leads to ambiguity. In a graph, the neighbors of a node have no ordering; however in the AST, the order of the children of a node reflects their order in the code. Therefore, we explicitly include the next sibling shortest paths. The ancestor shortest paths would not be required if we treated the AST as a directed graph; in this case, however, a leaf node could not reach any other node in the AST, and therefore both PPR and shortest path length are not useful in this case. Therefore, we model the AST as undirected and inject the ancestor / child edges to avoid ambiguity. ", + "bbox": [ + 232, + 359, + 823, + 470 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "3. Distance values are binned into 32 bins using area-based exponential binning with a growth factor of 1.3, i.e., the area of a bin’s rectangle ( $x$ : bin range, $y$ : number of values in bin) will be approximately 1.3 times bigger for the next bin (going away from the bin that contains the zero value). Additionally, for discrete distance measures (such as sequence distance or shortest path length), we hard-code 9 values around 0 to have their own bins. For instance, on the sequence distance the values $- 4 , - 3 , \\dots , 4$ have their individual bins, and around those values we employ the exponential binning. ", + "bbox": [ + 212, + 476, + 825, + 573 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "4. Punctuation tokens (such as points or brackets) are removed from the input sequence, as experiments showed that their presence does not improve performance but slows down training due to bigger input sizes. ", + "bbox": [ + 214, + 579, + 823, + 621 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "5. Snippets that are longer than MAX NUM TOKENS after punctuation tokens are removed are discarded from the training set. Throughout our experiments, we use MAX NUM TOKENS $= ~ 5 1 2$ . During evaluation on the test set, we use MAX NUM TOKENS $= ~ 1 0 0 0$ . ", + "bbox": [ + 215, + 626, + 823, + 667 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "A.4 INPUT EMBEDDINGS TO THE MODEL ", + "text_level": 1, + "bbox": [ + 176, + 686, + 467, + 700 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Besides its five subtokens (e.g., [‘get’, ‘data’, ‘[PAD]’, ‘[PAD]’, ‘[PAD]’]), each input token has a token type (coming from the Pygments tokenizer) and an AST node type. The AST node type is the type of the node assigned to each respective token, as described in Section A.3.2. We concatenate the embeddings of the five subtokens, the token type, and the AST node type. Then, we apply a linear layer (without activation function) to project down to the model’s embedding dimension. ", + "bbox": [ + 174, + 712, + 825, + 795 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "A.5 INPUT TO THE GREAT BASELINE", + "text_level": 1, + "bbox": [ + 176, + 813, + 449, + 828 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "As mentioned in the main text, we also compare with GREAT Hellendoorn et al. (2020). Since their preprocessing pipeline is proprietary and could not be shared with us even after contacting the authors, we provide to GREAT the same AST distances as our model. Since GREAT uses edges instead of distances to encode relations in the Structure, we essentially threshold the ancestor, sibling, and shortest-paths distances and provide the edges where the distances are equal to 1 (including their edge types) to the model. ", + "bbox": [ + 174, + 840, + 825, + 922 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/26b079a48f72f11ac5901761fadc58949615f228994d83e0ca015c1b803de70e.jpg", + "table_caption": [ + "(a) CODE TRANSFORMER ", + "Table 7: Code Summarization hyperparameters " + ], + "table_footnote": [], + "table_body": "
HyperparameterValue
ActivationGELU
Input Nonlinearitytanh
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
", + "bbox": [ + 272, + 122, + 483, + 250 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/d450c7940cdf06f24a236ed0b2ea7649815a8d485cbb72d5b067d40223c84355.jpg", + "table_caption": [ + "(b) GREAT (Hellendoorn et al., 2020) " + ], + "table_footnote": [], + "table_body": "
HyperparameterValue
ActivationGELU
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
", + "bbox": [ + 519, + 136, + 715, + 247 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "A.6 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 176, + 303, + 377, + 316 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Table 7 shows hyperparameters of our models for code summarization. For all our experiments, we use a Transformer Decoder with one layer and teacher forcing to generate 6 output sub tokens. We also employ label smoothing of 0.1. As optimizer, we use Adam with a learning rate of $8 e ^ { - 5 }$ and weight decay of $3 e ^ { - 5 }$ . Batch size during training is 8 with a simulated batch size of 128 achieved by gradient accumulation. ", + "bbox": [ + 173, + 328, + 825, + 398 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Apart from comparing the CODE TRANSFORMER to baselines, we performed the following hyperparameter comparisons and ablation studies: ", + "bbox": [ + 174, + 419, + 821, + 446 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "• CODE TRANSFORMER (structure-only) Using only AST information as input, i.e., masking all tokens that do not correspond to a leaf of the AST, and removing the token distance as a relation to be used by the model. Further, token types are not fed into the model. \n• CODE TRANSFORMER (context-only) Here, we do not include any information on the AST (i.e. node types and distances on the AST). This is effectively the XLNet backbone plus encoding of the token type returned by the tokenizer. \n• CODE TRANSFORMER (Max-Dist.) Applying a Max Distance Mask of 5 to the shortest paths distance (i.e., model cannot see a node that is more than 5 hops away no matter how small the other distances are). Early results showed that, as expected, results deteriorate substantially when limiting our model’s receptive field. Hence, we do not include these results in this work. \n• Using 16 and 64 bins instead of 32 bins. This had no noticeable effect on performance. ", + "bbox": [ + 217, + 460, + 825, + 670 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "A.7 CODE SUMMARIZATION EXAMPLES", + "text_level": 1, + "bbox": [ + 178, + 686, + 462, + 700 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "In the Tables 8, 9, 10, 11, 12, 13, 14 and 15 we present example functions from the Java-small dataset along with the different models’ predictions for the function name. ", + "bbox": [ + 173, + 712, + 825, + 741 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "public Summation next() { return parts $[ \\dot { 1 } + + ]$ ; } ", + "bbox": [ + 173, + 118, + 395, + 156 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/04303ffb667966783611a78b38db4a06a17d54b3303576e3d6f2082eb78c4e3f.jpg", + "table_caption": [ + "Table 8: The CODE TRANSFORMER is the only model to correctly identify the notion of getting the next entry. " + ], + "table_footnote": [], + "table_body": "
ModelPrediction
GREAT get x map
code2seq get parts
Ours w/o structureget
CODE TRANSFORMERget next
Ground Truthnext
", + "bbox": [ + 377, + 161, + 617, + 258 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "private Path findCacheFile(Path[] cacheFiles, String fileName) { if (cacheFiles $! =$ null && cacheFiles.length $> 0$ ) { for (Path file : cacheFiles) { if (file.getName().equals(fileName)) { return file; } return null; \n} ", + "bbox": [ + 173, + 343, + 740, + 467 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/7fce10876c7dbfae8cd1529bce06e79b9baef638e4b20f560a956fd0593414a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPrediction
GREATget path
code2seqfind file
Ours w/o structureget file
CODE TRANSFORMERfind cache
Ground Truthfind cache file
", + "bbox": [ + 367, + 474, + 629, + 573 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Table 9: The CODE TRANSFORMER is the only model to both recognize that the task is to find a file as well as the fact that it is about the cache. However, it did not correctly predict the file part of the method name. ", + "bbox": [ + 173, + 583, + 826, + 625 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "public int compare(Pair p1, Pair p2) { LoggedJob $\\ j 1 \\ = \\ \\mathtt { p 1 }$ .first(); LoggedJob $\\begin{array} { r } { \\dot { \\bf { \\ j } } 2 { \\bf { \\nu } } = { \\bf { \\ p } } 2 } \\end{array}$ .first(); return (j1.getSubmitTime() $<$ j2.getSubmitTime()) ? -1 : (j1.getSubmitTime() $= =$ j2.getSubmitTime()) ? 0 : 1; } ", + "bbox": [ + 173, + 671, + 748, + 758 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/fc7647e7d39190fcddd6003a2621a01161522b4db1837436041dcd5808f7523b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPrediction
GREATrun
code2seq get submit time
Ours w/o structurecompare
CODE TRANSFORMERcompare
Ground Truthcompare
", + "bbox": [ + 361, + 765, + 633, + 863 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Table 10: The CODE TRANSFORMER and the its context-only variant are the only models correctly recognizing the ‘compare’ template in the method body. ", + "bbox": [ + 174, + 872, + 823, + 901 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "public static MNTPROC fromValue(int value) { if (value $< ~ 0$ || value $> =$ values().length) { return null; } return values()[value]; \n} ", + "bbox": [ + 173, + 116, + 599, + 193 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/c731fa631d72aabd1f5e32632f679a09b671309e6980e8efeb5fd4f3681003b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPrediction
GREAT get value
code2seq get value
Ours w/o structureto
CODE TRANSFORMERfrom value
Ground Truthfrom value
", + "bbox": [ + 374, + 196, + 619, + 295 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Table 11: The CODE TRANSFORMER is the only model to recognize that the snippet is similar to a static factory method which is often preceded with from. ", + "bbox": [ + 171, + 305, + 825, + 334 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "private Iterable listRootBlobs(String aPrefix, ", + "bbox": [ + 174, + 376, + 700, + 388 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "boolean useFlatBlobListing, \nEnumSet listingDetails, \nBlobRequestOptions options, \nOperationContext opContext) ", + "bbox": [ + 565, + 388, + 947, + 439 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "throws StorageException, URISyntaxException { ", + "bbox": [ + 514, + 440, + 928, + 452 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "CloudBlobDirectoryWrapper directory $=$ this.container.getDirectoryReference(aPrefix); return directory.listBlobs(null, useFlatBlobListing, listingDetails, options, opContext); ", + "bbox": [ + 186, + 453, + 946, + 491 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "} ", + "bbox": [ + 174, + 491, + 184, + 502 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/4184371f3c5b40e5706dc3536eb808edba7ea9235136630720ca50cbd799e93d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPrediction
GREATlist blobs
code2seqlist blobs
Ours w/o structurelist blobs
CODE TRANSFORMERlist blobs by prefix
Ground Truthlist root blobs
", + "bbox": [ + 351, + 507, + 643, + 604 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Table 12: All models could correctly identify the listBlobs() call in the return statement. However, the CODE TRANSFORMER additionally comprehended that the specified prefix is quite important. ", + "bbox": [ + 171, + 614, + 823, + 643 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "private static void dumpOpCounts(EnumMap $<$ FSEditLogOpCodes, Holder $<$ > opCounts) { StringBuilder sb $=$ new StringBuilder(); sb.append(\"Summary of operations loaded from edit log:\\n \"); Joiner.on(\"\\n \").withKeyValueSeparator(\" $^ { 1 1 } =$ \").appendTo(sb, opCounts); FSImage.LOG.debug(sb.toString()); \n} ", + "bbox": [ + 169, + 685, + 942, + 757 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/3539e0325fcec95f6f2a244b7e8480d60028213ed3f48a53d7797b35801c9743.jpg", + "table_caption": [ + "Table 13: Only the CODE TRANSFORMER could correctly identify that it is the op counts that should be logged. " + ], + "table_footnote": [], + "table_body": "
ModelPrediction
GREATappend
code2seqadd
Ours w/o structurelog
CODE TRANSFORMERlog op counts
Ground Truthdump op counts
", + "bbox": [ + 359, + 765, + 633, + 864 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "static String execCommand(File f, String... cmd) throws IOException { String[] args $=$ new String[cmd.length $^ +$ 1]; System.arraycopy(cmd, 0, args, 0, cmd.length); args[cmd.length] $=$ f.getCanonicalPath(); String output $=$ Shell.execCommand(args); return output; ", + "bbox": [ + 173, + 190, + 782, + 268 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/3c5bfe70032cf08db67ee0f3135ec46e80e42b50a65cc9210b7b528695c79ab7.jpg", + "table_caption": [ + "Table 14: Only the CODE TRANSFORMER and code2seq could identify that the relevant part of the method is concerned with executing a command instead of returning something. " + ], + "table_footnote": [], + "table_body": "
ModelPrediction
GREATget canonical path
code2seqexec
Ours w/o structure get output
CODE TRANSFORMERexec command
Ground Truthexec command
", + "bbox": [ + 352, + 284, + 642, + 382 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "protected void subView(Class cls) { ", + "text_level": 1, + "bbox": [ + 173, + 612, + 650, + 625 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "indent(of(ENDTAG)); \nsb.setLength(0); \nout.print(sb.append('[').append(cls.getName()).append(']').toString()); \nout.println(); ", + "bbox": [ + 178, + 621, + 834, + 678 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/0c4c2219f8a76155c82fd9cf1ddec51733b172b4b12ee945147b97301cb67ad3.jpg", + "table_caption": [ + "Table 15: Only the CODE TRANSFORMER was able to link the print functionality to the object that should be printed, which can only be inferred from the object’s class in the method parameters. " + ], + "table_footnote": [], + "table_body": "
ModelPrediction
GREATprint
code2seqprint
Ours w/o structureprint
CODE TRANSFORMERprint sub view
Ground Truthsub view
", + "bbox": [ + 364, + 691, + 629, + 790 + ], + "page_idx": 20 + }, + { + "type": "image", + "img_path": "images/1225dc2212652b0900da96594dc1baedeb731793c2a7f99b2f39f0a306a759b8.jpg", + "image_caption": [ + "Figure 7: Share of tokens in the labels also occurring in the bodies of methods. " + ], + "image_footnote": [], + "bbox": [ + 276, + 106, + 722, + 315 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/f9846de1ac74fbb68c2e86b16559380e7f8fc1f6b4b397c66475b197f4726324.jpg", + "table_caption": [ + "Table 16: Code summarization results on the CSN dataset (sample-F1). " + ], + "table_footnote": [], + "table_body": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq111-11=1111-
GREAT34.9331.1231.6129.6924.2425.5525.6921.4922.1848.3845.9745.71
Ours w/o structure36.8732.1732.9731.3025.0326.6431.4325.3426.6349.7846.7346.69
Ours w/o pointer net38.7731.7233.2732.7025.5027.3332.1230.1729.3653.0948.7049.26
Ours36.6833.8633.8433.3627.5529.0231.5324.7226.4352.0047.3547.93
code2seq (Multilanguage)-====-
GREAT (Multilanguage)35.7330.8131.7431.4926.1727.4129.7224.2025.4350.3247.9447.66
Ours w/o structure (Mult.)36.7829.9231.5832.6026.0227.7431.7126.0727.2451.9147.5848.15
Ours w/o pointer (Mult.)37.1830.5232.0433.9525.9228.1132.7625.0427.0153.5048.5449.35
Ours (Multilanguage)38.1033.3234.1834.2928.6930.0833.3028.3329.2953.8650.4650.61
Ours (Mult. + Finetune)38.2932.4133.6534.4328.2829.9132.8927.1528.4953.8550.85
Ours (Mult.+ LMPretrain)50.81
38.9734.7735.3435.2330.2631.3833.7329.1529.9455.3152.0352.13
", + "bbox": [ + 176, + 361, + 823, + 523 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "A.8 ESTIMATION OF POINTER NETWORK POTENTIAL ", + "text_level": 1, + "bbox": [ + 174, + 575, + 557, + 590 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "In Table 2 we observe that the pointer network improves the F1 score for all languages except Go, where counterintuitively it leads to reduced performance as measured by F1 score on the test set (while it improves by about 3 points on validation). To investigate this, in Figure 7 we plot the share of tokens in the labels also occurring in the bodies of methods in the different languages. Intuitively, this gives an indication on how much gain we can expect from using a pointer network. If the share were zero, then no token in the labels ever occur in the bodies of the methods, so the pointer network cannot improve the prediction by pointing at the input. We see that for Go, there is a strong mismatch between the test partition and the train/validation partitions, which much fewer tokens from the labels occurring in the bodies of methods on test compared to train/validation. Thus, we attribute the drop in performance observed by adding a pointer network on Go to this apparent violation of the i.i.d. assumption. ", + "bbox": [ + 173, + 602, + 825, + 755 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "A.9 CODE SUMMARIZATION RESULTS ON THE CSN DATASET (SAMPLE-F1) ", + "text_level": 1, + "bbox": [ + 174, + 771, + 710, + 786 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "In Table 16, we present our results on the CSN dataset as measured by the sample-F1 score. ", + "bbox": [ + 171, + 797, + 771, + 813 + ], + "page_idx": 21 + } +] \ No newline at end of file diff --git a/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_middle.json b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..f2b8b946a678057b9ea2b12f82a1cf15d06d52b6 --- /dev/null +++ b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_middle.json @@ -0,0 +1,53618 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "LANGUAGE-AGNOSTIC REPRESENTATION LEARNING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 495, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 495, + 118 + ], + "score": 1.0, + "content": "OF SOURCE CODE FROM STRUCTURE AND CONTEXT", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 285, + 169 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 259, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 259, + 147 + ], + "score": 1.0, + "content": "Daniel Zugner, Tobias Kirschstein ¨", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 242, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 242, + 158 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 157, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 113, + 157, + 286, + 170 + ], + "score": 1.0, + "content": "{zuegnerd,kirschto}@in.tum.de", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 365, + 135, + 498, + 168 + ], + "lines": [ + { + "bbox": [ + 364, + 135, + 437, + 147 + ], + "spans": [ + { + "bbox": [ + 364, + 135, + 437, + 147 + ], + "score": 1.0, + "content": "Michele Catasta", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 364, + 145, + 446, + 158 + ], + "spans": [ + { + "bbox": [ + 364, + 145, + 446, + 158 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 364, + 158, + 498, + 168 + ], + "spans": [ + { + "bbox": [ + 364, + 158, + 498, + 168 + ], + "score": 1.0, + "content": "pirroh@cs.stanford.edu", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 113, + 186, + 234, + 218 + ], + "lines": [ + { + "bbox": [ + 111, + 185, + 176, + 197 + ], + "spans": [ + { + "bbox": [ + 111, + 185, + 176, + 197 + ], + "score": 1.0, + "content": "Jure Leskovec", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 195, + 194, + 209 + ], + "spans": [ + { + "bbox": [ + 111, + 195, + 194, + 209 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 208, + 234, + 219 + ], + "spans": [ + { + "bbox": [ + 113, + 208, + 234, + 219 + ], + "score": 1.0, + "content": "jure@cs.stanford.edu", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 365, + 186, + 494, + 218 + ], + "lines": [ + { + "bbox": [ + 365, + 186, + 458, + 197 + ], + "spans": [ + { + "bbox": [ + 365, + 186, + 458, + 197 + ], + "score": 1.0, + "content": "Stephan Gunnemann ¨", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 197, + 494, + 208 + ], + "spans": [ + { + "bbox": [ + 364, + 197, + 494, + 208 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 208, + 486, + 219 + ], + "spans": [ + { + "bbox": [ + 364, + 208, + 486, + 219 + ], + "score": 1.0, + "content": "guennemann@in.tum.de", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 278, + 248, + 333, + 259 + ], + "lines": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "spans": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 143, + 271, + 468, + 424 + ], + "lines": [ + { + "bbox": [ + 142, + 271, + 470, + 283 + ], + "spans": [ + { + "bbox": [ + 142, + 271, + 470, + 283 + ], + "score": 1.0, + "content": "Source code (Context) and its parsed abstract syntax tree (AST; Structure) are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "score": 1.0, + "content": "two complementary representations of the same computer program. Tradition-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 293, + 469, + 305 + ], + "spans": [ + { + "bbox": [ + 141, + 293, + 469, + 305 + ], + "score": 1.0, + "content": "ally, designers of machine learning models have relied predominantly either on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "score": 1.0, + "content": "Structure or Context. We propose a new model, which jointly learns on Context", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 315, + 470, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 470, + 327 + ], + "score": 1.0, + "content": "and Structure of source code. In contrast to previous approaches, our model uses", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "score": 1.0, + "content": "only language-agnostic features, i.e., source code and features that can be com-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 337, + 470, + 349 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 470, + 349 + ], + "score": 1.0, + "content": "puted directly from the AST. Besides obtaining state-of-the-art on monolingual", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 348, + 470, + 360 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 470, + 360 + ], + "score": 1.0, + "content": "code summarization on all five programming languages considered in this work,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 359, + 469, + 371 + ], + "spans": [ + { + "bbox": [ + 141, + 359, + 469, + 371 + ], + "score": 1.0, + "content": "we propose the first multilingual code summarization model. We show that jointly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 370, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 141, + 370, + 469, + 382 + ], + "score": 1.0, + "content": "training on non-parallel data from multiple programming languages improves re-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 380, + 470, + 393 + ], + "spans": [ + { + "bbox": [ + 141, + 380, + 470, + 393 + ], + "score": 1.0, + "content": "sults on all individual languages, where the strongest gains are on low-resource", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 392, + 470, + 404 + ], + "spans": [ + { + "bbox": [ + 141, + 392, + 470, + 404 + ], + "score": 1.0, + "content": "languages. Remarkably, multilingual training only from Context does not lead", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 402, + 470, + 415 + ], + "spans": [ + { + "bbox": [ + 141, + 402, + 470, + 415 + ], + "score": 1.0, + "content": "to the same improvements, highlighting the benefits of combining Structure and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 414, + 321, + 426 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 321, + 426 + ], + "score": 1.0, + "content": "Context for representation learning on code.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 206, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 208, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 208, + 459 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "Machine learning for code is an active and growing area of research which aims at building models", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "that can learn semantically meaningful representations of programs. These embeddings can be used", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "on downstream tasks, such as code generation, bug detection, or code summarization. We focus", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "our work on two complementary data representations of programs: the source code (referred to as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "Context in this work), and the abstract syntax tree (AST; referred to as Structure). Traditionally,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "researchers and practitioners have decided to predominantly leverage either Structure or Context in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "their machine learning models. In this work, we show that jointly learning on Context and Structure", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 544, + 352, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 352, + 557 + ], + "score": 1.0, + "content": "improves representation learning on source code (see Fig. 1).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "The source code representation naturally lends itself to models from natural language process-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "ing (NLP), e.g., long short-term memory networks (Hochreiter & Schmidhuber, 1997) (LSTM) or", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "Transformers (Vaswani et al., 2017; Radford et al., 2019; Dai et al., 2019; Yang et al., 2019; Shaw", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "et al., 2018). On the other hand, models leveraging the structure representations are typically based", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "on graph neural networks (GNNs) (Kipf & Welling, 2017; Xu et al., 2019; Velickovi ˇ c et al., 2018; ´", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "You et al., 2019; Hamilton et al., 2017; Li et al., 2015; Klicpera et al., 2020). While the AST repre-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "sentation makes the highly structured nature of source code explicit to the models, since most GNNs", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "use the message-passing framework, their learned representations are inherently local and struggle", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 649, + 250, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 250, + 662 + ], + "score": 1.0, + "content": "to leverage long-range interactions.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "Recently, Hellendoorn et al. (2020) have explored models that can leverage several representations,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "including both Structure and Context. Their Graph Relational Embedding Attention Transformer", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "(GREAT) extends Shaw et al. (2018), which biases the self-attention computation in a localized", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "way given the underlying graph. The language-specific representations used by GREAT include a", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "combination of the data flow graph, control flow graph, syntactic edges (inspired by Allamanis et al.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 459, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 459, + 732 + ], + "score": 1.0, + "content": "(2018)), etc. which require specialized pipelines and static analysis tools to be obtained.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 79, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 98 + ], + "score": 1.0, + "content": "LANGUAGE-AGNOSTIC REPRESENTATION LEARNING", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 495, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 495, + 118 + ], + "score": 1.0, + "content": "OF SOURCE CODE FROM STRUCTURE AND CONTEXT", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 135, + 285, + 169 + ], + "lines": [ + { + "bbox": [ + 112, + 136, + 259, + 147 + ], + "spans": [ + { + "bbox": [ + 112, + 136, + 259, + 147 + ], + "score": 1.0, + "content": "Daniel Zugner, Tobias Kirschstein ¨", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 242, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 242, + 158 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 157, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 113, + 157, + 286, + 170 + ], + "score": 1.0, + "content": "{zuegnerd,kirschto}@in.tum.de", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 112, + 136, + 286, + 170 + ] + }, + { + "type": "list", + "bbox": [ + 365, + 135, + 498, + 168 + ], + "lines": [ + { + "bbox": [ + 364, + 135, + 437, + 147 + ], + "spans": [ + { + "bbox": [ + 364, + 135, + 437, + 147 + ], + "score": 1.0, + "content": "Michele Catasta", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 364, + 145, + 446, + 158 + ], + "spans": [ + { + "bbox": [ + 364, + 145, + 446, + 158 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 364, + 158, + 498, + 168 + ], + "spans": [ + { + "bbox": [ + 364, + 158, + 498, + 168 + ], + "score": 1.0, + "content": "pirroh@cs.stanford.edu", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + } + ], + "index": 5, + "bbox_fs": [ + 364, + 135, + 498, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 113, + 186, + 234, + 218 + ], + "lines": [ + { + "bbox": [ + 111, + 185, + 176, + 197 + ], + "spans": [ + { + "bbox": [ + 111, + 185, + 176, + 197 + ], + "score": 1.0, + "content": "Jure Leskovec", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 195, + 194, + 209 + ], + "spans": [ + { + "bbox": [ + 111, + 195, + 194, + 209 + ], + "score": 1.0, + "content": "Stanford University", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 208, + 234, + 219 + ], + "spans": [ + { + "bbox": [ + 113, + 208, + 234, + 219 + ], + "score": 1.0, + "content": "jure@cs.stanford.edu", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 111, + 185, + 234, + 219 + ] + }, + { + "type": "text", + "bbox": [ + 365, + 186, + 494, + 218 + ], + "lines": [ + { + "bbox": [ + 365, + 186, + 458, + 197 + ], + "spans": [ + { + "bbox": [ + 365, + 186, + 458, + 197 + ], + "score": 1.0, + "content": "Stephan Gunnemann ¨", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 197, + 494, + 208 + ], + "spans": [ + { + "bbox": [ + 364, + 197, + 494, + 208 + ], + "score": 1.0, + "content": "Technical University of Munich", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 208, + 486, + 219 + ], + "spans": [ + { + "bbox": [ + 364, + 208, + 486, + 219 + ], + "score": 1.0, + "content": "guennemann@in.tum.de", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 364, + 186, + 494, + 219 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 248, + 333, + 259 + ], + "lines": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "spans": [ + { + "bbox": [ + 276, + 246, + 336, + 262 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 143, + 271, + 468, + 424 + ], + "lines": [ + { + "bbox": [ + 142, + 271, + 470, + 283 + ], + "spans": [ + { + "bbox": [ + 142, + 271, + 470, + 283 + ], + "score": 1.0, + "content": "Source code (Context) and its parsed abstract syntax tree (AST; Structure) are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "spans": [ + { + "bbox": [ + 141, + 281, + 469, + 294 + ], + "score": 1.0, + "content": "two complementary representations of the same computer program. Tradition-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 293, + 469, + 305 + ], + "spans": [ + { + "bbox": [ + 141, + 293, + 469, + 305 + ], + "score": 1.0, + "content": "ally, designers of machine learning models have relied predominantly either on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 469, + 316 + ], + "score": 1.0, + "content": "Structure or Context. We propose a new model, which jointly learns on Context", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 315, + 470, + 327 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 470, + 327 + ], + "score": 1.0, + "content": "and Structure of source code. In contrast to previous approaches, our model uses", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 470, + 339 + ], + "score": 1.0, + "content": "only language-agnostic features, i.e., source code and features that can be com-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 337, + 470, + 349 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 470, + 349 + ], + "score": 1.0, + "content": "puted directly from the AST. Besides obtaining state-of-the-art on monolingual", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 348, + 470, + 360 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 470, + 360 + ], + "score": 1.0, + "content": "code summarization on all five programming languages considered in this work,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 359, + 469, + 371 + ], + "spans": [ + { + "bbox": [ + 141, + 359, + 469, + 371 + ], + "score": 1.0, + "content": "we propose the first multilingual code summarization model. We show that jointly", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 370, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 141, + 370, + 469, + 382 + ], + "score": 1.0, + "content": "training on non-parallel data from multiple programming languages improves re-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 380, + 470, + 393 + ], + "spans": [ + { + "bbox": [ + 141, + 380, + 470, + 393 + ], + "score": 1.0, + "content": "sults on all individual languages, where the strongest gains are on low-resource", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 392, + 470, + 404 + ], + "spans": [ + { + "bbox": [ + 141, + 392, + 470, + 404 + ], + "score": 1.0, + "content": "languages. Remarkably, multilingual training only from Context does not lead", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 402, + 470, + 415 + ], + "spans": [ + { + "bbox": [ + 141, + 402, + 470, + 415 + ], + "score": 1.0, + "content": "to the same improvements, highlighting the benefits of combining Structure and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 414, + 321, + 426 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 321, + 426 + ], + "score": 1.0, + "content": "Context for representation learning on code.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 21.5, + "bbox_fs": [ + 141, + 271, + 470, + 426 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 443, + 206, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 441, + 208, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 208, + 459 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "Machine learning for code is an active and growing area of research which aims at building models", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "that can learn semantically meaningful representations of programs. These embeddings can be used", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "on downstream tasks, such as code generation, bug detection, or code summarization. We focus", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "our work on two complementary data representations of programs: the source code (referred to as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "Context in this work), and the abstract syntax tree (AST; referred to as Structure). Traditionally,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "researchers and practitioners have decided to predominantly leverage either Structure or Context in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "their machine learning models. In this work, we show that jointly learning on Context and Structure", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 544, + 352, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 352, + 557 + ], + "score": 1.0, + "content": "improves representation learning on source code (see Fig. 1).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 468, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "The source code representation naturally lends itself to models from natural language process-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "ing (NLP), e.g., long short-term memory networks (Hochreiter & Schmidhuber, 1997) (LSTM) or", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "Transformers (Vaswani et al., 2017; Radford et al., 2019; Dai et al., 2019; Yang et al., 2019; Shaw", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 607 + ], + "score": 1.0, + "content": "et al., 2018). On the other hand, models leveraging the structure representations are typically based", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "on graph neural networks (GNNs) (Kipf & Welling, 2017; Xu et al., 2019; Velickovi ˇ c et al., 2018; ´", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "You et al., 2019; Hamilton et al., 2017; Li et al., 2015; Klicpera et al., 2020). While the AST repre-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "sentation makes the highly structured nature of source code explicit to the models, since most GNNs", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "use the message-passing framework, their learned representations are inherently local and struggle", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 649, + 250, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 250, + 662 + ], + "score": 1.0, + "content": "to leverage long-range interactions.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42, + "bbox_fs": [ + 104, + 561, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "Recently, Hellendoorn et al. (2020) have explored models that can leverage several representations,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "including both Structure and Context. Their Graph Relational Embedding Attention Transformer", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "(GREAT) extends Shaw et al. (2018), which biases the self-attention computation in a localized", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "way given the underlying graph. The language-specific representations used by GREAT include a", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "combination of the data flow graph, control flow graph, syntactic edges (inspired by Allamanis et al.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 459, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 459, + 732 + ], + "score": 1.0, + "content": "(2018)), etc. which require specialized pipelines and static analysis tools to be obtained.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 665, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 80, + 446, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 166, + 80, + 446, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 166, + 80, + 446, + 186 + ], + "spans": [ + { + "bbox": [ + 166, + 80, + 446, + 186 + ], + "score": 0.923, + "type": "image", + "image_path": "7701ebd563c4ccbab0fb3932736f0a3f5ce8408d760ee36d22e78173cbc12d18.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 166, + 80, + 446, + 115.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 166, + 115.33333333333334, + 446, + 150.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 166, + 150.66666666666669, + 446, + 186.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 195, + 506, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "score": 1.0, + "content": "Figure 1: Context and Structure both encapsulate valuable information about source code. In this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "realistic example, token 1 and 4 are distant in the sequence of tokens (Context), but only 5 hops", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "away when traversing the Abstract Syntax Tree (Structure). As such, a method that relies only", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "on the sequence of tokens could neglect the relationship between a method name and its return", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "variable. Conversely, token 1 and 2 showcase the opposite setting. Hence, unifying Structure and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 251, + 363, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 363, + 263 + ], + "score": 1.0, + "content": "Context leads to a more powerful representation of source code.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "score": 1.0, + "content": "We propose the CODE TRANSFORMER1, which combines distances computed on Structure and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "Context in the self-attention operation. In contrast to the localized treatment via edges described", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "above, we make the full Structure accessible to the model at each layer by computing pairwise", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "distances on the AST, such as shortest path lengths. To this end, we draw inspiration from the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "XLNet architecture (Yang et al., 2019), which uses relative distances instead of absolute positions", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "in the attention computation. Importantly, all our features are language-agnostic2, i.e., can easily be", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 350, + 417, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 417, + 362 + ], + "score": 1.0, + "content": "computed for any programming language based on the source code and AST.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "We use two datasets comprising 5 different programming languages in total, and evaluate the rep-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "score": 1.0, + "content": "resentations learned by our model on the task of code summarization, where the model predicts a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "method’s name based on its body. Besides setting the state-of-the-art on all five languages for single-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "language training, we also train the first multilingual model for code summarization. This is enabled", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 410, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 504, + 424 + ], + "score": 1.0, + "content": "by the fact that our model uses only language-agnostic features that can easily be obtained for any", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "programming language. Remarkably, training our model on multiple programming languages sub-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 433, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 504, + 444 + ], + "score": 1.0, + "content": "stantially improves the performance on all languages. Moreover, multilingual training only from", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "Context does not lead to the same improvements, highlighting the benefits of combining Structure", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 455, + 301, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 301, + 468 + ], + "score": 1.0, + "content": "and Context for representation learning on code.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 482, + 208, + 495 + ], + "lines": [ + { + "bbox": [ + 104, + 481, + 211, + 497 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 211, + 497 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "Machine Learning for Code. Early research learned language models on raw text data, e.g., (Wang", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 521, + 504, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 504, + 535 + ], + "score": 1.0, + "content": "et al., 2016; Raychev et al., 2014; Dam et al., 2016), providing evidence for the naturalness assump-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "tion (Hindle et al., 2012). For example, Allamanis et al. (2015) learned distributed representations", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "of variables and methods, finding that they were indeed able to encode common semantic properties", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "from the regularities present in source code. Alon et al. (2019b) also found evidence of semantic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "arithmetic in their embedding space, dubbed code2vec. These representations—and their variants", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "like (Mou et al., 2016)—can then be used to predict sequences of identifier sub-tokens (Allamanis", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "et al., 2015) or API calls (Acharya et al., 2007; Nguyen et al., 2017). They can be used as advanced", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "auto-completion tools (Hindle et al., 2012; Bhoopchand et al., 2016), including for user-provided", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "tokens like Variable Names (Raychev et al., 2014; Allamanis et al., 2014). These are useful for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 621, + 389, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 389, + 633 + ], + "score": 1.0, + "content": "deobfuscating Android applications (Bichsel et al., 2016) for example.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 504, + 649 + ], + "score": 1.0, + "content": "Several works leverage structured graphical models for probabilistic models of source code, usually", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "through parse trees (Maddison & Tarlow, 2014; Bielik et al., 2016). Unlike previous works where", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "hand-crafted features were used as node features (Raychev et al., 2014) or as explicit semantic edges", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "(Allamanis et al., 2018), our work does not augment the existing syntactic relationships between the", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 690, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 689, + 488, + 703 + ], + "spans": [ + { + "bbox": [ + 119, + 689, + 488, + 703 + ], + "score": 1.0, + "content": "1Code at www.daml.in.tum.de/code-transformer, demo at code-transformer.org.", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 698, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 117, + 698, + 506, + 716 + ], + "score": 1.0, + "content": "2We use the term language-agnostic to highlight that our model does not rely on language-specific features", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "(e.g., program analysis edges), thus facilitating multi-language training, as it is possible to generate unified", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 720, + 320, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 320, + 734 + ], + "score": 1.0, + "content": "AST representations for different programming languages.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 166, + 80, + 446, + 186 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 166, + 80, + 446, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 166, + 80, + 446, + 186 + ], + "spans": [ + { + "bbox": [ + 166, + 80, + 446, + 186 + ], + "score": 0.923, + "type": "image", + "image_path": "7701ebd563c4ccbab0fb3932736f0a3f5ce8408d760ee36d22e78173cbc12d18.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 166, + 80, + 446, + 115.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 166, + 115.33333333333334, + 446, + 150.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 166, + 150.66666666666669, + 446, + 186.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 195, + 506, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 208 + ], + "score": 1.0, + "content": "Figure 1: Context and Structure both encapsulate valuable information about source code. In this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "realistic example, token 1 and 4 are distant in the sequence of tokens (Context), but only 5 hops", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "away when traversing the Abstract Syntax Tree (Structure). As such, a method that relies only", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 229, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 505, + 242 + ], + "score": 1.0, + "content": "on the sequence of tokens could neglect the relationship between a method name and its return", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "variable. Conversely, token 1 and 2 showcase the opposite setting. Hence, unifying Structure and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 251, + 363, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 363, + 263 + ], + "score": 1.0, + "content": "Context leads to a more powerful representation of source code.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 506, + 297 + ], + "score": 1.0, + "content": "We propose the CODE TRANSFORMER1, which combines distances computed on Structure and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 307 + ], + "score": 1.0, + "content": "Context in the self-attention operation. In contrast to the localized treatment via edges described", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "above, we make the full Structure accessible to the model at each layer by computing pairwise", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "distances on the AST, such as shortest path lengths. To this end, we draw inspiration from the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "XLNet architecture (Yang et al., 2019), which uses relative distances instead of absolute positions", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 352 + ], + "score": 1.0, + "content": "in the attention computation. Importantly, all our features are language-agnostic2, i.e., can easily be", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 350, + 417, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 417, + 362 + ], + "score": 1.0, + "content": "computed for any programming language based on the source code and AST.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 283, + 506, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "We use two datasets comprising 5 different programming languages in total, and evaluate the rep-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 506, + 390 + ], + "score": 1.0, + "content": "resentations learned by our model on the task of code summarization, where the model predicts a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "method’s name based on its body. Besides setting the state-of-the-art on all five languages for single-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "language training, we also train the first multilingual model for code summarization. This is enabled", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 410, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 504, + 424 + ], + "score": 1.0, + "content": "by the fact that our model uses only language-agnostic features that can easily be obtained for any", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 435 + ], + "score": 1.0, + "content": "programming language. Remarkably, training our model on multiple programming languages sub-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 433, + 504, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 504, + 444 + ], + "score": 1.0, + "content": "stantially improves the performance on all languages. Moreover, multilingual training only from", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "Context does not lead to the same improvements, highlighting the benefits of combining Structure", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 455, + 301, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 301, + 468 + ], + "score": 1.0, + "content": "and Context for representation learning on code.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 366, + 506, + 468 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 482, + 208, + 495 + ], + "lines": [ + { + "bbox": [ + 104, + 481, + 211, + 497 + ], + "spans": [ + { + "bbox": [ + 104, + 481, + 211, + 497 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 510, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "Machine Learning for Code. Early research learned language models on raw text data, e.g., (Wang", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 521, + 504, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 504, + 535 + ], + "score": 1.0, + "content": "et al., 2016; Raychev et al., 2014; Dam et al., 2016), providing evidence for the naturalness assump-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "tion (Hindle et al., 2012). For example, Allamanis et al. (2015) learned distributed representations", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "of variables and methods, finding that they were indeed able to encode common semantic properties", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "from the regularities present in source code. Alon et al. (2019b) also found evidence of semantic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "arithmetic in their embedding space, dubbed code2vec. These representations—and their variants", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "like (Mou et al., 2016)—can then be used to predict sequences of identifier sub-tokens (Allamanis", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "et al., 2015) or API calls (Acharya et al., 2007; Nguyen et al., 2017). They can be used as advanced", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "auto-completion tools (Hindle et al., 2012; Bhoopchand et al., 2016), including for user-provided", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "tokens like Variable Names (Raychev et al., 2014; Allamanis et al., 2014). These are useful for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 621, + 389, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 389, + 633 + ], + "score": 1.0, + "content": "deobfuscating Android applications (Bichsel et al., 2016) for example.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 509, + 506, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 637, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 504, + 649 + ], + "score": 1.0, + "content": "Several works leverage structured graphical models for probabilistic models of source code, usually", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 660 + ], + "score": 1.0, + "content": "through parse trees (Maddison & Tarlow, 2014; Bielik et al., 2016). Unlike previous works where", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "hand-crafted features were used as node features (Raychev et al., 2014) or as explicit semantic edges", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "(Allamanis et al., 2018), our work does not augment the existing syntactic relationships between the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "different elements to enhance the predictive capabilities of the model. Other approaches (Alon et al.,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 498, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 498, + 106 + ], + "score": 1.0, + "content": "2018; Li et al., 2017) also leverage the AST structure, but linearize the graph by first traversing it.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 637, + 505, + 682 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "different elements to enhance the predictive capabilities of the model. Other approaches (Alon et al.,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 498, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 498, + 106 + ], + "score": 1.0, + "content": "2018; Li et al., 2017) also leverage the AST structure, but linearize the graph by first traversing it.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 113, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "score": 1.0, + "content": "Learning representations of structured languages. While models of language have dramatically", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 1.0, + "content": "improved in their ability to learn structure (syntax) and semantics from scratch, it can be argued that", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 148 + ], + "score": 1.0, + "content": "directly providing the model with the underlying structure of the language can help with generaliza-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 160 + ], + "score": 1.0, + "content": "tion (Battaglia et al., 2018), managing long-ranging dependencies (Tai et al., 2015), or representing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 158, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 505, + 170 + ], + "score": 1.0, + "content": "the compositional aspect of natural language (Socher et al., 2013). Notably, tree structures have", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "score": 1.0, + "content": "shown promising results and inspired new architectures (Shen et al., 2019), including in the domain", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "of source code (Fernandes et al., 2019), where the underlying syntax is directly available. Our work", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "pursues this line of research, showing the benefits of explicitly integrating structural information as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 200, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 506, + 214 + ], + "score": 1.0, + "content": "an inductive bias. Shiv & Quirk (2019) propose positional encodings for nodes on trees; however,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "score": 1.0, + "content": "their approach assumes regular trees, which is an unrealistic assumption when working with Ab-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "stract Syntax Trees, as an AST node can have arbitrarily many children, e.g., the arguments of a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 233, + 145, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 145, + 246 + ], + "score": 1.0, + "content": "function.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "Graph Neural Networks. GNNs provide a powerful tool for machine learning on graphs, thanks to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "their ability to recursively incorporate information from neighboring nodes in the network (Battaglia", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "et al., 2018), naturally capturing the graph structure simultaneously with the nodes’ features. (Gori", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "et al., 2005; Scarselli et al., 2008) are able to learn vector representations of nodes and graphs in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "an end-to-end fashion, encoding structural and feature information in the embedding space. Under", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "this model, GNNs have achieved state-of-the-art performance across a variety of tasks, such as node", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 331 + ], + "score": 1.0, + "content": "classification (Kipf & Welling, 2017; Hamilton et al., 2017; Klicpera et al., 2019a), link prediction", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 329, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 344 + ], + "score": 1.0, + "content": "(Zhang & Chen, 2018; Schlichtkrull et al., 2018), graph clustering (Defferrard et al., 2016; Ying", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 341, + 484, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 484, + 354 + ], + "score": 1.0, + "content": "et al., 2018) or graph classification (Ying et al., 2018; Dai et al., 2016; Duvenaud et al., 2015).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 104, + 368, + 493, + 382 + ], + "lines": [ + { + "bbox": [ + 104, + 367, + 495, + 383 + ], + "spans": [ + { + "bbox": [ + 104, + 367, + 495, + 383 + ], + "score": 1.0, + "content": "3 INTEGRATING STRUCTURE AND CONTEXT IN THE CODE TRANSFORMER", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 504, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "score": 1.0, + "content": "Self-attention is the core operation powering the Transformer. It enables the model to selectively", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 405, + 491, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 491, + 416 + ], + "score": 1.0, + "content": "focus on relevant parts of the input. The matrix form equation for attention with a single head is", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 419, + 400, + 447 + ], + "lines": [ + { + "bbox": [ + 209, + 419, + 400, + 447 + ], + "spans": [ + { + "bbox": [ + 209, + 419, + 400, + 447 + ], + "score": 0.94, + "content": "{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { k } } } } \\right) V ,", + "type": "interline_equation", + "image_path": "bf5062f76acf55f459a4ae89adb1abed3d6a40f640c4bd316cd5ed6cd5cde38d.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 419, + 400, + 433.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 209, + 433.0, + 400, + 447.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 134, + 464 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 450, + 201, + 462 + ], + "score": 0.91, + "content": "Q , K \\in \\mathbb { R } ^ { N \\times d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 448, + 219, + 464 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 450, + 272, + 461 + ], + "score": 0.9, + "content": "V \\in \\mathbb { R } ^ { N \\times d _ { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 448, + 278, + 464 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 278, + 451, + 289, + 461 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 448, + 413, + 464 + ], + "score": 1.0, + "content": "is the number of input tokens,", + "type": "text" + }, + { + "bbox": [ + 414, + 451, + 425, + 462 + ], + "score": 0.89, + "content": "d _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "the key dimension,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 123, + 475 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 462, + 135, + 473 + ], + "score": 0.88, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 461, + 296, + 475 + ], + "score": 1.0, + "content": "the value dimension (typically we have", + "type": "text" + }, + { + "bbox": [ + 296, + 462, + 332, + 473 + ], + "score": 0.92, + "content": "d _ { k } = d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 461, + 456, + 475 + ], + "score": 1.0, + "content": "). The attention score of query", + "type": "text" + }, + { + "bbox": [ + 457, + 462, + 470, + 474 + ], + "score": 0.89, + "content": "Q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "and key", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 473, + 195, + 486 + ], + "spans": [ + { + "bbox": [ + 107, + 473, + 121, + 486 + ], + "score": 0.89, + "content": "K _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 473, + 195, + 485 + ], + "score": 1.0, + "content": "before softmax is", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 483, + 374, + 500 + ], + "lines": [ + { + "bbox": [ + 236, + 483, + 374, + 500 + ], + "spans": [ + { + "bbox": [ + 236, + 483, + 374, + 500 + ], + "score": 0.93, + "content": "\\pmb { A } _ { i j } = \\pmb { Q } _ { i } ^ { T } \\pmb { K } _ { j } = \\pmb { E } _ { i } ^ { T } \\pmb { W } _ { q } ^ { T } \\pmb { W } _ { k } \\pmb { E } _ { j } ,", + "type": "interline_equation", + "image_path": "a084e175bb3c6930841d0180a4202d6eff0f84ecb6a8a9cc250c5cd2833a474e.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 236, + 483, + 374, + 500 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 504, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 507, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 133, + 516 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 501, + 186, + 514 + ], + "score": 0.93, + "content": "E _ { i } , E _ { j } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 498, + 216, + 516 + ], + "score": 1.0, + "content": "are the", + "type": "text" + }, + { + "bbox": [ + 217, + 502, + 223, + 512 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 498, + 365, + 516 + ], + "score": 1.0, + "content": "-dimensional embeddings of tokens", + "type": "text" + }, + { + "bbox": [ + 365, + 503, + 370, + 512 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 498, + 387, + 516 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 388, + 502, + 393, + 514 + ], + "score": 0.78, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 498, + 414, + 516 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 414, + 502, + 430, + 514 + ], + "score": 0.63, + "content": "W _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 498, + 433, + 516 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 433, + 501, + 489, + 514 + ], + "score": 0.61, + "content": "W _ { k } \\in \\mathbb { R } ^ { d _ { k } \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 498, + 507, + 516 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 513, + 313, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 313, + 525 + ], + "score": 1.0, + "content": "the query and key projection matrices, respectively.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 529, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Observe that Eq. (2) contains no assumption about potential structure in the input domain: in the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 539, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 554 + ], + "score": 1.0, + "content": "attention operation we compute all dot products of query and key vectors equally, effectively viewing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "them as unordered sets of vectors. This means, however, that the model is oblivious to structured", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "inputs (such as text or graphs) and therefore is unable to distinguish, for example, a variable name", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 574, + 371, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 371, + 586 + ], + "score": 1.0, + "content": "occurring as an argument and in the return statement of a method.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "score": 1.0, + "content": "In NLP, it is common to bias Transformers towards sequential inputs by adding positional encodings", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "to the token embeddings. These positional encodings are obtained by applying an encoding function", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 156, + 624 + ], + "score": 0.92, + "content": "\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 612, + 253, + 625 + ], + "score": 1.0, + "content": "to each token’s position", + "type": "text" + }, + { + "bbox": [ + 254, + 614, + 263, + 624 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 612, + 505, + 625 + ], + "score": 1.0, + "content": ". These positional encodings make the information about the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 623, + 349, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 349, + 636 + ], + "score": 1.0, + "content": "sequence of tokens available to the model. Eq. (2) becomes:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 637, + 398, + 654 + ], + "lines": [ + { + "bbox": [ + 211, + 637, + 398, + 654 + ], + "spans": [ + { + "bbox": [ + 211, + 637, + 398, + 654 + ], + "score": 0.92, + "content": "A _ { i j } = ( E _ { i } + \\phi ( p _ { i } ) ) ^ { T } W _ { q } ^ { T } W _ { k } ( E _ { j } + \\phi ( p _ { j } ) ) , ,", + "type": "interline_equation", + "image_path": "b96beefee59c632a287223572ec7702368e163a81b493e1a921e5c4d74a409f9.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 211, + 637, + 398, + 654 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 656, + 192, + 667 + ], + "lines": [ + { + "bbox": [ + 107, + 656, + 192, + 667 + ], + "spans": [ + { + "bbox": [ + 107, + 656, + 192, + 667 + ], + "score": 1.0, + "content": "which factorizes into", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "interline_equation", + "bbox": [ + 118, + 669, + 482, + 704 + ], + "lines": [ + { + "bbox": [ + 118, + 669, + 482, + 704 + ], + "spans": [ + { + "bbox": [ + 118, + 669, + 482, + 704 + ], + "score": 0.94, + "content": "A _ { i j } = \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } } _ { ( \\mathrm { u } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { b } ) { A } _ { i j } ^ { \\mathrm { c p } } } } + \\underbrace { { \\phi } ( p _ { i } ) ^ { T } W _ { q } ^ { T } W _ { k } { E } _ { j } } _ { ( \\mathrm { c } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { { \\phi } ( p _ { i } ) } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { d } ) { A } _ { i j } ^ { \\mathrm { p } } } } .", + "type": "interline_equation", + "image_path": "df07c855af71840c08fb5fcac8d0a6450e30cf45761ee10da8da265d06d28ade.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 118, + 669, + 482, + 680.6666666666666 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 118, + 680.6666666666666, + 482, + 692.3333333333333 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 118, + 692.3333333333333, + 482, + 703.9999999999999 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 707, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 107, + 707, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 107, + 707, + 304, + 721 + ], + "score": 1.0, + "content": "We can interpret the terms (a)-(d) as follows. (a)", + "type": "text" + }, + { + "bbox": [ + 305, + 708, + 321, + 721 + ], + "score": 0.91, + "content": "{ \\cal { A } } _ { i j } ^ { \\mathrm { c c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 707, + 505, + 721 + ], + "score": 1.0, + "content": "is the contribution from the ‘match’ between", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 240, + 734 + ], + "score": 1.0, + "content": "the content embeddings of tokens", + "type": "text" + }, + { + "bbox": [ + 241, + 721, + 245, + 730 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 719, + 262, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 721, + 268, + 732 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 719, + 286, + 734 + ], + "score": 1.0, + "content": "; (b)", + "type": "text" + }, + { + "bbox": [ + 286, + 720, + 303, + 734 + ], + "score": 0.9, + "content": "A _ { i j } ^ { \\mathrm { c p } }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "steers the attention towards certain positions based", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 503, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 504, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 113, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "score": 1.0, + "content": "Learning representations of structured languages. While models of language have dramatically", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 506, + 137 + ], + "score": 1.0, + "content": "improved in their ability to learn structure (syntax) and semantics from scratch, it can be argued that", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 135, + 505, + 148 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 505, + 148 + ], + "score": 1.0, + "content": "directly providing the model with the underlying structure of the language can help with generaliza-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 160 + ], + "score": 1.0, + "content": "tion (Battaglia et al., 2018), managing long-ranging dependencies (Tai et al., 2015), or representing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 158, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 158, + 505, + 170 + ], + "score": 1.0, + "content": "the compositional aspect of natural language (Socher et al., 2013). Notably, tree structures have", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 181 + ], + "score": 1.0, + "content": "shown promising results and inspired new architectures (Shen et al., 2019), including in the domain", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "of source code (Fernandes et al., 2019), where the underlying syntax is directly available. Our work", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 506, + 204 + ], + "score": 1.0, + "content": "pursues this line of research, showing the benefits of explicitly integrating structural information as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 200, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 104, + 200, + 506, + 214 + ], + "score": 1.0, + "content": "an inductive bias. Shiv & Quirk (2019) propose positional encodings for nodes on trees; however,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 505, + 225 + ], + "score": 1.0, + "content": "their approach assumes regular trees, which is an unrealistic assumption when working with Ab-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "stract Syntax Trees, as an AST node can have arbitrarily many children, e.g., the arguments of a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 233, + 145, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 145, + 246 + ], + "score": 1.0, + "content": "function.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 7.5, + "bbox_fs": [ + 104, + 113, + 506, + 246 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 353 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "Graph Neural Networks. GNNs provide a powerful tool for machine learning on graphs, thanks to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "their ability to recursively incorporate information from neighboring nodes in the network (Battaglia", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "et al., 2018), naturally capturing the graph structure simultaneously with the nodes’ features. (Gori", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "et al., 2005; Scarselli et al., 2008) are able to learn vector representations of nodes and graphs in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "an end-to-end fashion, encoding structural and feature information in the embedding space. Under", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 321 + ], + "score": 1.0, + "content": "this model, GNNs have achieved state-of-the-art performance across a variety of tasks, such as node", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 331 + ], + "score": 1.0, + "content": "classification (Kipf & Welling, 2017; Hamilton et al., 2017; Klicpera et al., 2019a), link prediction", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 329, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 344 + ], + "score": 1.0, + "content": "(Zhang & Chen, 2018; Schlichtkrull et al., 2018), graph clustering (Defferrard et al., 2016; Ying", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 341, + 484, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 484, + 354 + ], + "score": 1.0, + "content": "et al., 2018) or graph classification (Ying et al., 2018; Dai et al., 2016; Duvenaud et al., 2015).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 254, + 506, + 354 + ] + }, + { + "type": "title", + "bbox": [ + 104, + 368, + 493, + 382 + ], + "lines": [ + { + "bbox": [ + 104, + 367, + 495, + 383 + ], + "spans": [ + { + "bbox": [ + 104, + 367, + 495, + 383 + ], + "score": 1.0, + "content": "3 INTEGRATING STRUCTURE AND CONTEXT IN THE CODE TRANSFORMER", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 504, + 416 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "score": 1.0, + "content": "Self-attention is the core operation powering the Transformer. It enables the model to selectively", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 405, + 491, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 491, + 416 + ], + "score": 1.0, + "content": "focus on relevant parts of the input. The matrix form equation for attention with a single head is", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 392, + 505, + 416 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 209, + 419, + 400, + 447 + ], + "lines": [ + { + "bbox": [ + 209, + 419, + 400, + 447 + ], + "spans": [ + { + "bbox": [ + 209, + 419, + 400, + 447 + ], + "score": 0.94, + "content": "{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { k } } } } \\right) V ,", + "type": "interline_equation", + "image_path": "bf5062f76acf55f459a4ae89adb1abed3d6a40f640c4bd316cd5ed6cd5cde38d.jpg" + } + ] + } + ], + "index": 26.5, + "virtual_lines": [ + { + "bbox": [ + 209, + 419, + 400, + 433.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 209, + 433.0, + 400, + 447.0 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 134, + 464 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 450, + 201, + 462 + ], + "score": 0.91, + "content": "Q , K \\in \\mathbb { R } ^ { N \\times d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 448, + 219, + 464 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 220, + 450, + 272, + 461 + ], + "score": 0.9, + "content": "V \\in \\mathbb { R } ^ { N \\times d _ { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 448, + 278, + 464 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 278, + 451, + 289, + 461 + ], + "score": 0.76, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 448, + 413, + 464 + ], + "score": 1.0, + "content": "is the number of input tokens,", + "type": "text" + }, + { + "bbox": [ + 414, + 451, + 425, + 462 + ], + "score": 0.89, + "content": "d _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "the key dimension,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 123, + 475 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 462, + 135, + 473 + ], + "score": 0.88, + "content": "d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 461, + 296, + 475 + ], + "score": 1.0, + "content": "the value dimension (typically we have", + "type": "text" + }, + { + "bbox": [ + 296, + 462, + 332, + 473 + ], + "score": 0.92, + "content": "d _ { k } = d _ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 461, + 456, + 475 + ], + "score": 1.0, + "content": "). The attention score of query", + "type": "text" + }, + { + "bbox": [ + 457, + 462, + 470, + 474 + ], + "score": 0.89, + "content": "Q _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "and key", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 473, + 195, + 486 + ], + "spans": [ + { + "bbox": [ + 107, + 473, + 121, + 486 + ], + "score": 0.89, + "content": "K _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 473, + 195, + 485 + ], + "score": 1.0, + "content": "before softmax is", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 448, + 505, + 486 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 483, + 374, + 500 + ], + "lines": [ + { + "bbox": [ + 236, + 483, + 374, + 500 + ], + "spans": [ + { + "bbox": [ + 236, + 483, + 374, + 500 + ], + "score": 0.93, + "content": "\\pmb { A } _ { i j } = \\pmb { Q } _ { i } ^ { T } \\pmb { K } _ { j } = \\pmb { E } _ { i } ^ { T } \\pmb { W } _ { q } ^ { T } \\pmb { W } _ { k } \\pmb { E } _ { j } ,", + "type": "interline_equation", + "image_path": "a084e175bb3c6930841d0180a4202d6eff0f84ecb6a8a9cc250c5cd2833a474e.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 236, + 483, + 374, + 500 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 504, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 507, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 133, + 516 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 501, + 186, + 514 + ], + "score": 0.93, + "content": "E _ { i } , E _ { j } \\in \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 498, + 216, + 516 + ], + "score": 1.0, + "content": "are the", + "type": "text" + }, + { + "bbox": [ + 217, + 502, + 223, + 512 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 498, + 365, + 516 + ], + "score": 1.0, + "content": "-dimensional embeddings of tokens", + "type": "text" + }, + { + "bbox": [ + 365, + 503, + 370, + 512 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 498, + 387, + 516 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 388, + 502, + 393, + 514 + ], + "score": 0.78, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 498, + 414, + 516 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 414, + 502, + 430, + 514 + ], + "score": 0.63, + "content": "W _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 498, + 433, + 516 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 433, + 501, + 489, + 514 + ], + "score": 0.61, + "content": "W _ { k } \\in \\mathbb { R } ^ { d _ { k } \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 498, + 507, + 516 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 513, + 313, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 313, + 525 + ], + "score": 1.0, + "content": "the query and key projection matrices, respectively.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 498, + 507, + 525 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 529, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Observe that Eq. (2) contains no assumption about potential structure in the input domain: in the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 539, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 554 + ], + "score": 1.0, + "content": "attention operation we compute all dot products of query and key vectors equally, effectively viewing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 564 + ], + "score": 1.0, + "content": "them as unordered sets of vectors. This means, however, that the model is oblivious to structured", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 576 + ], + "score": 1.0, + "content": "inputs (such as text or graphs) and therefore is unable to distinguish, for example, a variable name", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 574, + 371, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 371, + 586 + ], + "score": 1.0, + "content": "occurring as an argument and in the return statement of a method.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 530, + 506, + 586 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 590, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "score": 1.0, + "content": "In NLP, it is common to bias Transformers towards sequential inputs by adding positional encodings", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "to the token embeddings. These positional encodings are obtained by applying an encoding function", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 107, + 612, + 156, + 624 + ], + "score": 0.92, + "content": "\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 612, + 253, + 625 + ], + "score": 1.0, + "content": "to each token’s position", + "type": "text" + }, + { + "bbox": [ + 254, + 614, + 263, + 624 + ], + "score": 0.84, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 612, + 505, + 625 + ], + "score": 1.0, + "content": ". These positional encodings make the information about the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 623, + 349, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 349, + 636 + ], + "score": 1.0, + "content": "sequence of tokens available to the model. Eq. (2) becomes:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 589, + 505, + 636 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 211, + 637, + 398, + 654 + ], + "lines": [ + { + "bbox": [ + 211, + 637, + 398, + 654 + ], + "spans": [ + { + "bbox": [ + 211, + 637, + 398, + 654 + ], + "score": 0.92, + "content": "A _ { i j } = ( E _ { i } + \\phi ( p _ { i } ) ) ^ { T } W _ { q } ^ { T } W _ { k } ( E _ { j } + \\phi ( p _ { j } ) ) , ,", + "type": "interline_equation", + "image_path": "b96beefee59c632a287223572ec7702368e163a81b493e1a921e5c4d74a409f9.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 211, + 637, + 398, + 654 + ], + "spans": [], + "index": 43 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 656, + 192, + 667 + ], + "lines": [ + { + "bbox": [ + 107, + 656, + 192, + 667 + ], + "spans": [ + { + "bbox": [ + 107, + 656, + 192, + 667 + ], + "score": 1.0, + "content": "which factorizes into", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44, + "bbox_fs": [ + 107, + 656, + 192, + 667 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 118, + 669, + 482, + 704 + ], + "lines": [ + { + "bbox": [ + 118, + 669, + 482, + 704 + ], + "spans": [ + { + "bbox": [ + 118, + 669, + 482, + 704 + ], + "score": 0.94, + "content": "A _ { i j } = \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } } _ { ( \\mathrm { u } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { b } ) { A } _ { i j } ^ { \\mathrm { c p } } } } + \\underbrace { { \\phi } ( p _ { i } ) ^ { T } W _ { q } ^ { T } W _ { k } { E } _ { j } } _ { ( \\mathrm { c } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { { \\phi } ( p _ { i } ) } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { d } ) { A } _ { i j } ^ { \\mathrm { p } } } } .", + "type": "interline_equation", + "image_path": "df07c855af71840c08fb5fcac8d0a6450e30cf45761ee10da8da265d06d28ade.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 118, + 669, + 482, + 680.6666666666666 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 118, + 680.6666666666666, + 482, + 692.3333333333333 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 118, + 692.3333333333333, + 482, + 703.9999999999999 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 707, + 505, + 734 + ], + "lines": [ + { + "bbox": [ + 107, + 707, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 107, + 707, + 304, + 721 + ], + "score": 1.0, + "content": "We can interpret the terms (a)-(d) as follows. (a)", + "type": "text" + }, + { + "bbox": [ + 305, + 708, + 321, + 721 + ], + "score": 0.91, + "content": "{ \\cal { A } } _ { i j } ^ { \\mathrm { c c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 707, + 505, + 721 + ], + "score": 1.0, + "content": "is the contribution from the ‘match’ between", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 240, + 734 + ], + "score": 1.0, + "content": "the content embeddings of tokens", + "type": "text" + }, + { + "bbox": [ + 241, + 721, + 245, + 730 + ], + "score": 0.79, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 719, + 262, + 734 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 263, + 721, + 268, + 732 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 719, + 286, + 734 + ], + "score": 1.0, + "content": "; (b)", + "type": "text" + }, + { + "bbox": [ + 286, + 720, + 303, + 734 + ], + "score": 0.9, + "content": "A _ { i j } ^ { \\mathrm { c p } }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 719, + 505, + 734 + ], + "score": 1.0, + "content": "steers the attention towards certain positions based", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5, + "bbox_fs": [ + 106, + 707, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 504, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 200, + 96 + ], + "score": 1.0, + "content": "on the content of token", + "type": "text" + }, + { + "bbox": [ + 200, + 83, + 205, + 93 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 80, + 222, + 96 + ], + "score": 1.0, + "content": "; (c)", + "type": "text" + }, + { + "bbox": [ + 223, + 82, + 239, + 96 + ], + "score": 0.92, + "content": "A _ { i j } ^ { \\mathrm { p c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "biases towards content embeddings based on the position of token", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 405, + 110 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 111, + 106 + ], + "score": 0.67, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 94, + 129, + 109 + ], + "score": 1.0, + "content": "; (d)", + "type": "text" + }, + { + "bbox": [ + 129, + 95, + 147, + 110 + ], + "score": 0.9, + "content": "A _ { i j } ^ { \\mathrm { p p } }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 94, + 405, + 109 + ], + "score": 1.0, + "content": "controls which positions should attend to which other positions.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 104, + 112, + 505, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "In our model, we adopt the formulation of Dai et al. (2019); Yang et al. (2019). They modify Eq. (4)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 123, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 287, + 137 + ], + "score": 1.0, + "content": "by replacing the absolute position encodings", + "type": "text" + }, + { + "bbox": [ + 287, + 124, + 310, + 136 + ], + "score": 0.92, + "content": "\\phi ( p _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 123, + 442, + 137 + ], + "score": 1.0, + "content": "with relative position encodings", + "type": "text" + }, + { + "bbox": [ + 442, + 124, + 476, + 136 + ], + "score": 0.92, + "content": "\\phi ( r _ { i \\to j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 123, + 480, + 137 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 145, + 139, + 465, + 155 + ], + "lines": [ + { + "bbox": [ + 145, + 139, + 465, + 155 + ], + "spans": [ + { + "bbox": [ + 145, + 139, + 465, + 155 + ], + "score": 0.9, + "content": "\\begin{array} { r } { A _ { i j } ^ { \\mathrm { r e l } } = \\ E _ { i } ^ { T } W _ { q } ^ { T } W _ { k } E _ { j } + E _ { i } ^ { T } W _ { q } ^ { T } W _ { r } \\phi ( r _ { i j } ) + u ^ { T } W _ { k } E _ { j } + v ^ { T } W _ { r } \\phi ( r _ { i j } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "4abc95bdea1392a92734f005afce75f982c5226661a1db2266b0c52e16c233d4.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 145, + 139, + 465, + 155 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 133, + 173 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 163, + 154, + 172 + ], + "score": 0.9, + "content": "r _ { i \\to j }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 160, + 291, + 173 + ], + "score": 1.0, + "content": "is the relative distance from token", + "type": "text" + }, + { + "bbox": [ + 291, + 162, + 296, + 170 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 160, + 331, + 173 + ], + "score": 1.0, + "content": "to token", + "type": "text" + }, + { + "bbox": [ + 331, + 161, + 337, + 172 + ], + "score": 0.84, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 160, + 403, + 173 + ], + "score": 1.0, + "content": "in the sequence,", + "type": "text" + }, + { + "bbox": [ + 404, + 159, + 450, + 171 + ], + "score": 0.92, + "content": "\\boldsymbol { u } , \\boldsymbol { v } \\in \\mathbb { R } ^ { d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "are learnable", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 175, + 185 + ], + "score": 1.0, + "content": "bias vectors, and", + "type": "text" + }, + { + "bbox": [ + 176, + 171, + 192, + 182 + ], + "score": 0.9, + "content": "W _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "is a key projection matrix for the relative distances. Besides fixing issues with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "absolute position encodings such as ambiguity when processing two sentences at a time, Eq. (5)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 194, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 205 + ], + "score": 1.0, + "content": "enables native application of the powerful self-attention operation on domains such as graphs, where", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "absolute coordinates are not available. We adopt the (non-trainable) sinusoidal encoding function", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "proposed by (Vaswani et al., 2017) for all relations; see Appendix A.1 for details on the distance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 183, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 183, + 239 + ], + "score": 1.0, + "content": "encoding function.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 108, + 250, + 451, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 249, + 453, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 453, + 263 + ], + "score": 1.0, + "content": "3.1 INTEGRATING SOURCE CODE AND AST REPRESENTATIONS OF PROGRAMS.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "score": 1.0, + "content": "To enable the model to integrate information both the Context and Structure of programs, we modify", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Eq. (5) to be able to incorporate multiple different relations. To this end, we use one key projection", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 102, + 292, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 102, + 292, + 164, + 311 + ], + "score": 1.0, + "content": "matrix W (s)r", + "type": "text" + }, + { + "bbox": [ + 160, + 295, + 213, + 308 + ], + "score": 1.0, + "content": "per relation", + "type": "text" + }, + { + "bbox": [ + 213, + 298, + 219, + 306 + ], + "score": 0.58, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 295, + 505, + 308 + ], + "score": 1.0, + "content": ", and sum their contributions in the raw attention score. This enables", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "the CODE TRANSFORMER to combine information from multiple relations between tokens in the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 318, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 329 + ], + "score": 1.0, + "content": "attention computation. Besides the token distance in the Context, we include pairwise relations", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 328, + 506, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 341 + ], + "score": 1.0, + "content": "based on the AST as described in the following. See Fig. 2 for a visualization of the Structure", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 339, + 178, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 178, + 352 + ], + "score": 1.0, + "content": "distances we use.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 267, + 447 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 268, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 268, + 371 + ], + "score": 1.0, + "content": "Shortest path length. We include the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 370, + 268, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 268, + 382 + ], + "score": 1.0, + "content": "number of hops required to reach node", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 382, + 268, + 393 + ], + "spans": [ + { + "bbox": [ + 107, + 382, + 113, + 393 + ], + "score": 0.78, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 382, + 195, + 392 + ], + "score": 1.0, + "content": "starting from node", + "type": "text" + }, + { + "bbox": [ + 196, + 382, + 201, + 391 + ], + "score": 0.7, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 382, + 268, + 392 + ], + "score": 1.0, + "content": "and vice versa.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 392, + 268, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 268, + 403 + ], + "score": 1.0, + "content": "Here, we treat the AST as an undirected", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 403, + 268, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 268, + 414 + ], + "score": 1.0, + "content": "graph, since otherwise most distances", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 414, + 268, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 268, + 426 + ], + "score": 1.0, + "content": "would be undefined: e.g., all other nodes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 268, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 268, + 437 + ], + "score": 1.0, + "content": "in the AST would be unreachable from", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 435, + 151, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 151, + 448 + ], + "score": 1.0, + "content": "the leaves.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 267, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 268, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 268, + 464 + ], + "score": 1.0, + "content": "Similar to the distance of two tokens on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 464, + 268, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 268, + 475 + ], + "score": 1.0, + "content": "the source code sequence, the shortest-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 475, + 269, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 269, + 486 + ], + "score": 1.0, + "content": "path length is a global distance. This", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 486, + 268, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 268, + 498 + ], + "score": 1.0, + "content": "makes the whole graph structure acces-", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "image", + "bbox": [ + 278, + 372, + 502, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 278, + 372, + 502, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 278, + 372, + 502, + 469 + ], + "spans": [ + { + "bbox": [ + 278, + 372, + 502, + 469 + ], + "score": 0.962, + "type": "image", + "image_path": "971f2174ff96f3b5b7aea4b5cb4f59f9af4bf791dc6a632edbb218691a5d22cf.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 278, + 372, + 502, + 385.85714285714283 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 278, + 385.85714285714283, + 502, + 399.71428571428567 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 278, + 399.71428571428567, + 502, + 413.5714285714285 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 278, + 413.5714285714285, + 502, + 427.42857142857133 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 278, + 427.42857142857133, + 502, + 441.28571428571416 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 278, + 441.28571428571416, + 502, + 455.142857142857 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 278, + 455.142857142857, + 502, + 468.99999999999983 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 293, + 478, + 487, + 489 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 291, + 476, + 488, + 491 + ], + "spans": [ + { + "bbox": [ + 291, + 476, + 488, + 491 + ], + "score": 1.0, + "content": "Figure 2: Structure distances used by our model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + } + ], + "index": 37.0 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "sible to the model at each layer. In contrast, Hellendoorn et al. (2020) add bias terms to the attention", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "computation only for edges (i.e. shortest-path distance of 1), which is a local operation that only", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "exchanges information between immediate neighbors (similar to message passing in GNNs). The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "equivalent localized operation on the source code sequence would be to treat the sequence as a chain", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "graph and only compute attention terms for neighboring tokens, which in turn highlights the benefit", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 552, + 241, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 241, + 564 + ], + "score": 1.0, + "content": "of non-local attention operations.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 106, + 571, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 504, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 504, + 583 + ], + "score": 1.0, + "content": "Ancestor distance. Since we treat the ASTs as undirected for the computation of the shortest-path", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "length, we lose the direction information of the edges. To avoid this, we also include the distance", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 591, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 505, + 607 + ], + "score": 1.0, + "content": "on the ordered set of ancestors and descendants of a node in the AST (red arrow in Fig. 2). Again,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "we include number of (vertical) hops to avoid locality in the attention computation. For example, a", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 614, + 432, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 128, + 628 + ], + "score": 1.0, + "content": "node", + "type": "text" + }, + { + "bbox": [ + 128, + 616, + 167, + 628 + ], + "score": 0.92, + "content": "r _ { i \\to j } = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 614, + 252, + 628 + ], + "score": 1.0, + "content": "for “grand-children”", + "type": "text" + }, + { + "bbox": [ + 253, + 616, + 259, + 627 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 614, + 270, + 628 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 271, + 616, + 275, + 625 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 614, + 295, + 628 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 296, + 616, + 343, + 628 + ], + "score": 0.92, + "content": "r _ { j i } = - 2", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 614, + 432, + 628 + ], + "score": 1.0, + "content": "in the other direction.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "Sibling distance. The neighbor sets in graphs are typically considered to be unordered, but in an", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "score": 1.0, + "content": "AST, the order of children encodes their order of occurrence in the source code. To avoid information", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 473, + 669 + ], + "score": 1.0, + "content": "loss when encoding the AST, we further include the distance on the ordered set of siblings", + "type": "text" + }, + { + "bbox": [ + 474, + 657, + 493, + 669 + ], + "score": 0.91, + "content": "\\{ v _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 103, + 666, + 507, + 683 + ], + "spans": [ + { + "bbox": [ + 103, + 666, + 432, + 683 + ], + "score": 1.0, + "content": "a node, where we again avoid locality by encoding the number of hops, i.e.", + "type": "text" + }, + { + "bbox": [ + 432, + 669, + 486, + 680 + ], + "score": 0.91, + "content": "r _ { v _ { 1 } v _ { 3 } } ~ = ~ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 666, + 507, + 683 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 679, + 167, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 679, + 167, + 693 + ], + "score": 1.0, + "content": "rv3→v1 = −2.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Personalized PageRank (Page et al., 1999) (PPR). PPR is a well-studied proximity measure which", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "has been shown to be very effective in learning with graphs (Klicpera et al., 2019a;b; Bojchevski", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 422, + 734 + ], + "score": 1.0, + "content": "et al., 2020). PPR captures the local graph structure around a pair of nodes", + "type": "text" + }, + { + "bbox": [ + 422, + 721, + 443, + 732 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 720, + 482, + 734 + ], + "score": 1.0, + "content": ". E.g., if", + "type": "text" + }, + { + "bbox": [ + 482, + 721, + 487, + 730 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "has", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 57 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 81, + 504, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 200, + 96 + ], + "score": 1.0, + "content": "on the content of token", + "type": "text" + }, + { + "bbox": [ + 200, + 83, + 205, + 93 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 80, + 222, + 96 + ], + "score": 1.0, + "content": "; (c)", + "type": "text" + }, + { + "bbox": [ + 223, + 82, + 239, + 96 + ], + "score": 0.92, + "content": "A _ { i j } ^ { \\mathrm { p c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 80, + 506, + 96 + ], + "score": 1.0, + "content": "biases towards content embeddings based on the position of token", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 405, + 110 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 111, + 106 + ], + "score": 0.67, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 94, + 129, + 109 + ], + "score": 1.0, + "content": "; (d)", + "type": "text" + }, + { + "bbox": [ + 129, + 95, + 147, + 110 + ], + "score": 0.9, + "content": "A _ { i j } ^ { \\mathrm { p p } }", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 94, + 405, + 109 + ], + "score": 1.0, + "content": "controls which positions should attend to which other positions.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 80, + 506, + 110 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 112, + 505, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "In our model, we adopt the formulation of Dai et al. (2019); Yang et al. (2019). They modify Eq. (4)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 123, + 480, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 287, + 137 + ], + "score": 1.0, + "content": "by replacing the absolute position encodings", + "type": "text" + }, + { + "bbox": [ + 287, + 124, + 310, + 136 + ], + "score": 0.92, + "content": "\\phi ( p _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 123, + 442, + 137 + ], + "score": 1.0, + "content": "with relative position encodings", + "type": "text" + }, + { + "bbox": [ + 442, + 124, + 476, + 136 + ], + "score": 0.92, + "content": "\\phi ( r _ { i \\to j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 123, + 480, + 137 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 505, + 137 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 145, + 139, + 465, + 155 + ], + "lines": [ + { + "bbox": [ + 145, + 139, + 465, + 155 + ], + "spans": [ + { + "bbox": [ + 145, + 139, + 465, + 155 + ], + "score": 0.9, + "content": "\\begin{array} { r } { A _ { i j } ^ { \\mathrm { r e l } } = \\ E _ { i } ^ { T } W _ { q } ^ { T } W _ { k } E _ { j } + E _ { i } ^ { T } W _ { q } ^ { T } W _ { r } \\phi ( r _ { i j } ) + u ^ { T } W _ { k } E _ { j } + v ^ { T } W _ { r } \\phi ( r _ { i j } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "4abc95bdea1392a92734f005afce75f982c5226661a1db2266b0c52e16c233d4.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 145, + 139, + 465, + 155 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 238 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 133, + 173 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 163, + 154, + 172 + ], + "score": 0.9, + "content": "r _ { i \\to j }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 160, + 291, + 173 + ], + "score": 1.0, + "content": "is the relative distance from token", + "type": "text" + }, + { + "bbox": [ + 291, + 162, + 296, + 170 + ], + "score": 0.78, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 160, + 331, + 173 + ], + "score": 1.0, + "content": "to token", + "type": "text" + }, + { + "bbox": [ + 331, + 161, + 337, + 172 + ], + "score": 0.84, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 160, + 403, + 173 + ], + "score": 1.0, + "content": "in the sequence,", + "type": "text" + }, + { + "bbox": [ + 404, + 159, + 450, + 171 + ], + "score": 0.92, + "content": "\\boldsymbol { u } , \\boldsymbol { v } \\in \\mathbb { R } ^ { d _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "are learnable", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 175, + 185 + ], + "score": 1.0, + "content": "bias vectors, and", + "type": "text" + }, + { + "bbox": [ + 176, + 171, + 192, + 182 + ], + "score": 0.9, + "content": "W _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "is a key projection matrix for the relative distances. Besides fixing issues with", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "absolute position encodings such as ambiguity when processing two sentences at a time, Eq. (5)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 194, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 205 + ], + "score": 1.0, + "content": "enables native application of the powerful self-attention operation on domains such as graphs, where", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "absolute coordinates are not available. We adopt the (non-trainable) sinusoidal encoding function", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 506, + 228 + ], + "score": 1.0, + "content": "proposed by (Vaswani et al., 2017) for all relations; see Appendix A.1 for details on the distance", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 183, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 183, + 239 + ], + "score": 1.0, + "content": "encoding function.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 159, + 506, + 239 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 250, + 451, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 249, + 453, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 453, + 263 + ], + "score": 1.0, + "content": "3.1 INTEGRATING SOURCE CODE AND AST REPRESENTATIONS OF PROGRAMS.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 284 + ], + "score": 1.0, + "content": "To enable the model to integrate information both the Context and Structure of programs, we modify", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "Eq. (5) to be able to incorporate multiple different relations. To this end, we use one key projection", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 102, + 292, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 102, + 292, + 164, + 311 + ], + "score": 1.0, + "content": "matrix W (s)r", + "type": "text" + }, + { + "bbox": [ + 160, + 295, + 213, + 308 + ], + "score": 1.0, + "content": "per relation", + "type": "text" + }, + { + "bbox": [ + 213, + 298, + 219, + 306 + ], + "score": 0.58, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 295, + 505, + 308 + ], + "score": 1.0, + "content": ", and sum their contributions in the raw attention score. This enables", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "the CODE TRANSFORMER to combine information from multiple relations between tokens in the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 318, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 505, + 329 + ], + "score": 1.0, + "content": "attention computation. Besides the token distance in the Context, we include pairwise relations", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 328, + 506, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 506, + 341 + ], + "score": 1.0, + "content": "based on the AST as described in the following. See Fig. 2 for a visualization of the Structure", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 339, + 178, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 178, + 352 + ], + "score": 1.0, + "content": "distances we use.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 102, + 270, + 506, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 360, + 267, + 447 + ], + "lines": [ + { + "bbox": [ + 106, + 359, + 268, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 268, + 371 + ], + "score": 1.0, + "content": "Shortest path length. We include the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 370, + 268, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 268, + 382 + ], + "score": 1.0, + "content": "number of hops required to reach node", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 107, + 382, + 268, + 393 + ], + "spans": [ + { + "bbox": [ + 107, + 382, + 113, + 393 + ], + "score": 0.78, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 382, + 195, + 392 + ], + "score": 1.0, + "content": "starting from node", + "type": "text" + }, + { + "bbox": [ + 196, + 382, + 201, + 391 + ], + "score": 0.7, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 382, + 268, + 392 + ], + "score": 1.0, + "content": "and vice versa.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 392, + 268, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 268, + 403 + ], + "score": 1.0, + "content": "Here, we treat the AST as an undirected", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 403, + 268, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 268, + 414 + ], + "score": 1.0, + "content": "graph, since otherwise most distances", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 414, + 268, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 268, + 426 + ], + "score": 1.0, + "content": "would be undefined: e.g., all other nodes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 425, + 268, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 268, + 437 + ], + "score": 1.0, + "content": "in the AST would be unreachable from", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 435, + 151, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 151, + 448 + ], + "score": 1.0, + "content": "the leaves.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 359, + 268, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 267, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 268, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 268, + 464 + ], + "score": 1.0, + "content": "Similar to the distance of two tokens on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 464, + 268, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 268, + 475 + ], + "score": 1.0, + "content": "the source code sequence, the shortest-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 475, + 269, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 269, + 486 + ], + "score": 1.0, + "content": "path length is a global distance. This", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 486, + 268, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 268, + 498 + ], + "score": 1.0, + "content": "makes the whole graph structure acces-", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 452, + 269, + 498 + ] + }, + { + "type": "image", + "bbox": [ + 278, + 372, + 502, + 469 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 278, + 372, + 502, + 469 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 278, + 372, + 502, + 469 + ], + "spans": [ + { + "bbox": [ + 278, + 372, + 502, + 469 + ], + "score": 0.962, + "type": "image", + "image_path": "971f2174ff96f3b5b7aea4b5cb4f59f9af4bf791dc6a632edbb218691a5d22cf.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 278, + 372, + 502, + 385.85714285714283 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 278, + 385.85714285714283, + 502, + 399.71428571428567 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 278, + 399.71428571428567, + 502, + 413.5714285714285 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 278, + 413.5714285714285, + 502, + 427.42857142857133 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 278, + 427.42857142857133, + 502, + 441.28571428571416 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 278, + 441.28571428571416, + 502, + 455.142857142857 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 278, + 455.142857142857, + 502, + 468.99999999999983 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 293, + 478, + 487, + 489 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 291, + 476, + 488, + 491 + ], + "spans": [ + { + "bbox": [ + 291, + 476, + 488, + 491 + ], + "score": 1.0, + "content": "Figure 2: Structure distances used by our model.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + } + ], + "index": 37.0 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "sible to the model at each layer. In contrast, Hellendoorn et al. (2020) add bias terms to the attention", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "computation only for edges (i.e. shortest-path distance of 1), which is a local operation that only", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 532 + ], + "score": 1.0, + "content": "exchanges information between immediate neighbors (similar to message passing in GNNs). The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "equivalent localized operation on the source code sequence would be to treat the sequence as a chain", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "graph and only compute attention terms for neighboring tokens, which in turn highlights the benefit", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 552, + 241, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 241, + 564 + ], + "score": 1.0, + "content": "of non-local attention operations.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 497, + 506, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 571, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 571, + 504, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 504, + 583 + ], + "score": 1.0, + "content": "Ancestor distance. Since we treat the ASTs as undirected for the computation of the shortest-path", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 595 + ], + "score": 1.0, + "content": "length, we lose the direction information of the edges. To avoid this, we also include the distance", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 591, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 505, + 607 + ], + "score": 1.0, + "content": "on the ordered set of ancestors and descendants of a node in the AST (red arrow in Fig. 2). Again,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "we include number of (vertical) hops to avoid locality in the attention computation. For example, a", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 614, + 432, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 128, + 628 + ], + "score": 1.0, + "content": "node", + "type": "text" + }, + { + "bbox": [ + 128, + 616, + 167, + 628 + ], + "score": 0.92, + "content": "r _ { i \\to j } = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 614, + 252, + 628 + ], + "score": 1.0, + "content": "for “grand-children”", + "type": "text" + }, + { + "bbox": [ + 253, + 616, + 259, + 627 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 614, + 270, + 628 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 271, + 616, + 275, + 625 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 614, + 295, + 628 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 296, + 616, + 343, + 628 + ], + "score": 0.92, + "content": "r _ { j i } = - 2", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 614, + 432, + 628 + ], + "score": 1.0, + "content": "in the other direction.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48, + "bbox_fs": [ + 104, + 571, + 506, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 635, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "Sibling distance. The neighbor sets in graphs are typically considered to be unordered, but in an", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 505, + 658 + ], + "score": 1.0, + "content": "AST, the order of children encodes their order of occurrence in the source code. To avoid information", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 473, + 669 + ], + "score": 1.0, + "content": "loss when encoding the AST, we further include the distance on the ordered set of siblings", + "type": "text" + }, + { + "bbox": [ + 474, + 657, + 493, + 669 + ], + "score": 0.91, + "content": "\\{ v _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 103, + 666, + 507, + 683 + ], + "spans": [ + { + "bbox": [ + 103, + 666, + 432, + 683 + ], + "score": 1.0, + "content": "a node, where we again avoid locality by encoding the number of hops, i.e.", + "type": "text" + }, + { + "bbox": [ + 432, + 669, + 486, + 680 + ], + "score": 0.91, + "content": "r _ { v _ { 1 } v _ { 3 } } ~ = ~ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 666, + 507, + 683 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 104, + 679, + 167, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 679, + 167, + 693 + ], + "score": 1.0, + "content": "rv3→v1 = −2.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 53, + "bbox_fs": [ + 103, + 635, + 507, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Personalized PageRank (Page et al., 1999) (PPR). PPR is a well-studied proximity measure which", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "has been shown to be very effective in learning with graphs (Klicpera et al., 2019a;b; Bojchevski", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 422, + 734 + ], + "score": 1.0, + "content": "et al., 2020). PPR captures the local graph structure around a pair of nodes", + "type": "text" + }, + { + "bbox": [ + 422, + 721, + 443, + 732 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 720, + 482, + 734 + ], + "score": 1.0, + "content": ". E.g., if", + "type": "text" + }, + { + "bbox": [ + 482, + 721, + 487, + 730 + ], + "score": 0.75, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "has", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 250, + 365 + ], + "score": 1.0, + "content": "many neighbors, its PPR score for", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 250, + 353, + 257, + 363 + ], + "score": 0.84, + "content": "j", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 257, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "will be low even when they are only few hops apart, which", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 363, + 357, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 357, + 375 + ], + "score": 1.0, + "content": "complements the purely hop-based distances described above.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 57, + "bbox_fs": [ + 105, + 698, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 80, + 498, + 249 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 80, + 498, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 80, + 498, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 80, + 498, + 249 + ], + "score": 0.974, + "type": "image", + "image_path": "2ae4d77322414ffc49b55c0b5e39c6d3b3d0f5b89a9cfca8c5f81d4390780d66.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 80, + 498, + 136.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 136.33333333333334, + 498, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 192.66666666666669, + 498, + 249.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 261, + 505, + 327 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "Figure 3: Left: Sequence (Context) and AST (Structure) representation of an input code snippet.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "Center: The CODE TRANSFORMER jointly leverages the sequence of tokens and the Abstract Syn-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "tax Tree to learn expressive representations of source code. In addition to the input token and node", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "embeddings the model uses different distances between the tokens, e.g., shortest paths on the AST", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "or personalized PageRank, to reason about their relative positions. The output embeddings can be", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 316, + 361, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 361, + 328 + ], + "score": 1.0, + "content": "used for downstream tasks such as code summarization (right).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 250, + 365 + ], + "score": 1.0, + "content": "many neighbors, its PPR score for", + "type": "text" + }, + { + "bbox": [ + 250, + 353, + 257, + 363 + ], + "score": 0.84, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 352, + 505, + 365 + ], + "score": 1.0, + "content": "will be low even when they are only few hops apart, which", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 363, + 357, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 357, + 375 + ], + "score": 1.0, + "content": "complements the purely hop-based distances described above.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "score": 1.0, + "content": "Input embeddings to the model. To combine the Context and Structure information, we assign", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "each token in the sequence to an AST node by selecting the AST node whose range in the source", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 405, + 504, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 504, + 417 + ], + "score": 1.0, + "content": "code is the shortest one containing the token. We concatenate the (sub-) token embeddings with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "the embedding of the token’s assigned AST node type as well as the token type returned by the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "tokenizer. That is, among all the internal nodes, we use as input only those corresponding to a token", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "in the sequence; however, the remaining internal nodes can used by the model since their presence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 448, + 496, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 496, + 461 + ], + "score": 1.0, + "content": "affects the distances between the remaining AST nodes. See Appendices A.3 and A.4 for details.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 477, + 338, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 341, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 341, + 489 + ], + "score": 1.0, + "content": "3.2 EFFICIENT RELATIVE ATTENTION COMPUTATION.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 376, + 510 + ], + "score": 1.0, + "content": "Na¨ıvely, we need to compute and materialize a tensor of dimension", + "type": "text" + }, + { + "bbox": [ + 376, + 498, + 425, + 509 + ], + "score": 0.91, + "content": "N \\times N \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "to hold all pairwise", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 216, + 523 + ], + "score": 1.0, + "content": "relative position encodings", + "type": "text" + }, + { + "bbox": [ + 217, + 509, + 250, + 522 + ], + "score": 0.92, + "content": "\\phi ( r _ { i \\to j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 509, + 325, + 523 + ], + "score": 1.0, + "content": "in Eq. (5) , where", + "type": "text" + }, + { + "bbox": [ + 325, + 510, + 336, + 519 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "is the input length. This is prohibitive for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "score": 1.0, + "content": "fast GPU training. While for discrete distance values (e.g., sequence distance or shortest-path length", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "score": 1.0, + "content": "on a graph) we only need to compute unique distance values occurring in the input, this does not gen-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "eralize to continuous distances such as PPR. Therefore, we propose a constant-time approximation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 552, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 373, + 567 + ], + "score": 1.0, + "content": "of the relational attention computation by grouping the values into", + "type": "text" + }, + { + "bbox": [ + 373, + 553, + 409, + 564 + ], + "score": 0.91, + "content": "k \\ll N ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 552, + 505, + 567 + ], + "score": 1.0, + "content": "bins. Since closer sam-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "ples are typically more relevant for a query sample, we increase the bin widths exponentially with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "score": 1.0, + "content": "growing distance values. Throughout our experiments we have found the CODE TRANSFORMER to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 586, + 450, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 340, + 599 + ], + "score": 1.0, + "content": "be relatively insensitive to the number of bins; we thus set", + "type": "text" + }, + { + "bbox": [ + 340, + 586, + 370, + 596 + ], + "score": 0.9, + "content": "k = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 587, + 450, + 599 + ], + "score": 1.0, + "content": "in our experiments.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 617, + 242, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 244, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 244, + 631 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "Code summarization is one of the most popular tasks in machine learning for code. Given the body", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "of a function, the task is to predict the function’s name. As observed by Alon et al. (2019b) and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "Allamanis et al. (2016), this is a useful benchmark as method names in open-source projects tend to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "be precise and descriptive, and functions typically form complete logical units. See Fig. 3 (right) for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "a visual overview of the task. We use two complementary representations of programs: the source", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "code as a sequence of tokens (Context) and the AST (Structure). As shown in Fig. 3 (left), tokens", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "that are far away on the sequence may be very close on the AST and vice versa. In this task we make", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "use of the CODE TRANSFORMER’s ability jointly leverage both Structure and Context and show that", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 80, + 498, + 249 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 80, + 498, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 80, + 498, + 249 + ], + "spans": [ + { + "bbox": [ + 115, + 80, + 498, + 249 + ], + "score": 0.974, + "type": "image", + "image_path": "2ae4d77322414ffc49b55c0b5e39c6d3b3d0f5b89a9cfca8c5f81d4390780d66.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 80, + 498, + 136.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 136.33333333333334, + 498, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 192.66666666666669, + 498, + 249.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 261, + 505, + 327 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 505, + 274 + ], + "score": 1.0, + "content": "Figure 3: Left: Sequence (Context) and AST (Structure) representation of an input code snippet.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "Center: The CODE TRANSFORMER jointly leverages the sequence of tokens and the Abstract Syn-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 296 + ], + "score": 1.0, + "content": "tax Tree to learn expressive representations of source code. In addition to the input token and node", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 505, + 306 + ], + "score": 1.0, + "content": "embeddings the model uses different distances between the tokens, e.g., shortest paths on the AST", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "or personalized PageRank, to reason about their relative positions. The output embeddings can be", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 316, + 361, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 361, + 328 + ], + "score": 1.0, + "content": "used for downstream tasks such as code summarization (right).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 107, + 351, + 504, + 374 + ], + "lines": [], + "index": 9.5, + "bbox_fs": [ + 105, + 352, + 505, + 375 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 396 + ], + "score": 1.0, + "content": "Input embeddings to the model. To combine the Context and Structure information, we assign", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "score": 1.0, + "content": "each token in the sequence to an AST node by selecting the AST node whose range in the source", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 405, + 504, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 504, + 417 + ], + "score": 1.0, + "content": "code is the shortest one containing the token. We concatenate the (sub-) token embeddings with", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 428 + ], + "score": 1.0, + "content": "the embedding of the token’s assigned AST node type as well as the token type returned by the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "tokenizer. That is, among all the internal nodes, we use as input only those corresponding to a token", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "in the sequence; however, the remaining internal nodes can used by the model since their presence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 448, + 496, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 496, + 461 + ], + "score": 1.0, + "content": "affects the distances between the remaining AST nodes. See Appendices A.3 and A.4 for details.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 381, + 505, + 461 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 477, + 338, + 488 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 341, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 341, + 489 + ], + "score": 1.0, + "content": "3.2 EFFICIENT RELATIVE ATTENTION COMPUTATION.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 498, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 376, + 510 + ], + "score": 1.0, + "content": "Na¨ıvely, we need to compute and materialize a tensor of dimension", + "type": "text" + }, + { + "bbox": [ + 376, + 498, + 425, + 509 + ], + "score": 0.91, + "content": "N \\times N \\times d", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "to hold all pairwise", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 509, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 216, + 523 + ], + "score": 1.0, + "content": "relative position encodings", + "type": "text" + }, + { + "bbox": [ + 217, + 509, + 250, + 522 + ], + "score": 0.92, + "content": "\\phi ( r _ { i \\to j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 509, + 325, + 523 + ], + "score": 1.0, + "content": "in Eq. (5) , where", + "type": "text" + }, + { + "bbox": [ + 325, + 510, + 336, + 519 + ], + "score": 0.81, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 509, + 505, + 523 + ], + "score": 1.0, + "content": "is the input length. This is prohibitive for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 533 + ], + "score": 1.0, + "content": "fast GPU training. While for discrete distance values (e.g., sequence distance or shortest-path length", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 545 + ], + "score": 1.0, + "content": "on a graph) we only need to compute unique distance values occurring in the input, this does not gen-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "eralize to continuous distances such as PPR. Therefore, we propose a constant-time approximation", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 552, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 373, + 567 + ], + "score": 1.0, + "content": "of the relational attention computation by grouping the values into", + "type": "text" + }, + { + "bbox": [ + 373, + 553, + 409, + 564 + ], + "score": 0.91, + "content": "k \\ll N ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 552, + 505, + 567 + ], + "score": 1.0, + "content": "bins. Since closer sam-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "ples are typically more relevant for a query sample, we increase the bin widths exponentially with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 506, + 588 + ], + "score": 1.0, + "content": "growing distance values. Throughout our experiments we have found the CODE TRANSFORMER to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 586, + 450, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 340, + 599 + ], + "score": 1.0, + "content": "be relatively insensitive to the number of bins; we thus set", + "type": "text" + }, + { + "bbox": [ + 340, + 586, + 370, + 596 + ], + "score": 0.9, + "content": "k = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 371, + 587, + 450, + 599 + ], + "score": 1.0, + "content": "in our experiments.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 497, + 506, + 599 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 617, + 242, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 244, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 244, + 631 + ], + "score": 1.0, + "content": "4 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "Code summarization is one of the most popular tasks in machine learning for code. Given the body", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "of a function, the task is to predict the function’s name. As observed by Alon et al. (2019b) and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "Allamanis et al. (2016), this is a useful benchmark as method names in open-source projects tend to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "be precise and descriptive, and functions typically form complete logical units. See Fig. 3 (right) for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "a visual overview of the task. We use two complementary representations of programs: the source", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "code as a sequence of tokens (Context) and the AST (Structure). As shown in Fig. 3 (left), tokens", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "that are far away on the sequence may be very close on the AST and vice versa. In this task we make", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "use of the CODE TRANSFORMER’s ability jointly leverage both Structure and Context and show that", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "it improves learning. Further, we show the benefit of using only language-agnostic features in our", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 391, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 391, + 106 + ], + "score": 1.0, + "content": "model by training the first multilingual model for code summarization.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 643, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "it improves learning. Further, we show the benefit of using only language-agnostic features in our", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 391, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 391, + 106 + ], + "score": 1.0, + "content": "model by training the first multilingual model for code summarization.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 113, + 504, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "Datasets. To highlight the benefit of only relying on language-agnostic representations such as", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 504, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 504, + 137 + ], + "score": 1.0, + "content": "source code and abstract syntax trees, we evaluate on challenging datasets in four programming lan-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "score": 1.0, + "content": "guages introduced in the CodeSearchNet (CSN) Challenge (Husain et al., 2019): Python, Javascript,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "Go, and Ruby. Similar to Java-small, the datasets from CodeSearchNet have been carefully dedu-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 495, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 495, + 170 + ], + "score": 1.0, + "content": "plicated by the creators to avoid data leakage from the training set, e.g., via copy-and-paste code.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 324, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 324, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 324, + 186 + ], + "score": 1.0, + "content": "We further evaluate on Java-small (Allamanis et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 325, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 325, + 198 + ], + "score": 1.0, + "content": "2016), a popular and challenging code summarization", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 196, + 325, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 325, + 208 + ], + "score": 1.0, + "content": "dataset. It contains 11 open-source Java projects. We", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 207, + 325, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 325, + 218 + ], + "score": 1.0, + "content": "use the split as in Alon et al. (2019a), where 9 of these", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 325, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 325, + 230 + ], + "score": 1.0, + "content": "projects are used for training, one for validation, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 228, + 325, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 325, + 240 + ], + "score": 1.0, + "content": "one for test. The dataset contains roughly 700K sam-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 240, + 325, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 325, + 252 + ], + "score": 1.0, + "content": "ples (function definitions). Moreover, we also experi-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 325, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 325, + 263 + ], + "score": 1.0, + "content": "ment with pre-training our model on Java-medium and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 262, + 325, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 325, + 275 + ], + "score": 1.0, + "content": "Java-large (Alon et al., 2019a) before fine-tuning on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 325, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 325, + 285 + ], + "score": 1.0, + "content": "Java-small, making sure to avoid leakage by removing", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 284, + 325, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 325, + 295 + ], + "score": 1.0, + "content": "the test and validation projects of Java-small from the", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 334, + 187, + 505, + 269 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 334, + 187, + 505, + 269 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 334, + 187, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 334, + 187, + 505, + 269 + ], + "score": 0.977, + "html": "
Samples per partition
DatasetTrainVal.Test
CSN-Python412,17823,10722,176
CSN-Javascript123,8898,2536,483
CSN-Ruby48,7912,2092,279
CSN-Go317,83214,24214,291
Java-small691,97423,84457,088
", + "type": "table", + "image_path": "b4225af91868e1a8a78db0e52181e077f82a2809a40ec936d5ee0af636f96fb7.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 334, + 187, + 505, + 200.66666666666666 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 334, + 200.66666666666666, + 505, + 214.33333333333331 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 334, + 214.33333333333331, + 505, + 227.99999999999997 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 334, + 227.99999999999997, + 505, + 241.66666666666663 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 334, + 241.66666666666663, + 505, + 255.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 334, + 255.3333333333333, + 505, + 268.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 366, + 277, + 471, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 365, + 276, + 472, + 288 + ], + "spans": [ + { + "bbox": [ + 365, + 276, + 472, + 288 + ], + "score": 1.0, + "content": "Table 1: Dataset statistics.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 22.25 + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 439, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 441, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 441, + 307 + ], + "score": 1.0, + "content": "pre-training dataset. See Table 1 for a summary of the datasets we use in this work.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 314, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 328 + ], + "score": 1.0, + "content": "Preprocessing. Each token of the source code is split into subtokens respective to code naming", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 324, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 339 + ], + "score": 1.0, + "content": "conventions, i.e., get TrainingData is converted to [get, training, data]. Following", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 337, + 504, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 504, + 348 + ], + "score": 1.0, + "content": "Alon et al. (2019a) we use at most six subtokens for the method names, truncating longer function", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "names if necessary. In addition to the tokenized source code we produce an AST for each method", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "using the open-source AST parser Semantic3. We limit the vocabulary to subtokens with at least 100", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "score": 1.0, + "content": "occurrences in the training set, and only consider snippets with 512 or fewer tokens (after removing", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 379, + 484, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 484, + 394 + ], + "score": 1.0, + "content": "punctuation). We refer the reader to the appendix for further details on the data preprocessing.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 399, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Pointer network. We add a pointer network (Vinyals et al., 2015) (as described in Fernandes et al.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "(2019)) to the decoders of all Transformer-based models. This enables them to enhance their predic-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "tions by pointing at positions in the input sequence. For instance, when predicting the method name", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "get url, the model can point directly to occurrences of the variable url. This often improves", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "results for less frequent tokens, and even enables the model to predict tokens which are not in the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 455, + 320, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 320, + 468 + ], + "score": 1.0, + "content": "vocabulary by pointing at their positions in the input.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Baselines. We compare with code2seq (Alon et al., 2019a), the Graph Relational Embedding At-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 432, + 498 + ], + "score": 1.0, + "content": "tention Transformer (GREAT) (Hellendoorn et al., 2020), and the BiLSTM+GNN", + "type": "text" + }, + { + "bbox": [ + 433, + 486, + 442, + 495 + ], + "score": 0.75, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "LSTM+Pointer", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "model presented in Fernandes et al. (2019). Code2seq is a non-Transformer model and state of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "art for code summarization using only AST information. GREAT is a recent Transformer model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 518, + 504, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 504, + 531 + ], + "score": 1.0, + "content": "using the framework presented in (Shaw et al., 2018) to bias the attention via edges. In the origi-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "nal formulation, GREAT additionally uses hand-crafted, language-specific edges such as dataflow,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 540, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 506, + 554 + ], + "score": 1.0, + "content": "‘computed from’, or ‘next lexical use’ edges, which require specialized preprocessing and static", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 550, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 565 + ], + "score": 1.0, + "content": "analysis tools. While this approach of leveraging language-specific features can certainly improve", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "results on specific tasks and programming languages, our goal is to have a flexible model that can", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 573, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 587 + ], + "score": 1.0, + "content": "be used on any programming language. Since the specialized preprocessing used by GREAT is pro-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 584, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 506, + 597 + ], + "score": 1.0, + "content": "prietary and not public, we produce the results for GREAT using edges from the AST instead, i.e. it", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "has access to the same information as our proposed model. Note that the preprocessing of Fernandes", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 606, + 496, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 496, + 619 + ], + "score": 1.0, + "content": "et al. (2019) is language specific, which is why we only compare with their results on Java-small.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45 + }, + { + "type": "title", + "bbox": [ + 107, + 633, + 172, + 646 + ], + "lines": [ + { + "bbox": [ + 104, + 632, + 174, + 649 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 174, + 649 + ], + "score": 1.0, + "content": "5 RESULTS", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 52 + }, + { + "type": "title", + "bbox": [ + 109, + 659, + 300, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 302, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 302, + 671 + ], + "score": 1.0, + "content": "5.1 MONOLINGUAL CODE SUMMARIZATION", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 108, + 679, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "CSN dataset. First, we study the performance (measured by F1 score) of our model and the base-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "score": 1.0, + "content": "lines on the traditional setting, where training and evaluation are performed on a single programming", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 702, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 713 + ], + "score": 1.0, + "content": "language. The results are shown in the upper part of Table 2. The CODE TRANSFORMER (without", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 55 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 722, + 308, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 308, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 308, + 733 + ], + "score": 1.0, + "content": "3https://github.com/github/semantic", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 113, + 504, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "Datasets. To highlight the benefit of only relying on language-agnostic representations such as", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 504, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 504, + 137 + ], + "score": 1.0, + "content": "source code and abstract syntax trees, we evaluate on challenging datasets in four programming lan-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 506, + 149 + ], + "score": 1.0, + "content": "guages introduced in the CodeSearchNet (CSN) Challenge (Husain et al., 2019): Python, Javascript,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "Go, and Ruby. Similar to Java-small, the datasets from CodeSearchNet have been carefully dedu-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 495, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 495, + 170 + ], + "score": 1.0, + "content": "plicated by the creators to avoid data leakage from the training set, e.g., via copy-and-paste code.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 114, + 506, + 170 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 174, + 324, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 324, + 186 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 324, + 186 + ], + "score": 1.0, + "content": "We further evaluate on Java-small (Allamanis et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 185, + 325, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 325, + 198 + ], + "score": 1.0, + "content": "2016), a popular and challenging code summarization", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 196, + 325, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 325, + 208 + ], + "score": 1.0, + "content": "dataset. It contains 11 open-source Java projects. We", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 207, + 325, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 325, + 218 + ], + "score": 1.0, + "content": "use the split as in Alon et al. (2019a), where 9 of these", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 325, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 325, + 230 + ], + "score": 1.0, + "content": "projects are used for training, one for validation, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 228, + 325, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 325, + 240 + ], + "score": 1.0, + "content": "one for test. The dataset contains roughly 700K sam-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 240, + 325, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 325, + 252 + ], + "score": 1.0, + "content": "ples (function definitions). Moreover, we also experi-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 251, + 325, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 325, + 263 + ], + "score": 1.0, + "content": "ment with pre-training our model on Java-medium and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 262, + 325, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 325, + 275 + ], + "score": 1.0, + "content": "Java-large (Alon et al., 2019a) before fine-tuning on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 272, + 325, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 325, + 285 + ], + "score": 1.0, + "content": "Java-small, making sure to avoid leakage by removing", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 284, + 325, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 325, + 295 + ], + "score": 1.0, + "content": "the test and validation projects of Java-small from the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 294, + 441, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 441, + 307 + ], + "score": 1.0, + "content": "pre-training dataset. See Table 1 for a summary of the datasets we use in this work.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 174, + 325, + 295 + ] + }, + { + "type": "table", + "bbox": [ + 334, + 187, + 505, + 269 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 334, + 187, + 505, + 269 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 334, + 187, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 334, + 187, + 505, + 269 + ], + "score": 0.977, + "html": "
Samples per partition
DatasetTrainVal.Test
CSN-Python412,17823,10722,176
CSN-Javascript123,8898,2536,483
CSN-Ruby48,7912,2092,279
CSN-Go317,83214,24214,291
Java-small691,97423,84457,088
", + "type": "table", + "image_path": "b4225af91868e1a8a78db0e52181e077f82a2809a40ec936d5ee0af636f96fb7.jpg" + } + ] + } + ], + "index": 20.5, + "virtual_lines": [ + { + "bbox": [ + 334, + 187, + 505, + 200.66666666666666 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 334, + 200.66666666666666, + 505, + 214.33333333333331 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 334, + 214.33333333333331, + 505, + 227.99999999999997 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 334, + 227.99999999999997, + 505, + 241.66666666666663 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 334, + 241.66666666666663, + 505, + 255.3333333333333 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 334, + 255.3333333333333, + 505, + 268.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 366, + 277, + 471, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 365, + 276, + 472, + 288 + ], + "spans": [ + { + "bbox": [ + 365, + 276, + 472, + 288 + ], + "score": 1.0, + "content": "Table 1: Dataset statistics.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + } + ], + "index": 22.25 + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 439, + 306 + ], + "lines": [], + "index": 25, + "bbox_fs": [ + 105, + 294, + 441, + 307 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 314, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 105, + 313, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 328 + ], + "score": 1.0, + "content": "Preprocessing. Each token of the source code is split into subtokens respective to code naming", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 324, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 339 + ], + "score": 1.0, + "content": "conventions, i.e., get TrainingData is converted to [get, training, data]. Following", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 337, + 504, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 504, + 348 + ], + "score": 1.0, + "content": "Alon et al. (2019a) we use at most six subtokens for the method names, truncating longer function", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "names if necessary. In addition to the tokenized source code we produce an AST for each method", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "using the open-source AST parser Semantic3. We limit the vocabulary to subtokens with at least 100", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 506, + 383 + ], + "score": 1.0, + "content": "occurrences in the training set, and only consider snippets with 512 or fewer tokens (after removing", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 379, + 484, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 484, + 394 + ], + "score": 1.0, + "content": "punctuation). We refer the reader to the appendix for further details on the data preprocessing.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 313, + 506, + 394 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 399, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "Pointer network. We add a pointer network (Vinyals et al., 2015) (as described in Fernandes et al.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "(2019)) to the decoders of all Transformer-based models. This enables them to enhance their predic-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 505, + 434 + ], + "score": 1.0, + "content": "tions by pointing at positions in the input sequence. For instance, when predicting the method name", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 446 + ], + "score": 1.0, + "content": "get url, the model can point directly to occurrences of the variable url. This often improves", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "results for less frequent tokens, and even enables the model to predict tokens which are not in the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 455, + 320, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 320, + 468 + ], + "score": 1.0, + "content": "vocabulary by pointing at their positions in the input.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 400, + 506, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "Baselines. We compare with code2seq (Alon et al., 2019a), the Graph Relational Embedding At-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 432, + 498 + ], + "score": 1.0, + "content": "tention Transformer (GREAT) (Hellendoorn et al., 2020), and the BiLSTM+GNN", + "type": "text" + }, + { + "bbox": [ + 433, + 486, + 442, + 495 + ], + "score": 0.75, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "LSTM+Pointer", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "model presented in Fernandes et al. (2019). Code2seq is a non-Transformer model and state of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "art for code summarization using only AST information. GREAT is a recent Transformer model", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 518, + 504, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 504, + 531 + ], + "score": 1.0, + "content": "using the framework presented in (Shaw et al., 2018) to bias the attention via edges. In the origi-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "nal formulation, GREAT additionally uses hand-crafted, language-specific edges such as dataflow,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 540, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 506, + 554 + ], + "score": 1.0, + "content": "‘computed from’, or ‘next lexical use’ edges, which require specialized preprocessing and static", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 550, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 565 + ], + "score": 1.0, + "content": "analysis tools. While this approach of leveraging language-specific features can certainly improve", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "results on specific tasks and programming languages, our goal is to have a flexible model that can", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 573, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 587 + ], + "score": 1.0, + "content": "be used on any programming language. Since the specialized preprocessing used by GREAT is pro-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 584, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 506, + 597 + ], + "score": 1.0, + "content": "prietary and not public, we produce the results for GREAT using edges from the AST instead, i.e. it", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "score": 1.0, + "content": "has access to the same information as our proposed model. Note that the preprocessing of Fernandes", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 606, + 496, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 496, + 619 + ], + "score": 1.0, + "content": "et al. (2019) is language specific, which is why we only compare with their results on Java-small.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 474, + 506, + 619 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 633, + 172, + 646 + ], + "lines": [ + { + "bbox": [ + 104, + 632, + 174, + 649 + ], + "spans": [ + { + "bbox": [ + 104, + 632, + 174, + 649 + ], + "score": 1.0, + "content": "5 RESULTS", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 52 + }, + { + "type": "title", + "bbox": [ + 109, + 659, + 300, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 659, + 302, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 659, + 302, + 671 + ], + "score": 1.0, + "content": "5.1 MONOLINGUAL CODE SUMMARIZATION", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 108, + 679, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "CSN dataset. First, we study the performance (measured by F1 score) of our model and the base-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "score": 1.0, + "content": "lines on the traditional setting, where training and evaluation are performed on a single programming", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 702, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 702, + 505, + 713 + ], + "score": 1.0, + "content": "language. The results are shown in the upper part of Table 2. The CODE TRANSFORMER (without", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 234, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 504, + 246 + ], + "score": 1.0, + "content": "multi-language training) substantially outperforms all other models on all but one language, high-", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "lighting the effectiveness of jointly learning from Structure and Context. The only exception is Ruby,", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "where it performs on par with its Context-only variant. We attribute this to the fact that there are", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "relatively few samples in the Ruby dataset, and that Ruby is an dynamically typed language, which", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "score": 1.0, + "content": "could make the Structure less powerful for learning. Interestingly, the Context-only CODE TRANS-", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "FORMER outperforms GREAT on all languages. We attribute this to the fact that GREAT uses the", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "Structure of the programs only in a localized way (see Sec. 3.1). Another noteworthy finding is", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 310, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 323 + ], + "score": 1.0, + "content": "that code2seq performs comparably to the Transformer-based baselines on Go. We hypothesize that", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 320, + 449, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 449, + 335 + ], + "score": 1.0, + "content": "ASTs are more informative on Go since it is a compiled and strongly typed language.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 55, + "bbox_fs": [ + 105, + 678, + 506, + 713 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 63, + 503, + 192 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 63, + 503, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 63, + 503, + 192 + ], + "spans": [ + { + "bbox": [ + 108, + 63, + 503, + 192 + ], + "score": 0.982, + "html": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq35.7924.8529.3430.1819.8823.9723.2310.3114.2852.3043.4347.45
GREAT35.0731.5933.2431.2026.8428.8624.6422.2323.3850.0146.5148.20
Ours w/o structure37.3831.9834.4733.1726.7029.5929.8525.8727.7251.7847.5749.59
Ours w/o pointer net37.7431.8534.5533.1228.7030.7523.3225.2124.2354.3150.1252.13
Ours36.4033.6634.9735.0629.6132.1131.4224.4627.5055.1048.0551.34
code2seq (Multilanguage)34.4925.4929.3231.6222.1626.0623.9717.0619.9352.7044.3648.17
GREAT (Multilanguage)36.7531.5433.9433.5827.7830.4130.0524.3326.8952.6548.3050.38
Ours w/o structure (Mult.)38.4830.1433.8035.3827.4130.8932.6126.7629.4055.0347.3450.90
Ours w/o pointer (Mult.)38.9133.1235.7837.2129.7533.0734.5227.3130.5056.0750.7653.28
Ours (Multilanguage)38.8933.8236.1836.9529.9833.1033.9328.9431.2456.0050.4453.07
Ours (Mult.+ Finetune)39.8532.7935.9837.0029.7933.0035.8527.7531.2855.6351.1253.28
Ours (Mult. + LM Pretrain)39.6735.2937.3537.0631.9434.3135.1929.3632.0157.7351.8954.65
", + "type": "table", + "image_path": "6ea8249b1a6b65be5a059f68d87e3dba30b23b79c16802b7d266349476c02871.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 63, + 503, + 106.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 106.0, + 503, + 149.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 149.0, + 503, + 192.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 167, + 200, + 442, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 198, + 443, + 213 + ], + "spans": [ + { + "bbox": [ + 167, + 198, + 443, + 213 + ], + "score": 1.0, + "content": "Table 2: Code summarization results on the CSN dataset (micro F1).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 234, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 504, + 246 + ], + "score": 1.0, + "content": "multi-language training) substantially outperforms all other models on all but one language, high-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "lighting the effectiveness of jointly learning from Structure and Context. The only exception is Ruby,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "where it performs on par with its Context-only variant. We attribute this to the fact that there are", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 279 + ], + "score": 1.0, + "content": "relatively few samples in the Ruby dataset, and that Ruby is an dynamically typed language, which", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 278, + 505, + 289 + ], + "score": 1.0, + "content": "could make the Structure less powerful for learning. Interestingly, the Context-only CODE TRANS-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "FORMER outperforms GREAT on all languages. We attribute this to the fact that GREAT uses the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "Structure of the programs only in a localized way (see Sec. 3.1). Another noteworthy finding is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 310, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 323 + ], + "score": 1.0, + "content": "that code2seq performs comparably to the Transformer-based baselines on Go. We hypothesize that", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 320, + 449, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 449, + 335 + ], + "score": 1.0, + "content": "ASTs are more informative on Go since it is a compiled and strongly typed language.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Java-small results. In Table 3 we present code summarization results on the Java-small dataset.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Among all models equipped with a pointer network, the CODE TRANSFORMER (without pre-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "training) obtains state-of-the-art on code summarization, outperforming all baselines, including the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "score": 1.0, + "content": "previous state-of-the-art on Java-small proposed by Fernandes et al. (2019). Further, pre-training", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 388, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 402 + ], + "score": 1.0, + "content": "on Java-medium and Java-large on the permutation language modeling objective (Yang et al., 2019)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "score": 1.0, + "content": "substantially improves precision, recall, and F1 score after fine-tuning on Java-small. To avoid leak-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 411, + 495, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 495, + 425 + ], + "score": 1.0, + "content": "age, we exclude the projects used in the validation and test splits of Java-small from pre-training.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 504, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "Ablation study. We further perform ablations where we remove our model’s access to the Context", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 446, + 263, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 263, + 457 + ], + "score": 1.0, + "content": "or Structure, also presented in Table 3.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 295, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 296, + 474 + ], + "score": 1.0, + "content": "With pointer network. We find that both ab-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 473, + 297, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 297, + 485 + ], + "score": 1.0, + "content": "lations lead to a substantial drop in perfor-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 484, + 297, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 297, + 497 + ], + "score": 1.0, + "content": "mance, highlighting the benefit of learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 495, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 297, + 507 + ], + "score": 1.0, + "content": "jointly from Structure and Context. Interest-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 506, + 297, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 297, + 518 + ], + "score": 1.0, + "content": "ingly, the model without access to the Struc-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 518, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 297, + 529 + ], + "score": 1.0, + "content": "ture performs slightly better than the variant", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 528, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 297, + 539 + ], + "score": 1.0, + "content": "without Context. Note that our model with-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 538, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 297, + 552 + ], + "score": 1.0, + "content": "out Structure is related to the XLNet (Yang", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 550, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 297, + 562 + ], + "score": 1.0, + "content": "et al., 2019) model, where we add a pointer net-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 561, + 297, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 297, + 573 + ], + "score": 1.0, + "content": "work to the decoder and concatenate the token", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 572, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 297, + 585 + ], + "score": 1.0, + "content": "types to their respective input tokens (see Ap-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 583, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 297, + 594 + ], + "score": 1.0, + "content": "pendix A.4). Without pointer network. We re-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 595, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 297, + 606 + ], + "score": 1.0, + "content": "peat the ablation on the variants without pointer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 605, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 297, + 617 + ], + "score": 1.0, + "content": "network. Here, the variant without Context per-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 615, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 297, + 628 + ], + "score": 1.0, + "content": "forms better than the variant without Structure,", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 315, + 462, + 496, + 603 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 315, + 462, + 496, + 603 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 315, + 462, + 496, + 603 + ], + "spans": [ + { + "bbox": [ + 315, + 462, + 496, + 603 + ], + "score": 0.98, + "html": "
ModelPrec.Rec.F1
Without pointer net code2seq Ours w/o structure Ours w/o context51.23 50.70 51.8137.31 45.49 46.0443.18 47.96
Ours With pointer net50.3346.8048.75 48.50
Fernandes et al. (2019)=151.4
GREAT Ours w/o structure53.6046.4149.75
Ours w/o context55.4846.0750.34
54.4545.2949.45
Ours54.8549.8452.22
Ours + Pretrain57.0250.8753.77
", + "type": "table", + "image_path": "93c3db634ac78ceb6a3e42a79f6c881109ae26412d90ebf329ef70021d85d788.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 315, + 462, + 496, + 474.8181818181818 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 315, + 474.8181818181818, + 496, + 487.6363636363636 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 315, + 487.6363636363636, + 496, + 500.45454545454544 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 315, + 500.45454545454544, + 496, + 513.2727272727273 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 315, + 513.2727272727273, + 496, + 526.0909090909091 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 315, + 526.0909090909091, + 496, + 538.909090909091 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 315, + 538.909090909091, + 496, + 551.7272727272729 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 315, + 551.7272727272729, + 496, + 564.5454545454547 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 315, + 564.5454545454547, + 496, + 577.3636363636366 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 315, + 577.3636363636366, + 496, + 590.1818181818185 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 315, + 590.1818181818185, + 496, + 603.0000000000003 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 304, + 612, + 504, + 622 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 303, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 303, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "Table 3: Results on Java-small and ablation study.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + } + ], + "index": 45.0 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "indicating that the pointer network helps to compensate for the lack of access to Structure. The", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 107, + 639, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 639, + 504, + 649 + ], + "score": 1.0, + "content": "Structure-only variant (w/o pointer net) of our model even outperforms the full variant in this sce-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "nario. Inspection of the results revealed that the Structure-only variant has better performance on", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 660, + 489, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 489, + 671 + ], + "score": 1.0, + "content": "longer method names, which have an outsize influence on the micro-F1 score used in this work.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 688 + ], + "score": 1.0, + "content": "Ablation of the AST-based distances. In Table 4 we compare the performance of our model when", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "trained with each of the four different AST distances (sibling shortest paths, ancestor shortest paths,", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "shortest paths, personalized PageRank; see Section 3.1). Here, the model is trained on Java-small", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "in the Structure-only setting and without pointer network. For reference, we also show the results", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "of training our model using all four AST distance functions (c.f. Table 3). We find that, while the", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 55 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 63, + 503, + 192 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 63, + 503, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 63, + 503, + 192 + ], + "spans": [ + { + "bbox": [ + 108, + 63, + 503, + 192 + ], + "score": 0.982, + "html": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq35.7924.8529.3430.1819.8823.9723.2310.3114.2852.3043.4347.45
GREAT35.0731.5933.2431.2026.8428.8624.6422.2323.3850.0146.5148.20
Ours w/o structure37.3831.9834.4733.1726.7029.5929.8525.8727.7251.7847.5749.59
Ours w/o pointer net37.7431.8534.5533.1228.7030.7523.3225.2124.2354.3150.1252.13
Ours36.4033.6634.9735.0629.6132.1131.4224.4627.5055.1048.0551.34
code2seq (Multilanguage)34.4925.4929.3231.6222.1626.0623.9717.0619.9352.7044.3648.17
GREAT (Multilanguage)36.7531.5433.9433.5827.7830.4130.0524.3326.8952.6548.3050.38
Ours w/o structure (Mult.)38.4830.1433.8035.3827.4130.8932.6126.7629.4055.0347.3450.90
Ours w/o pointer (Mult.)38.9133.1235.7837.2129.7533.0734.5227.3130.5056.0750.7653.28
Ours (Multilanguage)38.8933.8236.1836.9529.9833.1033.9328.9431.2456.0050.4453.07
Ours (Mult.+ Finetune)39.8532.7935.9837.0029.7933.0035.8527.7531.2855.6351.1253.28
Ours (Mult. + LM Pretrain)39.6735.2937.3537.0631.9434.3135.1929.3632.0157.7351.8954.65
", + "type": "table", + "image_path": "6ea8249b1a6b65be5a059f68d87e3dba30b23b79c16802b7d266349476c02871.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 63, + 503, + 106.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 106.0, + 503, + 149.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 149.0, + 503, + 192.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 167, + 200, + 442, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 198, + 443, + 213 + ], + "spans": [ + { + "bbox": [ + 167, + 198, + 443, + 213 + ], + "score": 1.0, + "content": "Table 2: Code summarization results on the CSN dataset (micro F1).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 333 + ], + "lines": [], + "index": 8, + "bbox_fs": [ + 105, + 234, + 506, + 335 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Java-small results. In Table 3 we present code summarization results on the Java-small dataset.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "Among all models equipped with a pointer network, the CODE TRANSFORMER (without pre-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "training) obtains state-of-the-art on code summarization, outperforming all baselines, including the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 392 + ], + "score": 1.0, + "content": "previous state-of-the-art on Java-small proposed by Fernandes et al. (2019). Further, pre-training", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 388, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 402 + ], + "score": 1.0, + "content": "on Java-medium and Java-large on the permutation language modeling objective (Yang et al., 2019)", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 413 + ], + "score": 1.0, + "content": "substantially improves precision, recall, and F1 score after fine-tuning on Java-small. To avoid leak-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 411, + 495, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 495, + 425 + ], + "score": 1.0, + "content": "age, we exclude the projects used in the validation and test splits of Java-small from pre-training.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 344, + 506, + 425 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 504, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "Ablation study. We further perform ablations where we remove our model’s access to the Context", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 446, + 263, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 263, + 457 + ], + "score": 1.0, + "content": "or Structure, also presented in Table 3.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 434, + 505, + 457 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 295, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 296, + 474 + ], + "score": 1.0, + "content": "With pointer network. We find that both ab-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 473, + 297, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 297, + 485 + ], + "score": 1.0, + "content": "lations lead to a substantial drop in perfor-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 484, + 297, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 297, + 497 + ], + "score": 1.0, + "content": "mance, highlighting the benefit of learning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 495, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 297, + 507 + ], + "score": 1.0, + "content": "jointly from Structure and Context. Interest-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 506, + 297, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 297, + 518 + ], + "score": 1.0, + "content": "ingly, the model without access to the Struc-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 518, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 297, + 529 + ], + "score": 1.0, + "content": "ture performs slightly better than the variant", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 528, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 297, + 539 + ], + "score": 1.0, + "content": "without Context. Note that our model with-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 538, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 297, + 552 + ], + "score": 1.0, + "content": "out Structure is related to the XLNet (Yang", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 550, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 297, + 562 + ], + "score": 1.0, + "content": "et al., 2019) model, where we add a pointer net-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 561, + 297, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 297, + 573 + ], + "score": 1.0, + "content": "work to the decoder and concatenate the token", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 572, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 297, + 585 + ], + "score": 1.0, + "content": "types to their respective input tokens (see Ap-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 583, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 297, + 594 + ], + "score": 1.0, + "content": "pendix A.4). Without pointer network. We re-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 595, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 297, + 606 + ], + "score": 1.0, + "content": "peat the ablation on the variants without pointer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 605, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 297, + 617 + ], + "score": 1.0, + "content": "network. Here, the variant without Context per-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 615, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 297, + 628 + ], + "score": 1.0, + "content": "forms better than the variant without Structure,", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 462, + 297, + 628 + ] + }, + { + "type": "table", + "bbox": [ + 315, + 462, + 496, + 603 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 315, + 462, + 496, + 603 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 315, + 462, + 496, + 603 + ], + "spans": [ + { + "bbox": [ + 315, + 462, + 496, + 603 + ], + "score": 0.98, + "html": "
ModelPrec.Rec.F1
Without pointer net code2seq Ours w/o structure Ours w/o context51.23 50.70 51.8137.31 45.49 46.0443.18 47.96
Ours With pointer net50.3346.8048.75 48.50
Fernandes et al. (2019)=151.4
GREAT Ours w/o structure53.6046.4149.75
Ours w/o context55.4846.0750.34
54.4545.2949.45
Ours54.8549.8452.22
Ours + Pretrain57.0250.8753.77
", + "type": "table", + "image_path": "93c3db634ac78ceb6a3e42a79f6c881109ae26412d90ebf329ef70021d85d788.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 315, + 462, + 496, + 474.8181818181818 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 315, + 474.8181818181818, + 496, + 487.6363636363636 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 315, + 487.6363636363636, + 496, + 500.45454545454544 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 315, + 500.45454545454544, + 496, + 513.2727272727273 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 315, + 513.2727272727273, + 496, + 526.0909090909091 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 315, + 526.0909090909091, + 496, + 538.909090909091 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 315, + 538.909090909091, + 496, + 551.7272727272729 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 315, + 551.7272727272729, + 496, + 564.5454545454547 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 315, + 564.5454545454547, + 496, + 577.3636363636366 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 315, + 577.3636363636366, + 496, + 590.1818181818185 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 315, + 590.1818181818185, + 496, + 603.0000000000003 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 304, + 612, + 504, + 622 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 303, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 303, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "Table 3: Results on Java-small and ablation study.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + } + ], + "index": 45.0 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "indicating that the pointer network helps to compensate for the lack of access to Structure. The", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 107, + 639, + 504, + 649 + ], + "spans": [ + { + "bbox": [ + 107, + 639, + 504, + 649 + ], + "score": 1.0, + "content": "Structure-only variant (w/o pointer net) of our model even outperforms the full variant in this sce-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "nario. Inspection of the results revealed that the Structure-only variant has better performance on", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 660, + 489, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 489, + 671 + ], + "score": 1.0, + "content": "longer method names, which have an outsize influence on the micro-F1 score used in this work.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50.5, + "bbox_fs": [ + 105, + 627, + 505, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 688 + ], + "score": 1.0, + "content": "Ablation of the AST-based distances. In Table 4 we compare the performance of our model when", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "trained with each of the four different AST distances (sibling shortest paths, ancestor shortest paths,", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "shortest paths, personalized PageRank; see Section 3.1). Here, the model is trained on Java-small", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "in the Structure-only setting and without pointer network. For reference, we also show the results", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "of training our model using all four AST distance functions (c.f. Table 3). We find that, while the", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "personalized PageRank distance performs best on its own, each of the individual distances on their", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "own performs substantially worse than their combination, highlighting the usefulness of combining", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 370, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 370, + 117 + ], + "score": 1.0, + "content": "the distances in our model as well as their complementary nature.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 55, + "bbox_fs": [ + 105, + 678, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "personalized PageRank distance performs best on its own, each of the individual distances on their", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "own performs substantially worse than their combination, highlighting the usefulness of combining", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 370, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 370, + 117 + ], + "score": 1.0, + "content": "the distances in our model as well as their complementary nature.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 129, + 300, + 140 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 302, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 302, + 141 + ], + "score": 1.0, + "content": "5.2 MULTILINGUAL CODE SUMMARIZATION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 338, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 339, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 339, + 161 + ], + "score": 1.0, + "content": "Setup. A key contribution of our proposed architecture", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 160, + 339, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 339, + 173 + ], + "score": 1.0, + "content": "is that it only uses language-agnostic features, i.e. the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 339, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 339, + 183 + ], + "score": 1.0, + "content": "source code and features that can be directly computed", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 339, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 339, + 194 + ], + "score": 1.0, + "content": "from the AST. We use this fact to study the first multi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 339, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 339, + 205 + ], + "score": 1.0, + "content": "language code summarization model. We train our model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 340, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 340, + 216 + ], + "score": 1.0, + "content": "jointly on Python, Javascript, Ruby, and Go. The shared", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 339, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 339, + 227 + ], + "score": 1.0, + "content": "sub-token vocabulary is the union of the individual vocab-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 340, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 340, + 239 + ], + "score": 1.0, + "content": "ularies, enabling us to evaluate the multi-language model", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5 + }, + { + "type": "table", + "bbox": [ + 354, + 146, + 496, + 212 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 354, + 146, + 496, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 354, + 146, + 496, + 212 + ], + "spans": [ + { + "bbox": [ + 354, + 146, + 496, + 212 + ], + "score": 0.971, + "html": "
AST distanceF1 score
Sibling shortest paths46.17
Ancestor shortest paths47.89
Shortest paths47.76
Personalized PageRank48.47
All the above (c.f. Table 3)48.75
", + "type": "table", + "image_path": "44d93ce3802abbb255784a0d5f66ef881ac7c557a4abf8c4d5752cc699deb37d.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 354, + 146, + 496, + 179.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 354, + 179.0, + 496, + 212.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 351, + 221, + 501, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 349, + 218, + 501, + 234 + ], + "spans": [ + { + "bbox": [ + 349, + 218, + 501, + 234 + ], + "score": 1.0, + "content": "Table 4: AST distance ablation study.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 238, + 505, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "on the individual languages and compare with the single-language models. As proposed by Conneau", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 246, + 439, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 439, + 262 + ], + "score": 1.0, + "content": "& Lample (2019), we add a learned language embedding to each input embedding.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "Results. In the lower part of Table 2 we can see the results of training our CODE TRANSFORMER", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "jointly on all four programming languages. Our multi-lingual variants substantially outperform", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "the mono-lingual models on all languages. The strongest improvement is on Ruby, which is also", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "the programming language with the smallest number of samples in the dataset. Fine-tuning on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "the individual languages after joint training on code summarization only has a marginal effect on", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 325, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 340 + ], + "score": 1.0, + "content": "performance, indicating that the multilingual objective is well-aligned with the individual languages.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "In the last row, we have a variant of our model where we pre-train on the multi-lingual masked", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "language modeling task, followed by finetuning on code summarization on the individual languages.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "Further, we observe that similar to the results on Java-small, removing the pointer network generally", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "score": 1.0, + "content": "leads to weaker performance. One notable exception is Go, where the variant without the pointer", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "network performs better in terms of F1 score. Our investigation revealed that there seems to be some", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "violation of the i.i.d. assumption in the split provided by the creators of the dataset. In Figure 7 we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "show that in the test partition of the Go dataset, the share of tokens from the labels also occurring in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "the methods’ bodies – exactly the scenario where the pointer network can improve predictions – is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 431, + 357, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 357, + 443 + ], + "score": 1.0, + "content": "substantially lower compared to the train/validation partitions.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 448, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "Remarkably, the multi-language Context-only variant (i.e. without access to the Structure) performs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "substantially worse than the full multi-language variant. This highlights that Structure is crucial to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "exploit the commonalities of different programming languages. Also notably, the GREAT baseline’s", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "results also improve substantially when trained in the multi-language setting, though it is still outper-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 505 + ], + "score": 1.0, + "content": "formed by our model. However, our results indicate that any representation learning model for code", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 502, + 495, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 495, + 516 + ], + "score": 1.0, + "content": "can benefit from multi-language training, especially when evaluating on low-resource languages.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 108, + 519, + 504, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "In Table 16 we present results using the sample-F1 score. At the time of submission, our mono-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 422, + 543 + ], + "score": 1.0, + "content": "lingual model on Python outperforms the state of the art on the ogbg-code", + "type": "text" + }, + { + "bbox": [ + 422, + 530, + 433, + 541 + ], + "score": 0.67, + "content": "2 ^ { 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "(Hu et al., 2020)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 542, + 503, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 169, + 554 + ], + "score": 1.0, + "content": "leaderboard by", + "type": "text" + }, + { + "bbox": [ + 169, + 542, + 193, + 552 + ], + "score": 0.86, + "content": "112 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 542, + 474, + 554 + ], + "score": 1.0, + "content": ", and our multilanguage variant with LM pretraining outperforms it by", + "type": "text" + }, + { + "bbox": [ + 475, + 542, + 499, + 552 + ], + "score": 0.9, + "content": "122 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 542, + 503, + 554 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "Qualitative analysis of multilingual representations. Learning the CODE TRANSFORMER on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "multiple programming languages jointly provides us with embeddings in a shared representation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "space. In Fig. 4 we show a t-SNE (Maaten & Hinton, 2008) visualization of the ca. 40,000 snippets", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 596, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 612 + ], + "score": 1.0, + "content": "from the validation sets of four programming languages from the CSN dataset. For the embedding", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "of a snippet, we use the representation of the method name in the final layer of the encoder. Note", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "that the true method names are masked, i.e., inaccessible to the model. Further, note that in contrast", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "score": 1.0, + "content": "to the monolingual embeddings learned by Kanade et al. (2020), the embeddings we evaluate are", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "score": 1.0, + "content": "learned on the task of code summarization (though a similar study could be performed by using our", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "score": 1.0, + "content": "model that was trained on the traditional language modeling pretraining task on multiple languages).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "While snippets from the same language tend to be grouped together, there are interesting intersec-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "tions of the different programming languages. For example, we highlight all methods whose names", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "start with the subtoken parse or main. We see that snippets starting with parse are predom-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 703, + 505, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 715 + ], + "score": 1.0, + "content": "inantly in an intersection region of Python and Javascript. From these snippets, we display the", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 51.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 117, + 722, + 399, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 400, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 400, + 734 + ], + "score": 1.0, + "content": "4https://ogb.stanford.edu/docs/graphprop/#ogbg-code2", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 82, + 506, + 117 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 129, + 300, + 140 + ], + "lines": [ + { + "bbox": [ + 106, + 128, + 302, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 302, + 141 + ], + "score": 1.0, + "content": "5.2 MULTILINGUAL CODE SUMMARIZATION", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 149, + 338, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 339, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 339, + 161 + ], + "score": 1.0, + "content": "Setup. A key contribution of our proposed architecture", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 160, + 339, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 339, + 173 + ], + "score": 1.0, + "content": "is that it only uses language-agnostic features, i.e. the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 339, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 339, + 183 + ], + "score": 1.0, + "content": "source code and features that can be directly computed", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 339, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 339, + 194 + ], + "score": 1.0, + "content": "from the AST. We use this fact to study the first multi-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 194, + 339, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 339, + 205 + ], + "score": 1.0, + "content": "language code summarization model. We train our model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 340, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 340, + 216 + ], + "score": 1.0, + "content": "jointly on Python, Javascript, Ruby, and Go. The shared", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 215, + 339, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 339, + 227 + ], + "score": 1.0, + "content": "sub-token vocabulary is the union of the individual vocab-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 227, + 340, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 340, + 239 + ], + "score": 1.0, + "content": "ularies, enabling us to evaluate the multi-language model", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "on the individual languages and compare with the single-language models. As proposed by Conneau", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 246, + 439, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 439, + 262 + ], + "score": 1.0, + "content": "& Lample (2019), we add a learned language embedding to each input embedding.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 149, + 340, + 239 + ] + }, + { + "type": "table", + "bbox": [ + 354, + 146, + 496, + 212 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 354, + 146, + 496, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 354, + 146, + 496, + 212 + ], + "spans": [ + { + "bbox": [ + 354, + 146, + 496, + 212 + ], + "score": 0.971, + "html": "
AST distanceF1 score
Sibling shortest paths46.17
Ancestor shortest paths47.89
Shortest paths47.76
Personalized PageRank48.47
All the above (c.f. Table 3)48.75
", + "type": "table", + "image_path": "44d93ce3802abbb255784a0d5f66ef881ac7c557a4abf8c4d5752cc699deb37d.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 354, + 146, + 496, + 179.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 354, + 179.0, + 496, + 212.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 351, + 221, + 501, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 349, + 218, + 501, + 234 + ], + "spans": [ + { + "bbox": [ + 349, + 218, + 501, + 234 + ], + "score": 1.0, + "content": "Table 4: AST distance ablation study.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 238, + 505, + 259 + ], + "lines": [], + "index": 15.5, + "bbox_fs": [ + 105, + 236, + 505, + 262 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "Results. In the lower part of Table 2 we can see the results of training our CODE TRANSFORMER", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "jointly on all four programming languages. Our multi-lingual variants substantially outperform", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "the mono-lingual models on all languages. The strongest improvement is on Ruby, which is also", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "score": 1.0, + "content": "the programming language with the smallest number of samples in the dataset. Fine-tuning on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "the individual languages after joint training on code summarization only has a marginal effect on", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 325, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 340 + ], + "score": 1.0, + "content": "performance, indicating that the multilingual objective is well-aligned with the individual languages.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 351 + ], + "score": 1.0, + "content": "In the last row, we have a variant of our model where we pre-train on the multi-lingual masked", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 362 + ], + "score": 1.0, + "content": "language modeling task, followed by finetuning on code summarization on the individual languages.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 271, + 506, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "Further, we observe that similar to the results on Java-small, removing the pointer network generally", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "score": 1.0, + "content": "leads to weaker performance. One notable exception is Go, where the variant without the pointer", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "network performs better in terms of F1 score. Our investigation revealed that there seems to be some", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 411 + ], + "score": 1.0, + "content": "violation of the i.i.d. assumption in the split provided by the creators of the dataset. In Figure 7 we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "show that in the test partition of the Go dataset, the share of tokens from the labels also occurring in", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "the methods’ bodies – exactly the scenario where the pointer network can improve predictions – is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 431, + 357, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 357, + 443 + ], + "score": 1.0, + "content": "substantially lower compared to the train/validation partitions.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 364, + 506, + 443 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 448, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "Remarkably, the multi-language Context-only variant (i.e. without access to the Structure) performs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "substantially worse than the full multi-language variant. This highlights that Structure is crucial to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "exploit the commonalities of different programming languages. Also notably, the GREAT baseline’s", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 493 + ], + "score": 1.0, + "content": "results also improve substantially when trained in the multi-language setting, though it is still outper-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 505 + ], + "score": 1.0, + "content": "formed by our model. However, our results indicate that any representation learning model for code", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 502, + 495, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 495, + 516 + ], + "score": 1.0, + "content": "can benefit from multi-language training, especially when evaluating on low-resource languages.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 448, + 506, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 519, + 504, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "In Table 16 we present results using the sample-F1 score. At the time of submission, our mono-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 422, + 543 + ], + "score": 1.0, + "content": "lingual model on Python outperforms the state of the art on the ogbg-code", + "type": "text" + }, + { + "bbox": [ + 422, + 530, + 433, + 541 + ], + "score": 0.67, + "content": "2 ^ { 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "(Hu et al., 2020)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 542, + 503, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 169, + 554 + ], + "score": 1.0, + "content": "leaderboard by", + "type": "text" + }, + { + "bbox": [ + 169, + 542, + 193, + 552 + ], + "score": 0.86, + "content": "112 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 542, + 474, + 554 + ], + "score": 1.0, + "content": ", and our multilanguage variant with LM pretraining outperforms it by", + "type": "text" + }, + { + "bbox": [ + 475, + 542, + 499, + 552 + ], + "score": 0.9, + "content": "122 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 542, + 503, + 554 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 520, + 506, + 554 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "Qualitative analysis of multilingual representations. Learning the CODE TRANSFORMER on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 505, + 588 + ], + "score": 1.0, + "content": "multiple programming languages jointly provides us with embeddings in a shared representation", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 600 + ], + "score": 1.0, + "content": "space. In Fig. 4 we show a t-SNE (Maaten & Hinton, 2008) visualization of the ca. 40,000 snippets", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 596, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 612 + ], + "score": 1.0, + "content": "from the validation sets of four programming languages from the CSN dataset. For the embedding", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "of a snippet, we use the representation of the method name in the final layer of the encoder. Note", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 632 + ], + "score": 1.0, + "content": "that the true method names are masked, i.e., inaccessible to the model. Further, note that in contrast", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 506, + 644 + ], + "score": 1.0, + "content": "to the monolingual embeddings learned by Kanade et al. (2020), the embeddings we evaluate are", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "score": 1.0, + "content": "learned on the task of code summarization (though a similar study could be performed by using our", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "score": 1.0, + "content": "model that was trained on the traditional language modeling pretraining task on multiple languages).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 45, + "bbox_fs": [ + 104, + 564, + 506, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 669, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "While snippets from the same language tend to be grouped together, there are interesting intersec-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 506, + 694 + ], + "score": 1.0, + "content": "tions of the different programming languages. For example, we highlight all methods whose names", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 692, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 505, + 704 + ], + "score": 1.0, + "content": "start with the subtoken parse or main. We see that snippets starting with parse are predom-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 703, + 505, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 715 + ], + "score": 1.0, + "content": "inantly in an intersection region of Python and Javascript. From these snippets, we display the", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "cross-language pair with smallest Euclidean embedding distance in Fig. 5. Remarkably, both snip-", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "pets are effectively the same method in Javascript and Python – it is worth reminding that the model", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "score": 1.0, + "content": "has never seen any parallel data during training. On the other hand, snippets starting with main tend", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "score": 1.0, + "content": "to lie at an intersectional region of Python, Javascript, and Go. In Table 6 in the appendix we show", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "additional cross-lingual pairs with similar embeddings, including a failure case of a main function,", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "where embedding distance is not representative of semantic similarity. We attribute this to the fact", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "that we used the encoder output embedding of the masked method name – the representation used", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "by the decoder to predict the method name – as a snippet’s representation. Thus, snippets with com-", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "pletely different semantics (as is to be expected for very generic method names starting with main)", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 416, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 416, + 534 + ], + "score": 1.0, + "content": "have similar representations because they are predictive of the method name.", + "type": "text", + "cross_page": true + } + ], + "index": 20 + } + ], + "index": 51.5, + "bbox_fs": [ + 106, + 669, + 506, + 715 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 64, + 474, + 241 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 137, + 64, + 474, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 64, + 474, + 241 + ], + "spans": [ + { + "bbox": [ + 137, + 64, + 474, + 241 + ], + "score": 0.972, + "type": "image", + "image_path": "e875f6f92fb5baf876c7eada99cff2f2f6a1ae4c2907de273c6e3de07dd8a6a8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 137, + 64, + 474, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 137, + 123.0, + 474, + 182.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 137, + 182.0, + 474, + 241.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 247, + 502, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 246, + 504, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 504, + 261 + ], + "score": 1.0, + "content": "Figure 4: t-SNE visualization of the CODE TRANSFORMER’s learned multilingual representations.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 125, + 270, + 475, + 349 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 270, + 475, + 349 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 125, + 270, + 475, + 348 + ], + "spans": [ + { + "bbox": [ + 125, + 270, + 475, + 348 + ], + "score": 0.771, + "type": "image", + "image_path": "b4f73ecaf41d7a935e9b22345a3d4464c58d6b6b17f19a5180425db1c0483a08.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 125, + 270, + 475, + 296.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 125, + 296.3333333333333, + 475, + 322.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 125, + 322.66666666666663, + 475, + 348.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 357, + 505, + 401 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "Figure 5: Example snippet starting with parse (left) and its best embedding match from other", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "languages (right). Both methods parse an input string to convert it into a boolean value. Note", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "that even though they are semantically very similar, their method names are not; nonetheless, their", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 390, + 452, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 452, + 403 + ], + "score": 1.0, + "content": "representations in the CODE TRANSFORMER encoder reflect their semantic similarity.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ], + "index": 6.75 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 505, + 437 + ], + "score": 1.0, + "content": "cross-language pair with smallest Euclidean embedding distance in Fig. 5. Remarkably, both snip-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 434, + 506, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 506, + 447 + ], + "score": 1.0, + "content": "pets are effectively the same method in Javascript and Python – it is worth reminding that the model", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 505, + 458 + ], + "score": 1.0, + "content": "has never seen any parallel data during training. On the other hand, snippets starting with main tend", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 468 + ], + "score": 1.0, + "content": "to lie at an intersectional region of Python, Javascript, and Go. In Table 6 in the appendix we show", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "additional cross-lingual pairs with similar embeddings, including a failure case of a main function,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "where embedding distance is not representative of semantic similarity. We attribute this to the fact", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "that we used the encoder output embedding of the masked method name – the representation used", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "by the decoder to predict the method name – as a snippet’s representation. Thus, snippets with com-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "pletely different semantics (as is to be expected for very generic method names starting with main)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 522, + 416, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 416, + 534 + ], + "score": 1.0, + "content": "have similar representations because they are predictive of the method name.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 108, + 538, + 504, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "As another qualitative insight into the representations learned by the CODE TRANSFORMER we have", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "found that the language embeddings of languages with similar roots in language design are close;", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 560, + 501, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 501, + 574 + ], + "score": 1.0, + "content": "see Table 5 in the appendix for the pairwise similarity matrix of the learned language embeddings.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 107, + 588, + 195, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 197, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 197, + 604 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "We present the CODE TRANSFORMER, which learns jointly from Structure and Context of programs", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "while only relying on language-agnostic features. Our model obtains state-of-the-art performance", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "on code summarization on five different programming languages. Besides these results for train-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "ing on individual languages, the language-agnostic nature of our model allows us to train it jointly", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "on multiple programming languages. The resulting multilingual model substantially outperforms", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "its mono-lingual variant on all programming languages, setting the state of the art on each lan-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 678, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 506, + 693 + ], + "score": 1.0, + "content": "guage. We observe the largest improvement from multilingual training on the language with fewest", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "score": 1.0, + "content": "resources, indicating that multilingual training can improve learning for less widely used program-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "ming languages. Remarkably, multilingual training only from Context does not lead to the same", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 712, + 416, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 416, + 725 + ], + "score": 1.0, + "content": "improvements, highlighting the benefits of combining Structure and Context.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 29.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 64, + 474, + 241 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 137, + 64, + 474, + 241 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 64, + 474, + 241 + ], + "spans": [ + { + "bbox": [ + 137, + 64, + 474, + 241 + ], + "score": 0.972, + "type": "image", + "image_path": "e875f6f92fb5baf876c7eada99cff2f2f6a1ae4c2907de273c6e3de07dd8a6a8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 137, + 64, + 474, + 123.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 137, + 123.0, + 474, + 182.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 137, + 182.0, + 474, + 241.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 247, + 502, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 246, + 504, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 504, + 261 + ], + "score": 1.0, + "content": "Figure 4: t-SNE visualization of the CODE TRANSFORMER’s learned multilingual representations.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 125, + 270, + 475, + 349 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 270, + 475, + 349 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 125, + 270, + 475, + 348 + ], + "spans": [ + { + "bbox": [ + 125, + 270, + 475, + 348 + ], + "score": 0.771, + "type": "image", + "image_path": "b4f73ecaf41d7a935e9b22345a3d4464c58d6b6b17f19a5180425db1c0483a08.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 125, + 270, + 475, + 296.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 125, + 296.3333333333333, + 475, + 322.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 125, + 322.66666666666663, + 475, + 348.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 357, + 505, + 401 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 370 + ], + "score": 1.0, + "content": "Figure 5: Example snippet starting with parse (left) and its best embedding match from other", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 367, + 506, + 381 + ], + "score": 1.0, + "content": "languages (right). Both methods parse an input string to convert it into a boolean value. Note", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "that even though they are semantically very similar, their method names are not; nonetheless, their", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 390, + 452, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 452, + 403 + ], + "score": 1.0, + "content": "representations in the CODE TRANSFORMER encoder reflect their semantic similarity.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + } + ], + "index": 6.75 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 533 + ], + "lines": [], + "index": 15.5, + "bbox_fs": [ + 104, + 422, + 506, + 534 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 538, + 504, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "As another qualitative insight into the representations learned by the CODE TRANSFORMER we have", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "found that the language embeddings of languages with similar roots in language design are close;", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 560, + 501, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 501, + 574 + ], + "score": 1.0, + "content": "see Table 5 in the appendix for the pairwise similarity matrix of the learned language embeddings.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 537, + 505, + 574 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 588, + 195, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 197, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 197, + 604 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 613, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 626 + ], + "score": 1.0, + "content": "We present the CODE TRANSFORMER, which learns jointly from Structure and Context of programs", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "while only relying on language-agnostic features. Our model obtains state-of-the-art performance", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "on code summarization on five different programming languages. Besides these results for train-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "ing on individual languages, the language-agnostic nature of our model allows us to train it jointly", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 670 + ], + "score": 1.0, + "content": "on multiple programming languages. The resulting multilingual model substantially outperforms", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "its mono-lingual variant on all programming languages, setting the state of the art on each lan-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 678, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 506, + 693 + ], + "score": 1.0, + "content": "guage. We observe the largest improvement from multilingual training on the language with fewest", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "score": 1.0, + "content": "resources, indicating that multilingual training can improve learning for less widely used program-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 701, + 505, + 714 + ], + "score": 1.0, + "content": "ming languages. Remarkably, multilingual training only from Context does not lead to the same", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 712, + 416, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 712, + 416, + 725 + ], + "score": 1.0, + "content": "improvements, highlighting the benefits of combining Structure and Context.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 613, + 506, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 166 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Mithun Acharya, Tao Xie, Jian Pei, and Jun Xu. Mining api patterns as partial orders from source", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 111, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 117, + 111, + 504, + 123 + ], + "score": 1.0, + "content": "code: From usage scenarios to specifications. In Proceedings of the the 6th Joint Meeting of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "European Software Engineering Conference and the ACM SIGSOFT Symposium on The Founda-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "tions of Software Engineering, ESEC-FSE ’07, pp. 25–34, New York, NY, USA, 2007. Associ-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "ation for Computing Machinery. ISBN 9781595938114. doi: 10.1145/1287624.1287630. URL", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 155, + 354, + 167 + ], + "spans": [ + { + "bbox": [ + 116, + 155, + 354, + 167 + ], + "score": 1.0, + "content": "https://doi.org/10.1145/1287624.1287630.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 505, + 229 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 188 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Learning natural coding", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "conventions. In Proceedings of the 22nd ACM SIGSOFT International Symposium on Founda-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "tions of Software Engineering, FSE 2014, pp. 281–293, New York, NY, USA, 2014. Associa-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 506, + 219 + ], + "score": 1.0, + "content": "tion for Computing Machinery. ISBN 9781450330565. doi: 10.1145/2635868.2635883. URL", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 217, + 355, + 230 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 355, + 230 + ], + "score": 1.0, + "content": "https://doi.org/10.1145/2635868.2635883.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Suggesting accurate method", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "and class names. In Proceedings of the 2015 10th Joint Meeting on Foundations of Software", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 256, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 115, + 256, + 506, + 273 + ], + "score": 1.0, + "content": "Engineering, ESEC/FSE 2015, pp. 38–49, New York, NY, USA, 2015. Association for Computing", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 268, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 115, + 268, + 504, + 282 + ], + "score": 1.0, + "content": "Machinery. ISBN 9781450336758. doi: 10.1145/2786805.2786849. URL https://doi.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 280, + 284, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 284, + 293 + ], + "score": 1.0, + "content": "org/10.1145/2786805.2786849.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 503, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Hao Peng, and Charles Sutton. A convolutional attention network for extreme", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "summarization of source code. In International conference on machine learning, pp. 2091–2100,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 319, + 143, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 319, + 143, + 333 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 353 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 351, + 504, + 363 + ], + "spans": [ + { + "bbox": [ + 116, + 351, + 504, + 363 + ], + "score": 1.0, + "content": "with graphs. In International Conference on Learning Representations, 2018. URL https:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 362, + 332, + 374 + ], + "spans": [ + { + "bbox": [ + 117, + 362, + 266, + 374 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 363, + 273, + 371 + ], + "score": 0.52, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 362, + 332, + 374 + ], + "score": 1.0, + "content": "BJOFETxR-.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "score": 1.0, + "content": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "tion for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Confer-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "ence on Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 116, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450356985. doi:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 424, + 491, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 491, + 437 + ], + "score": 1.0, + "content": "10.1145/3192366.3192412. URL https://doi.org/10.1145/3192366.3192412.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 457 + ], + "score": 1.0, + "content": "Uri Alon, Shaked Brody, Omer Levy, and Eran Yahav. code2seq: Generating sequences from struc-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "tured representations of code. In International Conference on Learning Representations, 2019a.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 464, + 396, + 479 + ], + "spans": [ + { + "bbox": [ + 115, + 464, + 325, + 479 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 466, + 331, + 474 + ], + "score": 0.47, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 464, + 396, + 479 + ], + "score": 1.0, + "content": "H1gKYo09tX.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 504, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. code2vec: Learning distributed repre-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 494, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 115, + 494, + 505, + 508 + ], + "score": 1.0, + "content": "sentations of code. Proceedings of the ACM on Programming Languages, 3(POPL):1–29, 2019b.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 525 + ], + "score": 1.0, + "content": "Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 535, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 114, + 535, + 506, + 549 + ], + "score": 1.0, + "content": "Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 545, + 143, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 143, + 559 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "Avishkar Bhoopchand, Tim Rocktaschel, Earl Barr, and Sebastian Riedel. Learning python code ¨", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 577, + 448, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 448, + 588 + ], + "score": 1.0, + "content": "suggestion with a sparse pointer network. arXiv preprint arXiv:1611.08307, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "Benjamin Bichsel, Veselin Raychev, Petar Tsankov, and Martin Vechev. Statistical deobfuscation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "of android applications. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "Communications Security, CCS ’16, pp. 343–355, New York, NY, USA, 2016. Association for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 627, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 627, + 505, + 642 + ], + "score": 1.0, + "content": "Computing Machinery. ISBN 9781450341394. doi: 10.1145/2976749.2978422. URL https:", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 117, + 639, + 318, + 651 + ], + "spans": [ + { + "bbox": [ + 117, + 639, + 318, + 651 + ], + "score": 1.0, + "content": "//doi.org/10.1145/2976749.2978422.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 104, + 657, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "Pavol Bielik, Veselin Raychev, and Martin Vechev. Phog: probabilistic model for code. In Interna-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 669, + 371, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 371, + 681 + ], + "score": 1.0, + "content": "tional Conference on Machine Learning, pp. 2933–2942, 2016.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Aleksandar Bojchevski, Johannes Klicpera, Bryan Perozzi, Amol Kapoor, Martin Blais, Benedek", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Rozemberczki, Michal Lukasik, and Stephan G ´ unnemann. Scaling graph neural networks with ¨", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "approximate pagerank. In Proceedings of the 26th ACM SIGKDD International Conference on", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "Knowledge Discovery amp; Data Mining, KDD ’20, pp. 2464–2473, New York, NY, USA, 2020.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 166 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Mithun Acharya, Tao Xie, Jian Pei, and Jun Xu. Mining api patterns as partial orders from source", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 117, + 111, + 504, + 123 + ], + "spans": [ + { + "bbox": [ + 117, + 111, + 504, + 123 + ], + "score": 1.0, + "content": "code: From usage scenarios to specifications. In Proceedings of the the 6th Joint Meeting of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "European Software Engineering Conference and the ACM SIGSOFT Symposium on The Founda-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 132, + 505, + 146 + ], + "score": 1.0, + "content": "tions of Software Engineering, ESEC-FSE ’07, pp. 25–34, New York, NY, USA, 2007. Associ-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "ation for Computing Machinery. ISBN 9781595938114. doi: 10.1145/1287624.1287630. URL", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 155, + 354, + 167 + ], + "spans": [ + { + "bbox": [ + 116, + 155, + 354, + 167 + ], + "score": 1.0, + "content": "https://doi.org/10.1145/1287624.1287630.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 99, + 506, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 173, + 505, + 229 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 188 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Learning natural coding", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 185, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 185, + 505, + 198 + ], + "score": 1.0, + "content": "conventions. In Proceedings of the 22nd ACM SIGSOFT International Symposium on Founda-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "tions of Software Engineering, FSE 2014, pp. 281–293, New York, NY, USA, 2014. Associa-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 207, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 207, + 506, + 219 + ], + "score": 1.0, + "content": "tion for Computing Machinery. ISBN 9781450330565. doi: 10.1145/2635868.2635883. URL", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 217, + 355, + 230 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 355, + 230 + ], + "score": 1.0, + "content": "https://doi.org/10.1145/2635868.2635883.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 172, + 506, + 230 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 236, + 505, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Suggesting accurate method", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 247, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 116, + 247, + 505, + 261 + ], + "score": 1.0, + "content": "and class names. In Proceedings of the 2015 10th Joint Meeting on Foundations of Software", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 256, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 115, + 256, + 506, + 273 + ], + "score": 1.0, + "content": "Engineering, ESEC/FSE 2015, pp. 38–49, New York, NY, USA, 2015. Association for Computing", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 268, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 115, + 268, + 504, + 282 + ], + "score": 1.0, + "content": "Machinery. ISBN 9781450336758. doi: 10.1145/2786805.2786849. URL https://doi.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 280, + 284, + 293 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 284, + 293 + ], + "score": 1.0, + "content": "org/10.1145/2786805.2786849.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 236, + 506, + 293 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 503, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Hao Peng, and Charles Sutton. A convolutional attention network for extreme", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "summarization of source code. In International conference on machine learning, pp. 2091–2100,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 319, + 143, + 333 + ], + "spans": [ + { + "bbox": [ + 115, + 319, + 143, + 333 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 298, + 505, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 339, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 353 + ], + "score": 1.0, + "content": "Miltiadis Allamanis, Marc Brockschmidt, and Mahmoud Khademi. Learning to represent programs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 351, + 504, + 363 + ], + "spans": [ + { + "bbox": [ + 116, + 351, + 504, + 363 + ], + "score": 1.0, + "content": "with graphs. In International Conference on Learning Representations, 2018. URL https:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 362, + 332, + 374 + ], + "spans": [ + { + "bbox": [ + 117, + 362, + 266, + 374 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 363, + 273, + 371 + ], + "score": 0.52, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 362, + 332, + 374 + ], + "score": 1.0, + "content": "BJOFETxR-.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 338, + 505, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 504, + 393 + ], + "score": 1.0, + "content": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. A general path-based representa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "tion for predicting program properties. In Proceedings of the 39th ACM SIGPLAN Confer-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 115, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "ence on Programming Language Design and Implementation, PLDI 2018, pp. 404–419, New", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 116, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "York, NY, USA, 2018. Association for Computing Machinery. ISBN 9781450356985. doi:", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 424, + 491, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 424, + 491, + 437 + ], + "score": 1.0, + "content": "10.1145/3192366.3192412. URL https://doi.org/10.1145/3192366.3192412.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 381, + 505, + 437 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 505, + 457 + ], + "score": 1.0, + "content": "Uri Alon, Shaked Brody, Omer Levy, and Eran Yahav. code2seq: Generating sequences from struc-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 116, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "tured representations of code. In International Conference on Learning Representations, 2019a.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 464, + 396, + 479 + ], + "spans": [ + { + "bbox": [ + 115, + 464, + 325, + 479 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 466, + 331, + 474 + ], + "score": 0.47, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 464, + 396, + 479 + ], + "score": 1.0, + "content": "H1gKYo09tX.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 442, + 505, + 479 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 504, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "Uri Alon, Meital Zilberstein, Omer Levy, and Eran Yahav. code2vec: Learning distributed repre-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 494, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 115, + 494, + 505, + 508 + ], + "score": 1.0, + "content": "sentations of code. Proceedings of the ACM on Programming Languages, 3(POPL):1–29, 2019b.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 483, + 505, + 508 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 505, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 525 + ], + "score": 1.0, + "content": "Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 535, + 506, + 549 + ], + "spans": [ + { + "bbox": [ + 114, + 535, + 506, + 549 + ], + "score": 1.0, + "content": "Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 545, + 143, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 143, + 559 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 514, + 506, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "Avishkar Bhoopchand, Tim Rocktaschel, Earl Barr, and Sebastian Riedel. Learning python code ¨", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 577, + 448, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 448, + 588 + ], + "score": 1.0, + "content": "suggestion with a sparse pointer network. arXiv preprint arXiv:1611.08307, 2016.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 565, + 505, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 595, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "Benjamin Bichsel, Veselin Raychev, Petar Tsankov, and Martin Vechev. Statistical deobfuscation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 115, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "of android applications. In Proceedings of the 2016 ACM SIGSAC Conference on Computer and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "Communications Security, CCS ’16, pp. 343–355, New York, NY, USA, 2016. Association for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 627, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 627, + 505, + 642 + ], + "score": 1.0, + "content": "Computing Machinery. ISBN 9781450341394. doi: 10.1145/2976749.2978422. URL https:", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 117, + 639, + 318, + 651 + ], + "spans": [ + { + "bbox": [ + 117, + 639, + 318, + 651 + ], + "score": 1.0, + "content": "//doi.org/10.1145/2976749.2978422.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 595, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 657, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "Pavol Bielik, Veselin Raychev, and Martin Vechev. Phog: probabilistic model for code. In Interna-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 669, + 371, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 669, + 371, + 681 + ], + "score": 1.0, + "content": "tional Conference on Machine Learning, pp. 2933–2942, 2016.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 657, + 505, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "Aleksandar Bojchevski, Johannes Klicpera, Bryan Perozzi, Amol Kapoor, Martin Blais, Benedek", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 115, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "Rozemberczki, Michal Lukasik, and Stephan G ´ unnemann. Scaling graph neural networks with ¨", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "approximate pagerank. In Proceedings of the 26th ACM SIGKDD International Conference on", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "Knowledge Discovery amp; Data Mining, KDD ’20, pp. 2464–2473, New York, NY, USA, 2020.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5, + "bbox_fs": [ + 106, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 112, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 116, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 116, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Association for Computing Machinery. ISBN 9781450379984. doi: 10.1145/3394486.3403296.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 94, + 377, + 106 + ], + "spans": [ + { + "bbox": [ + 117, + 94, + 377, + 106 + ], + "score": 1.0, + "content": "URL https://doi.org/10.1145/3394486.3403296.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 113, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Dylan Bourgeois. Learning representations of source code from structure and context. 2019. URL", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 361, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 361, + 135 + ], + "score": 1.0, + "content": "http://infoscience.epfl.ch/record/277163.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 299, + 154 + ], + "score": 1.0, + "content": "Alexis Conneau and Guillaume Lample.", + "type": "text" + }, + { + "bbox": [ + 324, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "Cross-lingual language model pretrain-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 115, + 154, + 136, + 166 + ], + "score": 1.0, + "content": "ing.", + "type": "text" + }, + { + "bbox": [ + 147, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 7059–", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 174, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 115, + 174, + 291, + 188 + ], + "score": 1.0, + "content": "7069. Curran Associates, Inc., 2019.", + "type": "text" + }, + { + "bbox": [ + 304, + 174, + 505, + 189 + ], + "score": 1.0, + "content": "URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 417, + 200 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 417, + 200 + ], + "score": 1.0, + "content": "8928-cross-lingual-language-model-pretraining.pdf.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 205, + 503, + 229 + ], + "lines": [ + { + "bbox": [ + 106, + 206, + 504, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 504, + 217 + ], + "score": 1.0, + "content": "Hanjun Dai, Bo Dai, and Le Song. Discriminative embeddings of latent variable models for struc-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 216, + 451, + 230 + ], + "spans": [ + { + "bbox": [ + 115, + 216, + 451, + 230 + ], + "score": 1.0, + "content": "tured data. In International conference on machine learning, pp. 2702–2711, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 235, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 504, + 248 + ], + "score": 1.0, + "content": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "nov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 257, + 219, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 257, + 219, + 269 + ], + "score": 1.0, + "content": "arXiv:1901.02860, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "Hoa Khanh Dam, Truyen Tran, and Trang Pham. A deep language model for software code. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 116, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "FSE 2016: Proceedings of the Foundations Software Engineering International Symposium [The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 299, + 196, + 312 + ], + "spans": [ + { + "bbox": [ + 117, + 299, + 196, + 312 + ], + "score": 1.0, + "content": "Conference], 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 105, + 317, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 329, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 506, + 343 + ], + "score": 1.0, + "content": "graphs with fast localized spectral filtering. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 339, + 207, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 207, + 352 + ], + "score": 1.0, + "content": "pp. 3844–3852, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 108, + 358, + 503, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 358, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 372 + ], + "score": 1.0, + "content": "David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 380, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 479, + 394 + ], + "score": 1.0, + "content": "fingerprints. In Advances in neural information processing systems, pp. 2224–2232, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 108, + 399, + 502, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 504, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 504, + 412 + ], + "score": 1.0, + "content": "Patrick Fernandes, Miltiadis Allamanis, and Marc Brockschmidt. Structured neural summa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 411, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 116, + 411, + 504, + 424 + ], + "score": 1.0, + "content": "rization. In International Conference on Learning Representations, 2019. URL https:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 422, + 336, + 434 + ], + "spans": [ + { + "bbox": [ + 117, + 422, + 266, + 434 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 424, + 273, + 431 + ], + "score": 0.51, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 422, + 336, + 434 + ], + "score": 1.0, + "content": "H1ersoRqtm.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 504, + 453 + ], + "score": 1.0, + "content": "Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 116, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 463, + 223, + 474 + ], + "spans": [ + { + "bbox": [ + 114, + 463, + 223, + 474 + ], + "score": 1.0, + "content": "pp. 729–734. IEEE, 2005.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 503, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "score": 1.0, + "content": "William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 493, + 212, + 505 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 212, + 505 + ], + "score": 1.0, + "content": "graphs. In NIPS, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "Vincent J. Hellendoorn, Charles Sutton, Rishabh Singh, Petros Maniatis, and David Bieber. Global", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "relational models of source code. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 533, + 396, + 548 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 325, + 548 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 536, + 331, + 544 + ], + "score": 0.44, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 533, + 396, + 548 + ], + "score": 1.0, + "content": "B1lnbRNtwr.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 504, + 565 + ], + "score": 1.0, + "content": "Abram Hindle, Earl T. Barr, Zhendong Su, Mark Gabel, and Premkumar Devanbu. On the natural-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "ness of software. In Proceedings of the 34th International Conference on Software Engineering,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 117, + 576, + 385, + 588 + ], + "spans": [ + { + "bbox": [ + 117, + 576, + 385, + 588 + ], + "score": 1.0, + "content": "ICSE ’12, pp. 837–847. IEEE Press, 2012. ISBN 9781467310673.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 594, + 503, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 605, + 191, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 191, + 617 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 116, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 115, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "graphs. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin (eds.), Ad-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 658, + 504, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 504, + 669 + ], + "score": 1.0, + "content": "vances in Neural Information Processing Systems, volume 33, pp. 22118–22133. Curran As-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "sociates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/file/", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 678, + 373, + 692 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 373, + 692 + ], + "score": 1.0, + "content": "fb60d411a5c5b72b2e7d3527cfc84fd0-Paper.pdf.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 116, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "Codesearchnet challenge: Evaluating the state of semantic code search. arXiv preprint", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 220, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 220, + 732 + ], + "score": 1.0, + "content": "arXiv:1909.09436, 2019.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 112, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 116, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 116, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Association for Computing Machinery. ISBN 9781450379984. doi: 10.1145/3394486.3403296.", + "type": "text" + } + ], + "index": 0, + "is_list_end_line": true + }, + { + "bbox": [ + 117, + 94, + 377, + 106 + ], + "spans": [ + { + "bbox": [ + 117, + 94, + 377, + 106 + ], + "score": 1.0, + "content": "URL https://doi.org/10.1145/3394486.3403296.", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 0.5, + "bbox_fs": [ + 116, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 113, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 125 + ], + "score": 1.0, + "content": "Dylan Bourgeois. Learning representations of source code from structure and context. 2019. URL", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 361, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 361, + 135 + ], + "score": 1.0, + "content": "http://infoscience.epfl.ch/record/277163.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 112, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 142, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 299, + 154 + ], + "score": 1.0, + "content": "Alexis Conneau and Guillaume Lample.", + "type": "text" + }, + { + "bbox": [ + 324, + 142, + 505, + 155 + ], + "score": 1.0, + "content": "Cross-lingual language model pretrain-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 152, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 115, + 154, + 136, + 166 + ], + "score": 1.0, + "content": "ing.", + "type": "text" + }, + { + "bbox": [ + 147, + 152, + 505, + 166 + ], + "score": 1.0, + "content": "In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alche-Buc, E. Fox, and ´", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 506, + 177 + ], + "score": 1.0, + "content": "R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 7059–", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 174, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 115, + 174, + 291, + 188 + ], + "score": 1.0, + "content": "7069. Curran Associates, Inc., 2019.", + "type": "text" + }, + { + "bbox": [ + 304, + 174, + 505, + 189 + ], + "score": 1.0, + "content": "URL http://papers.nips.cc/paper/", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 417, + 200 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 417, + 200 + ], + "score": 1.0, + "content": "8928-cross-lingual-language-model-pretraining.pdf.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 142, + 506, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 205, + 503, + 229 + ], + "lines": [ + { + "bbox": [ + 106, + 206, + 504, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 504, + 217 + ], + "score": 1.0, + "content": "Hanjun Dai, Bo Dai, and Le Song. Discriminative embeddings of latent variable models for struc-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 216, + 451, + 230 + ], + "spans": [ + { + "bbox": [ + 115, + 216, + 451, + 230 + ], + "score": 1.0, + "content": "tured data. In International conference on machine learning, pp. 2702–2711, 2016.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 206, + 504, + 230 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 235, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 504, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 504, + 248 + ], + "score": 1.0, + "content": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "nov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 257, + 219, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 257, + 219, + 269 + ], + "score": 1.0, + "content": "arXiv:1901.02860, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 106, + 235, + 505, + 269 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "Hoa Khanh Dam, Truyen Tran, and Trang Pham. A deep language model for software code. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 288, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 116, + 288, + 505, + 300 + ], + "score": 1.0, + "content": "FSE 2016: Proceedings of the Foundations Software Engineering International Symposium [The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 299, + 196, + 312 + ], + "spans": [ + { + "bbox": [ + 117, + 299, + 196, + 312 + ], + "score": 1.0, + "content": "Conference], 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 276, + 505, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 317, + 504, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on ¨", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 329, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 506, + 343 + ], + "score": 1.0, + "content": "graphs with fast localized spectral filtering. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 339, + 207, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 207, + 352 + ], + "score": 1.0, + "content": "pp. 3844–3852, 2016.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 317, + 506, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 358, + 503, + 393 + ], + "lines": [ + { + "bbox": [ + 106, + 358, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 372 + ], + "score": 1.0, + "content": "David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bombarell, Timothy Hirzel, Alan´", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 116, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "Aspuru-Guzik, and Ryan P Adams. Convolutional networks on graphs for learning molecular", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 380, + 479, + 394 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 479, + 394 + ], + "score": 1.0, + "content": "fingerprints. In Advances in neural information processing systems, pp. 2224–2232, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 358, + 505, + 394 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 399, + 502, + 434 + ], + "lines": [ + { + "bbox": [ + 105, + 399, + 504, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 504, + 412 + ], + "score": 1.0, + "content": "Patrick Fernandes, Miltiadis Allamanis, and Marc Brockschmidt. Structured neural summa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 411, + 504, + 424 + ], + "spans": [ + { + "bbox": [ + 116, + 411, + 504, + 424 + ], + "score": 1.0, + "content": "rization. In International Conference on Learning Representations, 2019. URL https:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 422, + 336, + 434 + ], + "spans": [ + { + "bbox": [ + 117, + 422, + 266, + 434 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 424, + 273, + 431 + ], + "score": 0.51, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 422, + 336, + 434 + ], + "score": 1.0, + "content": "H1ersoRqtm.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 399, + 504, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 504, + 453 + ], + "score": 1.0, + "content": "Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 116, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "In Proceedings. 2005 IEEE International Joint Conference on Neural Networks, 2005., volume 2,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 463, + 223, + 474 + ], + "spans": [ + { + "bbox": [ + 114, + 463, + 223, + 474 + ], + "score": 1.0, + "content": "pp. 729–734. IEEE, 2005.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 440, + 505, + 474 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 503, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 496 + ], + "score": 1.0, + "content": "William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 493, + 212, + 505 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 212, + 505 + ], + "score": 1.0, + "content": "graphs. In NIPS, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 480, + 505, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "Vincent J. Hellendoorn, Charles Sutton, Rishabh Singh, Petros Maniatis, and David Bieber. Global", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "relational models of source code. In International Conference on Learning Representations, 2020.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 533, + 396, + 548 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 325, + 548 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 536, + 331, + 544 + ], + "score": 0.44, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 533, + 396, + 548 + ], + "score": 1.0, + "content": "B1lnbRNtwr.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 512, + 505, + 548 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 553, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 553, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 504, + 565 + ], + "score": 1.0, + "content": "Abram Hindle, Earl T. Barr, Zhendong Su, Mark Gabel, and Premkumar Devanbu. On the natural-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 564, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 505, + 578 + ], + "score": 1.0, + "content": "ness of software. In Proceedings of the 34th International Conference on Software Engineering,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 117, + 576, + 385, + 588 + ], + "spans": [ + { + "bbox": [ + 117, + 576, + 385, + 588 + ], + "score": 1.0, + "content": "ICSE ’12, pp. 837–847. IEEE Press, 2012. ISBN 9781467310673.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 553, + 505, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 594, + 503, + 617 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 608 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 605, + 191, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 191, + 617 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 594, + 505, + 617 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 116, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 115, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "graphs. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin (eds.), Ad-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 658, + 504, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 504, + 669 + ], + "score": 1.0, + "content": "vances in Neural Information Processing Systems, volume 33, pp. 22118–22133. Curran As-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "sociates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/file/", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 678, + 373, + 692 + ], + "spans": [ + { + "bbox": [ + 116, + 678, + 373, + 692 + ], + "score": 1.0, + "content": "fb60d411a5c5b72b2e7d3527cfc84fd0-Paper.pdf.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 624, + 506, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis Allamanis, and Marc Brockschmidt.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 116, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "Codesearchnet challenge: Evaluating the state of semantic code search. arXiv preprint", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 720, + 220, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 220, + 732 + ], + "score": 1.0, + "content": "arXiv:1909.09436, 2019.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Aditya Kanade, Petros Maniatis, Gogul Balakrishnan, and Kensen Shi. Learning and evaluating", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "contextual embedding of source code. In Proceedings of the 37th International Conference on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 220, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 220, + 117 + ], + "score": 1.0, + "content": "Machine Learning, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 104, + 123, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "score": 1.0, + "content": "Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 437, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 437, + 146 + ], + "score": 1.0, + "content": "works. In International Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "Johannes Klicpera, Aleksandar Bojchevski, and Stephan Gunnemann. Predict then propagate: ¨", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 505, + 177 + ], + "score": 1.0, + "content": "Graph neural networks meet personalized pagerank. In International Conference on Learning", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 245, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 245, + 187 + ], + "score": 1.0, + "content": "Representations (ICLR), 2019a.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 104, + 192, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 504, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 504, + 206 + ], + "score": 1.0, + "content": "Johannes Klicpera, Stefan Weißenberger, and Stephan Gunnemann. Diffusion improves graph learn-¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 443, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 443, + 216 + ], + "score": 1.0, + "content": "ing. In Conference on Neural Information Processing Systems (NeurIPS), 2019b.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Johannes Klicpera, Janek Groß, and Stephan Gunnemann. Directional message passing for molec- ¨", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 233, + 457, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 457, + 246 + ], + "score": 1.0, + "content": "ular graphs. In International Conference on Learning Representations (ICLR), 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 105, + 252, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "Jian Li, Yue Wang, Michael R Lyu, and Irwin King. Code completion with neural attention and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 264, + 352, + 275 + ], + "spans": [ + { + "bbox": [ + 115, + 264, + 352, + 275 + ], + "score": 1.0, + "content": "pointer networks. arXiv preprint arXiv:1711.09573, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 504, + 305 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 293, + 321, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 321, + 305 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1511.05493, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 105, + 311, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 322, + 298, + 335 + ], + "spans": [ + { + "bbox": [ + 116, + 322, + 298, + 335 + ], + "score": 1.0, + "content": "learning research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 105, + 341, + 503, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "Chris Maddison and Daniel Tarlow. Structured generative models of natural source code. In Inter-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 352, + 371, + 364 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 371, + 364 + ], + "score": 1.0, + "content": "national Conference on Machine Learning, pp. 649–657, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 504, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 504, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 504, + 384 + ], + "score": 1.0, + "content": "Lili Mou, Ge Li, Lu Zhang, Tao Wang, and Zhi Jin. Convolutional neural networks over tree struc-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "tures for programming language processing. In Proceedings of the Thirtieth AAAI Conference on", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 393, + 393, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 393, + 405 + ], + "score": 1.0, + "content": "Artificial Intelligence, AAAI’16, pp. 1287–1293. AAAI Press, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 425 + ], + "score": 1.0, + "content": "Trong Duc Nguyen, Anh Tuan Nguyen, Hung Dang Phan, and Tien N. Nguyen. Exploring api", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 423, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 423, + 506, + 435 + ], + "score": 1.0, + "content": "embedding for api usages and applications. In Proceedings of the 39th International Conference", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 434, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 434, + 505, + 445 + ], + "score": 1.0, + "content": "on Software Engineering, ICSE ’17, pp. 438–449. IEEE Press, 2017. ISBN 9781538638682. doi:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 444, + 457, + 456 + ], + "spans": [ + { + "bbox": [ + 117, + 444, + 457, + 456 + ], + "score": 1.0, + "content": "10.1109/ICSE.2017.47. URL https://doi.org/10.1109/ICSE.2017.47.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 477 + ], + "score": 1.0, + "content": "Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 474, + 394, + 487 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 394, + 487 + ], + "score": 1.0, + "content": "Bringing order to the web. Technical report, Stanford InfoLab, 1999.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 492, + 504, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 507 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 503, + 407, + 516 + ], + "spans": [ + { + "bbox": [ + 115, + 503, + 407, + 516 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "score": 1.0, + "content": "Veselin Raychev, Martin Vechev, and Eran Yahav. Code completion with statistical language models.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "In Proceedings of the 35th ACM SIGPLAN Conference on Programming Language Design and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 542, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 542, + 505, + 559 + ], + "score": 1.0, + "content": "Implementation, PLDI ’14, pp. 419–428, New York, NY, USA, 2014. Association for Computing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 554, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 504, + 568 + ], + "score": 1.0, + "content": "Machinery. ISBN 9781450327848. doi: 10.1145/2594291.2594321. URL https://doi.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 566, + 284, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 284, + 578 + ], + "score": 1.0, + "content": "org/10.1145/2594291.2594321.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 584, + 503, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 595, + 495, + 608 + ], + "spans": [ + { + "bbox": [ + 117, + 595, + 495, + 608 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 624, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 115, + 624, + 506, + 638 + ], + "score": 1.0, + "content": "Welling. Modeling relational data with graph convolutional networks. In Aldo Gangemi, Roberto", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Navigli, Maria-Esther Vidal, Pascal Hitzler, Raphael Troncy, Laura Hollink, Anna Tordai, and ¨", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 645, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 114, + 645, + 505, + 660 + ], + "score": 1.0, + "content": "Mehwish Alam (eds.), The Semantic Web, pp. 593–607, Cham, 2018. Springer International Pub-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 658, + 258, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 258, + 671 + ], + "score": 1.0, + "content": "lishing. ISBN 978-3-319-93417-4.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "tions. In Proceedings of the 2018 Conference of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 697, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 114, + 697, + 505, + 713 + ], + "score": 1.0, + "content": "for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pp.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "464–468, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. doi:", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 719, + 489, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 489, + 734 + ], + "score": 1.0, + "content": "10.18653/v1/N18-2074. URL https://www.aclweb.org/anthology/N18-2074.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Aditya Kanade, Petros Maniatis, Gogul Balakrishnan, and Kensen Shi. Learning and evaluating", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "contextual embedding of source code. In Proceedings of the 37th International Conference on", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 220, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 220, + 117 + ], + "score": 1.0, + "content": "Machine Learning, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 123, + 503, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 136 + ], + "score": 1.0, + "content": "Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 437, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 437, + 146 + ], + "score": 1.0, + "content": "works. In International Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 123, + 504, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "Johannes Klicpera, Aleksandar Bojchevski, and Stephan Gunnemann. Predict then propagate: ¨", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 505, + 177 + ], + "score": 1.0, + "content": "Graph neural networks meet personalized pagerank. In International Conference on Learning", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 245, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 245, + 187 + ], + "score": 1.0, + "content": "Representations (ICLR), 2019a.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 151, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 192, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 504, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 504, + 206 + ], + "score": 1.0, + "content": "Johannes Klicpera, Stefan Weißenberger, and Stephan Gunnemann. Diffusion improves graph learn-¨", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 443, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 443, + 216 + ], + "score": 1.0, + "content": "ing. In Conference on Neural Information Processing Systems (NeurIPS), 2019b.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 193, + 504, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Johannes Klicpera, Janek Groß, and Stephan Gunnemann. Directional message passing for molec- ¨", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 233, + 457, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 233, + 457, + 246 + ], + "score": 1.0, + "content": "ular graphs. In International Conference on Learning Representations (ICLR), 2020.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 223, + 505, + 246 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 252, + 504, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "Jian Li, Yue Wang, Michael R Lyu, and Irwin King. Code completion with neural attention and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 264, + 352, + 275 + ], + "spans": [ + { + "bbox": [ + 115, + 264, + 352, + 275 + ], + "score": 1.0, + "content": "pointer networks. arXiv preprint arXiv:1711.09573, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 252, + 505, + 275 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 281, + 504, + 305 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 294 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 293, + 321, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 321, + 305 + ], + "score": 1.0, + "content": "networks. arXiv preprint arXiv:1511.05493, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 281, + 505, + 305 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 311, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 322, + 298, + 335 + ], + "spans": [ + { + "bbox": [ + 116, + 322, + 298, + 335 + ], + "score": 1.0, + "content": "learning research, 9(Nov):2579–2605, 2008.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 311, + 505, + 335 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 341, + 503, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 505, + 354 + ], + "score": 1.0, + "content": "Chris Maddison and Daniel Tarlow. Structured generative models of natural source code. In Inter-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 352, + 371, + 364 + ], + "spans": [ + { + "bbox": [ + 115, + 352, + 371, + 364 + ], + "score": 1.0, + "content": "national Conference on Machine Learning, pp. 649–657, 2014.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 341, + 505, + 364 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 504, + 405 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 504, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 504, + 384 + ], + "score": 1.0, + "content": "Lili Mou, Ge Li, Lu Zhang, Tao Wang, and Zhi Jin. Convolutional neural networks over tree struc-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "tures for programming language processing. In Proceedings of the Thirtieth AAAI Conference on", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 393, + 393, + 405 + ], + "spans": [ + { + "bbox": [ + 115, + 393, + 393, + 405 + ], + "score": 1.0, + "content": "Artificial Intelligence, AAAI’16, pp. 1287–1293. AAAI Press, 2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 370, + 505, + 405 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 425 + ], + "score": 1.0, + "content": "Trong Duc Nguyen, Anh Tuan Nguyen, Hung Dang Phan, and Tien N. Nguyen. Exploring api", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 423, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 115, + 423, + 506, + 435 + ], + "score": 1.0, + "content": "embedding for api usages and applications. In Proceedings of the 39th International Conference", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 434, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 434, + 505, + 445 + ], + "score": 1.0, + "content": "on Software Engineering, ICSE ’17, pp. 438–449. IEEE Press, 2017. ISBN 9781538638682. doi:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 444, + 457, + 456 + ], + "spans": [ + { + "bbox": [ + 117, + 444, + 457, + 456 + ], + "score": 1.0, + "content": "10.1109/ICSE.2017.47. URL https://doi.org/10.1109/ICSE.2017.47.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 410, + 506, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 504, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 461, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 477 + ], + "score": 1.0, + "content": "Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking:", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 474, + 394, + 487 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 394, + 487 + ], + "score": 1.0, + "content": "Bringing order to the web. Technical report, Stanford InfoLab, 1999.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 461, + 505, + 487 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 492, + 504, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 490, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 507 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 503, + 407, + 516 + ], + "spans": [ + { + "bbox": [ + 115, + 503, + 407, + 516 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 490, + 505, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 504, + 534 + ], + "score": 1.0, + "content": "Veselin Raychev, Martin Vechev, and Eran Yahav. Code completion with statistical language models.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 546 + ], + "score": 1.0, + "content": "In Proceedings of the 35th ACM SIGPLAN Conference on Programming Language Design and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 542, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 115, + 542, + 505, + 559 + ], + "score": 1.0, + "content": "Implementation, PLDI ’14, pp. 419–428, New York, NY, USA, 2014. Association for Computing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 554, + 504, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 504, + 568 + ], + "score": 1.0, + "content": "Machinery. ISBN 9781450327848. doi: 10.1145/2594291.2594321. URL https://doi.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 566, + 284, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 566, + 284, + 578 + ], + "score": 1.0, + "content": "org/10.1145/2594291.2594321.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 522, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 584, + 503, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 595, + 495, + 608 + ], + "spans": [ + { + "bbox": [ + 117, + 595, + 495, + 608 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2008.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 584, + 505, + 608 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 505, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 624, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 115, + 624, + 506, + 638 + ], + "score": 1.0, + "content": "Welling. Modeling relational data with graph convolutional networks. In Aldo Gangemi, Roberto", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 636, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 636, + 505, + 649 + ], + "score": 1.0, + "content": "Navigli, Maria-Esther Vidal, Pascal Hitzler, Raphael Troncy, Laura Hollink, Anna Tordai, and ¨", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 645, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 114, + 645, + 505, + 660 + ], + "score": 1.0, + "content": "Mehwish Alam (eds.), The Semantic Web, pp. 593–607, Cham, 2018. Springer International Pub-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 658, + 258, + 671 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 258, + 671 + ], + "score": 1.0, + "content": "lishing. ISBN 978-3-319-93417-4.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 106, + 613, + 506, + 671 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "tions. In Proceedings of the 2018 Conference of the North American Chapter of the Association", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 697, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 114, + 697, + 505, + 713 + ], + "score": 1.0, + "content": "for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pp.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "464–468, New Orleans, Louisiana, June 2018. Association for Computational Linguistics. doi:", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 719, + 489, + 734 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 489, + 734 + ], + "score": 1.0, + "content": "10.18653/v1/N18-2074. URL https://www.aclweb.org/anthology/N18-2074.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 676, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 97 + ], + "score": 1.0, + "content": "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. Ordered neurons: Integrating", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tree structures into recurrent neural networks. In International Conference on Learning Repre-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 467, + 118 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 396, + 118 + ], + "score": 1.0, + "content": "sentations, 2019. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 396, + 106, + 402, + 114 + ], + "score": 0.4, + "content": "{ . } =", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 104, + 467, + 118 + ], + "score": 1.0, + "content": "B1l6qiR5F7.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Vighnesh Shiv and Chris Quirk. Novel positional encodings to enable tree-based transformers. In", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 430, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 430, + 147 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 12081–12091, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 505, + 188 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 conference on empirical methods in natural language pro-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 187, + 240, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 187, + 240, + 198 + ], + "score": 1.0, + "content": "cessing, pp. 1631–1642, 2013.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 503, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 504, + 228 + ], + "score": 1.0, + "content": "from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 257, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 115, + 257, + 287, + 268 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li ´ o, and Yoshua `", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "Bengio. Graph Attention Networks. International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 297, + 416, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 297, + 351, + 311 + ], + "score": 1.0, + "content": "2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 351, + 299, + 357, + 307 + ], + "score": 0.44, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 297, + 416, + 311 + ], + "score": 1.0, + "content": "rJXMpikCZ.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 504, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 506, + 329 + ], + "score": 1.0, + "content": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in neural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 327, + 336, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 327, + 336, + 340 + ], + "score": 1.0, + "content": "information processing systems, pp. 2692–2700, 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 504, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 504, + 359 + ], + "score": 1.0, + "content": "Song Wang, Devin Chollak, Dana Movshovitz-Attias, and Lin Tan. Bugram: Bug detection with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 357, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 114, + 357, + 506, + 372 + ], + "score": 1.0, + "content": "n-gram language models. In Proceedings of the 31st IEEE/ACM International Conference on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 506, + 381 + ], + "score": 1.0, + "content": "Automated Software Engineering, ASE 2016, pp. 708–719, New York, NY, USA, 2016. Associ-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 115, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "ation for Computing Machinery. ISBN 9781450338455. doi: 10.1145/2970276.2970341. URL", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 390, + 355, + 403 + ], + "spans": [ + { + "bbox": [ + 115, + 390, + 355, + 403 + ], + "score": 1.0, + "content": "https://doi.org/10.1145/2970276.2970341.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 409, + 504, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "score": 1.0, + "content": "Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 114, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "networks? In International Conference on Learning Representations, 2019. URL https:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 432, + 337, + 444 + ], + "spans": [ + { + "bbox": [ + 117, + 432, + 266, + 444 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 433, + 273, + 441 + ], + "score": 0.66, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 432, + 337, + 444 + ], + "score": 1.0, + "content": "ryGs6iA5Km.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 108, + 450, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 463 + ], + "score": 1.0, + "content": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 116, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 472, + 336, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 472, + 336, + 484 + ], + "score": 1.0, + "content": "information processing systems, pp. 5754–5764, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 504 + ], + "score": 1.0, + "content": "Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hi-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 501, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 505, + 515 + ], + "score": 1.0, + "content": "erarchical graph representation learning with differentiable pooling. In Advances in neural infor-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 513, + 317, + 525 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 317, + 525 + ], + "score": 1.0, + "content": "mation processing systems, pp. 4800–4810, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "Jiaxuan You, Rex Ying, and Jure Leskovec. Position-aware graph neural networks. In Kamalika", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 542, + 506, + 556 + ], + "score": 1.0, + "content": "Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 554, + 504, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 504, + 567 + ], + "score": 1.0, + "content": "on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 7134–", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 563, + 504, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 504, + 578 + ], + "score": 1.0, + "content": "7143. PMLR, 09–15 Jun 2019. URL http://proceedings.mlr.press/v97/you19b.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 576, + 146, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 576, + 146, + 588 + ], + "score": 1.0, + "content": "html.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 108, + 595, + 504, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 607 + ], + "score": 1.0, + "content": "Muhan Zhang and Yixin Chen. Link prediction based on graph neural networks. In Advances in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 605, + 369, + 618 + ], + "spans": [ + { + "bbox": [ + 115, + 605, + 369, + 618 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 5165–5175, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 97 + ], + "score": 1.0, + "content": "Yikang Shen, Shawn Tan, Alessandro Sordoni, and Aaron Courville. Ordered neurons: Integrating", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tree structures into recurrent neural networks. In International Conference on Learning Repre-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 467, + 118 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 396, + 118 + ], + "score": 1.0, + "content": "sentations, 2019. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 396, + 106, + 402, + 114 + ], + "score": 0.4, + "content": "{ . } =", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 104, + 467, + 118 + ], + "score": 1.0, + "content": "B1l6qiR5F7.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 80, + 505, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Vighnesh Shiv and Chris Quirk. Novel positional encodings to enable tree-based transformers. In", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 430, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 430, + 147 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 12081–12091, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 123, + 505, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 505, + 198 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 167 + ], + "score": 1.0, + "content": "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 505, + 188 + ], + "score": 1.0, + "content": "treebank. In Proceedings of the 2013 conference on empirical methods in natural language pro-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 187, + 240, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 187, + 240, + 198 + ], + "score": 1.0, + "content": "cessing, pp. 1631–1642, 2013.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 152, + 505, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 503, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "Kai Sheng Tai, Richard Socher, and Christopher D Manning. Improved semantic representations", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 504, + 228 + ], + "score": 1.0, + "content": "from tree-structured long short-term memory networks. arXiv preprint arXiv:1503.00075, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 204, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 257, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 115, + 257, + 287, + 268 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 233, + 505, + 268 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li ´ o, and Yoshua `", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "Bengio. Graph Attention Networks. International Conference on Learning Representations,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 297, + 416, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 297, + 351, + 311 + ], + "score": 1.0, + "content": "2018. URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 351, + 299, + 357, + 307 + ], + "score": 0.44, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 297, + 416, + 311 + ], + "score": 1.0, + "content": "rJXMpikCZ.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 275, + 505, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 504, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 506, + 329 + ], + "score": 1.0, + "content": "Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in neural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 327, + 336, + 340 + ], + "spans": [ + { + "bbox": [ + 115, + 327, + 336, + 340 + ], + "score": 1.0, + "content": "information processing systems, pp. 2692–2700, 2015.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 315, + 506, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 504, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 504, + 359 + ], + "score": 1.0, + "content": "Song Wang, Devin Chollak, Dana Movshovitz-Attias, and Lin Tan. Bugram: Bug detection with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 357, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 114, + 357, + 506, + 372 + ], + "score": 1.0, + "content": "n-gram language models. In Proceedings of the 31st IEEE/ACM International Conference on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 368, + 506, + 381 + ], + "score": 1.0, + "content": "Automated Software Engineering, ASE 2016, pp. 708–719, New York, NY, USA, 2016. Associ-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 115, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "ation for Computing Machinery. ISBN 9781450338455. doi: 10.1145/2970276.2970341. URL", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 390, + 355, + 403 + ], + "spans": [ + { + "bbox": [ + 115, + 390, + 355, + 403 + ], + "score": 1.0, + "content": "https://doi.org/10.1145/2970276.2970341.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 347, + 506, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 409, + 504, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 505, + 421 + ], + "score": 1.0, + "content": "Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 114, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 114, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "networks? In International Conference on Learning Representations, 2019. URL https:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 117, + 432, + 337, + 444 + ], + "spans": [ + { + "bbox": [ + 117, + 432, + 266, + 444 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 433, + 273, + 441 + ], + "score": 0.66, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 432, + 337, + 444 + ], + "score": 1.0, + "content": "ryGs6iA5Km.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 409, + 505, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 450, + 504, + 484 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 463 + ], + "score": 1.0, + "content": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 461, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 116, + 461, + 506, + 474 + ], + "score": 1.0, + "content": "Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 472, + 336, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 472, + 336, + 484 + ], + "score": 1.0, + "content": "information processing systems, pp. 5754–5764, 2019.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 449, + 506, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 491, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 505, + 504 + ], + "score": 1.0, + "content": "Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hi-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 501, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 116, + 501, + 505, + 515 + ], + "score": 1.0, + "content": "erarchical graph representation learning with differentiable pooling. In Advances in neural infor-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 513, + 317, + 525 + ], + "spans": [ + { + "bbox": [ + 116, + 513, + 317, + 525 + ], + "score": 1.0, + "content": "mation processing systems, pp. 4800–4810, 2018.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 490, + 505, + 525 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 531, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "Jiaxuan You, Rex Ying, and Jure Leskovec. Position-aware graph neural networks. In Kamalika", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 542, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 542, + 506, + 556 + ], + "score": 1.0, + "content": "Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 554, + 504, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 554, + 504, + 567 + ], + "score": 1.0, + "content": "on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pp. 7134–", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 563, + 504, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 504, + 578 + ], + "score": 1.0, + "content": "7143. PMLR, 09–15 Jun 2019. URL http://proceedings.mlr.press/v97/you19b.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 576, + 146, + 588 + ], + "spans": [ + { + "bbox": [ + 115, + 576, + 146, + 588 + ], + "score": 1.0, + "content": "html.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 532, + 506, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 595, + 504, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 607 + ], + "score": 1.0, + "content": "Muhan Zhang and Yixin Chen. Link prediction based on graph neural networks. In Advances in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 605, + 369, + 618 + ], + "spans": [ + { + "bbox": [ + 115, + 605, + 369, + 618 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pp. 5165–5175, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 595, + 506, + 618 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 80, + 410, + 148 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 201, + 80, + 410, + 148 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 201, + 80, + 410, + 148 + ], + "spans": [ + { + "bbox": [ + 201, + 80, + 410, + 148 + ], + "score": 0.975, + "html": "
PythonJavascriptGoRuby
Python1.000.430.420.79
Javascript0.431.000.840.39
Go0.430.841.000.38
Ruby0.790.390.381.00
", + "type": "table", + "image_path": "26bec4900b3fdc755545a32837761c235cd99ba3567c283f36c397e15113a344.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 201, + 80, + 410, + 93.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 201, + 93.6, + 410, + 107.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 201, + 107.19999999999999, + 410, + 120.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 201, + 120.79999999999998, + 410, + 134.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 201, + 134.39999999999998, + 410, + 147.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 156, + 503, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "Table 5: Pairwise cosine similarities of the learned language embeddings of the CODE TRANS-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 147, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 147, + 178 + ], + "score": 1.0, + "content": "FORMER.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "title", + "bbox": [ + 108, + 200, + 225, + 212 + ], + "lines": [ + { + "bbox": [ + 107, + 200, + 226, + 213 + ], + "spans": [ + { + "bbox": [ + 107, + 200, + 226, + 213 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 224, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "We are grateful to Dylan Bourgeois for having paved the way to this research contribution with", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 236, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 247 + ], + "score": 1.0, + "content": "his thesis work (Bourgeois, 2019). We further thank Simon Geisler for his helpful suggestions and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "proofreading the paper, as well as the anonymous reviewers for their constructive feedback and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 257, + 187, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 187, + 269 + ], + "score": 1.0, + "content": "fruitful discussions.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 274, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "This research was supported by the TUM International Graduate School of Science and Engineering", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "(IGSSE). Stanford University is supported by DARPA under Nos. N660011924033 (MCS); ARO", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "under Nos. W911NF-16-1- 0342 (MURI), W911NF-16-1-0171 (DURIP); NSF under Nos. OAC-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 306, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 321 + ], + "score": 1.0, + "content": "1835598 (CINES), OAC-1934578 (HDR), CCF-1918940 (Expeditions), IIS-2030477 (RAPID);", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "Stanford Data Science Initiative, Wu Tsai Neurosciences Institute, Chan Zuckerberg Biohub, Ama-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "zon, JPMorgan Chase, Docomo, Hitachi, JD.com, KDDI, NVIDIA, Dell, Toshiba, Intel, and Unit-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 340, + 406, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 406, + 355 + ], + "score": 1.0, + "content": "edHealth Group. Jure Leskovec is a Chan Zuckerberg Biohub investigator.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 368, + 183, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 184, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 184, + 383 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 272, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 274, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 274, + 406 + ], + "score": 1.0, + "content": "A.1 DISTANCE ENCODING FUNCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 504, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 417, + 426 + ], + "score": 1.0, + "content": "For encoding scalar relation values via vectors we employ encoding functions", + "type": "text" + }, + { + "bbox": [ + 417, + 413, + 467, + 425 + ], + "score": 0.92, + "content": "\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 413, + 497, + 426 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 497, + 415, + 504, + 424 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "is the model’s embedding dimension. We choose the popular sinusoidal encoding function presented", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 436, + 207, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 207, + 448 + ], + "score": 1.0, + "content": "in Vaswani et al. (2017):", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 452, + 452, + 474 + ], + "lines": [ + { + "bbox": [ + 158, + 452, + 452, + 474 + ], + "spans": [ + { + "bbox": [ + 158, + 452, + 452, + 474 + ], + "score": 0.93, + "content": "\\phi ( r _ { i j } ) _ { 2 k } = \\sin ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) \\qquad \\phi ( r _ { i j } ) _ { 2 k + 1 } = \\cos ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) ,", + "type": "interline_equation", + "image_path": "f46894c730c22bd189c627fee0b9dffb06a45fdfb55c7f527bb4fe0d0c5327e8.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 158, + 452, + 452, + 474 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 480, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 134, + 493 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 480, + 197, + 492 + ], + "score": 0.92, + "content": "1 \\leq k < d / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 479, + 375, + 493 + ], + "score": 1.0, + "content": "is the position in the encoding vector and", + "type": "text" + }, + { + "bbox": [ + 375, + 480, + 388, + 490 + ], + "score": 0.84, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "is some constant; we adopt", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 107, + 492, + 163, + 503 + ], + "score": 0.86, + "content": "M = 1 0 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "as chosen by (Vaswani et al., 2017). Note that the distance encoding functions have", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 206, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 206, + 515 + ], + "score": 1.0, + "content": "no trainable parameters.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 108, + 528, + 322, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 325, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 325, + 540 + ], + "score": 1.0, + "content": "A.2 MULTILINGUAL REPRESENTATION ANALYSIS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "In Table 5, we show the pairwise cosine similarities of the learned language embeddings of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "CODE TRANSFORMER. We can see that the pairs Python-Ruby and Javascript-Go have similar", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "language embeddings. This aligns well with roots of language design and common use cases of the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 580, + 153, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 153, + 596 + ], + "score": 1.0, + "content": "languages.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "Moreover, in Table 6, we show selected snippets starting with is, main, or load (left) and their", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 610, + 324, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 324, + 621 + ], + "score": 1.0, + "content": "best embedding matches from other languages (right).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 108, + 634, + 230, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 230, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 230, + 646 + ], + "score": 1.0, + "content": "A.3 DATA PREPROCESSING", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 655, + 320, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 320, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 320, + 668 + ], + "score": 1.0, + "content": "A.3.1 TEXTUAL CODE SNIPPET PREPROCESSING", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 129, + 672, + 414, + 684 + ], + "lines": [ + { + "bbox": [ + 130, + 671, + 414, + 686 + ], + "spans": [ + { + "bbox": [ + 130, + 671, + 414, + 686 + ], + "score": 1.0, + "content": "1. Tokenize code snippets with Pygments language-specific tokenizer.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 131, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 128, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 128, + 687, + 329, + 700 + ], + "score": 1.0, + "content": "2. Remove comments (both multi-line,", + "type": "text" + }, + { + "bbox": [ + 340, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "single-line and doc comments).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 141, + 698, + 162, + 710 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 174, + 700, + 217, + 710 + ], + "score": 1.0, + "content": "comment", + "type": "text" + }, + { + "bbox": [ + 226, + 699, + 295, + 711 + ], + "score": 1.0, + "content": "token types.", + "type": "text" + }, + { + "bbox": [ + 338, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "pygments.token.Comment and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 141, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "pygments.token.Literal.String.Doc that are generated by Pygments are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 141, + 721, + 253, + 732 + ], + "spans": [ + { + "bbox": [ + 141, + 721, + 253, + 732 + ], + "score": 1.0, + "content": "used to identify comments.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 201, + 80, + 410, + 148 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 201, + 80, + 410, + 148 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 201, + 80, + 410, + 148 + ], + "spans": [ + { + "bbox": [ + 201, + 80, + 410, + 148 + ], + "score": 0.975, + "html": "
PythonJavascriptGoRuby
Python1.000.430.420.79
Javascript0.431.000.840.39
Go0.430.841.000.38
Ruby0.790.390.381.00
", + "type": "table", + "image_path": "26bec4900b3fdc755545a32837761c235cd99ba3567c283f36c397e15113a344.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 201, + 80, + 410, + 93.6 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 201, + 93.6, + 410, + 107.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 201, + 107.19999999999999, + 410, + 120.79999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 201, + 120.79999999999998, + 410, + 134.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 201, + 134.39999999999998, + 410, + 147.99999999999997 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 156, + 503, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "Table 5: Pairwise cosine similarities of the learned language embeddings of the CODE TRANS-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 147, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 147, + 178 + ], + "score": 1.0, + "content": "FORMER.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 3.75 + }, + { + "type": "title", + "bbox": [ + 108, + 200, + 225, + 212 + ], + "lines": [ + { + "bbox": [ + 107, + 200, + 226, + 213 + ], + "spans": [ + { + "bbox": [ + 107, + 200, + 226, + 213 + ], + "score": 1.0, + "content": "ACKNOWLEDGEMENTS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 224, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "We are grateful to Dylan Bourgeois for having paved the way to this research contribution with", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 236, + 505, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 247 + ], + "score": 1.0, + "content": "his thesis work (Bourgeois, 2019). We further thank Simon Geisler for his helpful suggestions and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "proofreading the paper, as well as the anonymous reviewers for their constructive feedback and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 257, + 187, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 187, + 269 + ], + "score": 1.0, + "content": "fruitful discussions.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 225, + 505, + 269 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 274, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "This research was supported by the TUM International Graduate School of Science and Engineering", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "(IGSSE). Stanford University is supported by DARPA under Nos. N660011924033 (MCS); ARO", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "under Nos. W911NF-16-1- 0342 (MURI), W911NF-16-1-0171 (DURIP); NSF under Nos. OAC-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 306, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 321 + ], + "score": 1.0, + "content": "1835598 (CINES), OAC-1934578 (HDR), CCF-1918940 (Expeditions), IIS-2030477 (RAPID);", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 331 + ], + "score": 1.0, + "content": "Stanford Data Science Initiative, Wu Tsai Neurosciences Institute, Chan Zuckerberg Biohub, Ama-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "zon, JPMorgan Chase, Docomo, Hitachi, JD.com, KDDI, NVIDIA, Dell, Toshiba, Intel, and Unit-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 340, + 406, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 406, + 355 + ], + "score": 1.0, + "content": "edHealth Group. Jure Leskovec is a Chan Zuckerberg Biohub investigator.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 273, + 505, + 355 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 368, + 183, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 184, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 184, + 383 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 272, + 405 + ], + "lines": [ + { + "bbox": [ + 106, + 393, + 274, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 274, + 406 + ], + "score": 1.0, + "content": "A.1 DISTANCE ENCODING FUNCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 504, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 417, + 426 + ], + "score": 1.0, + "content": "For encoding scalar relation values via vectors we employ encoding functions", + "type": "text" + }, + { + "bbox": [ + 417, + 413, + 467, + 425 + ], + "score": 0.92, + "content": "\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 467, + 413, + 497, + 426 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 497, + 415, + 504, + 424 + ], + "score": 0.78, + "content": "d", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "is the model’s embedding dimension. We choose the popular sinusoidal encoding function presented", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 436, + 207, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 207, + 448 + ], + "score": 1.0, + "content": "in Vaswani et al. (2017):", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 413, + 505, + 448 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 452, + 452, + 474 + ], + "lines": [ + { + "bbox": [ + 158, + 452, + 452, + 474 + ], + "spans": [ + { + "bbox": [ + 158, + 452, + 452, + 474 + ], + "score": 0.93, + "content": "\\phi ( r _ { i j } ) _ { 2 k } = \\sin ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) \\qquad \\phi ( r _ { i j } ) _ { 2 k + 1 } = \\cos ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) ,", + "type": "interline_equation", + "image_path": "f46894c730c22bd189c627fee0b9dffb06a45fdfb55c7f527bb4fe0d0c5327e8.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 158, + 452, + 452, + 474 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 480, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 134, + 493 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 480, + 197, + 492 + ], + "score": 0.92, + "content": "1 \\leq k < d / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 479, + 375, + 493 + ], + "score": 1.0, + "content": "is the position in the encoding vector and", + "type": "text" + }, + { + "bbox": [ + 375, + 480, + 388, + 490 + ], + "score": 0.84, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "is some constant; we adopt", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 107, + 491, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 107, + 492, + 163, + 503 + ], + "score": 0.86, + "content": "M = 1 0 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 491, + 505, + 504 + ], + "score": 1.0, + "content": "as chosen by (Vaswani et al., 2017). Note that the distance encoding functions have", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 502, + 206, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 206, + 515 + ], + "score": 1.0, + "content": "no trainable parameters.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 479, + 505, + 515 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 528, + 322, + 539 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 325, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 325, + 540 + ], + "score": 1.0, + "content": "A.2 MULTILINGUAL REPRESENTATION ANALYSIS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 547, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "In Table 5, we show the pairwise cosine similarities of the learned language embeddings of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "CODE TRANSFORMER. We can see that the pairs Python-Ruby and Javascript-Go have similar", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "language embeddings. This aligns well with roots of language design and common use cases of the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 580, + 153, + 596 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 153, + 596 + ], + "score": 1.0, + "content": "languages.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 104, + 548, + 506, + 596 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 598, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 505, + 610 + ], + "score": 1.0, + "content": "Moreover, in Table 6, we show selected snippets starting with is, main, or load (left) and their", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 610, + 324, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 324, + 621 + ], + "score": 1.0, + "content": "best embedding matches from other languages (right).", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 597, + 505, + 621 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 634, + 230, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 635, + 230, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 230, + 646 + ], + "score": 1.0, + "content": "A.3 DATA PREPROCESSING", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 655, + 320, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 320, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 320, + 668 + ], + "score": 1.0, + "content": "A.3.1 TEXTUAL CODE SNIPPET PREPROCESSING", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 129, + 672, + 414, + 684 + ], + "lines": [ + { + "bbox": [ + 130, + 671, + 414, + 686 + ], + "spans": [ + { + "bbox": [ + 130, + 671, + 414, + 686 + ], + "score": 1.0, + "content": "1. Tokenize code snippets with Pygments language-specific tokenizer.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37, + "bbox_fs": [ + 130, + 671, + 414, + 686 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 128, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 128, + 687, + 329, + 700 + ], + "score": 1.0, + "content": "2. Remove comments (both multi-line,", + "type": "text" + }, + { + "bbox": [ + 340, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "single-line and doc comments).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 141, + 698, + 162, + 710 + ], + "score": 1.0, + "content": "The", + "type": "text" + }, + { + "bbox": [ + 174, + 700, + 217, + 710 + ], + "score": 1.0, + "content": "comment", + "type": "text" + }, + { + "bbox": [ + 226, + 699, + 295, + 711 + ], + "score": 1.0, + "content": "token types.", + "type": "text" + }, + { + "bbox": [ + 338, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "pygments.token.Comment and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 141, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "pygments.token.Literal.String.Doc that are generated by Pygments are", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 141, + 721, + 253, + 732 + ], + "spans": [ + { + "bbox": [ + 141, + 721, + 253, + 732 + ], + "score": 1.0, + "content": "used to identify comments.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 128, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 12, + 530, + 685 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 12, + 530, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 12, + 530, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 12, + 530, + 685 + ], + "score": 0.422, + "html": "
Published as a conference paper at ICLR 2021
function isEqualArray(a,b){ if (a return true;
if ((a === undefined) undefined)){ return false;func areSameFloat32Array(a,b []float32) bool if len(a) != len(b){ return false
var =a.length; if (i !== b.length){ return false;for i := O;i < len(a);i++{ if a[i]!=b[i]{ return false J
while (i--){ if(a[i] !== b[i]){ return false; }return true
return true;
function main() var rawData $(' .HeaderTexture[data-login- user-email]') .data() ;func TaskSayHello(t *tasking.T){
if (rawData) 1 me={username := t.Flags.String("name") if username user, := user.Current ()
name:rawData.loginUserName, mail: rawData.loginUserEmail };username = user.Name
getCartId(function (cart){ me.cart = cart;} ift.Flags.Bool("verbose")
injectMenu(); refreshUsers(actions.updateUsers); listenOnChanges(onChange);t.Logf("Hello %s, the time now is %s\\n", username,time.Now())
listenonorderConfirm(onConfirm); );}else{ t.Logf("Hello %s\\n",username)
} else{ caliback('no user');
func Backup(filename string)error{
info,err := os.Stat(filename) if err != nil
if os.IsNotExist(err){
return nil
return err
def _load_rule_file(self, filename):if info.Size() == 0{ return nil
if not (os.path.exists(filename)):}
sys.stderr.write(files,err := filepath.Glob(
"rflint: %s: No such file or "directory\\n" % filenamefilename + _BACKUP_SUFFIX
return) if err nil 1
try:return err
basename = os.path.basename(filename)}
(name,ext)= os.path.splitext(basename) imp.load_source(name,filename) numBackup := byte(1)
except Exception as e: sys.stderr.write(if len(files) !=0{
"rflint: %s: exception whilelastFile := files[len(files)-1]
"loading:%s\\n" % (filename, str(e))numBackup = lastFile[len(lastFile)-2] + 1
if numBackup >
numBackup = '1'
else{
'1'
numBackup =
return Copy (filename,
fmt.Sprintf("%s+%s~",
string(numBackup)))
", + "type": "table", + "image_path": "dfeb9bd55674c1d88bd8727a7ae6e6c8b0e9d1b458288922467db71461c08de5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 12, + 530, + 236.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 236.33333333333334, + 530, + 460.6666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 460.6666666666667, + 530, + 685.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 685, + 505, + 708 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 683, + 505, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 505, + 697 + ], + "score": 1.0, + "content": "Table 6: Selected snippets starting with is, main, or load (left) and their best embedding matches", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 694, + 224, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 224, + 709 + ], + "score": 1.0, + "content": "from other languages (right).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 12, + 530, + 685 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 12, + 530, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 12, + 530, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 12, + 530, + 685 + ], + "score": 0.422, + "html": "
Published as a conference paper at ICLR 2021
function isEqualArray(a,b){ if (a return true;
if ((a === undefined) undefined)){ return false;func areSameFloat32Array(a,b []float32) bool if len(a) != len(b){ return false
var =a.length; if (i !== b.length){ return false;for i := O;i < len(a);i++{ if a[i]!=b[i]{ return false J
while (i--){ if(a[i] !== b[i]){ return false; }return true
return true;
function main() var rawData $(' .HeaderTexture[data-login- user-email]') .data() ;func TaskSayHello(t *tasking.T){
if (rawData) 1 me={username := t.Flags.String("name") if username user, := user.Current ()
name:rawData.loginUserName, mail: rawData.loginUserEmail };username = user.Name
getCartId(function (cart){ me.cart = cart;} ift.Flags.Bool("verbose")
injectMenu(); refreshUsers(actions.updateUsers); listenOnChanges(onChange);t.Logf("Hello %s, the time now is %s\\n", username,time.Now())
listenonorderConfirm(onConfirm); );}else{ t.Logf("Hello %s\\n",username)
} else{ caliback('no user');
func Backup(filename string)error{
info,err := os.Stat(filename) if err != nil
if os.IsNotExist(err){
return nil
return err
def _load_rule_file(self, filename):if info.Size() == 0{ return nil
if not (os.path.exists(filename)):}
sys.stderr.write(files,err := filepath.Glob(
"rflint: %s: No such file or "directory\\n" % filenamefilename + _BACKUP_SUFFIX
return) if err nil 1
try:return err
basename = os.path.basename(filename)}
(name,ext)= os.path.splitext(basename) imp.load_source(name,filename) numBackup := byte(1)
except Exception as e: sys.stderr.write(if len(files) !=0{
"rflint: %s: exception whilelastFile := files[len(files)-1]
"loading:%s\\n" % (filename, str(e))numBackup = lastFile[len(lastFile)-2] + 1
if numBackup >
numBackup = '1'
else{
'1'
numBackup =
return Copy (filename,
fmt.Sprintf("%s+%s~",
string(numBackup)))
", + "type": "table", + "image_path": "dfeb9bd55674c1d88bd8727a7ae6e6c8b0e9d1b458288922467db71461c08de5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 12, + 530, + 236.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 236.33333333333334, + 530, + 460.6666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 460.6666666666667, + 530, + 685.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 685, + 505, + 708 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 683, + 505, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 505, + 697 + ], + "score": 1.0, + "content": "Table 6: Selected snippets starting with is, main, or load (left) and their best embedding matches", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 694, + 224, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 224, + 709 + ], + "score": 1.0, + "content": "from other languages (right).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 79, + 420, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 184, + 79, + 420, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 79, + 420, + 264 + ], + "spans": [ + { + "bbox": [ + 184, + 79, + 420, + 264 + ], + "score": 0.967, + "type": "image", + "image_path": "e37f9cb44d930542340c5ad3d88619f8535c0f5512daf8c0c8d8a050432bff4b.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 79, + 420, + 92.21428571428571 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 184, + 92.21428571428571, + 420, + 105.42857142857142 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 184, + 105.42857142857142, + 420, + 118.64285714285712 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 184, + 118.64285714285712, + 420, + 131.85714285714283 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 184, + 131.85714285714283, + 420, + 145.07142857142856 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 184, + 145.07142857142856, + 420, + 158.28571428571428 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 184, + 158.28571428571428, + 420, + 171.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 184, + 171.5, + 420, + 184.71428571428572 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 184, + 184.71428571428572, + 420, + 197.92857142857144 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 184, + 197.92857142857144, + 420, + 211.14285714285717 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 184, + 211.14285714285717, + 420, + 224.3571428571429 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 184, + 224.3571428571429, + 420, + 237.5714285714286 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 184, + 237.5714285714286, + 420, + 250.78571428571433 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 184, + 250.78571428571433, + 420, + 264.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 271, + 480, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 271, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 129, + 271, + 481, + 285 + ], + "score": 1.0, + "content": "Figure 6: Example snippet and its corresponding AST obtained from GitHub Semantic.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 10.25 + }, + { + "type": "text", + "bbox": [ + 131, + 308, + 248, + 319 + ], + "lines": [ + { + "bbox": [ + 129, + 307, + 250, + 321 + ], + "spans": [ + { + "bbox": [ + 129, + 307, + 250, + 321 + ], + "score": 1.0, + "content": "3. Empty lines are removed.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 129, + 326, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 129, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 129, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "4. Hard coded strings and numbers are replaced with a special [MASK STRING] and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 144, + 337, + 247, + 348 + ], + "spans": [ + { + "bbox": [ + 144, + 337, + 247, + 348 + ], + "score": 1.0, + "content": "[MASK NUMBER] token.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 130, + 354, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 128, + 353, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 128, + 353, + 505, + 368 + ], + "score": 1.0, + "content": "5. Indentation style of the code snippet is detected and whitespace characters at the beginning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "of a line are replaced with a single [INDENT] or [DEDENT] token when indentation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 140, + 374, + 181, + 391 + ], + "spans": [ + { + "bbox": [ + 140, + 374, + 181, + 391 + ], + "score": 1.0, + "content": "changes.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 130, + 394, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 129, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 129, + 394, + 420, + 407 + ], + "score": 1.0, + "content": "6. Tokens are further split into sub tokens, e.g., setBottomHeight", + "type": "text" + }, + { + "bbox": [ + 420, + 396, + 434, + 405 + ], + "score": 0.6, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "[‘set’, ‘bottom’,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 143, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 143, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "‘height’]. Throughout our experiments, we use 5 input sub tokens. If a token consists", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 415, + 484, + 430 + ], + "spans": [ + { + "bbox": [ + 141, + 415, + 484, + 430 + ], + "score": 1.0, + "content": "of less than 5 sub tokens, the remaining spaces are filled with a special [PAD] token.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 129, + 434, + 503, + 457 + ], + "lines": [ + { + "bbox": [ + 129, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 129, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "7. Any remaining tokens that only consist of white spaces are removed. The only white", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 445, + 344, + 458 + ], + "spans": [ + { + "bbox": [ + 141, + 445, + 322, + 458 + ], + "score": 1.0, + "content": "space characters that are kept are line breaks", + "type": "text" + }, + { + "bbox": [ + 323, + 445, + 339, + 457 + ], + "score": 0.85, + "content": "^ { \\bullet } \\backslash \\boldsymbol { \\mathrm { n } } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 445, + 344, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 127, + 463, + 492, + 475 + ], + "lines": [ + { + "bbox": [ + 128, + 463, + 492, + 476 + ], + "spans": [ + { + "bbox": [ + 128, + 463, + 492, + 476 + ], + "score": 1.0, + "content": "8. Any code snippets where the Pygments tokenizer cannot parse a token are discarded.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 107, + 490, + 360, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 361, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 361, + 504 + ], + "score": 1.0, + "content": "A.3.2 STAGE 1 PREPROCESSING (GENERATION OF ASTS)", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 131, + 511, + 504, + 544 + ], + "lines": [ + { + "bbox": [ + 129, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 129, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "1. Stripped code snippets are used to generate language-specific ASTs. For Java, we use the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 521, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 141, + 521, + 505, + 535 + ], + "score": 1.0, + "content": "AST parser from the java-parser project. The ASTs contain node types and source ranges.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 532, + 368, + 545 + ], + "spans": [ + { + "bbox": [ + 142, + 532, + 368, + 545 + ], + "score": 1.0, + "content": "For Python, JavaScript, Ruby and Go, we use semantic.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 130, + 551, + 371, + 562 + ], + "lines": [ + { + "bbox": [ + 129, + 550, + 372, + 564 + ], + "spans": [ + { + "bbox": [ + 129, + 550, + 372, + 564 + ], + "score": 1.0, + "content": "2. Snippets that lead to an AST parse error are discarded.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 129, + 569, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 129, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 129, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "3. We calculate a mapping between tokens and nodes in the AST. Every token is assigned", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 141, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "to the node in the AST with shortest source range that still encompasses the source range", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 141, + 589, + 195, + 603 + ], + "spans": [ + { + "bbox": [ + 141, + 589, + 195, + 603 + ], + "score": 1.0, + "content": "of the token.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 142, + 602, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 141, + 601, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 141, + 601, + 506, + 614 + ], + "score": 1.0, + "content": "To find such a node, we originally intended to make use of the assumption that source", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 142, + 613, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 142, + 613, + 506, + 624 + ], + "score": 1.0, + "content": "ranges of child nodes do not overlap. Then, one could easily find the node with smallest", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 141, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "encompassing source range by greedily selecting at every layer in the AST the child that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 141, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "encompasses the token’s source range (there can only be at most one child that fulfills this).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 141, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "However, this assumption does not hold for all ASTs (see Figure 6 for an example). As a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 141, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "heuristic, we greedily select the child node with the shorter source range in case there were", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 142, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "multiple child nodes with encompassing source ranges. This approximation seems to be", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 142, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 142, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "sufficient in our case, and limits runtime as we do not have to consider multiple paths in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 141, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 141, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "the AST. It is also sufficient to stop when no child node encompasses the source range of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 141, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 141, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "the token, as in ASTs the source ranges of child nodes are always contained in the source", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 141, + 712, + 232, + 723 + ], + "spans": [ + { + "bbox": [ + 141, + 712, + 232, + 723 + ], + "score": 1.0, + "content": "ranges of their parent.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 184, + 79, + 420, + 264 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 184, + 79, + 420, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 184, + 79, + 420, + 264 + ], + "spans": [ + { + "bbox": [ + 184, + 79, + 420, + 264 + ], + "score": 0.967, + "type": "image", + "image_path": "e37f9cb44d930542340c5ad3d88619f8535c0f5512daf8c0c8d8a050432bff4b.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 184, + 79, + 420, + 92.21428571428571 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 184, + 92.21428571428571, + 420, + 105.42857142857142 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 184, + 105.42857142857142, + 420, + 118.64285714285712 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 184, + 118.64285714285712, + 420, + 131.85714285714283 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 184, + 131.85714285714283, + 420, + 145.07142857142856 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 184, + 145.07142857142856, + 420, + 158.28571428571428 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 184, + 158.28571428571428, + 420, + 171.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 184, + 171.5, + 420, + 184.71428571428572 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 184, + 184.71428571428572, + 420, + 197.92857142857144 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 184, + 197.92857142857144, + 420, + 211.14285714285717 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 184, + 211.14285714285717, + 420, + 224.3571428571429 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 184, + 224.3571428571429, + 420, + 237.5714285714286 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 184, + 237.5714285714286, + 420, + 250.78571428571433 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 184, + 250.78571428571433, + 420, + 264.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 271, + 480, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 271, + 481, + 285 + ], + "spans": [ + { + "bbox": [ + 129, + 271, + 481, + 285 + ], + "score": 1.0, + "content": "Figure 6: Example snippet and its corresponding AST obtained from GitHub Semantic.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 10.25 + }, + { + "type": "text", + "bbox": [ + 131, + 308, + 248, + 319 + ], + "lines": [ + { + "bbox": [ + 129, + 307, + 250, + 321 + ], + "spans": [ + { + "bbox": [ + 129, + 307, + 250, + 321 + ], + "score": 1.0, + "content": "3. Empty lines are removed.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 129, + 307, + 250, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 326, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 129, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 129, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "4. Hard coded strings and numbers are replaced with a special [MASK STRING] and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 144, + 337, + 247, + 348 + ], + "spans": [ + { + "bbox": [ + 144, + 337, + 247, + 348 + ], + "score": 1.0, + "content": "[MASK NUMBER] token.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 129, + 325, + 505, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 354, + 505, + 388 + ], + "lines": [ + { + "bbox": [ + 128, + 353, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 128, + 353, + 505, + 368 + ], + "score": 1.0, + "content": "5. Indentation style of the code snippet is detected and whitespace characters at the beginning", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "of a line are replaced with a single [INDENT] or [DEDENT] token when indentation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 140, + 374, + 181, + 391 + ], + "spans": [ + { + "bbox": [ + 140, + 374, + 181, + 391 + ], + "score": 1.0, + "content": "changes.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 128, + 353, + 505, + 391 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 394, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 129, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 129, + 394, + 420, + 407 + ], + "score": 1.0, + "content": "6. Tokens are further split into sub tokens, e.g., setBottomHeight", + "type": "text" + }, + { + "bbox": [ + 420, + 396, + 434, + 405 + ], + "score": 0.6, + "content": "", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "[‘set’, ‘bottom’,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 143, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 143, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "‘height’]. Throughout our experiments, we use 5 input sub tokens. If a token consists", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 415, + 484, + 430 + ], + "spans": [ + { + "bbox": [ + 141, + 415, + 484, + 430 + ], + "score": 1.0, + "content": "of less than 5 sub tokens, the remaining spaces are filled with a special [PAD] token.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 129, + 394, + 505, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 434, + 503, + 457 + ], + "lines": [ + { + "bbox": [ + 129, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 129, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "7. Any remaining tokens that only consist of white spaces are removed. The only white", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 445, + 344, + 458 + ], + "spans": [ + { + "bbox": [ + 141, + 445, + 322, + 458 + ], + "score": 1.0, + "content": "space characters that are kept are line breaks", + "type": "text" + }, + { + "bbox": [ + 323, + 445, + 339, + 457 + ], + "score": 0.85, + "content": "^ { \\bullet } \\backslash \\boldsymbol { \\mathrm { n } } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 445, + 344, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 129, + 433, + 505, + 458 + ] + }, + { + "type": "text", + "bbox": [ + 127, + 463, + 492, + 475 + ], + "lines": [ + { + "bbox": [ + 128, + 463, + 492, + 476 + ], + "spans": [ + { + "bbox": [ + 128, + 463, + 492, + 476 + ], + "score": 1.0, + "content": "8. Any code snippets where the Pygments tokenizer cannot parse a token are discarded.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 128, + 463, + 492, + 476 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 490, + 360, + 502 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 361, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 361, + 504 + ], + "score": 1.0, + "content": "A.3.2 STAGE 1 PREPROCESSING (GENERATION OF ASTS)", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 131, + 511, + 504, + 544 + ], + "lines": [ + { + "bbox": [ + 129, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 129, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "1. Stripped code snippets are used to generate language-specific ASTs. For Java, we use the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 521, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 141, + 521, + 505, + 535 + ], + "score": 1.0, + "content": "AST parser from the java-parser project. The ASTs contain node types and source ranges.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 142, + 532, + 368, + 545 + ], + "spans": [ + { + "bbox": [ + 142, + 532, + 368, + 545 + ], + "score": 1.0, + "content": "For Python, JavaScript, Ruby and Go, we use semantic.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 129, + 510, + 505, + 545 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 551, + 371, + 562 + ], + "lines": [ + { + "bbox": [ + 129, + 550, + 372, + 564 + ], + "spans": [ + { + "bbox": [ + 129, + 550, + 372, + 564 + ], + "score": 1.0, + "content": "2. Snippets that lead to an AST parse error are discarded.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 129, + 550, + 372, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 569, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 129, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 129, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "3. We calculate a mapping between tokens and nodes in the AST. Every token is assigned", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 141, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "to the node in the AST with shortest source range that still encompasses the source range", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 141, + 589, + 195, + 603 + ], + "spans": [ + { + "bbox": [ + 141, + 589, + 195, + 603 + ], + "score": 1.0, + "content": "of the token.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 129, + 569, + 505, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 142, + 602, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 141, + 601, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 141, + 601, + 506, + 614 + ], + "score": 1.0, + "content": "To find such a node, we originally intended to make use of the assumption that source", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 142, + 613, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 142, + 613, + 506, + 624 + ], + "score": 1.0, + "content": "ranges of child nodes do not overlap. Then, one could easily find the node with smallest", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 141, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "encompassing source range by greedily selecting at every layer in the AST the child that", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 141, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "encompasses the token’s source range (there can only be at most one child that fulfills this).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 141, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 141, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "However, this assumption does not hold for all ASTs (see Figure 6 for an example). As a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 141, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "heuristic, we greedily select the child node with the shorter source range in case there were", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 142, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "multiple child nodes with encompassing source ranges. This approximation seems to be", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 142, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 142, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "sufficient in our case, and limits runtime as we do not have to consider multiple paths in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 141, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 141, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "the AST. It is also sufficient to stop when no child node encompasses the source range of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 141, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 141, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "the token, as in ASTs the source ranges of child nodes are always contained in the source", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 141, + 712, + 232, + 723 + ], + "spans": [ + { + "bbox": [ + 141, + 712, + 232, + 723 + ], + "score": 1.0, + "content": "ranges of their parent.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40, + "bbox_fs": [ + 141, + 601, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 129, + 99, + 505, + 122 + ], + "lines": [ + { + "bbox": [ + 129, + 98, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 129, + 98, + 505, + 112 + ], + "score": 1.0, + "content": "1. Tokens are vocabularized. Any token occurring less than 100 times in the training set is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 111, + 267, + 123 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 231, + 123 + ], + "score": 1.0, + "content": "replaced by an ", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 111, + 267, + 123 + ], + "score": 1.0, + "content": "token.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 130, + 126, + 422, + 137 + ], + "lines": [ + { + "bbox": [ + 129, + 125, + 422, + 139 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 422, + 139 + ], + "score": 1.0, + "content": "2. We calculate multiple pair-wise relations between nodes in the AST:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 154, + 141, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 153, + 141, + 288, + 154 + ], + "spans": [ + { + "bbox": [ + 153, + 141, + 288, + 154 + ], + "score": 1.0, + "content": "• Personalized Page Rank (PPR)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 152, + 504, + 165 + ], + "spans": [ + { + "bbox": [ + 161, + 152, + 504, + 165 + ], + "score": 1.0, + "content": "We interpret the negative logarithm of PPR as a distance. We use a teleport probability", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 161, + 163, + 173, + 176 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 173, + 164, + 212, + 174 + ], + "score": 0.89, + "content": "\\alpha = 0 . 1 5", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 163, + 287, + 176 + ], + "score": 1.0, + "content": "and a threshold of", + "type": "text" + }, + { + "bbox": [ + 288, + 163, + 303, + 174 + ], + "score": 0.89, + "content": "e ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 163, + 381, + 176 + ], + "score": 1.0, + "content": ", i.e., anything with", + "type": "text" + }, + { + "bbox": [ + 382, + 164, + 449, + 175 + ], + "score": 0.89, + "content": "- \\log P P R > 5", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "is considered", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 162, + 176, + 213, + 185 + ], + "spans": [ + { + "bbox": [ + 162, + 176, + 213, + 185 + ], + "score": 1.0, + "content": "unreachable", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 155, + 188, + 323, + 199 + ], + "spans": [ + { + "bbox": [ + 155, + 188, + 323, + 199 + ], + "score": 1.0, + "content": "• Shortest path length between two nodes", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 154, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 154, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "• Ancestor shortest paths (bidirectional). That is, the parent has an ancestor shortest", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 161, + 212, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 161, + 212, + 506, + 223 + ], + "score": 1.0, + "content": "path distance of 1 to all its children and the child has a distance of -1 to its parents. We", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 161, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 161, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "consider nodes that are not ancestors or descendants of a node (i.e. not reachable by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 162, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 162, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "following only parent or only child relations) as not connected in the ancestor shortest", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 160, + 245, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 160, + 245, + 505, + 257 + ], + "score": 1.0, + "content": "paths relation. We encode this with a very large value in their distance; we have found", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 161, + 256, + 330, + 268 + ], + "spans": [ + { + "bbox": [ + 161, + 257, + 204, + 268 + ], + "score": 1.0, + "content": "a value of", + "type": "text" + }, + { + "bbox": [ + 204, + 256, + 230, + 267 + ], + "score": 0.25, + "content": "1 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 257, + 330, + 268 + ], + "score": 1.0, + "content": "to work well in practice.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 154, + 268, + 493, + 281 + ], + "spans": [ + { + "bbox": [ + 154, + 268, + 493, + 281 + ], + "score": 1.0, + "content": "• Next sibling shortest paths (bidirectional, analogous to the ancestor shortest paths)", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 142, + 285, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 141, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "Note that the ancestor shortest paths and next sibling shortest paths are required because", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 295, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 506, + 309 + ], + "score": 1.0, + "content": "treating the AST as a normal graph leads to ambiguity. In a graph, the neighbors of a node", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "have no ordering; however in the AST, the order of the children of a node reflects their", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "order in the code. Therefore, we explicitly include the next sibling shortest paths. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "ancestor shortest paths would not be required if we treated the AST as a directed graph; in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "this case, however, a leaf node could not reach any other node in the AST, and therefore", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 141, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "both PPR and shortest path length are not useful in this case. Therefore, we model the AST", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 361, + 421, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 421, + 375 + ], + "score": 1.0, + "content": "as undirected and inject the ancestor / child edges to avoid ambiguity.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 130, + 377, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 128, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 128, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "3. Distance values are binned into 32 bins using area-based exponential binning with a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 371, + 401 + ], + "score": 1.0, + "content": "growth factor of 1.3, i.e., the area of a bin’s rectangle (", + "type": "text" + }, + { + "bbox": [ + 371, + 390, + 378, + 398 + ], + "score": 0.48, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 388, + 428, + 401 + ], + "score": 1.0, + "content": ": bin range,", + "type": "text" + }, + { + "bbox": [ + 429, + 390, + 435, + 400 + ], + "score": 0.62, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 388, + 505, + 401 + ], + "score": 1.0, + "content": ": number of val-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 142, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "ues in bin) will be approximately 1.3 times bigger for the next bin (going away from the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "bin that contains the zero value). Additionally, for discrete distance measures (such as se-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "quence distance or shortest path length), we hard-code 9 values around 0 to have their own", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 142, + 432, + 361, + 444 + ], + "score": 1.0, + "content": "bins. For instance, on the sequence distance the values", + "type": "text" + }, + { + "bbox": [ + 362, + 432, + 420, + 443 + ], + "score": 0.9, + "content": "- 4 , - 3 , \\dots , 4", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "have their individual", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 441, + 407, + 457 + ], + "spans": [ + { + "bbox": [ + 141, + 441, + 407, + 457 + ], + "score": 1.0, + "content": "bins, and around those values we employ the exponential binning.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 131, + 459, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 128, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 128, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "4. Punctuation tokens (such as points or brackets) are removed from the input sequence, as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 141, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "experiments showed that their presence does not improve performance but slows down", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 480, + 278, + 493 + ], + "spans": [ + { + "bbox": [ + 142, + 480, + 278, + 493 + ], + "score": 1.0, + "content": "training due to bigger input sizes.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 132, + 496, + 504, + 529 + ], + "lines": [ + { + "bbox": [ + 129, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 129, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "5. Snippets that are longer than MAX NUM TOKENS after punctuation tokens are removed are", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 142, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 142, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "discarded from the training set. Throughout our experiments, we use MAX NUM TOKENS", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 142, + 518, + 468, + 529 + ], + "spans": [ + { + "bbox": [ + 142, + 518, + 173, + 529 + ], + "score": 0.76, + "content": "= ~ 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 518, + 424, + 529 + ], + "score": 1.0, + "content": ". During evaluation on the test set, we use MAX NUM TOKENS", + "type": "text" + }, + { + "bbox": [ + 424, + 518, + 464, + 528 + ], + "score": 0.76, + "content": "= ~ 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 518, + 468, + 529 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 108, + 544, + 286, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 288, + 556 + ], + "score": 1.0, + "content": "A.4 INPUT EMBEDDINGS TO THE MODEL", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "score": 1.0, + "content": "Besides its five subtokens (e.g., [‘get’, ‘data’, ‘[PAD]’, ‘[PAD]’, ‘[PAD]’]),", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "score": 1.0, + "content": "each input token has a token type (coming from the Pygments tokenizer) and an AST node type.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "The AST node type is the type of the node assigned to each respective token, as described in Section", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "A.3.2. We concatenate the embeddings of the five subtokens, the token type, and the AST node", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "type. Then, we apply a linear layer (without activation function) to project down to the model’s", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 619, + 200, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 200, + 632 + ], + "score": 1.0, + "content": "embedding dimension.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 108, + 644, + 275, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 277, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 277, + 657 + ], + "score": 1.0, + "content": "A.5 INPUT TO THE GREAT BASELINE", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "As mentioned in the main text, we also compare with GREAT Hellendoorn et al. (2020). Since", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "their preprocessing pipeline is proprietary and could not be shared with us even after contacting the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "authors, we provide to GREAT the same AST distances as our model. Since GREAT uses edges in-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "stead of distances to encode relations in the Structure, we essentially threshold the ancestor, sibling,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "and shortest-paths distances and provide the edges where the distances are equal to 1 (including their", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 210, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 210, + 733 + ], + "score": 1.0, + "content": "edge types) to the model.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 82, + 429, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 430, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 430, + 96 + ], + "score": 1.0, + "content": "A.3.3 STAGE 2 PREPROCESSING (CALCULATION OF DISTANCE MATRICES)", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 129, + 99, + 505, + 122 + ], + "lines": [ + { + "bbox": [ + 129, + 98, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 129, + 98, + 505, + 112 + ], + "score": 1.0, + "content": "1. Tokens are vocabularized. Any token occurring less than 100 times in the training set is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 141, + 111, + 267, + 123 + ], + "spans": [ + { + "bbox": [ + 141, + 111, + 231, + 123 + ], + "score": 1.0, + "content": "replaced by an ", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 111, + 267, + 123 + ], + "score": 1.0, + "content": "token.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 129, + 98, + 505, + 123 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 126, + 422, + 137 + ], + "lines": [ + { + "bbox": [ + 129, + 125, + 422, + 139 + ], + "spans": [ + { + "bbox": [ + 129, + 125, + 422, + 139 + ], + "score": 1.0, + "content": "2. We calculate multiple pair-wise relations between nodes in the AST:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 129, + 125, + 422, + 139 + ] + }, + { + "type": "list", + "bbox": [ + 154, + 141, + 505, + 281 + ], + "lines": [ + { + "bbox": [ + 153, + 141, + 288, + 154 + ], + "spans": [ + { + "bbox": [ + 153, + 141, + 288, + 154 + ], + "score": 1.0, + "content": "• Personalized Page Rank (PPR)", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 161, + 152, + 504, + 165 + ], + "spans": [ + { + "bbox": [ + 161, + 152, + 504, + 165 + ], + "score": 1.0, + "content": "We interpret the negative logarithm of PPR as a distance. We use a teleport probability", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 161, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 161, + 163, + 173, + 176 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 173, + 164, + 212, + 174 + ], + "score": 0.89, + "content": "\\alpha = 0 . 1 5", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 163, + 287, + 176 + ], + "score": 1.0, + "content": "and a threshold of", + "type": "text" + }, + { + "bbox": [ + 288, + 163, + 303, + 174 + ], + "score": 0.89, + "content": "e ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 163, + 381, + 176 + ], + "score": 1.0, + "content": ", i.e., anything with", + "type": "text" + }, + { + "bbox": [ + 382, + 164, + 449, + 175 + ], + "score": 0.89, + "content": "- \\log P P R > 5", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "is considered", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 162, + 176, + 213, + 185 + ], + "spans": [ + { + "bbox": [ + 162, + 176, + 213, + 185 + ], + "score": 1.0, + "content": "unreachable", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 155, + 188, + 323, + 199 + ], + "spans": [ + { + "bbox": [ + 155, + 188, + 323, + 199 + ], + "score": 1.0, + "content": "• Shortest path length between two nodes", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 154, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 154, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "• Ancestor shortest paths (bidirectional). That is, the parent has an ancestor shortest", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 212, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 161, + 212, + 506, + 223 + ], + "score": 1.0, + "content": "path distance of 1 to all its children and the child has a distance of -1 to its parents. We", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 161, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 161, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "consider nodes that are not ancestors or descendants of a node (i.e. not reachable by", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 162, + 234, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 162, + 234, + 506, + 246 + ], + "score": 1.0, + "content": "following only parent or only child relations) as not connected in the ancestor shortest", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 160, + 245, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 160, + 245, + 505, + 257 + ], + "score": 1.0, + "content": "paths relation. We encode this with a very large value in their distance; we have found", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 161, + 256, + 330, + 268 + ], + "spans": [ + { + "bbox": [ + 161, + 257, + 204, + 268 + ], + "score": 1.0, + "content": "a value of", + "type": "text" + }, + { + "bbox": [ + 204, + 256, + 230, + 267 + ], + "score": 0.25, + "content": "1 , 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 257, + 330, + 268 + ], + "score": 1.0, + "content": "to work well in practice.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 154, + 268, + 493, + 281 + ], + "spans": [ + { + "bbox": [ + 154, + 268, + 493, + 281 + ], + "score": 1.0, + "content": "• Next sibling shortest paths (bidirectional, analogous to the ancestor shortest paths)", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + } + ], + "index": 8.5, + "bbox_fs": [ + 153, + 141, + 506, + 281 + ] + }, + { + "type": "text", + "bbox": [ + 142, + 285, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 141, + 284, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 284, + 505, + 297 + ], + "score": 1.0, + "content": "Note that the ancestor shortest paths and next sibling shortest paths are required because", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 295, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 506, + 309 + ], + "score": 1.0, + "content": "treating the AST as a normal graph leads to ambiguity. In a graph, the neighbors of a node", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "have no ordering; however in the AST, the order of the children of a node reflects their", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 318, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 141, + 318, + 506, + 330 + ], + "score": 1.0, + "content": "order in the code. Therefore, we explicitly include the next sibling shortest paths. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "ancestor shortest paths would not be required if we treated the AST as a directed graph; in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "this case, however, a leaf node could not reach any other node in the AST, and therefore", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 141, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "both PPR and shortest path length are not useful in this case. Therefore, we model the AST", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 361, + 421, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 421, + 375 + ], + "score": 1.0, + "content": "as undirected and inject the ancestor / child edges to avoid ambiguity.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 141, + 284, + 506, + 375 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 377, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 128, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 128, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "3. Distance values are binned into 32 bins using area-based exponential binning with a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 388, + 371, + 401 + ], + "score": 1.0, + "content": "growth factor of 1.3, i.e., the area of a bin’s rectangle (", + "type": "text" + }, + { + "bbox": [ + 371, + 390, + 378, + 398 + ], + "score": 0.48, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 388, + 428, + 401 + ], + "score": 1.0, + "content": ": bin range,", + "type": "text" + }, + { + "bbox": [ + 429, + 390, + 435, + 400 + ], + "score": 0.62, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 388, + 505, + 401 + ], + "score": 1.0, + "content": ": number of val-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 142, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "ues in bin) will be approximately 1.3 times bigger for the next bin (going away from the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "bin that contains the zero value). Additionally, for discrete distance measures (such as se-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "quence distance or shortest path length), we hard-code 9 values around 0 to have their own", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 142, + 432, + 361, + 444 + ], + "score": 1.0, + "content": "bins. For instance, on the sequence distance the values", + "type": "text" + }, + { + "bbox": [ + 362, + 432, + 420, + 443 + ], + "score": 0.9, + "content": "- 4 , - 3 , \\dots , 4", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "have their individual", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 441, + 407, + 457 + ], + "spans": [ + { + "bbox": [ + 141, + 441, + 407, + 457 + ], + "score": 1.0, + "content": "bins, and around those values we employ the exponential binning.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26, + "bbox_fs": [ + 128, + 376, + 506, + 457 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 459, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 128, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 128, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "4. Punctuation tokens (such as points or brackets) are removed from the input sequence, as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 141, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "experiments showed that their presence does not improve performance but slows down", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 480, + 278, + 493 + ], + "spans": [ + { + "bbox": [ + 142, + 480, + 278, + 493 + ], + "score": 1.0, + "content": "training due to bigger input sizes.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 128, + 457, + 505, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 496, + 504, + 529 + ], + "lines": [ + { + "bbox": [ + 129, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 129, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "5. Snippets that are longer than MAX NUM TOKENS after punctuation tokens are removed are", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 142, + 507, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 142, + 507, + 505, + 519 + ], + "score": 1.0, + "content": "discarded from the training set. Throughout our experiments, we use MAX NUM TOKENS", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 142, + 518, + 468, + 529 + ], + "spans": [ + { + "bbox": [ + 142, + 518, + 173, + 529 + ], + "score": 0.76, + "content": "= ~ 5 1 2", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 518, + 424, + 529 + ], + "score": 1.0, + "content": ". During evaluation on the test set, we use MAX NUM TOKENS", + "type": "text" + }, + { + "bbox": [ + 424, + 518, + 464, + 528 + ], + "score": 0.76, + "content": "= ~ 1 0 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 518, + 468, + 529 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 129, + 495, + 505, + 529 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 544, + 286, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 288, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 288, + 556 + ], + "score": 1.0, + "content": "A.4 INPUT EMBEDDINGS TO THE MODEL", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 578 + ], + "score": 1.0, + "content": "Besides its five subtokens (e.g., [‘get’, ‘data’, ‘[PAD]’, ‘[PAD]’, ‘[PAD]’]),", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 589 + ], + "score": 1.0, + "content": "each input token has a token type (coming from the Pygments tokenizer) and an AST node type.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 599 + ], + "score": 1.0, + "content": "The AST node type is the type of the node assigned to each respective token, as described in Section", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 610 + ], + "score": 1.0, + "content": "A.3.2. We concatenate the embeddings of the five subtokens, the token type, and the AST node", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "type. Then, we apply a linear layer (without activation function) to project down to the model’s", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 619, + 200, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 200, + 632 + ], + "score": 1.0, + "content": "embedding dimension.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 564, + 506, + 632 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 644, + 275, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 277, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 277, + 657 + ], + "score": 1.0, + "content": "A.5 INPUT TO THE GREAT BASELINE", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "As mentioned in the main text, we also compare with GREAT Hellendoorn et al. (2020). Since", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "their preprocessing pipeline is proprietary and could not be shared with us even after contacting the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "authors, we provide to GREAT the same AST distances as our model. Since GREAT uses edges in-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "stead of distances to encode relations in the Structure, we essentially threshold the ancestor, sibling,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "and shortest-paths distances and provide the edges where the distances are equal to 1 (including their", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 210, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 210, + 733 + ], + "score": 1.0, + "content": "edge types) to the model.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 665, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 97, + 296, + 198 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 182, + 86, + 280, + 96 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 182, + 86, + 281, + 97 + ], + "spans": [ + { + "bbox": [ + 182, + 86, + 281, + 97 + ], + "score": 1.0, + "content": "(a) CODE TRANSFORMER", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 167, + 97, + 296, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 97, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 167, + 97, + 296, + 198 + ], + "score": 0.968, + "html": "
HyperparameterValue
ActivationGELU
Input Nonlinearitytanh
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
", + "type": "table", + "image_path": "26b079a48f72f11ac5901761fadc58949615f228994d83e0ca015c1b803de70e.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 167, + 97, + 296, + 147.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 167, + 147.5, + 296, + 198.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 210, + 207, + 401, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 209, + 204, + 401, + 222 + ], + "spans": [ + { + "bbox": [ + 209, + 204, + 401, + 222 + ], + "score": 1.0, + "content": "Table 7: Code Summarization hyperparameters", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 318, + 108, + 438, + 196 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 313, + 86, + 444, + 106 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 312, + 85, + 446, + 98 + ], + "spans": [ + { + "bbox": [ + 312, + 85, + 446, + 98 + ], + "score": 1.0, + "content": "(b) GREAT (Hellendoorn et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 311, + 95, + 338, + 108 + ], + "spans": [ + { + "bbox": [ + 311, + 95, + 338, + 108 + ], + "score": 1.0, + "content": "2020)", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 318, + 108, + 438, + 196 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 318, + 108, + 438, + 196 + ], + "spans": [ + { + "bbox": [ + 318, + 108, + 438, + 196 + ], + "score": 0.969, + "html": "
HyperparameterValue
ActivationGELU
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
", + "type": "table", + "image_path": "d450c7940cdf06f24a236ed0b2ea7649815a8d485cbb72d5b067d40223c84355.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 318, + 108, + 438, + 152.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 318, + 152.0, + 438, + 196.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 108, + 240, + 231, + 251 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 232, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 232, + 252 + ], + "score": 1.0, + "content": "A.6 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 260, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "Table 7 shows hyperparameters of our models for code summarization. For all our experiments, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "use a Transformer Decoder with one layer and teacher forcing to generate 6 output sub tokens. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 465, + 295 + ], + "score": 1.0, + "content": "also employ label smoothing of 0.1. As optimizer, we use Adam with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 465, + 282, + 487, + 293 + ], + "score": 0.9, + "content": "8 e ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 174, + 307 + ], + "score": 1.0, + "content": "weight decay of", + "type": "text" + }, + { + "bbox": [ + 174, + 293, + 195, + 304 + ], + "score": 0.9, + "content": "3 e ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 292, + 506, + 307 + ], + "score": 1.0, + "content": ". Batch size during training is 8 with a simulated batch size of 128 achieved", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 305, + 212, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 212, + 316 + ], + "score": 1.0, + "content": "by gradient accumulation.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 503, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "Apart from comparing the CODE TRANSFORMER to baselines, we performed the following hyper-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 344, + 285, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 285, + 356 + ], + "score": 1.0, + "content": "parameter comparisons and ablation studies:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 133, + 365, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 136, + 364, + 303, + 376 + ], + "spans": [ + { + "bbox": [ + 136, + 364, + 303, + 376 + ], + "score": 1.0, + "content": "• CODE TRANSFORMER (structure-only)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 375, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 506, + 387 + ], + "score": 1.0, + "content": "Using only AST information as input, i.e., masking all tokens that do not correspond to a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 387, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 505, + 398 + ], + "score": 1.0, + "content": "leaf of the AST, and removing the token distance as a relation to be used by the model.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 397, + 332, + 409 + ], + "spans": [ + { + "bbox": [ + 141, + 397, + 332, + 409 + ], + "score": 1.0, + "content": "Further, token types are not fed into the model.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 136, + 411, + 297, + 424 + ], + "spans": [ + { + "bbox": [ + 136, + 411, + 297, + 424 + ], + "score": 1.0, + "content": "• CODE TRANSFORMER (context-only)", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 423, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 423, + 505, + 434 + ], + "score": 1.0, + "content": "Here, we do not include any information on the AST (i.e. node types and distances on the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 141, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "AST). This is effectively the XLNet backbone plus encoding of the token type returned by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 445, + 198, + 457 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 198, + 457 + ], + "score": 1.0, + "content": "the tokenizer.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 136, + 459, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 136, + 459, + 288, + 472 + ], + "score": 1.0, + "content": "• CODE TRANSFORMER (Max-Dist.)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 471, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 141, + 471, + 506, + 483 + ], + "score": 1.0, + "content": "Applying a Max Distance Mask of 5 to the shortest paths distance (i.e., model cannot see", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 140, + 481, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 140, + 481, + 505, + 495 + ], + "score": 1.0, + "content": "a node that is more than 5 hops away no matter how small the other distances are). Early", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 494, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 142, + 494, + 505, + 504 + ], + "score": 1.0, + "content": "results showed that, as expected, results deteriorate substantially when limiting our model’s", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 504, + 412, + 515 + ], + "spans": [ + { + "bbox": [ + 141, + 504, + 412, + 515 + ], + "score": 1.0, + "content": "receptive field. Hence, we do not include these results in this work.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 132, + 518, + 492, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 492, + 532 + ], + "score": 1.0, + "content": "• Using 16 and 64 bins instead of 32 bins. This had no noticeable effect on performance.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 109, + 544, + 283, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 284, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 284, + 556 + ], + "score": 1.0, + "content": "A.7 CODE SUMMARIZATION EXAMPLES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "score": 1.0, + "content": "In the Tables 8, 9, 10, 11, 12, 13, 14 and 15 we present example functions from the Java-small", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 574, + 405, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 405, + 588 + ], + "score": 1.0, + "content": "dataset along with the different models’ predictions for the function name.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 97, + 296, + 198 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 182, + 86, + 280, + 96 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 182, + 86, + 281, + 97 + ], + "spans": [ + { + "bbox": [ + 182, + 86, + 281, + 97 + ], + "score": 1.0, + "content": "(a) CODE TRANSFORMER", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 167, + 97, + 296, + 198 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 97, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 167, + 97, + 296, + 198 + ], + "score": 0.968, + "html": "
HyperparameterValue
ActivationGELU
Input Nonlinearitytanh
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
", + "type": "table", + "image_path": "26b079a48f72f11ac5901761fadc58949615f228994d83e0ca015c1b803de70e.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 167, + 97, + 296, + 147.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 167, + 147.5, + 296, + 198.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 210, + 207, + 401, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 209, + 204, + 401, + 222 + ], + "spans": [ + { + "bbox": [ + 209, + 204, + 401, + 222 + ], + "score": 1.0, + "content": "Table 7: Code Summarization hyperparameters", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 318, + 108, + 438, + 196 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 313, + 86, + 444, + 106 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 312, + 85, + 446, + 98 + ], + "spans": [ + { + "bbox": [ + 312, + 85, + 446, + 98 + ], + "score": 1.0, + "content": "(b) GREAT (Hellendoorn et al.,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 311, + 95, + 338, + 108 + ], + "spans": [ + { + "bbox": [ + 311, + 95, + 338, + 108 + ], + "score": 1.0, + "content": "2020)", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 318, + 108, + 438, + 196 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 318, + 108, + 438, + 196 + ], + "spans": [ + { + "bbox": [ + 318, + 108, + 438, + 196 + ], + "score": 0.969, + "html": "
HyperparameterValue
ActivationGELU
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
", + "type": "table", + "image_path": "d450c7940cdf06f24a236ed0b2ea7649815a8d485cbb72d5b067d40223c84355.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 318, + 108, + 438, + 152.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 318, + 152.0, + 438, + 196.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 108, + 240, + 231, + 251 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 232, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 232, + 252 + ], + "score": 1.0, + "content": "A.6 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 260, + 505, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "Table 7 shows hyperparameters of our models for code summarization. For all our experiments, we", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "use a Transformer Decoder with one layer and teacher forcing to generate 6 output sub tokens. We", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 465, + 295 + ], + "score": 1.0, + "content": "also employ label smoothing of 0.1. As optimizer, we use Adam with a learning rate of", + "type": "text" + }, + { + "bbox": [ + 465, + 282, + 487, + 293 + ], + "score": 0.9, + "content": "8 e ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 292, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 174, + 307 + ], + "score": 1.0, + "content": "weight decay of", + "type": "text" + }, + { + "bbox": [ + 174, + 293, + 195, + 304 + ], + "score": 0.9, + "content": "3 e ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 292, + 506, + 307 + ], + "score": 1.0, + "content": ". Batch size during training is 8 with a simulated batch size of 128 achieved", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 305, + 212, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 212, + 316 + ], + "score": 1.0, + "content": "by gradient accumulation.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 260, + 506, + 316 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 503, + 354 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "Apart from comparing the CODE TRANSFORMER to baselines, we performed the following hyper-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 344, + 285, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 285, + 356 + ], + "score": 1.0, + "content": "parameter comparisons and ablation studies:", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 106, + 331, + 505, + 356 + ] + }, + { + "type": "list", + "bbox": [ + 133, + 365, + 505, + 531 + ], + "lines": [ + { + "bbox": [ + 136, + 364, + 303, + 376 + ], + "spans": [ + { + "bbox": [ + 136, + 364, + 303, + 376 + ], + "score": 1.0, + "content": "• CODE TRANSFORMER (structure-only)", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 141, + 375, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 375, + 506, + 387 + ], + "score": 1.0, + "content": "Using only AST information as input, i.e., masking all tokens that do not correspond to a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 387, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 141, + 387, + 505, + 398 + ], + "score": 1.0, + "content": "leaf of the AST, and removing the token distance as a relation to be used by the model.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 397, + 332, + 409 + ], + "spans": [ + { + "bbox": [ + 141, + 397, + 332, + 409 + ], + "score": 1.0, + "content": "Further, token types are not fed into the model.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 136, + 411, + 297, + 424 + ], + "spans": [ + { + "bbox": [ + 136, + 411, + 297, + 424 + ], + "score": 1.0, + "content": "• CODE TRANSFORMER (context-only)", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 141, + 423, + 505, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 423, + 505, + 434 + ], + "score": 1.0, + "content": "Here, we do not include any information on the AST (i.e. node types and distances on the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 141, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "AST). This is effectively the XLNet backbone plus encoding of the token type returned by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 445, + 198, + 457 + ], + "spans": [ + { + "bbox": [ + 142, + 445, + 198, + 457 + ], + "score": 1.0, + "content": "the tokenizer.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 136, + 459, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 136, + 459, + 288, + 472 + ], + "score": 1.0, + "content": "• CODE TRANSFORMER (Max-Dist.)", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 141, + 471, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 141, + 471, + 506, + 483 + ], + "score": 1.0, + "content": "Applying a Max Distance Mask of 5 to the shortest paths distance (i.e., model cannot see", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 140, + 481, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 140, + 481, + 505, + 495 + ], + "score": 1.0, + "content": "a node that is more than 5 hops away no matter how small the other distances are). Early", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 494, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 142, + 494, + 505, + 504 + ], + "score": 1.0, + "content": "results showed that, as expected, results deteriorate substantially when limiting our model’s", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 504, + 412, + 515 + ], + "spans": [ + { + "bbox": [ + 141, + 504, + 412, + 515 + ], + "score": 1.0, + "content": "receptive field. Hence, we do not include these results in this work.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 518, + 492, + 532 + ], + "spans": [ + { + "bbox": [ + 132, + 518, + 492, + 532 + ], + "score": 1.0, + "content": "• Using 16 and 64 bins instead of 32 bins. This had no noticeable effect on performance.", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 22.5, + "bbox_fs": [ + 132, + 364, + 506, + 532 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 544, + 283, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 543, + 284, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 284, + 556 + ], + "score": 1.0, + "content": "A.7 CODE SUMMARIZATION EXAMPLES", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 504, + 576 + ], + "score": 1.0, + "content": "In the Tables 8, 9, 10, 11, 12, 13, 14 and 15 we present example functions from the Java-small", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 574, + 405, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 405, + 588 + ], + "score": 1.0, + "content": "dataset along with the different models’ predictions for the function name.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 564, + 504, + 588 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 94, + 242, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 244, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 244, + 105 + ], + "score": 1.0, + "content": "public Summation next() {", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 105, + 225, + 114 + ], + "spans": [ + { + "bbox": [ + 128, + 105, + 195, + 114 + ], + "score": 1.0, + "content": "return parts", + "type": "text" + }, + { + "bbox": [ + 195, + 105, + 219, + 114 + ], + "score": 0.41, + "content": "[ \\dot { 1 } + + ]", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 105, + 225, + 114 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 117, + 113, + 124 + ], + "spans": [ + { + "bbox": [ + 108, + 117, + 113, + 124 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 231, + 128, + 378, + 205 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 231, + 128, + 378, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 231, + 128, + 378, + 205 + ], + "spans": [ + { + "bbox": [ + 231, + 128, + 378, + 205 + ], + "score": 0.969, + "html": "
ModelPrediction
GREAT get x map
code2seq get parts
Ours w/o structureget
CODE TRANSFORMERget next
Ground Truthnext
", + "type": "table", + "image_path": "04303ffb667966783611a78b38db4a06a17d54b3303576e3d6f2082eb78c4e3f.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 231, + 128, + 378, + 166.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 231, + 166.5, + 378, + 205.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 214, + 504, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "Table 8: The CODE TRANSFORMER is the only model to correctly identify the notion of getting the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 226, + 151, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 151, + 237 + ], + "score": 1.0, + "content": "next entry.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 272, + 453, + 370 + ], + "lines": [ + { + "bbox": [ + 104, + 272, + 453, + 285 + ], + "spans": [ + { + "bbox": [ + 104, + 272, + 453, + 285 + ], + "score": 1.0, + "content": "private Path findCacheFile(Path[] cacheFiles, String fileName) {", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 282, + 400, + 294 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 210, + 294 + ], + "score": 1.0, + "content": "if (cacheFiles", + "type": "text" + }, + { + "bbox": [ + 210, + 284, + 222, + 292 + ], + "score": 0.69, + "content": "! =", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 282, + 362, + 294 + ], + "score": 1.0, + "content": "null && cacheFiles.length", + "type": "text" + }, + { + "bbox": [ + 363, + 283, + 383, + 293 + ], + "score": 0.41, + "content": "> 0", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 282, + 400, + 294 + ], + "score": 1.0, + "content": ") {", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 149, + 292, + 314, + 303 + ], + "spans": [ + { + "bbox": [ + 149, + 292, + 169, + 303 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 172, + 293, + 314, + 303 + ], + "score": 1.0, + "content": "(Path file : cacheFiles) {", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 171, + 303, + 378, + 313 + ], + "spans": [ + { + "bbox": [ + 171, + 303, + 185, + 312 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 187, + 303, + 378, + 313 + ], + "score": 1.0, + "content": "(file.getName().equals(fileName)) {", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 192, + 312, + 259, + 324 + ], + "spans": [ + { + "bbox": [ + 192, + 312, + 259, + 324 + ], + "score": 1.0, + "content": "return file;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 171, + 324, + 178, + 333 + ], + "spans": [ + { + "bbox": [ + 171, + 324, + 178, + 333 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 128, + 352, + 192, + 363 + ], + "spans": [ + { + "bbox": [ + 128, + 352, + 192, + 363 + ], + "score": 1.0, + "content": "return null;", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 108, + 365, + 113, + 371 + ], + "spans": [ + { + "bbox": [ + 108, + 365, + 113, + 371 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 10.5 + }, + { + "type": "table", + "bbox": [ + 225, + 376, + 385, + 454 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 225, + 376, + 385, + 454 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 225, + 376, + 385, + 454 + ], + "spans": [ + { + "bbox": [ + 225, + 376, + 385, + 454 + ], + "score": 0.908, + "html": "
ModelPrediction
GREATget path
code2seqfind file
Ours w/o structureget file
CODE TRANSFORMERfind cache
Ground Truthfind cache file
", + "type": "table", + "image_path": "7fce10876c7dbfae8cd1529bce06e79b9baef638e4b20f560a956fd0593414a7.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 225, + 376, + 385, + 387.14285714285717 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 225, + 387.14285714285717, + 385, + 398.28571428571433 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 225, + 398.28571428571433, + 385, + 409.4285714285715 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 225, + 409.4285714285715, + 385, + 420.57142857142867 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 225, + 420.57142857142867, + 385, + 431.71428571428584 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 225, + 431.71428571428584, + 385, + 442.857142857143 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 225, + 442.857142857143, + 385, + 454.00000000000017 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 506, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Table 9: The CODE TRANSFORMER is the only model to both recognize that the task is to find a file", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "as well as the fact that it is about the cache. However, it did not correctly predict the file part of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 483, + 165, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 165, + 496 + ], + "score": 1.0, + "content": "method name.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 458, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 398, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 398, + 543 + ], + "score": 1.0, + "content": "public int compare(Pair p1,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 208, + 542, + 410, + 553 + ], + "spans": [ + { + "bbox": [ + 208, + 542, + 410, + 553 + ], + "score": 1.0, + "content": "Pair p2) {", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 551, + 270, + 563 + ], + "spans": [ + { + "bbox": [ + 127, + 551, + 182, + 563 + ], + "score": 1.0, + "content": "LoggedJob", + "type": "text" + }, + { + "bbox": [ + 182, + 552, + 221, + 562 + ], + "score": 0.57, + "content": "\\ j 1 \\ = \\ \\mathtt { p 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 551, + 270, + 563 + ], + "score": 1.0, + "content": ".first();", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 562, + 270, + 573 + ], + "spans": [ + { + "bbox": [ + 127, + 562, + 182, + 573 + ], + "score": 1.0, + "content": "LoggedJob", + "type": "text" + }, + { + "bbox": [ + 182, + 563, + 221, + 572 + ], + "score": 0.51, + "content": "\\begin{array} { r } { \\dot { \\bf { \\ j } } 2 { \\bf { \\nu } } = { \\bf { \\ p } } 2 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 562, + 270, + 573 + ], + "score": 1.0, + "content": ".first();", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 129, + 572, + 415, + 583 + ], + "spans": [ + { + "bbox": [ + 129, + 573, + 163, + 582 + ], + "score": 1.0, + "content": "return", + "type": "text" + }, + { + "bbox": [ + 166, + 572, + 272, + 583 + ], + "score": 1.0, + "content": "(j1.getSubmitTime()", + "type": "text" + }, + { + "bbox": [ + 272, + 573, + 282, + 581 + ], + "score": 0.54, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 572, + 415, + 583 + ], + "score": 1.0, + "content": "j2.getSubmitTime()) ? -1", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 170, + 582, + 457, + 593 + ], + "spans": [ + { + "bbox": [ + 170, + 582, + 288, + 593 + ], + "score": 1.0, + "content": ": (j1.getSubmitTime()", + "type": "text" + }, + { + "bbox": [ + 288, + 583, + 303, + 592 + ], + "score": 0.35, + "content": "= =", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 582, + 457, + 593 + ], + "score": 1.0, + "content": "j2.getSubmitTime()) ? 0 : 1;", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 128, + 593, + 135, + 602 + ], + "spans": [ + { + "bbox": [ + 128, + 593, + 135, + 602 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 221, + 606, + 388, + 684 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 221, + 606, + 388, + 684 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 221, + 606, + 388, + 684 + ], + "spans": [ + { + "bbox": [ + 221, + 606, + 388, + 684 + ], + "score": 0.96, + "html": "
ModelPrediction
GREATrun
code2seq get submit time
Ours w/o structurecompare
CODE TRANSFORMERcompare
Ground Truthcompare
", + "type": "table", + "image_path": "fc7647e7d39190fcddd6003a2621a01161522b4db1837436041dcd5808f7523b.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 221, + 606, + 388, + 617.1428571428571 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 221, + 617.1428571428571, + 388, + 628.2857142857142 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 221, + 628.2857142857142, + 388, + 639.4285714285713 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 221, + 639.4285714285713, + 388, + 650.5714285714284 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 221, + 650.5714285714284, + 388, + 661.7142857142856 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 221, + 661.7142857142856, + 388, + 672.8571428571427 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 221, + 672.8571428571427, + 388, + 683.9999999999998 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 705 + ], + "score": 1.0, + "content": "Table 10: The CODE TRANSFORMER and the its context-only variant are the only models correctly", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 702, + 333, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 333, + 716 + ], + "score": 1.0, + "content": "recognizing the ‘compare’ template in the method body.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 94, + 242, + 124 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 244, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 244, + 105 + ], + "score": 1.0, + "content": "public Summation next() {", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 105, + 225, + 114 + ], + "spans": [ + { + "bbox": [ + 128, + 105, + 195, + 114 + ], + "score": 1.0, + "content": "return parts", + "type": "text" + }, + { + "bbox": [ + 195, + 105, + 219, + 114 + ], + "score": 0.41, + "content": "[ \\dot { 1 } + + ]", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 105, + 225, + 114 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 117, + 113, + 124 + ], + "spans": [ + { + "bbox": [ + 108, + 117, + 113, + 124 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 94, + 244, + 124 + ] + }, + { + "type": "table", + "bbox": [ + 231, + 128, + 378, + 205 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 231, + 128, + 378, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 231, + 128, + 378, + 205 + ], + "spans": [ + { + "bbox": [ + 231, + 128, + 378, + 205 + ], + "score": 0.969, + "html": "
ModelPrediction
GREAT get x map
code2seq get parts
Ours w/o structureget
CODE TRANSFORMERget next
Ground Truthnext
", + "type": "table", + "image_path": "04303ffb667966783611a78b38db4a06a17d54b3303576e3d6f2082eb78c4e3f.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 231, + 128, + 378, + 166.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 231, + 166.5, + 378, + 205.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 214, + 504, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 226 + ], + "score": 1.0, + "content": "Table 8: The CODE TRANSFORMER is the only model to correctly identify the notion of getting the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 226, + 151, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 151, + 237 + ], + "score": 1.0, + "content": "next entry.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + } + ], + "index": 4.5 + }, + { + "type": "list", + "bbox": [ + 106, + 272, + 453, + 370 + ], + "lines": [ + { + "bbox": [ + 104, + 272, + 453, + 285 + ], + "spans": [ + { + "bbox": [ + 104, + 272, + 453, + 285 + ], + "score": 1.0, + "content": "private Path findCacheFile(Path[] cacheFiles, String fileName) {", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 282, + 400, + 294 + ], + "spans": [ + { + "bbox": [ + 127, + 282, + 210, + 294 + ], + "score": 1.0, + "content": "if (cacheFiles", + "type": "text" + }, + { + "bbox": [ + 210, + 284, + 222, + 292 + ], + "score": 0.69, + "content": "! =", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 282, + 362, + 294 + ], + "score": 1.0, + "content": "null && cacheFiles.length", + "type": "text" + }, + { + "bbox": [ + 363, + 283, + 383, + 293 + ], + "score": 0.41, + "content": "> 0", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 282, + 400, + 294 + ], + "score": 1.0, + "content": ") {", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 149, + 292, + 314, + 303 + ], + "spans": [ + { + "bbox": [ + 149, + 292, + 169, + 303 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 172, + 293, + 314, + 303 + ], + "score": 1.0, + "content": "(Path file : cacheFiles) {", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 171, + 303, + 378, + 313 + ], + "spans": [ + { + "bbox": [ + 171, + 303, + 185, + 312 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 187, + 303, + 378, + 313 + ], + "score": 1.0, + "content": "(file.getName().equals(fileName)) {", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 192, + 312, + 259, + 324 + ], + "spans": [ + { + "bbox": [ + 192, + 312, + 259, + 324 + ], + "score": 1.0, + "content": "return file;", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 171, + 324, + 178, + 333 + ], + "spans": [ + { + "bbox": [ + 171, + 324, + 178, + 333 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 352, + 192, + 363 + ], + "spans": [ + { + "bbox": [ + 128, + 352, + 192, + 363 + ], + "score": 1.0, + "content": "return null;", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 365, + 113, + 371 + ], + "spans": [ + { + "bbox": [ + 108, + 365, + 113, + 371 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 272, + 453, + 371 + ] + }, + { + "type": "table", + "bbox": [ + 225, + 376, + 385, + 454 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 225, + 376, + 385, + 454 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 225, + 376, + 385, + 454 + ], + "spans": [ + { + "bbox": [ + 225, + 376, + 385, + 454 + ], + "score": 0.908, + "html": "
ModelPrediction
GREATget path
code2seqfind file
Ours w/o structureget file
CODE TRANSFORMERfind cache
Ground Truthfind cache file
", + "type": "table", + "image_path": "7fce10876c7dbfae8cd1529bce06e79b9baef638e4b20f560a956fd0593414a7.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 225, + 376, + 385, + 387.14285714285717 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 225, + 387.14285714285717, + 385, + 398.28571428571433 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 225, + 398.28571428571433, + 385, + 409.4285714285715 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 225, + 409.4285714285715, + 385, + 420.57142857142867 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 225, + 420.57142857142867, + 385, + 431.71428571428584 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 225, + 431.71428571428584, + 385, + 442.857142857143 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 225, + 442.857142857143, + 385, + 454.00000000000017 + ], + "spans": [], + "index": 21 + } + ] + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 506, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 474 + ], + "score": 1.0, + "content": "Table 9: The CODE TRANSFORMER is the only model to both recognize that the task is to find a file", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 486 + ], + "score": 1.0, + "content": "as well as the fact that it is about the cache. However, it did not correctly predict the file part of the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 483, + 165, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 165, + 496 + ], + "score": 1.0, + "content": "method name.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 461, + 505, + 496 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 532, + 458, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 531, + 398, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 398, + 543 + ], + "score": 1.0, + "content": "public int compare(Pair p1,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 208, + 542, + 410, + 553 + ], + "spans": [ + { + "bbox": [ + 208, + 542, + 410, + 553 + ], + "score": 1.0, + "content": "Pair p2) {", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 551, + 270, + 563 + ], + "spans": [ + { + "bbox": [ + 127, + 551, + 182, + 563 + ], + "score": 1.0, + "content": "LoggedJob", + "type": "text" + }, + { + "bbox": [ + 182, + 552, + 221, + 562 + ], + "score": 0.57, + "content": "\\ j 1 \\ = \\ \\mathtt { p 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 551, + 270, + 563 + ], + "score": 1.0, + "content": ".first();", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 127, + 562, + 270, + 573 + ], + "spans": [ + { + "bbox": [ + 127, + 562, + 182, + 573 + ], + "score": 1.0, + "content": "LoggedJob", + "type": "text" + }, + { + "bbox": [ + 182, + 563, + 221, + 572 + ], + "score": 0.51, + "content": "\\begin{array} { r } { \\dot { \\bf { \\ j } } 2 { \\bf { \\nu } } = { \\bf { \\ p } } 2 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 562, + 270, + 573 + ], + "score": 1.0, + "content": ".first();", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 129, + 572, + 415, + 583 + ], + "spans": [ + { + "bbox": [ + 129, + 573, + 163, + 582 + ], + "score": 1.0, + "content": "return", + "type": "text" + }, + { + "bbox": [ + 166, + 572, + 272, + 583 + ], + "score": 1.0, + "content": "(j1.getSubmitTime()", + "type": "text" + }, + { + "bbox": [ + 272, + 573, + 282, + 581 + ], + "score": 0.54, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 572, + 415, + 583 + ], + "score": 1.0, + "content": "j2.getSubmitTime()) ? -1", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 170, + 582, + 457, + 593 + ], + "spans": [ + { + "bbox": [ + 170, + 582, + 288, + 593 + ], + "score": 1.0, + "content": ": (j1.getSubmitTime()", + "type": "text" + }, + { + "bbox": [ + 288, + 583, + 303, + 592 + ], + "score": 0.35, + "content": "= =", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 582, + 457, + 593 + ], + "score": 1.0, + "content": "j2.getSubmitTime()) ? 0 : 1;", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 128, + 593, + 135, + 602 + ], + "spans": [ + { + "bbox": [ + 128, + 593, + 135, + 602 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 531, + 457, + 602 + ] + }, + { + "type": "table", + "bbox": [ + 221, + 606, + 388, + 684 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 221, + 606, + 388, + 684 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 221, + 606, + 388, + 684 + ], + "spans": [ + { + "bbox": [ + 221, + 606, + 388, + 684 + ], + "score": 0.96, + "html": "
ModelPrediction
GREATrun
code2seq get submit time
Ours w/o structurecompare
CODE TRANSFORMERcompare
Ground Truthcompare
", + "type": "table", + "image_path": "fc7647e7d39190fcddd6003a2621a01161522b4db1837436041dcd5808f7523b.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 221, + 606, + 388, + 617.1428571428571 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 221, + 617.1428571428571, + 388, + 628.2857142857142 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 221, + 628.2857142857142, + 388, + 639.4285714285713 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 221, + 639.4285714285713, + 388, + 650.5714285714284 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 221, + 650.5714285714284, + 388, + 661.7142857142856 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 221, + 661.7142857142856, + 388, + 672.8571428571427 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 221, + 672.8571428571427, + 388, + 683.9999999999998 + ], + "spans": [], + "index": 38 + } + ] + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 691, + 504, + 714 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 705 + ], + "score": 1.0, + "content": "Table 10: The CODE TRANSFORMER and the its context-only variant are the only models correctly", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 702, + 333, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 333, + 716 + ], + "score": 1.0, + "content": "recognizing the ‘compare’ template in the method body.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 689, + 505, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 92, + 367, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 345, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 345, + 104 + ], + "score": 1.0, + "content": "public static MNTPROC fromValue(int value) {", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 101, + 367, + 114 + ], + "spans": [ + { + "bbox": [ + 126, + 101, + 181, + 114 + ], + "score": 1.0, + "content": "if (value", + "type": "text" + }, + { + "bbox": [ + 181, + 103, + 200, + 113 + ], + "score": 0.4, + "content": "< ~ 0", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 101, + 250, + 114 + ], + "score": 1.0, + "content": "|| value", + "type": "text" + }, + { + "bbox": [ + 250, + 104, + 266, + 112 + ], + "score": 0.76, + "content": "> =", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 101, + 367, + 114 + ], + "score": 1.0, + "content": "values().length) {", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 113, + 215, + 124 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 215, + 124 + ], + "score": 1.0, + "content": "return null;", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 126, + 134, + 131 + ], + "spans": [ + { + "bbox": [ + 130, + 126, + 134, + 131 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 132, + 253, + 144 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 253, + 144 + ], + "score": 1.0, + "content": "return values()[value];", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 145, + 112, + 150 + ], + "spans": [ + { + "bbox": [ + 109, + 145, + 112, + 150 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 229, + 156, + 379, + 234 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 229, + 156, + 379, + 234 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 229, + 156, + 379, + 234 + ], + "spans": [ + { + "bbox": [ + 229, + 156, + 379, + 234 + ], + "score": 0.966, + "html": "
ModelPrediction
GREAT get value
code2seq get value
Ours w/o structureto
CODE TRANSFORMERfrom value
Ground Truthfrom value
", + "type": "table", + "image_path": "c731fa631d72aabd1f5e32632f679a09b671309e6980e8efeb5fd4f3681003b9.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 156, + 379, + 195.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 229, + 195.0, + 379, + 234.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 105, + 242, + 505, + 265 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "Table 11: The CODE TRANSFORMER is the only model to recognize that the snippet is similar to a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 253, + 334, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 334, + 266 + ], + "score": 1.0, + "content": "static factory method which is often preceded with from.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 429, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 430, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 430, + 310 + ], + "score": 1.0, + "content": "private Iterable listRootBlobs(String aPrefix,", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 346, + 308, + 580, + 348 + ], + "lines": [ + { + "bbox": [ + 347, + 307, + 495, + 319 + ], + "spans": [ + { + "bbox": [ + 347, + 307, + 495, + 319 + ], + "score": 1.0, + "content": "boolean useFlatBlobListing,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 317, + 581, + 329 + ], + "spans": [ + { + "bbox": [ + 348, + 317, + 581, + 329 + ], + "score": 1.0, + "content": "EnumSet listingDetails,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 328, + 495, + 339 + ], + "spans": [ + { + "bbox": [ + 348, + 328, + 495, + 339 + ], + "score": 1.0, + "content": "BlobRequestOptions options,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 338, + 494, + 349 + ], + "spans": [ + { + "bbox": [ + 348, + 338, + 494, + 349 + ], + "score": 1.0, + "content": "OperationContext opContext)", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 315, + 349, + 568, + 358 + ], + "lines": [ + { + "bbox": [ + 321, + 347, + 566, + 360 + ], + "spans": [ + { + "bbox": [ + 321, + 347, + 566, + 360 + ], + "score": 1.0, + "content": "throws StorageException, URISyntaxException {", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 114, + 359, + 579, + 389 + ], + "lines": [ + { + "bbox": [ + 127, + 357, + 581, + 369 + ], + "spans": [ + { + "bbox": [ + 127, + 357, + 320, + 369 + ], + "score": 1.0, + "content": "CloudBlobDirectoryWrapper directory", + "type": "text" + }, + { + "bbox": [ + 320, + 359, + 330, + 367 + ], + "score": 0.68, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 357, + 581, + 369 + ], + "score": 1.0, + "content": "this.container.getDirectoryReference(aPrefix);", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 366, + 410, + 380 + ], + "spans": [ + { + "bbox": [ + 126, + 366, + 410, + 380 + ], + "score": 1.0, + "content": "return directory.listBlobs(null, useFlatBlobListing,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 272, + 377, + 468, + 389 + ], + "spans": [ + { + "bbox": [ + 272, + 377, + 468, + 389 + ], + "score": 1.0, + "content": "listingDetails, options, opContext);", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 113, + 398 + ], + "lines": [ + { + "bbox": [ + 107, + 390, + 113, + 398 + ], + "spans": [ + { + "bbox": [ + 107, + 390, + 113, + 398 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 215, + 402, + 394, + 479 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 215, + 402, + 394, + 479 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 215, + 402, + 394, + 479 + ], + "spans": [ + { + "bbox": [ + 215, + 402, + 394, + 479 + ], + "score": 0.972, + "html": "
ModelPrediction
GREATlist blobs
code2seqlist blobs
Ours w/o structurelist blobs
CODE TRANSFORMERlist blobs by prefix
Ground Truthlist root blobs
", + "type": "table", + "image_path": "4184371f3c5b40e5706dc3536eb808edba7ea9235136630720ca50cbd799e93d.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 402, + 394, + 414.8333333333333 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 215, + 414.8333333333333, + 394, + 427.66666666666663 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 215, + 427.66666666666663, + 394, + 440.49999999999994 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 215, + 440.49999999999994, + 394, + 453.33333333333326 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 215, + 453.33333333333326, + 394, + 466.1666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 215, + 466.1666666666666, + 394, + 478.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 105, + 487, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 501 + ], + "score": 1.0, + "content": "Table 12: All models could correctly identify the listBlobs() call in the return statement. However,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 497, + 499, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 499, + 511 + ], + "score": 1.0, + "content": "the CODE TRANSFORMER additionally comprehended that the specified prefix is quite important.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 104, + 543, + 577, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 577, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 322, + 555 + ], + "score": 1.0, + "content": "private static void dumpOpCounts(EnumMap", + "type": "text" + }, + { + "bbox": [ + 322, + 545, + 328, + 552 + ], + "score": 0.3, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 542, + 457, + 555 + ], + "score": 1.0, + "content": "FSEditLogOpCodes, Holder", + "type": "text" + }, + { + "bbox": [ + 457, + 545, + 463, + 552 + ], + "score": 0.33, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 542, + 577, + 555 + ], + "score": 1.0, + "content": "> opCounts) {", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 552, + 341, + 565 + ], + "spans": [ + { + "bbox": [ + 127, + 552, + 217, + 565 + ], + "score": 1.0, + "content": "StringBuilder sb", + "type": "text" + }, + { + "bbox": [ + 218, + 555, + 228, + 563 + ], + "score": 0.73, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 552, + 341, + 565 + ], + "score": 1.0, + "content": "new StringBuilder();", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 561, + 459, + 576 + ], + "spans": [ + { + "bbox": [ + 126, + 561, + 459, + 576 + ], + "score": 1.0, + "content": "sb.append(\"Summary of operations loaded from edit log:\\n \");", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 570, + 496, + 586 + ], + "spans": [ + { + "bbox": [ + 126, + 570, + 345, + 586 + ], + "score": 1.0, + "content": "Joiner.on(\"\\n \").withKeyValueSeparator(\"", + "type": "text" + }, + { + "bbox": [ + 345, + 574, + 356, + 582 + ], + "score": 0.3, + "content": "^ { 1 1 } =", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 570, + 496, + 586 + ], + "score": 1.0, + "content": "\").appendTo(sb, opCounts);", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 581, + 309, + 596 + ], + "spans": [ + { + "bbox": [ + 126, + 581, + 309, + 596 + ], + "score": 1.0, + "content": "FSImage.LOG.debug(sb.toString());", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 108, + 594, + 114, + 601 + ], + "spans": [ + { + "bbox": [ + 108, + 594, + 114, + 601 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "table", + "bbox": [ + 220, + 606, + 388, + 685 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 220, + 606, + 388, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 606, + 388, + 685 + ], + "spans": [ + { + "bbox": [ + 220, + 606, + 388, + 685 + ], + "score": 0.965, + "html": "
ModelPrediction
GREATappend
code2seqadd
Ours w/o structurelog
CODE TRANSFORMERlog op counts
Ground Truthdump op counts
", + "type": "table", + "image_path": "3539e0325fcec95f6f2a244b7e8480d60028213ed3f48a53d7797b35801c9743.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 606, + 388, + 619.1666666666666 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 220, + 619.1666666666666, + 388, + 632.3333333333333 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 220, + 632.3333333333333, + 388, + 645.4999999999999 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 220, + 645.4999999999999, + 388, + 658.6666666666665 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 220, + 658.6666666666665, + 388, + 671.8333333333331 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 220, + 671.8333333333331, + 388, + 684.9999999999998 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 693, + 505, + 716 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 691, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 705 + ], + "score": 1.0, + "content": "Table 13: Only the CODE TRANSFORMER could correctly identify that it is the op counts that should", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 702, + 151, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 151, + 718 + ], + "score": 1.0, + "content": "be logged.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "index": 38.5 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 106, + 92, + 367, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 345, + 104 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 345, + 104 + ], + "score": 1.0, + "content": "public static MNTPROC fromValue(int value) {", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 126, + 101, + 367, + 114 + ], + "spans": [ + { + "bbox": [ + 126, + 101, + 181, + 114 + ], + "score": 1.0, + "content": "if (value", + "type": "text" + }, + { + "bbox": [ + 181, + 103, + 200, + 113 + ], + "score": 0.4, + "content": "< ~ 0", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 101, + 250, + 114 + ], + "score": 1.0, + "content": "|| value", + "type": "text" + }, + { + "bbox": [ + 250, + 104, + 266, + 112 + ], + "score": 0.76, + "content": "> =", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 101, + 367, + 114 + ], + "score": 1.0, + "content": "values().length) {", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 113, + 215, + 124 + ], + "spans": [ + { + "bbox": [ + 149, + 113, + 215, + 124 + ], + "score": 1.0, + "content": "return null;", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 130, + 126, + 134, + 131 + ], + "spans": [ + { + "bbox": [ + 130, + 126, + 134, + 131 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 127, + 132, + 253, + 144 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 253, + 144 + ], + "score": 1.0, + "content": "return values()[value];", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 145, + 112, + 150 + ], + "spans": [ + { + "bbox": [ + 109, + 145, + 112, + 150 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 92, + 367, + 150 + ] + }, + { + "type": "table", + "bbox": [ + 229, + 156, + 379, + 234 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 229, + 156, + 379, + 234 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 229, + 156, + 379, + 234 + ], + "spans": [ + { + "bbox": [ + 229, + 156, + 379, + 234 + ], + "score": 0.966, + "html": "
ModelPrediction
GREAT get value
code2seq get value
Ours w/o structureto
CODE TRANSFORMERfrom value
Ground Truthfrom value
", + "type": "table", + "image_path": "c731fa631d72aabd1f5e32632f679a09b671309e6980e8efeb5fd4f3681003b9.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 156, + 379, + 195.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 229, + 195.0, + 379, + 234.0 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 105, + 242, + 505, + 265 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "Table 11: The CODE TRANSFORMER is the only model to recognize that the snippet is similar to a", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 253, + 334, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 334, + 266 + ], + "score": 1.0, + "content": "static factory method which is often preceded with from.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 241, + 506, + 266 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 429, + 308 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 430, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 430, + 310 + ], + "score": 1.0, + "content": "private Iterable listRootBlobs(String aPrefix,", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 298, + 430, + 310 + ] + }, + { + "type": "list", + "bbox": [ + 346, + 308, + 580, + 348 + ], + "lines": [ + { + "bbox": [ + 347, + 307, + 495, + 319 + ], + "spans": [ + { + "bbox": [ + 347, + 307, + 495, + 319 + ], + "score": 1.0, + "content": "boolean useFlatBlobListing,", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 348, + 317, + 581, + 329 + ], + "spans": [ + { + "bbox": [ + 348, + 317, + 581, + 329 + ], + "score": 1.0, + "content": "EnumSet listingDetails,", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 348, + 328, + 495, + 339 + ], + "spans": [ + { + "bbox": [ + 348, + 328, + 495, + 339 + ], + "score": 1.0, + "content": "BlobRequestOptions options,", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 348, + 338, + 494, + 349 + ], + "spans": [ + { + "bbox": [ + 348, + 338, + 494, + 349 + ], + "score": 1.0, + "content": "OperationContext opContext)", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + } + ], + "index": 12.5, + "bbox_fs": [ + 347, + 307, + 581, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 315, + 349, + 568, + 358 + ], + "lines": [ + { + "bbox": [ + 321, + 347, + 566, + 360 + ], + "spans": [ + { + "bbox": [ + 321, + 347, + 566, + 360 + ], + "score": 1.0, + "content": "throws StorageException, URISyntaxException {", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 321, + 347, + 566, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 114, + 359, + 579, + 389 + ], + "lines": [ + { + "bbox": [ + 127, + 357, + 581, + 369 + ], + "spans": [ + { + "bbox": [ + 127, + 357, + 320, + 369 + ], + "score": 1.0, + "content": "CloudBlobDirectoryWrapper directory", + "type": "text" + }, + { + "bbox": [ + 320, + 359, + 330, + 367 + ], + "score": 0.68, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 357, + 581, + 369 + ], + "score": 1.0, + "content": "this.container.getDirectoryReference(aPrefix);", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 366, + 410, + 380 + ], + "spans": [ + { + "bbox": [ + 126, + 366, + 410, + 380 + ], + "score": 1.0, + "content": "return directory.listBlobs(null, useFlatBlobListing,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 272, + 377, + 468, + 389 + ], + "spans": [ + { + "bbox": [ + 272, + 377, + 468, + 389 + ], + "score": 1.0, + "content": "listingDetails, options, opContext);", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 126, + 357, + 581, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 113, + 398 + ], + "lines": [ + { + "bbox": [ + 107, + 390, + 113, + 398 + ], + "spans": [ + { + "bbox": [ + 107, + 390, + 113, + 398 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 107, + 390, + 113, + 398 + ] + }, + { + "type": "table", + "bbox": [ + 215, + 402, + 394, + 479 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 215, + 402, + 394, + 479 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 215, + 402, + 394, + 479 + ], + "spans": [ + { + "bbox": [ + 215, + 402, + 394, + 479 + ], + "score": 0.972, + "html": "
ModelPrediction
GREATlist blobs
code2seqlist blobs
Ours w/o structurelist blobs
CODE TRANSFORMERlist blobs by prefix
Ground Truthlist root blobs
", + "type": "table", + "image_path": "4184371f3c5b40e5706dc3536eb808edba7ea9235136630720ca50cbd799e93d.jpg" + } + ] + } + ], + "index": 22.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 402, + 394, + 414.8333333333333 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 215, + 414.8333333333333, + 394, + 427.66666666666663 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 215, + 427.66666666666663, + 394, + 440.49999999999994 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 215, + 440.49999999999994, + 394, + 453.33333333333326 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 215, + 453.33333333333326, + 394, + 466.1666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 215, + 466.1666666666666, + 394, + 478.9999999999999 + ], + "spans": [], + "index": 25 + } + ] + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 105, + 487, + 504, + 510 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 501 + ], + "score": 1.0, + "content": "Table 12: All models could correctly identify the listBlobs() call in the return statement. However,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 497, + 499, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 499, + 511 + ], + "score": 1.0, + "content": "the CODE TRANSFORMER additionally comprehended that the specified prefix is quite important.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 486, + 505, + 511 + ] + }, + { + "type": "list", + "bbox": [ + 104, + 543, + 577, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 577, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 322, + 555 + ], + "score": 1.0, + "content": "private static void dumpOpCounts(EnumMap", + "type": "text" + }, + { + "bbox": [ + 322, + 545, + 328, + 552 + ], + "score": 0.3, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 542, + 457, + 555 + ], + "score": 1.0, + "content": "FSEditLogOpCodes, Holder", + "type": "text" + }, + { + "bbox": [ + 457, + 545, + 463, + 552 + ], + "score": 0.33, + "content": "<", + "type": "inline_equation" + }, + { + "bbox": [ + 463, + 542, + 577, + 555 + ], + "score": 1.0, + "content": "> opCounts) {", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 552, + 341, + 565 + ], + "spans": [ + { + "bbox": [ + 127, + 552, + 217, + 565 + ], + "score": 1.0, + "content": "StringBuilder sb", + "type": "text" + }, + { + "bbox": [ + 218, + 555, + 228, + 563 + ], + "score": 0.73, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 552, + 341, + 565 + ], + "score": 1.0, + "content": "new StringBuilder();", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 126, + 561, + 459, + 576 + ], + "spans": [ + { + "bbox": [ + 126, + 561, + 459, + 576 + ], + "score": 1.0, + "content": "sb.append(\"Summary of operations loaded from edit log:\\n \");", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 126, + 570, + 496, + 586 + ], + "spans": [ + { + "bbox": [ + 126, + 570, + 345, + 586 + ], + "score": 1.0, + "content": "Joiner.on(\"\\n \").withKeyValueSeparator(\"", + "type": "text" + }, + { + "bbox": [ + 345, + 574, + 356, + 582 + ], + "score": 0.3, + "content": "^ { 1 1 } =", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 570, + 496, + 586 + ], + "score": 1.0, + "content": "\").appendTo(sb, opCounts);", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 126, + 581, + 309, + 596 + ], + "spans": [ + { + "bbox": [ + 126, + 581, + 309, + 596 + ], + "score": 1.0, + "content": "FSImage.LOG.debug(sb.toString());", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 594, + 114, + 601 + ], + "spans": [ + { + "bbox": [ + 108, + 594, + 114, + 601 + ], + "score": 1.0, + "content": "}", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 542, + 577, + 601 + ] + }, + { + "type": "table", + "bbox": [ + 220, + 606, + 388, + 685 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 220, + 606, + 388, + 685 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 220, + 606, + 388, + 685 + ], + "spans": [ + { + "bbox": [ + 220, + 606, + 388, + 685 + ], + "score": 0.965, + "html": "
ModelPrediction
GREATappend
code2seqadd
Ours w/o structurelog
CODE TRANSFORMERlog op counts
Ground Truthdump op counts
", + "type": "table", + "image_path": "3539e0325fcec95f6f2a244b7e8480d60028213ed3f48a53d7797b35801c9743.jpg" + } + ] + } + ], + "index": 36.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 606, + 388, + 619.1666666666666 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 220, + 619.1666666666666, + 388, + 632.3333333333333 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 220, + 632.3333333333333, + 388, + 645.4999999999999 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 220, + 645.4999999999999, + 388, + 658.6666666666665 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 220, + 658.6666666666665, + 388, + 671.8333333333331 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 220, + 671.8333333333331, + 388, + 684.9999999999998 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 693, + 505, + 716 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 691, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 705 + ], + "score": 1.0, + "content": "Table 13: Only the CODE TRANSFORMER could correctly identify that it is the op counts that should", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 702, + 151, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 151, + 718 + ], + "score": 1.0, + "content": "be logged.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "index": 38.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 151, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 479, + 163 + ], + "score": 1.0, + "content": "static String execCommand(File f, String... cmd) throws IOException {", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 162, + 362, + 172 + ], + "spans": [ + { + "bbox": [ + 128, + 162, + 201, + 172 + ], + "score": 1.0, + "content": "String[] args", + "type": "text" + }, + { + "bbox": [ + 202, + 163, + 212, + 171 + ], + "score": 0.75, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 162, + 331, + 172 + ], + "score": 1.0, + "content": "new String[cmd.length", + "type": "text" + }, + { + "bbox": [ + 332, + 163, + 340, + 171 + ], + "score": 0.34, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 162, + 362, + 172 + ], + "score": 1.0, + "content": "1];", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 171, + 376, + 183 + ], + "spans": [ + { + "bbox": [ + 127, + 171, + 376, + 183 + ], + "score": 1.0, + "content": "System.arraycopy(cmd, 0, args, 0, cmd.length);", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 181, + 345, + 193 + ], + "spans": [ + { + "bbox": [ + 127, + 181, + 218, + 193 + ], + "score": 1.0, + "content": "args[cmd.length]", + "type": "text" + }, + { + "bbox": [ + 218, + 183, + 228, + 191 + ], + "score": 0.77, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 181, + 345, + 193 + ], + "score": 1.0, + "content": "f.getCanonicalPath();", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 191, + 343, + 202 + ], + "spans": [ + { + "bbox": [ + 128, + 191, + 201, + 202 + ], + "score": 1.0, + "content": "String output", + "type": "text" + }, + { + "bbox": [ + 201, + 193, + 212, + 200 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 191, + 343, + 202 + ], + "score": 1.0, + "content": "Shell.execCommand(args);", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 201, + 205, + 213 + ], + "spans": [ + { + "bbox": [ + 127, + 201, + 205, + 213 + ], + "score": 1.0, + "content": "return output;", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 216, + 225, + 393, + 303 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 216, + 225, + 393, + 303 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 216, + 225, + 393, + 303 + ], + "spans": [ + { + "bbox": [ + 216, + 225, + 393, + 303 + ], + "score": 0.97, + "html": "
ModelPrediction
GREATget canonical path
code2seqexec
Ours w/o structure get output
CODE TRANSFORMERexec command
Ground Truthexec command
", + "type": "table", + "image_path": "3c5bfe70032cf08db67ee0f3135ec46e80e42b50a65cc9210b7b528695c79ab7.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 216, + 225, + 393, + 238.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 216, + 238.0, + 393, + 251.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 216, + 251.0, + 393, + 264.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 216, + 264.0, + 393, + 277.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 216, + 277.0, + 393, + 290.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 216, + 290.0, + 393, + 303.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 104, + 311, + 505, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Table 14: Only the CODE TRANSFORMER and code2seq could identify that the relevant part of the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 320, + 429, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 429, + 336 + ], + "score": 1.0, + "content": "method is concerned with executing a command instead of returning something.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 106, + 485, + 398, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 400, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 400, + 497 + ], + "score": 1.0, + "content": "protected void subView(Class cls) {", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 109, + 492, + 511, + 537 + ], + "lines": [ + { + "bbox": [ + 127, + 493, + 232, + 506 + ], + "spans": [ + { + "bbox": [ + 127, + 493, + 232, + 506 + ], + "score": 1.0, + "content": "indent(of(ENDTAG));", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 504, + 217, + 516 + ], + "spans": [ + { + "bbox": [ + 127, + 504, + 217, + 516 + ], + "score": 1.0, + "content": "sb.setLength(0);", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 514, + 512, + 526 + ], + "spans": [ + { + "bbox": [ + 127, + 514, + 512, + 526 + ], + "score": 1.0, + "content": "out.print(sb.append('[').append(cls.getName()).append(']').toString());", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 127, + 524, + 205, + 536 + ], + "spans": [ + { + "bbox": [ + 127, + 524, + 205, + 536 + ], + "score": 1.0, + "content": "out.println();", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "table", + "bbox": [ + 223, + 548, + 385, + 626 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 223, + 548, + 385, + 626 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 223, + 548, + 385, + 626 + ], + "spans": [ + { + "bbox": [ + 223, + 548, + 385, + 626 + ], + "score": 0.962, + "html": "
ModelPrediction
GREATprint
code2seqprint
Ours w/o structureprint
CODE TRANSFORMERprint sub view
Ground Truthsub view
", + "type": "table", + "image_path": "0c4c2219f8a76155c82fd9cf1ddec51733b172b4b12ee945147b97301cb67ad3.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 548, + 385, + 561.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 223, + 561.0, + 385, + 574.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 223, + 574.0, + 385, + 587.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 223, + 587.0, + 385, + 600.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 223, + 600.0, + 385, + 613.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 223, + 613.0, + 385, + 626.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 634, + 505, + 657 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "Table 15: Only the CODE TRANSFORMER was able to link the print functionality to the object that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 644, + 488, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 488, + 657 + ], + "score": 1.0, + "content": "should be printed, which can only be inferred from the object’s class in the method parameters.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + } + ], + "index": 23.5 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 151, + 479, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 479, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 479, + 163 + ], + "score": 1.0, + "content": "static String execCommand(File f, String... cmd) throws IOException {", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 162, + 362, + 172 + ], + "spans": [ + { + "bbox": [ + 128, + 162, + 201, + 172 + ], + "score": 1.0, + "content": "String[] args", + "type": "text" + }, + { + "bbox": [ + 202, + 163, + 212, + 171 + ], + "score": 0.75, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 162, + 331, + 172 + ], + "score": 1.0, + "content": "new String[cmd.length", + "type": "text" + }, + { + "bbox": [ + 332, + 163, + 340, + 171 + ], + "score": 0.34, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 162, + 362, + 172 + ], + "score": 1.0, + "content": "1];", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 171, + 376, + 183 + ], + "spans": [ + { + "bbox": [ + 127, + 171, + 376, + 183 + ], + "score": 1.0, + "content": "System.arraycopy(cmd, 0, args, 0, cmd.length);", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 181, + 345, + 193 + ], + "spans": [ + { + "bbox": [ + 127, + 181, + 218, + 193 + ], + "score": 1.0, + "content": "args[cmd.length]", + "type": "text" + }, + { + "bbox": [ + 218, + 183, + 228, + 191 + ], + "score": 0.77, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 181, + 345, + 193 + ], + "score": 1.0, + "content": "f.getCanonicalPath();", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 191, + 343, + 202 + ], + "spans": [ + { + "bbox": [ + 128, + 191, + 201, + 202 + ], + "score": 1.0, + "content": "String output", + "type": "text" + }, + { + "bbox": [ + 201, + 193, + 212, + 200 + ], + "score": 0.8, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 191, + 343, + 202 + ], + "score": 1.0, + "content": "Shell.execCommand(args);", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 201, + 205, + 213 + ], + "spans": [ + { + "bbox": [ + 127, + 201, + 205, + 213 + ], + "score": 1.0, + "content": "return output;", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 151, + 479, + 213 + ] + }, + { + "type": "table", + "bbox": [ + 216, + 225, + 393, + 303 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 216, + 225, + 393, + 303 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 216, + 225, + 393, + 303 + ], + "spans": [ + { + "bbox": [ + 216, + 225, + 393, + 303 + ], + "score": 0.97, + "html": "
ModelPrediction
GREATget canonical path
code2seqexec
Ours w/o structure get output
CODE TRANSFORMERexec command
Ground Truthexec command
", + "type": "table", + "image_path": "3c5bfe70032cf08db67ee0f3135ec46e80e42b50a65cc9210b7b528695c79ab7.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 216, + 225, + 393, + 238.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 216, + 238.0, + 393, + 251.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 216, + 251.0, + 393, + 264.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 216, + 264.0, + 393, + 277.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 216, + 277.0, + 393, + 290.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 216, + 290.0, + 393, + 303.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 104, + 311, + 505, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Table 14: Only the CODE TRANSFORMER and code2seq could identify that the relevant part of the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 320, + 429, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 429, + 336 + ], + "score": 1.0, + "content": "method is concerned with executing a command instead of returning something.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 106, + 485, + 398, + 495 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 400, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 400, + 497 + ], + "score": 1.0, + "content": "protected void subView(Class cls) {", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "list", + "bbox": [ + 109, + 492, + 511, + 537 + ], + "lines": [ + { + "bbox": [ + 127, + 493, + 232, + 506 + ], + "spans": [ + { + "bbox": [ + 127, + 493, + 232, + 506 + ], + "score": 1.0, + "content": "indent(of(ENDTAG));", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 127, + 504, + 217, + 516 + ], + "spans": [ + { + "bbox": [ + 127, + 504, + 217, + 516 + ], + "score": 1.0, + "content": "sb.setLength(0);", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 127, + 514, + 512, + 526 + ], + "spans": [ + { + "bbox": [ + 127, + 514, + 512, + 526 + ], + "score": 1.0, + "content": "out.print(sb.append('[').append(cls.getName()).append(']').toString());", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 127, + 524, + 205, + 536 + ], + "spans": [ + { + "bbox": [ + 127, + 524, + 205, + 536 + ], + "score": 1.0, + "content": "out.println();", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 16.5, + "bbox_fs": [ + 127, + 493, + 512, + 536 + ] + }, + { + "type": "table", + "bbox": [ + 223, + 548, + 385, + 626 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 223, + 548, + 385, + 626 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 223, + 548, + 385, + 626 + ], + "spans": [ + { + "bbox": [ + 223, + 548, + 385, + 626 + ], + "score": 0.962, + "html": "
ModelPrediction
GREATprint
code2seqprint
Ours w/o structureprint
CODE TRANSFORMERprint sub view
Ground Truthsub view
", + "type": "table", + "image_path": "0c4c2219f8a76155c82fd9cf1ddec51733b172b4b12ee945147b97301cb67ad3.jpg" + } + ] + } + ], + "index": 21.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 548, + 385, + 561.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 223, + 561.0, + 385, + 574.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 223, + 574.0, + 385, + 587.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 223, + 587.0, + 385, + 600.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 223, + 600.0, + 385, + 613.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 223, + 613.0, + 385, + 626.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 634, + 505, + 657 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 634, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 506, + 647 + ], + "score": 1.0, + "content": "Table 15: Only the CODE TRANSFORMER was able to link the print functionality to the object that", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 644, + 488, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 488, + 657 + ], + "score": 1.0, + "content": "should be printed, which can only be inferred from the object’s class in the method parameters.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + } + ], + "index": 23.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 84, + 442, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 169, + 84, + 442, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 84, + 442, + 250 + ], + "spans": [ + { + "bbox": [ + 169, + 84, + 442, + 250 + ], + "score": 0.968, + "type": "image", + "image_path": "1225dc2212652b0900da96594dc1baedeb731793c2a7f99b2f39f0a306a759b8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 169, + 84, + 442, + 139.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 169, + 139.33333333333334, + 442, + 194.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 169, + 194.66666666666669, + 442, + 250.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 147, + 263, + 462, + 276 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 263, + 464, + 277 + ], + "spans": [ + { + "bbox": [ + 147, + 263, + 464, + 277 + ], + "score": 1.0, + "content": "Figure 7: Share of tokens in the labels also occurring in the bodies of methods.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 108, + 286, + 504, + 415 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 286, + 504, + 415 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 286, + 504, + 415 + ], + "spans": [ + { + "bbox": [ + 108, + 286, + 504, + 415 + ], + "score": 0.973, + "html": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq111-11=1111-
GREAT34.9331.1231.6129.6924.2425.5525.6921.4922.1848.3845.9745.71
Ours w/o structure36.8732.1732.9731.3025.0326.6431.4325.3426.6349.7846.7346.69
Ours w/o pointer net38.7731.7233.2732.7025.5027.3332.1230.1729.3653.0948.7049.26
Ours36.6833.8633.8433.3627.5529.0231.5324.7226.4352.0047.3547.93
code2seq (Multilanguage)-====-
GREAT (Multilanguage)35.7330.8131.7431.4926.1727.4129.7224.2025.4350.3247.9447.66
Ours w/o structure (Mult.)36.7829.9231.5832.6026.0227.7431.7126.0727.2451.9147.5848.15
Ours w/o pointer (Mult.)37.1830.5232.0433.9525.9228.1132.7625.0427.0153.5048.5449.35
Ours (Multilanguage)38.1033.3234.1834.2928.6930.0833.3028.3329.2953.8650.4650.61
Ours (Mult. + Finetune)38.2932.4133.6534.4328.2829.9132.8927.1528.4953.8550.85
Ours (Mult.+ LMPretrain)50.81
38.9734.7735.3435.2330.2631.3833.7329.1529.9455.3152.0352.13
", + "type": "table", + "image_path": "f9846de1ac74fbb68c2e86b16559380e7f8fc1f6b4b397c66475b197f4726324.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 108, + 286, + 504, + 329.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 329.0, + 504, + 372.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 372.0, + 504, + 415.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 163, + 423, + 448, + 435 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 161, + 421, + 448, + 437 + ], + "spans": [ + { + "bbox": [ + 161, + 421, + 448, + 437 + ], + "score": 1.0, + "content": "Table 16: Code summarization results on the CSN dataset (sample-F1).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "title", + "bbox": [ + 107, + 456, + 341, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 342, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 342, + 469 + ], + "score": 1.0, + "content": "A.8 ESTIMATION OF POINTER NETWORK POTENTIAL", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "In Table 2 we observe that the pointer network improves the F1 score for all languages except Go,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "where counterintuitively it leads to reduced performance as measured by F1 score on the test set", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "(while it improves by about 3 points on validation). To investigate this, in Figure 7 we plot the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 509, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 509, + 506, + 524 + ], + "score": 1.0, + "content": "share of tokens in the labels also occurring in the bodies of methods in the different languages.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "Intuitively, this gives an indication on how much gain we can expect from using a pointer network.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "If the share were zero, then no token in the labels ever occur in the bodies of the methods, so the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "score": 1.0, + "content": "pointer network cannot improve the prediction by pointing at the input. We see that for Go, there is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 554, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 104, + 554, + 506, + 566 + ], + "score": 1.0, + "content": "a strong mismatch between the test partition and the train/validation partitions, which much fewer", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "tokens from the labels occurring in the bodies of methods on test compared to train/validation. Thus,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 506, + 588 + ], + "score": 1.0, + "content": "we attribute the drop in performance observed by adding a pointer network on Go to this apparent", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 587, + 240, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 240, + 599 + ], + "score": 1.0, + "content": "violation of the i.i.d. assumption.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 611, + 435, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 610, + 436, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 436, + 625 + ], + "score": 1.0, + "content": "A.9 CODE SUMMARIZATION RESULTS ON THE CSN DATASET (SAMPLE-F1)", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 105, + 632, + 472, + 644 + ], + "lines": [ + { + "bbox": [ + 104, + 630, + 475, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 630, + 475, + 646 + ], + "score": 1.0, + "content": "In Table 16, we present our results on the CSN dataset as measured by the sample-F1 score.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "22", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 169, + 84, + 442, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 169, + 84, + 442, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 84, + 442, + 250 + ], + "spans": [ + { + "bbox": [ + 169, + 84, + 442, + 250 + ], + "score": 0.968, + "type": "image", + "image_path": "1225dc2212652b0900da96594dc1baedeb731793c2a7f99b2f39f0a306a759b8.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 169, + 84, + 442, + 139.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 169, + 139.33333333333334, + 442, + 194.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 169, + 194.66666666666669, + 442, + 250.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 147, + 263, + 462, + 276 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 263, + 464, + 277 + ], + "spans": [ + { + "bbox": [ + 147, + 263, + 464, + 277 + ], + "score": 1.0, + "content": "Figure 7: Share of tokens in the labels also occurring in the bodies of methods.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "table", + "bbox": [ + 108, + 286, + 504, + 415 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 286, + 504, + 415 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 286, + 504, + 415 + ], + "spans": [ + { + "bbox": [ + 108, + 286, + 504, + 415 + ], + "score": 0.973, + "html": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq111-11=1111-
GREAT34.9331.1231.6129.6924.2425.5525.6921.4922.1848.3845.9745.71
Ours w/o structure36.8732.1732.9731.3025.0326.6431.4325.3426.6349.7846.7346.69
Ours w/o pointer net38.7731.7233.2732.7025.5027.3332.1230.1729.3653.0948.7049.26
Ours36.6833.8633.8433.3627.5529.0231.5324.7226.4352.0047.3547.93
code2seq (Multilanguage)-====-
GREAT (Multilanguage)35.7330.8131.7431.4926.1727.4129.7224.2025.4350.3247.9447.66
Ours w/o structure (Mult.)36.7829.9231.5832.6026.0227.7431.7126.0727.2451.9147.5848.15
Ours w/o pointer (Mult.)37.1830.5232.0433.9525.9228.1132.7625.0427.0153.5048.5449.35
Ours (Multilanguage)38.1033.3234.1834.2928.6930.0833.3028.3329.2953.8650.4650.61
Ours (Mult. + Finetune)38.2932.4133.6534.4328.2829.9132.8927.1528.4953.8550.85
Ours (Mult.+ LMPretrain)50.81
38.9734.7735.3435.2330.2631.3833.7329.1529.9455.3152.0352.13
", + "type": "table", + "image_path": "f9846de1ac74fbb68c2e86b16559380e7f8fc1f6b4b397c66475b197f4726324.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 108, + 286, + 504, + 329.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 329.0, + 504, + 372.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 372.0, + 504, + 415.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 163, + 423, + 448, + 435 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 161, + 421, + 448, + 437 + ], + "spans": [ + { + "bbox": [ + 161, + 421, + 448, + 437 + ], + "score": 1.0, + "content": "Table 16: Code summarization results on the CSN dataset (sample-F1).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "title", + "bbox": [ + 107, + 456, + 341, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 342, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 342, + 469 + ], + "score": 1.0, + "content": "A.8 ESTIMATION OF POINTER NETWORK POTENTIAL", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 477, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "In Table 2 we observe that the pointer network improves the F1 score for all languages except Go,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 488, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 506, + 501 + ], + "score": 1.0, + "content": "where counterintuitively it leads to reduced performance as measured by F1 score on the test set", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "(while it improves by about 3 points on validation). To investigate this, in Figure 7 we plot the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 509, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 104, + 509, + 506, + 524 + ], + "score": 1.0, + "content": "share of tokens in the labels also occurring in the bodies of methods in the different languages.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 505, + 533 + ], + "score": 1.0, + "content": "Intuitively, this gives an indication on how much gain we can expect from using a pointer network.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "If the share were zero, then no token in the labels ever occur in the bodies of the methods, so the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 556 + ], + "score": 1.0, + "content": "pointer network cannot improve the prediction by pointing at the input. We see that for Go, there is", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 554, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 104, + 554, + 506, + 566 + ], + "score": 1.0, + "content": "a strong mismatch between the test partition and the train/validation partitions, which much fewer", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "tokens from the labels occurring in the bodies of methods on test compared to train/validation. Thus,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 576, + 506, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 506, + 588 + ], + "score": 1.0, + "content": "we attribute the drop in performance observed by adding a pointer network on Go to this apparent", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 587, + 240, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 240, + 599 + ], + "score": 1.0, + "content": "violation of the i.i.d. assumption.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 477, + 506, + 599 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 611, + 435, + 623 + ], + "lines": [ + { + "bbox": [ + 105, + 610, + 436, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 436, + 625 + ], + "score": 1.0, + "content": "A.9 CODE SUMMARIZATION RESULTS ON THE CSN DATASET (SAMPLE-F1)", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 105, + 632, + 472, + 644 + ], + "lines": [ + { + "bbox": [ + 104, + 630, + 475, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 630, + 475, + 646 + ], + "score": 1.0, + "content": "In Table 16, we present our results on the CSN dataset as measured by the sample-F1 score.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 104, + 630, + 475, + 646 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_model.json b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bdcddfb61f6c257233b1b8fb98bbae56835eb0e9 --- /dev/null +++ b/parse/train/Xh5eMZVONGF/Xh5eMZVONGF_model.json @@ -0,0 +1,24418 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 753, + 1302, + 753, + 1302, + 1180, + 398, + 1180 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1298, + 1403, + 1298, + 1403, + 1545, + 298, + 1545 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1560, + 1403, + 1560, + 1403, + 1835, + 298, + 1835 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 0, + "poly": [ + 299, + 220, + 1400, + 220, + 1400, + 324, + 299, + 324 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 314, + 517, + 650, + 517, + 650, + 608, + 314, + 608 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 1015, + 377, + 1385, + 377, + 1385, + 469, + 1015, + 469 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 1014, + 517, + 1373, + 517, + 1373, + 608, + 1014, + 608 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 314, + 376, + 794, + 376, + 794, + 470, + 314, + 470 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 302, + 1233, + 573, + 1233, + 573, + 1268, + 302, + 1268 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 773, + 689, + 927, + 689, + 927, + 722, + 773, + 722 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.609 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 216.0, + 1405.0, + 216.0, + 1405.0, + 273.0, + 292.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 273.0, + 1377.0, + 273.0, + 1377.0, + 328.0, + 293.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1226.0, + 580.0, + 1226.0, + 580.0, + 1277.0, + 293.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 685.0, + 934.0, + 685.0, + 934.0, + 728.0, + 768.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 753.0, + 1306.0, + 753.0, + 1306.0, + 788.0, + 395.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 783.0, + 1304.0, + 783.0, + 1304.0, + 818.0, + 393.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 814.0, + 1305.0, + 814.0, + 1305.0, + 849.0, + 394.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 844.0, + 1305.0, + 844.0, + 1305.0, + 878.0, + 393.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 876.0, + 1306.0, + 876.0, + 1306.0, + 911.0, + 394.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 905.0, + 1307.0, + 905.0, + 1307.0, + 943.0, + 393.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 937.0, + 1307.0, + 937.0, + 1307.0, + 972.0, + 393.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 967.0, + 1306.0, + 967.0, + 1306.0, + 1000.0, + 392.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 998.0, + 1304.0, + 998.0, + 1304.0, + 1033.0, + 394.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1028.0, + 1305.0, + 1028.0, + 1305.0, + 1062.0, + 394.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1057.0, + 1306.0, + 1057.0, + 1306.0, + 1092.0, + 394.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1089.0, + 1307.0, + 1089.0, + 1307.0, + 1124.0, + 394.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1119.0, + 1306.0, + 1119.0, + 1306.0, + 1154.0, + 394.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1150.0, + 893.0, + 1150.0, + 893.0, + 1185.0, + 394.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1300.0, + 1404.0, + 1300.0, + 1404.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1361.0, + 1405.0, + 1361.0, + 1405.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1390.0, + 1407.0, + 1390.0, + 1407.0, + 1428.0, + 293.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1423.0, + 1403.0, + 1423.0, + 1403.0, + 1456.0, + 296.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1487.0, + 294.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1518.0, + 293.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1513.0, + 980.0, + 1513.0, + 980.0, + 1549.0, + 293.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1559.0, + 1404.0, + 1559.0, + 1404.0, + 1595.0, + 296.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1589.0, + 1407.0, + 1589.0, + 1407.0, + 1626.0, + 294.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1619.0, + 1405.0, + 1619.0, + 1405.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1651.0, + 1407.0, + 1651.0, + 1407.0, + 1687.0, + 293.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1679.0, + 1408.0, + 1679.0, + 1408.0, + 1718.0, + 291.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1748.0, + 296.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1774.0, + 1405.0, + 1774.0, + 1405.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 695.0, + 1804.0, + 695.0, + 1841.0, + 294.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1407.0, + 1880.0, + 1407.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1946.0, + 296.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1407.0, + 1942.0, + 1407.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1276.0, + 2002.0, + 1276.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 515.0, + 489.0, + 515.0, + 489.0, + 549.0, + 311.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 542.0, + 539.0, + 542.0, + 539.0, + 582.0, + 311.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 579.0, + 652.0, + 579.0, + 652.0, + 611.0, + 316.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 375.0, + 1214.0, + 375.0, + 1214.0, + 409.0, + 1013.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 404.0, + 1239.0, + 404.0, + 1239.0, + 441.0, + 1013.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 440.0, + 1386.0, + 440.0, + 1386.0, + 469.0, + 1012.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 517.0, + 1273.0, + 517.0, + 1273.0, + 549.0, + 1014.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 548.0, + 1373.0, + 548.0, + 1373.0, + 578.0, + 1013.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 580.0, + 1352.0, + 580.0, + 1352.0, + 609.0, + 1013.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 378.0, + 720.0, + 378.0, + 720.0, + 410.0, + 313.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 408.0, + 674.0, + 408.0, + 674.0, + 439.0, + 312.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 437.0, + 796.0, + 437.0, + 796.0, + 473.0, + 314.0, + 473.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1418, + 1404, + 1418, + 1404, + 1756, + 297, + 1756 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1020, + 1404, + 1020, + 1404, + 1296, + 298, + 1296 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 789, + 1403, + 789, + 1403, + 1005, + 298, + 1005 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1771, + 1403, + 1771, + 1403, + 1894, + 298, + 1894 + ], + "score": 0.971 + }, + { + "category_id": 4, + "poly": [ + 295, + 544, + 1406, + 544, + 1406, + 728, + 295, + 728 + ], + "score": 0.961 + }, + { + "category_id": 2, + "poly": [ + 296, + 1919, + 1403, + 1919, + 1403, + 2034, + 296, + 2034 + ], + "score": 0.96 + }, + { + "category_id": 3, + "poly": [ + 462, + 224, + 1239, + 224, + 1239, + 517, + 462, + 517 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 301, + 1341, + 580, + 1341, + 580, + 1376, + 301, + 1376 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.736 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 546.0, + 1404.0, + 546.0, + 1404.0, + 578.0, + 295.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 575.0, + 1404.0, + 575.0, + 1404.0, + 612.0, + 293.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 606.0, + 1404.0, + 606.0, + 1404.0, + 641.0, + 293.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 637.0, + 1404.0, + 637.0, + 1404.0, + 673.0, + 295.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 669.0, + 1404.0, + 669.0, + 1404.0, + 701.0, + 295.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 699.0, + 1009.0, + 699.0, + 1009.0, + 731.0, + 297.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1915.0, + 1357.0, + 1915.0, + 1357.0, + 1954.0, + 332.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1940.0, + 1407.0, + 1940.0, + 1407.0, + 1989.0, + 327.0, + 1989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1975.0, + 1405.0, + 1975.0, + 1405.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 889.0, + 2002.0, + 889.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 228.0, + 799.0, + 228.0, + 799.0, + 256.0, + 459.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 228.0, + 1242.0, + 228.0, + 1242.0, + 256.0, + 897.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 291.0, + 713.0, + 291.0, + 713.0, + 315.0, + 686.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 316.0, + 714.0, + 316.0, + 714.0, + 345.0, + 515.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 311.0, + 776.0, + 311.0, + 776.0, + 334.0, + 725.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 338.0, + 709.0, + 338.0, + 709.0, + 368.0, + 536.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 338.0, + 784.0, + 338.0, + 784.0, + 375.0, + 745.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 341.0, + 1004.0, + 341.0, + 1004.0, + 370.0, + 972.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 363.0, + 649.0, + 363.0, + 649.0, + 389.0, + 540.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 362.0, + 957.0, + 362.0, + 957.0, + 380.0, + 910.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 381.0, + 508.0, + 381.0, + 508.0, + 419.0, + 470.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 383.0, + 746.0, + 383.0, + 746.0, + 409.0, + 608.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 404.0, + 604.0, + 404.0, + 604.0, + 431.0, + 539.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 411.0, + 726.0, + 411.0, + 726.0, + 429.0, + 710.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 409.0, + 1083.0, + 409.0, + 1083.0, + 439.0, + 1055.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 413.0, + 1200.0, + 413.0, + 1200.0, + 434.0, + 1178.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 430.0, + 724.0, + 430.0, + 724.0, + 450.0, + 561.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 422.0, + 784.0, + 422.0, + 784.0, + 455.0, + 749.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 433.0, + 1045.0, + 433.0, + 1045.0, + 456.0, + 993.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 491.0, + 672.0, + 491.0, + 672.0, + 518.0, + 591.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 488.0, + 1120.0, + 488.0, + 1120.0, + 520.0, + 1024.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 300.5, + 615.0, + 300.5, + 615.0, + 325.0, + 559.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.25, + 393.0, + 804.25, + 393.0, + 804.25, + 416.5, + 746.25, + 416.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1337.0, + 587.0, + 1337.0, + 587.0, + 1383.0, + 291.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1416.0, + 1404.0, + 1416.0, + 1404.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1448.0, + 1402.0, + 1448.0, + 1402.0, + 1487.0, + 292.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 1404.0, + 1480.0, + 1404.0, + 1515.0, + 294.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1514.0, + 1404.0, + 1514.0, + 1404.0, + 1545.0, + 295.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1541.0, + 1404.0, + 1541.0, + 1404.0, + 1575.0, + 292.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1606.0, + 295.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1637.0, + 292.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1405.0, + 1633.0, + 1405.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1663.0, + 1405.0, + 1663.0, + 1405.0, + 1698.0, + 294.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1694.0, + 1406.0, + 1694.0, + 1406.0, + 1729.0, + 294.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1725.0, + 1082.0, + 1725.0, + 1082.0, + 1760.0, + 295.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1018.0, + 1404.0, + 1018.0, + 1404.0, + 1058.0, + 295.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1052.0, + 1406.0, + 1052.0, + 1406.0, + 1085.0, + 294.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1082.0, + 1404.0, + 1082.0, + 1404.0, + 1115.0, + 294.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1111.0, + 1406.0, + 1111.0, + 1406.0, + 1148.0, + 294.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1141.0, + 1402.0, + 1141.0, + 1402.0, + 1178.0, + 294.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1171.0, + 1404.0, + 1171.0, + 1404.0, + 1210.0, + 292.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 1401.0, + 1204.0, + 1401.0, + 1236.0, + 294.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1231.0, + 1404.0, + 1231.0, + 1404.0, + 1270.0, + 293.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1264.0, + 838.0, + 1264.0, + 838.0, + 1300.0, + 294.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 787.0, + 1407.0, + 787.0, + 1407.0, + 827.0, + 293.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 819.0, + 1405.0, + 819.0, + 1405.0, + 855.0, + 293.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 851.0, + 1405.0, + 851.0, + 1405.0, + 886.0, + 294.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 1405.0, + 882.0, + 1405.0, + 917.0, + 294.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 912.0, + 1404.0, + 912.0, + 1404.0, + 946.0, + 294.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 944.0, + 1404.0, + 944.0, + 1404.0, + 978.0, + 294.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 974.0, + 1159.0, + 974.0, + 1159.0, + 1007.0, + 293.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1772.0, + 1402.0, + 1772.0, + 1402.0, + 1805.0, + 296.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1835.0, + 294.0, + 1835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1831.0, + 1403.0, + 1831.0, + 1403.0, + 1867.0, + 293.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1864.0, + 1403.0, + 1864.0, + 1403.0, + 1896.0, + 296.0, + 1896.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 315, + 1404, + 315, + 1404, + 682, + 297, + 682 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 703, + 1404, + 703, + 1404, + 983, + 297, + 983 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1471, + 1403, + 1471, + 1403, + 1627, + 298, + 1627 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1640, + 1403, + 1640, + 1403, + 1764, + 297, + 1764 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1251, + 1404, + 1251, + 1404, + 1345, + 299, + 1345 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1399, + 229, + 1399, + 293, + 294, + 293 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 300, + 1964, + 1404, + 1964, + 1404, + 2039, + 300, + 2039 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 299, + 1395, + 1400, + 1395, + 1400, + 1458, + 299, + 1458 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 584, + 1166, + 1114, + 1166, + 1114, + 1243, + 584, + 1243 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 298, + 1091, + 1400, + 1091, + 1400, + 1156, + 298, + 1156 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 327, + 1860, + 1335, + 1860, + 1335, + 1956, + 327, + 1956 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 1824, + 536, + 1824, + 536, + 1854, + 297, + 1854 + ], + "score": 0.927 + }, + { + "category_id": 8, + "poly": [ + 661, + 1345, + 1039, + 1345, + 1039, + 1388, + 661, + 1388 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 291, + 1023, + 1370, + 1023, + 1370, + 1062, + 291, + 1062 + ], + "score": 0.871 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1189, + 1400, + 1189, + 1400, + 1219, + 1366, + 1219 + ], + "score": 0.869 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1352, + 1400, + 1352, + 1400, + 1381, + 1366, + 1381 + ], + "score": 0.865 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1781, + 1400, + 1781, + 1400, + 1810, + 1366, + 1810 + ], + "score": 0.865 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1870, + 1400, + 1870, + 1400, + 1899, + 1366, + 1899 + ], + "score": 0.842 + }, + { + "category_id": 8, + "poly": [ + 590, + 1774, + 1110, + 1774, + 1110, + 1817, + 590, + 1817 + ], + "score": 0.655 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.649 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.542 + }, + { + "category_id": 8, + "poly": [ + 589, + 1774, + 1112, + 1774, + 1112, + 1817, + 589, + 1817 + ], + "score": 0.274 + }, + { + "category_id": 8, + "poly": [ + 660, + 1345, + 1038, + 1345, + 1038, + 1387, + 660, + 1387 + ], + "score": 0.109 + }, + { + "category_id": 14, + "poly": [ + 582, + 1164, + 1113, + 1164, + 1113, + 1243, + 582, + 1243 + ], + "score": 0.94, + "latex": "{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { k } } } } \\right) V ," + }, + { + "category_id": 14, + "poly": [ + 328, + 1859, + 1339, + 1859, + 1339, + 1958, + 328, + 1958 + ], + "score": 0.94, + "latex": "A _ { i j } = \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } } _ { ( \\mathrm { u } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { E } _ { i } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { b } ) { A } _ { i j } ^ { \\mathrm { c p } } } } + \\underbrace { { \\phi } ( p _ { i } ) ^ { T } W _ { q } ^ { T } W _ { k } { E } _ { j } } _ { ( \\mathrm { c } ) { A } _ { i j } ^ { \\mathrm { c e } } } { + \\underbrace { { { \\phi } ( p _ { i } ) } ^ { T } W _ { q } ^ { T } W _ { k } { \\phi } ( p _ { j } ) } _ { ( \\mathrm { d } ) { A } _ { i j } ^ { \\mathrm { p } } } } ." + }, + { + "category_id": 14, + "poly": [ + 658, + 1343, + 1040, + 1343, + 1040, + 1389, + 658, + 1389 + ], + "score": 0.93, + "latex": "\\pmb { A } _ { i j } = \\pmb { Q } _ { i } ^ { T } \\pmb { K } _ { j } = \\pmb { E } _ { i } ^ { T } \\pmb { W } _ { q } ^ { T } \\pmb { W } _ { k } \\pmb { E } _ { j } ," + }, + { + "category_id": 13, + "poly": [ + 373, + 1393, + 519, + 1393, + 519, + 1429, + 373, + 1429 + ], + "score": 0.93, + "latex": "E _ { i } , E _ { j } \\in \\mathbb { R } ^ { d }" + }, + { + "category_id": 14, + "poly": [ + 588, + 1772, + 1108, + 1772, + 1108, + 1818, + 588, + 1818 + ], + "score": 0.92, + "latex": "A _ { i j } = ( E _ { i } + \\phi ( p _ { i } ) ) ^ { T } W _ { q } ^ { T } W _ { k } ( E _ { j } + \\phi ( p _ { j } ) ) , ," + }, + { + "category_id": 13, + "poly": [ + 824, + 1285, + 923, + 1285, + 923, + 1316, + 824, + 1316 + ], + "score": 0.92, + "latex": "d _ { k } = d _ { v }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1701, + 436, + 1701, + 436, + 1734, + 298, + 1734 + ], + "score": 0.92, + "latex": "\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1252, + 559, + 1252, + 559, + 1286, + 374, + 1286 + ], + "score": 0.91, + "latex": "Q , K \\in \\mathbb { R } ^ { N \\times d _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 848, + 1967, + 894, + 1967, + 894, + 2003, + 848, + 2003 + ], + "score": 0.91, + "latex": "{ \\cal { A } } _ { i j } ^ { \\mathrm { c c } }" + }, + { + "category_id": 13, + "poly": [ + 796, + 2002, + 842, + 2002, + 842, + 2041, + 796, + 2041 + ], + "score": 0.9, + "latex": "A _ { i j } ^ { \\mathrm { c p } }" + }, + { + "category_id": 13, + "poly": [ + 612, + 1251, + 757, + 1251, + 757, + 1283, + 612, + 1283 + ], + "score": 0.9, + "latex": "V \\in \\mathbb { R } ^ { N \\times d _ { v } }" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1255, + 1181, + 1255, + 1181, + 1285, + 1150, + 1285 + ], + "score": 0.89, + "latex": "d _ { k }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1316, + 338, + 1316, + 338, + 1350, + 298, + 1350 + ], + "score": 0.89, + "latex": "K _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1270, + 1286, + 1306, + 1286, + 1306, + 1317, + 1270, + 1317 + ], + "score": 0.89, + "latex": "Q _ { i }" + }, + { + "category_id": 13, + "poly": [ + 345, + 1286, + 375, + 1286, + 375, + 1315, + 345, + 1315 + ], + "score": 0.88, + "latex": "d _ { v }" + }, + { + "category_id": 13, + "poly": [ + 706, + 1706, + 732, + 1706, + 732, + 1734, + 706, + 1734 + ], + "score": 0.84, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 603, + 1397, + 621, + 1397, + 621, + 1423, + 603, + 1423 + ], + "score": 0.82, + "latex": "d" + }, + { + "category_id": 13, + "poly": [ + 1016, + 1398, + 1029, + 1398, + 1029, + 1423, + 1016, + 1423 + ], + "score": 0.79, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 670, + 2005, + 683, + 2005, + 683, + 2030, + 670, + 2030 + ], + "score": 0.79, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1078, + 1397, + 1094, + 1397, + 1094, + 1428, + 1078, + 1428 + ], + "score": 0.78, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 731, + 2005, + 747, + 2005, + 747, + 2035, + 731, + 2035 + ], + "score": 0.77, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 774, + 1254, + 803, + 1254, + 803, + 1282, + 774, + 1282 + ], + "score": 0.76, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1152, + 1395, + 1196, + 1395, + 1196, + 1430, + 1152, + 1430 + ], + "score": 0.63, + "latex": "W _ { q }" + }, + { + "category_id": 13, + "poly": [ + 1204, + 1392, + 1360, + 1392, + 1360, + 1429, + 1204, + 1429 + ], + "score": 0.61, + "latex": "W _ { k } \\in \\mathbb { R } ^ { d _ { k } \\times d }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1022.0, + 1377.0, + 1022.0, + 1377.0, + 1064.0, + 291.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 315.0, + 1405.0, + 315.0, + 1405.0, + 353.0, + 294.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 345.0, + 1406.0, + 345.0, + 1406.0, + 383.0, + 294.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 376.0, + 1404.0, + 376.0, + 1404.0, + 412.0, + 294.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 404.0, + 1405.0, + 404.0, + 1405.0, + 446.0, + 292.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 439.0, + 1405.0, + 439.0, + 1405.0, + 473.0, + 295.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 469.0, + 1405.0, + 469.0, + 1405.0, + 503.0, + 295.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 500.0, + 1405.0, + 500.0, + 1405.0, + 534.0, + 294.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 529.0, + 1407.0, + 529.0, + 1407.0, + 567.0, + 294.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 557.0, + 1406.0, + 557.0, + 1406.0, + 597.0, + 291.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 588.0, + 1405.0, + 588.0, + 1405.0, + 626.0, + 292.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 620.0, + 1407.0, + 620.0, + 1407.0, + 658.0, + 292.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 648.0, + 403.0, + 648.0, + 403.0, + 685.0, + 295.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 706.0, + 1405.0, + 706.0, + 1405.0, + 739.0, + 296.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 737.0, + 1405.0, + 737.0, + 1405.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 802.0, + 292.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 795.0, + 1406.0, + 795.0, + 1406.0, + 832.0, + 294.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 827.0, + 1406.0, + 827.0, + 1406.0, + 864.0, + 292.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 855.0, + 1406.0, + 855.0, + 1406.0, + 893.0, + 292.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 889.0, + 1405.0, + 889.0, + 1405.0, + 922.0, + 295.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 914.0, + 1406.0, + 914.0, + 1406.0, + 958.0, + 292.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 949.0, + 1346.0, + 949.0, + 1346.0, + 984.0, + 292.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1473.0, + 1404.0, + 1473.0, + 1404.0, + 1506.0, + 297.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1541.0, + 292.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1534.0, + 1404.0, + 1534.0, + 1404.0, + 1568.0, + 296.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1562.0, + 1407.0, + 1562.0, + 1407.0, + 1600.0, + 294.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1596.0, + 1031.0, + 1596.0, + 1031.0, + 1628.0, + 294.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1638.0, + 1404.0, + 1638.0, + 1404.0, + 1678.0, + 292.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1670.0, + 1404.0, + 1670.0, + 1404.0, + 1706.0, + 292.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1702.0, + 297.0, + 1702.0, + 297.0, + 1738.0, + 293.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1702.0, + 705.0, + 1702.0, + 705.0, + 1738.0, + 437.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1702.0, + 1405.0, + 1702.0, + 1405.0, + 1738.0, + 733.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1732.0, + 971.0, + 1732.0, + 971.0, + 1768.0, + 293.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 373.0, + 1247.0, + 373.0, + 1291.0, + 293.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1247.0, + 611.0, + 1247.0, + 611.0, + 1291.0, + 560.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1247.0, + 773.0, + 1247.0, + 773.0, + 1291.0, + 758.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 1247.0, + 1149.0, + 1247.0, + 1149.0, + 1291.0, + 804.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1247.0, + 1405.0, + 1247.0, + 1405.0, + 1291.0, + 1182.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1281.0, + 344.0, + 1281.0, + 344.0, + 1321.0, + 293.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1281.0, + 823.0, + 1281.0, + 823.0, + 1321.0, + 376.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1281.0, + 1269.0, + 1281.0, + 1269.0, + 1321.0, + 924.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1281.0, + 1403.0, + 1281.0, + 1403.0, + 1321.0, + 1307.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1315.0, + 544.0, + 1315.0, + 544.0, + 1349.0, + 339.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 1401.0, + 229.0, + 1401.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1385.0, + 261.0, + 1385.0, + 297.0, + 295.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1965.0, + 847.0, + 1965.0, + 847.0, + 2003.0, + 298.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1965.0, + 1405.0, + 1965.0, + 1405.0, + 2003.0, + 895.0, + 2003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1999.0, + 669.0, + 1999.0, + 669.0, + 2041.0, + 295.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1999.0, + 730.0, + 1999.0, + 730.0, + 2041.0, + 684.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1999.0, + 795.0, + 1999.0, + 795.0, + 2041.0, + 748.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1999.0, + 1405.0, + 1999.0, + 1405.0, + 2041.0, + 843.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1386.0, + 372.0, + 1386.0, + 372.0, + 1434.0, + 293.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1386.0, + 602.0, + 1386.0, + 602.0, + 1434.0, + 520.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1386.0, + 1015.0, + 1386.0, + 1015.0, + 1434.0, + 622.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1386.0, + 1077.0, + 1386.0, + 1077.0, + 1434.0, + 1030.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1386.0, + 1151.0, + 1386.0, + 1151.0, + 1434.0, + 1095.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1386.0, + 1203.0, + 1386.0, + 1203.0, + 1434.0, + 1197.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1386.0, + 1409.0, + 1386.0, + 1409.0, + 1434.0, + 1361.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1425.0, + 872.0, + 1425.0, + 872.0, + 1461.0, + 295.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1090.0, + 1404.0, + 1090.0, + 1404.0, + 1129.0, + 294.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1125.0, + 1364.0, + 1125.0, + 1364.0, + 1157.0, + 296.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1823.0, + 535.0, + 1823.0, + 535.0, + 1855.0, + 298.0, + 1855.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 445, + 1404, + 445, + 1404, + 662, + 297, + 662 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 753, + 1405, + 753, + 1405, + 976, + 297, + 976 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1000, + 744, + 1000, + 744, + 1243, + 297, + 1243 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1764, + 1404, + 1764, + 1404, + 1919, + 297, + 1919 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1587, + 1404, + 1587, + 1404, + 1744, + 297, + 1744 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1382, + 1404, + 1382, + 1404, + 1565, + 298, + 1565 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1941, + 1402, + 1941, + 1402, + 2035, + 298, + 2035 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1260, + 744, + 1260, + 744, + 1382, + 298, + 1382 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 773, + 1034, + 1395, + 1034, + 1395, + 1305, + 773, + 1305 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 289, + 227, + 1402, + 227, + 1402, + 299, + 289, + 299 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 291, + 312, + 1403, + 312, + 1403, + 377, + 291, + 377 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 403, + 388, + 1296, + 388, + 1296, + 432, + 403, + 432 + ], + "score": 0.925 + }, + { + "category_id": 4, + "poly": [ + 814, + 1328, + 1353, + 1328, + 1353, + 1360, + 814, + 1360 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1366, + 396, + 1400, + 396, + 1400, + 426, + 1366, + 426 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.767 + }, + { + "category_id": 0, + "poly": [ + 302, + 695, + 1253, + 695, + 1253, + 729, + 302, + 729 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 302, + 695, + 1253, + 695, + 1253, + 729, + 302, + 729 + ], + "score": 0.171 + }, + { + "category_id": 13, + "poly": [ + 1229, + 345, + 1323, + 345, + 1323, + 379, + 1229, + 379 + ], + "score": 0.92, + "latex": "\\phi ( r _ { i \\to j } )" + }, + { + "category_id": 13, + "poly": [ + 799, + 345, + 862, + 345, + 862, + 378, + 799, + 378 + ], + "score": 0.92, + "latex": "\\phi ( p _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 823, + 1712, + 953, + 1712, + 953, + 1745, + 823, + 1745 + ], + "score": 0.92, + "latex": "r _ { j i } = - 2" + }, + { + "category_id": 13, + "poly": [ + 357, + 1712, + 466, + 1712, + 466, + 1745, + 357, + 1745 + ], + "score": 0.92, + "latex": "r _ { i \\to j } = 2" + }, + { + "category_id": 13, + "poly": [ + 1123, + 443, + 1251, + 443, + 1251, + 477, + 1123, + 477 + ], + "score": 0.92, + "latex": "\\boldsymbol { u } , \\boldsymbol { v } \\in \\mathbb { R } ^ { d _ { k } }" + }, + { + "category_id": 13, + "poly": [ + 620, + 229, + 666, + 229, + 666, + 269, + 620, + 269 + ], + "score": 0.92, + "latex": "A _ { i j } ^ { \\mathrm { p c } }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1860, + 1350, + 1860, + 1350, + 1890, + 1202, + 1890 + ], + "score": 0.91, + "latex": "r _ { v _ { 1 } v _ { 3 } } ~ = ~ 2" + }, + { + "category_id": 13, + "poly": [ + 1317, + 1826, + 1370, + 1826, + 1370, + 1860, + 1317, + 1860 + ], + "score": 0.91, + "latex": "\\{ v _ { i } \\}" + }, + { + "category_id": 13, + "poly": [ + 379, + 815, + 445, + 815, + 445, + 853, + 379, + 853 + ], + "score": 0.91, + "latex": "W _ { r } ^ { ( s ) }" + }, + { + "category_id": 13, + "poly": [ + 1174, + 2003, + 1231, + 2003, + 1231, + 2036, + 1174, + 2036 + ], + "score": 0.91, + "latex": "( i , j )" + }, + { + "category_id": 13, + "poly": [ + 361, + 264, + 409, + 264, + 409, + 306, + 361, + 306 + ], + "score": 0.9, + "latex": "A _ { i j } ^ { \\mathrm { p p } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1888, + 453, + 1888, + 453, + 1922, + 297, + 1922 + ], + "score": 0.9, + "latex": "r _ { v _ { 3 } v _ { 1 } } = - 2" + }, + { + "category_id": 13, + "poly": [ + 489, + 477, + 534, + 477, + 534, + 508, + 489, + 508 + ], + "score": 0.9, + "latex": "W _ { r }" + }, + { + "category_id": 14, + "poly": [ + 405, + 387, + 1294, + 387, + 1294, + 433, + 405, + 433 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { A _ { i j } ^ { \\mathrm { r e l } } = \\ E _ { i } ^ { T } W _ { q } ^ { T } W _ { k } E _ { j } + E _ { i } ^ { T } W _ { q } ^ { T } W _ { r } \\phi ( r _ { i j } ) + u ^ { T } W _ { k } E _ { j } + v ^ { T } W _ { r } \\phi ( r _ { i j } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 371, + 453, + 428, + 453, + 428, + 480, + 371, + 480 + ], + "score": 0.9, + "latex": "r _ { i \\to j }" + }, + { + "category_id": 13, + "poly": [ + 921, + 449, + 937, + 449, + 937, + 479, + 921, + 479 + ], + "score": 0.84, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 703, + 1713, + 720, + 1713, + 720, + 1743, + 703, + 1743 + ], + "score": 0.81, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 810, + 450, + 823, + 450, + 823, + 474, + 810, + 474 + ], + "score": 0.78, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 298, + 1062, + 315, + 1062, + 315, + 1093, + 298, + 1093 + ], + "score": 0.78, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 753, + 1713, + 766, + 1713, + 766, + 1738, + 753, + 1738 + ], + "score": 0.76, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1341, + 2005, + 1355, + 2005, + 1355, + 2030, + 1341, + 2030 + ], + "score": 0.75, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 557, + 233, + 571, + 233, + 571, + 259, + 557, + 259 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 545, + 1063, + 560, + 1063, + 560, + 1088, + 545, + 1088 + ], + "score": 0.7, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 297, + 270, + 310, + 270, + 310, + 296, + 297, + 296 + ], + "score": 0.67, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 594, + 829, + 610, + 829, + 610, + 850, + 594, + 850 + ], + "score": 0.58, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1069.0, + 1276.0, + 1069.0, + 1276.0, + 1091.0, + 1200.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1084.0, + 965.0, + 1084.0, + 965.0, + 1139.0, + 941.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1091.0, + 1261.0, + 1091.0, + 1261.0, + 1114.0, + 1212.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1112.0, + 1269.0, + 1112.0, + 1269.0, + 1141.0, + 1206.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1151.0, + 842.0, + 1151.0, + 842.0, + 1181.0, + 773.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1168.0, + 941.0, + 1168.0, + 941.0, + 1193.0, + 917.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1176.0, + 1278.0, + 1176.0, + 1278.0, + 1198.0, + 1198.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1195.0, + 1278.0, + 1195.0, + 1278.0, + 1223.0, + 1196.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1223.0, + 1382.0, + 1223.0, + 1382.0, + 1253.0, + 1331.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1249.0, + 1269.0, + 1249.0, + 1269.0, + 1283.0, + 1202.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1250.0, + 1395.0, + 1250.0, + 1395.0, + 1275.0, + 1316.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1276.0, + 1278.0, + 1276.0, + 1278.0, + 1301.0, + 1198.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 1272.0, + 1391.0, + 1272.0, + 1391.0, + 1298.0, + 1320.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1323.0, + 1358.0, + 1323.0, + 1358.0, + 1366.0, + 810.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 694.0, + 1260.0, + 694.0, + 1260.0, + 732.0, + 294.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 445.0, + 370.0, + 445.0, + 370.0, + 482.0, + 293.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 445.0, + 809.0, + 445.0, + 809.0, + 482.0, + 429.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 445.0, + 920.0, + 445.0, + 920.0, + 482.0, + 824.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 445.0, + 1122.0, + 445.0, + 1122.0, + 482.0, + 938.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 445.0, + 1405.0, + 445.0, + 1405.0, + 482.0, + 1252.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 475.0, + 488.0, + 475.0, + 488.0, + 514.0, + 294.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 475.0, + 1405.0, + 475.0, + 1405.0, + 514.0, + 535.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 506.0, + 1406.0, + 506.0, + 1406.0, + 545.0, + 294.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 540.0, + 1404.0, + 540.0, + 1404.0, + 571.0, + 296.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 567.0, + 1405.0, + 567.0, + 1405.0, + 603.0, + 294.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 598.0, + 1407.0, + 598.0, + 1407.0, + 634.0, + 292.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 632.0, + 510.0, + 632.0, + 510.0, + 664.0, + 296.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 752.0, + 1403.0, + 752.0, + 1403.0, + 789.0, + 294.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 785.0, + 1405.0, + 785.0, + 1405.0, + 820.0, + 295.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 822.0, + 593.0, + 822.0, + 593.0, + 857.0, + 446.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 822.0, + 1405.0, + 822.0, + 1405.0, + 857.0, + 611.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 854.0, + 1403.0, + 854.0, + 1403.0, + 886.0, + 296.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 885.0, + 1403.0, + 885.0, + 1403.0, + 916.0, + 296.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 912.0, + 1406.0, + 912.0, + 1406.0, + 949.0, + 292.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 942.0, + 496.0, + 942.0, + 496.0, + 978.0, + 294.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 812.5, + 456.0, + 812.5, + 456.0, + 865.5, + 286.0, + 865.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 999.0, + 747.0, + 999.0, + 747.0, + 1032.0, + 296.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 747.0, + 1030.0, + 747.0, + 1063.0, + 295.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1062.0, + 297.0, + 1062.0, + 297.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1062.0, + 544.0, + 1062.0, + 544.0, + 1091.0, + 316.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1062.0, + 747.0, + 1062.0, + 747.0, + 1091.0, + 561.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1089.0, + 746.0, + 1089.0, + 746.0, + 1120.0, + 295.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1122.0, + 747.0, + 1122.0, + 747.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1151.0, + 747.0, + 1151.0, + 747.0, + 1184.0, + 294.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1181.0, + 747.0, + 1181.0, + 747.0, + 1214.0, + 294.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1210.0, + 422.0, + 1210.0, + 422.0, + 1245.0, + 295.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1766.0, + 1405.0, + 1766.0, + 1405.0, + 1799.0, + 296.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1795.0, + 1404.0, + 1795.0, + 1404.0, + 1828.0, + 296.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1824.0, + 1316.0, + 1824.0, + 1316.0, + 1861.0, + 295.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1371.0, + 1824.0, + 1406.0, + 1824.0, + 1406.0, + 1861.0, + 1371.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1851.0, + 1201.0, + 1851.0, + 1201.0, + 1899.0, + 287.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 1851.0, + 1410.0, + 1851.0, + 1410.0, + 1899.0, + 1351.0, + 1899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1887.0, + 465.0, + 1887.0, + 465.0, + 1926.0, + 290.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1587.0, + 1402.0, + 1587.0, + 1402.0, + 1621.0, + 296.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1617.0, + 1405.0, + 1617.0, + 1405.0, + 1653.0, + 295.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1644.0, + 1405.0, + 1644.0, + 1405.0, + 1688.0, + 291.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1678.0, + 1407.0, + 1678.0, + 1407.0, + 1716.0, + 295.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1706.0, + 356.0, + 1706.0, + 356.0, + 1747.0, + 294.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 1706.0, + 702.0, + 1706.0, + 702.0, + 1747.0, + 467.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1706.0, + 752.0, + 1706.0, + 752.0, + 1747.0, + 721.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1706.0, + 822.0, + 1706.0, + 822.0, + 1747.0, + 767.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1706.0, + 1201.0, + 1706.0, + 1201.0, + 1747.0, + 954.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1383.0, + 1404.0, + 1383.0, + 1404.0, + 1414.0, + 296.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1412.0, + 1404.0, + 1412.0, + 1404.0, + 1447.0, + 294.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1441.0, + 1407.0, + 1441.0, + 1407.0, + 1479.0, + 292.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1474.0, + 1404.0, + 1474.0, + 1404.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1503.0, + 1407.0, + 1503.0, + 1407.0, + 1539.0, + 292.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1535.0, + 671.0, + 1535.0, + 671.0, + 1567.0, + 296.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1173.0, + 2001.0, + 1173.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 2001.0, + 1340.0, + 2001.0, + 1340.0, + 2039.0, + 1232.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2039.0, + 1356.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1258.0, + 747.0, + 1258.0, + 747.0, + 1291.0, + 296.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1291.0, + 746.0, + 1291.0, + 746.0, + 1322.0, + 296.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1320.0, + 748.0, + 1320.0, + 748.0, + 1352.0, + 293.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1351.0, + 746.0, + 1351.0, + 746.0, + 1384.0, + 295.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 556.0, + 224.0, + 556.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 224.0, + 619.0, + 224.0, + 619.0, + 268.0, + 572.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 224.0, + 1407.0, + 224.0, + 1407.0, + 268.0, + 667.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 263.0, + 296.0, + 263.0, + 296.0, + 305.0, + 293.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 263.0, + 360.0, + 263.0, + 360.0, + 305.0, + 311.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 263.0, + 1126.0, + 263.0, + 1126.0, + 305.0, + 410.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 310.0, + 1405.0, + 310.0, + 1405.0, + 352.0, + 292.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 342.0, + 798.0, + 342.0, + 798.0, + 383.0, + 292.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 342.0, + 1228.0, + 342.0, + 1228.0, + 383.0, + 863.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 342.0, + 1336.0, + 342.0, + 1336.0, + 383.0, + 1324.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 694.0, + 1260.0, + 694.0, + 1260.0, + 732.0, + 294.0, + 732.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1384, + 1404, + 1384, + 1404, + 1661, + 298, + 1661 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1064, + 1404, + 1064, + 1404, + 1278, + 298, + 1278 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1403, + 1788, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 322, + 224, + 1386, + 224, + 1386, + 694, + 322, + 694 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 297, + 725, + 1404, + 725, + 1404, + 910, + 297, + 910 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 977, + 1400, + 977, + 1400, + 1040, + 298, + 1040 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 300, + 1714, + 674, + 1714, + 674, + 1749, + 300, + 1749 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 299, + 1326, + 941, + 1326, + 941, + 1356, + 299, + 1356 + ], + "score": 0.831 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.704 + }, + { + "category_id": 13, + "poly": [ + 603, + 1416, + 697, + 1416, + 697, + 1450, + 603, + 1450 + ], + "score": 0.92, + "latex": "\\phi ( r _ { i \\to j } )" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1537, + 1137, + 1537, + 1137, + 1567, + 1038, + 1567 + ], + "score": 0.91, + "latex": "k \\ll N ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1047, + 1386, + 1182, + 1386, + 1182, + 1415, + 1047, + 1415 + ], + "score": 0.91, + "latex": "N \\times N \\times d" + }, + { + "category_id": 13, + "poly": [ + 946, + 1630, + 1030, + 1630, + 1030, + 1658, + 946, + 1658 + ], + "score": 0.9, + "latex": "k = 3 2" + }, + { + "category_id": 13, + "poly": [ + 697, + 981, + 714, + 981, + 714, + 1011, + 697, + 1011 + ], + "score": 0.84, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 905, + 1417, + 934, + 1417, + 934, + 1444, + 905, + 1444 + ], + "score": 0.81, + "latex": "N" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 229.0, + 679.0, + 229.0, + 679.0, + 260.0, + 322.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 231.0, + 1064.0, + 231.0, + 1064.0, + 257.0, + 886.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 226.0, + 1337.0, + 226.0, + 1337.0, + 255.0, + 1274.0, + 255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 254.0, + 596.0, + 254.0, + 596.0, + 292.0, + 555.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 245.0, + 1382.0, + 245.0, + 1382.0, + 300.0, + 1228.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 287.0, + 593.0, + 287.0, + 593.0, + 338.0, + 380.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 305.0, + 908.0, + 305.0, + 908.0, + 334.0, + 805.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 288.0, + 1059.0, + 288.0, + 1059.0, + 337.0, + 922.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 294.0, + 1167.0, + 294.0, + 1167.0, + 326.0, + 1078.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 309.0, + 1281.0, + 309.0, + 1281.0, + 334.0, + 1230.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 308.0, + 1375.0, + 308.0, + 1375.0, + 335.0, + 1283.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 335.0, + 522.0, + 335.0, + 522.0, + 361.0, + 405.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 325.0, + 1162.0, + 325.0, + 1162.0, + 354.0, + 1080.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 343.0, + 1300.0, + 343.0, + 1300.0, + 358.0, + 1263.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 353.0, + 374.0, + 353.0, + 374.0, + 392.0, + 334.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 353.0, + 619.0, + 353.0, + 619.0, + 387.0, + 447.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 352.0, + 1166.0, + 352.0, + 1166.0, + 384.0, + 1076.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 358.0, + 1312.0, + 358.0, + 1312.0, + 377.0, + 1298.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 379.0, + 690.0, + 379.0, + 690.0, + 466.0, + 324.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 394.0, + 1043.0, + 394.0, + 1043.0, + 466.0, + 778.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 379.0, + 1170.0, + 379.0, + 1170.0, + 458.0, + 1073.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 392.0, + 1345.0, + 392.0, + 1345.0, + 455.0, + 1264.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 454.0, + 1326.0, + 454.0, + 1326.0, + 474.0, + 1284.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 474.0, + 879.0, + 474.0, + 879.0, + 490.0, + 865.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 478.0, + 936.0, + 478.0, + 936.0, + 491.0, + 918.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 480.0, + 978.0, + 480.0, + 978.0, + 490.0, + 964.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 473.0, + 1315.0, + 473.0, + 1315.0, + 493.0, + 1295.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 499.0, + 397.0, + 499.0, + 397.0, + 529.0, + 365.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 509.0, + 989.0, + 509.0, + 989.0, + 546.0, + 837.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 509.0, + 1077.0, + 509.0, + 1077.0, + 547.0, + 1036.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 497.0, + 1174.0, + 497.0, + 1174.0, + 518.0, + 1144.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 508.0, + 1345.0, + 508.0, + 1345.0, + 531.0, + 1264.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 537.0, + 1143.0, + 537.0, + 1143.0, + 559.0, + 1110.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 553.0, + 439.0, + 553.0, + 439.0, + 588.0, + 402.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 550.0, + 1099.0, + 550.0, + 1099.0, + 578.0, + 1021.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 588.0, + 862.0, + 588.0, + 862.0, + 600.0, + 852.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 588.0, + 976.0, + 588.0, + 976.0, + 599.0, + 967.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 574.0, + 1090.0, + 574.0, + 1090.0, + 598.0, + 1027.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 574.0, + 1391.0, + 574.0, + 1391.0, + 599.0, + 1212.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 593.0, + 1314.0, + 593.0, + 1314.0, + 618.0, + 1212.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 647.0, + 482.0, + 647.0, + 482.0, + 674.0, + 424.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 625.0, + 525.0, + 625.0, + 525.0, + 660.0, + 489.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 630.0, + 644.0, + 630.0, + 644.0, + 650.0, + 621.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 617.0, + 945.0, + 617.0, + 945.0, + 639.0, + 810.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 641.0, + 1175.0, + 641.0, + 1175.0, + 689.0, + 1046.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 612.0, + 1392.0, + 612.0, + 1392.0, + 679.0, + 1210.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.25, + 307.5, + 650.25, + 307.5, + 650.25, + 345.5, + 634.25, + 345.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 364.5, + 686.0, + 364.5, + 686.0, + 392.0, + 617.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 540.0, + 1324.0, + 540.0, + 1324.0, + 577.0, + 1285.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 725.0, + 1404.0, + 725.0, + 1404.0, + 762.0, + 294.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 755.0, + 1404.0, + 755.0, + 1404.0, + 793.0, + 294.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 787.0, + 1404.0, + 787.0, + 1404.0, + 823.0, + 295.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 816.0, + 1405.0, + 816.0, + 1405.0, + 851.0, + 294.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 847.0, + 1406.0, + 847.0, + 1406.0, + 885.0, + 291.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 880.0, + 1004.0, + 880.0, + 1004.0, + 912.0, + 296.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 678.0, + 1713.0, + 678.0, + 1754.0, + 294.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1324.0, + 948.0, + 1324.0, + 948.0, + 1360.0, + 292.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1382.0, + 1046.0, + 1382.0, + 1046.0, + 1419.0, + 294.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 1382.0, + 1405.0, + 1382.0, + 1405.0, + 1419.0, + 1183.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1416.0, + 602.0, + 1416.0, + 602.0, + 1453.0, + 293.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1416.0, + 904.0, + 1416.0, + 904.0, + 1453.0, + 698.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 1416.0, + 1405.0, + 1416.0, + 1405.0, + 1453.0, + 935.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1444.0, + 1404.0, + 1444.0, + 1404.0, + 1483.0, + 293.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1475.0, + 1404.0, + 1475.0, + 1404.0, + 1515.0, + 292.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1508.0, + 1404.0, + 1508.0, + 1404.0, + 1545.0, + 293.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1536.0, + 1037.0, + 1536.0, + 1037.0, + 1576.0, + 293.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1536.0, + 1404.0, + 1536.0, + 1404.0, + 1576.0, + 1138.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1568.0, + 1404.0, + 1568.0, + 1404.0, + 1605.0, + 293.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1598.0, + 1407.0, + 1598.0, + 1407.0, + 1636.0, + 292.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1631.0, + 945.0, + 1631.0, + 945.0, + 1664.0, + 296.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1631.0, + 1252.0, + 1631.0, + 1252.0, + 1664.0, + 1031.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1061.0, + 1405.0, + 1061.0, + 1405.0, + 1101.0, + 293.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1095.0, + 1404.0, + 1095.0, + 1404.0, + 1129.0, + 294.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1126.0, + 1402.0, + 1126.0, + 1402.0, + 1160.0, + 294.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1156.0, + 1404.0, + 1156.0, + 1404.0, + 1191.0, + 294.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1187.0, + 1404.0, + 1187.0, + 1404.0, + 1221.0, + 294.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1404.0, + 1216.0, + 1404.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1245.0, + 1379.0, + 1245.0, + 1379.0, + 1281.0, + 293.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1787.0, + 1403.0, + 1787.0, + 1403.0, + 1823.0, + 293.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1855.0, + 293.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1946.0, + 292.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1405.0, + 2002.0, + 1405.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 978.0, + 696.0, + 978.0, + 696.0, + 1014.0, + 296.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 978.0, + 1404.0, + 978.0, + 1404.0, + 1014.0, + 715.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1009.0, + 992.0, + 1009.0, + 992.0, + 1044.0, + 294.0, + 1044.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1318, + 1404, + 1318, + 1404, + 1715, + 298, + 1715 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 874, + 1404, + 874, + 1404, + 1089, + 298, + 1089 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1111, + 1402, + 1111, + 1402, + 1296, + 298, + 1296 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 929, + 521, + 1403, + 521, + 1403, + 748, + 929, + 748 + ], + "score": 0.977, + "html": "
Samples per partition
DatasetTrainVal.Test
CSN-Python412,17823,10722,176
CSN-Javascript123,8898,2536,483
CSN-Ruby48,7912,2092,279
CSN-Go317,83214,24214,291
Java-small691,97423,84457,088
" + }, + { + "category_id": 1, + "poly": [ + 298, + 315, + 1402, + 315, + 1402, + 469, + 298, + 469 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1888, + 1402, + 1888, + 1402, + 1982, + 300, + 1982 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 299, + 485, + 901, + 485, + 901, + 831, + 299, + 831 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1400, + 229, + 1400, + 292, + 294, + 292 + ], + "score": 0.953 + }, + { + "category_id": 2, + "poly": [ + 328, + 2006, + 856, + 2006, + 856, + 2034, + 328, + 2034 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 299, + 1761, + 480, + 1761, + 480, + 1797, + 299, + 1797 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 303, + 1833, + 834, + 1833, + 834, + 1861, + 303, + 1861 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 299, + 820, + 1221, + 820, + 1221, + 851, + 299, + 851 + ], + "score": 0.821 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.777 + }, + { + "category_id": 7, + "poly": [ + 1017, + 770, + 1309, + 770, + 1309, + 800, + 1017, + 800 + ], + "score": 0.308 + }, + { + "category_id": 13, + "poly": [ + 1203, + 1351, + 1230, + 1351, + 1230, + 1377, + 1203, + 1377 + ], + "score": 0.75, + "latex": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 2001.0, + 856.0, + 2001.0, + 856.0, + 2038.0, + 331.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1757.0, + 485.0, + 1757.0, + 485.0, + 1804.0, + 291.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1832.0, + 839.0, + 1832.0, + 839.0, + 1866.0, + 296.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 767.0, + 1312.0, + 767.0, + 1312.0, + 802.0, + 1014.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1318.0, + 1404.0, + 1318.0, + 1404.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1347.0, + 1202.0, + 1347.0, + 1202.0, + 1384.0, + 293.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1347.0, + 1406.0, + 1347.0, + 1406.0, + 1384.0, + 1231.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1379.0, + 1406.0, + 1379.0, + 1406.0, + 1416.0, + 294.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1409.0, + 1405.0, + 1409.0, + 1405.0, + 1444.0, + 292.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1440.0, + 1402.0, + 1440.0, + 1402.0, + 1476.0, + 294.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1471.0, + 1404.0, + 1471.0, + 1404.0, + 1507.0, + 294.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1500.0, + 1406.0, + 1500.0, + 1406.0, + 1539.0, + 296.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1529.0, + 1406.0, + 1529.0, + 1406.0, + 1571.0, + 292.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1563.0, + 1405.0, + 1563.0, + 1405.0, + 1599.0, + 293.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1592.0, + 1405.0, + 1592.0, + 1405.0, + 1632.0, + 293.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1623.0, + 1407.0, + 1623.0, + 1407.0, + 1659.0, + 292.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1652.0, + 1407.0, + 1652.0, + 1407.0, + 1691.0, + 293.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1685.0, + 1379.0, + 1685.0, + 1379.0, + 1721.0, + 293.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 871.0, + 1405.0, + 871.0, + 1405.0, + 913.0, + 293.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 900.0, + 1405.0, + 900.0, + 1405.0, + 943.0, + 292.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 937.0, + 1402.0, + 937.0, + 1402.0, + 968.0, + 296.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 966.0, + 1405.0, + 966.0, + 1405.0, + 1000.0, + 293.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 995.0, + 1405.0, + 995.0, + 1405.0, + 1029.0, + 293.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1023.0, + 1406.0, + 1023.0, + 1406.0, + 1064.0, + 292.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1055.0, + 1345.0, + 1055.0, + 1345.0, + 1096.0, + 292.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1113.0, + 1403.0, + 1113.0, + 1403.0, + 1145.0, + 294.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1144.0, + 1404.0, + 1144.0, + 1404.0, + 1176.0, + 296.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1174.0, + 1404.0, + 1174.0, + 1404.0, + 1206.0, + 296.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1202.0, + 1407.0, + 1202.0, + 1407.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1235.0, + 1404.0, + 1235.0, + 1404.0, + 1267.0, + 296.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1265.0, + 891.0, + 1265.0, + 891.0, + 1300.0, + 296.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 317.0, + 1404.0, + 317.0, + 1404.0, + 350.0, + 296.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 346.0, + 1402.0, + 346.0, + 1402.0, + 383.0, + 294.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 373.0, + 1406.0, + 373.0, + 1406.0, + 414.0, + 292.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 407.0, + 1404.0, + 407.0, + 1404.0, + 444.0, + 294.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 437.0, + 1375.0, + 437.0, + 1375.0, + 474.0, + 294.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1886.0, + 1406.0, + 1886.0, + 1406.0, + 1924.0, + 295.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1917.0, + 1404.0, + 1917.0, + 1404.0, + 1954.0, + 294.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1951.0, + 1404.0, + 1951.0, + 1404.0, + 1982.0, + 295.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 484.0, + 902.0, + 484.0, + 902.0, + 517.0, + 296.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 514.0, + 904.0, + 514.0, + 904.0, + 550.0, + 295.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 546.0, + 904.0, + 546.0, + 904.0, + 579.0, + 295.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 577.0, + 903.0, + 577.0, + 903.0, + 608.0, + 295.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 607.0, + 904.0, + 607.0, + 904.0, + 639.0, + 294.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 636.0, + 903.0, + 636.0, + 903.0, + 669.0, + 294.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 668.0, + 904.0, + 668.0, + 904.0, + 702.0, + 296.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 698.0, + 905.0, + 698.0, + 905.0, + 731.0, + 293.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 905.0, + 728.0, + 905.0, + 764.0, + 294.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 758.0, + 904.0, + 758.0, + 904.0, + 794.0, + 292.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 790.0, + 903.0, + 790.0, + 903.0, + 821.0, + 296.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 822.0, + 545.0, + 822.0, + 545.0, + 833.0, + 529.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 821.0, + 622.0, + 821.0, + 622.0, + 834.0, + 574.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 824.0, + 675.0, + 824.0, + 675.0, + 831.0, + 666.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 819.0, + 880.0, + 819.0, + 880.0, + 836.0, + 845.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1087.0, + 261.0, + 1087.0, + 296.0, + 296.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 817.0, + 1226.0, + 817.0, + 1226.0, + 855.0, + 292.0, + 855.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 649, + 1404, + 649, + 1404, + 925, + 297, + 925 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 300, + 177, + 1399, + 177, + 1399, + 534, + 300, + 534 + ], + "score": 0.982, + "html": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq35.7924.8529.3430.1819.8823.9723.2310.3114.2852.3043.4347.45
GREAT35.0731.5933.2431.2026.8428.8624.6422.2323.3850.0146.5148.20
Ours w/o structure37.3831.9834.4733.1726.7029.5929.8525.8727.7251.7847.5749.59
Ours w/o pointer net37.7431.8534.5533.1228.7030.7523.3225.2124.2354.3150.1252.13
Ours36.4033.6634.9735.0629.6132.1131.4224.4627.5055.1048.0551.34
code2seq (Multilanguage)34.4925.4929.3231.6222.1626.0623.9717.0619.9352.7044.3648.17
GREAT (Multilanguage)36.7531.5433.9433.5827.7830.4130.0524.3326.8952.6548.3050.38
Ours w/o structure (Mult.)38.4830.1433.8035.3827.4130.8932.6126.7629.4055.0347.3450.90
Ours w/o pointer (Mult.)38.9133.1235.7837.2129.7533.0734.5227.3130.5056.0750.7653.28
Ours (Multilanguage)38.8933.8236.1836.9529.9833.1033.9328.9431.2456.0050.4453.07
Ours (Mult.+ Finetune)39.8532.7935.9837.0029.7933.0035.8527.7531.2855.6351.1253.28
Ours (Mult. + LM Pretrain)39.6735.2937.3537.0631.9434.3135.1929.3632.0157.7351.8954.65
" + }, + { + "category_id": 5, + "poly": [ + 877, + 1285, + 1378, + 1285, + 1378, + 1677, + 877, + 1677 + ], + "score": 0.98, + "html": "
ModelPrec.Rec.F1
Without pointer net code2seq Ours w/o structure Ours w/o context51.23 50.70 51.8137.31 45.49 46.0443.18 47.96
Ours With pointer net50.3346.8048.75 48.50
Fernandes et al. (2019)=151.4
GREAT Ours w/o structure53.6046.4149.75
Ours w/o context55.4846.0750.34
54.4545.2949.45
Ours54.8549.8452.22
Ours + Pretrain57.0250.8753.77
" + }, + { + "category_id": 1, + "poly": [ + 298, + 959, + 1403, + 959, + 1403, + 1174, + 298, + 1174 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1287, + 822, + 1287, + 822, + 1741, + 299, + 1741 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1402, + 1881, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1743, + 1403, + 1743, + 1403, + 1865, + 299, + 1865 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 299, + 1208, + 1402, + 1208, + 1402, + 1270, + 299, + 1270 + ], + "score": 0.939 + }, + { + "category_id": 6, + "poly": [ + 464, + 557, + 1230, + 557, + 1230, + 589, + 464, + 589 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 105, + 300, + 105 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.655 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.234 + }, + { + "category_id": 7, + "poly": [ + 846, + 1700, + 1400, + 1700, + 1400, + 1730, + 846, + 1730 + ], + "score": 0.174 + }, + { + "category_id": 13, + "poly": [ + 416, + 486, + 433, + 486, + 433, + 504, + 416, + 504 + ], + "score": 0.65, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 416, + 508, + 433, + 508, + 433, + 527, + 416, + 527 + ], + "score": 0.41, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 552.0, + 1233.0, + 552.0, + 1233.0, + 594.0, + 464.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 294.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1696.0, + 1403.0, + 1696.0, + 1403.0, + 1734.0, + 843.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 651.0, + 1401.0, + 651.0, + 1401.0, + 684.0, + 295.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 679.0, + 1405.0, + 679.0, + 1405.0, + 717.0, + 294.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 1406.0, + 710.0, + 1406.0, + 748.0, + 294.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 740.0, + 1405.0, + 740.0, + 1405.0, + 777.0, + 294.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 773.0, + 1404.0, + 773.0, + 1404.0, + 805.0, + 295.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 801.0, + 1405.0, + 801.0, + 1405.0, + 838.0, + 294.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 833.0, + 1406.0, + 833.0, + 1406.0, + 869.0, + 295.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 863.0, + 1406.0, + 863.0, + 1406.0, + 899.0, + 294.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 890.0, + 1248.0, + 890.0, + 1248.0, + 932.0, + 294.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 958.0, + 1403.0, + 958.0, + 1403.0, + 993.0, + 294.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 990.0, + 1404.0, + 990.0, + 1404.0, + 1026.0, + 293.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1021.0, + 1404.0, + 1021.0, + 1404.0, + 1055.0, + 294.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1049.0, + 1407.0, + 1049.0, + 1407.0, + 1090.0, + 292.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1080.0, + 1405.0, + 1080.0, + 1405.0, + 1118.0, + 293.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1111.0, + 1404.0, + 1111.0, + 1404.0, + 1148.0, + 292.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1143.0, + 1375.0, + 1143.0, + 1375.0, + 1181.0, + 294.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1286.0, + 824.0, + 1286.0, + 824.0, + 1317.0, + 297.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1315.0, + 825.0, + 1315.0, + 825.0, + 1349.0, + 295.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1345.0, + 826.0, + 1345.0, + 826.0, + 1383.0, + 293.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1377.0, + 826.0, + 1377.0, + 826.0, + 1410.0, + 293.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1407.0, + 826.0, + 1407.0, + 826.0, + 1441.0, + 294.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1439.0, + 827.0, + 1439.0, + 827.0, + 1470.0, + 294.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1469.0, + 826.0, + 1469.0, + 826.0, + 1499.0, + 295.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1497.0, + 827.0, + 1497.0, + 827.0, + 1534.0, + 293.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1530.0, + 826.0, + 1530.0, + 826.0, + 1562.0, + 295.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1561.0, + 826.0, + 1561.0, + 826.0, + 1592.0, + 295.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1590.0, + 825.0, + 1590.0, + 825.0, + 1625.0, + 293.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1622.0, + 826.0, + 1622.0, + 826.0, + 1652.0, + 294.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1653.0, + 826.0, + 1653.0, + 826.0, + 1684.0, + 293.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 825.0, + 1681.0, + 825.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1710.0, + 826.0, + 1710.0, + 826.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1884.0, + 1403.0, + 1884.0, + 1403.0, + 1913.0, + 297.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1943.0, + 1403.0, + 1943.0, + 1403.0, + 1976.0, + 297.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2034.0, + 294.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1742.0, + 1405.0, + 1742.0, + 1405.0, + 1777.0, + 295.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1776.0, + 1402.0, + 1776.0, + 1402.0, + 1805.0, + 298.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 1404.0, + 1805.0, + 1404.0, + 1838.0, + 295.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1835.0, + 1361.0, + 1835.0, + 1361.0, + 1866.0, + 294.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1207.0, + 1404.0, + 1207.0, + 1404.0, + 1243.0, + 297.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 731.0, + 1239.0, + 731.0, + 1272.0, + 294.0, + 1272.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1246, + 1403, + 1246, + 1403, + 1429, + 298, + 1429 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1015, + 1403, + 1015, + 1403, + 1230, + 299, + 1230 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 755, + 1404, + 755, + 1404, + 1000, + 297, + 1000 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1570, + 1404, + 1570, + 1404, + 1845, + 298, + 1845 + ], + "score": 0.975 + }, + { + "category_id": 5, + "poly": [ + 985, + 406, + 1379, + 406, + 1379, + 591, + 985, + 591 + ], + "score": 0.971, + "html": "
AST distanceF1 score
Sibling shortest paths46.17
Ancestor shortest paths47.89
Shortest paths47.76
Personalized PageRank48.47
All the above (c.f. Table 3)48.75
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1444, + 1402, + 1444, + 1402, + 1537, + 300, + 1537 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 299, + 416, + 940, + 416, + 940, + 661, + 299, + 661 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 300, + 1861, + 1401, + 1861, + 1401, + 1985, + 300, + 1985 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 301, + 229, + 1402, + 229, + 1402, + 323, + 301, + 323 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 662, + 1403, + 662, + 1403, + 722, + 299, + 722 + ], + "score": 0.942 + }, + { + "category_id": 2, + "poly": [ + 327, + 2006, + 1110, + 2006, + 1110, + 2034, + 327, + 2034 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 301, + 360, + 836, + 360, + 836, + 389, + 301, + 389 + ], + "score": 0.857 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.794 + }, + { + "category_id": 6, + "poly": [ + 976, + 614, + 1393, + 614, + 1393, + 645, + 976, + 645 + ], + "score": 0.727 + }, + { + "category_id": 13, + "poly": [ + 1320, + 1506, + 1387, + 1506, + 1387, + 1535, + 1320, + 1535 + ], + "score": 0.9, + "latex": "122 \\%" + }, + { + "category_id": 13, + "poly": [ + 471, + 1506, + 537, + 1506, + 537, + 1535, + 471, + 1535 + ], + "score": 0.86, + "latex": "112 \\%" + }, + { + "category_id": 13, + "poly": [ + 1174, + 1474, + 1204, + 1474, + 1204, + 1504, + 1174, + 1504 + ], + "score": 0.67, + "latex": "2 ^ { 4 }" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1999.0, + 1113.0, + 1999.0, + 1113.0, + 2040.0, + 329.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 358.0, + 840.0, + 358.0, + 840.0, + 394.0, + 295.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 607.0, + 1394.0, + 607.0, + 1394.0, + 651.0, + 970.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1247.0, + 1404.0, + 1247.0, + 1404.0, + 1278.0, + 296.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1276.0, + 1404.0, + 1276.0, + 1404.0, + 1307.0, + 296.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1303.0, + 1405.0, + 1303.0, + 1405.0, + 1343.0, + 293.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1336.0, + 1403.0, + 1336.0, + 1403.0, + 1372.0, + 293.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1365.0, + 1407.0, + 1365.0, + 1407.0, + 1404.0, + 293.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1396.0, + 1375.0, + 1396.0, + 1375.0, + 1435.0, + 293.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1013.0, + 1403.0, + 1013.0, + 1403.0, + 1050.0, + 293.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1049.0, + 1404.0, + 1049.0, + 1404.0, + 1079.0, + 295.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1077.0, + 1405.0, + 1077.0, + 1405.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1108.0, + 1405.0, + 1108.0, + 1405.0, + 1142.0, + 295.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1134.0, + 1407.0, + 1134.0, + 1407.0, + 1174.0, + 293.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1168.0, + 1405.0, + 1168.0, + 1405.0, + 1202.0, + 295.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1199.0, + 994.0, + 1199.0, + 994.0, + 1233.0, + 295.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 754.0, + 1404.0, + 754.0, + 1404.0, + 789.0, + 294.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 1404.0, + 788.0, + 1404.0, + 821.0, + 294.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 815.0, + 1406.0, + 815.0, + 1406.0, + 852.0, + 294.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 845.0, + 1406.0, + 845.0, + 1406.0, + 884.0, + 292.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 875.0, + 1406.0, + 875.0, + 1406.0, + 913.0, + 292.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 904.0, + 1406.0, + 904.0, + 1406.0, + 946.0, + 292.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 937.0, + 1406.0, + 937.0, + 1406.0, + 975.0, + 294.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 967.0, + 1404.0, + 967.0, + 1404.0, + 1006.0, + 292.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1569.0, + 1405.0, + 1569.0, + 1405.0, + 1606.0, + 296.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1602.0, + 1404.0, + 1602.0, + 1404.0, + 1635.0, + 296.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1629.0, + 1406.0, + 1629.0, + 1406.0, + 1668.0, + 293.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1700.0, + 292.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1692.0, + 1406.0, + 1692.0, + 1406.0, + 1728.0, + 293.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1721.0, + 1406.0, + 1721.0, + 1406.0, + 1758.0, + 293.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1752.0, + 1406.0, + 1752.0, + 1406.0, + 1790.0, + 292.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1778.0, + 1405.0, + 1778.0, + 1405.0, + 1822.0, + 291.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1812.0, + 1404.0, + 1812.0, + 1404.0, + 1851.0, + 293.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1445.0, + 1404.0, + 1445.0, + 1404.0, + 1478.0, + 295.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1474.0, + 1173.0, + 1474.0, + 1173.0, + 1510.0, + 293.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1474.0, + 1406.0, + 1474.0, + 1406.0, + 1510.0, + 1205.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1506.0, + 470.0, + 1506.0, + 470.0, + 1540.0, + 295.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 1506.0, + 1319.0, + 1506.0, + 1319.0, + 1540.0, + 538.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1388.0, + 1506.0, + 1398.0, + 1506.0, + 1398.0, + 1540.0, + 1388.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 416.0, + 943.0, + 416.0, + 943.0, + 449.0, + 295.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 446.0, + 943.0, + 446.0, + 943.0, + 481.0, + 294.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 476.0, + 944.0, + 476.0, + 944.0, + 511.0, + 293.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 507.0, + 944.0, + 507.0, + 944.0, + 540.0, + 294.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 944.0, + 539.0, + 944.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 570.0, + 945.0, + 570.0, + 945.0, + 601.0, + 292.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 599.0, + 944.0, + 599.0, + 944.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 631.0, + 945.0, + 631.0, + 945.0, + 664.0, + 296.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1860.0, + 1406.0, + 1860.0, + 1406.0, + 1896.0, + 295.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1891.0, + 1406.0, + 1891.0, + 1406.0, + 1928.0, + 295.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1923.0, + 1403.0, + 1923.0, + 1403.0, + 1956.0, + 296.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1955.0, + 1403.0, + 1955.0, + 1403.0, + 1987.0, + 296.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 230.0, + 1407.0, + 230.0, + 1407.0, + 264.0, + 293.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 297.0, + 295.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 289.0, + 1028.0, + 289.0, + 1028.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 658.0, + 1403.0, + 658.0, + 1403.0, + 696.0, + 294.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 1221.0, + 686.0, + 1221.0, + 728.0, + 293.0, + 728.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1703, + 1403, + 1703, + 1403, + 2010, + 298, + 2010 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1175, + 1403, + 1175, + 1403, + 1482, + 298, + 1482 + ], + "score": 0.981 + }, + { + "category_id": 3, + "poly": [ + 383, + 178, + 1319, + 178, + 1319, + 671, + 383, + 671 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 301, + 1497, + 1400, + 1497, + 1400, + 1590, + 301, + 1590 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 296, + 993, + 1405, + 993, + 1405, + 1115, + 296, + 1115 + ], + "score": 0.948 + }, + { + "category_id": 4, + "poly": [ + 297, + 688, + 1396, + 688, + 1396, + 721, + 297, + 721 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 299, + 1636, + 543, + 1636, + 543, + 1670, + 299, + 1670 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 812, + 76, + 812, + 104, + 298, + 104 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.787 + }, + { + "category_id": 3, + "poly": [ + 348, + 751, + 1320, + 751, + 1320, + 967, + 348, + 967 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 349, + 750, + 759, + 750, + 759, + 970, + 349, + 970 + ], + "score": 0.374 + }, + { + "category_id": 13, + "poly": [ + 929, + 830, + 999, + 830, + 999, + 848, + 929, + 848 + ], + "score": 0.64, + "latex": "\\scriptstyle { \\begin{array} { r l } { S } & { { } = = = } \\\\ { \\sqrt { \\phantom { } } } \\end{array} }" + }, + { + "category_id": 13, + "poly": [ + 923, + 807, + 985, + 807, + 985, + 826, + 923, + 826 + ], + "score": 0.62, + "latex": "\\scriptstyle { \\begin{array} { r l } { \\mathbf { S } } & { { } = = = } \\\\ { \\mathbf { S } } & { { } = = = } \\end{array} }" + }, + { + "category_id": 13, + "poly": [ + 1114, + 806, + 1216, + 806, + 1216, + 826, + 1114, + 826 + ], + "score": 0.53, + "latex": "s \\ \\mathrm { ~ \\ = = = ~ \\ ~ } \\cdot \\mathrm { ~ f ~ }" + }, + { + "category_id": 13, + "poly": [ + 930, + 852, + 997, + 852, + 997, + 871, + 930, + 871 + ], + "score": 0.5, + "latex": "\\scriptstyle { \\begin{array} { r l } { S } & { { } = = = } \\\\ { \\sqrt { \\phantom { } } } \\end{array} }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 850, + 1217, + 850, + 1217, + 872, + 1083, + 872 + ], + "score": 0.48, + "latex": "| \\mathrm { ~ ~ { ~ \\cal ~ s ~ } ~ } = = { } | \\mathrm { ~ ~ { ~ \\cal ~ O ~ } ~ } ^ { \\prime } |" + }, + { + "category_id": 13, + "poly": [ + 895, + 786, + 915, + 786, + 915, + 804, + 895, + 804 + ], + "score": 0.26, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 182.0, + 706.0, + 182.0, + 706.0, + 212.0, + 628.0, + 212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 750.0, + 194.0, + 760.0, + 194.0, + 760.0, + 204.0, + 750.0, + 204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 184.0, + 882.0, + 184.0, + 882.0, + 212.0, + 780.0, + 212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 190.0, + 943.0, + 190.0, + 943.0, + 208.0, + 920.0, + 208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 186.0, + 998.0, + 186.0, + 998.0, + 213.0, + 957.0, + 213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 181.0, + 1122.0, + 181.0, + 1122.0, + 215.0, + 1065.0, + 215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 240.0, + 1312.0, + 240.0, + 1312.0, + 271.0, + 1052.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 271.0, + 1301.0, + 271.0, + 1301.0, + 301.0, + 1050.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 993.0, + 1406.0, + 993.0, + 1406.0, + 1029.0, + 294.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1022.0, + 1407.0, + 1022.0, + 1407.0, + 1059.0, + 291.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1055.0, + 1406.0, + 1055.0, + 1406.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1084.0, + 1257.0, + 1084.0, + 1257.0, + 1120.0, + 293.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 684.0, + 1401.0, + 684.0, + 1401.0, + 727.0, + 297.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1629.0, + 549.0, + 1629.0, + 549.0, + 1679.0, + 292.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 747.0, + 552.0, + 747.0, + 552.0, + 775.0, + 347.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 769.0, + 527.0, + 769.0, + 527.0, + 797.0, + 370.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 761.0, + 1322.0, + 761.0, + 1322.0, + 785.0, + 852.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 792.0, + 555.0, + 792.0, + 555.0, + 819.0, + 370.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 792.0, + 612.0, + 792.0, + 612.0, + 819.0, + 568.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 793.0, + 690.0, + 793.0, + 690.0, + 818.0, + 626.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 783.0, + 894.0, + 783.0, + 894.0, + 807.0, + 871.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 783.0, + 1111.0, + 783.0, + 1111.0, + 807.0, + 916.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 815.0, + 566.0, + 815.0, + 566.0, + 863.0, + 369.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 837.0, + 624.0, + 837.0, + 624.0, + 862.0, + 581.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 837.0, + 702.0, + 837.0, + 702.0, + 862.0, + 637.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 803.0, + 922.0, + 803.0, + 922.0, + 851.0, + 871.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 803.0, + 1076.0, + 803.0, + 1076.0, + 851.0, + 1000.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 810.0, + 1156.0, + 810.0, + 1156.0, + 850.0, + 1086.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 804.0, + 1228.0, + 804.0, + 1228.0, + 851.0, + 1217.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 849.0, + 929.0, + 849.0, + 929.0, + 872.0, + 896.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 849.0, + 1065.0, + 849.0, + 1065.0, + 872.0, + 998.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 848.0, + 1082.0, + 848.0, + 1082.0, + 873.0, + 1071.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 848.0, + 1252.0, + 848.0, + 1252.0, + 873.0, + 1218.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 860.0, + 543.0, + 860.0, + 543.0, + 884.0, + 393.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 882.0, + 564.0, + 882.0, + 564.0, + 906.0, + 371.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 870.0, + 1052.0, + 870.0, + 1052.0, + 898.0, + 896.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 900.0, + 766.0, + 900.0, + 766.0, + 955.0, + 392.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 943.0, + 863.0, + 943.0, + 863.0, + 958.0, + 854.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 914.0, + 1017.0, + 914.0, + 1017.0, + 941.0, + 872.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 953.0, + 386.0, + 953.0, + 386.0, + 969.0, + 372.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1703.0, + 1404.0, + 1703.0, + 1404.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1736.0, + 1405.0, + 1736.0, + 1405.0, + 1771.0, + 296.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1765.0, + 1405.0, + 1765.0, + 1405.0, + 1801.0, + 293.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1797.0, + 1405.0, + 1797.0, + 1405.0, + 1832.0, + 294.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1826.0, + 1405.0, + 1826.0, + 1405.0, + 1862.0, + 293.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1857.0, + 1405.0, + 1857.0, + 1405.0, + 1891.0, + 292.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1885.0, + 1408.0, + 1885.0, + 1408.0, + 1925.0, + 291.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1918.0, + 1405.0, + 1918.0, + 1405.0, + 1953.0, + 294.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1948.0, + 1405.0, + 1948.0, + 1405.0, + 1984.0, + 294.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1979.0, + 1157.0, + 1979.0, + 1157.0, + 2014.0, + 294.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1173.0, + 1404.0, + 1173.0, + 1404.0, + 1214.0, + 292.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1206.0, + 1407.0, + 1206.0, + 1407.0, + 1243.0, + 291.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1237.0, + 1405.0, + 1237.0, + 1405.0, + 1273.0, + 293.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1270.0, + 1404.0, + 1270.0, + 1404.0, + 1302.0, + 296.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1405.0, + 1298.0, + 1405.0, + 1334.0, + 294.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 1407.0, + 1330.0, + 1407.0, + 1365.0, + 294.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1359.0, + 1405.0, + 1359.0, + 1405.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1425.0, + 294.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1419.0, + 1407.0, + 1419.0, + 1407.0, + 1458.0, + 293.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1451.0, + 1156.0, + 1451.0, + 1156.0, + 1486.0, + 293.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1494.0, + 1405.0, + 1494.0, + 1405.0, + 1534.0, + 295.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1528.0, + 1404.0, + 1528.0, + 1404.0, + 1564.0, + 295.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1558.0, + 1394.0, + 1558.0, + 1394.0, + 1595.0, + 295.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 750.0, + 552.0, + 750.0, + 552.0, + 775.0, + 348.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 772.0, + 527.0, + 772.0, + 527.0, + 797.0, + 369.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 793.0, + 561.0, + 793.0, + 561.0, + 818.0, + 371.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 794.0, + 612.0, + 794.0, + 612.0, + 815.0, + 569.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 792.0, + 689.0, + 792.0, + 689.0, + 817.0, + 624.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 815.0, + 529.0, + 815.0, + 529.0, + 840.0, + 393.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 837.0, + 575.0, + 837.0, + 575.0, + 862.0, + 371.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 834.0, + 701.0, + 834.0, + 701.0, + 863.0, + 578.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 860.0, + 543.0, + 860.0, + 543.0, + 885.0, + 393.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 882.0, + 564.0, + 882.0, + 564.0, + 907.0, + 371.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 903.0, + 694.0, + 903.0, + 694.0, + 931.0, + 395.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 905.0, + 760.0, + 905.0, + 760.0, + 929.0, + 697.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 922.0, + 686.0, + 922.0, + 686.0, + 954.0, + 393.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 957.0, + 384.0, + 957.0, + 384.0, + 966.0, + 375.0, + 966.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1231, + 1403, + 1231, + 1403, + 1326, + 298, + 1326 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 297, + 1426, + 1405, + 1426, + 1405, + 1551, + 297, + 1551 + ], + "score": 0.903 + }, + { + "category_id": 1, + "poly": [ + 298, + 1057, + 1404, + 1057, + 1404, + 1213, + 298, + 1213 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 105, + 300, + 105 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 297, + 279, + 1404, + 279, + 1404, + 463, + 297, + 463 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 296, + 1343, + 1401, + 1343, + 1401, + 1409, + 296, + 1409 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 297, + 1653, + 1404, + 1653, + 1404, + 1808, + 297, + 1808 + ], + "score": 0.865 + }, + { + "category_id": 1, + "poly": [ + 297, + 831, + 1399, + 831, + 1399, + 924, + 297, + 924 + ], + "score": 0.861 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.856 + }, + { + "category_id": 1, + "poly": [ + 296, + 944, + 1400, + 944, + 1400, + 1039, + 296, + 1039 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 296, + 657, + 1405, + 657, + 1405, + 811, + 296, + 811 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 300, + 1910, + 1403, + 1910, + 1403, + 2035, + 300, + 2035 + ], + "score": 0.841 + }, + { + "category_id": 1, + "poly": [ + 289, + 1827, + 1401, + 1827, + 1401, + 1893, + 289, + 1893 + ], + "score": 0.829 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.822 + }, + { + "category_id": 1, + "poly": [ + 297, + 483, + 1404, + 483, + 1404, + 637, + 297, + 637 + ], + "score": 0.815 + }, + { + "category_id": 1, + "poly": [ + 295, + 1570, + 1403, + 1570, + 1403, + 1635, + 295, + 1635 + ], + "score": 0.813 + }, + { + "category_id": 13, + "poly": [ + 742, + 1011, + 760, + 1011, + 760, + 1032, + 742, + 1032 + ], + "score": 0.52, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 904, + 1297, + 922, + 1297, + 922, + 1319, + 904, + 1319 + ], + "score": 0.47, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 294.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1229.0, + 1403.0, + 1229.0, + 1403.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1263.0, + 1404.0, + 1263.0, + 1404.0, + 1297.0, + 323.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1291.0, + 903.0, + 1291.0, + 903.0, + 1331.0, + 322.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1291.0, + 1100.0, + 1291.0, + 1100.0, + 1331.0, + 923.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1428.0, + 1403.0, + 1428.0, + 1403.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1458.0, + 1405.0, + 1458.0, + 1405.0, + 1491.0, + 323.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1487.0, + 1406.0, + 1487.0, + 1406.0, + 1526.0, + 319.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1515.0, + 399.0, + 1515.0, + 399.0, + 1555.0, + 320.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1059.0, + 1402.0, + 1059.0, + 1402.0, + 1092.0, + 297.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1088.0, + 1404.0, + 1088.0, + 1404.0, + 1126.0, + 322.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1120.0, + 1405.0, + 1120.0, + 1405.0, + 1155.0, + 321.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1186.0, + 323.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1178.0, + 1365.0, + 1178.0, + 1365.0, + 1216.0, + 322.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 277.0, + 1405.0, + 277.0, + 1405.0, + 314.0, + 292.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 310.0, + 1401.0, + 310.0, + 1401.0, + 342.0, + 325.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 341.0, + 1404.0, + 341.0, + 1404.0, + 376.0, + 322.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 368.0, + 1405.0, + 368.0, + 1405.0, + 406.0, + 320.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 400.0, + 1406.0, + 400.0, + 1406.0, + 434.0, + 322.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 433.0, + 986.0, + 433.0, + 986.0, + 465.0, + 324.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1380.0, + 296.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1373.0, + 1405.0, + 1373.0, + 1405.0, + 1412.0, + 320.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1653.0, + 1405.0, + 1653.0, + 1405.0, + 1691.0, + 295.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1684.0, + 1406.0, + 1684.0, + 1406.0, + 1723.0, + 321.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1716.0, + 1406.0, + 1716.0, + 1406.0, + 1753.0, + 324.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1742.0, + 1404.0, + 1742.0, + 1404.0, + 1786.0, + 322.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1777.0, + 886.0, + 1777.0, + 886.0, + 1810.0, + 325.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 829.0, + 1404.0, + 829.0, + 1404.0, + 867.0, + 293.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 863.0, + 1404.0, + 863.0, + 1404.0, + 897.0, + 322.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 888.0, + 398.0, + 888.0, + 398.0, + 927.0, + 320.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 940.0, + 1405.0, + 940.0, + 1405.0, + 983.0, + 294.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 976.0, + 1401.0, + 976.0, + 1401.0, + 1010.0, + 324.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1006.0, + 741.0, + 1006.0, + 741.0, + 1041.0, + 326.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1006.0, + 923.0, + 1006.0, + 923.0, + 1041.0, + 761.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 657.0, + 1405.0, + 657.0, + 1405.0, + 694.0, + 294.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 688.0, + 1405.0, + 688.0, + 1405.0, + 725.0, + 323.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 713.0, + 1406.0, + 713.0, + 1406.0, + 760.0, + 320.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 747.0, + 1402.0, + 747.0, + 1402.0, + 786.0, + 320.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 779.0, + 789.0, + 779.0, + 789.0, + 815.0, + 322.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1978.0, + 321.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2008.0, + 320.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2037.0, + 321.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1825.0, + 1403.0, + 1825.0, + 1403.0, + 1864.0, + 293.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1859.0, + 1032.0, + 1859.0, + 1032.0, + 1894.0, + 321.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 478.0, + 1406.0, + 478.0, + 1406.0, + 524.0, + 292.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 514.0, + 1404.0, + 514.0, + 1404.0, + 551.0, + 322.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 544.0, + 1404.0, + 544.0, + 1404.0, + 578.0, + 320.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 575.0, + 1406.0, + 575.0, + 1406.0, + 609.0, + 321.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 605.0, + 988.0, + 605.0, + 988.0, + 641.0, + 320.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1572.0, + 1404.0, + 1572.0, + 1404.0, + 1608.0, + 297.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1603.0, + 1245.0, + 1603.0, + 1245.0, + 1636.0, + 322.0, + 1636.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1402, + 1942, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 312, + 228, + 1398, + 228, + 1398, + 294, + 312, + 294 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.819 + }, + { + "category_id": 1, + "poly": [ + 295, + 314, + 1400, + 314, + 1400, + 377, + 295, + 377 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 294, + 1651, + 1399, + 1651, + 1399, + 1715, + 294, + 1715 + ], + "score": 0.782 + }, + { + "category_id": 1, + "poly": [ + 298, + 1734, + 1404, + 1734, + 1404, + 1920, + 298, + 1920 + ], + "score": 0.756 + }, + { + "category_id": 1, + "poly": [ + 295, + 1339, + 1398, + 1339, + 1398, + 1404, + 295, + 1404 + ], + "score": 0.749 + }, + { + "category_id": 1, + "poly": [ + 297, + 397, + 1404, + 397, + 1404, + 552, + 297, + 552 + ], + "score": 0.739 + }, + { + "category_id": 1, + "poly": [ + 293, + 882, + 1401, + 882, + 1401, + 977, + 293, + 977 + ], + "score": 0.737 + }, + { + "category_id": 1, + "poly": [ + 299, + 1225, + 1400, + 1225, + 1400, + 1319, + 299, + 1319 + ], + "score": 0.731 + }, + { + "category_id": 1, + "poly": [ + 298, + 1537, + 1404, + 1537, + 1404, + 1631, + 298, + 1631 + ], + "score": 0.73 + }, + { + "category_id": 1, + "poly": [ + 302, + 1110, + 1396, + 1110, + 1396, + 1206, + 302, + 1206 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 296, + 571, + 1398, + 571, + 1398, + 637, + 296, + 637 + ], + "score": 0.71 + }, + { + "category_id": 1, + "poly": [ + 302, + 996, + 1399, + 996, + 1399, + 1093, + 302, + 1093 + ], + "score": 0.688 + }, + { + "category_id": 1, + "poly": [ + 298, + 1422, + 1404, + 1422, + 1404, + 1517, + 298, + 1517 + ], + "score": 0.677 + }, + { + "category_id": 1, + "poly": [ + 295, + 654, + 1400, + 654, + 1400, + 749, + 295, + 749 + ], + "score": 0.619 + }, + { + "category_id": 1, + "poly": [ + 295, + 768, + 1405, + 768, + 1405, + 863, + 295, + 863 + ], + "score": 0.251 + }, + { + "category_id": 13, + "poly": [ + 742, + 1178, + 760, + 1178, + 760, + 1199, + 742, + 1199 + ], + "score": 0.51, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 904, + 1490, + 921, + 1490, + 921, + 1512, + 904, + 1512 + ], + "score": 0.44, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1402.0, + 1941.0, + 1402.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1971.0, + 1403.0, + 1971.0, + 1403.0, + 2005.0, + 324.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2001.0, + 612.0, + 2001.0, + 612.0, + 2035.0, + 322.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 229.0, + 1403.0, + 229.0, + 1403.0, + 264.0, + 323.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 262.0, + 1048.0, + 262.0, + 1048.0, + 295.0, + 325.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 312.0, + 1405.0, + 312.0, + 1405.0, + 348.0, + 295.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 346.0, + 1003.0, + 346.0, + 1003.0, + 377.0, + 323.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 1404.0, + 1650.0, + 1404.0, + 1689.0, + 294.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1682.0, + 533.0, + 1682.0, + 533.0, + 1715.0, + 324.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1735.0, + 1406.0, + 1735.0, + 1406.0, + 1771.0, + 296.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1766.0, + 1404.0, + 1766.0, + 1404.0, + 1801.0, + 323.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1795.0, + 1405.0, + 1795.0, + 1405.0, + 1832.0, + 321.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1828.0, + 1402.0, + 1828.0, + 1402.0, + 1860.0, + 322.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1856.0, + 1403.0, + 1856.0, + 1403.0, + 1893.0, + 321.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1886.0, + 1038.0, + 1886.0, + 1038.0, + 1923.0, + 323.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1335.0, + 1403.0, + 1335.0, + 1403.0, + 1379.0, + 293.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1372.0, + 589.0, + 1372.0, + 589.0, + 1403.0, + 321.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 396.0, + 833.0, + 396.0, + 833.0, + 429.0, + 297.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 395.0, + 1404.0, + 395.0, + 1404.0, + 432.0, + 901.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 429.0, + 380.0, + 429.0, + 380.0, + 463.0, + 321.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 424.0, + 1405.0, + 424.0, + 1405.0, + 463.0, + 409.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 456.0, + 1406.0, + 456.0, + 1406.0, + 493.0, + 320.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 486.0, + 810.0, + 486.0, + 810.0, + 524.0, + 321.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 486.0, + 1404.0, + 486.0, + 1404.0, + 525.0, + 847.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 518.0, + 1161.0, + 518.0, + 1161.0, + 556.0, + 322.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 919.0, + 293.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 914.0, + 1406.0, + 914.0, + 1406.0, + 954.0, + 321.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 944.0, + 575.0, + 944.0, + 575.0, + 978.0, + 321.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 1402.0, + 1223.0, + 1402.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1257.0, + 1404.0, + 1257.0, + 1404.0, + 1291.0, + 323.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1287.0, + 622.0, + 1287.0, + 622.0, + 1319.0, + 319.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1538.0, + 1402.0, + 1538.0, + 1402.0, + 1572.0, + 296.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1567.0, + 1405.0, + 1567.0, + 1405.0, + 1607.0, + 321.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1600.0, + 1071.0, + 1600.0, + 1071.0, + 1634.0, + 325.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1110.0, + 1401.0, + 1110.0, + 1401.0, + 1146.0, + 294.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1142.0, + 1400.0, + 1142.0, + 1400.0, + 1178.0, + 323.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1174.0, + 741.0, + 1174.0, + 741.0, + 1208.0, + 325.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1174.0, + 936.0, + 1174.0, + 936.0, + 1208.0, + 761.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 573.0, + 1402.0, + 573.0, + 1402.0, + 605.0, + 296.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 600.0, + 1254.0, + 600.0, + 1254.0, + 639.0, + 320.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 996.0, + 1403.0, + 996.0, + 1403.0, + 1035.0, + 296.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1030.0, + 1404.0, + 1030.0, + 1404.0, + 1064.0, + 324.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1058.0, + 1331.0, + 1058.0, + 1331.0, + 1095.0, + 322.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1423.0, + 1405.0, + 1423.0, + 1405.0, + 1457.0, + 297.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1454.0, + 1403.0, + 1454.0, + 1403.0, + 1489.0, + 322.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1482.0, + 903.0, + 1482.0, + 903.0, + 1523.0, + 322.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1482.0, + 1101.0, + 1482.0, + 1101.0, + 1523.0, + 922.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 655.0, + 1402.0, + 655.0, + 1402.0, + 689.0, + 295.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 687.0, + 1405.0, + 687.0, + 1405.0, + 721.0, + 322.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 716.0, + 611.0, + 716.0, + 611.0, + 748.0, + 321.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 768.0, + 1405.0, + 768.0, + 1405.0, + 806.0, + 294.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 800.0, + 1403.0, + 800.0, + 1403.0, + 836.0, + 323.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 831.0, + 545.0, + 831.0, + 545.0, + 867.0, + 326.0, + 867.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1399, + 229, + 1399, + 323, + 298, + 323 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 294, + 783, + 1402, + 783, + 1402, + 848, + 294, + 848 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 298, + 1880, + 1404, + 1880, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 297, + 1450, + 1404, + 1450, + 1404, + 1606, + 297, + 1606 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 299, + 424, + 1400, + 424, + 1400, + 518, + 299, + 518 + ], + "score": 0.834 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 293, + 948, + 1399, + 948, + 1399, + 1013, + 293, + 1013 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 292, + 701, + 1402, + 701, + 1402, + 766, + 292, + 766 + ], + "score": 0.814 + }, + { + "category_id": 1, + "poly": [ + 299, + 1286, + 1401, + 1286, + 1401, + 1350, + 299, + 1350 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 296, + 1367, + 1400, + 1367, + 1400, + 1432, + 296, + 1432 + ], + "score": 0.806 + }, + { + "category_id": 1, + "poly": [ + 296, + 1624, + 1398, + 1624, + 1398, + 1688, + 296, + 1688 + ], + "score": 0.804 + }, + { + "category_id": 1, + "poly": [ + 290, + 536, + 1400, + 536, + 1400, + 601, + 290, + 601 + ], + "score": 0.799 + }, + { + "category_id": 1, + "poly": [ + 299, + 1030, + 1400, + 1030, + 1400, + 1125, + 299, + 1125 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 297, + 1706, + 1404, + 1706, + 1404, + 1860, + 297, + 1860 + ], + "score": 0.793 + }, + { + "category_id": 1, + "poly": [ + 291, + 342, + 1399, + 342, + 1399, + 406, + 291, + 406 + ], + "score": 0.786 + }, + { + "category_id": 1, + "poly": [ + 297, + 1143, + 1404, + 1143, + 1404, + 1268, + 297, + 1268 + ], + "score": 0.781 + }, + { + "category_id": 1, + "poly": [ + 296, + 618, + 1400, + 618, + 1400, + 683, + 296, + 683 + ], + "score": 0.77 + }, + { + "category_id": 1, + "poly": [ + 293, + 866, + 1402, + 866, + 1402, + 930, + 293, + 930 + ], + "score": 0.704 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 295.0, + 323.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 291.0, + 612.0, + 291.0, + 612.0, + 326.0, + 323.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 783.0, + 1404.0, + 783.0, + 1404.0, + 819.0, + 297.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 816.0, + 894.0, + 816.0, + 894.0, + 848.0, + 321.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1403.0, + 1880.0, + 1403.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1948.0, + 321.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1938.0, + 1405.0, + 1938.0, + 1405.0, + 1982.0, + 317.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1969.0, + 1405.0, + 1969.0, + 1405.0, + 2009.0, + 320.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1999.0, + 1361.0, + 1999.0, + 1361.0, + 2039.0, + 322.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1452.0, + 1402.0, + 1452.0, + 1402.0, + 1485.0, + 297.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1517.0, + 322.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1507.0, + 1405.0, + 1507.0, + 1405.0, + 1553.0, + 320.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1540.0, + 1402.0, + 1540.0, + 1402.0, + 1580.0, + 321.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1573.0, + 789.0, + 1573.0, + 789.0, + 1608.0, + 322.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 421.0, + 1405.0, + 421.0, + 1405.0, + 461.0, + 293.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 453.0, + 1405.0, + 453.0, + 1405.0, + 493.0, + 322.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 487.0, + 682.0, + 487.0, + 682.0, + 521.0, + 324.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 948.0, + 1404.0, + 948.0, + 1404.0, + 984.0, + 296.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 979.0, + 1033.0, + 979.0, + 1033.0, + 1013.0, + 321.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 701.0, + 1405.0, + 701.0, + 1405.0, + 737.0, + 295.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 734.0, + 980.0, + 734.0, + 980.0, + 766.0, + 321.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1282.0, + 1405.0, + 1282.0, + 1405.0, + 1326.0, + 293.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1317.0, + 1095.0, + 1317.0, + 1095.0, + 1353.0, + 324.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1363.0, + 1405.0, + 1363.0, + 1405.0, + 1409.0, + 294.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1398.0, + 1131.0, + 1398.0, + 1131.0, + 1434.0, + 322.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1624.0, + 1403.0, + 1624.0, + 1403.0, + 1660.0, + 296.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1655.0, + 1377.0, + 1655.0, + 1377.0, + 1691.0, + 326.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 538.0, + 1402.0, + 538.0, + 1402.0, + 574.0, + 294.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 569.0, + 1231.0, + 569.0, + 1231.0, + 602.0, + 321.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1030.0, + 1402.0, + 1030.0, + 1402.0, + 1067.0, + 294.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1062.0, + 1405.0, + 1062.0, + 1405.0, + 1099.0, + 322.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1092.0, + 1092.0, + 1092.0, + 1092.0, + 1126.0, + 320.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1704.0, + 1405.0, + 1704.0, + 1405.0, + 1743.0, + 295.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1735.0, + 1406.0, + 1735.0, + 1406.0, + 1774.0, + 322.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1767.0, + 1405.0, + 1767.0, + 1405.0, + 1804.0, + 322.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1794.0, + 1405.0, + 1794.0, + 1405.0, + 1836.0, + 319.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1828.0, + 718.0, + 1828.0, + 718.0, + 1865.0, + 321.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 342.0, + 1402.0, + 342.0, + 1402.0, + 378.0, + 296.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 373.0, + 1216.0, + 373.0, + 1216.0, + 407.0, + 323.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1140.0, + 1406.0, + 1140.0, + 1406.0, + 1183.0, + 294.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1175.0, + 1406.0, + 1175.0, + 1406.0, + 1211.0, + 322.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1207.0, + 1404.0, + 1207.0, + 1404.0, + 1238.0, + 322.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1236.0, + 1272.0, + 1236.0, + 1272.0, + 1269.0, + 325.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 620.0, + 1404.0, + 620.0, + 1404.0, + 656.0, + 295.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 649.0, + 1272.0, + 649.0, + 1272.0, + 684.0, + 322.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 866.0, + 1404.0, + 866.0, + 1404.0, + 902.0, + 295.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 897.0, + 830.0, + 897.0, + 830.0, + 931.0, + 324.0, + 931.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.837 + }, + { + "category_id": 1, + "poly": [ + 303, + 229, + 1400, + 229, + 1400, + 324, + 303, + 324 + ], + "score": 0.818 + }, + { + "category_id": 1, + "poly": [ + 301, + 1653, + 1402, + 1653, + 1402, + 1716, + 301, + 1716 + ], + "score": 0.806 + }, + { + "category_id": 1, + "poly": [ + 296, + 879, + 1401, + 879, + 1401, + 944, + 296, + 944 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 297, + 765, + 1405, + 765, + 1405, + 861, + 297, + 861 + ], + "score": 0.762 + }, + { + "category_id": 1, + "poly": [ + 294, + 343, + 1402, + 343, + 1402, + 407, + 294, + 407 + ], + "score": 0.754 + }, + { + "category_id": 1, + "poly": [ + 297, + 961, + 1404, + 961, + 1404, + 1118, + 297, + 1118 + ], + "score": 0.73 + }, + { + "category_id": 1, + "poly": [ + 298, + 651, + 1405, + 651, + 1405, + 746, + 298, + 746 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 300, + 1250, + 1401, + 1250, + 1401, + 1345, + 300, + 1345 + ], + "score": 0.703 + }, + { + "category_id": 1, + "poly": [ + 296, + 1477, + 1404, + 1477, + 1404, + 1632, + 296, + 1632 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 295, + 569, + 1398, + 569, + 1398, + 634, + 295, + 634 + ], + "score": 0.693 + }, + { + "category_id": 1, + "poly": [ + 297, + 426, + 1404, + 426, + 1404, + 550, + 297, + 550 + ], + "score": 0.691 + }, + { + "category_id": 1, + "poly": [ + 297, + 1364, + 1403, + 1364, + 1403, + 1459, + 297, + 1459 + ], + "score": 0.669 + }, + { + "category_id": 1, + "poly": [ + 299, + 1137, + 1401, + 1137, + 1401, + 1232, + 299, + 1232 + ], + "score": 0.669 + }, + { + "category_id": 13, + "poly": [ + 742, + 1204, + 760, + 1204, + 760, + 1226, + 742, + 1226 + ], + "score": 0.66, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 976, + 833, + 994, + 833, + 994, + 855, + 976, + 855 + ], + "score": 0.44, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1101, + 295, + 1119, + 295, + 1119, + 318, + 1101, + 318 + ], + "score": 0.4, + "latex": "{ . } =" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 224.0, + 1405.0, + 224.0, + 1405.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 297.0, + 322.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 291.0, + 1100.0, + 291.0, + 1100.0, + 329.0, + 323.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 291.0, + 1299.0, + 291.0, + 1299.0, + 329.0, + 1120.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1653.0, + 1406.0, + 1653.0, + 1406.0, + 1688.0, + 295.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1682.0, + 1027.0, + 1682.0, + 1027.0, + 1718.0, + 321.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 877.0, + 1406.0, + 877.0, + 1406.0, + 915.0, + 296.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 910.0, + 936.0, + 910.0, + 936.0, + 945.0, + 321.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 766.0, + 1403.0, + 766.0, + 1403.0, + 801.0, + 295.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 796.0, + 1405.0, + 796.0, + 1405.0, + 834.0, + 321.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 825.0, + 975.0, + 825.0, + 975.0, + 866.0, + 321.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 825.0, + 1156.0, + 825.0, + 1156.0, + 866.0, + 995.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 343.0, + 1404.0, + 343.0, + 1404.0, + 379.0, + 297.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 374.0, + 1195.0, + 374.0, + 1195.0, + 410.0, + 321.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 965.0, + 1402.0, + 965.0, + 1402.0, + 998.0, + 296.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 993.0, + 1407.0, + 993.0, + 1407.0, + 1034.0, + 319.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1023.0, + 1406.0, + 1023.0, + 1406.0, + 1060.0, + 320.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1054.0, + 1406.0, + 1054.0, + 1406.0, + 1089.0, + 321.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1085.0, + 988.0, + 1085.0, + 988.0, + 1121.0, + 320.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 690.0, + 293.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 683.0, + 1403.0, + 683.0, + 1403.0, + 718.0, + 321.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 716.0, + 798.0, + 716.0, + 798.0, + 747.0, + 320.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1248.0, + 1405.0, + 1248.0, + 1405.0, + 1288.0, + 293.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1282.0, + 1406.0, + 1282.0, + 1406.0, + 1319.0, + 323.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1312.0, + 936.0, + 1312.0, + 936.0, + 1346.0, + 324.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1405.0, + 1479.0, + 1405.0, + 1512.0, + 295.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1507.0, + 1406.0, + 1507.0, + 1406.0, + 1545.0, + 323.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1539.0, + 1402.0, + 1539.0, + 1402.0, + 1576.0, + 322.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1566.0, + 1402.0, + 1566.0, + 1402.0, + 1606.0, + 320.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1600.0, + 407.0, + 1600.0, + 407.0, + 1634.0, + 320.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 568.0, + 1404.0, + 568.0, + 1404.0, + 605.0, + 294.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 600.0, + 1402.0, + 600.0, + 1402.0, + 635.0, + 323.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 424.0, + 1405.0, + 424.0, + 1405.0, + 464.0, + 294.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 456.0, + 1405.0, + 456.0, + 1405.0, + 492.0, + 322.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 487.0, + 1404.0, + 487.0, + 1404.0, + 523.0, + 322.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 521.0, + 669.0, + 521.0, + 669.0, + 550.0, + 322.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1363.0, + 1403.0, + 1363.0, + 1403.0, + 1401.0, + 295.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1394.0, + 1403.0, + 1394.0, + 1403.0, + 1432.0, + 324.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1426.0, + 881.0, + 1426.0, + 881.0, + 1460.0, + 324.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1138.0, + 1403.0, + 1138.0, + 1403.0, + 1172.0, + 297.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1165.0, + 1403.0, + 1165.0, + 1403.0, + 1205.0, + 319.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1200.0, + 741.0, + 1200.0, + 741.0, + 1234.0, + 325.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1200.0, + 938.0, + 1200.0, + 938.0, + 1234.0, + 761.0, + 1234.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 559, + 224, + 1139, + 224, + 1139, + 413, + 559, + 413 + ], + "score": 0.975, + "html": "
PythonJavascriptGoRuby
Python1.000.430.420.79
Javascript0.431.000.840.39
Go0.430.841.000.38
Ruby0.790.390.381.00
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1150, + 1405, + 1150, + 1405, + 1243, + 299, + 1243 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 624, + 1404, + 624, + 1404, + 746, + 298, + 746 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 762, + 1405, + 762, + 1405, + 979, + 297, + 979 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1522, + 1405, + 1522, + 1405, + 1649, + 297, + 1649 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1334, + 1404, + 1334, + 1404, + 1429, + 298, + 1429 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 366, + 1911, + 1403, + 1911, + 1403, + 2034, + 366, + 2034 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 1663, + 1400, + 1663, + 1400, + 1726, + 296, + 1726 + ], + "score": 0.951 + }, + { + "category_id": 8, + "poly": [ + 438, + 1258, + 1257, + 1258, + 1257, + 1321, + 438, + 1321 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 301, + 1094, + 757, + 1094, + 757, + 1125, + 301, + 1125 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 301, + 1024, + 509, + 1024, + 509, + 1058, + 301, + 1058 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 301, + 1763, + 640, + 1763, + 640, + 1796, + 301, + 1796 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 302, + 556, + 626, + 556, + 626, + 589, + 302, + 589 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 302, + 1467, + 897, + 1467, + 897, + 1498, + 302, + 1498 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 361, + 1867, + 1150, + 1867, + 1150, + 1902, + 361, + 1902 + ], + "score": 0.877 + }, + { + "category_id": 0, + "poly": [ + 300, + 1821, + 890, + 1821, + 890, + 1853, + 300, + 1853 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.834 + }, + { + "category_id": 6, + "poly": [ + 298, + 434, + 1399, + 434, + 1399, + 497, + 298, + 497 + ], + "score": 0.496 + }, + { + "category_id": 8, + "poly": [ + 443, + 1258, + 755, + 1258, + 755, + 1320, + 443, + 1320 + ], + "score": 0.472 + }, + { + "category_id": 8, + "poly": [ + 897, + 1259, + 1255, + 1259, + 1255, + 1320, + 897, + 1320 + ], + "score": 0.45 + }, + { + "category_id": 1, + "poly": [ + 298, + 434, + 1399, + 434, + 1399, + 497, + 298, + 497 + ], + "score": 0.419 + }, + { + "category_id": 14, + "poly": [ + 441, + 1257, + 1256, + 1257, + 1256, + 1319, + 441, + 1319 + ], + "score": 0.93, + "latex": "\\phi ( r _ { i j } ) _ { 2 k } = \\sin ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) \\qquad \\phi ( r _ { i j } ) _ { 2 k + 1 } = \\cos ( \\frac { r _ { i j } } { M ^ { 2 k / d } } ) ," + }, + { + "category_id": 13, + "poly": [ + 375, + 1334, + 549, + 1334, + 549, + 1368, + 375, + 1368 + ], + "score": 0.92, + "latex": "1 \\leq k < d / 2" + }, + { + "category_id": 13, + "poly": [ + 1161, + 1149, + 1298, + 1149, + 1298, + 1183, + 1161, + 1183 + ], + "score": 0.92, + "latex": "\\phi : \\mathbb { R } \\to \\mathbb { R } ^ { d }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1367, + 454, + 1367, + 454, + 1398, + 298, + 1398 + ], + "score": 0.86, + "latex": "M = 1 0 , 0 0 0" + }, + { + "category_id": 13, + "poly": [ + 1044, + 1336, + 1078, + 1336, + 1078, + 1363, + 1044, + 1363 + ], + "score": 0.84, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1153, + 1402, + 1153, + 1402, + 1179, + 1383, + 1179 + ], + "score": 0.78, + "latex": "d" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1094.0, + 762.0, + 1094.0, + 762.0, + 1129.0, + 296.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1019.0, + 513.0, + 1019.0, + 513.0, + 1065.0, + 295.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1764.0, + 640.0, + 1764.0, + 640.0, + 1797.0, + 297.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 557.0, + 628.0, + 557.0, + 628.0, + 593.0, + 299.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1467.0, + 904.0, + 1467.0, + 904.0, + 1502.0, + 296.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1821.0, + 890.0, + 1821.0, + 890.0, + 1856.0, + 296.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 432.0, + 1403.0, + 432.0, + 1403.0, + 468.0, + 296.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 468.0, + 411.0, + 468.0, + 411.0, + 496.0, + 294.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1148.0, + 1160.0, + 1148.0, + 1160.0, + 1186.0, + 293.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 1148.0, + 1382.0, + 1148.0, + 1382.0, + 1186.0, + 1299.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1148.0, + 1406.0, + 1148.0, + 1406.0, + 1186.0, + 1403.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1182.0, + 1405.0, + 1182.0, + 1405.0, + 1216.0, + 293.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1212.0, + 576.0, + 1212.0, + 576.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 625.0, + 1405.0, + 625.0, + 1405.0, + 661.0, + 294.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 656.0, + 1405.0, + 656.0, + 1405.0, + 688.0, + 294.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 685.0, + 1405.0, + 685.0, + 1405.0, + 721.0, + 292.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 716.0, + 522.0, + 716.0, + 522.0, + 749.0, + 295.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 760.0, + 1405.0, + 760.0, + 1405.0, + 802.0, + 292.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 791.0, + 1403.0, + 791.0, + 1403.0, + 830.0, + 295.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 825.0, + 1403.0, + 825.0, + 1403.0, + 859.0, + 295.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 851.0, + 1405.0, + 851.0, + 1405.0, + 893.0, + 294.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 884.0, + 1403.0, + 884.0, + 1403.0, + 921.0, + 294.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 916.0, + 1403.0, + 916.0, + 1403.0, + 951.0, + 295.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 945.0, + 1129.0, + 945.0, + 1129.0, + 987.0, + 294.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1524.0, + 1403.0, + 1524.0, + 1403.0, + 1557.0, + 294.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1554.0, + 1405.0, + 1554.0, + 1405.0, + 1589.0, + 294.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1583.0, + 1406.0, + 1583.0, + 1406.0, + 1621.0, + 291.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1613.0, + 425.0, + 1613.0, + 425.0, + 1657.0, + 291.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1333.0, + 374.0, + 1333.0, + 374.0, + 1372.0, + 293.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1333.0, + 1043.0, + 1333.0, + 1043.0, + 1372.0, + 550.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1333.0, + 1405.0, + 1333.0, + 1405.0, + 1372.0, + 1079.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1364.0, + 1405.0, + 1364.0, + 1405.0, + 1401.0, + 455.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1397.0, + 573.0, + 1397.0, + 573.0, + 1431.0, + 294.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1909.0, + 915.0, + 1909.0, + 915.0, + 1946.0, + 357.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1945.0, + 946.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1941.0, + 451.0, + 1941.0, + 451.0, + 1974.0, + 393.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1946.0, + 605.0, + 1946.0, + 605.0, + 1973.0, + 484.0, + 1973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 1942.0, + 822.0, + 1942.0, + 822.0, + 1977.0, + 630.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1945.0, + 1404.0, + 1945.0, + 1404.0, + 1974.0, + 940.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2007.0, + 392.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 2003.0, + 703.0, + 2003.0, + 703.0, + 2035.0, + 394.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1661.0, + 1404.0, + 1661.0, + 1404.0, + 1697.0, + 295.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1695.0, + 901.0, + 1695.0, + 901.0, + 1726.0, + 296.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1864.0, + 1151.0, + 1864.0, + 1151.0, + 1907.0, + 362.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 432.0, + 1403.0, + 432.0, + 1403.0, + 468.0, + 296.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 468.0, + 411.0, + 468.0, + 411.0, + 496.0, + 294.0, + 496.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.85 + }, + { + "category_id": 5, + "poly": [ + 296, + 34, + 1473, + 34, + 1473, + 1905, + 296, + 1905 + ], + "score": 0.422, + "html": "
Published as a conference paper at ICLR 2021
function isEqualArray(a,b){ if (a return true;
if ((a === undefined) undefined)){ return false;func areSameFloat32Array(a,b []float32) bool if len(a) != len(b){ return false
var =a.length; if (i !== b.length){ return false;for i := O;i < len(a);i++{ if a[i]!=b[i]{ return false J
while (i--){ if(a[i] !== b[i]){ return false; }return true
return true;
function main() var rawData $(' .HeaderTexture[data-login- user-email]') .data() ;func TaskSayHello(t *tasking.T){
if (rawData) 1 me={username := t.Flags.String("name") if username user, := user.Current ()
name:rawData.loginUserName, mail: rawData.loginUserEmail };username = user.Name
getCartId(function (cart){ me.cart = cart;} ift.Flags.Bool("verbose")
injectMenu(); refreshUsers(actions.updateUsers); listenOnChanges(onChange);t.Logf("Hello %s, the time now is %s\\n", username,time.Now())
listenonorderConfirm(onConfirm); );}else{ t.Logf("Hello %s\\n",username)
} else{ caliback('no user');
func Backup(filename string)error{
info,err := os.Stat(filename) if err != nil
if os.IsNotExist(err){
return nil
return err
def _load_rule_file(self, filename):if info.Size() == 0{ return nil
if not (os.path.exists(filename)):}
sys.stderr.write(files,err := filepath.Glob(
"rflint: %s: No such file or "directory\\n" % filenamefilename + _BACKUP_SUFFIX
return) if err nil 1
try:return err
basename = os.path.basename(filename)}
(name,ext)= os.path.splitext(basename) imp.load_source(name,filename) numBackup := byte(1)
except Exception as e: sys.stderr.write(if len(files) !=0{
"rflint: %s: exception whilelastFile := files[len(files)-1]
"loading:%s\\n" % (filename, str(e))numBackup = lastFile[len(lastFile)-2] + 1
if numBackup >
numBackup = '1'
else{
'1'
numBackup =
return Copy (filename,
fmt.Sprintf("%s+%s~",
string(numBackup)))
" + }, + { + "category_id": 7, + "poly": [ + 299, + 1903, + 1403, + 1903, + 1403, + 1968, + 299, + 1968 + ], + "score": 0.397 + }, + { + "category_id": 13, + "poly": [ + 1037, + 851, + 1059, + 851, + 1059, + 867, + 1037, + 867 + ], + "score": 0.73, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1037, + 830, + 1060, + 830, + 1060, + 845, + 1037, + 845 + ], + "score": 0.71, + "latex": ": =" + }, + { + "category_id": 13, + "poly": [ + 1048, + 1625, + 1071, + 1625, + 1071, + 1641, + 1048, + 1641 + ], + "score": 0.69, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 543, + 1509, + 564, + 1509, + 564, + 1525, + 543, + 1525 + ], + "score": 0.68, + "latex": "^ { - }" + }, + { + "category_id": 13, + "poly": [ + 530, + 903, + 553, + 903, + 553, + 920, + 530, + 920 + ], + "score": 0.67, + "latex": "^ { - }" + }, + { + "category_id": 13, + "poly": [ + 1050, + 807, + 1082, + 807, + 1082, + 824, + 1050, + 824 + ], + "score": 0.65, + "latex": "= -" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1604, + 1071, + 1604, + 1071, + 1620, + 1051, + 1620 + ], + "score": 0.57, + "latex": "^ { - }" + }, + { + "category_id": 13, + "poly": [ + 1050, + 1394, + 1071, + 1394, + 1071, + 1410, + 1050, + 1410 + ], + "score": 0.56, + "latex": "{ } = { }" + }, + { + "category_id": 13, + "poly": [ + 962, + 461, + 1032, + 461, + 1032, + 481, + 962, + 481 + ], + "score": 0.54, + "latex": " { \\mathrm { ~ ~ \\dot { ~ } { ~ 1 ~ } ~ } } : = 0" + }, + { + "category_id": 13, + "poly": [ + 1072, + 1666, + 1092, + 1666, + 1092, + 1683, + 1072, + 1683 + ], + "score": 0.52, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1075, + 1581, + 1122, + 1581, + 1122, + 1601, + 1075, + 1601 + ], + "score": 0.47, + "latex": "! = 0" + }, + { + "category_id": 13, + "poly": [ + 1033, + 483, + 1058, + 483, + 1058, + 501, + 1033, + 501 + ], + "score": 0.47, + "latex": "! =" + }, + { + "category_id": 13, + "poly": [ + 1083, + 1645, + 1100, + 1645, + 1100, + 1662, + 1083, + 1662 + ], + "score": 0.46, + "latex": ">" + }, + { + "category_id": 13, + "poly": [ + 1026, + 787, + 1047, + 787, + 1047, + 804, + 1026, + 804 + ], + "score": 0.44, + "latex": ": =" + }, + { + "category_id": 13, + "poly": [ + 1032, + 398, + 1059, + 398, + 1059, + 417, + 1032, + 417 + ], + "score": 0.42, + "latex": "! =" + }, + { + "category_id": 13, + "poly": [ + 447, + 462, + 487, + 462, + 487, + 480, + 447, + 480 + ], + "score": 0.41, + "latex": "\\begin{array} { r l } { \\dot { \\perp } } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 575, + 1530, + 597, + 1530, + 597, + 1546, + 575, + 1546 + ], + "score": 0.4, + "latex": "^ { - }" + }, + { + "category_id": 13, + "poly": [ + 1054, + 461, + 1092, + 461, + 1092, + 481, + 1054, + 481 + ], + "score": 0.4, + "latex": "\\dot { \\perp } <" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1540, + 1059, + 1540, + 1059, + 1559, + 1032, + 1559 + ], + "score": 0.35, + "latex": ": =" + }, + { + "category_id": 13, + "poly": [ + 1184, + 460, + 1222, + 460, + 1222, + 481, + 1184, + 481 + ], + "score": 0.34, + "latex": "\\dot { \\perp } + +" + }, + { + "category_id": 13, + "poly": [ + 454, + 378, + 521, + 378, + 521, + 396, + 454, + 396 + ], + "score": 0.34, + "latex": "( \\mathsf { a } \\quad = =" + }, + { + "category_id": 13, + "poly": [ + 1039, + 1416, + 1057, + 1416, + 1057, + 1432, + 1039, + 1432 + ], + "score": 0.32, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 488, + 734, + 533, + 734, + 533, + 753, + 488, + 753 + ], + "score": 0.29, + "latex": "= ~ \\textsf { S } \\iota" + }, + { + "category_id": 13, + "poly": [ + 1048, + 1728, + 1101, + 1728, + 1101, + 1747, + 1048, + 1747 + ], + "score": 0.28, + "latex": "\\qquad = \\quad 1 1" + }, + { + "category_id": 13, + "poly": [ + 1040, + 1163, + 1059, + 1163, + 1059, + 1178, + 1040, + 1178 + ], + "score": 0.27, + "latex": "^ { - }" + }, + { + "category_id": 13, + "poly": [ + 456, + 399, + 520, + 399, + 520, + 418, + 456, + 418 + ], + "score": 0.26, + "latex": ": { \\mathrm { b } } \\ = = =" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1899.0, + 1405.0, + 1899.0, + 1405.0, + 1938.0, + 294.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1930.0, + 623.0, + 1930.0, + 623.0, + 1970.0, + 294.0, + 1970.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 513, + 222, + 1169, + 222, + 1169, + 734, + 513, + 734 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 364, + 1421, + 1402, + 1421, + 1402, + 1513, + 364, + 1513 + ], + "score": 0.952 + }, + { + "category_id": 4, + "poly": [ + 365, + 755, + 1335, + 755, + 1335, + 789, + 365, + 789 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 395, + 1674, + 1403, + 1674, + 1403, + 2008, + 395, + 2008 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 361, + 1582, + 1402, + 1582, + 1402, + 1671, + 361, + 1671 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 361, + 906, + 1403, + 906, + 1403, + 968, + 361, + 968 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 362, + 1531, + 1033, + 1531, + 1033, + 1563, + 362, + 1563 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 355, + 1288, + 1367, + 1288, + 1367, + 1321, + 355, + 1321 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 360, + 1207, + 1398, + 1207, + 1398, + 1271, + 360, + 1271 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 364, + 856, + 690, + 856, + 690, + 888, + 364, + 888 + ], + "score": 0.788 + }, + { + "category_id": 1, + "poly": [ + 362, + 1097, + 1403, + 1097, + 1403, + 1190, + 362, + 1190 + ], + "score": 0.785 + }, + { + "category_id": 1, + "poly": [ + 363, + 986, + 1405, + 986, + 1405, + 1078, + 363, + 1078 + ], + "score": 0.781 + }, + { + "category_id": 0, + "poly": [ + 299, + 1362, + 1001, + 1362, + 1001, + 1395, + 299, + 1395 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 299, + 1362, + 1001, + 1362, + 1001, + 1395, + 299, + 1395 + ], + "score": 0.258 + }, + { + "category_id": 13, + "poly": [ + 898, + 1238, + 944, + 1238, + 944, + 1272, + 898, + 1272 + ], + "score": 0.85, + "latex": "^ { \\bullet } \\backslash \\boldsymbol { \\mathrm { n } } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1169, + 1100, + 1206, + 1100, + 1206, + 1127, + 1169, + 1127 + ], + "score": 0.6, + "latex": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 226.0, + 694.0, + 226.0, + 694.0, + 259.0, + 535.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 263.0, + 798.0, + 263.0, + 798.0, + 292.0, + 529.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 253.0, + 973.0, + 253.0, + 973.0, + 285.0, + 863.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 283.0, + 561.0, + 283.0, + 561.0, + 314.0, + 527.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 288.0, + 653.0, + 288.0, + 653.0, + 315.0, + 599.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 302.0, + 1165.0, + 302.0, + 1165.0, + 327.0, + 1002.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 363.0, + 809.0, + 363.0, + 809.0, + 393.0, + 741.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 325.0, + 984.0, + 325.0, + 984.0, + 386.0, + 839.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 323.0, + 1174.0, + 323.0, + 1174.0, + 393.0, + 999.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 383.0, + 832.0, + 383.0, + 832.0, + 417.0, + 741.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 383.0, + 1110.0, + 383.0, + 1110.0, + 416.0, + 999.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 425.0, + 868.0, + 425.0, + 868.0, + 459.0, + 776.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 424.0, + 1069.0, + 424.0, + 1069.0, + 460.0, + 974.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 472.0, + 1134.0, + 472.0, + 1134.0, + 498.0, + 1042.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 511.0, + 770.0, + 511.0, + 770.0, + 544.0, + 676.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 512.0, + 979.0, + 512.0, + 979.0, + 544.0, + 868.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 572.0, + 932.0, + 572.0, + 932.0, + 587.0, + 915.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 594.0, + 957.0, + 594.0, + 957.0, + 637.0, + 889.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 654.0, + 934.0, + 654.0, + 934.0, + 674.0, + 912.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 682.0, + 961.0, + 682.0, + 961.0, + 723.0, + 887.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 753.0, + 1338.0, + 753.0, + 1338.0, + 793.0, + 361.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1358.0, + 1005.0, + 1358.0, + 1005.0, + 1402.0, + 293.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1456.0, + 361.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1448.0, + 1405.0, + 1448.0, + 1405.0, + 1487.0, + 393.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1480.0, + 1023.0, + 1480.0, + 1023.0, + 1516.0, + 395.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1670.0, + 1406.0, + 1670.0, + 1406.0, + 1707.0, + 394.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1736.0, + 395.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1735.0, + 1407.0, + 1735.0, + 1407.0, + 1768.0, + 392.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1763.0, + 1404.0, + 1763.0, + 1404.0, + 1797.0, + 393.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1792.0, + 1407.0, + 1792.0, + 1407.0, + 1830.0, + 392.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1823.0, + 1407.0, + 1823.0, + 1407.0, + 1860.0, + 392.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1856.0, + 1406.0, + 1856.0, + 1406.0, + 1891.0, + 395.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1886.0, + 1405.0, + 1886.0, + 1405.0, + 1918.0, + 395.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1913.0, + 1407.0, + 1913.0, + 1407.0, + 1951.0, + 392.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1946.0, + 1405.0, + 1946.0, + 1405.0, + 1980.0, + 394.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1979.0, + 646.0, + 1979.0, + 646.0, + 2010.0, + 392.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1582.0, + 1403.0, + 1582.0, + 1403.0, + 1615.0, + 360.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1610.0, + 1403.0, + 1610.0, + 1403.0, + 1647.0, + 393.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1638.0, + 544.0, + 1638.0, + 544.0, + 1675.0, + 393.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 905.0, + 1404.0, + 905.0, + 1404.0, + 941.0, + 359.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 937.0, + 687.0, + 937.0, + 687.0, + 969.0, + 401.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1528.0, + 1036.0, + 1528.0, + 1036.0, + 1568.0, + 359.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1287.0, + 1369.0, + 1287.0, + 1369.0, + 1324.0, + 357.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1204.0, + 1405.0, + 1204.0, + 1405.0, + 1244.0, + 360.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1238.0, + 897.0, + 1238.0, + 897.0, + 1274.0, + 394.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1238.0, + 957.0, + 1238.0, + 957.0, + 1274.0, + 945.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 853.0, + 695.0, + 853.0, + 695.0, + 893.0, + 359.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1095.0, + 1168.0, + 1095.0, + 1168.0, + 1133.0, + 360.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1095.0, + 1404.0, + 1095.0, + 1404.0, + 1133.0, + 1207.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1128.0, + 1404.0, + 1128.0, + 1404.0, + 1162.0, + 399.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1154.0, + 1347.0, + 1154.0, + 1347.0, + 1195.0, + 393.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 983.0, + 1404.0, + 983.0, + 1404.0, + 1024.0, + 358.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1015.0, + 1405.0, + 1015.0, + 1405.0, + 1052.0, + 394.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1041.0, + 504.0, + 1041.0, + 504.0, + 1088.0, + 391.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1358.0, + 1005.0, + 1358.0, + 1005.0, + 1402.0, + 293.0, + 1402.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1404, + 1850, + 1404, + 2033, + 298, + 2033 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1569, + 1404, + 1569, + 1404, + 1752, + 298, + 1752 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 428, + 394, + 1404, + 394, + 1404, + 781, + 428, + 781 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 361, + 277, + 1403, + 277, + 1403, + 340, + 361, + 340 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 302, + 1791, + 766, + 1791, + 766, + 1823, + 302, + 1823 + ], + "score": 0.881 + }, + { + "category_id": 0, + "poly": [ + 300, + 1512, + 797, + 1512, + 797, + 1542, + 300, + 1542 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.841 + }, + { + "category_id": 1, + "poly": [ + 396, + 793, + 1402, + 793, + 1402, + 1037, + 396, + 1037 + ], + "score": 0.819 + }, + { + "category_id": 1, + "poly": [ + 363, + 351, + 1173, + 351, + 1173, + 383, + 363, + 383 + ], + "score": 0.796 + }, + { + "category_id": 1, + "poly": [ + 363, + 1048, + 1403, + 1048, + 1403, + 1262, + 363, + 1262 + ], + "score": 0.594 + }, + { + "category_id": 1, + "poly": [ + 298, + 230, + 1193, + 230, + 1193, + 262, + 298, + 262 + ], + "score": 0.506 + }, + { + "category_id": 1, + "poly": [ + 368, + 1379, + 1401, + 1379, + 1401, + 1470, + 368, + 1470 + ], + "score": 0.429 + }, + { + "category_id": 1, + "poly": [ + 364, + 1275, + 1400, + 1275, + 1400, + 1367, + 364, + 1367 + ], + "score": 0.4 + }, + { + "category_id": 2, + "poly": [ + 298, + 230, + 1193, + 230, + 1193, + 262, + 298, + 262 + ], + "score": 0.164 + }, + { + "category_id": 13, + "poly": [ + 1006, + 1202, + 1168, + 1202, + 1168, + 1232, + 1006, + 1232 + ], + "score": 0.9, + "latex": "- 4 , - 3 , \\dots , 4" + }, + { + "category_id": 13, + "poly": [ + 800, + 455, + 844, + 455, + 844, + 484, + 800, + 484 + ], + "score": 0.89, + "latex": "e ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 483, + 456, + 590, + 456, + 590, + 485, + 483, + 485 + ], + "score": 0.89, + "latex": "\\alpha = 0 . 1 5" + }, + { + "category_id": 13, + "poly": [ + 1062, + 456, + 1249, + 456, + 1249, + 487, + 1062, + 487 + ], + "score": 0.89, + "latex": "- \\log P P R > 5" + }, + { + "category_id": 13, + "poly": [ + 1180, + 1441, + 1289, + 1441, + 1289, + 1469, + 1180, + 1469 + ], + "score": 0.76, + "latex": "= ~ 1 0 0 0" + }, + { + "category_id": 13, + "poly": [ + 395, + 1441, + 482, + 1441, + 482, + 1470, + 395, + 1470 + ], + "score": 0.76, + "latex": "= ~ 5 1 2" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1086, + 1210, + 1086, + 1210, + 1112, + 1192, + 1112 + ], + "score": 0.62, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1085, + 1050, + 1085, + 1050, + 1108, + 1032, + 1108 + ], + "score": 0.48, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 644, + 313, + 664, + 313, + 664, + 337, + 644, + 337 + ], + "score": 0.45, + "latex": ">" + }, + { + "category_id": 13, + "poly": [ + 568, + 713, + 639, + 713, + 639, + 744, + 568, + 744 + ], + "score": 0.25, + "latex": "1 , 0 0 0" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1791.0, + 771.0, + 1791.0, + 771.0, + 1827.0, + 297.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1511.0, + 801.0, + 1511.0, + 801.0, + 1545.0, + 296.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 226.0, + 1196.0, + 226.0, + 1196.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1404.0, + 1940.0, + 1404.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 585.0, + 2002.0, + 585.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1567.0, + 1406.0, + 1567.0, + 1406.0, + 1606.0, + 293.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1596.0, + 1405.0, + 1596.0, + 1405.0, + 1638.0, + 292.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1630.0, + 1405.0, + 1630.0, + 1405.0, + 1666.0, + 294.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1658.0, + 1404.0, + 1658.0, + 1404.0, + 1695.0, + 293.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1690.0, + 1407.0, + 1690.0, + 1407.0, + 1726.0, + 292.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1721.0, + 556.0, + 1721.0, + 556.0, + 1757.0, + 294.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 394.0, + 801.0, + 394.0, + 801.0, + 428.0, + 427.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 424.0, + 1402.0, + 424.0, + 1402.0, + 459.0, + 449.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 453.0, + 482.0, + 453.0, + 482.0, + 491.0, + 449.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 453.0, + 799.0, + 453.0, + 799.0, + 491.0, + 591.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 453.0, + 1061.0, + 453.0, + 1061.0, + 491.0, + 845.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 453.0, + 1405.0, + 453.0, + 1405.0, + 491.0, + 1250.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 490.0, + 592.0, + 490.0, + 592.0, + 515.0, + 451.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 523.0, + 898.0, + 523.0, + 898.0, + 554.0, + 431.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 559.0, + 1405.0, + 559.0, + 1405.0, + 592.0, + 430.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 591.0, + 1406.0, + 591.0, + 1406.0, + 622.0, + 449.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 620.0, + 1405.0, + 620.0, + 1405.0, + 655.0, + 448.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 652.0, + 1406.0, + 652.0, + 1406.0, + 685.0, + 450.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 682.0, + 1405.0, + 682.0, + 1405.0, + 714.0, + 447.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 714.0, + 567.0, + 714.0, + 567.0, + 745.0, + 448.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 714.0, + 918.0, + 714.0, + 918.0, + 745.0, + 640.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 747.0, + 1371.0, + 747.0, + 1371.0, + 782.0, + 428.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 274.0, + 1404.0, + 274.0, + 1404.0, + 312.0, + 361.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 309.0, + 643.0, + 309.0, + 643.0, + 344.0, + 394.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 309.0, + 742.0, + 309.0, + 742.0, + 344.0, + 665.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 791.0, + 1405.0, + 791.0, + 1405.0, + 827.0, + 393.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 822.0, + 1406.0, + 822.0, + 1406.0, + 859.0, + 393.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 854.0, + 1405.0, + 854.0, + 1405.0, + 888.0, + 394.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 884.0, + 1406.0, + 884.0, + 1406.0, + 917.0, + 394.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 915.0, + 1405.0, + 915.0, + 1405.0, + 948.0, + 394.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 945.0, + 1405.0, + 945.0, + 1405.0, + 978.0, + 394.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 974.0, + 1406.0, + 974.0, + 1406.0, + 1011.0, + 393.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1005.0, + 1172.0, + 1005.0, + 1172.0, + 1043.0, + 392.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 348.0, + 1173.0, + 348.0, + 1173.0, + 387.0, + 359.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1046.0, + 1406.0, + 1046.0, + 1406.0, + 1083.0, + 358.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1078.0, + 1031.0, + 1078.0, + 1031.0, + 1116.0, + 394.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1078.0, + 1191.0, + 1078.0, + 1191.0, + 1116.0, + 1051.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1078.0, + 1405.0, + 1078.0, + 1405.0, + 1116.0, + 1211.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1110.0, + 1404.0, + 1110.0, + 1404.0, + 1144.0, + 395.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1139.0, + 1404.0, + 1139.0, + 1404.0, + 1175.0, + 394.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1170.0, + 1405.0, + 1170.0, + 1405.0, + 1205.0, + 393.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1200.0, + 1005.0, + 1200.0, + 1005.0, + 1235.0, + 395.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1200.0, + 1405.0, + 1200.0, + 1405.0, + 1235.0, + 1169.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1227.0, + 1132.0, + 1227.0, + 1132.0, + 1270.0, + 394.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 226.0, + 1196.0, + 226.0, + 1196.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1414.0, + 360.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1409.0, + 1404.0, + 1409.0, + 1404.0, + 1444.0, + 395.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1441.0, + 1179.0, + 1441.0, + 1179.0, + 1472.0, + 483.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1441.0, + 1300.0, + 1441.0, + 1300.0, + 1472.0, + 1290.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1272.0, + 1405.0, + 1272.0, + 1405.0, + 1311.0, + 358.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1305.0, + 1405.0, + 1305.0, + 1405.0, + 1343.0, + 394.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1335.0, + 774.0, + 1335.0, + 774.0, + 1371.0, + 395.0, + 1371.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 371, + 1014, + 1404, + 1014, + 1404, + 1475, + 371, + 1475 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 724, + 1404, + 724, + 1404, + 879, + 297, + 879 + ], + "score": 0.972 + }, + { + "category_id": 5, + "poly": [ + 885, + 300, + 1218, + 300, + 1218, + 546, + 885, + 546 + ], + "score": 0.969, + "html": "
HyperparameterValue
ActivationGELU
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
" + }, + { + "category_id": 5, + "poly": [ + 466, + 272, + 823, + 272, + 823, + 552, + 466, + 552 + ], + "score": 0.968, + "html": "
HyperparameterValue
ActivationGELU
Input Nonlinearitytanh
Num. layers3
d1024
dFF2048
Pdropout0.2
Num.heads8
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1568, + 1404, + 1568, + 1404, + 1631, + 297, + 1631 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 299, + 923, + 1399, + 923, + 1399, + 986, + 299, + 986 + ], + "score": 0.945 + }, + { + "category_id": 6, + "poly": [ + 586, + 576, + 1115, + 576, + 1115, + 609, + 586, + 609 + ], + "score": 0.904 + }, + { + "category_id": 6, + "poly": [ + 870, + 240, + 1236, + 240, + 1236, + 295, + 870, + 295 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 303, + 1512, + 787, + 1512, + 787, + 1542, + 303, + 1542 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 301, + 668, + 644, + 668, + 644, + 699, + 301, + 699 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 813, + 76, + 813, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 6, + "poly": [ + 508, + 241, + 780, + 241, + 780, + 268, + 508, + 268 + ], + "score": 0.784 + }, + { + "category_id": 13, + "poly": [ + 485, + 815, + 543, + 815, + 543, + 845, + 485, + 845 + ], + "score": 0.9, + "latex": "3 e ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 1293, + 785, + 1353, + 785, + 1353, + 815, + 1293, + 815 + ], + "score": 0.9, + "latex": "8 e ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 485, + 450, + 539, + 450, + 539, + 481, + 485, + 481 + ], + "score": 0.33, + "latex": "d _ { F F }" + }, + { + "category_id": 13, + "poly": [ + 485, + 419, + 504, + 419, + 504, + 446, + 485, + 446 + ], + "score": 0.3, + "latex": "d" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 567.0, + 1116.0, + 567.0, + 1116.0, + 618.0, + 582.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 237.0, + 1241.0, + 237.0, + 1241.0, + 273.0, + 867.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 265.0, + 939.0, + 265.0, + 939.0, + 301.0, + 866.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1511.0, + 791.0, + 1511.0, + 791.0, + 1545.0, + 297.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 668.0, + 645.0, + 668.0, + 645.0, + 701.0, + 296.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 239.0, + 783.0, + 239.0, + 783.0, + 270.0, + 506.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1012.0, + 843.0, + 1012.0, + 843.0, + 1047.0, + 380.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1042.0, + 1407.0, + 1042.0, + 1407.0, + 1077.0, + 393.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1075.0, + 1404.0, + 1075.0, + 1404.0, + 1108.0, + 394.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1105.0, + 923.0, + 1105.0, + 923.0, + 1138.0, + 394.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1144.0, + 826.0, + 1144.0, + 826.0, + 1180.0, + 380.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1176.0, + 1403.0, + 1176.0, + 1403.0, + 1208.0, + 394.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1205.0, + 1403.0, + 1205.0, + 1403.0, + 1243.0, + 393.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1237.0, + 552.0, + 1237.0, + 552.0, + 1270.0, + 395.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 1277.0, + 801.0, + 1277.0, + 801.0, + 1312.0, + 379.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1309.0, + 1407.0, + 1309.0, + 1407.0, + 1344.0, + 393.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1338.0, + 1403.0, + 1338.0, + 1403.0, + 1375.0, + 391.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1373.0, + 1403.0, + 1373.0, + 1403.0, + 1402.0, + 395.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1402.0, + 1145.0, + 1402.0, + 1145.0, + 1433.0, + 393.0, + 1433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1440.0, + 1367.0, + 1440.0, + 1367.0, + 1478.0, + 369.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 724.0, + 1405.0, + 724.0, + 1405.0, + 759.0, + 296.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 754.0, + 1405.0, + 754.0, + 1405.0, + 790.0, + 295.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 784.0, + 1292.0, + 784.0, + 1292.0, + 820.0, + 294.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 784.0, + 1406.0, + 784.0, + 1406.0, + 820.0, + 1354.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 484.0, + 812.0, + 484.0, + 853.0, + 294.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 853.0, + 544.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 848.0, + 591.0, + 848.0, + 591.0, + 880.0, + 293.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1569.0, + 1402.0, + 1569.0, + 1402.0, + 1601.0, + 296.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1597.0, + 1127.0, + 1597.0, + 1127.0, + 1634.0, + 294.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 920.0, + 1403.0, + 920.0, + 1403.0, + 960.0, + 295.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 957.0, + 794.0, + 957.0, + 794.0, + 989.0, + 295.0, + 989.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 642, + 356, + 1051, + 356, + 1051, + 572, + 642, + 572 + ], + "score": 0.969, + "html": "
ModelPrediction
GREAT get x map
code2seq get parts
Ours w/o structureget
CODE TRANSFORMERget next
Ground Truthnext
" + }, + { + "category_id": 5, + "poly": [ + 615, + 1684, + 1079, + 1684, + 1079, + 1900, + 615, + 1900 + ], + "score": 0.96, + "html": "
ModelPrediction
GREATrun
code2seq get submit time
Ours w/o structurecompare
CODE TRANSFORMERcompare
Ground Truthcompare
" + }, + { + "category_id": 1, + "poly": [ + 295, + 1479, + 1274, + 1479, + 1274, + 1672, + 295, + 1672 + ], + "score": 0.928 + }, + { + "category_id": 5, + "poly": [ + 627, + 1046, + 1070, + 1046, + 1070, + 1262, + 627, + 1262 + ], + "score": 0.908, + "html": "
ModelPrediction
GREATget path
code2seqfind file
Ours w/o structureget file
CODE TRANSFORMERfind cache
Ground Truthfind cache file
" + }, + { + "category_id": 1, + "poly": [ + 295, + 263, + 674, + 263, + 674, + 346, + 295, + 346 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.859 + }, + { + "category_id": 1, + "poly": [ + 296, + 758, + 1259, + 758, + 1259, + 1030, + 296, + 1030 + ], + "score": 0.851 + }, + { + "category_id": 6, + "poly": [ + 295, + 595, + 1402, + 595, + 1402, + 658, + 295, + 658 + ], + "score": 0.742 + }, + { + "category_id": 1, + "poly": [ + 298, + 1921, + 1402, + 1921, + 1402, + 1986, + 298, + 1986 + ], + "score": 0.737 + }, + { + "category_id": 1, + "poly": [ + 297, + 1284, + 1406, + 1284, + 1406, + 1377, + 297, + 1377 + ], + "score": 0.647 + }, + { + "category_id": 6, + "poly": [ + 297, + 1284, + 1406, + 1284, + 1406, + 1377, + 297, + 1377 + ], + "score": 0.158 + }, + { + "category_id": 6, + "poly": [ + 298, + 1921, + 1402, + 1921, + 1402, + 1986, + 298, + 1986 + ], + "score": 0.145 + }, + { + "category_id": 13, + "poly": [ + 585, + 789, + 619, + 789, + 619, + 812, + 585, + 812 + ], + "score": 0.69, + "latex": "! =" + }, + { + "category_id": 13, + "poly": [ + 507, + 1536, + 614, + 1536, + 614, + 1563, + 507, + 1563 + ], + "score": 0.57, + "latex": "\\ j 1 \\ = \\ \\mathtt { p 1 }" + }, + { + "category_id": 13, + "poly": [ + 758, + 1592, + 784, + 1592, + 784, + 1616, + 758, + 1616 + ], + "score": 0.54, + "latex": "<" + }, + { + "category_id": 13, + "poly": [ + 508, + 1564, + 615, + 1564, + 615, + 1591, + 508, + 1591 + ], + "score": 0.51, + "latex": "\\begin{array} { r } { \\dot { \\bf { \\ j } } 2 { \\bf { \\nu } } = { \\bf { \\ p } } 2 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 543, + 293, + 611, + 293, + 611, + 319, + 543, + 319 + ], + "score": 0.41, + "latex": "[ \\dot { 1 } + + ]" + }, + { + "category_id": 13, + "poly": [ + 1009, + 788, + 1064, + 788, + 1064, + 814, + 1009, + 814 + ], + "score": 0.41, + "latex": "> 0" + }, + { + "category_id": 13, + "poly": [ + 802, + 1621, + 844, + 1621, + 844, + 1645, + 802, + 1645 + ], + "score": 0.35, + "latex": "= =" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2086.0, + 870.0, + 2086.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 592.0, + 1404.0, + 592.0, + 1404.0, + 630.0, + 294.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 628.0, + 420.0, + 628.0, + 420.0, + 661.0, + 294.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1283.0, + 1404.0, + 1283.0, + 1404.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1314.0, + 1404.0, + 1314.0, + 1404.0, + 1350.0, + 292.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 461.0, + 1344.0, + 461.0, + 1378.0, + 294.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1915.0, + 1403.0, + 1915.0, + 1403.0, + 1959.0, + 293.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1951.0, + 925.0, + 1951.0, + 925.0, + 1990.0, + 293.0, + 1990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1108.0, + 1477.0, + 1108.0, + 1511.0, + 293.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1508.0, + 1139.0, + 1508.0, + 1139.0, + 1537.0, + 578.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1532.0, + 506.0, + 1532.0, + 506.0, + 1566.0, + 354.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1532.0, + 751.0, + 1532.0, + 751.0, + 1566.0, + 615.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1562.0, + 507.0, + 1562.0, + 507.0, + 1594.0, + 355.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1562.0, + 750.0, + 1562.0, + 750.0, + 1594.0, + 616.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1594.0, + 454.0, + 1594.0, + 454.0, + 1617.0, + 360.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1589.0, + 757.0, + 1589.0, + 757.0, + 1621.0, + 462.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 1589.0, + 1155.0, + 1589.0, + 1155.0, + 1621.0, + 785.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 1617.0, + 801.0, + 1617.0, + 801.0, + 1649.0, + 474.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1617.0, + 1271.0, + 1617.0, + 1271.0, + 1649.0, + 845.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1649.0, + 375.0, + 1649.0, + 375.0, + 1673.0, + 357.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 263.0, + 678.0, + 263.0, + 678.0, + 292.0, + 294.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 292.0, + 542.0, + 292.0, + 542.0, + 319.0, + 356.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 292.0, + 627.0, + 292.0, + 627.0, + 319.0, + 612.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 325.0, + 316.0, + 325.0, + 316.0, + 345.0, + 300.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 756.0, + 1260.0, + 756.0, + 1260.0, + 792.0, + 291.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 785.0, + 584.0, + 785.0, + 584.0, + 818.0, + 355.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 785.0, + 1008.0, + 785.0, + 1008.0, + 818.0, + 620.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 785.0, + 1113.0, + 785.0, + 1113.0, + 818.0, + 1065.0, + 818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 813.0, + 471.0, + 813.0, + 471.0, + 844.0, + 414.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 815.0, + 874.0, + 815.0, + 874.0, + 844.0, + 480.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 843.0, + 515.0, + 843.0, + 515.0, + 868.0, + 476.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 843.0, + 1050.0, + 843.0, + 1050.0, + 872.0, + 520.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 867.0, + 720.0, + 867.0, + 720.0, + 901.0, + 534.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 902.0, + 495.0, + 902.0, + 495.0, + 925.0, + 477.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 980.0, + 536.0, + 980.0, + 536.0, + 1010.0, + 357.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1014.0, + 315.0, + 1014.0, + 315.0, + 1031.0, + 301.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1915.0, + 1403.0, + 1915.0, + 1403.0, + 1959.0, + 293.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1951.0, + 925.0, + 1951.0, + 925.0, + 1990.0, + 293.0, + 1990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1283.0, + 1404.0, + 1283.0, + 1404.0, + 1317.0, + 295.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1314.0, + 1404.0, + 1314.0, + 1404.0, + 1350.0, + 292.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 461.0, + 1344.0, + 461.0, + 1378.0, + 294.0, + 1378.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 599, + 1117, + 1095, + 1117, + 1095, + 1333, + 599, + 1333 + ], + "score": 0.972, + "html": "
ModelPrediction
GREATlist blobs
code2seqlist blobs
Ours w/o structurelist blobs
CODE TRANSFORMERlist blobs by prefix
Ground Truthlist root blobs
" + }, + { + "category_id": 5, + "poly": [ + 638, + 436, + 1053, + 436, + 1053, + 651, + 638, + 651 + ], + "score": 0.966, + "html": "
ModelPrediction
GREAT get value
code2seq get value
Ours w/o structureto
CODE TRANSFORMERfrom value
Ground Truthfrom value
" + }, + { + "category_id": 5, + "poly": [ + 612, + 1686, + 1080, + 1686, + 1080, + 1903, + 612, + 1903 + ], + "score": 0.965, + "html": "
ModelPrediction
GREATappend
code2seqadd
Ours w/o structurelog
CODE TRANSFORMERlog op counts
Ground Truthdump op counts
" + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 290, + 1510, + 1604, + 1510, + 1604, + 1667, + 290, + 1667 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 866, + 2087, + 866, + 2113, + 834, + 2113 + ], + "score": 0.866 + }, + { + "category_id": 1, + "poly": [ + 293, + 674, + 1403, + 674, + 1403, + 737, + 293, + 737 + ], + "score": 0.823 + }, + { + "category_id": 1, + "poly": [ + 318, + 998, + 1611, + 998, + 1611, + 1081, + 318, + 1081 + ], + "score": 0.797 + }, + { + "category_id": 1, + "poly": [ + 296, + 258, + 1022, + 258, + 1022, + 425, + 296, + 425 + ], + "score": 0.629 + }, + { + "category_id": 1, + "poly": [ + 293, + 1355, + 1401, + 1355, + 1401, + 1419, + 293, + 1419 + ], + "score": 0.605 + }, + { + "category_id": 1, + "poly": [ + 299, + 830, + 1194, + 830, + 1194, + 858, + 299, + 858 + ], + "score": 0.523 + }, + { + "category_id": 6, + "poly": [ + 295, + 1925, + 1403, + 1925, + 1403, + 1990, + 295, + 1990 + ], + "score": 0.366 + }, + { + "category_id": 6, + "poly": [ + 293, + 1355, + 1401, + 1355, + 1401, + 1419, + 293, + 1419 + ], + "score": 0.335 + }, + { + "category_id": 1, + "poly": [ + 962, + 858, + 1613, + 858, + 1613, + 968, + 962, + 968 + ], + "score": 0.184 + }, + { + "category_id": 1, + "poly": [ + 298, + 1081, + 314, + 1081, + 314, + 1108, + 298, + 1108 + ], + "score": 0.16 + }, + { + "category_id": 0, + "poly": [ + 297, + 260, + 962, + 260, + 962, + 286, + 297, + 286 + ], + "score": 0.138 + }, + { + "category_id": 1, + "poly": [ + 877, + 971, + 1580, + 971, + 1580, + 996, + 877, + 996 + ], + "score": 0.101 + }, + { + "category_id": 6, + "poly": [ + 293, + 674, + 1403, + 674, + 1403, + 737, + 293, + 737 + ], + "score": 0.097 + }, + { + "category_id": 13, + "poly": [ + 697, + 290, + 740, + 290, + 740, + 313, + 697, + 313 + ], + "score": 0.76, + "latex": "> =" + }, + { + "category_id": 13, + "poly": [ + 606, + 1542, + 635, + 1542, + 635, + 1564, + 606, + 1564 + ], + "score": 0.73, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 891, + 999, + 918, + 999, + 918, + 1021, + 891, + 1021 + ], + "score": 0.68, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 504, + 288, + 556, + 288, + 556, + 314, + 504, + 314 + ], + "score": 0.4, + "latex": "< ~ 0" + }, + { + "category_id": 13, + "poly": [ + 1271, + 1515, + 1287, + 1515, + 1287, + 1535, + 1271, + 1535 + ], + "score": 0.33, + "latex": "<" + }, + { + "category_id": 13, + "poly": [ + 897, + 1515, + 913, + 1515, + 913, + 1535, + 897, + 1535 + ], + "score": 0.3, + "latex": "<" + }, + { + "category_id": 13, + "poly": [ + 960, + 1597, + 989, + 1597, + 989, + 1618, + 960, + 1618 + ], + "score": 0.3, + "latex": "^ { 1 1 } =" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2124.0, + 828.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1920.0, + 1405.0, + 1920.0, + 1405.0, + 1961.0, + 294.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1952.0, + 422.0, + 1952.0, + 422.0, + 1995.0, + 292.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1351.0, + 1403.0, + 1351.0, + 1403.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1383.0, + 1388.0, + 1383.0, + 1388.0, + 1422.0, + 293.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 257.0, + 961.0, + 257.0, + 961.0, + 291.0, + 293.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 711.0, + 293.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 704.0, + 929.0, + 704.0, + 929.0, + 740.0, + 295.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1508.0, + 896.0, + 1508.0, + 896.0, + 1542.0, + 292.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1508.0, + 1270.0, + 1508.0, + 1270.0, + 1542.0, + 914.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1508.0, + 1604.0, + 1508.0, + 1604.0, + 1542.0, + 1288.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1535.0, + 605.0, + 1535.0, + 605.0, + 1571.0, + 353.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1535.0, + 948.0, + 1535.0, + 948.0, + 1571.0, + 636.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1559.0, + 1276.0, + 1559.0, + 1276.0, + 1601.0, + 352.0, + 1601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1585.0, + 959.0, + 1585.0, + 959.0, + 1629.0, + 351.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1585.0, + 1379.0, + 1585.0, + 1379.0, + 1629.0, + 990.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1616.0, + 859.0, + 1616.0, + 859.0, + 1658.0, + 351.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1650.0, + 317.0, + 1650.0, + 317.0, + 1670.0, + 300.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 711.0, + 293.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 704.0, + 929.0, + 704.0, + 929.0, + 740.0, + 295.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 994.0, + 890.0, + 994.0, + 890.0, + 1025.0, + 355.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 994.0, + 1616.0, + 994.0, + 1616.0, + 1025.0, + 919.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1017.0, + 1140.0, + 1017.0, + 1140.0, + 1057.0, + 352.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1048.0, + 1301.0, + 1048.0, + 1301.0, + 1082.0, + 757.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 256.0, + 967.0, + 256.0, + 967.0, + 289.0, + 294.0, + 289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 283.0, + 503.0, + 283.0, + 503.0, + 319.0, + 352.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 283.0, + 696.0, + 283.0, + 696.0, + 319.0, + 557.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 283.0, + 1022.0, + 283.0, + 1022.0, + 319.0, + 741.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 315.0, + 598.0, + 315.0, + 598.0, + 346.0, + 414.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 350.0, + 373.0, + 350.0, + 373.0, + 365.0, + 362.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 369.0, + 703.0, + 369.0, + 703.0, + 401.0, + 354.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 405.0, + 312.0, + 405.0, + 312.0, + 419.0, + 304.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1351.0, + 1403.0, + 1351.0, + 1403.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1383.0, + 1388.0, + 1383.0, + 1388.0, + 1422.0, + 293.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 829.0, + 1196.0, + 829.0, + 1196.0, + 863.0, + 293.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 854.0, + 1376.0, + 854.0, + 1376.0, + 888.0, + 965.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 883.0, + 1614.0, + 883.0, + 1614.0, + 916.0, + 967.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 913.0, + 1375.0, + 913.0, + 1375.0, + 943.0, + 968.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 941.0, + 1374.0, + 941.0, + 1374.0, + 970.0, + 969.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1084.0, + 315.0, + 1084.0, + 315.0, + 1107.0, + 298.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 965.0, + 1574.0, + 965.0, + 1574.0, + 1001.0, + 893.0, + 1001.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 601, + 626, + 1092, + 626, + 1092, + 844, + 601, + 844 + ], + "score": 0.97, + "html": "
ModelPrediction
GREATget canonical path
code2seqexec
Ours w/o structure get output
CODE TRANSFORMERexec command
Ground Truthexec command
" + }, + { + "category_id": 5, + "poly": [ + 621, + 1523, + 1071, + 1523, + 1071, + 1740, + 621, + 1740 + ], + "score": 0.962, + "html": "
ModelPrediction
GREATprint
code2seqprint
Ours w/o structureprint
CODE TRANSFORMERprint sub view
Ground Truthsub view
" + }, + { + "category_id": 1, + "poly": [ + 297, + 422, + 1332, + 422, + 1332, + 592, + 297, + 592 + ], + "score": 0.96 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 305, + 1369, + 1420, + 1369, + 1420, + 1494, + 305, + 1494 + ], + "score": 0.744 + }, + { + "category_id": 6, + "poly": [ + 296, + 1762, + 1404, + 1762, + 1404, + 1826, + 296, + 1826 + ], + "score": 0.625 + }, + { + "category_id": 6, + "poly": [ + 289, + 865, + 1404, + 865, + 1404, + 929, + 289, + 929 + ], + "score": 0.613 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.608 + }, + { + "category_id": 0, + "poly": [ + 295, + 1348, + 1106, + 1348, + 1106, + 1376, + 295, + 1376 + ], + "score": 0.391 + }, + { + "category_id": 1, + "poly": [ + 295, + 1348, + 1106, + 1348, + 1106, + 1376, + 295, + 1376 + ], + "score": 0.288 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.243 + }, + { + "category_id": 1, + "poly": [ + 296, + 1762, + 1404, + 1762, + 1404, + 1826, + 296, + 1826 + ], + "score": 0.228 + }, + { + "category_id": 1, + "poly": [ + 289, + 865, + 1404, + 865, + 1404, + 929, + 289, + 929 + ], + "score": 0.19 + }, + { + "category_id": 13, + "poly": [ + 561, + 537, + 590, + 537, + 590, + 558, + 561, + 558 + ], + "score": 0.8, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 607, + 509, + 635, + 509, + 635, + 532, + 607, + 532 + ], + "score": 0.77, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 562, + 454, + 590, + 454, + 590, + 475, + 562, + 475 + ], + "score": 0.75, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 923, + 454, + 946, + 454, + 946, + 475, + 923, + 475 + ], + "score": 0.34, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1762.0, + 1406.0, + 1762.0, + 1406.0, + 1798.0, + 296.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1791.0, + 1356.0, + 1791.0, + 1356.0, + 1827.0, + 295.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 865.0, + 1403.0, + 865.0, + 1403.0, + 901.0, + 296.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 891.0, + 1192.0, + 891.0, + 1192.0, + 934.0, + 292.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1345.0, + 1113.0, + 1345.0, + 1113.0, + 1381.0, + 292.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 420.0, + 1332.0, + 420.0, + 1332.0, + 453.0, + 295.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 450.0, + 561.0, + 450.0, + 561.0, + 480.0, + 356.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 450.0, + 922.0, + 450.0, + 922.0, + 480.0, + 591.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 450.0, + 1006.0, + 450.0, + 1006.0, + 480.0, + 947.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 477.0, + 1046.0, + 477.0, + 1046.0, + 511.0, + 355.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 503.0, + 606.0, + 503.0, + 606.0, + 538.0, + 353.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 503.0, + 960.0, + 503.0, + 960.0, + 538.0, + 636.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 532.0, + 560.0, + 532.0, + 560.0, + 562.0, + 356.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 532.0, + 955.0, + 532.0, + 955.0, + 562.0, + 591.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 559.0, + 570.0, + 559.0, + 570.0, + 593.0, + 355.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1372.0, + 646.0, + 1372.0, + 646.0, + 1406.0, + 354.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1402.0, + 603.0, + 1402.0, + 603.0, + 1435.0, + 354.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1430.0, + 1423.0, + 1430.0, + 1423.0, + 1463.0, + 354.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1456.0, + 570.0, + 1456.0, + 570.0, + 1489.0, + 354.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1345.0, + 1113.0, + 1345.0, + 1113.0, + 1381.0, + 292.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1762.0, + 1406.0, + 1762.0, + 1406.0, + 1798.0, + 296.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1791.0, + 1356.0, + 1791.0, + 1356.0, + 1827.0, + 295.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 865.0, + 1403.0, + 865.0, + 1403.0, + 901.0, + 296.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 891.0, + 1192.0, + 891.0, + 1192.0, + 934.0, + 292.0, + 934.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1326, + 1405, + 1326, + 1405, + 1662, + 297, + 1662 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 302, + 796, + 1400, + 796, + 1400, + 1155, + 302, + 1155 + ], + "score": 0.973, + "html": "
ModelPythonJavascriptRubyGo
Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1Prec.Rec.F1
code2seq111-11=1111-
GREAT34.9331.1231.6129.6924.2425.5525.6921.4922.1848.3845.9745.71
Ours w/o structure36.8732.1732.9731.3025.0326.6431.4325.3426.6349.7846.7346.69
Ours w/o pointer net38.7731.7233.2732.7025.5027.3332.1230.1729.3653.0948.7049.26
Ours36.6833.8633.8433.3627.5529.0231.5324.7226.4352.0047.3547.93
code2seq (Multilanguage)-====-
GREAT (Multilanguage)35.7330.8131.7431.4926.1727.4129.7224.2025.4350.3247.9447.66
Ours w/o structure (Mult.)36.7829.9231.5832.6026.0227.7431.7126.0727.2451.9147.5848.15
Ours w/o pointer (Mult.)37.1830.5232.0433.9525.9228.1132.7625.0427.0153.5048.5449.35
Ours (Multilanguage)38.1033.3234.1834.2928.6930.0833.3028.3329.2953.8650.4650.61
Ours (Mult. + Finetune)38.2932.4133.6534.4328.2829.9132.8927.1528.4953.8550.85
Ours (Mult.+ LMPretrain)50.81
38.9734.7735.3435.2330.2631.3833.7329.1529.9455.3152.0352.13
" + }, + { + "category_id": 3, + "poly": [ + 472, + 236, + 1229, + 236, + 1229, + 697, + 472, + 697 + ], + "score": 0.968 + }, + { + "category_id": 0, + "poly": [ + 298, + 1268, + 948, + 1268, + 948, + 1300, + 298, + 1300 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.892 + }, + { + "category_id": 1, + "poly": [ + 293, + 1757, + 1313, + 1757, + 1313, + 1789, + 293, + 1789 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.85 + }, + { + "category_id": 4, + "poly": [ + 410, + 733, + 1286, + 733, + 1286, + 767, + 410, + 767 + ], + "score": 0.766 + }, + { + "category_id": 6, + "poly": [ + 453, + 1177, + 1245, + 1177, + 1245, + 1209, + 453, + 1209 + ], + "score": 0.651 + }, + { + "category_id": 0, + "poly": [ + 299, + 1699, + 1209, + 1699, + 1209, + 1731, + 299, + 1731 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 299, + 1699, + 1209, + 1699, + 1209, + 1731, + 299, + 1731 + ], + "score": 0.321 + }, + { + "category_id": 6, + "poly": [ + 410, + 733, + 1286, + 733, + 1286, + 767, + 410, + 767 + ], + "score": 0.14 + }, + { + "category_id": 1, + "poly": [ + 453, + 1177, + 1245, + 1177, + 1245, + 1209, + 453, + 1209 + ], + "score": 0.103 + }, + { + "category_id": 13, + "poly": [ + 415, + 1106, + 433, + 1106, + 433, + 1123, + 415, + 1123 + ], + "score": 0.59, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 416, + 1128, + 433, + 1128, + 433, + 1147, + 416, + 1147 + ], + "score": 0.36, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 239.0, + 989.0, + 239.0, + 989.0, + 260.0, + 770.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 271.0, + 1213.0, + 271.0, + 1213.0, + 291.0, + 1172.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 288.0, + 1215.0, + 288.0, + 1215.0, + 312.0, + 1172.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 292.0, + 500.0, + 292.0, + 500.0, + 586.0, + 473.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 296.0, + 533.0, + 296.0, + 533.0, + 317.0, + 494.0, + 317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 312.0, + 1207.0, + 312.0, + 1207.0, + 329.0, + 1173.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 348.0, + 533.0, + 348.0, + 533.0, + 368.0, + 494.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 397.0, + 534.0, + 397.0, + 534.0, + 422.0, + 493.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 449.0, + 534.0, + 449.0, + 534.0, + 473.0, + 493.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 502.0, + 534.0, + 502.0, + 534.0, + 525.0, + 493.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 553.0, + 535.0, + 553.0, + 535.0, + 577.0, + 494.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 607.0, + 532.0, + 607.0, + 532.0, + 625.0, + 504.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 622.0, + 613.0, + 622.0, + 613.0, + 696.0, + 593.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 620.0, + 753.0, + 620.0, + 753.0, + 646.0, + 734.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 620.0, + 889.0, + 620.0, + 889.0, + 675.0, + 869.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 620.0, + 1028.0, + 620.0, + 1028.0, + 661.0, + 1006.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 619.0, + 1162.0, + 619.0, + 1162.0, + 694.0, + 1143.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1266.0, + 952.0, + 1266.0, + 952.0, + 1304.0, + 295.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2124.0, + 831.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 731.0, + 1290.0, + 731.0, + 1290.0, + 770.0, + 409.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1172.0, + 1247.0, + 1172.0, + 1247.0, + 1215.0, + 449.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1695.0, + 1213.0, + 1695.0, + 1213.0, + 1738.0, + 293.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 731.0, + 1290.0, + 731.0, + 1290.0, + 770.0, + 409.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1327.0, + 1403.0, + 1327.0, + 1403.0, + 1361.0, + 294.0, + 1361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1358.0, + 1407.0, + 1358.0, + 1407.0, + 1392.0, + 295.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1422.0, + 295.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1415.0, + 1406.0, + 1415.0, + 1406.0, + 1456.0, + 291.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1448.0, + 1403.0, + 1448.0, + 1403.0, + 1483.0, + 294.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1479.0, + 1405.0, + 1479.0, + 1405.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1510.0, + 1406.0, + 1510.0, + 1406.0, + 1545.0, + 294.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1541.0, + 1406.0, + 1541.0, + 1406.0, + 1573.0, + 291.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1572.0, + 1403.0, + 1572.0, + 1403.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1602.0, + 1406.0, + 1602.0, + 1406.0, + 1636.0, + 295.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1631.0, + 669.0, + 1631.0, + 669.0, + 1665.0, + 296.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1751.0, + 1321.0, + 1751.0, + 1321.0, + 1797.0, + 291.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1695.0, + 1213.0, + 1695.0, + 1213.0, + 1738.0, + 293.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1172.0, + 1247.0, + 1172.0, + 1247.0, + 1215.0, + 449.0, + 1215.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/rJerHlrYwH/rJerHlrYwH.md b/parse/train/rJerHlrYwH/rJerHlrYwH.md new file mode 100644 index 0000000000000000000000000000000000000000..9e976bb2c542b22bbfa2b334b47446ad5b7622b0 --- /dev/null +++ b/parse/train/rJerHlrYwH/rJerHlrYwH.md @@ -0,0 +1,298 @@ +# DATA-EFFICIENT IMAGE RECOGNITION WITH CONTRASTIVE PREDICTIVE CODING + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Human observers can learn to recognize new categories of objects from a handful of examples, yet doing so with machine perception remains an open challenge. We hypothesize that data-efficient recognition is enabled by representations which make the variability in natural signals more predictable, as suggested by recent perceptual evidence. We therefore revisit and improve Contrastive Predictive Coding, a recently-proposed unsupervised learning framework, and arrive at a representation which enables generalization from small amounts of labeled data. When provided with only $1 \%$ of ImageNet labels (i.e. 13 per class), this model retains a strong classification performance, $73 \%$ Top-5 accuracy, outperforming supervised networks by $28 \%$ (a $65 \%$ relative improvement) and state-of-the-art semisupervised methods by $14 \%$ . We also find this representation to serve as a useful substrate for object detection on the PASCAL-VOC 2007 dataset, approaching the performance of representations trained with a fully annotated ImageNet dataset. + +# 1 INTRODUCTION + +Deep neural networks excel at perceptual tasks when labeled data are abundant, yet their performance degrades substantially when provided with limited supervision (Fig. 1, red). In contrast, humans and animals can quickly learn about new classes of objects from few examples (Landau et al., 1988; Markman, 1989). What accounts for this monumental difference in data-efficiency between biological and machine vision? While highly-structured representations (e.g. as proposed by Lake et al., 2015) may improve data-efficiency, it remains unclear how to program explicit structures that capture the enormous complexity of real visual scenes like those in ImageNet (Russakovsky et al., 2015). An alternative hypothesis has proposed that intelligent systems need not be structured a priori, but can instead learn about the structure of the world in an unsupervised manner (Barlow, 1989; Hinton et al., 1999; LeCun et al., 2015). Choosing an appropriate training objective is an open problem, but a promising guiding principle has emerged recently: good representations should make the spatio-temporal variability in natural signals more predictable. Indeed, human perceptual representations have been shown to linearize + +![](images/7b8ed316aeaae9bdfe3d3d7e12974b01335c33b7eb790b24e724753c0422fb2e.jpg) +Figure 1: Data-efficient image recognition with Contrastive Predictive Coding. With decreasing amounts of labeled data, supervised networks trained on pixels fail to generalize (red). When trained on unsupervised representations learned with CPC, these networks retain a much higher accuracy in this low-data regime (blue). Equivalently, the accuracy of supervised networks can be matched with significantly fewer labels. + +(or ‘straighten’) the temporal transformations found in natural videos, a property lacking from current supervised image recognition models (Henaff et al., 2019), and theories of both spatial and tem- ´ poral predictability have succeeded in describing properties of early visual areas (Rao & Ballard, + +1999; Palmer et al., 2015). In this work, we hypothesize that spatially predictable representations may allow artificial systems to benefit from human-like data-efficiency. + +Contrastive Predictive Coding (CPC, van den Oord et al., 2018) is an unsupervised objective which learns such predictable representations. CPC is a general technique that only requires in its definition that observations be ordered along e.g. temporal or spatial dimensions, and as such has been applied to a variety of different modalities including speech, natural language and images. This generality, combined with the strong performance of its representations in downstream linear classification tasks, makes CPC a promising candidate for investigating the efficacy of predictable representations for data-efficient image recognition. + +Our work makes the following contributions: + +• We revisit CPC in terms of its architecture and training methodology, and arrive at a new implementation of CPC with dramatically-improved ability to linearly separate image classes $+ 1 7 \%$ Top-1 ImageNet classification accuracy). +We then train deep networks on top of the resulting CPC representations using very few labeled images (e.g. $1 \%$ of the ImageNet dataset), and demonstrate test-time classification accuracy far above networks trained on raw pixels ( $73 \%$ Top-5 accuracy, a $28 \%$ absolute improvement), outperforming all other unsupervised representation learning methods $( + 1 5 \%$ Top-5 accuracy over the previous state-of-the-art (Zhai et al., 2019)). Surprisingly, this representation also surpasses supervised methods when given the entire ImageNet dataset $+ 1 \%$ Top-5 accuracy). +We isolate the contributions of different components of the final model to such downstream tasks. Interestingly, we find that linear classification accuracy is not always predictive of low-data classification accuracy, emphasizing the importance of this metric as a stand-alone benchmark for unsupervised learning. +• Finally, we assess the generality of CPC representations by transferring them to a new task and dataset: object detection on PASCAL-VOC 2007. Consistent with the results from the previous section, we find CPC to give state-of-the-art performance in this setting. + +# 2 EXPERIMENTAL SETUP + +We first review the CPC architecture and learning objective in section 2.1, before detailing how we use its resulting representations for image recognition tasks in section 2.2. + +# 2.1 CONTRASTIVE PREDICTIVE CODING + +Contrastive Predictive Coding as formulated in (van den Oord et al., 2018) learns representations by training neural networks to predict the representations of future observations from those of past ones. When applied to images, the original formulation of CPC operates by predicting the representations of patches below a certain position from those above it (Fig. 2, left). These predictions are evaluated using a contrastive loss, in which the network must correctly classify the ‘future’ representation amongst a set of unrelated ‘negative’ representations. This avoids trivial solutions such as representing all patches with a constant vector, as would be the case with a mean squared error loss. + +In the CPC architecture, each input image is first divided into a set of overlapping patches $\mathbf { \delta } _ { \mathbf { x } _ { i , j } }$ , each of which is encoded with a neural network $f _ { \theta }$ into a single vector $z _ { i , j } = f _ { \theta } ( \pmb { x } _ { i , j } )$ . To make predictions, a masked convolutional network $g _ { \phi }$ is then applied to the grid of feature vectors. The masks are such that the receptive field of each resulting context vector $c _ { i , j }$ only includes feature vectors that lie above it in the image (i.e. $\{ z _ { u , v } \} _ { u \leq i , v } )$ . The prediction task then consists o predicting ‘future’ feature vectors $z _ { i + k , j }$ from current context vectors $c _ { i , j }$ , where $k > 0$ . The predictions are made linearly: given a context vector $c _ { i , j }$ , a prediction length $k > 0$ , and a prediction matrix $W _ { k }$ , the predicted feature vector is $\hat { \boldsymbol { z } } _ { i + k , j } = \boldsymbol { W } _ { k } \boldsymbol { c } _ { i , j }$ . + +The quality of this prediction is then evaluated using a contrastive loss. Specifically, the goal is to correctly recognize the target $z _ { i + k , j }$ among a set of randomly sampled feature vectors $\{ z _ { l } \}$ from the dataset. We compute the probability assigned to the target using a softmax, and evaluate this probability using the usual cross-entropy loss. Summing this loss over locations and prediction offsets, we arrive at the CPC objective as defined in (van den Oord et al., 2018): + +![](images/e21408d1a55db66ba2cac6a3f4d98b0d1a13c2f1211e3956f93d67bae89bbaae.jpg) +Figure 2: Overview of the framework for semi-supervised learning with Contrastive Predictive Coding. Left: unsupervised pre-training with the spatial prediction task (See Section 2.1). First, an image is divided into a grid of overlapping patches. Each patch is encoded independently from the rest with a feature extractor (blue) which terminates with a mean-pooling operation, yielding a single feature vector for that patch. Doing so for all patches yields a field of such feature vectors (wireframe vectors). Feature vectors above a certain level (in this case, the center of the image) are then aggregated with a context network (red), yielding a row of context vectors which are used to linearly predict features vectors below. Right: using the CPC representation for a classification task. Having trained the encoder network, the context network (red) is discarded and replaced by a classifier network (green) which can be trained in a supervised manner. For some experiments, we also fine-tune the encoder network (blue) for the classification task. When applying the encoder to cropped patches (as opposed to the full image) we refer to it as a patched ResNet in the figure. + +$$ +\mathcal { L } _ { \mathrm { C P C } } = - \sum _ { i , j , k } \log p ( z _ { i + k , j } | \hat { z } _ { i + k , j } , \{ z _ { l } \} ) = - \sum _ { i , j , k } \log \frac { \exp ( \hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) } { \exp ( \hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) + \sum _ { l } \exp ( \hat { z } _ { i + k , j } ^ { T } z _ { l } ) } +$$ + +The negative samples $\left\{ z _ { l } \right\}$ are taken from other locations in the image and other images in the minibatch. This loss is called InfoNCE (van den Oord et al., 2018) as it is inspired by Noise-Contrastive Estimation (Gutmann $\&$ Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013) and has been shown to ¨ maximize the mutual information between $\mathbf { } _ { c _ { i , j } }$ and $z _ { i + k , j }$ (van den Oord et al., 2018). + +# 2.2 EVALUATION PROTOCOL + +Having trained an encoder network $f _ { \theta }$ , a context network $g _ { \phi }$ , and a set of linear predictors $\{ W _ { k } \}$ using the CPC objective, we use the latents $z = f _ { \boldsymbol { \theta } } ( \pmb { x } )$ as a representation of new observations $x$ for downstream tasks, and discard the rest. We then train a model $h _ { \psi }$ to classify these representations given a dataset of labeled images. More formally, given a dataset of $N$ unlabeled images $\mathbb { D } _ { u } =$ $\left\{ x _ { n } \right\}$ , and a (potentially much smaller) dataset of $M$ labeled images $\mathbb { D } _ { l } = \{ x _ { m } , y _ { m } \}$ : + +$$ +\theta ^ { * } = \arg \operatorname* { m i n } _ { \theta } \frac { 1 } { N } \sum _ { n = 1 } ^ { N } \mathcal { L } _ { \mathrm { C P C } } [ f _ { \theta } ( x _ { n } ) ] , \quad \psi ^ { * } = \arg \operatorname* { m i n } _ { \psi } \frac { 1 } { M } \sum _ { m = 1 } ^ { M } \mathcal { L } _ { \mathrm { S u p } } [ h _ { \psi } \circ f _ { \theta ^ { * } } ( x _ { m } ) , y _ { m } ] +$$ + +In all cases, the dataset of unlabeled images $\mathbb { D } _ { u }$ we pre-train on is the full ImageNet ILSVRC 2012 training set (Russakovsky et al., 2015). We consider three labeled datasets $\mathbb { D } _ { l }$ for evaluation, each with an associated classifier $h _ { \psi }$ and supervised losse $\mathcal { L } _ { \mathrm { { S u p } } }$ (see Fig. 2, right). This protocol is sufficiently generic to allow us to later compare the CPC representation to other methods which have their own means of learning a feature extractor $f _ { \theta }$ . + +Linear classification is the standard benchmark for evaluating the quality of unsupervised image representations. In this regime, the classification network $h _ { \psi }$ is restricted to mean pooling followed by a single linear layer, and the parameters of $f _ { \theta }$ are kept fixed. The labeled dataset $\mathbb { D } _ { l }$ is the entire ImageNet dataset, and the supervised loss $\mathcal { L } _ { \mathrm { { S u p } } }$ is standard cross-entropy. We use the same dataaugmentation as in the unsupervised learning phase for training, and none at test time and evaluate with a single crop. + +Efficient classification directly tests whether the CPC representation enables visual learning from few labels. For this task, the classifier $h _ { \psi }$ is an arbitrary deep neural network (we use an 11-block ResNet architecture with 4096-dimensional feature maps and 1024-dimensional bottleneck layers). The labeled dataset $\mathbb { D } _ { l }$ is a subset of the ImageNet dataset: we investigated using $1 \%$ , $2 \%$ , $5 \%$ , $10 \%$ , $20 \%$ , $50 \%$ and $100 \%$ of the ImageNet dataset. The supervised loss $\mathcal { L } _ { \mathrm { { S u p } } }$ is again cross-entropy. In addition to random color-dropping we use the Inception data-augmentation scheme (Szegedy et al., 2014) for training, no augmentation at test-time and evaluate with a single crop. + +Transfer learning tests the generality of the representation by applying it to a new task and dataset. For this we chose image detection on the PASCAL-2007 dataset, a standard benchmark in computer vision (Everingham et al., 2007). As such $\mathbb { D } _ { l }$ is the entire PASCAL-2007 dataset (comprised of 5011 labeled images); $h _ { \psi }$ and $\mathcal { L } _ { \mathrm { { S u p } } }$ are the Faster-RCNN architecture and loss (Ren et al., 2015). In addition to color-dropping, we use scale-augmentation (Doersch et al., 2015) for training. + +For linear classification, we keep the feature extractor $f _ { \theta }$ fixed to assess the representation in absolute terms. For efficient classification and transfer learning, we additionally explore fine-tuning the feature extractor for the supervised objective. In this regime, we initialize the feature extractor and classifier with the solutions $\theta ^ { * } , \psi ^ { * }$ found in the previous learning phase, and train them both for the supervised objective. To ensure that the feature extractor does not deviate too much from the solution dictated by the CPC objective, we use a smaller learning rate and early-stopping. + +# 3 RELATED WORK + +Data-efficient learning has typically been approached by two complementary methods, both of which seek to make use of more plentiful unlabeled data: representation learning and semisupervised learning. The former formulates an objective to learn a feature extractor $f _ { \theta }$ in an unsupervised manner, whereas the latter directly constrains the classifier $h _ { \psi }$ using the unlabeled data. + +Representation learning saw early success using generative modeling (Kingma et al., 2014), but likelihood-based models have yet to generalize to more complex stimulus classes. Generative adversarial models have also been harnessed for representation learning (Donahue et al., 2016), and large-scale implementations have recently achieved corresponding gains in linear classification accuracy (Donahue & Simonyan, 2019). + +In contrast to generative models which require the reconstruction of observations, self-supervised techniques directly formulate tasks involving the learned representation. For example, simply asking a network to recognize the spatial layout of an image led to representations that transferred to popular vision tasks such as classification and detection (Doersch et al., 2015; Noroozi & Favaro, 2016). Other works showed that prediction of color (Zhang et al., 2016; Larsson et al., 2017) and image orientation (Gidaris et al., 2018), and invariance to data augmentation (Dosovitskiy et al., 2014) can provide useful self-supervised tasks. Beyond single images, works have leveraged video cues such as object tracking (Wang & Gupta, 2015), frame ordering (Misra et al., 2016), and object boundary cues (Li et al., 2016; Pathak et al., 2016). Non-visual information can be equally powerful; information about camera motion (Agrawal et al., 2015; Jayaraman & Grauman, 2015), scene geometry (Zamir et al., 2016), or sound (Arandjelovic & Zisserman, 2017; 2018) can all serve as natural sources of supervision. + +While many of these tasks require predicting fixed quantities computed from the data, another class of contrastive methods formulate their objectives in the learned representations themselves. CPC is a contrastive representation learning method that maximizes the mutual information between spatially removed latent representations with InfoNCE (van den Oord et al., 2018), a loss function based on Noise-Contrastive Estimation (Gutmann & Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013). Two ¨ other methods have recently been proposed using the same loss function, but with different associated prediction tasks. Contrastive Multiview Coding (Tian et al., 2019) maximizes the mutual information between representations of different views of the same observation. Augmented Multiscale Deep InfoMax (AMDIM, Bachman et al., 2019) is most similar to CPC in that it makes predictions across space, but differs in that it also predicts representations across layers in the model. In addition, AMDIM limits the receptive field of its representation, but does this by constraining the number of spatial convolutions in the network architecture rather than using image patches. + +A common alternative approach for improving data efficiency is label-propagation (Zhu & Ghahramani, 2002), where a classifier is trained on a subset of labeled data, then used to label parts of the unlabeled dataset, after which the process is repeated. This label-propagation can either be discrete (as in pseudo-labeling, Lee, 2013) or continuous (as in entropy minimization, Grandvalet & Bengio, 2005). The predictions of this classifier are often constrained to be smooth with respect to certain deformations, such as data-augmentation (Xie et al., 2019) or adversarial perturbation (Miyato et al., 2018). Representation learning and semi-supervised learning have been shown to be complementary and can be combined to great effect (Zhai et al., 2019), which is why we focus solely on representation learning in this paper. + +# 4 RESULTS + +When asking whether CPC enables data-efficient learning, we wish to use the best possible representative of this model class. Unfortunately, purely unsupervised metrics tell us little about downstream performance, and implementation details have been shown to matter enormously (Doersch & Zisserman, 2017; Kolesnikov et al., 2019). Since many design choices (e.g. network architecture and datapreprocessing) have been previously evaluated using linear classification, we use this benchmark in section 4.1 to align the CPC model with best practices in representation learning and compare to published results. In section 4.2 we select the best performing model from the previous section and assess whether it enables efficient classification. We also investigate to what extent the first, more common metric (linear classification accuracy) is predictive of efficient classification. Finally, in section 4.3 we investigate the generality of our results through transfer learning to PASCAL-2007. + +# 4.1 FROM CPC V1 TO CPC V2 + +The overarching principle behind our new model design is to increase the scale and efficiency of the encoder architecture while also maximizing the supervisory signal we obtain from each image. At the same time, it is important not to allow the network to solve the problem trivially, i.e., without learning semantics. To this end, we seek to remove low-level cues common across patches by augmenting individual patches independently, using standard stochastic data-processing techniques from supervised and self-supervised learning. + +We identify four axes for model capacity and task setup that could impact the model’s performance. The first axis increases model capacity by increasing depth and width, while the second improves training efficiency capacity by introducing layer normalization. The third axis increases task complexity by making predictions in all four directions, and the fourth does so by performing more extensive patch-based augmentation. + +Model capacity. Recent work has shown that networks and more effective training improves self-supervised learning (Doersch & Zisserman, 2017; Kolesnikov et al., 2019), but the original CPC model used only the first 3 stacks of a ResNet-101 (He et al., 2016a) architecture (i.e. a ResNet-92). Therefore, we converted the third residual stack of ResNet-101 (originally containing 23 blocks, 1024-dimensional feature maps, and 256-dimensional bottleneck layers), to use 46 blocks, with 4096-dimensional feature maps and 512-dimensional bottleneck layers. We call the resulting network ResNet-161. Consistent with prior results, this new architecture delivers better performance regardless of + +![](images/3eeded7c7663fc17771caa90292cc1343662124f1094d06f22d9ca255036f2a3.jpg) +Figure 3: Linear classification performance of new variants of CPC, which incrementally add a series of modifications. BU: bottum up spatial predictions. HF: randomly flipping patches horizontally. LN: layer normalization. RC: random color-dropping. TL: tuned prediction lengths. HP: horizontal spatial predictions. We use color to indicate the number of spatial predictions used (orange, green, blue for 1, 2 and 4 directions). + +Table 1: Linear classifier $h _ { \psi }$ trained with $100 \%$ of labels. Comparison to linear separability of other self-supervised methods. In all cases a feature extractor is optimized in an unsupervised manner, and a linear classifier is trained on top using all labels in the ImageNet dataset. + +
MethodTop-1Top-5
Motion Segmentation (MS) (Pathak et al., 2016) Exemplar (Ex) (Dosovitskiy et al.,2014) Relative Position (RP) (Doersch et al.,2015) Colorization (Col) (Zhang et al., 2016) Combination of MS + Ex + RP+ Col (Doersch & Zisserman,2017) CPC v1 (van den Oord et al., 2018) Rotation (Kolesnikov et al., 2019)27.6 31.5 36.2 39.6 1 48.7 55.448.3 53.1 59.2 62.5 69.3 73.6
CMC (Tian et al., 2019) Local Aggregation (Zhuang et al., 2019)60.1 60.282.8 1 =
BigBiGAN (Donahue & Simonyan,2019) AMDIM (Bachman et al., 2019) CPC v2 (ours) 65.961.3 68.181.9
+ +other design choices. Interestingly, a larger architecture delivers larger improvements with more efficient training, more self-supervised losses, and more patch-based augmentations (Fig. 3, $+5 \%$ Top-1 accuracy with original training scheme, $+ 1 0 \%$ accuracy with new one). + +Layer normalization. Large architectures are more difficult to train efficiently. Early works on context prediction with patches used batch normalization (Ioffe & Szegedy, 2015; Doersch et al., 2015) to speed training. However, with CPC we find that batch normalization actually harms downstream performance of large models. We hypothesize that batch normalization allows large models to find a trivial solution to CPC: it introduces a dependency between patches (through the batch statistics) that can be exploited to bypass the constraints on the receptive field. We find that we can reclaim much of batch normalization’s training efficiency using layer normalization (Ba et al., 2016), which leads to a small gain for the smaller architecture $+ 1 \%$ accuracy over equivalent architectures that use neither normalization) and a larger gain for the larger architecture $( + 2 . 5 \%$ accuracy). + +Prediction lengths and directions. Larger architectures also run a greater risk of overfitting. We address this by asking more from the network: specifically, whereas van den Oord et al. (2018) predicted each patch using only context from spatially beneath it, we repeatedly predict the patch using context from above, to the right, and to the left, resulting in up to four times as many prediction tasks. Combining top-to-bottom with bottom-to-top helps both model architectures $42 \%$ accuracy for both), but using all 4 spatial directions only benefits the larger model (an additional $+ 1 . 5 \%$ for the larger model, ${ \bf - 1 \% }$ for the smaller), consistent with the idea that model capacity and amount of supervision must go hand-in-hand. We also hypothesized that prediction “length”—i.e. offset between the predicted patch and the aggregated context—might affect performance, as distant patches might lie on distinct objects, encouraging the network to memorize images. Indeed, limiting the range of the prediction length $k$ to $\{ 2 , 3 \}$ performed better than $\{ 2 , \ldots , 5 \}$ as was used originally $+ 1 \%$ for the larger model). + +Patch-based augmentation. If the network can solve CPC using low-level patterns (e.g. straight lines continuing between patches, chromatic aberration), it need not learn semantically meaningful content. Augmenting the low-level variability across patches can remove such low level cues. The original CPC model spatially jitters individual patches independently. We further this logic by adopting the ‘color dropping’ method of Doersch et al. (2015), which randomly drops two of the three color channels in each patch, and find it to delivers systematic gains $( + 1 \%$ for the small model, $+3 \%$ for the larger one). We also randomly flip patches horizontally, but find it only benefits the smaller model $( + 1 \% )$ . + +Combined. Cumulatively, these fairly straightforward implementation changes lead to a substantial improvement to the original CPC model $6 5 . 9 \%$ Top-1 accuracy, a $17 \%$ improvement), making it competitive with recent approaches and outperforming prior methods (see table 1). Interestingly, if we train the same patch-based architecture from scratch in a fully supervised manner, we obtain $6 6 . 4 \%$ Top-1 accuracy (with batch normalization; $6 2 . 5 \%$ without), suggesting that CPC is now nearly saturating the architecture’s representational power despite not using labels. These results illustrate how architecture and data have an outsized impact on the linear classification performance of self-supervised representations, and are interesting to compare with with previous results. For example, in AMDIM, different settings of data augmentation alone can result in a nearly $10 \%$ absolute increase in performance on ImageNet linear classification. + +# 4.2 EFFICIENT IMAGE CLASSIFICATION + +toWe now turn to our original question of whether CPC can enable data-efficient image recognition. We start by evaluating the performance of purely-supervised networks as the size of the labeled dataset $\mathbb { D } _ { l }$ varies from $1 \%$ to $100 \%$ of ImageNet, training separate classifiers on each subset. We found that a ResNet-152 to works best across all data-regimes (see Appendix). Despite our efforts to tune the supervised model for low-data classification (including network depth, regularization, and optimization parameters), the accuracy of the best model only reaches $4 4 . 1 \%$ Top-5 accuracy when trained on $1 \%$ of the dataset (compared to $9 3 . 9 \%$ when trained on the entire dataset, see Fig. 1, red). + +Contrastive Predictive Coding. We now address our central question of whether CPC enables data-efficient learning. We follow the same paradigm as for the supervised baseline (training and evaluating a separate classifier for each size subset), stacking a neural network classifier on top of the CPC latents $\begin{array} { r l r } { z } & { { } = } & { f _ { \theta } ( { \pmb x } ) } \end{array}$ rather than the raw image pixels $_ { \textbf { \em x } }$ (see section 2.2, efficient classification, and Appendix). This representation, which we selected for its improved linear classification performance (CPC v2 in Fig. 3), leads to a significant increase in data-efficiency compared to purely supervised networks (Fig. 1, blue curve). This classifier yields $7 2 . 9 \%$ Top-5 accuracy with only $1 \%$ of the labels, a $2 9 \%$ absolute improvement $65 \%$ relative) over purely-supervised methods. Surprisingly, when given the entire dataset, this classifier reaches $8 0 . 6 \% / 9 5 . 2 \%$ Top1/Top5 accuracy, surpassing our supervised baseline (ResNet-152: $7 8 . 0 \% / 9 3 . 9 \%$ accuracy) and published results (ResNet-200: $7 9 . 9 \% / 9 5 . 2 \%$ , He et al. (2016b)). We find similar results in all other data-regimes we considered (see Fig. 1). + +How important are the model specifications described in Section 4.1 for low-data classification? We hypothesized that predictable representations might enable data-efficient classifi + +![](images/106f2252ad377e16a329cfb21360459a72fdc1d09e4f745b26572da1cf6d5d14.jpg) +Figure 4: Relationship between linear classification accuracy and low-data classification, for different variants of the CPC model. Left: CPC variants with the same architecture but different training protocols. Orange, green, and blue dots correspond to CPC models making predictions in 1, 2, and 4 spatial directions respectively. Within a color group, different models correspond to other implementation details (e.g. layer norm and patch augmentation, and combinations thereof). + +cation, and therefore expect that increasing the amount of ‘predictability’ in the representation should also increase its ability to learn from small amounts of data. Fig. 4 shows evidence for this by ablating model parameters and comparing linear classification performance against low-data classification. Consistent with our hypothesis, increasing the number of spatial directions in the CPC prediction task (which increased linear classification performance) systematically increases low-data classification performance (Fig. 4, left, different color groups). As a control, we asked if all modifications that improve linear classification also improve low-data classification. We did not find evidence in favor of this: improvements in linear classification as a result of changing other model parameters (patch-based data-augmentation, layer normalization, and combinations thereof) seem uncorrelated to performance in other tasks (Fig. 4, left, within green group: $R ^ { 2 } = 0 . 1 7 , p = 0 . 3 6$ ). Different architectural specifications also produced different changes in both tasks (Fig. 4, right). Whereas increasing the depth of the encoding network greatly improves both metrics, increasing the network width (and therefore the number of features used for linear classification) only improves linear classification accuracy. + +Table 2: ResNet classifier $h _ { \psi }$ trained with $1 \%$ or $10 \%$ of labels. Comparison to other methods for semi-supervised learning using $1 \%$ or $10 \%$ of labeled data. Representation learning methods use a classifier to discriminate an unsupervised representation, and optimize it solely for the supervised objective on labeled data. Label-propagation methods on the other hand further constrain the classifier with smoothness and entropy criteria on unlabeled data, making the additional assumption that all training images fit into a single (unknown) testing category. $\dagger$ denotes methods implemented in this work, fixed and fine-tuned denote whether the feature extractor is allowed to accommodate the supervised objective. + +
Labeled data 1% 10% 100% Method Top-5 accuracy
Supervised baseline 44.1 82.1 93.9 Methods using label-propagation:
Pseudolabeling (Zhai et al., 2019) 51.6 82.4 VAT + Entropy Minimization (Zhai et al., 2019) 47.0 83.4 Unsup. Data Augmentation (Xie et al., 2019) 88.5
Rotation+ VAT+Ent.Min. (Zhai et al.,2019) 91.2 95.0 =
Methods using representation learning only:
Instance Discrimination (Wu et al., 2018) 39.2 77.4
Rotation (Zhai et al., 2019) 57.5 86.4
+ResNet trained on BigBiGAN (fixed) 55.2 78.8 87.0
+ResNet trained on AMDIM (fixed) 67.4 85.8 92.2
+ResNet trained on CPC v2 (fixed) 72.3 89.1 94.4 +ResNet trained on CPC v2 (fine-tuned) 72.9 89.5 95.2
+ +Other unsupervised representations. How well does the CPC representation compare to other representations that have been learned in an unsupervised manner? If predictable representations are uniquely suited for efficient classification, we would expect other methods within this family to perform similarly, and other model classes less so. Table 2 compares our best model with other works on efficient recognition. We consider three objectives from different model classes: self-supervised learning with rotation prediction (Zhai et al., 2019), large-scale adversarial feature learning (BigBiGAN, Donahue & Simonyan, 2019), and another contrastive prediction objective (AMDIM, Bachman et al., 2019). Zhai et al. (2019) evaluate the low-data classification performance of representations learned with rotation prediction using a similar paradigm and architecture (ResNet-152), hence we report their results directly. Given $1 \%$ of ImageNet, their method achieves $5 7 . 5 \%$ Top-5 accuracy, consistently with the reduced accuracy of a linear classifier ( $5 5 . 4 \%$ vs $6 5 . 9 \%$ for CPC)1. + +Because BigBiGAN and AMDIM achieve stronger linear classification accuracy than rotation prediction $6 1 . 3 \%$ and $6 8 . 1 \%$ Top-1 accuracy, respectively), we might expect better performance on efficient classification as well. Since their authors do not report results on efficient classification we evaluated these representations using the same paradigm we used for evaluating CPC, stacking a ResNet classifier on top of the $7 \times 7 \times 8 1 9 2$ latents of the BigBiGAN and the $7 \times 7 \times 2 5 6 0$ grid of feature vectors of AMDIM. We found fine-tuned representations to yield only marginal gains over fixed ones $7 2 . 9 \%$ compared to $7 2 . 3 \%$ Top-5 accuracy given $1 \%$ of labels), hence for simplicity we evaluate BigBiGAN and AMDIM on this task while keeping them fixed. We re-tune the hyper-parameters of the classifier (including optimization, regularization, etc.) for each of these representations separately. Although these methods achieve similar performance in terms of linear classification, we find them to achieve very different results in efficient classification. Given $1 \%$ of ImageNet, classifiers trained on top of BigBiGAN achieve $5 5 . 2 \%$ Top-5 accuracy, similarly to rotation prediction $( 5 7 . 5 \% )$ , despite its increased linear classification accuracy $( + 6 \%$ relative to rotation prediction). In contrast, AMDIM (which also belongs to the family of contrastive prediction methods) achieves $6 7 . 4 \%$ on this same task. Again, its increased linear classification accuracy did not entail an increase in data-efficiency. Nevertheless, in line with our initial hypothesis, we find that contrastive prediction methods such as CPC surpass other approaches in our efficient classification experiments, and that linear classification performance is not perfectly correlated with these results. + +Other semi-supervised techniques A separate class of methods for low-data classification attempts to propagate the knowledge extracted from the subset of labeled examples to unlabeled examples while being invariant to augmentation or other perturbations. These methods generally depend on the quality of the classifier’s predictions, and as such tend to fare well when given intermediate amounts of data. Although not sufficient in themselves (Unsupervised Data Augmentation (Xie et al., 2019), Virtual Adversarial Training (Miyato et al., 2018) and entropy minimization (Grandvalet & Bengio, 2005), and pseudo-labeling (Lee, 2013) achieve $8 5 . 8 \%$ , $8 3 . 4 \%$ , and $8 2 . 4 \%$ Top-5 accuracy with $10 \%$ of labels, compared to our $8 9 . 4 \%$ ) when combined with representation learning (e.g. rotation prediction Zhai et al., 2019) they can provide considerable gains $9 1 . 2 \%$ Top5 accuracy). It is therefore surprising that CPC representations alone can enable accuracy that is comparable to that of these methods, and investigating to what extent they can be combined would be an interesting topic of future work. + +# 4.3 TRANSFER LEARNING: IMAGE DETECTION ON PASCAL VOC 2007 + +We next investigate transfer performance on object detection on the PASCAL-2007 dataset, which reflects the practical scenario where a representation must be trained on a dataset with different statistics than the dataset of interest. This dataset also tests the efficiency of the representation as it only contains 5011 labeled images to train from. In this setting, we replaced the neural network classifier $h _ { \psi }$ used previously with a Faster-RCNN (Ren et al., 2015) image detection architecture, and use the pre-trained feature extractor on ImageNet. As before, we first trained the Faster-RCNN model while keeping the feature extractor fixed, then fine-tuned the entire model end-to-end. Table 3 displays our results compared to other methods. Most competing methods, which optimize a single unsupervised objective on ImageNet before fine-tuning on PASCAL detection, attain around $65 \%$ mean average precision. Leveraging larger unlabeled datasets increases their performance up to $6 7 . 8 \%$ (Caron et al., 2019). Combining multiple forms of self-supervision enables them to reach $7 0 . 5 \%$ (Doersch & Zisserman, 2017). The proposed method, which learns only from ImageNet data using a single unsupervised objective, reaches $7 0 . 6 \%$ when equipped with a ResNet-101 feature extractor $f _ { \theta }$ (as for most competing methods (Doersch & Zisserman, 2017) but not all (Caron et al., 2018; 2019). Equipped with the more powerful ResNet-161 feature extractor $f _ { \theta }$ , our method reaches $7 2 . 7 \%$ . Importantly, this result is only $2 \%$ short of the performance attained by purely supervised transfer learning, which we obtain by using all ImageNet labels before transferring to PASCAL. + +# 5 DISCUSSION + +We asked whether CPC could enable data-efficient image recognition, and found that it indeed greatly improves the accuracy of classifiers and object detectors when given small amounts of labeled data. Surprisingly, CPC even improves results given ImageNet-scale labels. Our results show that there is still room for improvement using relatively straightforward changes such as augmentation, optimization, and network architecture. Furthermore, we found that the standard method for evaluating unsupervised representations—linear classification—is only partially predictive of efficient recognition performance, suggesting that further research should focus on efficient recognition as a standalone benchmark. Overall, these results open the door toward research on problems where data is naturally limited, e.g. medical imaging or robotics. + +Table 3: Faster-RCNN $h _ { \psi }$ trained with $100 \%$ of PASCAL labels. Comparison of PASCAL 2007 image detection accuracy to other transfer methods. The supervised baseline learns from the entire labeled ImageNet dataset and fine-tunes for PASCAL detection. The second class of methods learns from the same unlabeled images before transferring. All of these methods pre-train on the ImageNet dataset, except for DeeperCluster which learns from the larger, but uncurated, YFCC100M dataset (Thomee et al., 2015). All results are reported in terms of mean average precision (mAP). $\dagger$ denotes methods implemented in this work. + +
MethodmAP
Transfer from labeled data: Supervised - ResNet-15274.7
Transfer from unlabeled data: Exemplar (Ex) (Dosovitskiy et al., 2014) Motion Segmentation (MS) (Pathak et al.,2016)60.9
Colorization (Col) (Zhang et al.,2016) Relative Position (RP) (Doersch et al.,2015)61.1 65.5
Combination of Ex + MS + Col +RP (Doersch & Zisserman,2017)66.8 70.5
Instance Discrimination (Wu et al., 2018)65.4
Deep Cluster (Caron et al., 2018) Deeper Cluster (Caron et al., 2019)65.9
+ +Furthermore, images are far from the only domain where unsupervised representation learning is important: for example, unsupervised learning is already a critical step in language (Mikolov et al., 2013; Devlin et al., 2018), and shows promise in domains like audio (van den Oord et al., 2018; Arandjelovic & Zisserman, 2018; 2017), video (Jing & Tian, 2018; Misra et al., 2016), and robotic manipulation (Pinto & Gupta, 2016; Pinto et al., 2016; Sermanet et al., 2018). Currently much self-supervised work builds upon tasks tailored for a specific domain (often images), which may not be easily adapted to other domains. Contrastive prediction methods, including the techniques suggested in this paper, are task agnostic and could therefore serve as a unifying framework for integrating these tasks and modalities. This generality is particularly useful given that many realworld environments are inherently multimodal, e.g. robotic environments which can have vision, audio, touch, proprioception, action, and more over long temporal sequences. Given the importance of increasing the amounts of self-supervision (via additional directions of prediction), integrating these modalities and tasks could lead to unsupervised representations which rival the efficiency and effectiveness of biological ones. + +# REFERENCES + +Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. In ICCV, 2015. + +Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision, pp. 609–617, 2017. + +Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 435–451, 2018. + +Lei Jimmy Ba, Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. + +Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. + +H.B. Barlow. Unsupervised learning. Neural Computation, 1(3):295–311, 1989. doi: 10.1162/neco. 1989.1.3.295. + +Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In The European Conference on Computer Vision (ECCV), September 2018. + +Mathilde Caron, Piotr Bojanowski, Julien Mairal, and Armand Joulin. Leveraging large-scale uncurated data for unsupervised pre-training of visual features. 2019. + +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018. + +Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2051–2060, 2017. + +Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1422–1430, 2015. + +Jeff Donahue and Karen Simonyan. Large scale adversarial representation learning. arXiv preprint arXiv:1907.02544, 2019. + +Jeff Donahue, Philipp Krahenb ¨ uhl, and Trevor Darrell. Adversarial feature learning. ¨ arXiv preprint arXiv:1605.09782, 2016. + +Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. In Advances in neural information processing systems, pp. 766–774, 2014. + +Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge 2007 (voc2007) results. 2007. + +Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. + +Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in neural information processing systems, pp. 529–536, 2005. + +Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨ for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pp. 297–304, 2010. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016a. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pp. 630–645. Springer, 2016b. + +G.E. Hinton, T.J. Sejnowski, H.H.M.I.C.N.L.T.J. Sejnowski, and T.A. Poggio. Unsupervised Learning: Foundations of Neural Computation. A Bradford Book. MIT Press, 1999. ISBN 9780262581684. + +Olivier J. Henaff, Robbe L. T. Goris, and Eero P. Simoncelli. Perceptual straightening of natural ´ videos. Nature Neuroscience, Apr 2019. + +Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. + +Dinesh Jayaraman and Kristen Grauman. Learning image representations tied to ego-motion. In ICCV, 2015. + +Longlong Jing and Yingli Tian. Self-supervised spatiotemporal feature learning by video geometric transformations. arXiv preprint arXiv:1811.11387, 2018. + +Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. In Advances in neural information processing systems, pp. 3581–3589, 2014. + +Alexander Kolesnikov, Xiaohua Zhai, and Lucas Beyer. Revisiting self-supervised visual representation learning. CoRR, abs/1901.09005, 2019. URL http://arxiv.org/abs/1901.09005. + +Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332–1338, 2015. + +Barbara Landau, Linda B Smith, and Susan S Jones. The importance of shape in early lexical learning. Cognitive development, 3(3):299–321, 1988. + +Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In CVPR, pp. 6874–6883, 2017. + +Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436, 2015. + +Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on Challenges in Representation Learning, ICML, volume 3, pp. 2, 2013. + +Yin Li, Manohar Paluri, James M Rehg, and Piotr Dollar. Unsupervised learning of edges. In ´ CVPR, 2016. + +Ellen M Markman. Categorization and naming in children: Problems of induction. mit Press, 1989. + +Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 26, pp. 3111–3119. Curran Associates, Inc., 2013. + +Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In ECCV, 2016. + +Takeru Miyato, Shin-ichi Maeda, Shin Ishii, and Masanori Koyama. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 2018. + +Andriy Mnih and Koray Kavukcuoglu. Learning word embeddings efficiently with noise-contrastive estimation. In Advances in neural information processing systems, pp. 2265–2273, 2013. + +Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European Conference on Computer Vision, pp. 69–84. Springer, 2016. + +Stephanie E Palmer, Olivier Marre, Michael J Berry, and William Bialek. Predictive information in a sensory population. Proceedings of the National Academy of Sciences, 112(22):6908–6913, 2015. + +Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning fea-´ tures by watching objects move. arXiv preprint arXiv:1612.06370, 2016. + +Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016. + +Lerrel Pinto, James Davidson, and Abhinav Gupta. Supervision via competition: Robot adversaries for learning tasks. arXiv preprint arXiv:1610.01685, 2016. + +Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature neuroscience, 2(1):79, 1999. + +Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015. + +Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211–252, 2015. + +Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1134–1141. IEEE, 2018. + +Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014. + +Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. CoRR, abs/1409.4842, 2014. URL http://arxiv.org/abs/1409.4842. + +Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. arXiv preprint arXiv:1503.01817, 2015. + +Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019. + +Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. + +Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015. + +Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via nonparametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3733–3742, 2018. + +Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised Data Augmentation. arXiv e-prints, art. arXiv:1904.12848, Apr 2019. + +Amir R Zamir, Tilman Wekel, Pulkit Agrawal, Colin Wei, Jitendra Malik, and Silvio Savarese. Generic 3D representation via pose estimation and matching. In ECCV, 2016. + +Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. $S ^ { 4 } L$ : Self-supervised semisupervised learning. arXiv preprint arXiv:1905.03670, 2019. + +Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pp. 649–666. Springer, 2016. + +Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. In Technical Report CMU-CALD-02-107, Carnegie Mellon University, 2002. + +Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. arXiv preprint arXiv:1903.12355, 2019. + +# A APPENDIX + +# A.1 ADDITIONAL RESULTS + +Table 4: Data efficient classification results with Top-1 accuracy. + +
Labeled data Method1%2%5%10% Top-1 accuracy20%50%100%
Supervised trained on pixels23.134.850.260.067.674.478.0
ReseNet trained on CPC v2 (fine-tuned)46.354.764.169.773.077.380.6
+ +Table 5: Data efficient classification results with Top-5 accuracy (data in Fig. 1). + +
Labeled data Method1%2%5%10% Top-5 accuracy20%50%100%
Supervised trained on pixels44.159.975.282.187.991.893.8
ReseNet trained on CPC v2 (fine-tuned)72.979.686.189.591.493.695.2
+ +# A.2 INFONCE IMPLEMENTATION + +For completeness, we provide pseudo-code for the main calculations involved in the InfoNCE objective, loosely modeled after Tensorflow operations. We suppose we have just calculated a set of latents $z _ { i , j } = f _ { \theta } ( \pmb { x } _ { i , j } )$ for $i , j \in \{ 1 , . . . , 7 \}$ , each one being e.g. a 4096-dimensional vector. Assuming we do so for a batch of $B$ images $\{ { \pmb x } \}$ , the set of latents is a tensor of size $B \times 7 \times 7 \times 4 0 9 6$ . + +def pixelCNN(latents): # latents: [B, H, W, D] cres $=$ latents cres_dim $=$ cres.shape[-1] for in range(5): c $=$ Conv2D(output_channels $= 2 5 6$ , kernel_shape $=$ (1, 1))(cres) c = ReLU(c) c = Conv2D(output_channels $= 2 5 6$ , kernel_shape $=$ (1, 3))(c) c = Pad(c, [[0, 0], [1, 0], [0, 0], [0, 0]]) c = Conv2D(output_channels $= 2 5 6$ , kernel_shape $=$ (2, 1), type $= ^ { \parallel }$ VALID')(c) c $=$ ReLU(c) c $=$ Conv2D(output_channels $=$ cres_dim, kernel_shape $=$ (1, 1))(c) cres $=$ cres + c cres $=$ ReLU(cres) return cres + +# A.3 LINEAR CLASSIFICATION + +• Model architecture: Having extracted $8 0 \mathrm { x } 8 0$ patches with a stride of $3 2 \mathrm { x } 3 2 $ from a $2 4 0 \times 2 4 0$ shaped input image, we end up with a grid of 6x6 features (each of which is obtained from our ResNet-161 architecture). This gives us a [6,6,4096] tensor for the image. We then use a Batch-Normalization layer to normalize the features (without scale parameter) followed by a 1x1 convolution mapping each feature in the grid to the 1000 logits for ImageNet classification. We then spatially-mean-pool these logits to end up with the final log probabilities for the linear classification. +• We use the Inception preprocessing (Szegedy et al., 2014) to extract $2 4 0 \times 2 4 0$ crops from the raw image. The image is divided into subcrops as per CPC data-preprocessing used for CPC pre-training. +• Optimization details: We use Adam Optimizer with a learning rate of 5e-4. We train the model on a batch size of 512 images with 32 images per core spread over 16 workers. + +A.4 EFFICIENT CLASSIFICATION: PURELY SUPERVISED + +In order to find the best model within this class, we vary the following hyperparameters: + +• Model architecture: We investigate using ResNet-50, ResNet-101, and ResNet-152 model architectures, all of them using the $\mathbf { \hat { v } } _ { \mathbf { \Delta } } 2 \mathbf { \hat { \mu } }$ variant (He et al., 2016b), and find larger architecture to perform better, even when given smaller amounts of data. We insert a DropOut layer before the final linear classification layer (Srivastava et al., 2014). • Data pre-processing: We use the Inception pre-processing pipeline (Szegedy et al., 2014). • Optimization details: We vary the learning rate in $\{ 0 . 0 5 , 0 . 1 , 0 . 2 \}$ , the weight decay logarithmically from $1 0 ^ { - 5 }$ to $1 0 ^ { - 2 }$ , the DropOut linearly from 0 to 1, and the batch size per worker in $\{ 1 6 , 3 2 \}$ . + +We chose the best performing model for each training subset $\mathbb { D } _ { l }$ of labeled ImageNet (using a separate validation set), and report its accuracy on the test set (i.e. the publicly available ILSVRC-2012 validation set). \ No newline at end of file diff --git a/parse/train/rJerHlrYwH/rJerHlrYwH_content_list.json b/parse/train/rJerHlrYwH/rJerHlrYwH_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ca5ef137f102419892bab9a2854d8534e8ed2af --- /dev/null +++ b/parse/train/rJerHlrYwH/rJerHlrYwH_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "DATA-EFFICIENT IMAGE RECOGNITION WITH CONTRASTIVE PREDICTIVE CODING ", + "text_level": 1, + "bbox": [ + 174, + 98, + 684, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 170, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 234, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human observers can learn to recognize new categories of objects from a handful of examples, yet doing so with machine perception remains an open challenge. We hypothesize that data-efficient recognition is enabled by representations which make the variability in natural signals more predictable, as suggested by recent perceptual evidence. We therefore revisit and improve Contrastive Predictive Coding, a recently-proposed unsupervised learning framework, and arrive at a representation which enables generalization from small amounts of labeled data. When provided with only $1 \\%$ of ImageNet labels (i.e. 13 per class), this model retains a strong classification performance, $73 \\%$ Top-5 accuracy, outperforming supervised networks by $28 \\%$ (a $65 \\%$ relative improvement) and state-of-the-art semisupervised methods by $14 \\%$ . We also find this representation to serve as a useful substrate for object detection on the PASCAL-VOC 2007 dataset, approaching the performance of representations trained with a fully annotated ImageNet dataset. ", + "bbox": [ + 233, + 267, + 764, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 178, + 476, + 334, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks excel at perceptual tasks when labeled data are abundant, yet their performance degrades substantially when provided with limited supervision (Fig. 1, red). In contrast, humans and animals can quickly learn about new classes of objects from few examples (Landau et al., 1988; Markman, 1989). What accounts for this monumental difference in data-efficiency between biological and machine vision? While highly-structured representations (e.g. as proposed by Lake et al., 2015) may improve data-efficiency, it remains unclear how to program explicit structures that capture the enormous complexity of real visual scenes like those in ImageNet (Russakovsky et al., 2015). An alternative hypothesis has proposed that intelligent systems need not be structured a priori, but can instead learn about the structure of the world in an unsupervised manner (Barlow, 1989; Hinton et al., 1999; LeCun et al., 2015). Choosing an appropriate training objective is an open problem, but a promising guiding principle has emerged recently: good representations should make the spatio-temporal variability in natural signals more predictable. Indeed, human perceptual representations have been shown to linearize ", + "bbox": [ + 174, + 510, + 483, + 881 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b8ed316aeaae9bdfe3d3d7e12974b01335c33b7eb790b24e724753c0422fb2e.jpg", + "image_caption": [ + "Figure 1: Data-efficient image recognition with Contrastive Predictive Coding. With decreasing amounts of labeled data, supervised networks trained on pixels fail to generalize (red). When trained on unsupervised representations learned with CPC, these networks retain a much higher accuracy in this low-data regime (blue). Equivalently, the accuracy of supervised networks can be matched with significantly fewer labels. " + ], + "image_footnote": [], + "bbox": [ + 500, + 512, + 820, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(or ‘straighten’) the temporal transformations found in natural videos, a property lacking from current supervised image recognition models (Henaff et al., 2019), and theories of both spatial and tem- ´ poral predictability have succeeded in describing properties of early visual areas (Rao & Ballard, ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1999; Palmer et al., 2015). In this work, we hypothesize that spatially predictable representations may allow artificial systems to benefit from human-like data-efficiency. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Predictive Coding (CPC, van den Oord et al., 2018) is an unsupervised objective which learns such predictable representations. CPC is a general technique that only requires in its definition that observations be ordered along e.g. temporal or spatial dimensions, and as such has been applied to a variety of different modalities including speech, natural language and images. This generality, combined with the strong performance of its representations in downstream linear classification tasks, makes CPC a promising candidate for investigating the efficacy of predictable representations for data-efficient image recognition. ", + "bbox": [ + 174, + 138, + 825, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work makes the following contributions: ", + "bbox": [ + 174, + 243, + 468, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• We revisit CPC in terms of its architecture and training methodology, and arrive at a new implementation of CPC with dramatically-improved ability to linearly separate image classes $+ 1 7 \\%$ Top-1 ImageNet classification accuracy). \nWe then train deep networks on top of the resulting CPC representations using very few labeled images (e.g. $1 \\%$ of the ImageNet dataset), and demonstrate test-time classification accuracy far above networks trained on raw pixels ( $73 \\%$ Top-5 accuracy, a $28 \\%$ absolute improvement), outperforming all other unsupervised representation learning methods $( + 1 5 \\%$ Top-5 accuracy over the previous state-of-the-art (Zhai et al., 2019)). Surprisingly, this representation also surpasses supervised methods when given the entire ImageNet dataset $+ 1 \\%$ Top-5 accuracy). \nWe isolate the contributions of different components of the final model to such downstream tasks. Interestingly, we find that linear classification accuracy is not always predictive of low-data classification accuracy, emphasizing the importance of this metric as a stand-alone benchmark for unsupervised learning. \n• Finally, we assess the generality of CPC representations by transferring them to a new task and dataset: object detection on PASCAL-VOC 2007. Consistent with the results from the previous section, we find CPC to give state-of-the-art performance in this setting. ", + "bbox": [ + 217, + 270, + 825, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 174, + 541, + 398, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first review the CPC architecture and learning objective in section 2.1, before detailing how we use its resulting representations for image recognition tasks in section 2.2. ", + "bbox": [ + 176, + 574, + 825, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 CONTRASTIVE PREDICTIVE CODING ", + "text_level": 1, + "bbox": [ + 176, + 618, + 467, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Predictive Coding as formulated in (van den Oord et al., 2018) learns representations by training neural networks to predict the representations of future observations from those of past ones. When applied to images, the original formulation of CPC operates by predicting the representations of patches below a certain position from those above it (Fig. 2, left). These predictions are evaluated using a contrastive loss, in which the network must correctly classify the ‘future’ representation amongst a set of unrelated ‘negative’ representations. This avoids trivial solutions such as representing all patches with a constant vector, as would be the case with a mean squared error loss. ", + "bbox": [ + 174, + 645, + 825, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the CPC architecture, each input image is first divided into a set of overlapping patches $\\mathbf { \\delta } _ { \\mathbf { x } _ { i , j } }$ , each of which is encoded with a neural network $f _ { \\theta }$ into a single vector $z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )$ . To make predictions, a masked convolutional network $g _ { \\phi }$ is then applied to the grid of feature vectors. The masks are such that the receptive field of each resulting context vector $c _ { i , j }$ only includes feature vectors that lie above it in the image (i.e. $\\{ z _ { u , v } \\} _ { u \\leq i , v } )$ . The prediction task then consists o predicting ‘future’ feature vectors $z _ { i + k , j }$ from current context vectors $c _ { i , j }$ , where $k > 0$ . The predictions are made linearly: given a context vector $c _ { i , j }$ , a prediction length $k > 0$ , and a prediction matrix $W _ { k }$ , the predicted feature vector is $\\hat { \\boldsymbol { z } } _ { i + k , j } = \\boldsymbol { W } _ { k } \\boldsymbol { c } _ { i , j }$ . ", + "bbox": [ + 174, + 750, + 825, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The quality of this prediction is then evaluated using a contrastive loss. Specifically, the goal is to correctly recognize the target $z _ { i + k , j }$ among a set of randomly sampled feature vectors $\\{ z _ { l } \\}$ from the dataset. We compute the probability assigned to the target using a softmax, and evaluate this probability using the usual cross-entropy loss. Summing this loss over locations and prediction offsets, we arrive at the CPC objective as defined in (van den Oord et al., 2018): ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e21408d1a55db66ba2cac6a3f4d98b0d1a13c2f1211e3956f93d67bae89bbaae.jpg", + "image_caption": [ + "Figure 2: Overview of the framework for semi-supervised learning with Contrastive Predictive Coding. Left: unsupervised pre-training with the spatial prediction task (See Section 2.1). First, an image is divided into a grid of overlapping patches. Each patch is encoded independently from the rest with a feature extractor (blue) which terminates with a mean-pooling operation, yielding a single feature vector for that patch. Doing so for all patches yields a field of such feature vectors (wireframe vectors). Feature vectors above a certain level (in this case, the center of the image) are then aggregated with a context network (red), yielding a row of context vectors which are used to linearly predict features vectors below. Right: using the CPC representation for a classification task. Having trained the encoder network, the context network (red) is discarded and replaced by a classifier network (green) which can be trained in a supervised manner. For some experiments, we also fine-tune the encoder network (blue) for the classification task. When applying the encoder to cropped patches (as opposed to the full image) we refer to it as a patched ResNet in the figure. " + ], + "image_footnote": [], + "bbox": [ + 174, + 104, + 823, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 486, + 696, + 502 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/7113b15debaadf2fc32b7ca0c3e55fb0689eecd2935810917cee824de20f6604.jpg", + "text": "$$\n\\mathcal { L } _ { \\mathrm { C P C } } = - \\sum _ { i , j , k } \\log p ( z _ { i + k , j } | \\hat { z } _ { i + k , j } , \\{ z _ { l } \\} ) = - \\sum _ { i , j , k } \\log \\frac { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) } { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) + \\sum _ { l } \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { l } ) }\n$$", + "text_format": "latex", + "bbox": [ + 186, + 503, + 808, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The negative samples $\\left\\{ z _ { l } \\right\\}$ are taken from other locations in the image and other images in the minibatch. This loss is called InfoNCE (van den Oord et al., 2018) as it is inspired by Noise-Contrastive Estimation (Gutmann $\\&$ Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013) and has been shown to ¨ maximize the mutual information between $\\mathbf { } _ { c _ { i , j } }$ and $z _ { i + k , j }$ (van den Oord et al., 2018). ", + "bbox": [ + 174, + 547, + 825, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 EVALUATION PROTOCOL ", + "text_level": 1, + "bbox": [ + 174, + 621, + 383, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Having trained an encoder network $f _ { \\theta }$ , a context network $g _ { \\phi }$ , and a set of linear predictors $\\{ W _ { k } \\}$ using the CPC objective, we use the latents $z = f _ { \\boldsymbol { \\theta } } ( \\pmb { x } )$ as a representation of new observations $x$ for downstream tasks, and discard the rest. We then train a model $h _ { \\psi }$ to classify these representations given a dataset of labeled images. More formally, given a dataset of $N$ unlabeled images $\\mathbb { D } _ { u } =$ $\\left\\{ x _ { n } \\right\\}$ , and a (potentially much smaller) dataset of $M$ labeled images $\\mathbb { D } _ { l } = \\{ x _ { m } , y _ { m } \\}$ : ", + "bbox": [ + 174, + 646, + 825, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/c6df6158bc3640c41b4e80de8a77924da143cb3d171b2d5421986f303ad176ed.jpg", + "text": "$$\n\\theta ^ { * } = \\arg \\operatorname* { m i n } _ { \\theta } \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } \\mathcal { L } _ { \\mathrm { C P C } } [ f _ { \\theta } ( x _ { n } ) ] , \\quad \\psi ^ { * } = \\arg \\operatorname* { m i n } _ { \\psi } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { L } _ { \\mathrm { S u p } } [ h _ { \\psi } \\circ f _ { \\theta ^ { * } } ( x _ { m } ) , y _ { m } ]\n$$", + "text_format": "latex", + "bbox": [ + 217, + 718, + 781, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In all cases, the dataset of unlabeled images $\\mathbb { D } _ { u }$ we pre-train on is the full ImageNet ILSVRC 2012 training set (Russakovsky et al., 2015). We consider three labeled datasets $\\mathbb { D } _ { l }$ for evaluation, each with an associated classifier $h _ { \\psi }$ and supervised losse $\\mathcal { L } _ { \\mathrm { { S u p } } }$ (see Fig. 2, right). This protocol is sufficiently generic to allow us to later compare the CPC representation to other methods which have their own means of learning a feature extractor $f _ { \\theta }$ . ", + "bbox": [ + 173, + 763, + 825, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Linear classification is the standard benchmark for evaluating the quality of unsupervised image representations. In this regime, the classification network $h _ { \\psi }$ is restricted to mean pooling followed by a single linear layer, and the parameters of $f _ { \\theta }$ are kept fixed. The labeled dataset $\\mathbb { D } _ { l }$ is the entire ImageNet dataset, and the supervised loss $\\mathcal { L } _ { \\mathrm { { S u p } } }$ is standard cross-entropy. We use the same dataaugmentation as in the unsupervised learning phase for training, and none at test time and evaluate with a single crop. ", + "bbox": [ + 173, + 839, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient classification directly tests whether the CPC representation enables visual learning from few labels. For this task, the classifier $h _ { \\psi }$ is an arbitrary deep neural network (we use an 11-block ResNet architecture with 4096-dimensional feature maps and 1024-dimensional bottleneck layers). The labeled dataset $\\mathbb { D } _ { l }$ is a subset of the ImageNet dataset: we investigated using $1 \\%$ , $2 \\%$ , $5 \\%$ , $10 \\%$ , $20 \\%$ , $50 \\%$ and $100 \\%$ of the ImageNet dataset. The supervised loss $\\mathcal { L } _ { \\mathrm { { S u p } } }$ is again cross-entropy. In addition to random color-dropping we use the Inception data-augmentation scheme (Szegedy et al., 2014) for training, no augmentation at test-time and evaluate with a single crop. ", + "bbox": [ + 174, + 103, + 825, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transfer learning tests the generality of the representation by applying it to a new task and dataset. For this we chose image detection on the PASCAL-2007 dataset, a standard benchmark in computer vision (Everingham et al., 2007). As such $\\mathbb { D } _ { l }$ is the entire PASCAL-2007 dataset (comprised of 5011 labeled images); $h _ { \\psi }$ and $\\mathcal { L } _ { \\mathrm { { S u p } } }$ are the Faster-RCNN architecture and loss (Ren et al., 2015). In addition to color-dropping, we use scale-augmentation (Doersch et al., 2015) for training. ", + "bbox": [ + 174, + 208, + 825, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For linear classification, we keep the feature extractor $f _ { \\theta }$ fixed to assess the representation in absolute terms. For efficient classification and transfer learning, we additionally explore fine-tuning the feature extractor for the supervised objective. In this regime, we initialize the feature extractor and classifier with the solutions $\\theta ^ { * } , \\psi ^ { * }$ found in the previous learning phase, and train them both for the supervised objective. To ensure that the feature extractor does not deviate too much from the solution dictated by the CPC objective, we use a smaller learning rate and early-stopping. ", + "bbox": [ + 174, + 285, + 825, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 402, + 343, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data-efficient learning has typically been approached by two complementary methods, both of which seek to make use of more plentiful unlabeled data: representation learning and semisupervised learning. The former formulates an objective to learn a feature extractor $f _ { \\theta }$ in an unsupervised manner, whereas the latter directly constrains the classifier $h _ { \\psi }$ using the unlabeled data. ", + "bbox": [ + 174, + 444, + 823, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Representation learning saw early success using generative modeling (Kingma et al., 2014), but likelihood-based models have yet to generalize to more complex stimulus classes. Generative adversarial models have also been harnessed for representation learning (Donahue et al., 2016), and large-scale implementations have recently achieved corresponding gains in linear classification accuracy (Donahue & Simonyan, 2019). ", + "bbox": [ + 174, + 507, + 823, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to generative models which require the reconstruction of observations, self-supervised techniques directly formulate tasks involving the learned representation. For example, simply asking a network to recognize the spatial layout of an image led to representations that transferred to popular vision tasks such as classification and detection (Doersch et al., 2015; Noroozi & Favaro, 2016). Other works showed that prediction of color (Zhang et al., 2016; Larsson et al., 2017) and image orientation (Gidaris et al., 2018), and invariance to data augmentation (Dosovitskiy et al., 2014) can provide useful self-supervised tasks. Beyond single images, works have leveraged video cues such as object tracking (Wang & Gupta, 2015), frame ordering (Misra et al., 2016), and object boundary cues (Li et al., 2016; Pathak et al., 2016). Non-visual information can be equally powerful; information about camera motion (Agrawal et al., 2015; Jayaraman & Grauman, 2015), scene geometry (Zamir et al., 2016), or sound (Arandjelovic & Zisserman, 2017; 2018) can all serve as natural sources of supervision. ", + "bbox": [ + 174, + 583, + 825, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While many of these tasks require predicting fixed quantities computed from the data, another class of contrastive methods formulate their objectives in the learned representations themselves. CPC is a contrastive representation learning method that maximizes the mutual information between spatially removed latent representations with InfoNCE (van den Oord et al., 2018), a loss function based on Noise-Contrastive Estimation (Gutmann & Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013). Two ¨ other methods have recently been proposed using the same loss function, but with different associated prediction tasks. Contrastive Multiview Coding (Tian et al., 2019) maximizes the mutual information between representations of different views of the same observation. Augmented Multiscale Deep InfoMax (AMDIM, Bachman et al., 2019) is most similar to CPC in that it makes predictions across space, but differs in that it also predicts representations across layers in the model. In addition, AMDIM limits the receptive field of its representation, but does this by constraining the number of spatial convolutions in the network architecture rather than using image patches. ", + "bbox": [ + 174, + 757, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A common alternative approach for improving data efficiency is label-propagation (Zhu & Ghahramani, 2002), where a classifier is trained on a subset of labeled data, then used to label parts of the unlabeled dataset, after which the process is repeated. This label-propagation can either be discrete (as in pseudo-labeling, Lee, 2013) or continuous (as in entropy minimization, Grandvalet & Bengio, 2005). The predictions of this classifier are often constrained to be smooth with respect to certain deformations, such as data-augmentation (Xie et al., 2019) or adversarial perturbation (Miyato et al., 2018). Representation learning and semi-supervised learning have been shown to be complementary and can be combined to great effect (Zhai et al., 2019), which is why we focus solely on representation learning in this paper. ", + "bbox": [ + 174, + 103, + 825, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 251, + 281, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When asking whether CPC enables data-efficient learning, we wish to use the best possible representative of this model class. Unfortunately, purely unsupervised metrics tell us little about downstream performance, and implementation details have been shown to matter enormously (Doersch & Zisserman, 2017; Kolesnikov et al., 2019). Since many design choices (e.g. network architecture and datapreprocessing) have been previously evaluated using linear classification, we use this benchmark in section 4.1 to align the CPC model with best practices in representation learning and compare to published results. In section 4.2 we select the best performing model from the previous section and assess whether it enables efficient classification. We also investigate to what extent the first, more common metric (linear classification accuracy) is predictive of efficient classification. Finally, in section 4.3 we investigate the generality of our results through transfer learning to PASCAL-2007. ", + "bbox": [ + 174, + 284, + 825, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 FROM CPC V1 TO CPC V2 ", + "text_level": 1, + "bbox": [ + 176, + 441, + 400, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overarching principle behind our new model design is to increase the scale and efficiency of the encoder architecture while also maximizing the supervisory signal we obtain from each image. At the same time, it is important not to allow the network to solve the problem trivially, i.e., without learning semantics. To this end, we seek to remove low-level cues common across patches by augmenting individual patches independently, using standard stochastic data-processing techniques from supervised and self-supervised learning. ", + "bbox": [ + 174, + 468, + 825, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We identify four axes for model capacity and task setup that could impact the model’s performance. The first axis increases model capacity by increasing depth and width, while the second improves training efficiency capacity by introducing layer normalization. The third axis increases task complexity by making predictions in all four directions, and the fourth does so by performing more extensive patch-based augmentation. ", + "bbox": [ + 174, + 559, + 485, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model capacity. Recent work has shown that networks and more effective training improves self-supervised learning (Doersch & Zisserman, 2017; Kolesnikov et al., 2019), but the original CPC model used only the first 3 stacks of a ResNet-101 (He et al., 2016a) architecture (i.e. a ResNet-92). Therefore, we converted the third residual stack of ResNet-101 (originally containing 23 blocks, 1024-dimensional feature maps, and 256-dimensional bottleneck layers), to use 46 blocks, with 4096-dimensional feature maps and 512-dimensional bottleneck layers. We call the resulting network ResNet-161. Consistent with prior results, this new architecture delivers better performance regardless of ", + "bbox": [ + 174, + 715, + 483, + 922 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3eeded7c7663fc17771caa90292cc1343662124f1094d06f22d9ca255036f2a3.jpg", + "image_caption": [ + "Figure 3: Linear classification performance of new variants of CPC, which incrementally add a series of modifications. BU: bottum up spatial predictions. HF: randomly flipping patches horizontally. LN: layer normalization. RC: random color-dropping. TL: tuned prediction lengths. HP: horizontal spatial predictions. We use color to indicate the number of spatial predictions used (orange, green, blue for 1, 2 and 4 directions). " + ], + "image_footnote": [], + "bbox": [ + 501, + 566, + 818, + 777 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fa1f65edaf71df9800daca9c26a7999e93d74856116a87cc357e08cc713a4be7.jpg", + "table_caption": [ + "Table 1: Linear classifier $h _ { \\psi }$ trained with $100 \\%$ of labels. Comparison to linear separability of other self-supervised methods. In all cases a feature extractor is optimized in an unsupervised manner, and a linear classifier is trained on top using all labels in the ImageNet dataset. " + ], + "table_footnote": [], + "table_body": "
MethodTop-1Top-5
Motion Segmentation (MS) (Pathak et al., 2016) Exemplar (Ex) (Dosovitskiy et al.,2014) Relative Position (RP) (Doersch et al.,2015) Colorization (Col) (Zhang et al., 2016) Combination of MS + Ex + RP+ Col (Doersch & Zisserman,2017) CPC v1 (van den Oord et al., 2018) Rotation (Kolesnikov et al., 2019)27.6 31.5 36.2 39.6 1 48.7 55.448.3 53.1 59.2 62.5 69.3 73.6
CMC (Tian et al., 2019) Local Aggregation (Zhuang et al., 2019)60.1 60.282.8 1 =
BigBiGAN (Donahue & Simonyan,2019) AMDIM (Bachman et al., 2019) CPC v2 (ours) 65.961.3 68.181.9
", + "bbox": [ + 204, + 159, + 794, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "other design choices. Interestingly, a larger architecture delivers larger improvements with more efficient training, more self-supervised losses, and more patch-based augmentations (Fig. 3, $+5 \\%$ Top-1 accuracy with original training scheme, $+ 1 0 \\%$ accuracy with new one). ", + "bbox": [ + 176, + 410, + 823, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Layer normalization. Large architectures are more difficult to train efficiently. Early works on context prediction with patches used batch normalization (Ioffe & Szegedy, 2015; Doersch et al., 2015) to speed training. However, with CPC we find that batch normalization actually harms downstream performance of large models. We hypothesize that batch normalization allows large models to find a trivial solution to CPC: it introduces a dependency between patches (through the batch statistics) that can be exploited to bypass the constraints on the receptive field. We find that we can reclaim much of batch normalization’s training efficiency using layer normalization (Ba et al., 2016), which leads to a small gain for the smaller architecture $+ 1 \\%$ accuracy over equivalent architectures that use neither normalization) and a larger gain for the larger architecture $( + 2 . 5 \\%$ accuracy). ", + "bbox": [ + 174, + 474, + 825, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prediction lengths and directions. Larger architectures also run a greater risk of overfitting. We address this by asking more from the network: specifically, whereas van den Oord et al. (2018) predicted each patch using only context from spatially beneath it, we repeatedly predict the patch using context from above, to the right, and to the left, resulting in up to four times as many prediction tasks. Combining top-to-bottom with bottom-to-top helps both model architectures $42 \\%$ accuracy for both), but using all 4 spatial directions only benefits the larger model (an additional $+ 1 . 5 \\%$ for the larger model, ${ \\bf - 1 \\% }$ for the smaller), consistent with the idea that model capacity and amount of supervision must go hand-in-hand. We also hypothesized that prediction “length”—i.e. offset between the predicted patch and the aggregated context—might affect performance, as distant patches might lie on distinct objects, encouraging the network to memorize images. Indeed, limiting the range of the prediction length $k$ to $\\{ 2 , 3 \\}$ performed better than $\\{ 2 , \\ldots , 5 \\}$ as was used originally $+ 1 \\%$ for the larger model). ", + "bbox": [ + 173, + 622, + 825, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patch-based augmentation. If the network can solve CPC using low-level patterns (e.g. straight lines continuing between patches, chromatic aberration), it need not learn semantically meaningful content. Augmenting the low-level variability across patches can remove such low level cues. The original CPC model spatially jitters individual patches independently. We further this logic by adopting the ‘color dropping’ method of Doersch et al. (2015), which randomly drops two of the three color channels in each patch, and find it to delivers systematic gains $( + 1 \\%$ for the small model, $+3 \\%$ for the larger one). We also randomly flip patches horizontally, but find it only benefits the smaller model $( + 1 \\% )$ . ", + "bbox": [ + 174, + 811, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Combined. Cumulatively, these fairly straightforward implementation changes lead to a substantial improvement to the original CPC model $6 5 . 9 \\%$ Top-1 accuracy, a $17 \\%$ improvement), making it competitive with recent approaches and outperforming prior methods (see table 1). Interestingly, if we train the same patch-based architecture from scratch in a fully supervised manner, we obtain $6 6 . 4 \\%$ Top-1 accuracy (with batch normalization; $6 2 . 5 \\%$ without), suggesting that CPC is now nearly saturating the architecture’s representational power despite not using labels. These results illustrate how architecture and data have an outsized impact on the linear classification performance of self-supervised representations, and are interesting to compare with with previous results. For example, in AMDIM, different settings of data augmentation alone can result in a nearly $10 \\%$ absolute increase in performance on ImageNet linear classification. ", + "bbox": [ + 173, + 103, + 825, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 EFFICIENT IMAGE CLASSIFICATION ", + "text_level": 1, + "bbox": [ + 178, + 262, + 455, + 276 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "toWe now turn to our original question of whether CPC can enable data-efficient image recognition. We start by evaluating the performance of purely-supervised networks as the size of the labeled dataset $\\mathbb { D } _ { l }$ varies from $1 \\%$ to $100 \\%$ of ImageNet, training separate classifiers on each subset. We found that a ResNet-152 to works best across all data-regimes (see Appendix). Despite our efforts to tune the supervised model for low-data classification (including network depth, regularization, and optimization parameters), the accuracy of the best model only reaches $4 4 . 1 \\%$ Top-5 accuracy when trained on $1 \\%$ of the dataset (compared to $9 3 . 9 \\%$ when trained on the entire dataset, see Fig. 1, red). ", + "bbox": [ + 174, + 287, + 825, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contrastive Predictive Coding. We now address our central question of whether CPC enables data-efficient learning. We follow the same paradigm as for the supervised baseline (training and evaluating a separate classifier for each size subset), stacking a neural network classifier on top of the CPC latents $\\begin{array} { r l r } { z } & { { } = } & { f _ { \\theta } ( { \\pmb x } ) } \\end{array}$ rather than the raw image pixels $_ { \\textbf { \\em x } }$ (see section 2.2, efficient classification, and Appendix). This representation, which we selected for its improved linear classification performance (CPC v2 in Fig. 3), leads to a significant increase in data-efficiency compared to purely supervised networks (Fig. 1, blue curve). This classifier yields $7 2 . 9 \\%$ Top-5 accuracy with only $1 \\%$ of the labels, a $2 9 \\%$ absolute improvement $65 \\%$ relative) over purely-supervised methods. Surprisingly, when given the entire dataset, this classifier reaches $8 0 . 6 \\% / 9 5 . 2 \\%$ Top1/Top5 accuracy, surpassing our supervised baseline (ResNet-152: $7 8 . 0 \\% / 9 3 . 9 \\%$ accuracy) and published results (ResNet-200: $7 9 . 9 \\% / 9 5 . 2 \\%$ , He et al. (2016b)). We find similar results in all other data-regimes we considered (see Fig. 1). ", + "bbox": [ + 174, + 404, + 483, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How important are the model specifications described in Section 4.1 for low-data classification? We hypothesized that predictable representations might enable data-efficient classifi", + "bbox": [ + 174, + 757, + 482, + 813 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/106f2252ad377e16a329cfb21360459a72fdc1d09e4f745b26572da1cf6d5d14.jpg", + "image_caption": [ + "Figure 4: Relationship between linear classification accuracy and low-data classification, for different variants of the CPC model. Left: CPC variants with the same architecture but different training protocols. Orange, green, and blue dots correspond to CPC models making predictions in 1, 2, and 4 spatial directions respectively. Within a color group, different models correspond to other implementation details (e.g. layer norm and patch augmentation, and combinations thereof). " + ], + "image_footnote": [], + "bbox": [ + 503, + 409, + 810, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cation, and therefore expect that increasing the amount of ‘predictability’ in the representation should also increase its ability to learn from small amounts of data. Fig. 4 shows evidence for this by ablating model parameters and comparing linear classification performance against low-data classification. Consistent with our hypothesis, increasing the number of spatial directions in the CPC prediction task (which increased linear classification performance) systematically increases low-data classification performance (Fig. 4, left, different color groups). As a control, we asked if all modifications that improve linear classification also improve low-data classification. We did not find evidence in favor of this: improvements in linear classification as a result of changing other model parameters (patch-based data-augmentation, layer normalization, and combinations thereof) seem uncorrelated to performance in other tasks (Fig. 4, left, within green group: $R ^ { 2 } = 0 . 1 7 , p = 0 . 3 6$ ). Different architectural specifications also produced different changes in both tasks (Fig. 4, right). Whereas increasing the depth of the encoding network greatly improves both metrics, increasing the network width (and therefore the number of features used for linear classification) only improves linear classification accuracy. ", + "bbox": [ + 174, + 813, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e969618e28fc2ffa8c76067cea0310a604dd1bb350cd5a5534dd7265b81333a6.jpg", + "table_caption": [ + "Table 2: ResNet classifier $h _ { \\psi }$ trained with $1 \\%$ or $10 \\%$ of labels. Comparison to other methods for semi-supervised learning using $1 \\%$ or $10 \\%$ of labeled data. Representation learning methods use a classifier to discriminate an unsupervised representation, and optimize it solely for the supervised objective on labeled data. Label-propagation methods on the other hand further constrain the classifier with smoothness and entropy criteria on unlabeled data, making the additional assumption that all training images fit into a single (unknown) testing category. $\\dagger$ denotes methods implemented in this work, fixed and fine-tuned denote whether the feature extractor is allowed to accommodate the supervised objective. " + ], + "table_footnote": [], + "table_body": "
Labeled data 1% 10% 100% Method Top-5 accuracy
Supervised baseline 44.1 82.1 93.9 Methods using label-propagation:
Pseudolabeling (Zhai et al., 2019) 51.6 82.4 VAT + Entropy Minimization (Zhai et al., 2019) 47.0 83.4 Unsup. Data Augmentation (Xie et al., 2019) 88.5
Rotation+ VAT+Ent.Min. (Zhai et al.,2019) 91.2 95.0 =
Methods using representation learning only:
Instance Discrimination (Wu et al., 2018) 39.2 77.4
Rotation (Zhai et al., 2019) 57.5 86.4
+ResNet trained on BigBiGAN (fixed) 55.2 78.8 87.0
+ResNet trained on AMDIM (fixed) 67.4 85.8 92.2
+ResNet trained on CPC v2 (fixed) 72.3 89.1 94.4 +ResNet trained on CPC v2 (fine-tuned) 72.9 89.5 95.2
", + "bbox": [ + 250, + 228, + 745, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 531, + 825, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other unsupervised representations. How well does the CPC representation compare to other representations that have been learned in an unsupervised manner? If predictable representations are uniquely suited for efficient classification, we would expect other methods within this family to perform similarly, and other model classes less so. Table 2 compares our best model with other works on efficient recognition. We consider three objectives from different model classes: self-supervised learning with rotation prediction (Zhai et al., 2019), large-scale adversarial feature learning (BigBiGAN, Donahue & Simonyan, 2019), and another contrastive prediction objective (AMDIM, Bachman et al., 2019). Zhai et al. (2019) evaluate the low-data classification performance of representations learned with rotation prediction using a similar paradigm and architecture (ResNet-152), hence we report their results directly. Given $1 \\%$ of ImageNet, their method achieves $5 7 . 5 \\%$ Top-5 accuracy, consistently with the reduced accuracy of a linear classifier ( $5 5 . 4 \\%$ vs $6 5 . 9 \\%$ for CPC)1. ", + "bbox": [ + 173, + 631, + 825, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Because BigBiGAN and AMDIM achieve stronger linear classification accuracy than rotation prediction $6 1 . 3 \\%$ and $6 8 . 1 \\%$ Top-1 accuracy, respectively), we might expect better performance on efficient classification as well. Since their authors do not report results on efficient classification we evaluated these representations using the same paradigm we used for evaluating CPC, stacking a ResNet classifier on top of the $7 \\times 7 \\times 8 1 9 2$ latents of the BigBiGAN and the $7 \\times 7 \\times 2 5 6 0$ grid of feature vectors of AMDIM. We found fine-tuned representations to yield only marginal gains over fixed ones $7 2 . 9 \\%$ compared to $7 2 . 3 \\%$ Top-5 accuracy given $1 \\%$ of labels), hence for simplicity we evaluate BigBiGAN and AMDIM on this task while keeping them fixed. We re-tune the hyper-parameters of the classifier (including optimization, regularization, etc.) for each of these representations separately. Although these methods achieve similar performance in terms of linear classification, we find them to achieve very different results in efficient classification. Given $1 \\%$ of ImageNet, classifiers trained on top of BigBiGAN achieve $5 5 . 2 \\%$ Top-5 accuracy, similarly to rotation prediction $( 5 7 . 5 \\% )$ , despite its increased linear classification accuracy $( + 6 \\%$ relative to rotation prediction). In contrast, AMDIM (which also belongs to the family of contrastive prediction methods) achieves $6 7 . 4 \\%$ on this same task. Again, its increased linear classification accuracy did not entail an increase in data-efficiency. Nevertheless, in line with our initial hypothesis, we find that contrastive prediction methods such as CPC surpass other approaches in our efficient classification experiments, and that linear classification performance is not perfectly correlated with these results. ", + "bbox": [ + 173, + 803, + 825, + 888 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 270 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Other semi-supervised techniques A separate class of methods for low-data classification attempts to propagate the knowledge extracted from the subset of labeled examples to unlabeled examples while being invariant to augmentation or other perturbations. These methods generally depend on the quality of the classifier’s predictions, and as such tend to fare well when given intermediate amounts of data. Although not sufficient in themselves (Unsupervised Data Augmentation (Xie et al., 2019), Virtual Adversarial Training (Miyato et al., 2018) and entropy minimization (Grandvalet & Bengio, 2005), and pseudo-labeling (Lee, 2013) achieve $8 5 . 8 \\%$ , $8 3 . 4 \\%$ , and $8 2 . 4 \\%$ Top-5 accuracy with $10 \\%$ of labels, compared to our $8 9 . 4 \\%$ ) when combined with representation learning (e.g. rotation prediction Zhai et al., 2019) they can provide considerable gains $9 1 . 2 \\%$ Top5 accuracy). It is therefore surprising that CPC representations alone can enable accuracy that is comparable to that of these methods, and investigating to what extent they can be combined would be an interesting topic of future work. ", + "bbox": [ + 174, + 290, + 823, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.3 TRANSFER LEARNING: IMAGE DETECTION ON PASCAL VOC 2007 ", + "text_level": 1, + "bbox": [ + 176, + 477, + 681, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We next investigate transfer performance on object detection on the PASCAL-2007 dataset, which reflects the practical scenario where a representation must be trained on a dataset with different statistics than the dataset of interest. This dataset also tests the efficiency of the representation as it only contains 5011 labeled images to train from. In this setting, we replaced the neural network classifier $h _ { \\psi }$ used previously with a Faster-RCNN (Ren et al., 2015) image detection architecture, and use the pre-trained feature extractor on ImageNet. As before, we first trained the Faster-RCNN model while keeping the feature extractor fixed, then fine-tuned the entire model end-to-end. Table 3 displays our results compared to other methods. Most competing methods, which optimize a single unsupervised objective on ImageNet before fine-tuning on PASCAL detection, attain around $65 \\%$ mean average precision. Leveraging larger unlabeled datasets increases their performance up to $6 7 . 8 \\%$ (Caron et al., 2019). Combining multiple forms of self-supervision enables them to reach $7 0 . 5 \\%$ (Doersch & Zisserman, 2017). The proposed method, which learns only from ImageNet data using a single unsupervised objective, reaches $7 0 . 6 \\%$ when equipped with a ResNet-101 feature extractor $f _ { \\theta }$ (as for most competing methods (Doersch & Zisserman, 2017) but not all (Caron et al., 2018; 2019). Equipped with the more powerful ResNet-161 feature extractor $f _ { \\theta }$ , our method reaches $7 2 . 7 \\%$ . Importantly, this result is only $2 \\%$ short of the performance attained by purely supervised transfer learning, which we obtain by using all ImageNet labels before transferring to PASCAL. ", + "bbox": [ + 174, + 505, + 825, + 741 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 DISCUSSION ", + "text_level": 1, + "bbox": [ + 176, + 765, + 310, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We asked whether CPC could enable data-efficient image recognition, and found that it indeed greatly improves the accuracy of classifiers and object detectors when given small amounts of labeled data. Surprisingly, CPC even improves results given ImageNet-scale labels. Our results show that there is still room for improvement using relatively straightforward changes such as augmentation, optimization, and network architecture. Furthermore, we found that the standard method for evaluating unsupervised representations—linear classification—is only partially predictive of efficient recognition performance, suggesting that further research should focus on efficient recognition as a standalone benchmark. Overall, these results open the door toward research on problems where data is naturally limited, e.g. medical imaging or robotics. ", + "bbox": [ + 173, + 799, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/d005a0a972800a539ffdecef950a4659796c2a8300a3ab9b860336449e53c88e.jpg", + "table_caption": [ + "Table 3: Faster-RCNN $h _ { \\psi }$ trained with $100 \\%$ of PASCAL labels. Comparison of PASCAL 2007 image detection accuracy to other transfer methods. The supervised baseline learns from the entire labeled ImageNet dataset and fine-tunes for PASCAL detection. The second class of methods learns from the same unlabeled images before transferring. All of these methods pre-train on the ImageNet dataset, except for DeeperCluster which learns from the larger, but uncurated, YFCC100M dataset (Thomee et al., 2015). All results are reported in terms of mean average precision (mAP). $\\dagger$ denotes methods implemented in this work. " + ], + "table_footnote": [], + "table_body": "
MethodmAP
Transfer from labeled data: Supervised - ResNet-15274.7
Transfer from unlabeled data: Exemplar (Ex) (Dosovitskiy et al., 2014) Motion Segmentation (MS) (Pathak et al.,2016)60.9
Colorization (Col) (Zhang et al.,2016) Relative Position (RP) (Doersch et al.,2015)61.1 65.5
Combination of Ex + MS + Col +RP (Doersch & Zisserman,2017)66.8 70.5
Instance Discrimination (Wu et al., 2018)65.4
Deep Cluster (Caron et al., 2018) Deeper Cluster (Caron et al., 2019)65.9
", + "bbox": [ + 240, + 214, + 758, + 457 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Furthermore, images are far from the only domain where unsupervised representation learning is important: for example, unsupervised learning is already a critical step in language (Mikolov et al., 2013; Devlin et al., 2018), and shows promise in domains like audio (van den Oord et al., 2018; Arandjelovic & Zisserman, 2018; 2017), video (Jing & Tian, 2018; Misra et al., 2016), and robotic manipulation (Pinto & Gupta, 2016; Pinto et al., 2016; Sermanet et al., 2018). Currently much self-supervised work builds upon tasks tailored for a specific domain (often images), which may not be easily adapted to other domains. Contrastive prediction methods, including the techniques suggested in this paper, are task agnostic and could therefore serve as a unifying framework for integrating these tasks and modalities. This generality is particularly useful given that many realworld environments are inherently multimodal, e.g. robotic environments which can have vision, audio, touch, proprioception, action, and more over long temporal sequences. Given the importance of increasing the amounts of self-supervision (via additional directions of prediction), integrating these modalities and tasks could lead to unsupervised representations which rival the efficiency and effectiveness of biological ones. ", + "bbox": [ + 173, + 486, + 825, + 680 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 702, + 285, + 717 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. In ICCV, 2015. ", + "bbox": [ + 171, + 724, + 794, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE International Conference on Computer Vision, pp. 609–617, 2017. ", + "bbox": [ + 173, + 747, + 823, + 776 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European Conference on Computer Vision (ECCV), pp. 435–451, 2018. ", + "bbox": [ + 173, + 784, + 823, + 813 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Lei Jimmy Ba, Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. ", + "bbox": [ + 173, + 820, + 823, + 849 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing mutual information across views. arXiv preprint arXiv:1906.00910, 2019. ", + "bbox": [ + 173, + 858, + 823, + 887 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "H.B. Barlow. Unsupervised learning. Neural Computation, 1(3):295–311, 1989. doi: 10.1162/neco. 1989.1.3.295. ", + "bbox": [ + 174, + 895, + 821, + 922 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In The European Conference on Computer Vision (ECCV), September 2018. ", + "bbox": [ + 173, + 103, + 823, + 146 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mathilde Caron, Piotr Bojanowski, Julien Mairal, and Armand Joulin. Leveraging large-scale uncurated data for unsupervised pre-training of visual features. 2019. ", + "bbox": [ + 173, + 156, + 820, + 185 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018. ", + "bbox": [ + 173, + 195, + 821, + 224 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of the IEEE International Conference on Computer Vision, pp. 2051–2060, 2017. ", + "bbox": [ + 174, + 234, + 823, + 265 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1422–1430, 2015. ", + "bbox": [ + 173, + 273, + 826, + 318 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jeff Donahue and Karen Simonyan. Large scale adversarial representation learning. arXiv preprint arXiv:1907.02544, 2019. ", + "bbox": [ + 169, + 327, + 823, + 357 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jeff Donahue, Philipp Krahenb ¨ uhl, and Trevor Darrell. Adversarial feature learning. ¨ arXiv preprint arXiv:1605.09782, 2016. ", + "bbox": [ + 171, + 367, + 823, + 396 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with convolutional neural networks. In Advances in neural information processing systems, pp. 766–774, 2014. ", + "bbox": [ + 174, + 406, + 825, + 450 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge 2007 (voc2007) results. 2007. ", + "bbox": [ + 178, + 459, + 823, + 489 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. ", + "bbox": [ + 174, + 498, + 823, + 529 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in neural information processing systems, pp. 529–536, 2005. ", + "bbox": [ + 171, + 539, + 823, + 568 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨ for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pp. 297–304, 2010. ", + "bbox": [ + 176, + 577, + 825, + 621 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016a. ", + "bbox": [ + 174, + 631, + 825, + 674 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pp. 630–645. Springer, 2016b. ", + "bbox": [ + 173, + 684, + 825, + 714 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "G.E. Hinton, T.J. Sejnowski, H.H.M.I.C.N.L.T.J. Sejnowski, and T.A. Poggio. Unsupervised Learning: Foundations of Neural Computation. A Bradford Book. MIT Press, 1999. ISBN 9780262581684. ", + "bbox": [ + 173, + 723, + 825, + 766 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Olivier J. Henaff, Robbe L. T. Goris, and Eero P. Simoncelli. Perceptual straightening of natural ´ videos. Nature Neuroscience, Apr 2019. ", + "bbox": [ + 171, + 776, + 825, + 806 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. ", + "bbox": [ + 171, + 816, + 823, + 845 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Dinesh Jayaraman and Kristen Grauman. Learning image representations tied to ego-motion. In ICCV, 2015. ", + "bbox": [ + 171, + 856, + 823, + 885 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Longlong Jing and Yingli Tian. Self-supervised spatiotemporal feature learning by video geometric transformations. arXiv preprint arXiv:1811.11387, 2018. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised learning with deep generative models. In Advances in neural information processing systems, pp. 3581–3589, 2014. ", + "bbox": [ + 173, + 103, + 823, + 146 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alexander Kolesnikov, Xiaohua Zhai, and Lucas Beyer. Revisiting self-supervised visual representation learning. CoRR, abs/1901.09005, 2019. URL http://arxiv.org/abs/1901.09005. ", + "bbox": [ + 173, + 155, + 823, + 184 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning through probabilistic program induction. Science, 350(6266):1332–1338, 2015. ", + "bbox": [ + 173, + 193, + 823, + 222 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Barbara Landau, Linda B Smith, and Susan S Jones. The importance of shape in early lexical learning. Cognitive development, 3(3):299–321, 1988. ", + "bbox": [ + 171, + 231, + 823, + 261 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In CVPR, pp. 6874–6883, 2017. ", + "bbox": [ + 174, + 268, + 820, + 299 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436, 2015. ", + "bbox": [ + 173, + 306, + 810, + 323 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on Challenges in Representation Learning, ICML, volume 3, pp. 2, 2013. ", + "bbox": [ + 173, + 332, + 825, + 375 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yin Li, Manohar Paluri, James M Rehg, and Piotr Dollar. Unsupervised learning of edges. In ´ CVPR, 2016. ", + "bbox": [ + 174, + 382, + 821, + 412 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ellen M Markman. Categorization and naming in children: Problems of induction. mit Press, 1989. ", + "bbox": [ + 171, + 421, + 821, + 438 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 26, pp. 3111–3119. Curran Associates, Inc., 2013. ", + "bbox": [ + 173, + 444, + 825, + 502 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning using temporal order verification. In ECCV, 2016. ", + "bbox": [ + 174, + 511, + 823, + 540 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Takeru Miyato, Shin-ichi Maeda, Shin Ishii, and Masanori Koyama. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 2018. ", + "bbox": [ + 173, + 549, + 825, + 592 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Andriy Mnih and Koray Kavukcuoglu. Learning word embeddings efficiently with noise-contrastive estimation. In Advances in neural information processing systems, pp. 2265–2273, 2013. ", + "bbox": [ + 173, + 601, + 821, + 631 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European Conference on Computer Vision, pp. 69–84. Springer, 2016. ", + "bbox": [ + 173, + 638, + 821, + 669 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Stephanie E Palmer, Olivier Marre, Michael J Berry, and William Bialek. Predictive information in a sensory population. Proceedings of the National Academy of Sciences, 112(22):6908–6913, 2015. ", + "bbox": [ + 178, + 676, + 823, + 719 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning fea-´ tures by watching objects move. arXiv preprint arXiv:1612.06370, 2016. ", + "bbox": [ + 171, + 728, + 823, + 758 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016. ", + "bbox": [ + 171, + 767, + 825, + 796 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Lerrel Pinto, James Davidson, and Abhinav Gupta. Supervision via competition: Robot adversaries for learning tasks. arXiv preprint arXiv:1610.01685, 2016. ", + "bbox": [ + 171, + 805, + 823, + 834 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature neuroscience, 2(1):79, 1999. ", + "bbox": [ + 171, + 843, + 823, + 872 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pp. 91–99, 2015. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211–252, 2015. ", + "bbox": [ + 178, + 103, + 821, + 146 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1134–1141. IEEE, 2018. ", + "bbox": [ + 174, + 155, + 825, + 210 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014. ", + "bbox": [ + 176, + 219, + 823, + 263 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. CoRR, abs/1409.4842, 2014. URL http://arxiv.org/abs/1409.4842. ", + "bbox": [ + 173, + 271, + 825, + 315 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. arXiv preprint arXiv:1503.01817, 2015. ", + "bbox": [ + 174, + 323, + 823, + 366 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019. ", + "bbox": [ + 173, + 375, + 823, + 404 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. ", + "bbox": [ + 173, + 412, + 823, + 441 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In ICCV, 2015. ", + "bbox": [ + 173, + 450, + 823, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via nonparametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3733–3742, 2018. ", + "bbox": [ + 173, + 488, + 823, + 531 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised Data Augmentation. arXiv e-prints, art. arXiv:1904.12848, Apr 2019. ", + "bbox": [ + 169, + 540, + 823, + 569 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Amir R Zamir, Tilman Wekel, Pulkit Agrawal, Colin Wei, Jitendra Malik, and Silvio Savarese. Generic 3D representation via pose estimation and matching. In ECCV, 2016. ", + "bbox": [ + 169, + 577, + 821, + 607 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. $S ^ { 4 } L$ : Self-supervised semisupervised learning. arXiv preprint arXiv:1905.03670, 2019. ", + "bbox": [ + 173, + 614, + 821, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European conference on computer vision, pp. 649–666. Springer, 2016. ", + "bbox": [ + 173, + 652, + 821, + 683 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. In Technical Report CMU-CALD-02-107, Carnegie Mellon University, 2002. ", + "bbox": [ + 171, + 690, + 821, + 720 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning of visual embeddings. arXiv preprint arXiv:1903.12355, 2019. ", + "bbox": [ + 173, + 728, + 823, + 758 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 103, + 299, + 117 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.1 ADDITIONAL RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 133, + 377, + 148 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/cd16abc0799fc70d1a375836e8cc04ff4651be0d7ba68507f20886a5866fd0b3.jpg", + "table_caption": [ + "Table 4: Data efficient classification results with Top-1 accuracy. " + ], + "table_footnote": [], + "table_body": "
Labeled data Method1%2%5%10% Top-1 accuracy20%50%100%
Supervised trained on pixels23.134.850.260.067.674.478.0
ReseNet trained on CPC v2 (fine-tuned)46.354.764.169.773.077.380.6
", + "bbox": [ + 184, + 193, + 815, + 267 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/ef57c53b71083a6069724c7b7ef7e00273991c3de17438cac7824c7e459eff57.jpg", + "table_caption": [ + "Table 5: Data efficient classification results with Top-5 accuracy (data in Fig. 1). " + ], + "table_footnote": [], + "table_body": "
Labeled data Method1%2%5%10% Top-5 accuracy20%50%100%
Supervised trained on pixels44.159.975.282.187.991.893.8
ReseNet trained on CPC v2 (fine-tuned)72.979.686.189.591.493.695.2
", + "bbox": [ + 184, + 321, + 815, + 397 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.2 INFONCE IMPLEMENTATION ", + "text_level": 1, + "bbox": [ + 176, + 429, + 418, + 444 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For completeness, we provide pseudo-code for the main calculations involved in the InfoNCE objective, loosely modeled after Tensorflow operations. We suppose we have just calculated a set of latents $z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )$ for $i , j \\in \\{ 1 , . . . , 7 \\}$ , each one being e.g. a 4096-dimensional vector. Assuming we do so for a batch of $B$ images $\\{ { \\pmb x } \\}$ , the set of latents is a tensor of size $B \\times 7 \\times 7 \\times 4 0 9 6$ . ", + "bbox": [ + 173, + 455, + 825, + 512 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "def pixelCNN(latents): # latents: [B, H, W, D] cres $=$ latents cres_dim $=$ cres.shape[-1] for in range(5): c $=$ Conv2D(output_channels $= 2 5 6$ , kernel_shape $=$ (1, 1))(cres) c = ReLU(c) c = Conv2D(output_channels $= 2 5 6$ , kernel_shape $=$ (1, 3))(c) c = Pad(c, [[0, 0], [1, 0], [0, 0], [0, 0]]) c = Conv2D(output_channels $= 2 5 6$ , kernel_shape $=$ (2, 1), type $= ^ { \\parallel }$ VALID')(c) c $=$ ReLU(c) c $=$ Conv2D(output_channels $=$ cres_dim, kernel_shape $=$ (1, 1))(c) cres $=$ cres + c cres $=$ ReLU(cres) return cres ", + "bbox": [ + 176, + 119, + 663, + 398 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.3 LINEAR CLASSIFICATION ", + "text_level": 1, + "bbox": [ + 176, + 421, + 392, + 435 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "• Model architecture: Having extracted $8 0 \\mathrm { x } 8 0$ patches with a stride of $3 2 \\mathrm { x } 3 2 $ from a $2 4 0 \\times 2 4 0$ shaped input image, we end up with a grid of 6x6 features (each of which is obtained from our ResNet-161 architecture). This gives us a [6,6,4096] tensor for the image. We then use a Batch-Normalization layer to normalize the features (without scale parameter) followed by a 1x1 convolution mapping each feature in the grid to the 1000 logits for ImageNet classification. We then spatially-mean-pool these logits to end up with the final log probabilities for the linear classification. \n• We use the Inception preprocessing (Szegedy et al., 2014) to extract $2 4 0 \\times 2 4 0$ crops from the raw image. The image is divided into subcrops as per CPC data-preprocessing used for CPC pre-training. \n• Optimization details: We use Adam Optimizer with a learning rate of 5e-4. We train the model on a batch size of 512 images with 32 images per core spread over 16 workers. ", + "bbox": [ + 217, + 443, + 826, + 619 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.4 EFFICIENT CLASSIFICATION: PURELY SUPERVISED ", + "bbox": [ + 174, + 637, + 565, + 650 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "In order to find the best model within this class, we vary the following hyperparameters: ", + "bbox": [ + 171, + 661, + 750, + 676 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "• Model architecture: We investigate using ResNet-50, ResNet-101, and ResNet-152 model architectures, all of them using the $\\mathbf { \\hat { v } } _ { \\mathbf { \\Delta } } 2 \\mathbf { \\hat { \\mu } }$ variant (He et al., 2016b), and find larger architecture to perform better, even when given smaller amounts of data. We insert a DropOut layer before the final linear classification layer (Srivastava et al., 2014). • Data pre-processing: We use the Inception pre-processing pipeline (Szegedy et al., 2014). • Optimization details: We vary the learning rate in $\\{ 0 . 0 5 , 0 . 1 , 0 . 2 \\}$ , the weight decay logarithmically from $1 0 ^ { - 5 }$ to $1 0 ^ { - 2 }$ , the DropOut linearly from 0 to 1, and the batch size per worker in $\\{ 1 6 , 3 2 \\}$ . ", + "bbox": [ + 217, + 689, + 823, + 811 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We chose the best performing model for each training subset $\\mathbb { D } _ { l }$ of labeled ImageNet (using a separate validation set), and report its accuracy on the test set (i.e. the publicly available ILSVRC-2012 validation set). ", + "bbox": [ + 178, + 821, + 825, + 864 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/parse/train/rJerHlrYwH/rJerHlrYwH_middle.json b/parse/train/rJerHlrYwH/rJerHlrYwH_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..6e522952034fbfd0d54cc28c42936e4afaba012f --- /dev/null +++ b/parse/train/rJerHlrYwH/rJerHlrYwH_middle.json @@ -0,0 +1,43945 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 419, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 398, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 398, + 97 + ], + "score": 1.0, + "content": "DATA-EFFICIENT IMAGE RECOGNITION", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 420, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 420, + 117 + ], + "score": 1.0, + "content": "WITH CONTRASTIVE PREDICTIVE CODING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "score": 1.0, + "content": "Human observers can learn to recognize new categories of objects from a hand-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "ful of examples, yet doing so with machine perception remains an open challenge.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "score": 1.0, + "content": "We hypothesize that data-efficient recognition is enabled by representations which", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "score": 1.0, + "content": "make the variability in natural signals more predictable, as suggested by recent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "perceptual evidence. We therefore revisit and improve Contrastive Predictive Cod-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "ing, a recently-proposed unsupervised learning framework, and arrive at a repre-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "score": 1.0, + "content": "sentation which enables generalization from small amounts of labeled data. When", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 222, + 302 + ], + "score": 1.0, + "content": "provided with only", + "type": "text" + }, + { + "bbox": [ + 222, + 289, + 237, + 299 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 289, + 470, + 302 + ], + "score": 1.0, + "content": "of ImageNet labels (i.e. 13 per class), this model retains", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 290, + 312 + ], + "score": 1.0, + "content": "a strong classification performance,", + "type": "text" + }, + { + "bbox": [ + 291, + 300, + 311, + 310 + ], + "score": 0.86, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 300, + 469, + 312 + ], + "score": 1.0, + "content": "Top-5 accuracy, outperforming super-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 218, + 322 + ], + "score": 1.0, + "content": "vised networks by", + "type": "text" + }, + { + "bbox": [ + 219, + 311, + 239, + 321 + ], + "score": 0.87, + "content": "28 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 311, + 251, + 322 + ], + "score": 1.0, + "content": "(a", + "type": "text" + }, + { + "bbox": [ + 251, + 311, + 271, + 321 + ], + "score": 0.87, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "relative improvement) and state-of-the-art semi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 237, + 334 + ], + "score": 1.0, + "content": "supervised methods by", + "type": "text" + }, + { + "bbox": [ + 237, + 322, + 256, + 332 + ], + "score": 0.87, + "content": "14 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 322, + 470, + 334 + ], + "score": 1.0, + "content": ". We also find this representation to serve as a useful", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "substrate for object detection on the PASCAL-VOC 2007 dataset, approaching the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 344, + 463, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 463, + 357 + ], + "score": 1.0, + "content": "performance of representations trained with a fully annotated ImageNet dataset.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 109, + 377, + 205, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 208, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 208, + 393 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 296, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 296, + 415 + ], + "score": 1.0, + "content": "Deep neural networks excel at perceptual tasks", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 413, + 297, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 297, + 426 + ], + "score": 1.0, + "content": "when labeled data are abundant, yet their per-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 424, + 297, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 297, + 437 + ], + "score": 1.0, + "content": "formance degrades substantially when provided", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 435, + 297, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 297, + 448 + ], + "score": 1.0, + "content": "with limited supervision (Fig. 1, red). In con-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 446, + 297, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 297, + 459 + ], + "score": 1.0, + "content": "trast, humans and animals can quickly learn", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 457, + 297, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 297, + 469 + ], + "score": 1.0, + "content": "about new classes of objects from few exam-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 468, + 297, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 297, + 480 + ], + "score": 1.0, + "content": "ples (Landau et al., 1988; Markman, 1989).", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 479, + 297, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 297, + 491 + ], + "score": 1.0, + "content": "What accounts for this monumental difference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 490, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 297, + 502 + ], + "score": 1.0, + "content": "in data-efficiency between biological and ma-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 501, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 297, + 514 + ], + "score": 1.0, + "content": "chine vision? While highly-structured repre-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 512, + 297, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 297, + 524 + ], + "score": 1.0, + "content": "sentations (e.g. as proposed by Lake et al.,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 523, + 297, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 297, + 535 + ], + "score": 1.0, + "content": "2015) may improve data-efficiency, it remains", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 534, + 298, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 298, + 546 + ], + "score": 1.0, + "content": "unclear how to program explicit structures that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 297, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 297, + 557 + ], + "score": 1.0, + "content": "capture the enormous complexity of real visual", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 556, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 297, + 568 + ], + "score": 1.0, + "content": "scenes like those in ImageNet (Russakovsky", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 298, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 298, + 579 + ], + "score": 1.0, + "content": "et al., 2015). An alternative hypothesis has", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 578, + 298, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 298, + 590 + ], + "score": 1.0, + "content": "proposed that intelligent systems need not be", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 590, + 297, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 297, + 600 + ], + "score": 1.0, + "content": "structured a priori, but can instead learn about", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 601, + 297, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 297, + 612 + ], + "score": 1.0, + "content": "the structure of the world in an unsupervised", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 611, + 297, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 297, + 623 + ], + "score": 1.0, + "content": "manner (Barlow, 1989; Hinton et al., 1999; Le-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 621, + 297, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 297, + 635 + ], + "score": 1.0, + "content": "Cun et al., 2015). Choosing an appropriate", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 633, + 298, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 298, + 645 + ], + "score": 1.0, + "content": "training objective is an open problem, but a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 644, + 297, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 297, + 656 + ], + "score": 1.0, + "content": "promising guiding principle has emerged re-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 655, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 297, + 666 + ], + "score": 1.0, + "content": "cently: good representations should make the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "score": 1.0, + "content": "spatio-temporal variability in natural signals", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 297, + 689 + ], + "score": 1.0, + "content": "more predictable. Indeed, human perceptual", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 297, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 297, + 699 + ], + "score": 1.0, + "content": "representations have been shown to linearize", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 306, + 406, + 502, + 581 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 306, + 406, + 502, + 581 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 502, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 502, + 581 + ], + "score": 0.973, + "type": "image", + "image_path": "7b8ed316aeaae9bdfe3d3d7e12974b01335c33b7eb790b24e724753c0422fb2e.jpg" + } + ] + } + ], + "index": 52.5, + "virtual_lines": [ + { + "bbox": [ + 306, + 406, + 502, + 418.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 306, + 418.5, + 502, + 431.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 306, + 431.0, + 502, + 443.5 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 306, + 443.5, + 502, + 456.0 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 306, + 456.0, + 502, + 468.5 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 306, + 468.5, + 502, + 481.0 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 306, + 481.0, + 502, + 493.5 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 306, + 493.5, + 502, + 506.0 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 306, + 506.0, + 502, + 518.5 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 306, + 518.5, + 502, + 531.0 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 306, + 531.0, + 502, + 543.5 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 306, + 543.5, + 502, + 556.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 306, + 556.0, + 502, + 568.5 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 306, + 568.5, + 502, + 581.0 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 586, + 504, + 684 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 505, + 597 + ], + "score": 1.0, + "content": "Figure 1: Data-efficient image recognition with", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding. With decreas-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 303, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 303, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "ing amounts of labeled data, supervised networks", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 303, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 303, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "trained on pixels fail to generalize (red). When", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "trained on unsupervised representations learned", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 640, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 506, + 652 + ], + "score": 1.0, + "content": "with CPC, these networks retain a much higher", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "accuracy in this low-data regime (blue). Equiva-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "lently, the accuracy of supervised networks can be", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 673, + 466, + 685 + ], + "spans": [ + { + "bbox": [ + 303, + 673, + 466, + 685 + ], + "score": 1.0, + "content": "matched with significantly fewer labels.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 64 + } + ], + "index": 58.25 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "score": 1.0, + "content": "(or ‘straighten’) the temporal transformations found in natural videos, a property lacking from cur-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "rent supervised image recognition models (Henaff et al., 2019), and theories of both spatial and tem- ´", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "poral predictability have succeeded in describing properties of early visual areas (Rao & Ballard,", + "type": "text" + } + ], + "index": 71 + } + ], + "index": 70 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 419, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 398, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 398, + 97 + ], + "score": 1.0, + "content": "DATA-EFFICIENT IMAGE RECOGNITION", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 98, + 420, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 98, + 420, + 117 + ], + "score": 1.0, + "content": "WITH CONTRASTIVE PREDICTIVE CODING", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 158 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 354 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "score": 1.0, + "content": "Human observers can learn to recognize new categories of objects from a hand-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 236 + ], + "score": 1.0, + "content": "ful of examples, yet doing so with machine perception remains an open challenge.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "score": 1.0, + "content": "We hypothesize that data-efficient recognition is enabled by representations which", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 470, + 257 + ], + "score": 1.0, + "content": "make the variability in natural signals more predictable, as suggested by recent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "perceptual evidence. We therefore revisit and improve Contrastive Predictive Cod-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 279 + ], + "score": 1.0, + "content": "ing, a recently-proposed unsupervised learning framework, and arrive at a repre-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 289 + ], + "score": 1.0, + "content": "sentation which enables generalization from small amounts of labeled data. When", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 222, + 302 + ], + "score": 1.0, + "content": "provided with only", + "type": "text" + }, + { + "bbox": [ + 222, + 289, + 237, + 299 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 289, + 470, + 302 + ], + "score": 1.0, + "content": "of ImageNet labels (i.e. 13 per class), this model retains", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 290, + 312 + ], + "score": 1.0, + "content": "a strong classification performance,", + "type": "text" + }, + { + "bbox": [ + 291, + 300, + 311, + 310 + ], + "score": 0.86, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 300, + 469, + 312 + ], + "score": 1.0, + "content": "Top-5 accuracy, outperforming super-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 218, + 322 + ], + "score": 1.0, + "content": "vised networks by", + "type": "text" + }, + { + "bbox": [ + 219, + 311, + 239, + 321 + ], + "score": 0.87, + "content": "28 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 311, + 251, + 322 + ], + "score": 1.0, + "content": "(a", + "type": "text" + }, + { + "bbox": [ + 251, + 311, + 271, + 321 + ], + "score": 0.87, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 311, + 469, + 322 + ], + "score": 1.0, + "content": "relative improvement) and state-of-the-art semi-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 470, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 237, + 334 + ], + "score": 1.0, + "content": "supervised methods by", + "type": "text" + }, + { + "bbox": [ + 237, + 322, + 256, + 332 + ], + "score": 0.87, + "content": "14 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 322, + 470, + 334 + ], + "score": 1.0, + "content": ". We also find this representation to serve as a useful", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 469, + 345 + ], + "score": 1.0, + "content": "substrate for object detection on the PASCAL-VOC 2007 dataset, approaching the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 344, + 463, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 344, + 463, + 357 + ], + "score": 1.0, + "content": "performance of representations trained with a fully annotated ImageNet dataset.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11, + "bbox_fs": [ + 141, + 212, + 470, + 357 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 377, + 205, + 390 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 208, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 208, + 393 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 404, + 296, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 296, + 415 + ], + "score": 1.0, + "content": "Deep neural networks excel at perceptual tasks", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 413, + 297, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 297, + 426 + ], + "score": 1.0, + "content": "when labeled data are abundant, yet their per-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 424, + 297, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 297, + 437 + ], + "score": 1.0, + "content": "formance degrades substantially when provided", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 435, + 297, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 297, + 448 + ], + "score": 1.0, + "content": "with limited supervision (Fig. 1, red). In con-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 446, + 297, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 297, + 459 + ], + "score": 1.0, + "content": "trast, humans and animals can quickly learn", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 457, + 297, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 297, + 469 + ], + "score": 1.0, + "content": "about new classes of objects from few exam-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 468, + 297, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 297, + 480 + ], + "score": 1.0, + "content": "ples (Landau et al., 1988; Markman, 1989).", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 479, + 297, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 297, + 491 + ], + "score": 1.0, + "content": "What accounts for this monumental difference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 490, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 297, + 502 + ], + "score": 1.0, + "content": "in data-efficiency between biological and ma-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 501, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 297, + 514 + ], + "score": 1.0, + "content": "chine vision? While highly-structured repre-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 512, + 297, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 297, + 524 + ], + "score": 1.0, + "content": "sentations (e.g. as proposed by Lake et al.,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 523, + 297, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 297, + 535 + ], + "score": 1.0, + "content": "2015) may improve data-efficiency, it remains", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 534, + 298, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 298, + 546 + ], + "score": 1.0, + "content": "unclear how to program explicit structures that", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 297, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 297, + 557 + ], + "score": 1.0, + "content": "capture the enormous complexity of real visual", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 556, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 297, + 568 + ], + "score": 1.0, + "content": "scenes like those in ImageNet (Russakovsky", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 298, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 298, + 579 + ], + "score": 1.0, + "content": "et al., 2015). An alternative hypothesis has", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 578, + 298, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 298, + 590 + ], + "score": 1.0, + "content": "proposed that intelligent systems need not be", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 590, + 297, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 297, + 600 + ], + "score": 1.0, + "content": "structured a priori, but can instead learn about", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 601, + 297, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 297, + 612 + ], + "score": 1.0, + "content": "the structure of the world in an unsupervised", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 611, + 297, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 297, + 623 + ], + "score": 1.0, + "content": "manner (Barlow, 1989; Hinton et al., 1999; Le-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 621, + 297, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 621, + 297, + 635 + ], + "score": 1.0, + "content": "Cun et al., 2015). Choosing an appropriate", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 633, + 298, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 298, + 645 + ], + "score": 1.0, + "content": "training objective is an open problem, but a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 644, + 297, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 297, + 656 + ], + "score": 1.0, + "content": "promising guiding principle has emerged re-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 655, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 297, + 666 + ], + "score": 1.0, + "content": "cently: good representations should make the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "score": 1.0, + "content": "spatio-temporal variability in natural signals", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 677, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 297, + 689 + ], + "score": 1.0, + "content": "more predictable. Indeed, human perceptual", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 297, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 297, + 699 + ], + "score": 1.0, + "content": "representations have been shown to linearize", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 403, + 298, + 699 + ] + }, + { + "type": "image", + "bbox": [ + 306, + 406, + 502, + 581 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 306, + 406, + 502, + 581 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 502, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 502, + 581 + ], + "score": 0.973, + "type": "image", + "image_path": "7b8ed316aeaae9bdfe3d3d7e12974b01335c33b7eb790b24e724753c0422fb2e.jpg" + } + ] + } + ], + "index": 52.5, + "virtual_lines": [ + { + "bbox": [ + 306, + 406, + 502, + 418.5 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 306, + 418.5, + 502, + 431.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 306, + 431.0, + 502, + 443.5 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 306, + 443.5, + 502, + 456.0 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 306, + 456.0, + 502, + 468.5 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 306, + 468.5, + 502, + 481.0 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 306, + 481.0, + 502, + 493.5 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 306, + 493.5, + 502, + 506.0 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 306, + 506.0, + 502, + 518.5 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 306, + 518.5, + 502, + 531.0 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 306, + 531.0, + 502, + 543.5 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 306, + 543.5, + 502, + 556.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 306, + 556.0, + 502, + 568.5 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 306, + 568.5, + 502, + 581.0 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 586, + 504, + 684 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 505, + 597 + ], + "score": 1.0, + "content": "Figure 1: Data-efficient image recognition with", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding. With decreas-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 303, + 608, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 303, + 608, + 505, + 619 + ], + "score": 1.0, + "content": "ing amounts of labeled data, supervised networks", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 303, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 303, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "trained on pixels fail to generalize (red). When", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 304, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "trained on unsupervised representations learned", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 640, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 506, + 652 + ], + "score": 1.0, + "content": "with CPC, these networks retain a much higher", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "accuracy in this low-data regime (blue). Equiva-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "lently, the accuracy of supervised networks can be", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 673, + 466, + 685 + ], + "spans": [ + { + "bbox": [ + 303, + 673, + 466, + 685 + ], + "score": 1.0, + "content": "matched with significantly fewer labels.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 64 + } + ], + "index": 58.25 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 712 + ], + "score": 1.0, + "content": "(or ‘straighten’) the temporal transformations found in natural videos, a property lacking from cur-", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "rent supervised image recognition models (Henaff et al., 2019), and theories of both spatial and tem- ´", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "poral predictability have succeeded in describing properties of early visual areas (Rao & Ballard,", + "type": "text" + } + ], + "index": 71 + } + ], + "index": 70, + "bbox_fs": [ + 105, + 699, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "1999; Palmer et al., 2015). In this work, we hypothesize that spatially predictable representations", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 393, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 393, + 108 + ], + "score": 1.0, + "content": "may allow artificial systems to benefit from human-like data-efficiency.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 124 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding (CPC, van den Oord et al., 2018) is an unsupervised objective which", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 122, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 133 + ], + "score": 1.0, + "content": "learns such predictable representations. CPC is a general technique that only requires in its definition", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "that observations be ordered along e.g. temporal or spatial dimensions, and as such has been applied", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "to a variety of different modalities including speech, natural language and images. This generality,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "combined with the strong performance of its representations in downstream linear classification", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "tasks, makes CPC a promising candidate for investigating the efficacy of predictable representations", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 253, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 253, + 189 + ], + "score": 1.0, + "content": "for data-efficient image recognition.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 287, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 289, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 289, + 206 + ], + "score": 1.0, + "content": "Our work makes the following contributions:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 133, + 214, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 505, + 227 + ], + "score": 1.0, + "content": "• We revisit CPC in terms of its architecture and training methodology, and arrive at a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "new implementation of CPC with dramatically-improved ability to linearly separate im-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 236, + 384, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 191, + 249 + ], + "score": 1.0, + "content": "age classes", + "type": "text" + }, + { + "bbox": [ + 192, + 236, + 217, + 247 + ], + "score": 0.87, + "content": "+ 1 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 236, + 384, + 249 + ], + "score": 1.0, + "content": "Top-1 ImageNet classification accuracy).", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "We then train deep networks on top of the resulting CPC representations using very few la-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 261, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 261, + 216, + 276 + ], + "score": 1.0, + "content": "beled images (e.g.", + "type": "text" + }, + { + "bbox": [ + 216, + 262, + 230, + 273 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 261, + 505, + 276 + ], + "score": 1.0, + "content": "of the ImageNet dataset), and demonstrate test-time classification ac-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 339, + 285 + ], + "score": 1.0, + "content": "curacy far above networks trained on raw pixels (", + "type": "text" + }, + { + "bbox": [ + 339, + 273, + 358, + 284 + ], + "score": 0.85, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 273, + 432, + 285 + ], + "score": 1.0, + "content": "Top-5 accuracy, a", + "type": "text" + }, + { + "bbox": [ + 432, + 273, + 452, + 284 + ], + "score": 0.88, + "content": "28 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "absolute im-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 284, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 284, + 477, + 297 + ], + "score": 1.0, + "content": "provement), outperforming all other unsupervised representation learning methods", + "type": "text" + }, + { + "bbox": [ + 477, + 284, + 504, + 295 + ], + "score": 0.87, + "content": "( + 1 5 \\%", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "Top-5 accuracy over the previous state-of-the-art (Zhai et al., 2019)). Surprisingly, this", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "representation also surpasses supervised methods when given the entire ImageNet dataset", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 145, + 317, + 236, + 329 + ], + "spans": [ + { + "bbox": [ + 145, + 317, + 166, + 328 + ], + "score": 0.85, + "content": "+ 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 317, + 236, + 329 + ], + "score": 1.0, + "content": "Top-5 accuracy).", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 138, + 330, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 330, + 505, + 345 + ], + "score": 1.0, + "content": "We isolate the contributions of different components of the final model to such downstream", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "tasks. Interestingly, we find that linear classification accuracy is not always predictive of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "low-data classification accuracy, emphasizing the importance of this metric as a stand-alone", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 363, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 363, + 296, + 379 + ], + "score": 1.0, + "content": "benchmark for unsupervised learning.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 135, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 135, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "• Finally, we assess the generality of CPC representations by transferring them to a new task", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 142, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "and dataset: object detection on PASCAL-VOC 2007. Consistent with the results from the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 402, + 468, + 415 + ], + "spans": [ + { + "bbox": [ + 142, + 402, + 468, + 415 + ], + "score": 1.0, + "content": "previous section, we find CPC to give state-of-the-art performance in this setting.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 429, + 244, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 245, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 245, + 444 + ], + "score": 1.0, + "content": "2 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 455, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "We first review the CPC architecture and learning objective in section 2.1, before detailing how we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 466, + 405, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 405, + 478 + ], + "score": 1.0, + "content": "use its resulting representations for image recognition tasks in section 2.2.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "title", + "bbox": [ + 108, + 490, + 286, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 287, + 503 + ], + "score": 1.0, + "content": "2.1 CONTRASTIVE PREDICTIVE CODING", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding as formulated in (van den Oord et al., 2018) learns representations", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "by training neural networks to predict the representations of future observations from those of past", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "ones. When applied to images, the original formulation of CPC operates by predicting the represen-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "tations of patches below a certain position from those above it (Fig. 2, left). These predictions are", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "evaluated using a contrastive loss, in which the network must correctly classify the ‘future’ repre-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "sentation amongst a set of unrelated ‘negative’ representations. This avoids trivial solutions such as", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "representing all patches with a constant vector, as would be the case with a mean squared error loss.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 683 + ], + "lines": [ + { + "bbox": [ + 104, + 591, + 507, + 609 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 484, + 609 + ], + "score": 1.0, + "content": "In the CPC architecture, each input image is first divided into a set of overlapping patches", + "type": "text" + }, + { + "bbox": [ + 484, + 596, + 501, + 606 + ], + "score": 0.88, + "content": "\\mathbf { \\delta } _ { \\mathbf { x } _ { i , j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 591, + 507, + 609 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 304, + 618 + ], + "score": 1.0, + "content": "each of which is encoded with a neural network", + "type": "text" + }, + { + "bbox": [ + 304, + 605, + 315, + 616 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 604, + 397, + 618 + ], + "score": 1.0, + "content": "into a single vector", + "type": "text" + }, + { + "bbox": [ + 397, + 605, + 461, + 618 + ], + "score": 0.93, + "content": "z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 604, + 506, + 618 + ], + "score": 1.0, + "content": ". To make", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 289, + 629 + ], + "score": 1.0, + "content": "predictions, a masked convolutional network", + "type": "text" + }, + { + "bbox": [ + 290, + 618, + 301, + 628 + ], + "score": 0.87, + "content": "g _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "is then applied to the grid of feature vectors. The", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 399, + 640 + ], + "score": 1.0, + "content": "masks are such that the receptive field of each resulting context vector", + "type": "text" + }, + { + "bbox": [ + 400, + 628, + 415, + 639 + ], + "score": 0.88, + "content": "c _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "only includes feature", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 268, + 652 + ], + "score": 1.0, + "content": "vectors that lie above it in the image (i.e.", + "type": "text" + }, + { + "bbox": [ + 268, + 638, + 320, + 650 + ], + "score": 0.92, + "content": "\\{ z _ { u , v } \\} _ { u \\leq i , v } )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 637, + 506, + 652 + ], + "score": 1.0, + "content": ". The prediction task then consists o predicting", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 201, + 662 + ], + "score": 1.0, + "content": "‘future’ feature vectors", + "type": "text" + }, + { + "bbox": [ + 202, + 650, + 227, + 661 + ], + "score": 0.9, + "content": "z _ { i + k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 648, + 346, + 662 + ], + "score": 1.0, + "content": "from current context vectors", + "type": "text" + }, + { + "bbox": [ + 346, + 650, + 361, + 661 + ], + "score": 0.88, + "content": "c _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 648, + 392, + 662 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 393, + 649, + 419, + 659 + ], + "score": 0.9, + "content": "k > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 648, + 506, + 662 + ], + "score": 1.0, + "content": ". The predictions are", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 260, + 673 + ], + "score": 1.0, + "content": "made linearly: given a context vector", + "type": "text" + }, + { + "bbox": [ + 260, + 662, + 275, + 672 + ], + "score": 0.86, + "content": "c _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 660, + 358, + 673 + ], + "score": 1.0, + "content": ", a prediction length", + "type": "text" + }, + { + "bbox": [ + 358, + 660, + 384, + 670 + ], + "score": 0.9, + "content": "k > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 660, + 486, + 673 + ], + "score": 1.0, + "content": ", and a prediction matrix", + "type": "text" + }, + { + "bbox": [ + 486, + 660, + 501, + 671 + ], + "score": 0.88, + "content": "W _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 660, + 506, + 673 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 301, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 227, + 685 + ], + "score": 1.0, + "content": "the predicted feature vector is", + "type": "text" + }, + { + "bbox": [ + 228, + 672, + 297, + 683 + ], + "score": 0.91, + "content": "\\hat { \\boldsymbol { z } } _ { i + k , j } = \\boldsymbol { W } _ { k } \\boldsymbol { c } _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 671, + 301, + 685 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "The quality of this prediction is then evaluated using a contrastive loss. Specifically, the goal is to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 228, + 711 + ], + "score": 1.0, + "content": "correctly recognize the target", + "type": "text" + }, + { + "bbox": [ + 229, + 700, + 254, + 711 + ], + "score": 0.82, + "content": "z _ { i + k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 699, + 462, + 711 + ], + "score": 1.0, + "content": "among a set of randomly sampled feature vectors", + "type": "text" + }, + { + "bbox": [ + 462, + 699, + 481, + 711 + ], + "score": 0.92, + "content": "\\{ z _ { l } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "the dataset. We compute the probability assigned to the target using a softmax, and evaluate this", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "probability using the usual cross-entropy loss. Summing this loss over locations and prediction", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "1999; Palmer et al., 2015). In this work, we hypothesize that spatially predictable representations", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 393, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 393, + 108 + ], + "score": 1.0, + "content": "may allow artificial systems to benefit from human-like data-efficiency.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 108 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 506, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 124 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding (CPC, van den Oord et al., 2018) is an unsupervised objective which", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 122, + 505, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 133 + ], + "score": 1.0, + "content": "learns such predictable representations. CPC is a general technique that only requires in its definition", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "that observations be ordered along e.g. temporal or spatial dimensions, and as such has been applied", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "to a variety of different modalities including speech, natural language and images. This generality,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "combined with the strong performance of its representations in downstream linear classification", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "tasks, makes CPC a promising candidate for investigating the efficacy of predictable representations", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 253, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 253, + 189 + ], + "score": 1.0, + "content": "for data-efficient image recognition.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 109, + 506, + 189 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 287, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 289, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 289, + 206 + ], + "score": 1.0, + "content": "Our work makes the following contributions:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 191, + 289, + 206 + ] + }, + { + "type": "list", + "bbox": [ + 133, + 214, + 505, + 414 + ], + "lines": [ + { + "bbox": [ + 133, + 213, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 133, + 213, + 505, + 227 + ], + "score": 1.0, + "content": "• We revisit CPC in terms of its architecture and training methodology, and arrive at a", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 225, + 505, + 237 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 505, + 237 + ], + "score": 1.0, + "content": "new implementation of CPC with dramatically-improved ability to linearly separate im-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 236, + 384, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 191, + 249 + ], + "score": 1.0, + "content": "age classes", + "type": "text" + }, + { + "bbox": [ + 192, + 236, + 217, + 247 + ], + "score": 0.87, + "content": "+ 1 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 236, + 384, + 249 + ], + "score": 1.0, + "content": "Top-1 ImageNet classification accuracy).", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 138, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 138, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "We then train deep networks on top of the resulting CPC representations using very few la-", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 261, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 261, + 216, + 276 + ], + "score": 1.0, + "content": "beled images (e.g.", + "type": "text" + }, + { + "bbox": [ + 216, + 262, + 230, + 273 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 261, + 505, + 276 + ], + "score": 1.0, + "content": "of the ImageNet dataset), and demonstrate test-time classification ac-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 273, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 141, + 273, + 339, + 285 + ], + "score": 1.0, + "content": "curacy far above networks trained on raw pixels (", + "type": "text" + }, + { + "bbox": [ + 339, + 273, + 358, + 284 + ], + "score": 0.85, + "content": "73 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 273, + 432, + 285 + ], + "score": 1.0, + "content": "Top-5 accuracy, a", + "type": "text" + }, + { + "bbox": [ + 432, + 273, + 452, + 284 + ], + "score": 0.88, + "content": "28 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 273, + 505, + 285 + ], + "score": 1.0, + "content": "absolute im-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 284, + 504, + 297 + ], + "spans": [ + { + "bbox": [ + 141, + 284, + 477, + 297 + ], + "score": 1.0, + "content": "provement), outperforming all other unsupervised representation learning methods", + "type": "text" + }, + { + "bbox": [ + 477, + 284, + 504, + 295 + ], + "score": 0.87, + "content": "( + 1 5 \\%", + "type": "inline_equation" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 295, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 141, + 295, + 506, + 308 + ], + "score": 1.0, + "content": "Top-5 accuracy over the previous state-of-the-art (Zhai et al., 2019)). Surprisingly, this", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 141, + 307, + 505, + 319 + ], + "score": 1.0, + "content": "representation also surpasses supervised methods when given the entire ImageNet dataset", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 145, + 317, + 236, + 329 + ], + "spans": [ + { + "bbox": [ + 145, + 317, + 166, + 328 + ], + "score": 0.85, + "content": "+ 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 317, + 236, + 329 + ], + "score": 1.0, + "content": "Top-5 accuracy).", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 138, + 330, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 138, + 330, + 505, + 345 + ], + "score": 1.0, + "content": "We isolate the contributions of different components of the final model to such downstream", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 141, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "tasks. Interestingly, we find that linear classification accuracy is not always predictive of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "low-data classification accuracy, emphasizing the importance of this metric as a stand-alone", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 363, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 363, + 296, + 379 + ], + "score": 1.0, + "content": "benchmark for unsupervised learning.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 135, + 379, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 135, + 379, + 506, + 393 + ], + "score": 1.0, + "content": "• Finally, we assess the generality of CPC representations by transferring them to a new task", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 142, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "and dataset: object detection on PASCAL-VOC 2007. Consistent with the results from the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 402, + 468, + 415 + ], + "spans": [ + { + "bbox": [ + 142, + 402, + 468, + 415 + ], + "score": 1.0, + "content": "previous section, we find CPC to give state-of-the-art performance in this setting.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + } + ], + "index": 18, + "bbox_fs": [ + 133, + 213, + 506, + 415 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 429, + 244, + 442 + ], + "lines": [ + { + "bbox": [ + 105, + 428, + 245, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 245, + 444 + ], + "score": 1.0, + "content": "2 EXPERIMENTAL SETUP", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 455, + 505, + 477 + ], + "lines": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 468 + ], + "score": 1.0, + "content": "We first review the CPC architecture and learning objective in section 2.1, before detailing how we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 466, + 405, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 405, + 478 + ], + "score": 1.0, + "content": "use its resulting representations for image recognition tasks in section 2.2.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 453, + 505, + 478 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 490, + 286, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 490, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 490, + 287, + 503 + ], + "score": 1.0, + "content": "2.1 CONTRASTIVE PREDICTIVE CODING", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 511, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding as formulated in (van den Oord et al., 2018) learns representations", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "by training neural networks to predict the representations of future observations from those of past", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "ones. When applied to images, the original formulation of CPC operates by predicting the represen-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 556 + ], + "score": 1.0, + "content": "tations of patches below a certain position from those above it (Fig. 2, left). These predictions are", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 569 + ], + "score": 1.0, + "content": "evaluated using a contrastive loss, in which the network must correctly classify the ‘future’ repre-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "sentation amongst a set of unrelated ‘negative’ representations. This avoids trivial solutions such as", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "representing all patches with a constant vector, as would be the case with a mean squared error loss.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 510, + 506, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 683 + ], + "lines": [ + { + "bbox": [ + 104, + 591, + 507, + 609 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 484, + 609 + ], + "score": 1.0, + "content": "In the CPC architecture, each input image is first divided into a set of overlapping patches", + "type": "text" + }, + { + "bbox": [ + 484, + 596, + 501, + 606 + ], + "score": 0.88, + "content": "\\mathbf { \\delta } _ { \\mathbf { x } _ { i , j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 591, + 507, + 609 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 304, + 618 + ], + "score": 1.0, + "content": "each of which is encoded with a neural network", + "type": "text" + }, + { + "bbox": [ + 304, + 605, + 315, + 616 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 604, + 397, + 618 + ], + "score": 1.0, + "content": "into a single vector", + "type": "text" + }, + { + "bbox": [ + 397, + 605, + 461, + 618 + ], + "score": 0.93, + "content": "z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 604, + 506, + 618 + ], + "score": 1.0, + "content": ". To make", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 289, + 629 + ], + "score": 1.0, + "content": "predictions, a masked convolutional network", + "type": "text" + }, + { + "bbox": [ + 290, + 618, + 301, + 628 + ], + "score": 0.87, + "content": "g _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "is then applied to the grid of feature vectors. The", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 399, + 640 + ], + "score": 1.0, + "content": "masks are such that the receptive field of each resulting context vector", + "type": "text" + }, + { + "bbox": [ + 400, + 628, + 415, + 639 + ], + "score": 0.88, + "content": "c _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "only includes feature", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 637, + 506, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 268, + 652 + ], + "score": 1.0, + "content": "vectors that lie above it in the image (i.e.", + "type": "text" + }, + { + "bbox": [ + 268, + 638, + 320, + 650 + ], + "score": 0.92, + "content": "\\{ z _ { u , v } \\} _ { u \\leq i , v } )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 637, + 506, + 652 + ], + "score": 1.0, + "content": ". The prediction task then consists o predicting", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 648, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 648, + 201, + 662 + ], + "score": 1.0, + "content": "‘future’ feature vectors", + "type": "text" + }, + { + "bbox": [ + 202, + 650, + 227, + 661 + ], + "score": 0.9, + "content": "z _ { i + k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 228, + 648, + 346, + 662 + ], + "score": 1.0, + "content": "from current context vectors", + "type": "text" + }, + { + "bbox": [ + 346, + 650, + 361, + 661 + ], + "score": 0.88, + "content": "c _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 648, + 392, + 662 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 393, + 649, + 419, + 659 + ], + "score": 0.9, + "content": "k > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 648, + 506, + 662 + ], + "score": 1.0, + "content": ". The predictions are", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 660, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 260, + 673 + ], + "score": 1.0, + "content": "made linearly: given a context vector", + "type": "text" + }, + { + "bbox": [ + 260, + 662, + 275, + 672 + ], + "score": 0.86, + "content": "c _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 660, + 358, + 673 + ], + "score": 1.0, + "content": ", a prediction length", + "type": "text" + }, + { + "bbox": [ + 358, + 660, + 384, + 670 + ], + "score": 0.9, + "content": "k > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 660, + 486, + 673 + ], + "score": 1.0, + "content": ", and a prediction matrix", + "type": "text" + }, + { + "bbox": [ + 486, + 660, + 501, + 671 + ], + "score": 0.88, + "content": "W _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 660, + 506, + 673 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 671, + 301, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 227, + 685 + ], + "score": 1.0, + "content": "the predicted feature vector is", + "type": "text" + }, + { + "bbox": [ + 228, + 672, + 297, + 683 + ], + "score": 0.91, + "content": "\\hat { \\boldsymbol { z } } _ { i + k , j } = \\boldsymbol { W } _ { k } \\boldsymbol { c } _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 671, + 301, + 685 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5, + "bbox_fs": [ + 104, + 591, + 507, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "The quality of this prediction is then evaluated using a contrastive loss. Specifically, the goal is to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 228, + 711 + ], + "score": 1.0, + "content": "correctly recognize the target", + "type": "text" + }, + { + "bbox": [ + 229, + 700, + 254, + 711 + ], + "score": 0.82, + "content": "z _ { i + k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 699, + 462, + 711 + ], + "score": 1.0, + "content": "among a set of randomly sampled feature vectors", + "type": "text" + }, + { + "bbox": [ + 462, + 699, + 481, + 711 + ], + "score": 0.92, + "content": "\\{ z _ { l } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 482, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "the dataset. We compute the probability assigned to the target using a softmax, and evaluate this", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 734 + ], + "score": 1.0, + "content": "probability using the usual cross-entropy loss. Summing this loss over locations and prediction", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 385, + 427, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 427, + 399 + ], + "score": 1.0, + "content": "offsets, we arrive at the CPC objective as defined in (van den Oord et al., 2018):", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 688, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 83, + 504, + 233 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 83, + 504, + 233 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 83, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 504, + 233 + ], + "score": 0.97, + "type": "image", + "image_path": "e21408d1a55db66ba2cac6a3f4d98b0d1a13c2f1211e3956f93d67bae89bbaae.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 83, + 504, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 133.0, + 504, + 183.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 183.0, + 504, + 233.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 237, + 506, + 369 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 237, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 506, + 249 + ], + "score": 1.0, + "content": "Figure 2: Overview of the framework for semi-supervised learning with Contrastive Predictive", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "Coding. Left: unsupervised pre-training with the spatial prediction task (See Section 2.1). First,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 259, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 272 + ], + "score": 1.0, + "content": "an image is divided into a grid of overlapping patches. Each patch is encoded independently from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "score": 1.0, + "content": "the rest with a feature extractor (blue) which terminates with a mean-pooling operation, yielding a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "single feature vector for that patch. Doing so for all patches yields a field of such feature vectors", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "(wireframe vectors). Feature vectors above a certain level (in this case, the center of the image)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "score": 1.0, + "content": "are then aggregated with a context network (red), yielding a row of context vectors which are used", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "to linearly predict features vectors below. Right: using the CPC representation for a classification", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "task. Having trained the encoder network, the context network (red) is discarded and replaced by a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "classifier network (green) which can be trained in a supervised manner. For some experiments, we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "also fine-tune the encoder network (blue) for the classification task. When applying the encoder to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 358, + 485, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 485, + 370 + ], + "score": 1.0, + "content": "cropped patches (as opposed to the full image) we refer to it as a patched ResNet in the figure.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 426, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 427, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 427, + 399 + ], + "score": 1.0, + "content": "offsets, we arrive at the CPC objective as defined in (van den Oord et al., 2018):", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 114, + 399, + 495, + 433 + ], + "lines": [ + { + "bbox": [ + 114, + 399, + 495, + 433 + ], + "spans": [ + { + "bbox": [ + 114, + 399, + 495, + 433 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { \\mathrm { C P C } } = - \\sum _ { i , j , k } \\log p ( z _ { i + k , j } | \\hat { z } _ { i + k , j } , \\{ z _ { l } \\} ) = - \\sum _ { i , j , k } \\log \\frac { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) } { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) + \\sum _ { l } \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { l } ) }", + "type": "interline_equation", + "image_path": "7113b15debaadf2fc32b7ca0c3e55fb0689eecd2935810917cee824de20f6604.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 399, + 495, + 410.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 114, + 410.3333333333333, + 495, + 421.66666666666663 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 114, + 421.66666666666663, + 495, + 432.99999999999994 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 194, + 447 + ], + "score": 1.0, + "content": "The negative samples", + "type": "text" + }, + { + "bbox": [ + 195, + 435, + 214, + 447 + ], + "score": 0.92, + "content": "\\left\\{ z _ { l } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "are taken from other locations in the image and other images in the mini-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "batch. This loss is called InfoNCE (van den Oord et al., 2018) as it is inspired by Noise-Contrastive", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 197, + 469 + ], + "score": 1.0, + "content": "Estimation (Gutmann", + "type": "text" + }, + { + "bbox": [ + 198, + 457, + 207, + 467 + ], + "score": 0.43, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013) and has been shown to ¨", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 468, + 452, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 278, + 481 + ], + "score": 1.0, + "content": "maximize the mutual information between", + "type": "text" + }, + { + "bbox": [ + 278, + 470, + 294, + 480 + ], + "score": 0.91, + "content": "\\mathbf { } _ { c _ { i , j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 468, + 312, + 481 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 312, + 469, + 339, + 480 + ], + "score": 0.9, + "content": "z _ { i + k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 468, + 452, + 481 + ], + "score": 1.0, + "content": "(van den Oord et al., 2018).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 107, + 492, + 235, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 236, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 236, + 505 + ], + "score": 1.0, + "content": "2.2 EVALUATION PROTOCOL", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 504, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 252, + 525 + ], + "score": 1.0, + "content": "Having trained an encoder network", + "type": "text" + }, + { + "bbox": [ + 252, + 513, + 262, + 524 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 512, + 342, + 525 + ], + "score": 1.0, + "content": ", a context network", + "type": "text" + }, + { + "bbox": [ + 342, + 515, + 353, + 525 + ], + "score": 0.86, + "content": "g _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 512, + 477, + 525 + ], + "score": 1.0, + "content": ", and a set of linear predictors", + "type": "text" + }, + { + "bbox": [ + 477, + 513, + 504, + 525 + ], + "score": 0.91, + "content": "\\{ W _ { k } \\}", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 279, + 536 + ], + "score": 1.0, + "content": "using the CPC objective, we use the latents", + "type": "text" + }, + { + "bbox": [ + 279, + 524, + 323, + 536 + ], + "score": 0.93, + "content": "z = f _ { \\boldsymbol { \\theta } } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 524, + 483, + 536 + ], + "score": 1.0, + "content": "as a representation of new observations", + "type": "text" + }, + { + "bbox": [ + 483, + 527, + 490, + 534 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 360, + 547 + ], + "score": 1.0, + "content": "downstream tasks, and discard the rest. We then train a model", + "type": "text" + }, + { + "bbox": [ + 360, + 535, + 373, + 547 + ], + "score": 0.9, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "to classify these representations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 391, + 559 + ], + "score": 1.0, + "content": "given a dataset of labeled images. More formally, given a dataset of", + "type": "text" + }, + { + "bbox": [ + 392, + 546, + 402, + 556 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 545, + 477, + 559 + ], + "score": 1.0, + "content": "unlabeled images", + "type": "text" + }, + { + "bbox": [ + 478, + 546, + 505, + 557 + ], + "score": 0.89, + "content": "\\mathbb { D } _ { u } =", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 555, + 450, + 571 + ], + "spans": [ + { + "bbox": [ + 107, + 556, + 128, + 569 + ], + "score": 0.91, + "content": "\\left\\{ x _ { n } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 555, + 306, + 571 + ], + "score": 1.0, + "content": ", and a (potentially much smaller) dataset of", + "type": "text" + }, + { + "bbox": [ + 306, + 557, + 317, + 567 + ], + "score": 0.85, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 555, + 380, + 571 + ], + "score": 1.0, + "content": "labeled images", + "type": "text" + }, + { + "bbox": [ + 381, + 556, + 446, + 569 + ], + "score": 0.92, + "content": "\\mathbb { D } _ { l } = \\{ x _ { m } , y _ { m } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 555, + 450, + 571 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 569, + 478, + 604 + ], + "lines": [ + { + "bbox": [ + 133, + 569, + 478, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 569, + 478, + 604 + ], + "score": 0.94, + "content": "\\theta ^ { * } = \\arg \\operatorname* { m i n } _ { \\theta } \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } \\mathcal { L } _ { \\mathrm { C P C } } [ f _ { \\theta } ( x _ { n } ) ] , \\quad \\psi ^ { * } = \\arg \\operatorname* { m i n } _ { \\psi } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { L } _ { \\mathrm { S u p } } [ h _ { \\psi } \\circ f _ { \\theta ^ { * } } ( x _ { m } ) , y _ { m } ]", + "type": "interline_equation", + "image_path": "c6df6158bc3640c41b4e80de8a77924da143cb3d171b2d5421986f303ad176ed.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 133, + 569, + 478, + 580.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 133, + 580.6666666666666, + 478, + 592.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 133, + 592.3333333333333, + 478, + 603.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 295, + 618 + ], + "score": 1.0, + "content": "In all cases, the dataset of unlabeled images", + "type": "text" + }, + { + "bbox": [ + 295, + 605, + 309, + 616 + ], + "score": 0.89, + "content": "\\mathbb { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "we pre-train on is the full ImageNet ILSVRC", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 431, + 628 + ], + "score": 1.0, + "content": "2012 training set (Russakovsky et al., 2015). We consider three labeled datasets", + "type": "text" + }, + { + "bbox": [ + 431, + 616, + 443, + 627 + ], + "score": 0.87, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "for evaluation,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 244, + 640 + ], + "score": 1.0, + "content": "each with an associated classifier", + "type": "text" + }, + { + "bbox": [ + 245, + 627, + 258, + 639 + ], + "score": 0.9, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 627, + 347, + 640 + ], + "score": 1.0, + "content": "and supervised losse", + "type": "text" + }, + { + "bbox": [ + 347, + 627, + 367, + 639 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "(see Fig. 2, right). This protocol", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "is sufficiently generic to allow us to later compare the CPC representation to other methods which", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 648, + 330, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 316, + 662 + ], + "score": 1.0, + "content": "have their own means of learning a feature extractor", + "type": "text" + }, + { + "bbox": [ + 317, + 649, + 326, + 660 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 648, + 330, + 662 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "Linear classification is the standard benchmark for evaluating the quality of unsupervised image", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 336, + 689 + ], + "score": 1.0, + "content": "representations. In this regime, the classification network", + "type": "text" + }, + { + "bbox": [ + 337, + 677, + 350, + 689 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "is restricted to mean pooling followed", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 291, + 701 + ], + "score": 1.0, + "content": "by a single linear layer, and the parameters of", + "type": "text" + }, + { + "bbox": [ + 291, + 688, + 302, + 699 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 687, + 442, + 701 + ], + "score": 1.0, + "content": "are kept fixed. The labeled dataset", + "type": "text" + }, + { + "bbox": [ + 443, + 688, + 455, + 699 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "is the entire", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 280, + 712 + ], + "score": 1.0, + "content": "ImageNet dataset, and the supervised loss", + "type": "text" + }, + { + "bbox": [ + 281, + 700, + 300, + 711 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "is standard cross-entropy. We use the same data-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "augmentation as in the unsupervised learning phase for training, and none at test time and evaluate", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 183, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 183, + 734 + ], + "score": 1.0, + "content": "with a single crop.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 83, + 504, + 233 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 83, + 504, + 233 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 83, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 504, + 233 + ], + "score": 0.97, + "type": "image", + "image_path": "e21408d1a55db66ba2cac6a3f4d98b0d1a13c2f1211e3956f93d67bae89bbaae.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 83, + 504, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 133.0, + 504, + 183.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 183.0, + 504, + 233.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 237, + 506, + 369 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 237, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 506, + 249 + ], + "score": 1.0, + "content": "Figure 2: Overview of the framework for semi-supervised learning with Contrastive Predictive", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "Coding. Left: unsupervised pre-training with the spatial prediction task (See Section 2.1). First,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 259, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 272 + ], + "score": 1.0, + "content": "an image is divided into a grid of overlapping patches. Each patch is encoded independently from", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 506, + 283 + ], + "score": 1.0, + "content": "the rest with a feature extractor (blue) which terminates with a mean-pooling operation, yielding a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "single feature vector for that patch. Doing so for all patches yields a field of such feature vectors", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "(wireframe vectors). Feature vectors above a certain level (in this case, the center of the image)", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 315 + ], + "score": 1.0, + "content": "are then aggregated with a context network (red), yielding a row of context vectors which are used", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "to linearly predict features vectors below. Right: using the CPC representation for a classification", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "task. Having trained the encoder network, the context network (red) is discarded and replaced by a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 349 + ], + "score": 1.0, + "content": "classifier network (green) which can be trained in a supervised manner. For some experiments, we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "also fine-tune the encoder network (blue) for the classification task. When applying the encoder to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 358, + 485, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 485, + 370 + ], + "score": 1.0, + "content": "cropped patches (as opposed to the full image) we refer to it as a patched ResNet in the figure.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 8.5 + } + ], + "index": 4.75 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 426, + 398 + ], + "lines": [], + "index": 15, + "bbox_fs": [ + 106, + 385, + 427, + 399 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 114, + 399, + 495, + 433 + ], + "lines": [ + { + "bbox": [ + 114, + 399, + 495, + 433 + ], + "spans": [ + { + "bbox": [ + 114, + 399, + 495, + 433 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { \\mathrm { C P C } } = - \\sum _ { i , j , k } \\log p ( z _ { i + k , j } | \\hat { z } _ { i + k , j } , \\{ z _ { l } \\} ) = - \\sum _ { i , j , k } \\log \\frac { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) } { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) + \\sum _ { l } \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { l } ) }", + "type": "interline_equation", + "image_path": "7113b15debaadf2fc32b7ca0c3e55fb0689eecd2935810917cee824de20f6604.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 114, + 399, + 495, + 410.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 114, + 410.3333333333333, + 495, + 421.66666666666663 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 114, + 421.66666666666663, + 495, + 432.99999999999994 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 194, + 447 + ], + "score": 1.0, + "content": "The negative samples", + "type": "text" + }, + { + "bbox": [ + 195, + 435, + 214, + 447 + ], + "score": 0.92, + "content": "\\left\\{ z _ { l } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "are taken from other locations in the image and other images in the mini-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 459 + ], + "score": 1.0, + "content": "batch. This loss is called InfoNCE (van den Oord et al., 2018) as it is inspired by Noise-Contrastive", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 197, + 469 + ], + "score": 1.0, + "content": "Estimation (Gutmann", + "type": "text" + }, + { + "bbox": [ + 198, + 457, + 207, + 467 + ], + "score": 0.43, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013) and has been shown to ¨", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 468, + 452, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 278, + 481 + ], + "score": 1.0, + "content": "maximize the mutual information between", + "type": "text" + }, + { + "bbox": [ + 278, + 470, + 294, + 480 + ], + "score": 0.91, + "content": "\\mathbf { } _ { c _ { i , j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 468, + 312, + 481 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 312, + 469, + 339, + 480 + ], + "score": 0.9, + "content": "z _ { i + k , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 468, + 452, + 481 + ], + "score": 1.0, + "content": "(van den Oord et al., 2018).", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 434, + 506, + 481 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 492, + 235, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 236, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 236, + 505 + ], + "score": 1.0, + "content": "2.2 EVALUATION PROTOCOL", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 504, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 252, + 525 + ], + "score": 1.0, + "content": "Having trained an encoder network", + "type": "text" + }, + { + "bbox": [ + 252, + 513, + 262, + 524 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 512, + 342, + 525 + ], + "score": 1.0, + "content": ", a context network", + "type": "text" + }, + { + "bbox": [ + 342, + 515, + 353, + 525 + ], + "score": 0.86, + "content": "g _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 512, + 477, + 525 + ], + "score": 1.0, + "content": ", and a set of linear predictors", + "type": "text" + }, + { + "bbox": [ + 477, + 513, + 504, + 525 + ], + "score": 0.91, + "content": "\\{ W _ { k } \\}", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 279, + 536 + ], + "score": 1.0, + "content": "using the CPC objective, we use the latents", + "type": "text" + }, + { + "bbox": [ + 279, + 524, + 323, + 536 + ], + "score": 0.93, + "content": "z = f _ { \\boldsymbol { \\theta } } ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 524, + 483, + 536 + ], + "score": 1.0, + "content": "as a representation of new observations", + "type": "text" + }, + { + "bbox": [ + 483, + 527, + 490, + 534 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 490, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 360, + 547 + ], + "score": 1.0, + "content": "downstream tasks, and discard the rest. We then train a model", + "type": "text" + }, + { + "bbox": [ + 360, + 535, + 373, + 547 + ], + "score": 0.9, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "to classify these representations", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 391, + 559 + ], + "score": 1.0, + "content": "given a dataset of labeled images. More formally, given a dataset of", + "type": "text" + }, + { + "bbox": [ + 392, + 546, + 402, + 556 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 545, + 477, + 559 + ], + "score": 1.0, + "content": "unlabeled images", + "type": "text" + }, + { + "bbox": [ + 478, + 546, + 505, + 557 + ], + "score": 0.89, + "content": "\\mathbb { D } _ { u } =", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 555, + 450, + 571 + ], + "spans": [ + { + "bbox": [ + 107, + 556, + 128, + 569 + ], + "score": 0.91, + "content": "\\left\\{ x _ { n } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 555, + 306, + 571 + ], + "score": 1.0, + "content": ", and a (potentially much smaller) dataset of", + "type": "text" + }, + { + "bbox": [ + 306, + 557, + 317, + 567 + ], + "score": 0.85, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 555, + 380, + 571 + ], + "score": 1.0, + "content": "labeled images", + "type": "text" + }, + { + "bbox": [ + 381, + 556, + 446, + 569 + ], + "score": 0.92, + "content": "\\mathbb { D } _ { l } = \\{ x _ { m } , y _ { m } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 555, + 450, + 571 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 512, + 505, + 571 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 569, + 478, + 604 + ], + "lines": [ + { + "bbox": [ + 133, + 569, + 478, + 604 + ], + "spans": [ + { + "bbox": [ + 133, + 569, + 478, + 604 + ], + "score": 0.94, + "content": "\\theta ^ { * } = \\arg \\operatorname* { m i n } _ { \\theta } \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } \\mathcal { L } _ { \\mathrm { C P C } } [ f _ { \\theta } ( x _ { n } ) ] , \\quad \\psi ^ { * } = \\arg \\operatorname* { m i n } _ { \\psi } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { L } _ { \\mathrm { S u p } } [ h _ { \\psi } \\circ f _ { \\theta ^ { * } } ( x _ { m } ) , y _ { m } ]", + "type": "interline_equation", + "image_path": "c6df6158bc3640c41b4e80de8a77924da143cb3d171b2d5421986f303ad176ed.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 133, + 569, + 478, + 580.6666666666666 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 133, + 580.6666666666666, + 478, + 592.3333333333333 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 133, + 592.3333333333333, + 478, + 603.9999999999999 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 605, + 505, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 295, + 618 + ], + "score": 1.0, + "content": "In all cases, the dataset of unlabeled images", + "type": "text" + }, + { + "bbox": [ + 295, + 605, + 309, + 616 + ], + "score": 0.89, + "content": "\\mathbb { D } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "we pre-train on is the full ImageNet ILSVRC", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 431, + 628 + ], + "score": 1.0, + "content": "2012 training set (Russakovsky et al., 2015). We consider three labeled datasets", + "type": "text" + }, + { + "bbox": [ + 431, + 616, + 443, + 627 + ], + "score": 0.87, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "for evaluation,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 627, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 244, + 640 + ], + "score": 1.0, + "content": "each with an associated classifier", + "type": "text" + }, + { + "bbox": [ + 245, + 627, + 258, + 639 + ], + "score": 0.9, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 627, + 347, + 640 + ], + "score": 1.0, + "content": "and supervised losse", + "type": "text" + }, + { + "bbox": [ + 347, + 627, + 367, + 639 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 627, + 506, + 640 + ], + "score": 1.0, + "content": "(see Fig. 2, right). This protocol", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 104, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "is sufficiently generic to allow us to later compare the CPC representation to other methods which", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 648, + 330, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 316, + 662 + ], + "score": 1.0, + "content": "have their own means of learning a feature extractor", + "type": "text" + }, + { + "bbox": [ + 317, + 649, + 326, + 660 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 648, + 330, + 662 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34, + "bbox_fs": [ + 104, + 604, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 104, + 664, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 104, + 664, + 505, + 679 + ], + "score": 1.0, + "content": "Linear classification is the standard benchmark for evaluating the quality of unsupervised image", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 104, + 677, + 336, + 689 + ], + "score": 1.0, + "content": "representations. In this regime, the classification network", + "type": "text" + }, + { + "bbox": [ + 337, + 677, + 350, + 689 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "is restricted to mean pooling followed", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 291, + 701 + ], + "score": 1.0, + "content": "by a single linear layer, and the parameters of", + "type": "text" + }, + { + "bbox": [ + 291, + 688, + 302, + 699 + ], + "score": 0.88, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 687, + 442, + 701 + ], + "score": 1.0, + "content": "are kept fixed. The labeled dataset", + "type": "text" + }, + { + "bbox": [ + 443, + 688, + 455, + 699 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "is the entire", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 280, + 712 + ], + "score": 1.0, + "content": "ImageNet dataset, and the supervised loss", + "type": "text" + }, + { + "bbox": [ + 281, + 700, + 300, + 711 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "is standard cross-entropy. We use the same data-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "augmentation as in the unsupervised learning phase for training, and none at test time and evaluate", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 183, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 183, + 734 + ], + "score": 1.0, + "content": "with a single crop.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 104, + 664, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Efficient classification directly tests whether the CPC representation enables visual learning from", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 262, + 106 + ], + "score": 1.0, + "content": "few labels. For this task, the classifier", + "type": "text" + }, + { + "bbox": [ + 262, + 94, + 275, + 106 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 92, + 505, + 106 + ], + "score": 1.0, + "content": "is an arbitrary deep neural network (we use an 11-block", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "ResNet architecture with 4096-dimensional feature maps and 1024-dimensional bottleneck layers).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 185, + 128 + ], + "score": 1.0, + "content": "The labeled dataset", + "type": "text" + }, + { + "bbox": [ + 185, + 116, + 196, + 126 + ], + "score": 0.87, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 115, + 428, + 128 + ], + "score": 1.0, + "content": "is a subset of the ImageNet dataset: we investigated using", + "type": "text" + }, + { + "bbox": [ + 428, + 115, + 442, + 126 + ], + "score": 0.76, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 115, + 445, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 445, + 115, + 460, + 126 + ], + "score": 0.77, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 115, + 463, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 463, + 115, + 478, + 126 + ], + "score": 0.78, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 115, + 482, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 482, + 115, + 501, + 126 + ], + "score": 0.83, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 115, + 505, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 125, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 126, + 137 + ], + "score": 0.84, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 125, + 129, + 140 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 130, + 126, + 150, + 137 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 125, + 169, + 140 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 169, + 127, + 193, + 137 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 125, + 379, + 140 + ], + "score": 1.0, + "content": "of the ImageNet dataset. The supervised loss", + "type": "text" + }, + { + "bbox": [ + 379, + 127, + 399, + 138 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 125, + 506, + 140 + ], + "score": 1.0, + "content": "is again cross-entropy. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "score": 1.0, + "content": "addition to random color-dropping we use the Inception data-augmentation scheme (Szegedy et al.,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 426, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 426, + 162 + ], + "score": 1.0, + "content": "2014) for training, no augmentation at test-time and evaluate with a single crop.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Transfer learning tests the generality of the representation by applying it to a new task and dataset.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "For this we chose image detection on the PASCAL-2007 dataset, a standard benchmark in computer", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 284, + 200 + ], + "score": 1.0, + "content": "vision (Everingham et al., 2007). As such", + "type": "text" + }, + { + "bbox": [ + 285, + 187, + 296, + 198 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "is the entire PASCAL-2007 dataset (comprised of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 196, + 212 + ], + "score": 1.0, + "content": "5011 labeled images);", + "type": "text" + }, + { + "bbox": [ + 197, + 198, + 209, + 210 + ], + "score": 0.88, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 197, + 228, + 212 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 199, + 248, + 210 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 197, + 506, + 212 + ], + "score": 1.0, + "content": "are the Faster-RCNN architecture and loss (Ren et al., 2015). In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 466, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 466, + 223 + ], + "score": 1.0, + "content": "addition to color-dropping, we use scale-augmentation (Doersch et al., 2015) for training.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 292 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 331, + 239 + ], + "score": 1.0, + "content": "For linear classification, we keep the feature extractor", + "type": "text" + }, + { + "bbox": [ + 332, + 226, + 342, + 238 + ], + "score": 0.34, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "fixed to assess the representation in ab-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 251 + ], + "score": 1.0, + "content": "solute terms. For efficient classification and transfer learning, we additionally explore fine-tuning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "the feature extractor for the supervised objective. In this regime, we initialize the feature extractor", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 234, + 271 + ], + "score": 1.0, + "content": "and classifier with the solutions", + "type": "text" + }, + { + "bbox": [ + 234, + 259, + 260, + 270 + ], + "score": 0.4, + "content": "\\theta ^ { * } , \\psi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 259, + 506, + 271 + ], + "score": 1.0, + "content": "found in the previous learning phase, and train them both for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 270, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 505, + 281 + ], + "score": 1.0, + "content": "the supervised objective. To ensure that the feature extractor does not deviate too much from the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 280, + 465, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 465, + 294 + ], + "score": 1.0, + "content": "solution dictated by the CPC objective, we use a smaller learning rate and early-stopping.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 319, + 210, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 213, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 213, + 334 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 504, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "Data-efficient learning has typically been approached by two complementary methods, both of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "which seek to make use of more plentiful unlabeled data: representation learning and semi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 452, + 386 + ], + "score": 1.0, + "content": "supervised learning. The former formulates an objective to learn a feature extractor", + "type": "text" + }, + { + "bbox": [ + 453, + 374, + 463, + 385 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 375, + 505, + 386 + ], + "score": 1.0, + "content": "in an un-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 385, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 387, + 398 + ], + "score": 1.0, + "content": "supervised manner, whereas the latter directly constrains the classifier", + "type": "text" + }, + { + "bbox": [ + 387, + 385, + 400, + 397 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 385, + 504, + 398 + ], + "score": 1.0, + "content": "using the unlabeled data.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 504, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "Representation learning saw early success using generative modeling (Kingma et al., 2014), but", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "likelihood-based models have yet to generalize to more complex stimulus classes. Generative ad-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "versarial models have also been harnessed for representation learning (Donahue et al., 2016), and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "large-scale implementations have recently achieved corresponding gains in linear classification ac-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 446, + 261, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 261, + 459 + ], + "score": 1.0, + "content": "curacy (Donahue & Simonyan, 2019).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "In contrast to generative models which require the reconstruction of observations, self-supervised", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "score": 1.0, + "content": "techniques directly formulate tasks involving the learned representation. For example, simply ask-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "ing a network to recognize the spatial layout of an image led to representations that transferred to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "popular vision tasks such as classification and detection (Doersch et al., 2015; Noroozi & Favaro,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "2016). Other works showed that prediction of color (Zhang et al., 2016; Larsson et al., 2017) and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "image orientation (Gidaris et al., 2018), and invariance to data augmentation (Dosovitskiy et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "2014) can provide useful self-supervised tasks. Beyond single images, works have leveraged video", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "cues such as object tracking (Wang & Gupta, 2015), frame ordering (Misra et al., 2016), and object", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "boundary cues (Li et al., 2016; Pathak et al., 2016). Non-visual information can be equally power-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 104, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "ful; information about camera motion (Agrawal et al., 2015; Jayaraman & Grauman, 2015), scene", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "geometry (Zamir et al., 2016), or sound (Arandjelovic & Zisserman, 2017; 2018) can all serve as", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 584, + 230, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 230, + 595 + ], + "score": 1.0, + "content": "natural sources of supervision.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "While many of these tasks require predicting fixed quantities computed from the data, another class", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "of contrastive methods formulate their objectives in the learned representations themselves. CPC is", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "a contrastive representation learning method that maximizes the mutual information between spa-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "tially removed latent representations with InfoNCE (van den Oord et al., 2018), a loss function based", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 645, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 655 + ], + "score": 1.0, + "content": "on Noise-Contrastive Estimation (Gutmann & Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013). Two ¨", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "other methods have recently been proposed using the same loss function, but with different associ-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "ated prediction tasks. Contrastive Multiview Coding (Tian et al., 2019) maximizes the mutual infor-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "mation between representations of different views of the same observation. Augmented Multiscale", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "Deep InfoMax (AMDIM, Bachman et al., 2019) is most similar to CPC in that it makes predic-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "tions across space, but differs in that it also predicts representations across layers in the model. In", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "addition, AMDIM limits the receptive field of its representation, but does this by constraining the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 721, + 473, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 473, + 733 + ], + "score": 1.0, + "content": "number of spatial convolutions in the network architecture rather than using image patches.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Efficient classification directly tests whether the CPC representation enables visual learning from", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 262, + 106 + ], + "score": 1.0, + "content": "few labels. For this task, the classifier", + "type": "text" + }, + { + "bbox": [ + 262, + 94, + 275, + 106 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 92, + 505, + 106 + ], + "score": 1.0, + "content": "is an arbitrary deep neural network (we use an 11-block", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "ResNet architecture with 4096-dimensional feature maps and 1024-dimensional bottleneck layers).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 185, + 128 + ], + "score": 1.0, + "content": "The labeled dataset", + "type": "text" + }, + { + "bbox": [ + 185, + 116, + 196, + 126 + ], + "score": 0.87, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 115, + 428, + 128 + ], + "score": 1.0, + "content": "is a subset of the ImageNet dataset: we investigated using", + "type": "text" + }, + { + "bbox": [ + 428, + 115, + 442, + 126 + ], + "score": 0.76, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 115, + 445, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 445, + 115, + 460, + 126 + ], + "score": 0.77, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 115, + 463, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 463, + 115, + 478, + 126 + ], + "score": 0.78, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 115, + 482, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 482, + 115, + 501, + 126 + ], + "score": 0.83, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 115, + 505, + 128 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 125, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 126, + 137 + ], + "score": 0.84, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 125, + 129, + 140 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 130, + 126, + 150, + 137 + ], + "score": 0.86, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 125, + 169, + 140 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 169, + 127, + 193, + 137 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 125, + 379, + 140 + ], + "score": 1.0, + "content": "of the ImageNet dataset. The supervised loss", + "type": "text" + }, + { + "bbox": [ + 379, + 127, + 399, + 138 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 125, + 506, + 140 + ], + "score": 1.0, + "content": "is again cross-entropy. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 151 + ], + "score": 1.0, + "content": "addition to random color-dropping we use the Inception data-augmentation scheme (Szegedy et al.,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 426, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 426, + 162 + ], + "score": 1.0, + "content": "2014) for training, no augmentation at test-time and evaluate with a single crop.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 82, + 506, + 162 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "Transfer learning tests the generality of the representation by applying it to a new task and dataset.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "For this we chose image detection on the PASCAL-2007 dataset, a standard benchmark in computer", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 284, + 200 + ], + "score": 1.0, + "content": "vision (Everingham et al., 2007). As such", + "type": "text" + }, + { + "bbox": [ + 285, + 187, + 296, + 198 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "is the entire PASCAL-2007 dataset (comprised of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 196, + 212 + ], + "score": 1.0, + "content": "5011 labeled images);", + "type": "text" + }, + { + "bbox": [ + 197, + 198, + 209, + 210 + ], + "score": 0.88, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 197, + 228, + 212 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 228, + 199, + 248, + 210 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { \\mathrm { { S u p } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 197, + 506, + 212 + ], + "score": 1.0, + "content": "are the Faster-RCNN architecture and loss (Ren et al., 2015). In", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 466, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 466, + 223 + ], + "score": 1.0, + "content": "addition to color-dropping, we use scale-augmentation (Doersch et al., 2015) for training.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 165, + 506, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 292 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 331, + 239 + ], + "score": 1.0, + "content": "For linear classification, we keep the feature extractor", + "type": "text" + }, + { + "bbox": [ + 332, + 226, + 342, + 238 + ], + "score": 0.34, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "fixed to assess the representation in ab-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 251 + ], + "score": 1.0, + "content": "solute terms. For efficient classification and transfer learning, we additionally explore fine-tuning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 259 + ], + "score": 1.0, + "content": "the feature extractor for the supervised objective. In this regime, we initialize the feature extractor", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 234, + 271 + ], + "score": 1.0, + "content": "and classifier with the solutions", + "type": "text" + }, + { + "bbox": [ + 234, + 259, + 260, + 270 + ], + "score": 0.4, + "content": "\\theta ^ { * } , \\psi ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 259, + 506, + 271 + ], + "score": 1.0, + "content": "found in the previous learning phase, and train them both for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 270, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 505, + 281 + ], + "score": 1.0, + "content": "the supervised objective. To ensure that the feature extractor does not deviate too much from the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 280, + 465, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 465, + 294 + ], + "score": 1.0, + "content": "solution dictated by the CPC objective, we use a smaller learning rate and early-stopping.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 226, + 506, + 294 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 319, + 210, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 213, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 213, + 334 + ], + "score": 1.0, + "content": "3 RELATED WORK", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 504, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "Data-efficient learning has typically been approached by two complementary methods, both of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "which seek to make use of more plentiful unlabeled data: representation learning and semi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 452, + 386 + ], + "score": 1.0, + "content": "supervised learning. The former formulates an objective to learn a feature extractor", + "type": "text" + }, + { + "bbox": [ + 453, + 374, + 463, + 385 + ], + "score": 0.87, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 375, + 505, + 386 + ], + "score": 1.0, + "content": "in an un-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 385, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 387, + 398 + ], + "score": 1.0, + "content": "supervised manner, whereas the latter directly constrains the classifier", + "type": "text" + }, + { + "bbox": [ + 387, + 385, + 400, + 397 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 385, + 504, + 398 + ], + "score": 1.0, + "content": "using the unlabeled data.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 352, + 506, + 398 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 402, + 504, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "Representation learning saw early success using generative modeling (Kingma et al., 2014), but", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "likelihood-based models have yet to generalize to more complex stimulus classes. Generative ad-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 424, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 436 + ], + "score": 1.0, + "content": "versarial models have also been harnessed for representation learning (Donahue et al., 2016), and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 448 + ], + "score": 1.0, + "content": "large-scale implementations have recently achieved corresponding gains in linear classification ac-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 446, + 261, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 261, + 459 + ], + "score": 1.0, + "content": "curacy (Donahue & Simonyan, 2019).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 402, + 506, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 476 + ], + "score": 1.0, + "content": "In contrast to generative models which require the reconstruction of observations, self-supervised", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 487 + ], + "score": 1.0, + "content": "techniques directly formulate tasks involving the learned representation. For example, simply ask-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "ing a network to recognize the spatial layout of an image led to representations that transferred to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "popular vision tasks such as classification and detection (Doersch et al., 2015; Noroozi & Favaro,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 519 + ], + "score": 1.0, + "content": "2016). Other works showed that prediction of color (Zhang et al., 2016; Larsson et al., 2017) and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 530 + ], + "score": 1.0, + "content": "image orientation (Gidaris et al., 2018), and invariance to data augmentation (Dosovitskiy et al.,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "2014) can provide useful self-supervised tasks. Beyond single images, works have leveraged video", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "cues such as object tracking (Wang & Gupta, 2015), frame ordering (Misra et al., 2016), and object", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "boundary cues (Li et al., 2016; Pathak et al., 2016). Non-visual information can be equally power-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 104, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "ful; information about camera motion (Agrawal et al., 2015; Jayaraman & Grauman, 2015), scene", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "geometry (Zamir et al., 2016), or sound (Arandjelovic & Zisserman, 2017; 2018) can all serve as", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 584, + 230, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 230, + 595 + ], + "score": 1.0, + "content": "natural sources of supervision.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 462, + 506, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "While many of these tasks require predicting fixed quantities computed from the data, another class", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "of contrastive methods formulate their objectives in the learned representations themselves. CPC is", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "a contrastive representation learning method that maximizes the mutual information between spa-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "tially removed latent representations with InfoNCE (van den Oord et al., 2018), a loss function based", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 645, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 655 + ], + "score": 1.0, + "content": "on Noise-Contrastive Estimation (Gutmann & Hyvarinen, 2010; Mnih & Kavukcuoglu, 2013). Two ¨", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "other methods have recently been proposed using the same loss function, but with different associ-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "ated prediction tasks. Contrastive Multiview Coding (Tian et al., 2019) maximizes the mutual infor-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "mation between representations of different views of the same observation. Augmented Multiscale", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "Deep InfoMax (AMDIM, Bachman et al., 2019) is most similar to CPC in that it makes predic-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "tions across space, but differs in that it also predicts representations across layers in the model. In", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "addition, AMDIM limits the receptive field of its representation, but does this by constraining the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 721, + 473, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 473, + 733 + ], + "score": 1.0, + "content": "number of spatial convolutions in the network architecture rather than using image patches.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 599, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 94 + ], + "score": 1.0, + "content": "A common alternative approach for improving data efficiency is label-propagation (Zhu & Ghahra-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "mani, 2002), where a classifier is trained on a subset of labeled data, then used to label parts of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "unlabeled dataset, after which the process is repeated. This label-propagation can either be discrete", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "(as in pseudo-labeling, Lee, 2013) or continuous (as in entropy minimization, Grandvalet & Bengio,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "2005). The predictions of this classifier are often constrained to be smooth with respect to cer-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 504, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 504, + 150 + ], + "score": 1.0, + "content": "tain deformations, such as data-augmentation (Xie et al., 2019) or adversarial perturbation (Miyato", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "et al., 2018). Representation learning and semi-supervised learning have been shown to be comple-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "mentary and can be combined to great effect (Zhai et al., 2019), which is why we focus solely on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 253, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 253, + 183 + ], + "score": 1.0, + "content": "representation learning in this paper.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 199, + 172, + 212 + ], + "lines": [ + { + "bbox": [ + 104, + 197, + 174, + 214 + ], + "spans": [ + { + "bbox": [ + 104, + 197, + 174, + 214 + ], + "score": 1.0, + "content": "4 RESULTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "When asking whether CPC enables data-efficient learning, we wish to use the best possible represen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "tative of this model class. Unfortunately, purely unsupervised metrics tell us little about downstream", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "performance, and implementation details have been shown to matter enormously (Doersch & Zisser-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "man, 2017; Kolesnikov et al., 2019). Since many design choices (e.g. network architecture and data-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "preprocessing) have been previously evaluated using linear classification, we use this benchmark", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "in section 4.1 to align the CPC model with best practices in representation learning and compare to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 290, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 506, + 304 + ], + "score": 1.0, + "content": "published results. In section 4.2 we select the best performing model from the previous section and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "assess whether it enables efficient classification. We also investigate to what extent the first, more", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 325 + ], + "score": 1.0, + "content": "common metric (linear classification accuracy) is predictive of efficient classification. Finally, in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "section 4.3 we investigate the generality of our results through transfer learning to PASCAL-2007.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 350, + 245, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 247, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 247, + 363 + ], + "score": 1.0, + "content": "4.1 FROM CPC V1 TO CPC V2", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 383 + ], + "score": 1.0, + "content": "The overarching principle behind our new model design is to increase the scale and efficiency of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "score": 1.0, + "content": "encoder architecture while also maximizing the supervisory signal we obtain from each image. At", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 394, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 405 + ], + "score": 1.0, + "content": "the same time, it is important not to allow the network to solve the problem trivially, i.e., without", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "learning semantics. To this end, we seek to remove low-level cues common across patches by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "augmenting individual patches independently, using standard stochastic data-processing techniques", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 291, + 441 + ], + "score": 1.0, + "content": "from supervised and self-supervised learning.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 297, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 297, + 456 + ], + "score": 1.0, + "content": "We identify four axes for model capacity and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 455, + 297, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 297, + 466 + ], + "score": 1.0, + "content": "task setup that could impact the model’s perfor-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 297, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 297, + 478 + ], + "score": 1.0, + "content": "mance. The first axis increases model capacity", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 476, + 297, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 297, + 488 + ], + "score": 1.0, + "content": "by increasing depth and width, while the second", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 488, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 296, + 499 + ], + "score": 1.0, + "content": "improves training efficiency capacity by intro-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 297, + 510 + ], + "score": 1.0, + "content": "ducing layer normalization. The third axis in-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 509, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 297, + 521 + ], + "score": 1.0, + "content": "creases task complexity by making predictions", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 519, + 297, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 297, + 532 + ], + "score": 1.0, + "content": "in all four directions, and the fourth does so by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 531, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 297, + 544 + ], + "score": 1.0, + "content": "performing more extensive patch-based aug-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 542, + 154, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 154, + 553 + ], + "score": 1.0, + "content": "mentation.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 296, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 297, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 297, + 578 + ], + "score": 1.0, + "content": "Model capacity. Recent work has shown that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 577, + 297, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 297, + 591 + ], + "score": 1.0, + "content": "networks and more effective training improves", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 589, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 297, + 601 + ], + "score": 1.0, + "content": "self-supervised learning (Doersch & Zisser-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 600, + 297, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 297, + 612 + ], + "score": 1.0, + "content": "man, 2017; Kolesnikov et al., 2019), but the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 611, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 297, + 622 + ], + "score": 1.0, + "content": "original CPC model used only the first 3 stacks", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 622, + 297, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 297, + 633 + ], + "score": 1.0, + "content": "of a ResNet-101 (He et al., 2016a) architecture", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 633, + 297, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 297, + 644 + ], + "score": 1.0, + "content": "(i.e. a ResNet-92). Therefore, we converted the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 643, + 297, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 297, + 655 + ], + "score": 1.0, + "content": "third residual stack of ResNet-101 (originally", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 655, + 297, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 297, + 667 + ], + "score": 1.0, + "content": "containing 23 blocks, 1024-dimensional feature", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "score": 1.0, + "content": "maps, and 256-dimensional bottleneck layers),", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 677, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 297, + 688 + ], + "score": 1.0, + "content": "to use 46 blocks, with 4096-dimensional fea-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 687, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 297, + 700 + ], + "score": 1.0, + "content": "ture maps and 512-dimensional bottleneck lay-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 699, + 297, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 297, + 710 + ], + "score": 1.0, + "content": "ers. We call the resulting network ResNet-161.", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 709, + 297, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 297, + 721 + ], + "score": 1.0, + "content": "Consistent with prior results, this new architec-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 720, + 298, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 298, + 733 + ], + "score": 1.0, + "content": "ture delivers better performance regardless of", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 307, + 449, + 501, + 616 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 307, + 449, + 501, + 616 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 501, + 616 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 501, + 616 + ], + "score": 0.97, + "type": "image", + "image_path": "3eeded7c7663fc17771caa90292cc1343662124f1094d06f22d9ca255036f2a3.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 307, + 449, + 501, + 461.84615384615387 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 307, + 461.84615384615387, + 501, + 474.69230769230774 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 307, + 474.69230769230774, + 501, + 487.5384615384616 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 307, + 487.5384615384616, + 501, + 500.3846153846155 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 307, + 500.3846153846155, + 501, + 513.2307692307693 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 307, + 513.2307692307693, + 501, + 526.0769230769231 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 307, + 526.0769230769231, + 501, + 538.9230769230769 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 307, + 538.9230769230769, + 501, + 551.7692307692307 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 307, + 551.7692307692307, + 501, + 564.6153846153845 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 307, + 564.6153846153845, + 501, + 577.4615384615383 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 307, + 577.4615384615383, + 501, + 590.3076923076922 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 307, + 590.3076923076922, + 501, + 603.153846153846 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 307, + 603.153846153846, + 501, + 615.9999999999998 + ], + "spans": [], + "index": 64 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 620, + 504, + 719 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 506, + 632 + ], + "score": 1.0, + "content": "Figure 3: Linear classification performance of", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 303, + 631, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 303, + 631, + 506, + 643 + ], + "score": 1.0, + "content": "new variants of CPC, which incrementally add a", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "series of modifications. BU: bottum up spatial", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 653, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 303, + 653, + 505, + 665 + ], + "score": 1.0, + "content": "predictions. HF: randomly flipping patches hor-", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 664, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 505, + 675 + ], + "score": 1.0, + "content": "izontally. LN: layer normalization. RC: random", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 303, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 303, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "color-dropping. TL: tuned prediction lengths. HP:", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 686, + 505, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 686, + 505, + 697 + ], + "score": 1.0, + "content": "horizontal spatial predictions. We use color to in-", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 697, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "dicate the number of spatial predictions used (or-", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 708, + 478, + 719 + ], + "spans": [ + { + "bbox": [ + 303, + 708, + 478, + 719 + ], + "score": 1.0, + "content": "ange, green, blue for 1, 2 and 4 directions).", + "type": "text" + } + ], + "index": 73 + } + ], + "index": 69 + } + ], + "index": 56.0 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 504, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 504, + 94 + ], + "score": 1.0, + "content": "A common alternative approach for improving data efficiency is label-propagation (Zhu & Ghahra-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "mani, 2002), where a classifier is trained on a subset of labeled data, then used to label parts of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "unlabeled dataset, after which the process is repeated. This label-propagation can either be discrete", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "(as in pseudo-labeling, Lee, 2013) or continuous (as in entropy minimization, Grandvalet & Bengio,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "2005). The predictions of this classifier are often constrained to be smooth with respect to cer-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 138, + 504, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 504, + 150 + ], + "score": 1.0, + "content": "tain deformations, such as data-augmentation (Xie et al., 2019) or adversarial perturbation (Miyato", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "et al., 2018). Representation learning and semi-supervised learning have been shown to be comple-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "mentary and can be combined to great effect (Zhai et al., 2019), which is why we focus solely on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 171, + 253, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 253, + 183 + ], + "score": 1.0, + "content": "representation learning in this paper.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 82, + 505, + 183 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 199, + 172, + 212 + ], + "lines": [ + { + "bbox": [ + 104, + 197, + 174, + 214 + ], + "spans": [ + { + "bbox": [ + 104, + 197, + 174, + 214 + ], + "score": 1.0, + "content": "4 RESULTS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "When asking whether CPC enables data-efficient learning, we wish to use the best possible represen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 505, + 249 + ], + "score": 1.0, + "content": "tative of this model class. Unfortunately, purely unsupervised metrics tell us little about downstream", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 505, + 259 + ], + "score": 1.0, + "content": "performance, and implementation details have been shown to matter enormously (Doersch & Zisser-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "man, 2017; Kolesnikov et al., 2019). Since many design choices (e.g. network architecture and data-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 281 + ], + "score": 1.0, + "content": "preprocessing) have been previously evaluated using linear classification, we use this benchmark", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "in section 4.1 to align the CPC model with best practices in representation learning and compare to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 290, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 506, + 304 + ], + "score": 1.0, + "content": "published results. In section 4.2 we select the best performing model from the previous section and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 315 + ], + "score": 1.0, + "content": "assess whether it enables efficient classification. We also investigate to what extent the first, more", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 314, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 325 + ], + "score": 1.0, + "content": "common metric (linear classification accuracy) is predictive of efficient classification. Finally, in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "section 4.3 we investigate the generality of our results through transfer learning to PASCAL-2007.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 225, + 506, + 336 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 350, + 245, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 348, + 247, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 247, + 363 + ], + "score": 1.0, + "content": "4.1 FROM CPC V1 TO CPC V2", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 372, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 505, + 383 + ], + "score": 1.0, + "content": "The overarching principle behind our new model design is to increase the scale and efficiency of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 383, + 505, + 394 + ], + "score": 1.0, + "content": "encoder architecture while also maximizing the supervisory signal we obtain from each image. At", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 394, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 405 + ], + "score": 1.0, + "content": "the same time, it is important not to allow the network to solve the problem trivially, i.e., without", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "learning semantics. To this end, we seek to remove low-level cues common across patches by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "augmenting individual patches independently, using standard stochastic data-processing techniques", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 291, + 441 + ], + "score": 1.0, + "content": "from supervised and self-supervised learning.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 372, + 505, + 441 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 297, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 442, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 297, + 456 + ], + "score": 1.0, + "content": "We identify four axes for model capacity and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 455, + 297, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 297, + 466 + ], + "score": 1.0, + "content": "task setup that could impact the model’s perfor-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 465, + 297, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 297, + 478 + ], + "score": 1.0, + "content": "mance. The first axis increases model capacity", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 476, + 297, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 297, + 488 + ], + "score": 1.0, + "content": "by increasing depth and width, while the second", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 488, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 296, + 499 + ], + "score": 1.0, + "content": "improves training efficiency capacity by intro-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 297, + 510 + ], + "score": 1.0, + "content": "ducing layer normalization. The third axis in-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 509, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 297, + 521 + ], + "score": 1.0, + "content": "creases task complexity by making predictions", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 519, + 297, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 297, + 532 + ], + "score": 1.0, + "content": "in all four directions, and the fourth does so by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 531, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 297, + 544 + ], + "score": 1.0, + "content": "performing more extensive patch-based aug-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 542, + 154, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 154, + 553 + ], + "score": 1.0, + "content": "mentation.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 442, + 297, + 553 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 296, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 297, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 297, + 578 + ], + "score": 1.0, + "content": "Model capacity. Recent work has shown that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 577, + 297, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 297, + 591 + ], + "score": 1.0, + "content": "networks and more effective training improves", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 589, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 297, + 601 + ], + "score": 1.0, + "content": "self-supervised learning (Doersch & Zisser-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 600, + 297, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 297, + 612 + ], + "score": 1.0, + "content": "man, 2017; Kolesnikov et al., 2019), but the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 611, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 297, + 622 + ], + "score": 1.0, + "content": "original CPC model used only the first 3 stacks", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 622, + 297, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 297, + 633 + ], + "score": 1.0, + "content": "of a ResNet-101 (He et al., 2016a) architecture", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 633, + 297, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 297, + 644 + ], + "score": 1.0, + "content": "(i.e. a ResNet-92). Therefore, we converted the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 643, + 297, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 297, + 655 + ], + "score": 1.0, + "content": "third residual stack of ResNet-101 (originally", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 106, + 655, + 297, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 297, + 667 + ], + "score": 1.0, + "content": "containing 23 blocks, 1024-dimensional feature", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 297, + 678 + ], + "score": 1.0, + "content": "maps, and 256-dimensional bottleneck layers),", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 106, + 677, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 297, + 688 + ], + "score": 1.0, + "content": "to use 46 blocks, with 4096-dimensional fea-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 106, + 687, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 297, + 700 + ], + "score": 1.0, + "content": "ture maps and 512-dimensional bottleneck lay-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 699, + 297, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 297, + 710 + ], + "score": 1.0, + "content": "ers. We call the resulting network ResNet-161.", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 709, + 297, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 297, + 721 + ], + "score": 1.0, + "content": "Consistent with prior results, this new architec-", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 106, + 720, + 298, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 298, + 733 + ], + "score": 1.0, + "content": "ture delivers better performance regardless of", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 53, + "bbox_fs": [ + 105, + 567, + 298, + 733 + ] + }, + { + "type": "image", + "bbox": [ + 307, + 449, + 501, + 616 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 307, + 449, + 501, + 616 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 501, + 616 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 501, + 616 + ], + "score": 0.97, + "type": "image", + "image_path": "3eeded7c7663fc17771caa90292cc1343662124f1094d06f22d9ca255036f2a3.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 307, + 449, + 501, + 461.84615384615387 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 307, + 461.84615384615387, + 501, + 474.69230769230774 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 307, + 474.69230769230774, + 501, + 487.5384615384616 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 307, + 487.5384615384616, + 501, + 500.3846153846155 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 307, + 500.3846153846155, + 501, + 513.2307692307693 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 307, + 513.2307692307693, + 501, + 526.0769230769231 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 307, + 526.0769230769231, + 501, + 538.9230769230769 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 307, + 538.9230769230769, + 501, + 551.7692307692307 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 307, + 551.7692307692307, + 501, + 564.6153846153845 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 307, + 564.6153846153845, + 501, + 577.4615384615383 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 307, + 577.4615384615383, + 501, + 590.3076923076922 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 307, + 590.3076923076922, + 501, + 603.153846153846 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 307, + 603.153846153846, + 501, + 615.9999999999998 + ], + "spans": [], + "index": 64 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 620, + 504, + 719 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 506, + 632 + ], + "score": 1.0, + "content": "Figure 3: Linear classification performance of", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 303, + 631, + 506, + 643 + ], + "spans": [ + { + "bbox": [ + 303, + 631, + 506, + 643 + ], + "score": 1.0, + "content": "new variants of CPC, which incrementally add a", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "series of modifications. BU: bottum up spatial", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 653, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 303, + 653, + 505, + 665 + ], + "score": 1.0, + "content": "predictions. HF: randomly flipping patches hor-", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 304, + 664, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 505, + 675 + ], + "score": 1.0, + "content": "izontally. LN: layer normalization. RC: random", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 303, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 303, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "color-dropping. TL: tuned prediction lengths. HP:", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 304, + 686, + 505, + 697 + ], + "spans": [ + { + "bbox": [ + 304, + 686, + 505, + 697 + ], + "score": 1.0, + "content": "horizontal spatial predictions. We use color to in-", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 304, + 697, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 697, + 505, + 709 + ], + "score": 1.0, + "content": "dicate the number of spatial predictions used (or-", + "type": "text" + } + ], + "index": 72 + }, + { + "bbox": [ + 303, + 708, + 478, + 719 + ], + "spans": [ + { + "bbox": [ + 303, + 708, + 478, + 719 + ], + "score": 1.0, + "content": "ange, green, blue for 1, 2 and 4 directions).", + "type": "text" + } + ], + "index": 73 + } + ], + "index": 69 + } + ], + "index": 56.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 125, + 126, + 486, + 294 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 89, + 504, + 123 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 218, + 102 + ], + "score": 1.0, + "content": "Table 1: Linear classifier", + "type": "text" + }, + { + "bbox": [ + 218, + 90, + 231, + 101 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 88, + 291, + 102 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 291, + 89, + 317, + 100 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "of labels. Comparison to linear separability", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "score": 1.0, + "content": "of other self-supervised methods. In all cases a feature extractor is optimized in an unsupervised", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 111, + 456, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 456, + 124 + ], + "score": 1.0, + "content": "manner, and a linear classifier is trained on top using all labels in the ImageNet dataset.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 125, + 126, + 486, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 126, + 486, + 294 + ], + "spans": [ + { + "bbox": [ + 125, + 126, + 486, + 294 + ], + "score": 0.974, + "html": "
MethodTop-1Top-5
Motion Segmentation (MS) (Pathak et al., 2016) Exemplar (Ex) (Dosovitskiy et al.,2014) Relative Position (RP) (Doersch et al.,2015) Colorization (Col) (Zhang et al., 2016) Combination of MS + Ex + RP+ Col (Doersch & Zisserman,2017) CPC v1 (van den Oord et al., 2018) Rotation (Kolesnikov et al., 2019)27.6 31.5 36.2 39.6 1 48.7 55.448.3 53.1 59.2 62.5 69.3 73.6
CMC (Tian et al., 2019) Local Aggregation (Zhuang et al., 2019)60.1 60.282.8 1 =
BigBiGAN (Donahue & Simonyan,2019) AMDIM (Bachman et al., 2019) CPC v2 (ours) 65.961.3 68.181.9
", + "type": "table", + "image_path": "fa1f65edaf71df9800daca9c26a7999e93d74856116a87cc357e08cc713a4be7.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 125, + 126, + 486, + 182.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 125, + 182.0, + 486, + 238.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 125, + 238.0, + 486, + 294.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 325, + 504, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "other design choices. Interestingly, a larger architecture delivers larger improvements with more", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 335, + 503, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 481, + 349 + ], + "score": 1.0, + "content": "efficient training, more self-supervised losses, and more patch-based augmentations (Fig. 3,", + "type": "text" + }, + { + "bbox": [ + 482, + 336, + 503, + 347 + ], + "score": 0.84, + "content": "+5 \\%", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 347, + 420, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 292, + 360 + ], + "score": 1.0, + "content": "Top-1 accuracy with original training scheme,", + "type": "text" + }, + { + "bbox": [ + 292, + 347, + 318, + 357 + ], + "score": 0.88, + "content": "+ 1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 347, + 420, + 360 + ], + "score": 1.0, + "content": "accuracy with new one).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "Layer normalization. Large architectures are more difficult to train efficiently. Early works on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "context prediction with patches used batch normalization (Ioffe & Szegedy, 2015; Doersch et al.,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 397, + 504, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 504, + 411 + ], + "score": 1.0, + "content": "2015) to speed training. However, with CPC we find that batch normalization actually harms down-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "stream performance of large models. We hypothesize that batch normalization allows large models", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "to find a trivial solution to CPC: it introduces a dependency between patches (through the batch", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "statistics) that can be exploited to bypass the constraints on the receptive field. We find that we can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "reclaim much of batch normalization’s training efficiency using layer normalization (Ba et al., 2016),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 328, + 466 + ], + "score": 1.0, + "content": "which leads to a small gain for the smaller architecture", + "type": "text" + }, + { + "bbox": [ + 328, + 453, + 349, + 464 + ], + "score": 0.87, + "content": "+ 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "accuracy over equivalent architectures", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 462, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 406, + 478 + ], + "score": 1.0, + "content": "that use neither normalization) and a larger gain for the larger architecture", + "type": "text" + }, + { + "bbox": [ + 406, + 464, + 436, + 475 + ], + "score": 0.87, + "content": "( + 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 462, + 481, + 478 + ], + "score": 1.0, + "content": "accuracy).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "Prediction lengths and directions. Larger architectures also run a greater risk of overfitting. We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "address this by asking more from the network: specifically, whereas van den Oord et al. (2018) pre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "score": 1.0, + "content": "dicted each patch using only context from spatially beneath it, we repeatedly predict the patch using", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "context from above, to the right, and to the left, resulting in up to four times as many prediction", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 536, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 443, + 551 + ], + "score": 1.0, + "content": "tasks. Combining top-to-bottom with bottom-to-top helps both model architectures", + "type": "text" + }, + { + "bbox": [ + 443, + 538, + 465, + 548 + ], + "score": 0.87, + "content": "42 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 536, + 505, + 551 + ], + "score": 1.0, + "content": "accuracy", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 460, + 560 + ], + "score": 1.0, + "content": "for both), but using all 4 spatial directions only benefits the larger model (an additional", + "type": "text" + }, + { + "bbox": [ + 460, + 549, + 489, + 559 + ], + "score": 0.9, + "content": "+ 1 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 560, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 177, + 572 + ], + "score": 1.0, + "content": "the larger model,", + "type": "text" + }, + { + "bbox": [ + 177, + 560, + 196, + 570 + ], + "score": 0.85, + "content": "{ \\bf - 1 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 560, + 506, + 572 + ], + "score": 1.0, + "content": "for the smaller), consistent with the idea that model capacity and amount of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "supervision must go hand-in-hand. We also hypothesized that prediction “length”—i.e. offset be-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "tween the predicted patch and the aggregated context—might affect performance, as distant patches", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "might lie on distinct objects, encouraging the network to memorize images. Indeed, limiting the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 230, + 617 + ], + "score": 1.0, + "content": "range of the prediction length", + "type": "text" + }, + { + "bbox": [ + 230, + 604, + 237, + 613 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 603, + 249, + 617 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 250, + 603, + 275, + 615 + ], + "score": 0.93, + "content": "\\{ 2 , 3 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 603, + 367, + 617 + ], + "score": 1.0, + "content": "performed better than", + "type": "text" + }, + { + "bbox": [ + 368, + 603, + 411, + 615 + ], + "score": 0.93, + "content": "\\{ 2 , \\ldots , 5 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "as was used originally", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 614, + 220, + 627 + ], + "spans": [ + { + "bbox": [ + 110, + 614, + 131, + 625 + ], + "score": 0.86, + "content": "+ 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 614, + 220, + 627 + ], + "score": 1.0, + "content": "for the larger model).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "Patch-based augmentation. If the network can solve CPC using low-level patterns (e.g. straight", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "lines continuing between patches, chromatic aberration), it need not learn semantically meaning-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "ful content. Augmenting the low-level variability across patches can remove such low level cues.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "The original CPC model spatially jitters individual patches independently. We further this logic by", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "adopting the ‘color dropping’ method of Doersch et al. (2015), which randomly drops two of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 399, + 711 + ], + "score": 1.0, + "content": "three color channels in each patch, and find it to delivers systematic gains", + "type": "text" + }, + { + "bbox": [ + 400, + 699, + 422, + 709 + ], + "score": 0.86, + "content": "( + 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "for the small model,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 128, + 720 + ], + "score": 0.87, + "content": "+3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "for the larger one). We also randomly flip patches horizontally, but find it only benefits the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 199, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 166, + 733 + ], + "score": 1.0, + "content": "smaller model", + "type": "text" + }, + { + "bbox": [ + 166, + 721, + 194, + 732 + ], + "score": 0.9, + "content": "( + 1 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 720, + 199, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 125, + 126, + 486, + 294 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 89, + 504, + 123 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 218, + 102 + ], + "score": 1.0, + "content": "Table 1: Linear classifier", + "type": "text" + }, + { + "bbox": [ + 218, + 90, + 231, + 101 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 88, + 291, + 102 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 291, + 89, + 317, + 100 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "of labels. Comparison to linear separability", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 506, + 113 + ], + "score": 1.0, + "content": "of other self-supervised methods. In all cases a feature extractor is optimized in an unsupervised", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 111, + 456, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 456, + 124 + ], + "score": 1.0, + "content": "manner, and a linear classifier is trained on top using all labels in the ImageNet dataset.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "table_body", + "bbox": [ + 125, + 126, + 486, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 125, + 126, + 486, + 294 + ], + "spans": [ + { + "bbox": [ + 125, + 126, + 486, + 294 + ], + "score": 0.974, + "html": "
MethodTop-1Top-5
Motion Segmentation (MS) (Pathak et al., 2016) Exemplar (Ex) (Dosovitskiy et al.,2014) Relative Position (RP) (Doersch et al.,2015) Colorization (Col) (Zhang et al., 2016) Combination of MS + Ex + RP+ Col (Doersch & Zisserman,2017) CPC v1 (van den Oord et al., 2018) Rotation (Kolesnikov et al., 2019)27.6 31.5 36.2 39.6 1 48.7 55.448.3 53.1 59.2 62.5 69.3 73.6
CMC (Tian et al., 2019) Local Aggregation (Zhuang et al., 2019)60.1 60.282.8 1 =
BigBiGAN (Donahue & Simonyan,2019) AMDIM (Bachman et al., 2019) CPC v2 (ours) 65.961.3 68.181.9
", + "type": "table", + "image_path": "fa1f65edaf71df9800daca9c26a7999e93d74856116a87cc357e08cc713a4be7.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 125, + 126, + 486, + 182.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 125, + 182.0, + 486, + 238.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 125, + 238.0, + 486, + 294.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 325, + 504, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "other design choices. Interestingly, a larger architecture delivers larger improvements with more", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 335, + 503, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 481, + 349 + ], + "score": 1.0, + "content": "efficient training, more self-supervised losses, and more patch-based augmentations (Fig. 3,", + "type": "text" + }, + { + "bbox": [ + 482, + 336, + 503, + 347 + ], + "score": 0.84, + "content": "+5 \\%", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 347, + 420, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 292, + 360 + ], + "score": 1.0, + "content": "Top-1 accuracy with original training scheme,", + "type": "text" + }, + { + "bbox": [ + 292, + 347, + 318, + 357 + ], + "score": 0.88, + "content": "+ 1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 347, + 420, + 360 + ], + "score": 1.0, + "content": "accuracy with new one).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 324, + 506, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "Layer normalization. Large architectures are more difficult to train efficiently. Early works on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "context prediction with patches used batch normalization (Ioffe & Szegedy, 2015; Doersch et al.,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 397, + 504, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 504, + 411 + ], + "score": 1.0, + "content": "2015) to speed training. However, with CPC we find that batch normalization actually harms down-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 409, + 505, + 422 + ], + "score": 1.0, + "content": "stream performance of large models. We hypothesize that batch normalization allows large models", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "to find a trivial solution to CPC: it introduces a dependency between patches (through the batch", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 431, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 505, + 444 + ], + "score": 1.0, + "content": "statistics) that can be exploited to bypass the constraints on the receptive field. We find that we can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "reclaim much of batch normalization’s training efficiency using layer normalization (Ba et al., 2016),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 452, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 328, + 466 + ], + "score": 1.0, + "content": "which leads to a small gain for the smaller architecture", + "type": "text" + }, + { + "bbox": [ + 328, + 453, + 349, + 464 + ], + "score": 0.87, + "content": "+ 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 452, + 505, + 466 + ], + "score": 1.0, + "content": "accuracy over equivalent architectures", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 462, + 481, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 406, + 478 + ], + "score": 1.0, + "content": "that use neither normalization) and a larger gain for the larger architecture", + "type": "text" + }, + { + "bbox": [ + 406, + 464, + 436, + 475 + ], + "score": 0.87, + "content": "( + 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 462, + 481, + 478 + ], + "score": 1.0, + "content": "accuracy).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 376, + 506, + 478 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 493, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "Prediction lengths and directions. Larger architectures also run a greater risk of overfitting. We", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "address this by asking more from the network: specifically, whereas van den Oord et al. (2018) pre-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 529 + ], + "score": 1.0, + "content": "dicted each patch using only context from spatially beneath it, we repeatedly predict the patch using", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 539 + ], + "score": 1.0, + "content": "context from above, to the right, and to the left, resulting in up to four times as many prediction", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 536, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 443, + 551 + ], + "score": 1.0, + "content": "tasks. Combining top-to-bottom with bottom-to-top helps both model architectures", + "type": "text" + }, + { + "bbox": [ + 443, + 538, + 465, + 548 + ], + "score": 0.87, + "content": "42 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 536, + 505, + 551 + ], + "score": 1.0, + "content": "accuracy", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 460, + 560 + ], + "score": 1.0, + "content": "for both), but using all 4 spatial directions only benefits the larger model (an additional", + "type": "text" + }, + { + "bbox": [ + 460, + 549, + 489, + 559 + ], + "score": 0.9, + "content": "+ 1 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 560, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 177, + 572 + ], + "score": 1.0, + "content": "the larger model,", + "type": "text" + }, + { + "bbox": [ + 177, + 560, + 196, + 570 + ], + "score": 0.85, + "content": "{ \\bf - 1 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 560, + 506, + 572 + ], + "score": 1.0, + "content": "for the smaller), consistent with the idea that model capacity and amount of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "supervision must go hand-in-hand. We also hypothesized that prediction “length”—i.e. offset be-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "tween the predicted patch and the aggregated context—might affect performance, as distant patches", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "might lie on distinct objects, encouraging the network to memorize images. Indeed, limiting the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 230, + 617 + ], + "score": 1.0, + "content": "range of the prediction length", + "type": "text" + }, + { + "bbox": [ + 230, + 604, + 237, + 613 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 603, + 249, + 617 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 250, + 603, + 275, + 615 + ], + "score": 0.93, + "content": "\\{ 2 , 3 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 603, + 367, + 617 + ], + "score": 1.0, + "content": "performed better than", + "type": "text" + }, + { + "bbox": [ + 368, + 603, + 411, + 615 + ], + "score": 0.93, + "content": "\\{ 2 , \\ldots , 5 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "as was used originally", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 110, + 614, + 220, + 627 + ], + "spans": [ + { + "bbox": [ + 110, + 614, + 131, + 625 + ], + "score": 0.86, + "content": "+ 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 614, + 220, + 627 + ], + "score": 1.0, + "content": "for the larger model).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23.5, + "bbox_fs": [ + 104, + 493, + 506, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "Patch-based augmentation. If the network can solve CPC using low-level patterns (e.g. straight", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "lines continuing between patches, chromatic aberration), it need not learn semantically meaning-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "ful content. Augmenting the low-level variability across patches can remove such low level cues.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "The original CPC model spatially jitters individual patches independently. We further this logic by", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "adopting the ‘color dropping’ method of Doersch et al. (2015), which randomly drops two of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 698, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 399, + 711 + ], + "score": 1.0, + "content": "three color channels in each patch, and find it to delivers systematic gains", + "type": "text" + }, + { + "bbox": [ + 400, + 699, + 422, + 709 + ], + "score": 0.86, + "content": "( + 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 698, + 505, + 711 + ], + "score": 1.0, + "content": "for the small model,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 128, + 720 + ], + "score": 0.87, + "content": "+3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "for the larger one). We also randomly flip patches horizontally, but find it only benefits the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 720, + 199, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 166, + 733 + ], + "score": 1.0, + "content": "smaller model", + "type": "text" + }, + { + "bbox": [ + 166, + 721, + 194, + 732 + ], + "score": 0.9, + "content": "( + 1 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 720, + 199, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 643, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Combined. Cumulatively, these fairly straightforward implementation changes lead to a substan-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 288, + 107 + ], + "score": 1.0, + "content": "tial improvement to the original CPC model", + "type": "text" + }, + { + "bbox": [ + 289, + 94, + 316, + 105 + ], + "score": 0.83, + "content": "6 5 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 93, + 390, + 107 + ], + "score": 1.0, + "content": "Top-1 accuracy, a", + "type": "text" + }, + { + "bbox": [ + 390, + 94, + 410, + 104 + ], + "score": 0.87, + "content": "17 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "improvement), making", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "it competitive with recent approaches and outperforming prior methods (see table 1). Interestingly,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "if we train the same patch-based architecture from scratch in a fully supervised manner, we ob-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 124, + 139 + ], + "score": 1.0, + "content": "tain", + "type": "text" + }, + { + "bbox": [ + 124, + 126, + 151, + 137 + ], + "score": 0.86, + "content": "6 6 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 126, + 325, + 139 + ], + "score": 1.0, + "content": "Top-1 accuracy (with batch normalization;", + "type": "text" + }, + { + "bbox": [ + 325, + 126, + 352, + 137 + ], + "score": 0.87, + "content": "6 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "without), suggesting that CPC is now", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "nearly saturating the architecture’s representational power despite not using labels. These results", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "illustrate how architecture and data have an outsized impact on the linear classification performance", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "of self-supervised representations, and are interesting to compare with with previous results. For ex-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 449, + 183 + ], + "score": 1.0, + "content": "ample, in AMDIM, different settings of data augmentation alone can result in a nearly", + "type": "text" + }, + { + "bbox": [ + 449, + 171, + 468, + 181 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 169, + 506, + 183 + ], + "score": 1.0, + "content": "absolute", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 340, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 340, + 193 + ], + "score": 1.0, + "content": "increase in performance on ImageNet linear classification.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 109, + 208, + 279, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 281, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 281, + 220 + ], + "score": 1.0, + "content": "4.2 EFFICIENT IMAGE CLASSIFICATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "toWe now turn to our original question of whether CPC can enable data-efficient image recognition.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "score": 1.0, + "content": "We start by evaluating the performance of purely-supervised networks as the size of the labeled", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 137, + 263 + ], + "score": 1.0, + "content": "dataset", + "type": "text" + }, + { + "bbox": [ + 137, + 251, + 149, + 262 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 251, + 200, + 263 + ], + "score": 1.0, + "content": "varies from", + "type": "text" + }, + { + "bbox": [ + 200, + 251, + 214, + 261 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 251, + 227, + 263 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 228, + 251, + 252, + 262 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "of ImageNet, training separate classifiers on each subset. We", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "found that a ResNet-152 to works best across all data-regimes (see Appendix). Despite our efforts to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 272, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 286 + ], + "score": 1.0, + "content": "tune the supervised model for low-data classification (including network depth, regularization, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 388, + 296 + ], + "score": 1.0, + "content": "optimization parameters), the accuracy of the best model only reaches", + "type": "text" + }, + { + "bbox": [ + 389, + 284, + 416, + 294 + ], + "score": 0.86, + "content": "4 4 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "Top-5 accuracy when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 295, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 149, + 307 + ], + "score": 1.0, + "content": "trained on", + "type": "text" + }, + { + "bbox": [ + 149, + 295, + 163, + 305 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 295, + 274, + 307 + ], + "score": 1.0, + "content": "of the dataset (compared to", + "type": "text" + }, + { + "bbox": [ + 275, + 295, + 302, + 306 + ], + "score": 0.86, + "content": "9 3 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 295, + 505, + 307 + ], + "score": 1.0, + "content": "when trained on the entire dataset, see Fig. 1, red).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 296, + 594 + ], + "lines": [ + { + "bbox": [ + 107, + 320, + 297, + 332 + ], + "spans": [ + { + "bbox": [ + 107, + 320, + 297, + 332 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding. We now ad-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 331, + 297, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 297, + 343 + ], + "score": 1.0, + "content": "dress our central question of whether CPC en-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 342, + 297, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 297, + 354 + ], + "score": 1.0, + "content": "ables data-efficient learning. We follow the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 353, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 297, + 365 + ], + "score": 1.0, + "content": "same paradigm as for the supervised base-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 297, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 297, + 377 + ], + "score": 1.0, + "content": "line (training and evaluating a separate clas-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 374, + 297, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 297, + 387 + ], + "score": 1.0, + "content": "sifier for each size subset), stacking a neu-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 297, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 297, + 398 + ], + "score": 1.0, + "content": "ral network classifier on top of the CPC la-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 397, + 297, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 130, + 409 + ], + "score": 1.0, + "content": "tents", + "type": "text" + }, + { + "bbox": [ + 131, + 397, + 186, + 408 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { z } & { { } = } & { f _ { \\theta } ( { \\pmb x } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 397, + 297, + 409 + ], + "score": 1.0, + "content": "rather than the raw im-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 298, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 153, + 420 + ], + "score": 1.0, + "content": "age pixels", + "type": "text" + }, + { + "bbox": [ + 154, + 410, + 162, + 418 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 408, + 298, + 420 + ], + "score": 1.0, + "content": "(see section 2.2, efficient clas-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 418, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 297, + 431 + ], + "score": 1.0, + "content": "sification, and Appendix). This representa-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 429, + 297, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 297, + 441 + ], + "score": 1.0, + "content": "tion, which we selected for its improved linear", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 297, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 297, + 453 + ], + "score": 1.0, + "content": "classification performance (CPC v2 in Fig. 3),", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 451, + 297, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 297, + 464 + ], + "score": 1.0, + "content": "leads to a significant increase in data-efficiency", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 463, + 297, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 297, + 475 + ], + "score": 1.0, + "content": "compared to purely supervised networks (Fig.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 268, + 486 + ], + "score": 1.0, + "content": "1, blue curve). This classifier yields", + "type": "text" + }, + { + "bbox": [ + 268, + 474, + 296, + 484 + ], + "score": 0.85, + "content": "7 2 . 9 \\%", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 484, + 297, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 220, + 496 + ], + "score": 1.0, + "content": "Top-5 accuracy with only", + "type": "text" + }, + { + "bbox": [ + 220, + 485, + 235, + 495 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 484, + 297, + 496 + ], + "score": 1.0, + "content": "of the labels,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 115, + 507 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 115, + 496, + 136, + 506 + ], + "score": 0.87, + "content": "2 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 495, + 238, + 507 + ], + "score": 1.0, + "content": "absolute improvement", + "type": "text" + }, + { + "bbox": [ + 239, + 496, + 258, + 506 + ], + "score": 0.83, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 495, + 297, + 507 + ], + "score": 1.0, + "content": "relative)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "score": 1.0, + "content": "over purely-supervised methods. Surprisingly,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 517, + 297, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 297, + 528 + ], + "score": 1.0, + "content": "when given the entire dataset, this classifier", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 528, + 297, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 139, + 540 + ], + "score": 1.0, + "content": "reaches", + "type": "text" + }, + { + "bbox": [ + 139, + 528, + 194, + 539 + ], + "score": 0.9, + "content": "8 0 . 6 \\% / 9 5 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 528, + 297, + 540 + ], + "score": 1.0, + "content": "Top1/Top5 accuracy, sur-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 540, + 297, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 297, + 550 + ], + "score": 1.0, + "content": "passing our supervised baseline (ResNet-152:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 549, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 162, + 561 + ], + "score": 0.89, + "content": "7 8 . 0 \\% / 9 3 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 549, + 297, + 562 + ], + "score": 1.0, + "content": "accuracy) and published results", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 560, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 163, + 574 + ], + "score": 1.0, + "content": "(ResNet-200:", + "type": "text" + }, + { + "bbox": [ + 163, + 561, + 218, + 572 + ], + "score": 0.86, + "content": "7 9 . 9 \\% / 9 5 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 560, + 297, + 574 + ], + "score": 1.0, + "content": ", He et al. (2016b)).", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 571, + 297, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 297, + 584 + ], + "score": 1.0, + "content": "We find similar results in all other data-regimes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 582, + 217, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 217, + 595 + ], + "score": 1.0, + "content": "we considered (see Fig. 1).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 295, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 296, + 612 + ], + "score": 1.0, + "content": "How important are the model specifications de-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 297, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 297, + 623 + ], + "score": 1.0, + "content": "scribed in Section 4.1 for low-data classifica-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 622, + 297, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 297, + 633 + ], + "score": 1.0, + "content": "tion? We hypothesized that predictable repre-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 297, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 297, + 644 + ], + "score": 1.0, + "content": "sentations might enable data-efficient classifi-", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + }, + { + "type": "image", + "bbox": [ + 308, + 324, + 496, + 521 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 324, + 496, + 521 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 324, + 496, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 324, + 496, + 521 + ], + "score": 0.97, + "type": "image", + "image_path": "106f2252ad377e16a329cfb21360459a72fdc1d09e4f745b26572da1cf6d5d14.jpg" + } + ] + } + ], + "index": 54.5, + "virtual_lines": [ + { + "bbox": [ + 308, + 324, + 496, + 336.3125 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 308, + 336.3125, + 496, + 348.625 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 308, + 348.625, + 496, + 360.9375 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 308, + 360.9375, + 496, + 373.25 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 308, + 373.25, + 496, + 385.5625 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 308, + 385.5625, + 496, + 397.875 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 308, + 397.875, + 496, + 410.1875 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 308, + 410.1875, + 496, + 422.5 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 308, + 422.5, + 496, + 434.8125 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 308, + 434.8125, + 496, + 447.125 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 308, + 447.125, + 496, + 459.4375 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 308, + 459.4375, + 496, + 471.75 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 308, + 471.75, + 496, + 484.0625 + ], + "spans": [], + "index": 59 + }, + { + "bbox": [ + 308, + 484.0625, + 496, + 496.375 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 308, + 496.375, + 496, + 508.6875 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 308, + 508.6875, + 496, + 521.0 + ], + "spans": [], + "index": 62 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 526, + 504, + 636 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 505, + 537 + ], + "score": 1.0, + "content": "Figure 4: Relationship between linear classifica-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 303, + 537, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 303, + 537, + 505, + 548 + ], + "score": 1.0, + "content": "tion accuracy and low-data classification, for dif-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 548, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 505, + 559 + ], + "score": 1.0, + "content": "ferent variants of the CPC model. Left: CPC vari-", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 559, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 505, + 570 + ], + "score": 1.0, + "content": "ants with the same architecture but different train-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 570, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 303, + 570, + 505, + 581 + ], + "score": 1.0, + "content": "ing protocols. Orange, green, and blue dots cor-", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 303, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "respond to CPC models making predictions in 1,", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 303, + 592, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 303, + 592, + 506, + 604 + ], + "score": 1.0, + "content": "2, and 4 spatial directions respectively. Within a", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 303, + 603, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 303, + 603, + 505, + 614 + ], + "score": 1.0, + "content": "color group, different models correspond to other", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 303, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "implementation details (e.g. layer norm and patch", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 303, + 625, + 473, + 636 + ], + "spans": [ + { + "bbox": [ + 303, + 625, + 473, + 636 + ], + "score": 1.0, + "content": "augmentation, and combinations thereof).", + "type": "text" + } + ], + "index": 72 + } + ], + "index": 67.5 + } + ], + "index": 61.0 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "cation, and therefore expect that increasing the amount of ‘predictability’ in the representation", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "should also increase its ability to learn from small amounts of data. Fig. 4 shows evidence for", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "this by ablating model parameters and comparing linear classification performance against low-data", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "classification. Consistent with our hypothesis, increasing the number of spatial directions in the CPC", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "prediction task (which increased linear classification performance) systematically increases low-data", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "classification performance (Fig. 4, left, different color groups). As a control, we asked if all mod-", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "ifications that improve linear classification also improve low-data classification. We did not find", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "evidence in favor of this: improvements in linear classification as a result of changing other model", + "type": "text" + } + ], + "index": 80 + } + ], + "index": 76.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 192 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Combined. Cumulatively, these fairly straightforward implementation changes lead to a substan-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 288, + 107 + ], + "score": 1.0, + "content": "tial improvement to the original CPC model", + "type": "text" + }, + { + "bbox": [ + 289, + 94, + 316, + 105 + ], + "score": 0.83, + "content": "6 5 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 93, + 390, + 107 + ], + "score": 1.0, + "content": "Top-1 accuracy, a", + "type": "text" + }, + { + "bbox": [ + 390, + 94, + 410, + 104 + ], + "score": 0.87, + "content": "17 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "improvement), making", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 505, + 118 + ], + "score": 1.0, + "content": "it competitive with recent approaches and outperforming prior methods (see table 1). Interestingly,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "if we train the same patch-based architecture from scratch in a fully supervised manner, we ob-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 124, + 139 + ], + "score": 1.0, + "content": "tain", + "type": "text" + }, + { + "bbox": [ + 124, + 126, + 151, + 137 + ], + "score": 0.86, + "content": "6 6 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 126, + 325, + 139 + ], + "score": 1.0, + "content": "Top-1 accuracy (with batch normalization;", + "type": "text" + }, + { + "bbox": [ + 325, + 126, + 352, + 137 + ], + "score": 0.87, + "content": "6 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "without), suggesting that CPC is now", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "nearly saturating the architecture’s representational power despite not using labels. These results", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "illustrate how architecture and data have an outsized impact on the linear classification performance", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "of self-supervised representations, and are interesting to compare with with previous results. For ex-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 169, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 169, + 449, + 183 + ], + "score": 1.0, + "content": "ample, in AMDIM, different settings of data augmentation alone can result in a nearly", + "type": "text" + }, + { + "bbox": [ + 449, + 171, + 468, + 181 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 169, + 506, + 183 + ], + "score": 1.0, + "content": "absolute", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 181, + 340, + 193 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 340, + 193 + ], + "score": 1.0, + "content": "increase in performance on ImageNet linear classification.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 4.5, + "bbox_fs": [ + 104, + 82, + 506, + 193 + ] + }, + { + "type": "title", + "bbox": [ + 109, + 208, + 279, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 281, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 281, + 220 + ], + "score": 1.0, + "content": "4.2 EFFICIENT IMAGE CLASSIFICATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 228, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 105, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "toWe now turn to our original question of whether CPC can enable data-efficient image recognition.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 253 + ], + "score": 1.0, + "content": "We start by evaluating the performance of purely-supervised networks as the size of the labeled", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 137, + 263 + ], + "score": 1.0, + "content": "dataset", + "type": "text" + }, + { + "bbox": [ + 137, + 251, + 149, + 262 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 251, + 200, + 263 + ], + "score": 1.0, + "content": "varies from", + "type": "text" + }, + { + "bbox": [ + 200, + 251, + 214, + 261 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 251, + 227, + 263 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 228, + 251, + 252, + 262 + ], + "score": 0.87, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "of ImageNet, training separate classifiers on each subset. We", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 275 + ], + "score": 1.0, + "content": "found that a ResNet-152 to works best across all data-regimes (see Appendix). Despite our efforts to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 272, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 272, + 505, + 286 + ], + "score": 1.0, + "content": "tune the supervised model for low-data classification (including network depth, regularization, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 388, + 296 + ], + "score": 1.0, + "content": "optimization parameters), the accuracy of the best model only reaches", + "type": "text" + }, + { + "bbox": [ + 389, + 284, + 416, + 294 + ], + "score": 0.86, + "content": "4 4 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "Top-5 accuracy when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 295, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 149, + 307 + ], + "score": 1.0, + "content": "trained on", + "type": "text" + }, + { + "bbox": [ + 149, + 295, + 163, + 305 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 295, + 274, + 307 + ], + "score": 1.0, + "content": "of the dataset (compared to", + "type": "text" + }, + { + "bbox": [ + 275, + 295, + 302, + 306 + ], + "score": 0.86, + "content": "9 3 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 295, + 505, + 307 + ], + "score": 1.0, + "content": "when trained on the entire dataset, see Fig. 1, red).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 229, + 506, + 307 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 296, + 594 + ], + "lines": [ + { + "bbox": [ + 107, + 320, + 297, + 332 + ], + "spans": [ + { + "bbox": [ + 107, + 320, + 297, + 332 + ], + "score": 1.0, + "content": "Contrastive Predictive Coding. We now ad-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 331, + 297, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 297, + 343 + ], + "score": 1.0, + "content": "dress our central question of whether CPC en-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 342, + 297, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 297, + 354 + ], + "score": 1.0, + "content": "ables data-efficient learning. We follow the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 353, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 297, + 365 + ], + "score": 1.0, + "content": "same paradigm as for the supervised base-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 297, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 297, + 377 + ], + "score": 1.0, + "content": "line (training and evaluating a separate clas-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 374, + 297, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 297, + 387 + ], + "score": 1.0, + "content": "sifier for each size subset), stacking a neu-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 297, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 297, + 398 + ], + "score": 1.0, + "content": "ral network classifier on top of the CPC la-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 397, + 297, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 130, + 409 + ], + "score": 1.0, + "content": "tents", + "type": "text" + }, + { + "bbox": [ + 131, + 397, + 186, + 408 + ], + "score": 0.92, + "content": "\\begin{array} { r l r } { z } & { { } = } & { f _ { \\theta } ( { \\pmb x } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 397, + 297, + 409 + ], + "score": 1.0, + "content": "rather than the raw im-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 298, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 153, + 420 + ], + "score": 1.0, + "content": "age pixels", + "type": "text" + }, + { + "bbox": [ + 154, + 410, + 162, + 418 + ], + "score": 0.74, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 408, + 298, + 420 + ], + "score": 1.0, + "content": "(see section 2.2, efficient clas-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 418, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 297, + 431 + ], + "score": 1.0, + "content": "sification, and Appendix). This representa-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 429, + 297, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 297, + 441 + ], + "score": 1.0, + "content": "tion, which we selected for its improved linear", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 297, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 297, + 453 + ], + "score": 1.0, + "content": "classification performance (CPC v2 in Fig. 3),", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 451, + 297, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 297, + 464 + ], + "score": 1.0, + "content": "leads to a significant increase in data-efficiency", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 463, + 297, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 297, + 475 + ], + "score": 1.0, + "content": "compared to purely supervised networks (Fig.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 473, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 268, + 486 + ], + "score": 1.0, + "content": "1, blue curve). This classifier yields", + "type": "text" + }, + { + "bbox": [ + 268, + 474, + 296, + 484 + ], + "score": 0.85, + "content": "7 2 . 9 \\%", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 484, + 297, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 220, + 496 + ], + "score": 1.0, + "content": "Top-5 accuracy with only", + "type": "text" + }, + { + "bbox": [ + 220, + 485, + 235, + 495 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 484, + 297, + 496 + ], + "score": 1.0, + "content": "of the labels,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 115, + 507 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 115, + 496, + 136, + 506 + ], + "score": 0.87, + "content": "2 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 495, + 238, + 507 + ], + "score": 1.0, + "content": "absolute improvement", + "type": "text" + }, + { + "bbox": [ + 239, + 496, + 258, + 506 + ], + "score": 0.83, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 495, + 297, + 507 + ], + "score": 1.0, + "content": "relative)", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 297, + 519 + ], + "score": 1.0, + "content": "over purely-supervised methods. Surprisingly,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 517, + 297, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 297, + 528 + ], + "score": 1.0, + "content": "when given the entire dataset, this classifier", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 528, + 297, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 139, + 540 + ], + "score": 1.0, + "content": "reaches", + "type": "text" + }, + { + "bbox": [ + 139, + 528, + 194, + 539 + ], + "score": 0.9, + "content": "8 0 . 6 \\% / 9 5 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 528, + 297, + 540 + ], + "score": 1.0, + "content": "Top1/Top5 accuracy, sur-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 540, + 297, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 297, + 550 + ], + "score": 1.0, + "content": "passing our supervised baseline (ResNet-152:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 549, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 162, + 561 + ], + "score": 0.89, + "content": "7 8 . 0 \\% / 9 3 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 549, + 297, + 562 + ], + "score": 1.0, + "content": "accuracy) and published results", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 560, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 163, + 574 + ], + "score": 1.0, + "content": "(ResNet-200:", + "type": "text" + }, + { + "bbox": [ + 163, + 561, + 218, + 572 + ], + "score": 0.86, + "content": "7 9 . 9 \\% / 9 5 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 560, + 297, + 574 + ], + "score": 1.0, + "content": ", He et al. (2016b)).", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 571, + 297, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 297, + 584 + ], + "score": 1.0, + "content": "We find similar results in all other data-regimes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 582, + 217, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 217, + 595 + ], + "score": 1.0, + "content": "we considered (see Fig. 1).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 320, + 298, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 600, + 295, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 600, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 296, + 612 + ], + "score": 1.0, + "content": "How important are the model specifications de-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 297, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 297, + 623 + ], + "score": 1.0, + "content": "scribed in Section 4.1 for low-data classifica-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 622, + 297, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 297, + 633 + ], + "score": 1.0, + "content": "tion? We hypothesized that predictable repre-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 297, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 297, + 644 + ], + "score": 1.0, + "content": "sentations might enable data-efficient classifi-", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 600, + 297, + 644 + ] + }, + { + "type": "image", + "bbox": [ + 308, + 324, + 496, + 521 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 308, + 324, + 496, + 521 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 308, + 324, + 496, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 324, + 496, + 521 + ], + "score": 0.97, + "type": "image", + "image_path": "106f2252ad377e16a329cfb21360459a72fdc1d09e4f745b26572da1cf6d5d14.jpg" + } + ] + } + ], + "index": 54.5, + "virtual_lines": [ + { + "bbox": [ + 308, + 324, + 496, + 336.3125 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 308, + 336.3125, + 496, + 348.625 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 308, + 348.625, + 496, + 360.9375 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 308, + 360.9375, + 496, + 373.25 + ], + "spans": [], + "index": 50 + }, + { + "bbox": [ + 308, + 373.25, + 496, + 385.5625 + ], + "spans": [], + "index": 51 + }, + { + "bbox": [ + 308, + 385.5625, + 496, + 397.875 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 308, + 397.875, + 496, + 410.1875 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 308, + 410.1875, + 496, + 422.5 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 308, + 422.5, + 496, + 434.8125 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 308, + 434.8125, + 496, + 447.125 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 308, + 447.125, + 496, + 459.4375 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 308, + 459.4375, + 496, + 471.75 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 308, + 471.75, + 496, + 484.0625 + ], + "spans": [], + "index": 59 + }, + { + "bbox": [ + 308, + 484.0625, + 496, + 496.375 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 308, + 496.375, + 496, + 508.6875 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 308, + 508.6875, + 496, + 521.0 + ], + "spans": [], + "index": 62 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 526, + 504, + 636 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 505, + 537 + ], + "score": 1.0, + "content": "Figure 4: Relationship between linear classifica-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 303, + 537, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 303, + 537, + 505, + 548 + ], + "score": 1.0, + "content": "tion accuracy and low-data classification, for dif-", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 304, + 548, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 505, + 559 + ], + "score": 1.0, + "content": "ferent variants of the CPC model. Left: CPC vari-", + "type": "text" + } + ], + "index": 65 + }, + { + "bbox": [ + 304, + 559, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 505, + 570 + ], + "score": 1.0, + "content": "ants with the same architecture but different train-", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 303, + 570, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 303, + 570, + 505, + 581 + ], + "score": 1.0, + "content": "ing protocols. Orange, green, and blue dots cor-", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 303, + 581, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 303, + 581, + 505, + 593 + ], + "score": 1.0, + "content": "respond to CPC models making predictions in 1,", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 303, + 592, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 303, + 592, + 506, + 604 + ], + "score": 1.0, + "content": "2, and 4 spatial directions respectively. Within a", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 303, + 603, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 303, + 603, + 505, + 614 + ], + "score": 1.0, + "content": "color group, different models correspond to other", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 303, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 303, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "implementation details (e.g. layer norm and patch", + "type": "text" + } + ], + "index": 71 + }, + { + "bbox": [ + 303, + 625, + 473, + 636 + ], + "spans": [ + { + "bbox": [ + 303, + 625, + 473, + 636 + ], + "score": 1.0, + "content": "augmentation, and combinations thereof).", + "type": "text" + } + ], + "index": 72 + } + ], + "index": 67.5 + } + ], + "index": 61.0 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "cation, and therefore expect that increasing the amount of ‘predictability’ in the representation", + "type": "text" + } + ], + "index": 73 + }, + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 666 + ], + "score": 1.0, + "content": "should also increase its ability to learn from small amounts of data. Fig. 4 shows evidence for", + "type": "text" + } + ], + "index": 74 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "this by ablating model parameters and comparing linear classification performance against low-data", + "type": "text" + } + ], + "index": 75 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "classification. Consistent with our hypothesis, increasing the number of spatial directions in the CPC", + "type": "text" + } + ], + "index": 76 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "prediction task (which increased linear classification performance) systematically increases low-data", + "type": "text" + } + ], + "index": 77 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "classification performance (Fig. 4, left, different color groups). As a control, we asked if all mod-", + "type": "text" + } + ], + "index": 78 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "ifications that improve linear classification also improve low-data classification. We did not find", + "type": "text" + } + ], + "index": 79 + }, + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "evidence in favor of this: improvements in linear classification as a result of changing other model", + "type": "text" + } + ], + "index": 80 + }, + { + "bbox": [ + 105, + 421, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 434 + ], + "score": 1.0, + "content": "parameters (patch-based data-augmentation, layer normalization, and combinations thereof) seem", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 432, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 411, + 445 + ], + "score": 1.0, + "content": "uncorrelated to performance in other tasks (Fig. 4, left, within green group:", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 412, + 432, + 498, + 444 + ], + "score": 0.78, + "content": "R ^ { 2 } = 0 . 1 7 , p = 0 . 3 6", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 499, + 432, + 505, + 445 + ], + "score": 1.0, + "content": ").", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 443, + 504, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 504, + 456 + ], + "score": 1.0, + "content": "Different architectural specifications also produced different changes in both tasks (Fig. 4, right).", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "score": 1.0, + "content": "Whereas increasing the depth of the encoding network greatly improves both metrics, increasing the", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "network width (and therefore the number of features used for linear classification) only improves", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 475, + 226, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 226, + 489 + ], + "score": 1.0, + "content": "linear classification accuracy.", + "type": "text", + "cross_page": true + } + ], + "index": 28 + } + ], + "index": 76.5, + "bbox_fs": [ + 105, + 644, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 181, + 456, + 386 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 88, + 505, + 177 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 216, + 101 + ], + "score": 1.0, + "content": "Table 2: ResNet classifier", + "type": "text" + }, + { + "bbox": [ + 217, + 90, + 229, + 101 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 89, + 286, + 101 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 287, + 90, + 303, + 100 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 89, + 317, + 101 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 317, + 90, + 338, + 100 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "of labels. Comparison to other methods", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 250, + 113 + ], + "score": 1.0, + "content": "for semi-supervised learning using", + "type": "text" + }, + { + "bbox": [ + 250, + 101, + 265, + 111 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 101, + 279, + 113 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 279, + 101, + 298, + 111 + ], + "score": 0.88, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 101, + 505, + 113 + ], + "score": 1.0, + "content": "of labeled data. Representation learning methods", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 125 + ], + "score": 1.0, + "content": "use a classifier to discriminate an unsupervised representation, and optimize it solely for the super-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "vised objective on labeled data. Label-propagation methods on the other hand further constrain the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "classifier with smoothness and entropy criteria on unlabeled data, making the additional assumption", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 374, + 156 + ], + "score": 1.0, + "content": "that all training images fit into a single (unknown) testing category.", + "type": "text" + }, + { + "bbox": [ + 375, + 145, + 381, + 156 + ], + "score": 0.76, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "denotes methods implemented", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "in this work, fixed and fine-tuned denote whether the feature extractor is allowed to accommodate", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 208, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 208, + 179 + ], + "score": 1.0, + "content": "the supervised objective.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "table_body", + "bbox": [ + 153, + 181, + 456, + 386 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 181, + 456, + 386 + ], + "spans": [ + { + "bbox": [ + 153, + 181, + 456, + 386 + ], + "score": 0.91, + "html": "
Labeled data 1% 10% 100% Method Top-5 accuracy
Supervised baseline 44.1 82.1 93.9 Methods using label-propagation:
Pseudolabeling (Zhai et al., 2019) 51.6 82.4 VAT + Entropy Minimization (Zhai et al., 2019) 47.0 83.4 Unsup. Data Augmentation (Xie et al., 2019) 88.5
Rotation+ VAT+Ent.Min. (Zhai et al.,2019) 91.2 95.0 =
Methods using representation learning only:
Instance Discrimination (Wu et al., 2018) 39.2 77.4
Rotation (Zhai et al., 2019) 57.5 86.4
+ResNet trained on BigBiGAN (fixed) 55.2 78.8 87.0
+ResNet trained on AMDIM (fixed) 67.4 85.8 92.2
+ResNet trained on CPC v2 (fixed) 72.3 89.1 94.4 +ResNet trained on CPC v2 (fine-tuned) 72.9 89.5 95.2
", + "type": "table", + "image_path": "e969618e28fc2ffa8c76067cea0310a604dd1bb350cd5a5534dd7265b81333a6.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 153, + 181, + 456, + 194.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 153, + 194.66666666666666, + 456, + 208.33333333333331 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 153, + 208.33333333333331, + 456, + 221.99999999999997 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 153, + 221.99999999999997, + 456, + 235.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 153, + 235.66666666666663, + 456, + 249.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 153, + 249.3333333333333, + 456, + 262.99999999999994 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 153, + 262.99999999999994, + 456, + 276.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 153, + 276.66666666666663, + 456, + 290.3333333333333 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 153, + 290.3333333333333, + 456, + 304.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 153, + 304.0, + 456, + 317.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 153, + 317.6666666666667, + 456, + 331.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 153, + 331.33333333333337, + 456, + 345.00000000000006 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 153, + 345.00000000000006, + 456, + 358.66666666666674 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 153, + 358.66666666666674, + 456, + 372.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 153, + 372.3333333333334, + 456, + 386.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 107, + 421, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 434 + ], + "score": 1.0, + "content": "parameters (patch-based data-augmentation, layer normalization, and combinations thereof) seem", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 432, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 411, + 445 + ], + "score": 1.0, + "content": "uncorrelated to performance in other tasks (Fig. 4, left, within green group:", + "type": "text" + }, + { + "bbox": [ + 412, + 432, + 498, + 444 + ], + "score": 0.78, + "content": "R ^ { 2 } = 0 . 1 7 , p = 0 . 3 6", + "type": "inline_equation" + }, + { + "bbox": [ + 499, + 432, + 505, + 445 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 443, + 504, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 504, + 456 + ], + "score": 1.0, + "content": "Different architectural specifications also produced different changes in both tasks (Fig. 4, right).", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 467 + ], + "score": 1.0, + "content": "Whereas increasing the depth of the encoding network greatly improves both metrics, increasing the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "network width (and therefore the number of features used for linear classification) only improves", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 475, + 226, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 226, + 489 + ], + "score": 1.0, + "content": "linear classification accuracy.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "Other unsupervised representations. How well does the CPC representation compare to other", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 104, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "representations that have been learned in an unsupervised manner? If predictable representations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 522, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 504, + 534 + ], + "score": 1.0, + "content": "are uniquely suited for efficient classification, we would expect other methods within this fam-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "ily to perform similarly, and other model classes less so. Table 2 compares our best model with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "other works on efficient recognition. We consider three objectives from different model classes:", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "self-supervised learning with rotation prediction (Zhai et al., 2019), large-scale adversarial feature", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "learning (BigBiGAN, Donahue & Simonyan, 2019), and another contrastive prediction objective", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "(AMDIM, Bachman et al., 2019). Zhai et al. (2019) evaluate the low-data classification perfor-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "mance of representations learned with rotation prediction using a similar paradigm and architecture", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 597, + 504, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 344, + 610 + ], + "score": 1.0, + "content": "(ResNet-152), hence we report their results directly. Given", + "type": "text" + }, + { + "bbox": [ + 344, + 598, + 359, + 609 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 597, + 504, + 610 + ], + "score": 1.0, + "content": "of ImageNet, their method achieves", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 608, + 504, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 133, + 620 + ], + "score": 0.85, + "content": "5 7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 608, + 438, + 622 + ], + "score": 1.0, + "content": "Top-5 accuracy, consistently with the reduced accuracy of a linear classifier (", + "type": "text" + }, + { + "bbox": [ + 438, + 609, + 465, + 620 + ], + "score": 0.82, + "content": "5 5 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 608, + 477, + 622 + ], + "score": 1.0, + "content": "vs", + "type": "text" + }, + { + "bbox": [ + 477, + 609, + 504, + 620 + ], + "score": 0.85, + "content": "6 5 . 9 \\%", + "type": "inline_equation" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 619, + 152, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 152, + 632 + ], + "score": 1.0, + "content": "for CPC)1.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "Because BigBiGAN and AMDIM achieve stronger linear classification accuracy than rotation pre-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 140, + 661 + ], + "score": 1.0, + "content": "diction", + "type": "text" + }, + { + "bbox": [ + 141, + 648, + 168, + 659 + ], + "score": 0.83, + "content": "6 1 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 648, + 187, + 661 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 188, + 648, + 215, + 659 + ], + "score": 0.86, + "content": "6 8 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "Top-1 accuracy, respectively), we might expect better performance on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "efficient classification as well. Since their authors do not report results on efficient classification we", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 670, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 506, + 683 + ], + "score": 1.0, + "content": "evaluated these representations using the same paradigm we used for evaluating CPC, stacking a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 228, + 694 + ], + "score": 1.0, + "content": "ResNet classifier on top of the", + "type": "text" + }, + { + "bbox": [ + 229, + 681, + 276, + 691 + ], + "score": 0.9, + "content": "7 \\times 7 \\times 8 1 9 2", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 680, + 410, + 694 + ], + "score": 1.0, + "content": "latents of the BigBiGAN and the", + "type": "text" + }, + { + "bbox": [ + 410, + 681, + 457, + 692 + ], + "score": 0.9, + "content": "7 \\times 7 \\times 2 5 6 0", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "grid of fea-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "ture vectors of AMDIM. We found fine-tuned representations to yield only marginal gains over fixed", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 712, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 118, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "1Although note that linear classification accuracy is reported using a separate model that was designed for", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 318, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 318, + 732 + ], + "score": 1.0, + "content": "linear classification performance (Kolesnikov et al., 2019)", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 181, + 456, + 386 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 88, + 505, + 177 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 216, + 101 + ], + "score": 1.0, + "content": "Table 2: ResNet classifier", + "type": "text" + }, + { + "bbox": [ + 217, + 90, + 229, + 101 + ], + "score": 0.89, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 89, + 286, + 101 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 287, + 90, + 303, + 100 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 89, + 317, + 101 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 317, + 90, + 338, + 100 + ], + "score": 0.85, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "of labels. Comparison to other methods", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 101, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 101, + 250, + 113 + ], + "score": 1.0, + "content": "for semi-supervised learning using", + "type": "text" + }, + { + "bbox": [ + 250, + 101, + 265, + 111 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 101, + 279, + 113 + ], + "score": 1.0, + "content": "or", + "type": "text" + }, + { + "bbox": [ + 279, + 101, + 298, + 111 + ], + "score": 0.88, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 101, + 505, + 113 + ], + "score": 1.0, + "content": "of labeled data. Representation learning methods", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 125 + ], + "score": 1.0, + "content": "use a classifier to discriminate an unsupervised representation, and optimize it solely for the super-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 135 + ], + "score": 1.0, + "content": "vised objective on labeled data. Label-propagation methods on the other hand further constrain the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 505, + 145 + ], + "score": 1.0, + "content": "classifier with smoothness and entropy criteria on unlabeled data, making the additional assumption", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 374, + 156 + ], + "score": 1.0, + "content": "that all training images fit into a single (unknown) testing category.", + "type": "text" + }, + { + "bbox": [ + 375, + 145, + 381, + 156 + ], + "score": 0.76, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "denotes methods implemented", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "in this work, fixed and fine-tuned denote whether the feature extractor is allowed to accommodate", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 208, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 208, + 179 + ], + "score": 1.0, + "content": "the supervised objective.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "table_body", + "bbox": [ + 153, + 181, + 456, + 386 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 153, + 181, + 456, + 386 + ], + "spans": [ + { + "bbox": [ + 153, + 181, + 456, + 386 + ], + "score": 0.91, + "html": "
Labeled data 1% 10% 100% Method Top-5 accuracy
Supervised baseline 44.1 82.1 93.9 Methods using label-propagation:
Pseudolabeling (Zhai et al., 2019) 51.6 82.4 VAT + Entropy Minimization (Zhai et al., 2019) 47.0 83.4 Unsup. Data Augmentation (Xie et al., 2019) 88.5
Rotation+ VAT+Ent.Min. (Zhai et al.,2019) 91.2 95.0 =
Methods using representation learning only:
Instance Discrimination (Wu et al., 2018) 39.2 77.4
Rotation (Zhai et al., 2019) 57.5 86.4
+ResNet trained on BigBiGAN (fixed) 55.2 78.8 87.0
+ResNet trained on AMDIM (fixed) 67.4 85.8 92.2
+ResNet trained on CPC v2 (fixed) 72.3 89.1 94.4 +ResNet trained on CPC v2 (fine-tuned) 72.9 89.5 95.2
", + "type": "table", + "image_path": "e969618e28fc2ffa8c76067cea0310a604dd1bb350cd5a5534dd7265b81333a6.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 153, + 181, + 456, + 194.66666666666666 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 153, + 194.66666666666666, + 456, + 208.33333333333331 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 153, + 208.33333333333331, + 456, + 221.99999999999997 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 153, + 221.99999999999997, + 456, + 235.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 153, + 235.66666666666663, + 456, + 249.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 153, + 249.3333333333333, + 456, + 262.99999999999994 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 153, + 262.99999999999994, + 456, + 276.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 153, + 276.66666666666663, + 456, + 290.3333333333333 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 153, + 290.3333333333333, + 456, + 304.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 153, + 304.0, + 456, + 317.6666666666667 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 153, + 317.6666666666667, + 456, + 331.33333333333337 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 153, + 331.33333333333337, + 456, + 345.00000000000006 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 153, + 345.00000000000006, + 456, + 358.66666666666674 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 153, + 358.66666666666674, + 456, + 372.3333333333334 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 153, + 372.3333333333334, + 456, + 386.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + } + ], + "index": 9.25 + }, + { + "type": "text", + "bbox": [ + 107, + 421, + 505, + 487 + ], + "lines": [], + "index": 25.5, + "bbox_fs": [ + 104, + 421, + 506, + 489 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 512 + ], + "score": 1.0, + "content": "Other unsupervised representations. How well does the CPC representation compare to other", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 104, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "representations that have been learned in an unsupervised manner? If predictable representations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 522, + 504, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 504, + 534 + ], + "score": 1.0, + "content": "are uniquely suited for efficient classification, we would expect other methods within this fam-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "ily to perform similarly, and other model classes less so. Table 2 compares our best model with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "other works on efficient recognition. We consider three objectives from different model classes:", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "self-supervised learning with rotation prediction (Zhai et al., 2019), large-scale adversarial feature", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "learning (BigBiGAN, Donahue & Simonyan, 2019), and another contrastive prediction objective", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 589 + ], + "score": 1.0, + "content": "(AMDIM, Bachman et al., 2019). Zhai et al. (2019) evaluate the low-data classification perfor-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "mance of representations learned with rotation prediction using a similar paradigm and architecture", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 597, + 504, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 597, + 344, + 610 + ], + "score": 1.0, + "content": "(ResNet-152), hence we report their results directly. Given", + "type": "text" + }, + { + "bbox": [ + 344, + 598, + 359, + 609 + ], + "score": 0.85, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 597, + 504, + 610 + ], + "score": 1.0, + "content": "of ImageNet, their method achieves", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 608, + 504, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 133, + 620 + ], + "score": 0.85, + "content": "5 7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 608, + 438, + 622 + ], + "score": 1.0, + "content": "Top-5 accuracy, consistently with the reduced accuracy of a linear classifier (", + "type": "text" + }, + { + "bbox": [ + 438, + 609, + 465, + 620 + ], + "score": 0.82, + "content": "5 5 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 608, + 477, + 622 + ], + "score": 1.0, + "content": "vs", + "type": "text" + }, + { + "bbox": [ + 477, + 609, + 504, + 620 + ], + "score": 0.85, + "content": "6 5 . 9 \\%", + "type": "inline_equation" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 619, + 152, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 152, + 632 + ], + "score": 1.0, + "content": "for CPC)1.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 500, + 506, + 632 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 650 + ], + "score": 1.0, + "content": "Because BigBiGAN and AMDIM achieve stronger linear classification accuracy than rotation pre-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 140, + 661 + ], + "score": 1.0, + "content": "diction", + "type": "text" + }, + { + "bbox": [ + 141, + 648, + 168, + 659 + ], + "score": 0.83, + "content": "6 1 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 648, + 187, + 661 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 188, + 648, + 215, + 659 + ], + "score": 0.86, + "content": "6 8 . 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "Top-1 accuracy, respectively), we might expect better performance on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 672 + ], + "score": 1.0, + "content": "efficient classification as well. Since their authors do not report results on efficient classification we", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 670, + 506, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 506, + 683 + ], + "score": 1.0, + "content": "evaluated these representations using the same paradigm we used for evaluating CPC, stacking a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 680, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 228, + 694 + ], + "score": 1.0, + "content": "ResNet classifier on top of the", + "type": "text" + }, + { + "bbox": [ + 229, + 681, + 276, + 691 + ], + "score": 0.9, + "content": "7 \\times 7 \\times 8 1 9 2", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 680, + 410, + 694 + ], + "score": 1.0, + "content": "latents of the BigBiGAN and the", + "type": "text" + }, + { + "bbox": [ + 410, + 681, + 457, + 692 + ], + "score": 0.9, + "content": "7 \\times 7 \\times 2 5 6 0", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 680, + 505, + 694 + ], + "score": 1.0, + "content": "grid of fea-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "ture vectors of AMDIM. We found fine-tuned representations to yield only marginal gains over fixed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 130, + 95 + ], + "score": 1.0, + "content": "ones", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 131, + 83, + 158, + 93 + ], + "score": 0.84, + "content": "7 2 . 9 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 158, + 83, + 209, + 95 + ], + "score": 1.0, + "content": "compared to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 210, + 83, + 237, + 93 + ], + "score": 0.89, + "content": "7 2 . 3 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 237, + 83, + 325, + 95 + ], + "score": 1.0, + "content": "Top-5 accuracy given", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 325, + 83, + 340, + 93 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 340, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "of labels), hence for simplicity we evalu-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "ate BigBiGAN and AMDIM on this task while keeping them fixed. We re-tune the hyper-parameters", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "of the classifier (including optimization, regularization, etc.) for each of these representations sep-", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "arately. Although these methods achieve similar performance in terms of linear classification, we", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 411, + 139 + ], + "score": 1.0, + "content": "find them to achieve very different results in efficient classification. Given", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 411, + 127, + 426, + 137 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 426, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "of ImageNet, clas-", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 284, + 150 + ], + "score": 1.0, + "content": "sifiers trained on top of BigBiGAN achieve", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 284, + 137, + 312, + 148 + ], + "score": 0.88, + "content": "5 5 . 2 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 312, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "Top-5 accuracy, similarly to rotation prediction", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 107, + 149, + 140, + 160 + ], + "score": 0.87, + "content": "( 5 7 . 5 \\% )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 140, + 149, + 354, + 161 + ], + "score": 1.0, + "content": ", despite its increased linear classification accuracy", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 354, + 149, + 375, + 159 + ], + "score": 0.87, + "content": "( + 6 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 376, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "relative to rotation prediction).", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "In contrast, AMDIM (which also belongs to the family of contrastive prediction methods) achieves", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 134, + 181 + ], + "score": 0.88, + "content": "6 7 . 4 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 134, + 170, + 505, + 183 + ], + "score": 1.0, + "content": "on this same task. Again, its increased linear classification accuracy did not entail an increase", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "in data-efficiency. Nevertheless, in line with our initial hypothesis, we find that contrastive predic-", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "tion methods such as CPC surpass other approaches in our efficient classification experiments, and", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 435, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 435, + 216 + ], + "score": 1.0, + "content": "that linear classification performance is not perfectly correlated with these results.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 636, + 506, + 705 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 214 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 130, + 95 + ], + "score": 1.0, + "content": "ones", + "type": "text" + }, + { + "bbox": [ + 131, + 83, + 158, + 93 + ], + "score": 0.84, + "content": "7 2 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 83, + 209, + 95 + ], + "score": 1.0, + "content": "compared to", + "type": "text" + }, + { + "bbox": [ + 210, + 83, + 237, + 93 + ], + "score": 0.89, + "content": "7 2 . 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 83, + 325, + 95 + ], + "score": 1.0, + "content": "Top-5 accuracy given", + "type": "text" + }, + { + "bbox": [ + 325, + 83, + 340, + 93 + ], + "score": 0.86, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "of labels), hence for simplicity we evalu-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "ate BigBiGAN and AMDIM on this task while keeping them fixed. We re-tune the hyper-parameters", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "of the classifier (including optimization, regularization, etc.) for each of these representations sep-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "arately. Although these methods achieve similar performance in terms of linear classification, we", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 411, + 139 + ], + "score": 1.0, + "content": "find them to achieve very different results in efficient classification. Given", + "type": "text" + }, + { + "bbox": [ + 411, + 127, + 426, + 137 + ], + "score": 0.84, + "content": "1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "of ImageNet, clas-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 284, + 150 + ], + "score": 1.0, + "content": "sifiers trained on top of BigBiGAN achieve", + "type": "text" + }, + { + "bbox": [ + 284, + 137, + 312, + 148 + ], + "score": 0.88, + "content": "5 5 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "Top-5 accuracy, similarly to rotation prediction", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 107, + 149, + 140, + 160 + ], + "score": 0.87, + "content": "( 5 7 . 5 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 149, + 354, + 161 + ], + "score": 1.0, + "content": ", despite its increased linear classification accuracy", + "type": "text" + }, + { + "bbox": [ + 354, + 149, + 375, + 159 + ], + "score": 0.87, + "content": "( + 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "relative to rotation prediction).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "In contrast, AMDIM (which also belongs to the family of contrastive prediction methods) achieves", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 134, + 181 + ], + "score": 0.88, + "content": "6 7 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 170, + 505, + 183 + ], + "score": 1.0, + "content": "on this same task. Again, its increased linear classification accuracy did not entail an increase", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "in data-efficiency. Nevertheless, in line with our initial hypothesis, we find that contrastive predic-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "tion methods such as CPC surpass other approaches in our efficient classification experiments, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 435, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 435, + 216 + ], + "score": 1.0, + "content": "that linear classification performance is not perfectly correlated with these results.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "Other semi-supervised techniques A separate class of methods for low-data classification at-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "score": 1.0, + "content": "tempts to propagate the knowledge extracted from the subset of labeled examples to unlabeled", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "score": 1.0, + "content": "examples while being invariant to augmentation or other perturbations. These methods generally", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 263, + 504, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 504, + 275 + ], + "score": 1.0, + "content": "depend on the quality of the classifier’s predictions, and as such tend to fare well when given in-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "termediate amounts of data. Although not sufficient in themselves (Unsupervised Data Augmenta-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "tion (Xie et al., 2019), Virtual Adversarial Training (Miyato et al., 2018) and entropy minimization", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 295, + 504, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 397, + 308 + ], + "score": 1.0, + "content": "(Grandvalet & Bengio, 2005), and pseudo-labeling (Lee, 2013) achieve", + "type": "text" + }, + { + "bbox": [ + 397, + 296, + 424, + 306 + ], + "score": 0.86, + "content": "8 5 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 296, + 428, + 308 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 428, + 296, + 455, + 306 + ], + "score": 0.83, + "content": "8 3 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 296, + 477, + 308 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 477, + 295, + 504, + 306 + ], + "score": 0.86, + "content": "8 2 . 4 \\%", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 193, + 320 + ], + "score": 1.0, + "content": "Top-5 accuracy with", + "type": "text" + }, + { + "bbox": [ + 193, + 307, + 213, + 317 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 306, + 325, + 320 + ], + "score": 1.0, + "content": "of labels, compared to our", + "type": "text" + }, + { + "bbox": [ + 325, + 307, + 353, + 317 + ], + "score": 0.88, + "content": "8 9 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 306, + 506, + 320 + ], + "score": 1.0, + "content": ") when combined with representation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 455, + 331 + ], + "score": 1.0, + "content": "learning (e.g. rotation prediction Zhai et al., 2019) they can provide considerable gains", + "type": "text" + }, + { + "bbox": [ + 455, + 318, + 483, + 329 + ], + "score": 0.8, + "content": "9 1 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 329, + 506, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 341 + ], + "score": 1.0, + "content": "5 accuracy). It is therefore surprising that CPC representations alone can enable accuracy that is", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "comparable to that of these methods, and investigating to what extent they can be combined would", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 351, + 259, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 259, + 362 + ], + "score": 1.0, + "content": "be an interesting topic of future work.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 378, + 417, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 378, + 419, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 419, + 391 + ], + "score": 1.0, + "content": "4.3 TRANSFER LEARNING: IMAGE DETECTION ON PASCAL VOC 2007", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 413 + ], + "score": 1.0, + "content": "We next investigate transfer performance on object detection on the PASCAL-2007 dataset, which", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "reflects the practical scenario where a representation must be trained on a dataset with different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "statistics than the dataset of interest. This dataset also tests the efficiency of the representation as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 504, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 504, + 445 + ], + "score": 1.0, + "content": "it only contains 5011 labeled images to train from. In this setting, we replaced the neural network", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 145, + 456 + ], + "score": 1.0, + "content": "classifier", + "type": "text" + }, + { + "bbox": [ + 145, + 444, + 158, + 456 + ], + "score": 0.9, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "used previously with a Faster-RCNN (Ren et al., 2015) image detection architecture,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "and use the pre-trained feature extractor on ImageNet. As before, we first trained the Faster-RCNN", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "model while keeping the feature extractor fixed, then fine-tuned the entire model end-to-end. Table", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "3 displays our results compared to other methods. Most competing methods, which optimize a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 487, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 506, + 502 + ], + "score": 1.0, + "content": "single unsupervised objective on ImageNet before fine-tuning on PASCAL detection, attain around", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 498, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 126, + 509 + ], + "score": 0.86, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 498, + 506, + 513 + ], + "score": 1.0, + "content": "mean average precision. Leveraging larger unlabeled datasets increases their performance up", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 117, + 522 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 510, + 144, + 520 + ], + "score": 0.86, + "content": "6 7 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "(Caron et al., 2019). Combining multiple forms of self-supervision enables them to reach", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 519, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 134, + 532 + ], + "score": 0.86, + "content": "7 0 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 519, + 506, + 534 + ], + "score": 1.0, + "content": "(Doersch & Zisserman, 2017). The proposed method, which learns only from ImageNet data", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 299, + 545 + ], + "score": 1.0, + "content": "using a single unsupervised objective, reaches", + "type": "text" + }, + { + "bbox": [ + 299, + 532, + 327, + 542 + ], + "score": 0.87, + "content": "7 0 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "when equipped with a ResNet-101 feature", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 144, + 555 + ], + "score": 1.0, + "content": "extractor", + "type": "text" + }, + { + "bbox": [ + 145, + 543, + 155, + 554 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "(as for most competing methods (Doersch & Zisserman, 2017) but not all (Caron et al.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 411, + 566 + ], + "score": 1.0, + "content": "2018; 2019). Equipped with the more powerful ResNet-161 feature extractor", + "type": "text" + }, + { + "bbox": [ + 411, + 554, + 421, + 565 + ], + "score": 0.89, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 553, + 505, + 566 + ], + "score": 1.0, + "content": ", our method reaches", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 133, + 575 + ], + "score": 0.87, + "content": "7 2 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 563, + 264, + 577 + ], + "score": 1.0, + "content": ". Importantly, this result is only", + "type": "text" + }, + { + "bbox": [ + 264, + 565, + 279, + 575 + ], + "score": 0.84, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "short of the performance attained by purely supervised", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 576, + 491, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 491, + 588 + ], + "score": 1.0, + "content": "transfer learning, which we obtain by using all ImageNet labels before transferring to PASCAL.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 33 + }, + { + "type": "title", + "bbox": [ + 108, + 606, + 190, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 192, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 192, + 621 + ], + "score": 1.0, + "content": "5 DISCUSSION", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "We asked whether CPC could enable data-efficient image recognition, and found that it indeed", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "greatly improves the accuracy of classifiers and object detectors when given small amounts of la-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "beled data. Surprisingly, CPC even improves results given ImageNet-scale labels. Our results show", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "that there is still room for improvement using relatively straightforward changes such as augmenta-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "tion, optimization, and network architecture. Furthermore, we found that the standard method for", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "evaluating unsupervised representations—linear classification—is only partially predictive of effi-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "cient recognition performance, suggesting that further research should focus on efficient recognition", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "as a standalone benchmark. Overall, these results open the door toward research on problems where", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 340, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 340, + 733 + ], + "score": 1.0, + "content": "data is naturally limited, e.g. medical imaging or robotics.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 214 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 83, + 506, + 216 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "Other semi-supervised techniques A separate class of methods for low-data classification at-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 253 + ], + "score": 1.0, + "content": "tempts to propagate the knowledge extracted from the subset of labeled examples to unlabeled", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "score": 1.0, + "content": "examples while being invariant to augmentation or other perturbations. These methods generally", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 263, + 504, + 275 + ], + "spans": [ + { + "bbox": [ + 106, + 263, + 504, + 275 + ], + "score": 1.0, + "content": "depend on the quality of the classifier’s predictions, and as such tend to fare well when given in-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 505, + 287 + ], + "score": 1.0, + "content": "termediate amounts of data. Although not sufficient in themselves (Unsupervised Data Augmenta-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "tion (Xie et al., 2019), Virtual Adversarial Training (Miyato et al., 2018) and entropy minimization", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 295, + 504, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 397, + 308 + ], + "score": 1.0, + "content": "(Grandvalet & Bengio, 2005), and pseudo-labeling (Lee, 2013) achieve", + "type": "text" + }, + { + "bbox": [ + 397, + 296, + 424, + 306 + ], + "score": 0.86, + "content": "8 5 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 296, + 428, + 308 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 428, + 296, + 455, + 306 + ], + "score": 0.83, + "content": "8 3 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 296, + 477, + 308 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 477, + 295, + 504, + 306 + ], + "score": 0.86, + "content": "8 2 . 4 \\%", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 193, + 320 + ], + "score": 1.0, + "content": "Top-5 accuracy with", + "type": "text" + }, + { + "bbox": [ + 193, + 307, + 213, + 317 + ], + "score": 0.87, + "content": "10 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 306, + 325, + 320 + ], + "score": 1.0, + "content": "of labels, compared to our", + "type": "text" + }, + { + "bbox": [ + 325, + 307, + 353, + 317 + ], + "score": 0.88, + "content": "8 9 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 306, + 506, + 320 + ], + "score": 1.0, + "content": ") when combined with representation", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 455, + 331 + ], + "score": 1.0, + "content": "learning (e.g. rotation prediction Zhai et al., 2019) they can provide considerable gains", + "type": "text" + }, + { + "bbox": [ + 455, + 318, + 483, + 329 + ], + "score": 0.8, + "content": "9 1 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 329, + 506, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 506, + 341 + ], + "score": 1.0, + "content": "5 accuracy). It is therefore surprising that CPC representations alone can enable accuracy that is", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 353 + ], + "score": 1.0, + "content": "comparable to that of these methods, and investigating to what extent they can be combined would", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 351, + 259, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 259, + 362 + ], + "score": 1.0, + "content": "be an interesting topic of future work.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 230, + 506, + 362 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 378, + 417, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 378, + 419, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 419, + 391 + ], + "score": 1.0, + "content": "4.3 TRANSFER LEARNING: IMAGE DETECTION ON PASCAL VOC 2007", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 400, + 505, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 400, + 505, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 505, + 413 + ], + "score": 1.0, + "content": "We next investigate transfer performance on object detection on the PASCAL-2007 dataset, which", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 506, + 424 + ], + "score": 1.0, + "content": "reflects the practical scenario where a representation must be trained on a dataset with different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 435 + ], + "score": 1.0, + "content": "statistics than the dataset of interest. This dataset also tests the efficiency of the representation as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 434, + 504, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 504, + 445 + ], + "score": 1.0, + "content": "it only contains 5011 labeled images to train from. In this setting, we replaced the neural network", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 145, + 456 + ], + "score": 1.0, + "content": "classifier", + "type": "text" + }, + { + "bbox": [ + 145, + 444, + 158, + 456 + ], + "score": 0.9, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "used previously with a Faster-RCNN (Ren et al., 2015) image detection architecture,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "and use the pre-trained feature extractor on ImageNet. As before, we first trained the Faster-RCNN", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "model while keeping the feature extractor fixed, then fine-tuned the entire model end-to-end. Table", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "3 displays our results compared to other methods. Most competing methods, which optimize a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 487, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 506, + 502 + ], + "score": 1.0, + "content": "single unsupervised objective on ImageNet before fine-tuning on PASCAL detection, attain around", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 498, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 126, + 509 + ], + "score": 0.86, + "content": "65 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 127, + 498, + 506, + 513 + ], + "score": 1.0, + "content": "mean average precision. Leveraging larger unlabeled datasets increases their performance up", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 117, + 522 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 510, + 144, + 520 + ], + "score": 0.86, + "content": "6 7 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "(Caron et al., 2019). Combining multiple forms of self-supervision enables them to reach", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 519, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 134, + 532 + ], + "score": 0.86, + "content": "7 0 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 519, + 506, + 534 + ], + "score": 1.0, + "content": "(Doersch & Zisserman, 2017). The proposed method, which learns only from ImageNet data", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 299, + 545 + ], + "score": 1.0, + "content": "using a single unsupervised objective, reaches", + "type": "text" + }, + { + "bbox": [ + 299, + 532, + 327, + 542 + ], + "score": 0.87, + "content": "7 0 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "when equipped with a ResNet-101 feature", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 144, + 555 + ], + "score": 1.0, + "content": "extractor", + "type": "text" + }, + { + "bbox": [ + 145, + 543, + 155, + 554 + ], + "score": 0.86, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "(as for most competing methods (Doersch & Zisserman, 2017) but not all (Caron et al.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 411, + 566 + ], + "score": 1.0, + "content": "2018; 2019). Equipped with the more powerful ResNet-161 feature extractor", + "type": "text" + }, + { + "bbox": [ + 411, + 554, + 421, + 565 + ], + "score": 0.89, + "content": "f _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 553, + 505, + 566 + ], + "score": 1.0, + "content": ", our method reaches", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 133, + 575 + ], + "score": 0.87, + "content": "7 2 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 563, + 264, + 577 + ], + "score": 1.0, + "content": ". Importantly, this result is only", + "type": "text" + }, + { + "bbox": [ + 264, + 565, + 279, + 575 + ], + "score": 0.84, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "short of the performance attained by purely supervised", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 576, + 491, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 491, + 588 + ], + "score": 1.0, + "content": "transfer learning, which we obtain by using all ImageNet labels before transferring to PASCAL.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 33, + "bbox_fs": [ + 104, + 400, + 506, + 588 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 606, + 190, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 192, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 192, + 621 + ], + "score": 1.0, + "content": "5 DISCUSSION", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 106, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "We asked whether CPC could enable data-efficient image recognition, and found that it indeed", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "greatly improves the accuracy of classifiers and object detectors when given small amounts of la-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "beled data. Surprisingly, CPC even improves results given ImageNet-scale labels. Our results show", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "that there is still room for improvement using relatively straightforward changes such as augmenta-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "tion, optimization, and network architecture. Furthermore, we found that the standard method for", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "evaluating unsupervised representations—linear classification—is only partially predictive of effi-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "cient recognition performance, suggesting that further research should focus on efficient recognition", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "as a standalone benchmark. Overall, these results open the door toward research on problems where", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 340, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 340, + 733 + ], + "score": 1.0, + "content": "data is naturally limited, e.g. medical imaging or robotics.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 632, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 170, + 464, + 362 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 166 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 202, + 101 + ], + "score": 1.0, + "content": "Table 3: Faster-RCNN", + "type": "text" + }, + { + "bbox": [ + 203, + 90, + 216, + 101 + ], + "score": 0.88, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 89, + 271, + 101 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 271, + 89, + 297, + 100 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "of PASCAL labels. Comparison of PASCAL 2007", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "image detection accuracy to other transfer methods. The supervised baseline learns from the entire", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "labeled ImageNet dataset and fine-tunes for PASCAL detection. The second class of methods learns", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 123, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 134 + ], + "score": 1.0, + "content": "from the same unlabeled images before transferring. All of these methods pre-train on the ImageNet", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "dataset, except for DeeperCluster which learns from the larger, but uncurated, YFCC100M dataset", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 464, + 157 + ], + "score": 1.0, + "content": "(Thomee et al., 2015). All results are reported in terms of mean average precision (mAP).", + "type": "text" + }, + { + "bbox": [ + 465, + 145, + 471, + 156 + ], + "score": 0.73, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 144, + 505, + 157 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 156, + 249, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 249, + 167 + ], + "score": 1.0, + "content": "methods implemented in this work.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 147, + 170, + 464, + 362 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 170, + 464, + 362 + ], + "spans": [ + { + "bbox": [ + 147, + 170, + 464, + 362 + ], + "score": 0.527, + "html": "
MethodmAP
Transfer from labeled data: Supervised - ResNet-15274.7
Transfer from unlabeled data: Exemplar (Ex) (Dosovitskiy et al., 2014) Motion Segmentation (MS) (Pathak et al.,2016)60.9
Colorization (Col) (Zhang et al.,2016) Relative Position (RP) (Doersch et al.,2015)61.1 65.5
Combination of Ex + MS + Col +RP (Doersch & Zisserman,2017)66.8 70.5
Instance Discrimination (Wu et al., 2018)65.4
Deep Cluster (Caron et al., 2018) Deeper Cluster (Caron et al., 2019)65.9
", + "type": "table", + "image_path": "d005a0a972800a539ffdecef950a4659796c2a8300a3ab9b860336449e53c88e.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 147, + 170, + 464, + 234.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 147, + 234.0, + 464, + 298.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 147, + 298.0, + 464, + 362.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "Furthermore, images are far from the only domain where unsupervised representation learning is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "important: for example, unsupervised learning is already a critical step in language (Mikolov et al.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "2013; Devlin et al., 2018), and shows promise in domains like audio (van den Oord et al., 2018;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "Arandjelovic & Zisserman, 2018; 2017), video (Jing & Tian, 2018; Misra et al., 2016), and robotic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "manipulation (Pinto & Gupta, 2016; Pinto et al., 2016; Sermanet et al., 2018). Currently much", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 439, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 453 + ], + "score": 1.0, + "content": "self-supervised work builds upon tasks tailored for a specific domain (often images), which may", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "not be easily adapted to other domains. Contrastive prediction methods, including the techniques", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "suggested in this paper, are task agnostic and could therefore serve as a unifying framework for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "integrating these tasks and modalities. This generality is particularly useful given that many real-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "world environments are inherently multimodal, e.g. robotic environments which can have vision,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "audio, touch, proprioception, action, and more over long temporal sequences. Given the importance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "score": 1.0, + "content": "of increasing the amounts of self-supervision (via additional directions of prediction), integrating", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "these modalities and tasks could lead to unsupervised representations which rival the efficiency and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 528, + 237, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 237, + 541 + ], + "score": 1.0, + "content": "effectiveness of biological ones.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 107, + 556, + 175, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 176, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 176, + 569 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 574, + 486, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 487, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 487, + 588 + ], + "score": 1.0, + "content": "Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. In ICCV, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 603, + 384, + 616 + ], + "spans": [ + { + "bbox": [ + 116, + 603, + 384, + 616 + ], + "score": 1.0, + "content": "International Conference on Computer Vision, pp. 609–617, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 620, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 635 + ], + "score": 1.0, + "content": "Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 632, + 364, + 645 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 364, + 645 + ], + "score": 1.0, + "content": "Conference on Computer Vision (ECCV), pp. 435–451, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 650, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 664 + ], + "score": 1.0, + "content": "Lei Jimmy Ba, Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 114, + 659, + 144, + 675 + ], + "spans": [ + { + "bbox": [ + 114, + 659, + 144, + 675 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 691, + 415, + 704 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 415, + 704 + ], + "score": 1.0, + "content": "mutual information across views. arXiv preprint arXiv:1906.00910, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "H.B. Barlow. Unsupervised learning. Neural Computation, 1(3):295–311, 1989. doi: 10.1162/neco.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 720, + 173, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 173, + 732 + ], + "score": 1.0, + "content": "1989.1.3.295.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 170, + 464, + 362 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 166 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 202, + 101 + ], + "score": 1.0, + "content": "Table 3: Faster-RCNN", + "type": "text" + }, + { + "bbox": [ + 203, + 90, + 216, + 101 + ], + "score": 0.88, + "content": "h _ { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 89, + 271, + 101 + ], + "score": 1.0, + "content": "trained with", + "type": "text" + }, + { + "bbox": [ + 271, + 89, + 297, + 100 + ], + "score": 0.86, + "content": "100 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "of PASCAL labels. Comparison of PASCAL 2007", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "image detection accuracy to other transfer methods. The supervised baseline learns from the entire", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "labeled ImageNet dataset and fine-tunes for PASCAL detection. The second class of methods learns", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 123, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 134 + ], + "score": 1.0, + "content": "from the same unlabeled images before transferring. All of these methods pre-train on the ImageNet", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "dataset, except for DeeperCluster which learns from the larger, but uncurated, YFCC100M dataset", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 464, + 157 + ], + "score": 1.0, + "content": "(Thomee et al., 2015). All results are reported in terms of mean average precision (mAP).", + "type": "text" + }, + { + "bbox": [ + 465, + 145, + 471, + 156 + ], + "score": 0.73, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 144, + 505, + 157 + ], + "score": 1.0, + "content": "denotes", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 156, + 249, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 249, + 167 + ], + "score": 1.0, + "content": "methods implemented in this work.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "table_body", + "bbox": [ + 147, + 170, + 464, + 362 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 170, + 464, + 362 + ], + "spans": [ + { + "bbox": [ + 147, + 170, + 464, + 362 + ], + "score": 0.527, + "html": "
MethodmAP
Transfer from labeled data: Supervised - ResNet-15274.7
Transfer from unlabeled data: Exemplar (Ex) (Dosovitskiy et al., 2014) Motion Segmentation (MS) (Pathak et al.,2016)60.9
Colorization (Col) (Zhang et al.,2016) Relative Position (RP) (Doersch et al.,2015)61.1 65.5
Combination of Ex + MS + Col +RP (Doersch & Zisserman,2017)66.8 70.5
Instance Discrimination (Wu et al., 2018)65.4
Deep Cluster (Caron et al., 2018) Deeper Cluster (Caron et al., 2019)65.9
", + "type": "table", + "image_path": "d005a0a972800a539ffdecef950a4659796c2a8300a3ab9b860336449e53c88e.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 147, + 170, + 464, + 234.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 147, + 234.0, + 464, + 298.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 147, + 298.0, + 464, + 362.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "Furthermore, images are far from the only domain where unsupervised representation learning is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "important: for example, unsupervised learning is already a critical step in language (Mikolov et al.,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "2013; Devlin et al., 2018), and shows promise in domains like audio (van den Oord et al., 2018;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "Arandjelovic & Zisserman, 2018; 2017), video (Jing & Tian, 2018; Misra et al., 2016), and robotic", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 505, + 442 + ], + "score": 1.0, + "content": "manipulation (Pinto & Gupta, 2016; Pinto et al., 2016; Sermanet et al., 2018). Currently much", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 439, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 453 + ], + "score": 1.0, + "content": "self-supervised work builds upon tasks tailored for a specific domain (often images), which may", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 450, + 506, + 465 + ], + "score": 1.0, + "content": "not be easily adapted to other domains. Contrastive prediction methods, including the techniques", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "suggested in this paper, are task agnostic and could therefore serve as a unifying framework for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "integrating these tasks and modalities. This generality is particularly useful given that many real-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "world environments are inherently multimodal, e.g. robotic environments which can have vision,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "audio, touch, proprioception, action, and more over long temporal sequences. Given the importance", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 520 + ], + "score": 1.0, + "content": "of increasing the amounts of self-supervision (via additional directions of prediction), integrating", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "these modalities and tasks could lead to unsupervised representations which rival the efficiency and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 528, + 237, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 237, + 541 + ], + "score": 1.0, + "content": "effectiveness of biological ones.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 385, + 506, + 541 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 556, + 175, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 176, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 176, + 569 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 574, + 486, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 572, + 487, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 487, + 588 + ], + "score": 1.0, + "content": "Pulkit Agrawal, Joao Carreira, and Jitendra Malik. Learning to see by moving. In ICCV, 2015.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 572, + 487, + 588 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 504, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 592, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 605 + ], + "score": 1.0, + "content": "Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 603, + 384, + 616 + ], + "spans": [ + { + "bbox": [ + 116, + 603, + 384, + 616 + ], + "score": 1.0, + "content": "International Conference on Computer Vision, pp. 609–617, 2017.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 592, + 505, + 616 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 620, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 635 + ], + "score": 1.0, + "content": "Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 632, + 364, + 645 + ], + "spans": [ + { + "bbox": [ + 116, + 632, + 364, + 645 + ], + "score": 1.0, + "content": "Conference on Computer Vision (ECCV), pp. 435–451, 2018.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 620, + 505, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 650, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 664 + ], + "score": 1.0, + "content": "Lei Jimmy Ba, Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 114, + 659, + 144, + 675 + ], + "spans": [ + { + "bbox": [ + 114, + 659, + 144, + 675 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 649, + 505, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 680, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 505, + 694 + ], + "score": 1.0, + "content": "Philip Bachman, R Devon Hjelm, and William Buchwalter. Learning representations by maximizing", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 691, + 415, + 704 + ], + "spans": [ + { + "bbox": [ + 116, + 691, + 415, + 704 + ], + "score": 1.0, + "content": "mutual information across views. arXiv preprint arXiv:1906.00910, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 679, + 505, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "H.B. Barlow. Unsupervised learning. Neural Computation, 1(3):295–311, 1989. doi: 10.1162/neco.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 720, + 173, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 173, + 732 + ], + "score": 1.0, + "content": "1989.1.3.295.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsu-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "pervised learning of visual features. In The European Conference on Computer Vision (ECCV),", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 187, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 187, + 117 + ], + "score": 1.0, + "content": "September 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 124, + 502, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 503, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 503, + 138 + ], + "score": 1.0, + "content": "Mathilde Caron, Piotr Bojanowski, Julien Mairal, and Armand Joulin. Leveraging large-scale uncu-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 136, + 376, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 136, + 376, + 148 + ], + "score": 1.0, + "content": "rated data for unsupervised pre-training of visual features. 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 503, + 178 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 169 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 166, + 459, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 166, + 459, + 179 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 504, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 198, + 433, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 198, + 433, + 210 + ], + "score": 1.0, + "content": "the IEEE International Conference on Computer Vision, pp. 2051–2060, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 506, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 232 + ], + "score": 1.0, + "content": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 228, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 115, + 228, + 505, + 243 + ], + "score": 1.0, + "content": "context prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 240, + 192, + 251 + ], + "spans": [ + { + "bbox": [ + 116, + 240, + 192, + 251 + ], + "score": 1.0, + "content": "1422–1430, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 104, + 259, + 504, + 283 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "score": 1.0, + "content": "Jeff Donahue and Karen Simonyan. Large scale adversarial representation learning. arXiv preprint", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 271, + 219, + 282 + ], + "spans": [ + { + "bbox": [ + 116, + 271, + 219, + 282 + ], + "score": 1.0, + "content": "arXiv:1907.02544, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 305 + ], + "score": 1.0, + "content": "Jeff Donahue, Philipp Krahenb ¨ uhl, and Trevor Darrell. Adversarial feature learning. ¨ arXiv preprint", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 302, + 219, + 315 + ], + "spans": [ + { + "bbox": [ + 115, + 302, + 219, + 315 + ], + "score": 1.0, + "content": "arXiv:1605.09782, 2016.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 322, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 504, + 335 + ], + "score": 1.0, + "content": "Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discrimina-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "tive unsupervised feature learning with convolutional neural networks. In Advances in neural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 344, + 326, + 357 + ], + "spans": [ + { + "bbox": [ + 116, + 344, + 326, + 357 + ], + "score": 1.0, + "content": "information processing systems, pp. 766–774, 2014.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 109, + 364, + 504, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 375, + 408, + 388 + ], + "spans": [ + { + "bbox": [ + 116, + 375, + 408, + 388 + ], + "score": 1.0, + "content": "The pascal visual object classes challenge 2007 (voc2007) results. 2007.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 504, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 410 + ], + "score": 1.0, + "content": "Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 407, + 389, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 389, + 419 + ], + "score": 1.0, + "content": "predicting image rotations. arXiv preprint arXiv:1803.07728, 2018.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 105, + 427, + 504, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Ad-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 438, + 394, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 394, + 451 + ], + "score": 1.0, + "content": "vances in neural information processing systems, pp. 529–536, 2005.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 108, + 457, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "score": 1.0, + "content": "for unnormalized statistical models. In Proceedings of the Thirteenth International Conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 481, + 354, + 492 + ], + "spans": [ + { + "bbox": [ + 116, + 481, + 354, + 492 + ], + "score": 1.0, + "content": "on Artificial Intelligence and Statistics, pp. 297–304, 2010.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 520, + 186, + 535 + ], + "spans": [ + { + "bbox": [ + 116, + 520, + 186, + 535 + ], + "score": 1.0, + "content": "770–778, 2016a.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 553, + 462, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 553, + 462, + 567 + ], + "score": 1.0, + "content": "networks. In European conference on computer vision, pp. 630–645. Springer, 2016b.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 573, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "G.E. Hinton, T.J. Sejnowski, H.H.M.I.C.N.L.T.J. Sejnowski, and T.A. Poggio. Unsupervised", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 585, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 505, + 596 + ], + "score": 1.0, + "content": "Learning: Foundations of Neural Computation. A Bradford Book. MIT Press, 1999. ISBN", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 596, + 186, + 607 + ], + "spans": [ + { + "bbox": [ + 116, + 596, + 186, + 607 + ], + "score": 1.0, + "content": "9780262581684.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 105, + 615, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 629 + ], + "score": 1.0, + "content": "Olivier J. Henaff, Robbe L. T. Goris, and Eero P. Simoncelli. Perceptual straightening of natural ´", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 627, + 280, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 280, + 640 + ], + "score": 1.0, + "content": "videos. Nature Neuroscience, Apr 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 105, + 647, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 661 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 658, + 412, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 412, + 670 + ], + "score": 1.0, + "content": "reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "Dinesh Jayaraman and Kristen Grauman. Learning image representations tied to ego-motion. In", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 689, + 169, + 701 + ], + "spans": [ + { + "bbox": [ + 116, + 689, + 169, + 701 + ], + "score": 1.0, + "content": "ICCV, 2015.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Longlong Jing and Yingli Tian. Self-supervised spatiotemporal feature learning by video geometric", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 347, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 347, + 732 + ], + "score": 1.0, + "content": "transformations. arXiv preprint arXiv:1811.11387, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsu-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "pervised learning of visual features. In The European Conference on Computer Vision (ECCV),", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 187, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 187, + 117 + ], + "score": 1.0, + "content": "September 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 124, + 502, + 147 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 503, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 503, + 138 + ], + "score": 1.0, + "content": "Mathilde Caron, Piotr Bojanowski, Julien Mairal, and Armand Joulin. Leveraging large-scale uncu-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 136, + 376, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 136, + 376, + 148 + ], + "score": 1.0, + "content": "rated data for unsupervised pre-training of visual features. 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 123, + 503, + 148 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 155, + 503, + 178 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 169 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 166, + 459, + 179 + ], + "spans": [ + { + "bbox": [ + 115, + 166, + 459, + 179 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. CoRR, abs/1810.04805, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 155, + 505, + 179 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 504, + 210 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "Carl Doersch and Andrew Zisserman. Multi-task self-supervised visual learning. In Proceedings of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 198, + 433, + 210 + ], + "spans": [ + { + "bbox": [ + 116, + 198, + 433, + 210 + ], + "score": 1.0, + "content": "the IEEE International Conference on Computer Vision, pp. 2051–2060, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 186, + 505, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 506, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 232 + ], + "score": 1.0, + "content": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 228, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 115, + 228, + 505, + 243 + ], + "score": 1.0, + "content": "context prediction. In Proceedings of the IEEE International Conference on Computer Vision, pp.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 240, + 192, + 251 + ], + "spans": [ + { + "bbox": [ + 116, + 240, + 192, + 251 + ], + "score": 1.0, + "content": "1422–1430, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 216, + 505, + 251 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 259, + 504, + 283 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "score": 1.0, + "content": "Jeff Donahue and Karen Simonyan. Large scale adversarial representation learning. arXiv preprint", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 271, + 219, + 282 + ], + "spans": [ + { + "bbox": [ + 116, + 271, + 219, + 282 + ], + "score": 1.0, + "content": "arXiv:1907.02544, 2019.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 258, + 505, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 289, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 505, + 305 + ], + "score": 1.0, + "content": "Jeff Donahue, Philipp Krahenb ¨ uhl, and Trevor Darrell. Adversarial feature learning. ¨ arXiv preprint", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 302, + 219, + 315 + ], + "spans": [ + { + "bbox": [ + 115, + 302, + 219, + 315 + ], + "score": 1.0, + "content": "arXiv:1605.09782, 2016.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 289, + 505, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 322, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 322, + 504, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 504, + 335 + ], + "score": 1.0, + "content": "Alexey Dosovitskiy, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discrimina-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 334, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 115, + 334, + 505, + 346 + ], + "score": 1.0, + "content": "tive unsupervised feature learning with convolutional neural networks. In Advances in neural", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 344, + 326, + 357 + ], + "spans": [ + { + "bbox": [ + 116, + 344, + 326, + 357 + ], + "score": 1.0, + "content": "information processing systems, pp. 766–774, 2014.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 322, + 505, + 357 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 364, + 504, + 388 + ], + "lines": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 375, + 408, + 388 + ], + "spans": [ + { + "bbox": [ + 116, + 375, + 408, + 388 + ], + "score": 1.0, + "content": "The pascal visual object classes challenge 2007 (voc2007) results. 2007.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 106, + 364, + 505, + 388 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 504, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 410 + ], + "score": 1.0, + "content": "Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 407, + 389, + 419 + ], + "spans": [ + { + "bbox": [ + 115, + 407, + 389, + 419 + ], + "score": 1.0, + "content": "predicting image rotations. arXiv preprint arXiv:1803.07728, 2018.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 395, + 505, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 427, + 504, + 450 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Ad-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 438, + 394, + 451 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 394, + 451 + ], + "score": 1.0, + "content": "vances in neural information processing systems, pp. 529–536, 2005.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 426, + 505, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 457, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "score": 1.0, + "content": "Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle ¨", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 505, + 482 + ], + "score": 1.0, + "content": "for unnormalized statistical models. In Proceedings of the Thirteenth International Conference", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 481, + 354, + 492 + ], + "spans": [ + { + "bbox": [ + 116, + 481, + 354, + 492 + ], + "score": 1.0, + "content": "on Artificial Intelligence and Statistics, pp. 297–304, 2010.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 457, + 505, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 500, + 505, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 499, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 505, + 514 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 520, + 186, + 535 + ], + "spans": [ + { + "bbox": [ + 116, + 520, + 186, + 535 + ], + "score": 1.0, + "content": "770–778, 2016a.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 499, + 505, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 542, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 553, + 462, + 567 + ], + "spans": [ + { + "bbox": [ + 115, + 553, + 462, + 567 + ], + "score": 1.0, + "content": "networks. In European conference on computer vision, pp. 630–645. Springer, 2016b.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 542, + 505, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 573, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "G.E. Hinton, T.J. Sejnowski, H.H.M.I.C.N.L.T.J. Sejnowski, and T.A. Poggio. Unsupervised", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 585, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 585, + 505, + 596 + ], + "score": 1.0, + "content": "Learning: Foundations of Neural Computation. A Bradford Book. MIT Press, 1999. ISBN", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 596, + 186, + 607 + ], + "spans": [ + { + "bbox": [ + 116, + 596, + 186, + 607 + ], + "score": 1.0, + "content": "9780262581684.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 106, + 573, + 505, + 607 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 615, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 505, + 629 + ], + "score": 1.0, + "content": "Olivier J. Henaff, Robbe L. T. Goris, and Eero P. Simoncelli. Perceptual straightening of natural ´", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 627, + 280, + 640 + ], + "spans": [ + { + "bbox": [ + 116, + 627, + 280, + 640 + ], + "score": 1.0, + "content": "videos. Nature Neuroscience, Apr 2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 614, + 505, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 647, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 661 + ], + "score": 1.0, + "content": "Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 658, + 412, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 658, + 412, + 670 + ], + "score": 1.0, + "content": "reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 646, + 505, + 670 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "Dinesh Jayaraman and Kristen Grauman. Learning image representations tied to ego-motion. In", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 689, + 169, + 701 + ], + "spans": [ + { + "bbox": [ + 116, + 689, + 169, + 701 + ], + "score": 1.0, + "content": "ICCV, 2015.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 677, + 505, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Longlong Jing and Yingli Tian. Self-supervised spatiotemporal feature learning by video geometric", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 347, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 347, + 732 + ], + "score": 1.0, + "content": "transformations. arXiv preprint arXiv:1811.11387, 2018.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "learning with deep generative models. In Advances in neural information processing systems, pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 191, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 191, + 115 + ], + "score": 1.0, + "content": "3581–3589, 2014.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Alexander Kolesnikov, Xiaohua Zhai, and Lucas Beyer. Revisiting self-supervised visual representa-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "score": 1.0, + "content": "tion learning. CoRR, abs/1901.09005, 2019. URL http://arxiv.org/abs/1901.09005.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 506, + 168 + ], + "score": 1.0, + "content": "Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 437, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 437, + 177 + ], + "score": 1.0, + "content": "through probabilistic program induction. Science, 350(6266):1332–1338, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 105, + 183, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "score": 1.0, + "content": "Barbara Landau, Linda B Smith, and Susan S Jones. The importance of shape in early lexical", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 195, + 335, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 335, + 207 + ], + "score": 1.0, + "content": "learning. Cognitive development, 3(3):299–321, 1988.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 213, + 502, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 504, + 226 + ], + "score": 1.0, + "content": "Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 225, + 309, + 237 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 309, + 237 + ], + "score": 1.0, + "content": "understanding. In CVPR, pp. 6874–6883, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 496, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 242, + 498, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 498, + 258 + ], + "score": 1.0, + "content": "Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 277 + ], + "score": 1.0, + "content": "Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 114, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "neural networks. In Workshop on Challenges in Representation Learning, ICML, volume 3, pp.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 284, + 152, + 297 + ], + "spans": [ + { + "bbox": [ + 114, + 284, + 152, + 297 + ], + "score": 1.0, + "content": "2, 2013.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 503, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 318 + ], + "score": 1.0, + "content": "Yin Li, Manohar Paluri, James M Rehg, and Piotr Dollar. Unsupervised learning of edges. In ´ CVPR,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 314, + 143, + 328 + ], + "spans": [ + { + "bbox": [ + 114, + 314, + 143, + 328 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 105, + 334, + 503, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "Ellen M Markman. Categorization and naming in children: Problems of induction. mit Press, 1989.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 352, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representa-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 363, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 505, + 379 + ], + "score": 1.0, + "content": "tions of words and phrases and their compositionality. In C. J. C. Burges, L. Bottou, M. Welling,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "Z. Ghahramani, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 386, + 319, + 399 + ], + "spans": [ + { + "bbox": [ + 115, + 386, + 319, + 399 + ], + "score": 1.0, + "content": "26, pp. 3111–3119. Curran Associates, Inc., 2013.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 405, + 504, + 428 + ], + "lines": [ + { + "bbox": [ + 104, + 403, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 506, + 420 + ], + "score": 1.0, + "content": "Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 417, + 319, + 428 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 319, + 428 + ], + "score": 1.0, + "content": "using temporal order verification. In ECCV, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "score": 1.0, + "content": "Takeru Miyato, Shin-ichi Maeda, Shin Ishii, and Masanori Koyama. Virtual adversarial training: a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "regularization method for supervised and semi-supervised learning. IEEE transactions on pattern", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 457, + 281, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 457, + 281, + 470 + ], + "score": 1.0, + "content": "analysis and machine intelligence, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 503, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Andriy Mnih and Koray Kavukcuoglu. Learning word embeddings efficiently with noise-contrastive", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 487, + 474, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 474, + 501 + ], + "score": 1.0, + "content": "estimation. In Advances in neural information processing systems, pp. 2265–2273, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 503, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "score": 1.0, + "content": "Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 518, + 447, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 518, + 447, + 531 + ], + "score": 1.0, + "content": "puzzles. In European Conference on Computer Vision, pp. 69–84. Springer, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 109, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "Stephanie E Palmer, Olivier Marre, Michael J Berry, and William Bialek. Predictive information", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "in a sensory population. Proceedings of the National Academy of Sciences, 112(22):6908–6913,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 558, + 143, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 143, + 571 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 105, + 577, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "score": 1.0, + "content": "Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning fea-´", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 589, + 410, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 589, + 410, + 601 + ], + "score": 1.0, + "content": "tures by watching objects move. arXiv preprint arXiv:1612.06370, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 105, + 608, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 104, + 607, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 104, + 607, + 506, + 622 + ], + "score": 1.0, + "content": "Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 619, + 248, + 631 + ], + "spans": [ + { + "bbox": [ + 117, + 619, + 248, + 631 + ], + "score": 1.0, + "content": "700 robot hours. In ICRA, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 105, + 638, + 504, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "score": 1.0, + "content": "Lerrel Pinto, James Davidson, and Abhinav Gupta. Supervision via competition: Robot adversaries", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 649, + 354, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 354, + 662 + ], + "score": 1.0, + "content": "for learning tasks. arXiv preprint arXiv:1610.01685, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "score": 1.0, + "content": "Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpreta-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 680, + 466, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 466, + 691 + ], + "score": 1.0, + "content": "tion of some extra-classical receptive-field effects. Nature neuroscience, 2(1):79, 1999.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "detection with region proposal networks. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 720, + 187, + 732 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 187, + 732 + ], + "score": 1.0, + "content": "pp. 91–99, 2015.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Durk P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-supervised", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "learning with deep generative models. In Advances in neural information processing systems, pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 191, + 115 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 191, + 115 + ], + "score": 1.0, + "content": "3581–3589, 2014.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Alexander Kolesnikov, Xiaohua Zhai, and Lucas Beyer. Revisiting self-supervised visual representa-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 505, + 147 + ], + "score": 1.0, + "content": "tion learning. CoRR, abs/1901.09005, 2019. URL http://arxiv.org/abs/1901.09005.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 123, + 505, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 151, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 151, + 506, + 168 + ], + "score": 1.0, + "content": "Brenden M Lake, Ruslan Salakhutdinov, and Joshua B Tenenbaum. Human-level concept learning", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 165, + 437, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 437, + 177 + ], + "score": 1.0, + "content": "through probabilistic program induction. Science, 350(6266):1332–1338, 2015.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 151, + 506, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 183, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "score": 1.0, + "content": "Barbara Landau, Linda B Smith, and Susan S Jones. The importance of shape in early lexical", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 195, + 335, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 195, + 335, + 207 + ], + "score": 1.0, + "content": "learning. Cognitive development, 3(3):299–321, 1988.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 182, + 505, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 213, + 502, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 504, + 226 + ], + "score": 1.0, + "content": "Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 225, + 309, + 237 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 309, + 237 + ], + "score": 1.0, + "content": "understanding. In CVPR, pp. 6874–6883, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 213, + 504, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 496, + 256 + ], + "lines": [ + { + "bbox": [ + 106, + 242, + 498, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 498, + 258 + ], + "score": 1.0, + "content": "Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 242, + 498, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 263, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 505, + 277 + ], + "score": 1.0, + "content": "Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 114, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "neural networks. In Workshop on Challenges in Representation Learning, ICML, volume 3, pp.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 284, + 152, + 297 + ], + "spans": [ + { + "bbox": [ + 114, + 284, + 152, + 297 + ], + "score": 1.0, + "content": "2, 2013.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 262, + 505, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 503, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 318 + ], + "score": 1.0, + "content": "Yin Li, Manohar Paluri, James M Rehg, and Piotr Dollar. Unsupervised learning of edges. In ´ CVPR,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 314, + 143, + 328 + ], + "spans": [ + { + "bbox": [ + 114, + 314, + 143, + 328 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 302, + 505, + 328 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 334, + 503, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "Ellen M Markman. Categorization and naming in children: Problems of induction. mit Press, 1989.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 106, + 334, + 505, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 352, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representa-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 363, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 505, + 379 + ], + "score": 1.0, + "content": "tions of words and phrases and their compositionality. In C. J. C. Burges, L. Bottou, M. Welling,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "Z. Ghahramani, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 386, + 319, + 399 + ], + "spans": [ + { + "bbox": [ + 115, + 386, + 319, + 399 + ], + "score": 1.0, + "content": "26, pp. 3111–3119. Curran Associates, Inc., 2013.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 353, + 505, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 405, + 504, + 428 + ], + "lines": [ + { + "bbox": [ + 104, + 403, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 506, + 420 + ], + "score": 1.0, + "content": "Ishan Misra, C Lawrence Zitnick, and Martial Hebert. Shuffle and learn: unsupervised learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 417, + 319, + 428 + ], + "spans": [ + { + "bbox": [ + 116, + 417, + 319, + 428 + ], + "score": 1.0, + "content": "using temporal order verification. In ECCV, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 403, + 506, + 428 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 506, + 449 + ], + "score": 1.0, + "content": "Takeru Miyato, Shin-ichi Maeda, Shin Ishii, and Masanori Koyama. Virtual adversarial training: a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 446, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 446, + 505, + 460 + ], + "score": 1.0, + "content": "regularization method for supervised and semi-supervised learning. IEEE transactions on pattern", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 457, + 281, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 457, + 281, + 470 + ], + "score": 1.0, + "content": "analysis and machine intelligence, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 434, + 506, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 503, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 489 + ], + "score": 1.0, + "content": "Andriy Mnih and Koray Kavukcuoglu. Learning word embeddings efficiently with noise-contrastive", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 487, + 474, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 474, + 501 + ], + "score": 1.0, + "content": "estimation. In Advances in neural information processing systems, pp. 2265–2273, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 476, + 505, + 501 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 506, + 503, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 520 + ], + "score": 1.0, + "content": "Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 518, + 447, + 531 + ], + "spans": [ + { + "bbox": [ + 115, + 518, + 447, + 531 + ], + "score": 1.0, + "content": "puzzles. In European Conference on Computer Vision, pp. 69–84. Springer, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 505, + 505, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 536, + 504, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "Stephanie E Palmer, Olivier Marre, Michael J Berry, and William Bialek. Predictive information", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 505, + 560 + ], + "score": 1.0, + "content": "in a sensory population. Proceedings of the National Academy of Sciences, 112(22):6908–6913,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 558, + 143, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 558, + 143, + 571 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 536, + 505, + 571 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 577, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 504, + 591 + ], + "score": 1.0, + "content": "Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning fea-´", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 589, + 410, + 601 + ], + "spans": [ + { + "bbox": [ + 115, + 589, + 410, + 601 + ], + "score": 1.0, + "content": "tures by watching objects move. arXiv preprint arXiv:1612.06370, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 578, + 504, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 608, + 505, + 631 + ], + "lines": [ + { + "bbox": [ + 104, + 607, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 104, + 607, + 506, + 622 + ], + "score": 1.0, + "content": "Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 117, + 619, + 248, + 631 + ], + "spans": [ + { + "bbox": [ + 117, + 619, + 248, + 631 + ], + "score": 1.0, + "content": "700 robot hours. In ICRA, 2016.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 607, + 506, + 631 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 638, + 504, + 661 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 505, + 652 + ], + "score": 1.0, + "content": "Lerrel Pinto, James Davidson, and Abhinav Gupta. Supervision via competition: Robot adversaries", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 649, + 354, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 649, + 354, + 662 + ], + "score": 1.0, + "content": "for learning tasks. arXiv preprint arXiv:1610.01685, 2016.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 636, + 505, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 504, + 681 + ], + "score": 1.0, + "content": "Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpreta-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 680, + 466, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 466, + 691 + ], + "score": 1.0, + "content": "tion of some extra-classical receptive-field effects. Nature neuroscience, 2(1):79, 1999.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 668, + 504, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "detection with region proposal networks. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 114, + 720, + 187, + 732 + ], + "spans": [ + { + "bbox": [ + 114, + 720, + 187, + 732 + ], + "score": 1.0, + "content": "pp. 91–99, 2015.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 469, + 118 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 469, + 118 + ], + "score": 1.0, + "content": "recognition challenge. International journal of computer vision, 115(3):211–252, 2015.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 506, + 158 + ], + "score": 1.0, + "content": "2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1134–1141. IEEE,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 143, + 169 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 143, + 169 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 108, + 174, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 187, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 116, + 187, + 504, + 198 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 197, + 291, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 291, + 209 + ], + "score": 1.0, + "content": "learning research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 229 + ], + "score": 1.0, + "content": "Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 227, + 504, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 227, + 504, + 239 + ], + "score": 1.0, + "content": "Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 237, + 442, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 442, + 250 + ], + "score": 1.0, + "content": "CoRR, abs/1409.4842, 2014. URL http://arxiv.org/abs/1409.4842.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 256, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. arXiv preprint", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 278, + 220, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 278, + 220, + 290 + ], + "score": 1.0, + "content": "arXiv:1503.01817, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "score": 1.0, + "content": "Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 309, + 219, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 309, + 219, + 320 + ], + "score": 1.0, + "content": "arXiv:1906.05849, 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 504, + 350 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 504, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 504, + 340 + ], + "score": 1.0, + "content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 338, + 329, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 329, + 351 + ], + "score": 1.0, + "content": "tive coding. arXiv preprint arXiv:1807.03748, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 368, + 180, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 180, + 380 + ], + "score": 1.0, + "content": "In ICCV, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 504, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "score": 1.0, + "content": "Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 398, + 506, + 411 + ], + "score": 1.0, + "content": "parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 410, + 308, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 308, + 421 + ], + "score": 1.0, + "content": "and Pattern Recognition, pp. 3733–3742, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 107, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 107, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised Data", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 439, + 376, + 451 + ], + "spans": [ + { + "bbox": [ + 116, + 439, + 376, + 451 + ], + "score": 1.0, + "content": "Augmentation. arXiv e-prints, art. arXiv:1904.12848, Apr 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 104, + 457, + 503, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 504, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 504, + 471 + ], + "score": 1.0, + "content": "Amir R Zamir, Tilman Wekel, Pulkit Agrawal, Colin Wei, Jitendra Malik, and Silvio Savarese.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 469, + 430, + 481 + ], + "spans": [ + { + "bbox": [ + 116, + 469, + 430, + 481 + ], + "score": 1.0, + "content": "Generic 3D representation via pose estimation and matching. In ECCV, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 503, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 391, + 501 + ], + "score": 1.0, + "content": "Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer.", + "type": "text" + }, + { + "bbox": [ + 391, + 487, + 411, + 498 + ], + "score": 0.87, + "content": "S ^ { 4 } L", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 486, + 505, + 501 + ], + "score": 1.0, + "content": ": Self-supervised semi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 499, + 362, + 511 + ], + "spans": [ + { + "bbox": [ + 115, + 499, + 362, + 511 + ], + "score": 1.0, + "content": "supervised learning. arXiv preprint arXiv:1905.03670, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 503, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 532 + ], + "score": 1.0, + "content": "Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European confer-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 529, + 337, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 337, + 541 + ], + "score": 1.0, + "content": "ence on computer vision, pp. 649–666. Springer, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 105, + 547, + 503, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propa-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 559, + 457, + 572 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 457, + 572 + ], + "score": 1.0, + "content": "gation. In Technical Report CMU-CALD-02-107, Carnegie Mellon University, 2002.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 592 + ], + "score": 1.0, + "content": "Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 588, + 370, + 600 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 370, + 600 + ], + "score": 1.0, + "content": "of visual embeddings. arXiv preprint arXiv:1903.12355, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 109, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 469, + 118 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 469, + 118 + ], + "score": 1.0, + "content": "recognition challenge. International journal of computer vision, 115(3):211–252, 2015.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 81, + 505, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 137 + ], + "score": 1.0, + "content": "Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 506, + 158 + ], + "score": 1.0, + "content": "2018 IEEE International Conference on Robotics and Automation (ICRA), pp. 1134–1141. IEEE,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 143, + 169 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 143, + 169 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 122, + 506, + 169 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 174, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 505, + 188 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 187, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 116, + 187, + 504, + 198 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 197, + 291, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 197, + 291, + 209 + ], + "score": 1.0, + "content": "learning research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 106, + 174, + 505, + 209 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 215, + 505, + 250 + ], + "lines": [ + { + "bbox": [ + 106, + 214, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 229 + ], + "score": 1.0, + "content": "Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 227, + 504, + 239 + ], + "spans": [ + { + "bbox": [ + 116, + 227, + 504, + 239 + ], + "score": 1.0, + "content": "Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 237, + 442, + 250 + ], + "spans": [ + { + "bbox": [ + 116, + 237, + 442, + 250 + ], + "score": 1.0, + "content": "CoRR, abs/1409.4842, 2014. URL http://arxiv.org/abs/1409.4842.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 106, + 214, + 505, + 250 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 256, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 269 + ], + "score": 1.0, + "content": "Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 267, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 267, + 505, + 280 + ], + "score": 1.0, + "content": "Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. arXiv preprint", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 278, + 220, + 290 + ], + "spans": [ + { + "bbox": [ + 115, + 278, + 220, + 290 + ], + "score": 1.0, + "content": "arXiv:1503.01817, 2015.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 257, + 505, + 290 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 297, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 311 + ], + "score": 1.0, + "content": "Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 309, + 219, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 309, + 219, + 320 + ], + "score": 1.0, + "content": "arXiv:1906.05849, 2019.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 296, + 505, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 327, + 504, + 350 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 504, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 504, + 340 + ], + "score": 1.0, + "content": "Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predic-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 338, + 329, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 329, + 351 + ], + "score": 1.0, + "content": "tive coding. arXiv preprint arXiv:1807.03748, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 327, + 504, + 351 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 368, + 180, + 380 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 180, + 380 + ], + "score": 1.0, + "content": "In ICCV, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 356, + 505, + 380 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 387, + 504, + 421 + ], + "lines": [ + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 505, + 401 + ], + "score": 1.0, + "content": "Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 398, + 506, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 398, + 506, + 411 + ], + "score": 1.0, + "content": "parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 410, + 308, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 308, + 421 + ], + "score": 1.0, + "content": "and Pattern Recognition, pp. 3733–3742, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 386, + 506, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 428, + 504, + 451 + ], + "lines": [ + { + "bbox": [ + 107, + 428, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 107, + 428, + 505, + 441 + ], + "score": 1.0, + "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised Data", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 439, + 376, + 451 + ], + "spans": [ + { + "bbox": [ + 116, + 439, + 376, + 451 + ], + "score": 1.0, + "content": "Augmentation. arXiv e-prints, art. arXiv:1904.12848, Apr 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 107, + 428, + 505, + 451 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 457, + 503, + 481 + ], + "lines": [ + { + "bbox": [ + 105, + 457, + 504, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 504, + 471 + ], + "score": 1.0, + "content": "Amir R Zamir, Tilman Wekel, Pulkit Agrawal, Colin Wei, Jitendra Malik, and Silvio Savarese.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 469, + 430, + 481 + ], + "spans": [ + { + "bbox": [ + 116, + 469, + 430, + 481 + ], + "score": 1.0, + "content": "Generic 3D representation via pose estimation and matching. In ECCV, 2016.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 457, + 504, + 481 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 503, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 391, + 501 + ], + "score": 1.0, + "content": "Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer.", + "type": "text" + }, + { + "bbox": [ + 391, + 487, + 411, + 498 + ], + "score": 0.87, + "content": "S ^ { 4 } L", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 486, + 505, + 501 + ], + "score": 1.0, + "content": ": Self-supervised semi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 499, + 362, + 511 + ], + "spans": [ + { + "bbox": [ + 115, + 499, + 362, + 511 + ], + "score": 1.0, + "content": "supervised learning. arXiv preprint arXiv:1905.03670, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 486, + 505, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 503, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 532 + ], + "score": 1.0, + "content": "Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In European confer-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 529, + 337, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 337, + 541 + ], + "score": 1.0, + "content": "ence on computer vision, pp. 649–666. Springer, 2016.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 516, + 505, + 541 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 547, + 503, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 504, + 560 + ], + "score": 1.0, + "content": "Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propa-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 559, + 457, + 572 + ], + "spans": [ + { + "bbox": [ + 115, + 559, + 457, + 572 + ], + "score": 1.0, + "content": "gation. In Technical Report CMU-CALD-02-107, Carnegie Mellon University, 2002.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 547, + 504, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 504, + 601 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 505, + 592 + ], + "score": 1.0, + "content": "Chengxu Zhuang, Alex Lin Zhai, and Daniel Yamins. Local aggregation for unsupervised learning", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 588, + 370, + 600 + ], + "spans": [ + { + "bbox": [ + 115, + 588, + 370, + 600 + ], + "score": 1.0, + "content": "of visual embeddings. arXiv preprint arXiv:1903.12355, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 106, + 575, + 505, + 600 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 183, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 231, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "score": 1.0, + "content": "A.1 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 113, + 153, + 499, + 212 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 174, + 138, + 435, + 150 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 136, + 437, + 153 + ], + "spans": [ + { + "bbox": [ + 174, + 136, + 437, + 153 + ], + "score": 1.0, + "content": "Table 4: Data efficient classification results with Top-1 accuracy.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 113, + 153, + 499, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 153, + 499, + 212 + ], + "spans": [ + { + "bbox": [ + 113, + 153, + 499, + 212 + ], + "score": 0.947, + "html": "
Labeled data Method1%2%5%10% Top-1 accuracy20%50%100%
Supervised trained on pixels23.134.850.260.067.674.478.0
ReseNet trained on CPC v2 (fine-tuned)46.354.764.169.773.077.380.6
", + "type": "table", + "image_path": "cd16abc0799fc70d1a375836e8cc04ff4651be0d7ba68507f20886a5866fd0b3.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 113, + 153, + 499, + 172.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 113, + 172.66666666666666, + 499, + 192.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 192.33333333333331, + 499, + 211.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 113, + 255, + 499, + 315 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 144, + 241, + 466, + 254 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 142, + 239, + 467, + 255 + ], + "spans": [ + { + "bbox": [ + 142, + 239, + 467, + 255 + ], + "score": 1.0, + "content": "Table 5: Data efficient classification results with Top-5 accuracy (data in Fig. 1).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 113, + 255, + 499, + 315 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 255, + 499, + 315 + ], + "spans": [ + { + "bbox": [ + 113, + 255, + 499, + 315 + ], + "score": 0.956, + "html": "
Labeled data Method1%2%5%10% Top-5 accuracy20%50%100%
Supervised trained on pixels44.159.975.282.187.991.893.8
ReseNet trained on CPC v2 (fine-tuned)72.979.686.189.591.493.695.2
", + "type": "table", + "image_path": "ef57c53b71083a6069724c7b7ef7e00273991c3de17438cac7824c7e459eff57.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 113, + 255, + 499, + 275.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 113, + 275.0, + 499, + 295.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 113, + 295.0, + 499, + 315.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "title", + "bbox": [ + 108, + 340, + 256, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 257, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 257, + 353 + ], + "score": 1.0, + "content": "A.2 INFONCE IMPLEMENTATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 505, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "score": 1.0, + "content": "For completeness, we provide pseudo-code for the main calculations involved in the InfoNCE ob-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 372, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 506, + 384 + ], + "score": 1.0, + "content": "jective, loosely modeled after Tensorflow operations. We suppose we have just calculated a set of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 136, + 396 + ], + "score": 1.0, + "content": "latents", + "type": "text" + }, + { + "bbox": [ + 136, + 383, + 200, + 396 + ], + "score": 0.94, + "content": "z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 383, + 217, + 396 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 218, + 383, + 287, + 395 + ], + "score": 0.92, + "content": "i , j \\in \\{ 1 , . . . , 7 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 383, + 505, + 396 + ], + "score": 1.0, + "content": ", each one being e.g. a 4096-dimensional vector. As-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 231, + 406 + ], + "score": 1.0, + "content": "suming we do so for a batch of", + "type": "text" + }, + { + "bbox": [ + 231, + 395, + 240, + 404 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 394, + 271, + 406 + ], + "score": 1.0, + "content": "images", + "type": "text" + }, + { + "bbox": [ + 272, + 395, + 289, + 406 + ], + "score": 0.9, + "content": "\\{ { \\pmb x } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 394, + 430, + 406 + ], + "score": 1.0, + "content": ", the set of latents is a tensor of size", + "type": "text" + }, + { + "bbox": [ + 430, + 394, + 501, + 405 + ], + "score": 0.91, + "content": "B \\times 7 \\times 7 \\times 4 0 9 6", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 394, + 505, + 406 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 105, + 425, + 492, + 715 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 183, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 231, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 231, + 119 + ], + "score": 1.0, + "content": "A.1 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 113, + 153, + 499, + 212 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 174, + 138, + 435, + 150 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 136, + 437, + 153 + ], + "spans": [ + { + "bbox": [ + 174, + 136, + 437, + 153 + ], + "score": 1.0, + "content": "Table 4: Data efficient classification results with Top-1 accuracy.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 113, + 153, + 499, + 212 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 153, + 499, + 212 + ], + "spans": [ + { + "bbox": [ + 113, + 153, + 499, + 212 + ], + "score": 0.947, + "html": "
Labeled data Method1%2%5%10% Top-1 accuracy20%50%100%
Supervised trained on pixels23.134.850.260.067.674.478.0
ReseNet trained on CPC v2 (fine-tuned)46.354.764.169.773.077.380.6
", + "type": "table", + "image_path": "cd16abc0799fc70d1a375836e8cc04ff4651be0d7ba68507f20886a5866fd0b3.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 113, + 153, + 499, + 172.66666666666666 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 113, + 172.66666666666666, + 499, + 192.33333333333331 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 113, + 192.33333333333331, + 499, + 211.99999999999997 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 113, + 255, + 499, + 315 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 144, + 241, + 466, + 254 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 142, + 239, + 467, + 255 + ], + "spans": [ + { + "bbox": [ + 142, + 239, + 467, + 255 + ], + "score": 1.0, + "content": "Table 5: Data efficient classification results with Top-5 accuracy (data in Fig. 1).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 113, + 255, + 499, + 315 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 113, + 255, + 499, + 315 + ], + "spans": [ + { + "bbox": [ + 113, + 255, + 499, + 315 + ], + "score": 0.956, + "html": "
Labeled data Method1%2%5%10% Top-5 accuracy20%50%100%
Supervised trained on pixels44.159.975.282.187.991.893.8
ReseNet trained on CPC v2 (fine-tuned)72.979.686.189.591.493.695.2
", + "type": "table", + "image_path": "ef57c53b71083a6069724c7b7ef7e00273991c3de17438cac7824c7e459eff57.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 113, + 255, + 499, + 275.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 113, + 275.0, + 499, + 295.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 113, + 295.0, + 499, + 315.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 7.0 + }, + { + "type": "title", + "bbox": [ + 108, + 340, + 256, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 340, + 257, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 257, + 353 + ], + "score": 1.0, + "content": "A.2 INFONCE IMPLEMENTATION", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 505, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 373 + ], + "score": 1.0, + "content": "For completeness, we provide pseudo-code for the main calculations involved in the InfoNCE ob-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 372, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 506, + 384 + ], + "score": 1.0, + "content": "jective, loosely modeled after Tensorflow operations. We suppose we have just calculated a set of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 136, + 396 + ], + "score": 1.0, + "content": "latents", + "type": "text" + }, + { + "bbox": [ + 136, + 383, + 200, + 396 + ], + "score": 0.94, + "content": "z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 383, + 217, + 396 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 218, + 383, + 287, + 395 + ], + "score": 0.92, + "content": "i , j \\in \\{ 1 , . . . , 7 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 383, + 505, + 396 + ], + "score": 1.0, + "content": ", each one being e.g. a 4096-dimensional vector. As-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 394, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 231, + 406 + ], + "score": 1.0, + "content": "suming we do so for a batch of", + "type": "text" + }, + { + "bbox": [ + 231, + 395, + 240, + 404 + ], + "score": 0.8, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 394, + 271, + 406 + ], + "score": 1.0, + "content": "images", + "type": "text" + }, + { + "bbox": [ + 272, + 395, + 289, + 406 + ], + "score": 0.9, + "content": "\\{ { \\pmb x } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 394, + 430, + 406 + ], + "score": 1.0, + "content": ", the set of latents is a tensor of size", + "type": "text" + }, + { + "bbox": [ + 430, + 394, + 501, + 405 + ], + "score": 0.91, + "content": "B \\times 7 \\times 7 \\times 4 0 9 6", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 394, + 505, + 406 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 361, + 506, + 406 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 95, + 406, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 94, + 240, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 240, + 105 + ], + "score": 1.0, + "content": "def pixelCNN(latents):", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 105, + 269, + 117 + ], + "spans": [ + { + "bbox": [ + 129, + 105, + 269, + 117 + ], + "score": 1.0, + "content": "# latents: [B, H, W, D]", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 117, + 217, + 127 + ], + "spans": [ + { + "bbox": [ + 131, + 117, + 158, + 127 + ], + "score": 1.0, + "content": "cres", + "type": "text" + }, + { + "bbox": [ + 159, + 117, + 170, + 126 + ], + "score": 0.71, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 117, + 217, + 127 + ], + "score": 1.0, + "content": "latents", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 127, + 281, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 127, + 182, + 138 + ], + "score": 1.0, + "content": "cres_dim", + "type": "text" + }, + { + "bbox": [ + 182, + 128, + 194, + 137 + ], + "score": 0.74, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 127, + 281, + 138 + ], + "score": 1.0, + "content": "cres.shape[-1]", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 137, + 240, + 150 + ], + "spans": [ + { + "bbox": [ + 131, + 138, + 152, + 149 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 164, + 137, + 240, + 150 + ], + "score": 1.0, + "content": "in range(5):", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 149, + 329, + 160 + ], + "spans": [ + { + "bbox": [ + 141, + 149, + 152, + 160 + ], + "score": 1.0, + "content": "c", + "type": "text" + }, + { + "bbox": [ + 152, + 150, + 164, + 159 + ], + "score": 0.63, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 149, + 298, + 160 + ], + "score": 1.0, + "content": "Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 149, + 323, + 159 + ], + "score": 0.4, + "content": "= 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 149, + 329, + 160 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 207, + 159, + 364, + 172 + ], + "spans": [ + { + "bbox": [ + 207, + 159, + 280, + 172 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 162, + 288, + 169 + ], + "score": 0.66, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 159, + 364, + 172 + ], + "score": 1.0, + "content": "(1, 1))(cres)", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 171, + 209, + 182 + ], + "spans": [ + { + "bbox": [ + 142, + 171, + 209, + 182 + ], + "score": 1.0, + "content": "c = ReLU(c)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 182, + 329, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 182, + 297, + 193 + ], + "score": 1.0, + "content": "c = Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 182, + 323, + 192 + ], + "score": 0.42, + "content": "= 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 182, + 329, + 193 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 207, + 193, + 347, + 204 + ], + "spans": [ + { + "bbox": [ + 207, + 193, + 280, + 204 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 194, + 288, + 202 + ], + "score": 0.74, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 193, + 347, + 204 + ], + "score": 1.0, + "content": "(1, 3))(c)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 203, + 406, + 216 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 406, + 216 + ], + "score": 1.0, + "content": "c = Pad(c, [[0, 0], [1, 0], [0, 0], [0, 0]])", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 214, + 329, + 227 + ], + "spans": [ + { + "bbox": [ + 142, + 214, + 297, + 227 + ], + "score": 1.0, + "content": "c = Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 215, + 323, + 225 + ], + "score": 0.29, + "content": "= 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 214, + 329, + 227 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 207, + 225, + 329, + 237 + ], + "spans": [ + { + "bbox": [ + 207, + 225, + 280, + 237 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 227, + 288, + 235 + ], + "score": 0.77, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 225, + 329, + 237 + ], + "score": 1.0, + "content": "(2, 1),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 208, + 236, + 304, + 248 + ], + "spans": [ + { + "bbox": [ + 208, + 236, + 232, + 248 + ], + "score": 1.0, + "content": "type", + "type": "text" + }, + { + "bbox": [ + 232, + 238, + 244, + 246 + ], + "score": 0.65, + "content": "= ^ { \\parallel }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 236, + 304, + 248 + ], + "score": 1.0, + "content": "VALID')(c)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 248, + 208, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 153, + 258 + ], + "score": 1.0, + "content": "c", + "type": "text" + }, + { + "bbox": [ + 153, + 249, + 164, + 257 + ], + "score": 0.31, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 248, + 208, + 258 + ], + "score": 1.0, + "content": "ReLU(c)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 258, + 358, + 271 + ], + "spans": [ + { + "bbox": [ + 142, + 258, + 153, + 271 + ], + "score": 1.0, + "content": "c", + "type": "text" + }, + { + "bbox": [ + 153, + 259, + 164, + 268 + ], + "score": 0.28, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 258, + 298, + 271 + ], + "score": 1.0, + "content": "Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 261, + 305, + 268 + ], + "score": 0.69, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 258, + 358, + 271 + ], + "score": 1.0, + "content": "cres_dim,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 207, + 269, + 346, + 281 + ], + "spans": [ + { + "bbox": [ + 207, + 269, + 280, + 281 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 271, + 288, + 279 + ], + "score": 0.7, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 269, + 346, + 281 + ], + "score": 1.0, + "content": "(1, 1))(c)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 281, + 235, + 291 + ], + "spans": [ + { + "bbox": [ + 142, + 281, + 171, + 291 + ], + "score": 1.0, + "content": "cres", + "type": "text" + }, + { + "bbox": [ + 171, + 282, + 182, + 290 + ], + "score": 0.38, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 281, + 235, + 291 + ], + "score": 1.0, + "content": "cres + c", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 291, + 233, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 159, + 303 + ], + "score": 1.0, + "content": "cres", + "type": "text" + }, + { + "bbox": [ + 159, + 293, + 169, + 301 + ], + "score": 0.59, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 291, + 233, + 303 + ], + "score": 1.0, + "content": "ReLU(cres)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 302, + 199, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 199, + 313 + ], + "score": 1.0, + "content": "return cres", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 334, + 240, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 333, + 241, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 241, + 347 + ], + "score": 1.0, + "content": "A.3 LINEAR CLASSIFICATION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 133, + 351, + 506, + 491 + ], + "lines": [ + { + "bbox": [ + 135, + 351, + 504, + 363 + ], + "spans": [ + { + "bbox": [ + 135, + 351, + 293, + 363 + ], + "score": 1.0, + "content": "• Model architecture: Having extracted", + "type": "text" + }, + { + "bbox": [ + 293, + 351, + 320, + 362 + ], + "score": 0.64, + "content": "8 0 \\mathrm { x } 8 0", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 351, + 412, + 363 + ], + "score": 1.0, + "content": "patches with a stride of", + "type": "text" + }, + { + "bbox": [ + 413, + 351, + 439, + 361 + ], + "score": 0.33, + "content": "3 2 \\mathrm { x } 3 2 ", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 351, + 467, + 363 + ], + "score": 1.0, + "content": "from a", + "type": "text" + }, + { + "bbox": [ + 468, + 351, + 504, + 362 + ], + "score": 0.68, + "content": "2 4 0 \\times 2 4 0", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "shaped input image, we end up with a grid of 6x6 features (each of which is obtained", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 373, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 373, + 506, + 386 + ], + "score": 1.0, + "content": "from our ResNet-161 architecture). This gives us a [6,6,4096] tensor for the image. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "then use a Batch-Normalization layer to normalize the features (without scale parameter)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "followed by a 1x1 convolution mapping each feature in the grid to the 1000 logits for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "ImageNet classification. We then spatially-mean-pool these logits to end up with the final", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 417, + 321, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 417, + 321, + 429 + ], + "score": 1.0, + "content": "log probabilities for the linear classification.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 134, + 430, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 134, + 430, + 420, + 446 + ], + "score": 1.0, + "content": "• We use the Inception preprocessing (Szegedy et al., 2014) to extract", + "type": "text" + }, + { + "bbox": [ + 421, + 432, + 457, + 442 + ], + "score": 0.62, + "content": "2 4 0 \\times 2 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 430, + 506, + 446 + ], + "score": 1.0, + "content": "crops from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "the raw image. The image is divided into subcrops as per CPC data-preprocessing used for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 453, + 216, + 468 + ], + "spans": [ + { + "bbox": [ + 141, + 453, + 216, + 468 + ], + "score": 1.0, + "content": "CPC pre-training.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 133, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "• Optimization details: We use Adam Optimizer with a learning rate of 5e-4. We train the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 479, + 486, + 492 + ], + "spans": [ + { + "bbox": [ + 141, + 479, + 486, + 492 + ], + "score": 1.0, + "content": "model on a batch size of 512 images with 32 images per core spread over 16 workers.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 346, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 347, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 347, + 516 + ], + "score": 1.0, + "content": "A.4 EFFICIENT CLASSIFICATION: PURELY SUPERVISED", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 105, + 524, + 459, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 523, + 461, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 461, + 538 + ], + "score": 1.0, + "content": "In order to find the best model within this class, we vary the following hyperparameters:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 133, + 546, + 504, + 643 + ], + "lines": [ + { + "bbox": [ + 133, + 546, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 506, + 558 + ], + "score": 1.0, + "content": "• Model architecture: We investigate using ResNet-50, ResNet-101, and ResNet-152 model", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 142, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 142, + 557, + 288, + 569 + ], + "score": 1.0, + "content": "architectures, all of them using the", + "type": "text" + }, + { + "bbox": [ + 288, + 558, + 306, + 568 + ], + "score": 0.42, + "content": "\\mathbf { \\hat { v } } _ { \\mathbf { \\Delta } } 2 \\mathbf { \\hat { \\mu } }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "variant (He et al., 2016b), and find larger archi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 141, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "tecture to perform better, even when given smaller amounts of data. We insert a DropOut", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 142, + 579, + 428, + 591 + ], + "spans": [ + { + "bbox": [ + 142, + 579, + 428, + 591 + ], + "score": 1.0, + "content": "layer before the final linear classification layer (Srivastava et al., 2014).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 133, + 593, + 502, + 607 + ], + "spans": [ + { + "bbox": [ + 133, + 593, + 502, + 607 + ], + "score": 1.0, + "content": "• Data pre-processing: We use the Inception pre-processing pipeline (Szegedy et al., 2014).", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 133, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 608, + 346, + 622 + ], + "score": 1.0, + "content": "• Optimization details: We vary the learning rate in", + "type": "text" + }, + { + "bbox": [ + 347, + 608, + 411, + 621 + ], + "score": 0.7, + "content": "\\{ 0 . 0 5 , 0 . 1 , 0 . 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 608, + 505, + 622 + ], + "score": 1.0, + "content": ", the weight decay log-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 141, + 618, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 141, + 618, + 216, + 633 + ], + "score": 1.0, + "content": "arithmically from", + "type": "text" + }, + { + "bbox": [ + 216, + 619, + 237, + 630 + ], + "score": 0.9, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 618, + 250, + 633 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 250, + 619, + 272, + 630 + ], + "score": 0.89, + "content": "1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 618, + 506, + 633 + ], + "score": 1.0, + "content": ", the DropOut linearly from 0 to 1, and the batch size per", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 141, + 630, + 224, + 644 + ], + "spans": [ + { + "bbox": [ + 141, + 630, + 183, + 644 + ], + "score": 1.0, + "content": "worker in", + "type": "text" + }, + { + "bbox": [ + 183, + 631, + 219, + 643 + ], + "score": 0.77, + "content": "\\{ 1 6 , 3 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 630, + 224, + 644 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 109, + 651, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 350, + 666 + ], + "score": 1.0, + "content": "We chose the best performing model for each training subset", + "type": "text" + }, + { + "bbox": [ + 351, + 652, + 362, + 663 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 650, + 505, + 666 + ], + "score": 1.0, + "content": "of labeled ImageNet (using a sepa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "rate validation set), and report its accuracy on the test set (i.e. the publicly available ILSVRC-2012", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 674, + 167, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 167, + 685 + ], + "score": 1.0, + "content": "validation set).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 95, + 406, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 94, + 240, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 240, + 105 + ], + "score": 1.0, + "content": "def pixelCNN(latents):", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 129, + 105, + 269, + 117 + ], + "spans": [ + { + "bbox": [ + 129, + 105, + 269, + 117 + ], + "score": 1.0, + "content": "# latents: [B, H, W, D]", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 131, + 117, + 217, + 127 + ], + "spans": [ + { + "bbox": [ + 131, + 117, + 158, + 127 + ], + "score": 1.0, + "content": "cres", + "type": "text" + }, + { + "bbox": [ + 159, + 117, + 170, + 126 + ], + "score": 0.71, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 117, + 217, + 127 + ], + "score": 1.0, + "content": "latents", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 127, + 281, + 138 + ], + "spans": [ + { + "bbox": [ + 130, + 127, + 182, + 138 + ], + "score": 1.0, + "content": "cres_dim", + "type": "text" + }, + { + "bbox": [ + 182, + 128, + 194, + 137 + ], + "score": 0.74, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 127, + 281, + 138 + ], + "score": 1.0, + "content": "cres.shape[-1]", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 137, + 240, + 150 + ], + "spans": [ + { + "bbox": [ + 131, + 138, + 152, + 149 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 164, + 137, + 240, + 150 + ], + "score": 1.0, + "content": "in range(5):", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 149, + 329, + 160 + ], + "spans": [ + { + "bbox": [ + 141, + 149, + 152, + 160 + ], + "score": 1.0, + "content": "c", + "type": "text" + }, + { + "bbox": [ + 152, + 150, + 164, + 159 + ], + "score": 0.63, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 149, + 298, + 160 + ], + "score": 1.0, + "content": "Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 149, + 323, + 159 + ], + "score": 0.4, + "content": "= 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 149, + 329, + 160 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 207, + 159, + 364, + 172 + ], + "spans": [ + { + "bbox": [ + 207, + 159, + 280, + 172 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 162, + 288, + 169 + ], + "score": 0.66, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 159, + 364, + 172 + ], + "score": 1.0, + "content": "(1, 1))(cres)", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 142, + 171, + 209, + 182 + ], + "spans": [ + { + "bbox": [ + 142, + 171, + 209, + 182 + ], + "score": 1.0, + "content": "c = ReLU(c)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 182, + 329, + 193 + ], + "spans": [ + { + "bbox": [ + 141, + 182, + 297, + 193 + ], + "score": 1.0, + "content": "c = Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 182, + 323, + 192 + ], + "score": 0.42, + "content": "= 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 182, + 329, + 193 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 207, + 193, + 347, + 204 + ], + "spans": [ + { + "bbox": [ + 207, + 193, + 280, + 204 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 194, + 288, + 202 + ], + "score": 0.74, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 193, + 347, + 204 + ], + "score": 1.0, + "content": "(1, 3))(c)", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 203, + 406, + 216 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 406, + 216 + ], + "score": 1.0, + "content": "c = Pad(c, [[0, 0], [1, 0], [0, 0], [0, 0]])", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 214, + 329, + 227 + ], + "spans": [ + { + "bbox": [ + 142, + 214, + 297, + 227 + ], + "score": 1.0, + "content": "c = Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 215, + 323, + 225 + ], + "score": 0.29, + "content": "= 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 214, + 329, + 227 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 207, + 225, + 329, + 237 + ], + "spans": [ + { + "bbox": [ + 207, + 225, + 280, + 237 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 227, + 288, + 235 + ], + "score": 0.77, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 225, + 329, + 237 + ], + "score": 1.0, + "content": "(2, 1),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 208, + 236, + 304, + 248 + ], + "spans": [ + { + "bbox": [ + 208, + 236, + 232, + 248 + ], + "score": 1.0, + "content": "type", + "type": "text" + }, + { + "bbox": [ + 232, + 238, + 244, + 246 + ], + "score": 0.65, + "content": "= ^ { \\parallel }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 236, + 304, + 248 + ], + "score": 1.0, + "content": "VALID')(c)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 248, + 208, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 153, + 258 + ], + "score": 1.0, + "content": "c", + "type": "text" + }, + { + "bbox": [ + 153, + 249, + 164, + 257 + ], + "score": 0.31, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 248, + 208, + 258 + ], + "score": 1.0, + "content": "ReLU(c)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 258, + 358, + 271 + ], + "spans": [ + { + "bbox": [ + 142, + 258, + 153, + 271 + ], + "score": 1.0, + "content": "c", + "type": "text" + }, + { + "bbox": [ + 153, + 259, + 164, + 268 + ], + "score": 0.28, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 258, + 298, + 271 + ], + "score": 1.0, + "content": "Conv2D(output_channels", + "type": "text" + }, + { + "bbox": [ + 298, + 261, + 305, + 268 + ], + "score": 0.69, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 258, + 358, + 271 + ], + "score": 1.0, + "content": "cres_dim,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 207, + 269, + 346, + 281 + ], + "spans": [ + { + "bbox": [ + 207, + 269, + 280, + 281 + ], + "score": 1.0, + "content": "kernel_shape", + "type": "text" + }, + { + "bbox": [ + 280, + 271, + 288, + 279 + ], + "score": 0.7, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 269, + 346, + 281 + ], + "score": 1.0, + "content": "(1, 1))(c)", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 281, + 235, + 291 + ], + "spans": [ + { + "bbox": [ + 142, + 281, + 171, + 291 + ], + "score": 1.0, + "content": "cres", + "type": "text" + }, + { + "bbox": [ + 171, + 282, + 182, + 290 + ], + "score": 0.38, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 281, + 235, + 291 + ], + "score": 1.0, + "content": "cres + c", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 130, + 291, + 233, + 303 + ], + "spans": [ + { + "bbox": [ + 130, + 291, + 159, + 303 + ], + "score": 1.0, + "content": "cres", + "type": "text" + }, + { + "bbox": [ + 159, + 293, + 169, + 301 + ], + "score": 0.59, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 291, + 233, + 303 + ], + "score": 1.0, + "content": "ReLU(cres)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 130, + 302, + 199, + 313 + ], + "spans": [ + { + "bbox": [ + 130, + 302, + 199, + 313 + ], + "score": 1.0, + "content": "return cres", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 94, + 406, + 313 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 334, + 240, + 345 + ], + "lines": [ + { + "bbox": [ + 106, + 333, + 241, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 333, + 241, + 347 + ], + "score": 1.0, + "content": "A.3 LINEAR CLASSIFICATION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "list", + "bbox": [ + 133, + 351, + 506, + 491 + ], + "lines": [ + { + "bbox": [ + 135, + 351, + 504, + 363 + ], + "spans": [ + { + "bbox": [ + 135, + 351, + 293, + 363 + ], + "score": 1.0, + "content": "• Model architecture: Having extracted", + "type": "text" + }, + { + "bbox": [ + 293, + 351, + 320, + 362 + ], + "score": 0.64, + "content": "8 0 \\mathrm { x } 8 0", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 351, + 412, + 363 + ], + "score": 1.0, + "content": "patches with a stride of", + "type": "text" + }, + { + "bbox": [ + 413, + 351, + 439, + 361 + ], + "score": 0.33, + "content": "3 2 \\mathrm { x } 3 2 ", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 351, + 467, + 363 + ], + "score": 1.0, + "content": "from a", + "type": "text" + }, + { + "bbox": [ + 468, + 351, + 504, + 362 + ], + "score": 0.68, + "content": "2 4 0 \\times 2 4 0", + "type": "inline_equation" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 361, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 506, + 375 + ], + "score": 1.0, + "content": "shaped input image, we end up with a grid of 6x6 features (each of which is obtained", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 373, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 373, + 506, + 386 + ], + "score": 1.0, + "content": "from our ResNet-161 architecture). This gives us a [6,6,4096] tensor for the image. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 141, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "then use a Batch-Normalization layer to normalize the features (without scale parameter)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "followed by a 1x1 convolution mapping each feature in the grid to the 1000 logits for", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "ImageNet classification. We then spatially-mean-pool these logits to end up with the final", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 417, + 321, + 429 + ], + "spans": [ + { + "bbox": [ + 141, + 417, + 321, + 429 + ], + "score": 1.0, + "content": "log probabilities for the linear classification.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 134, + 430, + 506, + 446 + ], + "spans": [ + { + "bbox": [ + 134, + 430, + 420, + 446 + ], + "score": 1.0, + "content": "• We use the Inception preprocessing (Szegedy et al., 2014) to extract", + "type": "text" + }, + { + "bbox": [ + 421, + 432, + 457, + 442 + ], + "score": 0.62, + "content": "2 4 0 \\times 2 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 430, + 506, + 446 + ], + "score": 1.0, + "content": "crops from", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "the raw image. The image is divided into subcrops as per CPC data-preprocessing used for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 453, + 216, + 468 + ], + "spans": [ + { + "bbox": [ + 141, + 453, + 216, + 468 + ], + "score": 1.0, + "content": "CPC pre-training.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 468, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 133, + 468, + 505, + 482 + ], + "score": 1.0, + "content": "• Optimization details: We use Adam Optimizer with a learning rate of 5e-4. We train the", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 479, + 486, + 492 + ], + "spans": [ + { + "bbox": [ + 141, + 479, + 486, + 492 + ], + "score": 1.0, + "content": "model on a batch size of 512 images with 32 images per core spread over 16 workers.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + } + ], + "index": 26.5, + "bbox_fs": [ + 133, + 351, + 506, + 492 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 346, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 347, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 347, + 516 + ], + "score": 1.0, + "content": "A.4 EFFICIENT CLASSIFICATION: PURELY SUPERVISED", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 504, + 347, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 524, + 459, + 536 + ], + "lines": [ + { + "bbox": [ + 105, + 523, + 461, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 461, + 538 + ], + "score": 1.0, + "content": "In order to find the best model within this class, we vary the following hyperparameters:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 523, + 461, + 538 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 546, + 504, + 643 + ], + "lines": [ + { + "bbox": [ + 133, + 546, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 133, + 546, + 506, + 558 + ], + "score": 1.0, + "content": "• Model architecture: We investigate using ResNet-50, ResNet-101, and ResNet-152 model", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 142, + 557, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 142, + 557, + 288, + 569 + ], + "score": 1.0, + "content": "architectures, all of them using the", + "type": "text" + }, + { + "bbox": [ + 288, + 558, + 306, + 568 + ], + "score": 0.42, + "content": "\\mathbf { \\hat { v } } _ { \\mathbf { \\Delta } } 2 \\mathbf { \\hat { \\mu } }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 557, + 505, + 569 + ], + "score": 1.0, + "content": "variant (He et al., 2016b), and find larger archi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 141, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 141, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "tecture to perform better, even when given smaller amounts of data. We insert a DropOut", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 142, + 579, + 428, + 591 + ], + "spans": [ + { + "bbox": [ + 142, + 579, + 428, + 591 + ], + "score": 1.0, + "content": "layer before the final linear classification layer (Srivastava et al., 2014).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 133, + 593, + 502, + 607 + ], + "spans": [ + { + "bbox": [ + 133, + 593, + 502, + 607 + ], + "score": 1.0, + "content": "• Data pre-processing: We use the Inception pre-processing pipeline (Szegedy et al., 2014).", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 133, + 608, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 133, + 608, + 346, + 622 + ], + "score": 1.0, + "content": "• Optimization details: We vary the learning rate in", + "type": "text" + }, + { + "bbox": [ + 347, + 608, + 411, + 621 + ], + "score": 0.7, + "content": "\\{ 0 . 0 5 , 0 . 1 , 0 . 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 608, + 505, + 622 + ], + "score": 1.0, + "content": ", the weight decay log-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 141, + 618, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 141, + 618, + 216, + 633 + ], + "score": 1.0, + "content": "arithmically from", + "type": "text" + }, + { + "bbox": [ + 216, + 619, + 237, + 630 + ], + "score": 0.9, + "content": "1 0 ^ { - 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 618, + 250, + 633 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 250, + 619, + 272, + 630 + ], + "score": 0.89, + "content": "1 0 ^ { - 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 618, + 506, + 633 + ], + "score": 1.0, + "content": ", the DropOut linearly from 0 to 1, and the batch size per", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 141, + 630, + 224, + 644 + ], + "spans": [ + { + "bbox": [ + 141, + 630, + 183, + 644 + ], + "score": 1.0, + "content": "worker in", + "type": "text" + }, + { + "bbox": [ + 183, + 631, + 219, + 643 + ], + "score": 0.77, + "content": "\\{ 1 6 , 3 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 630, + 224, + 644 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5, + "bbox_fs": [ + 133, + 546, + 506, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 651, + 505, + 685 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 350, + 666 + ], + "score": 1.0, + "content": "We chose the best performing model for each training subset", + "type": "text" + }, + { + "bbox": [ + 351, + 652, + 362, + 663 + ], + "score": 0.88, + "content": "\\mathbb { D } _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 650, + 505, + 666 + ], + "score": 1.0, + "content": "of labeled ImageNet (using a sepa-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 675 + ], + "score": 1.0, + "content": "rate validation set), and report its accuracy on the test set (i.e. the publicly available ILSVRC-2012", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 674, + 167, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 167, + 685 + ], + "score": 1.0, + "content": "validation set).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 650, + 505, + 685 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/rJerHlrYwH/rJerHlrYwH_model.json b/parse/train/rJerHlrYwH/rJerHlrYwH_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9198570e889173181978158a20fdf40a0e195ad6 --- /dev/null +++ b/parse/train/rJerHlrYwH/rJerHlrYwH_model.json @@ -0,0 +1,22521 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 591, + 1302, + 591, + 1302, + 986, + 398, + 986 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1123, + 824, + 1123, + 824, + 1941, + 298, + 1941 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 852, + 1129, + 1395, + 1129, + 1395, + 1614, + 852, + 1614 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 847, + 1628, + 1401, + 1628, + 1401, + 1902, + 847, + 1902 + ], + "score": 0.97 + }, + { + "category_id": 0, + "poly": [ + 299, + 218, + 1166, + 218, + 1166, + 324, + 299, + 324 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 300, + 1944, + 1401, + 1944, + 1401, + 2034, + 300, + 2034 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 313, + 377, + 680, + 377, + 680, + 437, + 313, + 437 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 303, + 1049, + 572, + 1049, + 572, + 1084, + 303, + 1084 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 773, + 518, + 926, + 518, + 926, + 553, + 773, + 553 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 857, + 2089, + 857, + 2112, + 841, + 2112 + ], + "score": 0.666 + }, + { + "category_id": 13, + "poly": [ + 699, + 864, + 754, + 864, + 754, + 893, + 699, + 893 + ], + "score": 0.87, + "latex": "65 \\%" + }, + { + "category_id": 13, + "poly": [ + 609, + 864, + 665, + 864, + 665, + 893, + 609, + 893 + ], + "score": 0.87, + "latex": "28 \\%" + }, + { + "category_id": 13, + "poly": [ + 660, + 895, + 712, + 895, + 712, + 924, + 660, + 924 + ], + "score": 0.87, + "latex": "14 \\%" + }, + { + "category_id": 13, + "poly": [ + 809, + 834, + 865, + 834, + 865, + 863, + 809, + 863 + ], + "score": 0.86, + "latex": "73 \\%" + }, + { + "category_id": 13, + "poly": [ + 619, + 804, + 660, + 804, + 660, + 833, + 619, + 833 + ], + "score": 0.86, + "latex": "1 \\%" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1175.0, + 919.0, + 1175.0, + 919.0, + 1210.0, + 869.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1175.0, + 1302.0, + 1175.0, + 1302.0, + 1184.0, + 1290.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1189.0, + 1161.0, + 1189.0, + 1161.0, + 1206.0, + 1143.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1204.0, + 1221.0, + 1204.0, + 1221.0, + 1216.0, + 1206.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1194.0, + 1373.0, + 1194.0, + 1373.0, + 1223.0, + 1281.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1212.0, + 890.0, + 1212.0, + 890.0, + 1480.0, + 848.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1212.0, + 1356.0, + 1212.0, + 1356.0, + 1244.0, + 1297.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1215.0, + 1099.0, + 1215.0, + 1099.0, + 1230.0, + 1081.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1145.0, + 1246.0, + 1160.0, + 1246.0, + 1160.0, + 1258.0, + 1145.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1251.0, + 915.0, + 1251.0, + 915.0, + 1281.0, + 872.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1293.0, + 1100.0, + 1293.0, + 1100.0, + 1311.0, + 1082.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1310.0, + 956.0, + 1310.0, + 956.0, + 1327.0, + 939.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1324.0, + 915.0, + 1324.0, + 915.0, + 1353.0, + 870.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1319.0, + 1060.0, + 1319.0, + 1060.0, + 1348.0, + 969.0, + 1348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1339.0, + 1042.0, + 1339.0, + 1042.0, + 1367.0, + 986.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1397.0, + 916.0, + 1397.0, + 916.0, + 1427.0, + 871.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1466.0, + 922.0, + 1466.0, + 922.0, + 1503.0, + 868.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1484.0, + 1375.0, + 1484.0, + 1375.0, + 1516.0, + 1134.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1520.0, + 956.0, + 1520.0, + 956.0, + 1535.0, + 939.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1514.0, + 1383.0, + 1514.0, + 1383.0, + 1543.0, + 1137.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1539.0, + 916.0, + 1539.0, + 916.0, + 1573.0, + 870.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1564.0, + 955.0, + 1564.0, + 955.0, + 1585.0, + 940.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1560.0, + 1021.0, + 1560.0, + 1021.0, + 1589.0, + 997.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1560.0, + 1102.0, + 1560.0, + 1102.0, + 1589.0, + 1079.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1558.0, + 1173.0, + 1558.0, + 1173.0, + 1590.0, + 1135.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1559.0, + 1232.0, + 1559.0, + 1232.0, + 1590.0, + 1194.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 1556.0, + 1383.0, + 1556.0, + 1383.0, + 1594.0, + 1274.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1588.0, + 1278.0, + 1588.0, + 1278.0, + 1617.0, + 1034.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.75, + 1264.5, + 1025.75, + 1264.5, + 1025.75, + 1279.0, + 988.75, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.75, + 1307.5, + 964.75, + 1307.5, + 964.75, + 1315.5, + 947.75, + 1315.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.75, + 1413.5, + 1022.75, + 1413.5, + 1022.75, + 1425.0, + 984.75, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1628.0, + 1403.0, + 1628.0, + 1403.0, + 1661.0, + 845.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1657.0, + 1404.0, + 1657.0, + 1404.0, + 1691.0, + 845.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1689.0, + 1403.0, + 1689.0, + 1403.0, + 1721.0, + 844.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1718.0, + 1405.0, + 1718.0, + 1405.0, + 1751.0, + 844.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1748.0, + 1405.0, + 1748.0, + 1405.0, + 1782.0, + 845.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1780.0, + 1406.0, + 1780.0, + 1406.0, + 1813.0, + 845.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1811.0, + 1404.0, + 1811.0, + 1404.0, + 1844.0, + 845.0, + 1844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1841.0, + 1405.0, + 1841.0, + 1405.0, + 1874.0, + 845.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1871.0, + 1296.0, + 1871.0, + 1296.0, + 1904.0, + 844.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 216.0, + 1106.0, + 216.0, + 1106.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 274.0, + 1168.0, + 274.0, + 1168.0, + 325.0, + 297.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1045.0, + 578.0, + 1045.0, + 578.0, + 1092.0, + 294.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 518.0, + 932.0, + 518.0, + 932.0, + 556.0, + 769.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 861.0, + 2088.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 591.0, + 1303.0, + 591.0, + 1303.0, + 626.0, + 394.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 618.0, + 1304.0, + 618.0, + 1304.0, + 657.0, + 393.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 648.0, + 1306.0, + 648.0, + 1306.0, + 686.0, + 393.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 681.0, + 1306.0, + 681.0, + 1306.0, + 715.0, + 394.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 712.0, + 1306.0, + 712.0, + 1306.0, + 745.0, + 392.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 744.0, + 1305.0, + 744.0, + 1305.0, + 776.0, + 393.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 773.0, + 1305.0, + 773.0, + 1305.0, + 805.0, + 393.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 804.0, + 618.0, + 804.0, + 618.0, + 839.0, + 394.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 804.0, + 1306.0, + 804.0, + 1306.0, + 839.0, + 661.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 836.0, + 808.0, + 836.0, + 808.0, + 868.0, + 394.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 836.0, + 1303.0, + 836.0, + 1303.0, + 868.0, + 866.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 865.0, + 608.0, + 865.0, + 608.0, + 897.0, + 395.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 865.0, + 698.0, + 865.0, + 698.0, + 897.0, + 666.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 865.0, + 1305.0, + 865.0, + 1305.0, + 897.0, + 755.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 895.0, + 659.0, + 895.0, + 659.0, + 929.0, + 393.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 895.0, + 1306.0, + 895.0, + 1306.0, + 929.0, + 713.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 925.0, + 1305.0, + 925.0, + 1305.0, + 960.0, + 394.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 956.0, + 1288.0, + 956.0, + 1288.0, + 993.0, + 392.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1120.0, + 824.0, + 1120.0, + 824.0, + 1154.0, + 296.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 827.0, + 1148.0, + 827.0, + 1185.0, + 295.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1180.0, + 826.0, + 1180.0, + 826.0, + 1214.0, + 296.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1211.0, + 827.0, + 1211.0, + 827.0, + 1246.0, + 296.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1240.0, + 825.0, + 1240.0, + 825.0, + 1276.0, + 294.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1272.0, + 825.0, + 1272.0, + 825.0, + 1304.0, + 296.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1302.0, + 827.0, + 1302.0, + 827.0, + 1336.0, + 295.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1332.0, + 826.0, + 1332.0, + 826.0, + 1366.0, + 295.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1362.0, + 827.0, + 1362.0, + 827.0, + 1397.0, + 294.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1392.0, + 825.0, + 1392.0, + 825.0, + 1428.0, + 295.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1424.0, + 825.0, + 1424.0, + 825.0, + 1458.0, + 294.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 827.0, + 1453.0, + 827.0, + 1488.0, + 294.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1486.0, + 828.0, + 1486.0, + 828.0, + 1518.0, + 294.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1516.0, + 827.0, + 1516.0, + 827.0, + 1548.0, + 294.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1547.0, + 826.0, + 1547.0, + 826.0, + 1580.0, + 294.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 828.0, + 1574.0, + 828.0, + 1610.0, + 293.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1608.0, + 828.0, + 1608.0, + 828.0, + 1640.0, + 293.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1639.0, + 827.0, + 1639.0, + 827.0, + 1668.0, + 294.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1670.0, + 827.0, + 1670.0, + 827.0, + 1700.0, + 296.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 826.0, + 1698.0, + 826.0, + 1731.0, + 295.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1726.0, + 827.0, + 1726.0, + 827.0, + 1764.0, + 295.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1761.0, + 828.0, + 1761.0, + 828.0, + 1794.0, + 295.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 827.0, + 1790.0, + 827.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1822.0, + 827.0, + 1822.0, + 827.0, + 1852.0, + 294.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 827.0, + 1851.0, + 827.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1881.0, + 827.0, + 1881.0, + 827.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 827.0, + 1912.0, + 827.0, + 1943.0, + 294.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1402.0, + 1942.0, + 1402.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1974.0, + 1403.0, + 1974.0, + 1403.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 378.0, + 560.0, + 378.0, + 560.0, + 410.0, + 315.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 406.0, + 681.0, + 406.0, + 681.0, + 441.0, + 312.0, + 441.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 307, + 1403, + 307, + 1403, + 522, + 299, + 522 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1420, + 1403, + 1420, + 1403, + 1636, + 298, + 1636 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1650, + 1403, + 1650, + 1403, + 1899, + 298, + 1899 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1401, + 1911, + 1401, + 2034, + 299, + 2034 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 370, + 595, + 1403, + 595, + 1403, + 1150, + 370, + 1150 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1400, + 229, + 1400, + 292, + 294, + 292 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 300, + 1264, + 1403, + 1264, + 1403, + 1326, + 300, + 1326 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 299, + 537, + 798, + 537, + 798, + 568, + 299, + 568 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 299, + 1194, + 679, + 1194, + 679, + 1230, + 299, + 1230 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 300, + 1363, + 796, + 1363, + 796, + 1395, + 300, + 1395 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.727 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.703 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.176 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.173 + }, + { + "category_id": 13, + "poly": [ + 1105, + 1682, + 1283, + 1682, + 1283, + 1717, + 1105, + 1717 + ], + "score": 0.93, + "latex": "z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )" + }, + { + "category_id": 13, + "poly": [ + 746, + 1773, + 890, + 1773, + 890, + 1808, + 746, + 1808 + ], + "score": 0.92, + "latex": "\\{ z _ { u , v } \\} _ { u \\leq i , v } )" + }, + { + "category_id": 13, + "poly": [ + 1286, + 1943, + 1338, + 1943, + 1338, + 1976, + 1286, + 1976 + ], + "score": 0.92, + "latex": "\\{ z _ { l } \\}" + }, + { + "category_id": 13, + "poly": [ + 634, + 1868, + 826, + 1868, + 826, + 1899, + 634, + 1899 + ], + "score": 0.91, + "latex": "\\hat { \\boldsymbol { z } } _ { i + k , j } = \\boldsymbol { W } _ { k } \\boldsymbol { c } _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 1092, + 1804, + 1166, + 1804, + 1166, + 1832, + 1092, + 1832 + ], + "score": 0.9, + "latex": "k > 0" + }, + { + "category_id": 13, + "poly": [ + 996, + 1836, + 1069, + 1836, + 1069, + 1863, + 996, + 1863 + ], + "score": 0.9, + "latex": "k > 0" + }, + { + "category_id": 13, + "poly": [ + 562, + 1808, + 633, + 1808, + 633, + 1838, + 562, + 1838 + ], + "score": 0.9, + "latex": "z _ { i + k , j }" + }, + { + "category_id": 13, + "poly": [ + 847, + 1683, + 876, + 1683, + 876, + 1713, + 847, + 1713 + ], + "score": 0.88, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1112, + 1747, + 1153, + 1747, + 1153, + 1776, + 1112, + 1776 + ], + "score": 0.88, + "latex": "c _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 1351, + 1835, + 1392, + 1835, + 1392, + 1865, + 1351, + 1865 + ], + "score": 0.88, + "latex": "W _ { k }" + }, + { + "category_id": 13, + "poly": [ + 963, + 1808, + 1004, + 1808, + 1004, + 1837, + 963, + 1837 + ], + "score": 0.88, + "latex": "c _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 760, + 1257, + 760, + 1257, + 789, + 1202, + 789 + ], + "score": 0.88, + "latex": "28 \\%" + }, + { + "category_id": 13, + "poly": [ + 1346, + 1658, + 1393, + 1658, + 1393, + 1685, + 1346, + 1685 + ], + "score": 0.88, + "latex": "\\mathbf { \\delta } _ { \\mathbf { x } _ { i , j } }" + }, + { + "category_id": 13, + "poly": [ + 1327, + 791, + 1401, + 791, + 1401, + 821, + 1327, + 821 + ], + "score": 0.87, + "latex": "( + 1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 534, + 657, + 605, + 657, + 605, + 687, + 534, + 687 + ], + "score": 0.87, + "latex": "+ 1 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 806, + 1717, + 838, + 1717, + 838, + 1746, + 806, + 1746 + ], + "score": 0.87, + "latex": "g _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 724, + 1840, + 765, + 1840, + 765, + 1868, + 724, + 1868 + ], + "score": 0.86, + "latex": "c _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 405, + 882, + 462, + 882, + 462, + 912, + 405, + 912 + ], + "score": 0.85, + "latex": "+ 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 943, + 760, + 997, + 760, + 997, + 790, + 943, + 790 + ], + "score": 0.85, + "latex": "73 \\%" + }, + { + "category_id": 13, + "poly": [ + 602, + 730, + 641, + 730, + 641, + 759, + 602, + 759 + ], + "score": 0.85, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 637, + 1947, + 707, + 1947, + 707, + 1977, + 637, + 1977 + ], + "score": 0.82, + "latex": "z _ { i + k , j }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1191.0, + 683.0, + 1191.0, + 683.0, + 1234.0, + 292.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1363.0, + 799.0, + 1363.0, + 799.0, + 1399.0, + 295.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 305.0, + 1407.0, + 305.0, + 1407.0, + 346.0, + 293.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 341.0, + 1404.0, + 341.0, + 1404.0, + 371.0, + 295.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1404.0, + 369.0, + 1404.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 398.0, + 1404.0, + 398.0, + 1404.0, + 435.0, + 293.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 429.0, + 1404.0, + 429.0, + 1404.0, + 464.0, + 295.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 460.0, + 1407.0, + 460.0, + 1407.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 490.0, + 703.0, + 490.0, + 703.0, + 527.0, + 294.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1419.0, + 1405.0, + 1419.0, + 1405.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1487.0, + 294.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1484.0, + 1403.0, + 1484.0, + 1403.0, + 1518.0, + 294.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1516.0, + 1404.0, + 1516.0, + 1404.0, + 1547.0, + 296.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1542.0, + 1407.0, + 1542.0, + 1407.0, + 1581.0, + 292.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1405.0, + 1574.0, + 1405.0, + 1607.0, + 293.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1605.0, + 1407.0, + 1605.0, + 1407.0, + 1642.0, + 292.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1644.0, + 1345.0, + 1644.0, + 1345.0, + 1692.0, + 291.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1644.0, + 1409.0, + 1644.0, + 1409.0, + 1692.0, + 1394.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1679.0, + 846.0, + 1679.0, + 846.0, + 1719.0, + 292.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1679.0, + 1104.0, + 1679.0, + 1104.0, + 1719.0, + 877.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1679.0, + 1407.0, + 1679.0, + 1407.0, + 1719.0, + 1284.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 805.0, + 1711.0, + 805.0, + 1749.0, + 293.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1711.0, + 1407.0, + 1711.0, + 1407.0, + 1749.0, + 839.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1742.0, + 1111.0, + 1742.0, + 1111.0, + 1779.0, + 292.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1742.0, + 1407.0, + 1742.0, + 1407.0, + 1779.0, + 1154.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 745.0, + 1771.0, + 745.0, + 1812.0, + 292.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1812.0, + 891.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1800.0, + 561.0, + 1800.0, + 561.0, + 1841.0, + 296.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1800.0, + 962.0, + 1800.0, + 962.0, + 1841.0, + 634.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1800.0, + 1091.0, + 1800.0, + 1091.0, + 1841.0, + 1005.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1800.0, + 1407.0, + 1800.0, + 1407.0, + 1841.0, + 1167.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1834.0, + 723.0, + 1834.0, + 723.0, + 1871.0, + 293.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1834.0, + 995.0, + 1834.0, + 995.0, + 1871.0, + 766.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1834.0, + 1350.0, + 1834.0, + 1350.0, + 1871.0, + 1070.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1834.0, + 1407.0, + 1834.0, + 1407.0, + 1871.0, + 1393.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1864.0, + 633.0, + 1864.0, + 633.0, + 1903.0, + 294.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1864.0, + 838.0, + 1864.0, + 838.0, + 1903.0, + 827.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1912.0, + 1403.0, + 1912.0, + 1403.0, + 1945.0, + 297.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 636.0, + 1942.0, + 636.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1942.0, + 1285.0, + 1942.0, + 1285.0, + 1977.0, + 708.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1977.0, + 1339.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1405.0, + 2003.0, + 1405.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 594.0, + 1405.0, + 594.0, + 1405.0, + 632.0, + 370.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 627.0, + 1403.0, + 627.0, + 1403.0, + 661.0, + 393.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 658.0, + 533.0, + 658.0, + 533.0, + 692.0, + 392.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 658.0, + 1069.0, + 658.0, + 1069.0, + 692.0, + 606.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 699.0, + 1403.0, + 699.0, + 1403.0, + 733.0, + 384.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 727.0, + 601.0, + 727.0, + 601.0, + 767.0, + 392.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 727.0, + 1405.0, + 727.0, + 1405.0, + 767.0, + 642.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 759.0, + 942.0, + 759.0, + 942.0, + 794.0, + 392.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 759.0, + 1201.0, + 759.0, + 1201.0, + 794.0, + 998.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 759.0, + 1405.0, + 759.0, + 1405.0, + 794.0, + 1258.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 789.0, + 1326.0, + 789.0, + 1326.0, + 826.0, + 392.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 789.0, + 1405.0, + 789.0, + 1405.0, + 826.0, + 1402.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 820.0, + 1406.0, + 820.0, + 1406.0, + 857.0, + 393.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 853.0, + 1405.0, + 853.0, + 1405.0, + 887.0, + 393.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 882.0, + 404.0, + 882.0, + 404.0, + 916.0, + 393.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 882.0, + 658.0, + 882.0, + 658.0, + 916.0, + 463.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 919.0, + 1405.0, + 919.0, + 1405.0, + 959.0, + 386.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 954.0, + 1405.0, + 954.0, + 1405.0, + 988.0, + 394.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 983.0, + 1405.0, + 983.0, + 1405.0, + 1021.0, + 393.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1010.0, + 823.0, + 1010.0, + 823.0, + 1053.0, + 392.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1053.0, + 1406.0, + 1053.0, + 1406.0, + 1093.0, + 377.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1088.0, + 1405.0, + 1088.0, + 1405.0, + 1122.0, + 396.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1119.0, + 1301.0, + 1119.0, + 1301.0, + 1153.0, + 396.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 257.0, + 1093.0, + 257.0, + 1093.0, + 300.0, + 293.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1260.0, + 1405.0, + 1260.0, + 1405.0, + 1301.0, + 294.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1295.0, + 1125.0, + 1295.0, + 1125.0, + 1328.0, + 294.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 532.0, + 803.0, + 532.0, + 803.0, + 574.0, + 294.0, + 574.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1849, + 1404, + 1849, + 1404, + 2035, + 296, + 2035 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1424, + 1405, + 1424, + 1405, + 1580, + 298, + 1580 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 1682, + 1404, + 1682, + 1404, + 1837, + 296, + 1837 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1208, + 1403, + 1208, + 1403, + 1335, + 298, + 1335 + ], + "score": 0.975 + }, + { + "category_id": 4, + "poly": [ + 296, + 659, + 1406, + 659, + 1406, + 1027, + 296, + 1027 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 299, + 232, + 1400, + 232, + 1400, + 648, + 299, + 648 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 367, + 1587, + 1326, + 1587, + 1326, + 1677, + 367, + 1677 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 317, + 1114, + 1380, + 1114, + 1380, + 1202, + 317, + 1202 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 1072, + 1184, + 1072, + 1184, + 1106, + 297, + 1106 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 299, + 1368, + 654, + 1368, + 654, + 1399, + 299, + 1399 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.767 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.659 + }, + { + "category_id": 2, + "poly": [ + 300, + 77, + 856, + 77, + 856, + 103, + 300, + 103 + ], + "score": 0.437 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.343 + }, + { + "category_id": 14, + "poly": [ + 370, + 1583, + 1329, + 1583, + 1329, + 1679, + 370, + 1679 + ], + "score": 0.94, + "latex": "\\theta ^ { * } = \\arg \\operatorname* { m i n } _ { \\theta } \\frac { 1 } { N } \\sum _ { n = 1 } ^ { N } \\mathcal { L } _ { \\mathrm { C P C } } [ f _ { \\theta } ( x _ { n } ) ] , \\quad \\psi ^ { * } = \\arg \\operatorname* { m i n } _ { \\psi } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } \\mathcal { L } _ { \\mathrm { S u p } } [ h _ { \\psi } \\circ f _ { \\theta ^ { * } } ( x _ { m } ) , y _ { m } ]" + }, + { + "category_id": 14, + "poly": [ + 319, + 1111, + 1377, + 1111, + 1377, + 1205, + 319, + 1205 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { \\mathrm { C P C } } = - \\sum _ { i , j , k } \\log p ( z _ { i + k , j } | \\hat { z } _ { i + k , j } , \\{ z _ { l } \\} ) = - \\sum _ { i , j , k } \\log \\frac { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) } { \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { i + k , j } ) + \\sum _ { l } \\exp ( \\hat { z } _ { i + k , j } ^ { T } z _ { l } ) }" + }, + { + "category_id": 13, + "poly": [ + 777, + 1456, + 898, + 1456, + 898, + 1490, + 777, + 1490 + ], + "score": 0.93, + "latex": "z = f _ { \\boldsymbol { \\theta } } ( \\pmb { x } )" + }, + { + "category_id": 13, + "poly": [ + 1059, + 1547, + 1239, + 1547, + 1239, + 1581, + 1059, + 1581 + ], + "score": 0.92, + "latex": "\\mathbb { D } _ { l } = \\{ x _ { m } , y _ { m } \\}" + }, + { + "category_id": 13, + "poly": [ + 542, + 1210, + 595, + 1210, + 595, + 1243, + 542, + 1243 + ], + "score": 0.92, + "latex": "\\left\\{ z _ { l } \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 298, + 1547, + 358, + 1547, + 358, + 1581, + 298, + 1581 + ], + "score": 0.91, + "latex": "\\left\\{ x _ { n } \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 1327, + 1425, + 1400, + 1425, + 1400, + 1459, + 1327, + 1459 + ], + "score": 0.91, + "latex": "\\{ W _ { k } \\}" + }, + { + "category_id": 13, + "poly": [ + 774, + 1306, + 818, + 1306, + 818, + 1335, + 774, + 1335 + ], + "score": 0.91, + "latex": "\\mathbf { } _ { c _ { i , j } }" + }, + { + "category_id": 13, + "poly": [ + 868, + 1305, + 943, + 1305, + 943, + 1336, + 868, + 1336 + ], + "score": 0.9, + "latex": "z _ { i + k , j }" + }, + { + "category_id": 13, + "poly": [ + 781, + 1945, + 835, + 1945, + 835, + 1976, + 781, + 1976 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { \\mathrm { { S u p } } }" + }, + { + "category_id": 13, + "poly": [ + 965, + 1743, + 1020, + 1743, + 1020, + 1776, + 965, + 1776 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { \\mathrm { { S u p } } }" + }, + { + "category_id": 13, + "poly": [ + 1002, + 1488, + 1038, + 1488, + 1038, + 1520, + 1002, + 1520 + ], + "score": 0.9, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 681, + 1744, + 717, + 1744, + 717, + 1776, + 681, + 1776 + ], + "score": 0.9, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 821, + 1683, + 859, + 1683, + 859, + 1713, + 821, + 1713 + ], + "score": 0.89, + "latex": "\\mathbb { D } _ { u }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 1518, + 1403, + 1518, + 1403, + 1549, + 1328, + 1549 + ], + "score": 0.89, + "latex": "\\mathbb { D } _ { u } =" + }, + { + "category_id": 13, + "poly": [ + 937, + 1883, + 973, + 1883, + 973, + 1915, + 937, + 1915 + ], + "score": 0.89, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 1231, + 1913, + 1264, + 1913, + 1264, + 1942, + 1231, + 1942 + ], + "score": 0.88, + "latex": "\\mathbb { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 811, + 1913, + 839, + 1913, + 839, + 1943, + 811, + 1943 + ], + "score": 0.88, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 1199, + 1713, + 1232, + 1713, + 1232, + 1743, + 1199, + 1743 + ], + "score": 0.87, + "latex": "\\mathbb { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 881, + 1805, + 908, + 1805, + 908, + 1836, + 881, + 1836 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 701, + 1427, + 730, + 1427, + 730, + 1458, + 701, + 1458 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 952, + 1431, + 982, + 1431, + 982, + 1459, + 952, + 1459 + ], + "score": 0.86, + "latex": "g _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 851, + 1549, + 883, + 1549, + 883, + 1575, + 851, + 1575 + ], + "score": 0.85, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1089, + 1519, + 1118, + 1519, + 1118, + 1545, + 1089, + 1545 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1343, + 1464, + 1362, + 1464, + 1362, + 1484, + 1343, + 1484 + ], + "score": 0.78, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 550, + 1272, + 577, + 1272, + 577, + 1299, + 550, + 1299 + ], + "score": 0.43, + "latex": "\\&" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 660.0, + 1406.0, + 660.0, + 1406.0, + 694.0, + 295.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 691.0, + 1404.0, + 691.0, + 1404.0, + 725.0, + 295.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 720.0, + 1406.0, + 720.0, + 1406.0, + 758.0, + 293.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 748.0, + 1406.0, + 748.0, + 1406.0, + 787.0, + 294.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 781.0, + 1407.0, + 781.0, + 1407.0, + 819.0, + 293.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 813.0, + 1404.0, + 813.0, + 1404.0, + 847.0, + 295.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 843.0, + 1406.0, + 843.0, + 1406.0, + 876.0, + 293.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 871.0, + 1407.0, + 871.0, + 1407.0, + 908.0, + 291.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 900.0, + 1407.0, + 900.0, + 1407.0, + 941.0, + 293.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 932.0, + 1406.0, + 932.0, + 1406.0, + 970.0, + 293.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 966.0, + 1404.0, + 966.0, + 1404.0, + 997.0, + 295.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 995.0, + 1348.0, + 995.0, + 1348.0, + 1029.0, + 295.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 255.0, + 572.0, + 255.0, + 572.0, + 296.0, + 464.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 242.0, + 892.0, + 242.0, + 892.0, + 302.0, + 756.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 249.0, + 975.0, + 249.0, + 975.0, + 295.0, + 913.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 239.0, + 1048.0, + 239.0, + 1048.0, + 297.0, + 992.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 249.0, + 1123.0, + 249.0, + 1123.0, + 296.0, + 1066.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 234.0, + 1196.0, + 234.0, + 1196.0, + 299.0, + 1142.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 248.0, + 1271.0, + 248.0, + 1271.0, + 294.0, + 1214.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 261.0, + 1341.0, + 261.0, + 1341.0, + 282.0, + 1290.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 224.0, + 1398.0, + 224.0, + 1398.0, + 324.0, + 1373.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 319.0, + 1071.0, + 319.0, + 1071.0, + 342.0, + 970.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 330.0, + 355.0, + 330.0, + 355.0, + 351.0, + 310.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 372.0, + 407.0, + 372.0, + 407.0, + 388.0, + 393.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 379.0, + 589.0, + 379.0, + 589.0, + 389.0, + 577.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 338.0, + 613.0, + 338.0, + 613.0, + 348.0, + 604.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 369.0, + 714.0, + 369.0, + 714.0, + 389.0, + 618.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 342.0, + 912.0, + 342.0, + 912.0, + 386.0, + 749.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 335.0, + 1123.0, + 335.0, + 1123.0, + 383.0, + 913.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 333.0, + 1189.0, + 333.0, + 1189.0, + 354.0, + 1150.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 333.0, + 1267.0, + 333.0, + 1267.0, + 383.0, + 1217.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 354.0, + 1290.0, + 354.0, + 1290.0, + 363.0, + 1277.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 339.0, + 1337.0, + 339.0, + 1337.0, + 377.0, + 1298.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 397.0, + 1061.0, + 397.0, + 1061.0, + 429.0, + 981.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 400.0, + 1396.0, + 400.0, + 1396.0, + 492.0, + 1367.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 423.0, + 654.0, + 423.0, + 654.0, + 433.0, + 644.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 421.0, + 973.0, + 421.0, + 973.0, + 441.0, + 913.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 423.0, + 1048.0, + 423.0, + 1048.0, + 439.0, + 991.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 421.0, + 1128.0, + 421.0, + 1128.0, + 441.0, + 1061.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 421.0, + 1197.0, + 421.0, + 1197.0, + 441.0, + 1141.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 423.0, + 1266.0, + 423.0, + 1266.0, + 440.0, + 1220.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 431.0, + 917.0, + 431.0, + 917.0, + 455.0, + 742.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 431.0, + 1335.0, + 431.0, + 1335.0, + 449.0, + 1297.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 450.0, + 915.0, + 450.0, + 915.0, + 470.0, + 746.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 452.0, + 960.0, + 452.0, + 960.0, + 467.0, + 938.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 449.0, + 1031.0, + 449.0, + 1031.0, + 470.0, + 1008.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 449.0, + 1106.0, + 449.0, + 1106.0, + 470.0, + 1067.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 449.0, + 1256.0, + 449.0, + 1256.0, + 470.0, + 1217.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 442.0, + 1331.0, + 442.0, + 1331.0, + 461.0, + 1302.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 483.0, + 1055.0, + 483.0, + 1055.0, + 503.0, + 986.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 496.0, + 1061.0, + 496.0, + 1061.0, + 516.0, + 980.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 508.0, + 967.0, + 508.0, + 967.0, + 529.0, + 921.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 509.0, + 1050.0, + 509.0, + 1050.0, + 526.0, + 991.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 508.0, + 1201.0, + 508.0, + 1201.0, + 528.0, + 1051.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 508.0, + 1261.0, + 508.0, + 1261.0, + 529.0, + 1225.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 517.0, + 898.0, + 517.0, + 898.0, + 544.0, + 761.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 517.0, + 1334.0, + 517.0, + 1334.0, + 538.0, + 1299.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 537.0, + 898.0, + 537.0, + 898.0, + 557.0, + 762.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 538.0, + 958.0, + 538.0, + 958.0, + 557.0, + 936.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 534.0, + 1033.0, + 534.0, + 1033.0, + 559.0, + 1006.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 536.0, + 1106.0, + 536.0, + 1106.0, + 557.0, + 1068.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 533.0, + 1184.0, + 533.0, + 1184.0, + 559.0, + 1155.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 536.0, + 1256.0, + 536.0, + 1256.0, + 557.0, + 1216.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1299.0, + 528.0, + 1334.0, + 528.0, + 1334.0, + 549.0, + 1299.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 599.0, + 911.0, + 599.0, + 911.0, + 631.0, + 747.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 595.0, + 975.0, + 595.0, + 975.0, + 615.0, + 913.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 593.0, + 1199.0, + 593.0, + 1199.0, + 612.0, + 1140.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 594.0, + 1267.0, + 594.0, + 1267.0, + 615.0, + 1219.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 602.0, + 1336.0, + 602.0, + 1336.0, + 624.0, + 1297.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 574.0, + 1396.0, + 574.0, + 1396.0, + 652.0, + 1367.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 620.0, + 916.0, + 620.0, + 916.0, + 640.0, + 744.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 625.0, + 956.0, + 625.0, + 956.0, + 641.0, + 938.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 620.0, + 1184.0, + 620.0, + 1184.0, + 646.0, + 1155.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 622.0, + 1256.0, + 622.0, + 1256.0, + 643.0, + 1217.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 615.0, + 1331.0, + 615.0, + 1331.0, + 637.0, + 1300.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 348.5, + 720.0, + 348.5, + 720.0, + 376.5, + 610.0, + 376.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 360.0, + 1181.0, + 360.0, + 1181.0, + 385.0, + 1156.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 448.0, + 1181.0, + 448.0, + 1181.0, + 471.5, + 1157.0, + 471.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 519.5, + 440.0, + 519.5, + 440.0, + 553.0, + 358.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1367.0, + 657.0, + 1367.0, + 657.0, + 1403.0, + 295.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1888.0, + 291.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1882.0, + 936.0, + 1882.0, + 936.0, + 1916.0, + 291.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 974.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 810.0, + 1909.0, + 810.0, + 1948.0, + 293.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1909.0, + 1230.0, + 1909.0, + 1230.0, + 1948.0, + 840.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1265.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1948.0, + 1265.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 780.0, + 1941.0, + 780.0, + 1980.0, + 293.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1980.0, + 836.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 509.0, + 2001.0, + 509.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1423.0, + 700.0, + 1423.0, + 700.0, + 1461.0, + 295.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 731.0, + 1423.0, + 951.0, + 1423.0, + 951.0, + 1461.0, + 731.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 1423.0, + 1326.0, + 1423.0, + 1326.0, + 1461.0, + 983.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1457.0, + 776.0, + 1457.0, + 776.0, + 1490.0, + 296.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1457.0, + 1342.0, + 1457.0, + 1342.0, + 1490.0, + 899.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 1457.0, + 1403.0, + 1457.0, + 1403.0, + 1490.0, + 1363.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1483.0, + 1001.0, + 1483.0, + 1001.0, + 1522.0, + 295.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1483.0, + 1403.0, + 1483.0, + 1403.0, + 1522.0, + 1039.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1514.0, + 1088.0, + 1514.0, + 1088.0, + 1555.0, + 293.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1514.0, + 1327.0, + 1514.0, + 1327.0, + 1555.0, + 1119.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1543.0, + 850.0, + 1543.0, + 850.0, + 1587.0, + 359.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1543.0, + 1058.0, + 1543.0, + 1058.0, + 1587.0, + 884.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1543.0, + 1251.0, + 1543.0, + 1251.0, + 1587.0, + 1240.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1680.0, + 820.0, + 1680.0, + 820.0, + 1717.0, + 294.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1680.0, + 1406.0, + 1680.0, + 1406.0, + 1717.0, + 860.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1710.0, + 1198.0, + 1710.0, + 1198.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1710.0, + 1404.0, + 1710.0, + 1404.0, + 1746.0, + 1233.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 680.0, + 1742.0, + 680.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1742.0, + 964.0, + 1742.0, + 964.0, + 1779.0, + 718.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1779.0, + 1021.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1771.0, + 1406.0, + 1771.0, + 1406.0, + 1807.0, + 291.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1801.0, + 880.0, + 1801.0, + 880.0, + 1841.0, + 293.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1801.0, + 919.0, + 1801.0, + 919.0, + 1841.0, + 909.0, + 1841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1208.0, + 541.0, + 1208.0, + 541.0, + 1244.0, + 293.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1208.0, + 1403.0, + 1208.0, + 1403.0, + 1244.0, + 596.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 1407.0, + 1240.0, + 1407.0, + 1276.0, + 293.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1271.0, + 549.0, + 1271.0, + 549.0, + 1305.0, + 294.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1271.0, + 1405.0, + 1271.0, + 1405.0, + 1305.0, + 578.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1301.0, + 773.0, + 1301.0, + 773.0, + 1338.0, + 293.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1301.0, + 867.0, + 1301.0, + 867.0, + 1338.0, + 819.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1301.0, + 1258.0, + 1301.0, + 1258.0, + 1338.0, + 944.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1071.0, + 1188.0, + 1071.0, + 1188.0, + 1110.0, + 295.0, + 1110.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1668, + 1403, + 1668, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1286, + 1404, + 1286, + 1404, + 1651, + 298, + 1651 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 629, + 1403, + 629, + 1403, + 813, + 299, + 813 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 445, + 298, + 445 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 460, + 1403, + 460, + 1403, + 614, + 299, + 614 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 978, + 1402, + 978, + 1402, + 1101, + 299, + 1101 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1117, + 1402, + 1117, + 1402, + 1270, + 299, + 1270 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 301, + 888, + 586, + 888, + 586, + 924, + 301, + 924 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.754 + }, + { + "category_id": 13, + "poly": [ + 635, + 553, + 689, + 553, + 689, + 585, + 635, + 585 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { \\mathrm { { S u p } } }" + }, + { + "category_id": 13, + "poly": [ + 1055, + 353, + 1110, + 353, + 1110, + 385, + 1055, + 385 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { \\mathrm { { S u p } } }" + }, + { + "category_id": 13, + "poly": [ + 730, + 262, + 766, + 262, + 766, + 295, + 730, + 295 + ], + "score": 0.89, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 1077, + 1071, + 1113, + 1071, + 1113, + 1105, + 1077, + 1105 + ], + "score": 0.89, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 548, + 552, + 583, + 552, + 583, + 585, + 548, + 585 + ], + "score": 0.88, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 792, + 522, + 824, + 522, + 824, + 552, + 792, + 552 + ], + "score": 0.88, + "latex": "\\mathbb { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 516, + 323, + 547, + 323, + 547, + 352, + 516, + 352 + ], + "score": 0.87, + "latex": "\\mathbb { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1259, + 1041, + 1288, + 1041, + 1288, + 1072, + 1259, + 1072 + ], + "score": 0.87, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 471, + 353, + 538, + 353, + 538, + 382, + 471, + 382 + ], + "score": 0.87, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 362, + 352, + 418, + 352, + 418, + 382, + 362, + 382 + ], + "score": 0.86, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 352, + 351, + 352, + 351, + 382, + 296, + 382 + ], + "score": 0.84, + "latex": "20 \\%" + }, + { + "category_id": 13, + "poly": [ + 1340, + 322, + 1393, + 322, + 1393, + 352, + 1340, + 352 + ], + "score": 0.83, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 1288, + 322, + 1329, + 322, + 1329, + 352, + 1288, + 352 + ], + "score": 0.78, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1238, + 322, + 1279, + 322, + 1279, + 351, + 1238, + 351 + ], + "score": 0.77, + "latex": "2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1190, + 322, + 1229, + 322, + 1229, + 351, + 1190, + 351 + ], + "score": 0.76, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 651, + 721, + 681, + 721, + 681, + 750, + 651, + 750 + ], + "score": 0.61, + "latex": "\\theta ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 690, + 721, + 725, + 721, + 725, + 752, + 690, + 752 + ], + "score": 0.55, + "latex": "\\psi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 651, + 721, + 723, + 721, + 723, + 752, + 651, + 752 + ], + "score": 0.4, + "latex": "\\theta ^ { * } , \\psi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 923, + 630, + 951, + 630, + 951, + 662, + 923, + 662 + ], + "score": 0.34, + "latex": "f _ { \\theta }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 886.0, + 592.0, + 886.0, + 592.0, + 929.0, + 292.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1665.0, + 1407.0, + 1665.0, + 1407.0, + 1705.0, + 293.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1699.0, + 1404.0, + 1699.0, + 1404.0, + 1730.0, + 296.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1730.0, + 1405.0, + 1730.0, + 1405.0, + 1764.0, + 292.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1760.0, + 1405.0, + 1760.0, + 1405.0, + 1793.0, + 293.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1792.0, + 1404.0, + 1792.0, + 1404.0, + 1822.0, + 296.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1821.0, + 1404.0, + 1821.0, + 1404.0, + 1855.0, + 293.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1403.0, + 1851.0, + 1403.0, + 1884.0, + 293.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1914.0, + 293.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1948.0, + 293.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1977.0, + 294.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1407.0, + 1970.0, + 1407.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1315.0, + 2003.0, + 1315.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1323.0, + 293.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1316.0, + 1405.0, + 1316.0, + 1405.0, + 1353.0, + 293.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 1405.0, + 1349.0, + 1405.0, + 1383.0, + 294.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1376.0, + 1406.0, + 1376.0, + 1406.0, + 1414.0, + 291.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1408.0, + 1405.0, + 1408.0, + 1405.0, + 1442.0, + 293.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1439.0, + 1405.0, + 1439.0, + 1405.0, + 1473.0, + 293.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1467.0, + 1405.0, + 1467.0, + 1405.0, + 1504.0, + 293.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 1405.0, + 1500.0, + 1405.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1567.0, + 292.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1557.0, + 1405.0, + 1557.0, + 1405.0, + 1597.0, + 291.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1591.0, + 1406.0, + 1591.0, + 1406.0, + 1626.0, + 292.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1623.0, + 640.0, + 1623.0, + 640.0, + 1654.0, + 294.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 628.0, + 922.0, + 628.0, + 922.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 628.0, + 1404.0, + 628.0, + 1404.0, + 664.0, + 952.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 654.0, + 1407.0, + 654.0, + 1407.0, + 698.0, + 292.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 691.0, + 1405.0, + 691.0, + 1405.0, + 722.0, + 295.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 650.0, + 720.0, + 650.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 720.0, + 1407.0, + 720.0, + 1407.0, + 755.0, + 726.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 751.0, + 1404.0, + 751.0, + 1404.0, + 783.0, + 295.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 779.0, + 1294.0, + 779.0, + 1294.0, + 819.0, + 293.0, + 819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 729.0, + 258.0, + 729.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 258.0, + 1405.0, + 258.0, + 1405.0, + 296.0, + 767.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1403.0, + 289.0, + 1403.0, + 326.0, + 292.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 320.0, + 515.0, + 320.0, + 515.0, + 356.0, + 294.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 320.0, + 1189.0, + 320.0, + 1189.0, + 356.0, + 548.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 320.0, + 1237.0, + 320.0, + 1237.0, + 356.0, + 1230.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 320.0, + 1287.0, + 320.0, + 1287.0, + 356.0, + 1280.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 320.0, + 1339.0, + 320.0, + 1339.0, + 356.0, + 1330.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 356.0, + 1394.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 349.0, + 295.0, + 349.0, + 295.0, + 390.0, + 292.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 349.0, + 361.0, + 349.0, + 361.0, + 390.0, + 352.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 349.0, + 470.0, + 349.0, + 470.0, + 390.0, + 419.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 349.0, + 1054.0, + 349.0, + 1054.0, + 390.0, + 539.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 349.0, + 1407.0, + 349.0, + 1407.0, + 390.0, + 1111.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 380.0, + 1405.0, + 380.0, + 1405.0, + 420.0, + 293.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 412.0, + 1186.0, + 412.0, + 1186.0, + 451.0, + 295.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 496.0, + 297.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 524.0, + 297.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 791.0, + 520.0, + 791.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 520.0, + 1407.0, + 520.0, + 1407.0, + 557.0, + 825.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 549.0, + 547.0, + 549.0, + 547.0, + 590.0, + 293.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 549.0, + 634.0, + 549.0, + 634.0, + 590.0, + 584.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 549.0, + 1407.0, + 549.0, + 1407.0, + 590.0, + 690.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 578.0, + 1295.0, + 578.0, + 1295.0, + 622.0, + 292.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 978.0, + 1407.0, + 978.0, + 1407.0, + 1014.0, + 294.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1010.0, + 1404.0, + 1010.0, + 1404.0, + 1043.0, + 295.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1042.0, + 1258.0, + 1042.0, + 1258.0, + 1073.0, + 294.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 1042.0, + 1404.0, + 1042.0, + 1404.0, + 1073.0, + 1289.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1071.0, + 1076.0, + 1071.0, + 1076.0, + 1106.0, + 293.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1071.0, + 1401.0, + 1071.0, + 1401.0, + 1106.0, + 1114.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1117.0, + 1406.0, + 1117.0, + 1406.0, + 1154.0, + 294.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1404.0, + 1146.0, + 1404.0, + 1183.0, + 294.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1179.0, + 1404.0, + 1179.0, + 1404.0, + 1212.0, + 295.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1207.0, + 1406.0, + 1207.0, + 1406.0, + 1245.0, + 293.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1240.0, + 725.0, + 1240.0, + 725.0, + 1275.0, + 294.0, + 1275.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 625, + 1404, + 625, + 1404, + 932, + 298, + 932 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1404, + 229, + 1404, + 507, + 298, + 507 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1576, + 823, + 1576, + 823, + 2033, + 299, + 2033 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1233, + 825, + 1233, + 825, + 1537, + 298, + 1537 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1032, + 1404, + 1032, + 1404, + 1217, + 298, + 1217 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 854, + 1248, + 1393, + 1248, + 1393, + 1713, + 854, + 1713 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 847, + 1724, + 1402, + 1724, + 1402, + 1999, + 847, + 1999 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 300, + 973, + 682, + 973, + 682, + 1004, + 300, + 1004 + ], + "score": 0.892 + }, + { + "category_id": 0, + "poly": [ + 300, + 554, + 480, + 554, + 480, + 589, + 300, + 589 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 857, + 76, + 857, + 104, + 298, + 104 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.732 + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1248.0, + 1063.0, + 1248.0, + 1063.0, + 1297.0, + 867.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 1258.0, + 1370.0, + 1258.0, + 1370.0, + 1285.0, + 1341.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 938.0, + 1299.0, + 1070.0, + 1299.0, + 1070.0, + 1331.0, + 938.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1318.0, + 1235.0, + 1318.0, + 1235.0, + 1342.0, + 1208.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1353.0, + 930.0, + 1353.0, + 930.0, + 1387.0, + 869.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1363.0, + 956.0, + 1363.0, + 956.0, + 1613.0, + 824.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1440.0, + 1026.0, + 1440.0, + 1026.0, + 1447.0, + 1018.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1451.0, + 928.0, + 1451.0, + 928.0, + 1488.0, + 869.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1477.0, + 969.0, + 1477.0, + 969.0, + 1501.0, + 941.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1501.0, + 929.0, + 1501.0, + 929.0, + 1534.0, + 868.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1509.0, + 1170.0, + 1509.0, + 1170.0, + 1535.0, + 1139.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 1507.0, + 1364.0, + 1507.0, + 1364.0, + 1522.0, + 1345.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1557.0, + 1039.0, + 1557.0, + 1039.0, + 1587.0, + 1004.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1598.0, + 935.0, + 1598.0, + 935.0, + 1640.0, + 869.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1604.0, + 972.0, + 1604.0, + 972.0, + 1636.0, + 937.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1640.0, + 996.0, + 1640.0, + 996.0, + 1677.0, + 911.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 1641.0, + 1396.0, + 1641.0, + 1396.0, + 1675.0, + 1312.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1674.0, + 1014.0, + 1674.0, + 1014.0, + 1683.0, + 1004.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1671.0, + 1287.0, + 1671.0, + 1287.0, + 1686.0, + 1267.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1668.0, + 1360.0, + 1668.0, + 1360.0, + 1690.0, + 1329.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1684.0, + 1347.0, + 1684.0, + 1347.0, + 1718.0, + 956.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.25, + 1279.5, + 882.25, + 1279.5, + 882.25, + 1409.5, + 878.25, + 1409.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 1533.0, + 1101.0, + 1533.0, + 1101.0, + 1548.0, + 1076.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1724.0, + 1406.0, + 1724.0, + 1406.0, + 1758.0, + 846.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1754.0, + 1406.0, + 1754.0, + 1406.0, + 1787.0, + 844.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1785.0, + 1403.0, + 1785.0, + 1403.0, + 1819.0, + 845.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1816.0, + 1404.0, + 1816.0, + 1404.0, + 1848.0, + 844.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1846.0, + 1403.0, + 1846.0, + 1403.0, + 1876.0, + 845.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1877.0, + 1403.0, + 1877.0, + 1403.0, + 1908.0, + 844.0, + 1908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1907.0, + 1403.0, + 1907.0, + 1403.0, + 1938.0, + 845.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1937.0, + 1405.0, + 1937.0, + 1405.0, + 1970.0, + 845.0, + 1970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1969.0, + 1330.0, + 1969.0, + 1330.0, + 1999.0, + 844.0, + 1999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 969.0, + 687.0, + 969.0, + 687.0, + 1009.0, + 292.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 549.0, + 486.0, + 549.0, + 486.0, + 597.0, + 291.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 625.0, + 1404.0, + 625.0, + 1404.0, + 662.0, + 295.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 658.0, + 1405.0, + 658.0, + 1405.0, + 693.0, + 294.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 687.0, + 1404.0, + 687.0, + 1404.0, + 721.0, + 293.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 719.0, + 1404.0, + 719.0, + 1404.0, + 754.0, + 294.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 749.0, + 1405.0, + 749.0, + 1405.0, + 782.0, + 292.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 780.0, + 1405.0, + 780.0, + 1405.0, + 814.0, + 293.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 808.0, + 1407.0, + 808.0, + 1407.0, + 845.0, + 291.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 841.0, + 1406.0, + 841.0, + 1406.0, + 876.0, + 293.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 873.0, + 1404.0, + 873.0, + 1404.0, + 904.0, + 294.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 901.0, + 1405.0, + 901.0, + 1405.0, + 935.0, + 293.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1401.0, + 230.0, + 1401.0, + 263.0, + 296.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 318.0, + 1405.0, + 318.0, + 1405.0, + 360.0, + 292.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 384.0, + 1402.0, + 384.0, + 1402.0, + 417.0, + 296.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 1404.0, + 413.0, + 1404.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 441.0, + 1405.0, + 441.0, + 1405.0, + 480.0, + 292.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 477.0, + 705.0, + 477.0, + 705.0, + 510.0, + 296.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1576.0, + 826.0, + 1576.0, + 826.0, + 1608.0, + 297.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 826.0, + 1605.0, + 826.0, + 1643.0, + 294.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1637.0, + 826.0, + 1637.0, + 826.0, + 1671.0, + 296.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1668.0, + 825.0, + 1668.0, + 825.0, + 1700.0, + 295.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 825.0, + 1698.0, + 825.0, + 1730.0, + 295.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1729.0, + 826.0, + 1729.0, + 826.0, + 1760.0, + 296.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1759.0, + 826.0, + 1759.0, + 826.0, + 1791.0, + 296.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1788.0, + 826.0, + 1788.0, + 826.0, + 1822.0, + 295.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1821.0, + 827.0, + 1821.0, + 827.0, + 1853.0, + 296.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 827.0, + 1850.0, + 827.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1881.0, + 825.0, + 1881.0, + 825.0, + 1913.0, + 296.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 825.0, + 1911.0, + 825.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 825.0, + 1943.0, + 825.0, + 1974.0, + 296.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 825.0, + 1972.0, + 825.0, + 2004.0, + 296.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2002.0, + 828.0, + 2002.0, + 828.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1230.0, + 827.0, + 1230.0, + 827.0, + 1267.0, + 297.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1264.0, + 826.0, + 1264.0, + 826.0, + 1297.0, + 296.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1292.0, + 825.0, + 1292.0, + 825.0, + 1328.0, + 294.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1324.0, + 826.0, + 1324.0, + 826.0, + 1357.0, + 295.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1357.0, + 824.0, + 1357.0, + 824.0, + 1387.0, + 296.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1384.0, + 825.0, + 1384.0, + 825.0, + 1417.0, + 294.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1416.0, + 826.0, + 1416.0, + 826.0, + 1449.0, + 295.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1443.0, + 825.0, + 1443.0, + 825.0, + 1479.0, + 294.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1475.0, + 827.0, + 1475.0, + 827.0, + 1513.0, + 292.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1507.0, + 430.0, + 1507.0, + 430.0, + 1538.0, + 295.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1034.0, + 1404.0, + 1034.0, + 1404.0, + 1066.0, + 297.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1065.0, + 1405.0, + 1065.0, + 1405.0, + 1097.0, + 296.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1095.0, + 1405.0, + 1095.0, + 1405.0, + 1127.0, + 296.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1124.0, + 1404.0, + 1124.0, + 1404.0, + 1160.0, + 294.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1156.0, + 1405.0, + 1156.0, + 1405.0, + 1192.0, + 294.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1179.0, + 811.0, + 1179.0, + 811.0, + 1225.0, + 292.0, + 1225.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1372, + 1404, + 1372, + 1404, + 1739, + 297, + 1739 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1045, + 1404, + 1045, + 1404, + 1322, + 298, + 1322 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1403, + 1788, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 348, + 350, + 1351, + 350, + 1351, + 819, + 348, + 819 + ], + "score": 0.974, + "html": "
MethodTop-1Top-5
Motion Segmentation (MS) (Pathak et al., 2016) Exemplar (Ex) (Dosovitskiy et al.,2014) Relative Position (RP) (Doersch et al.,2015) Colorization (Col) (Zhang et al., 2016) Combination of MS + Ex + RP+ Col (Doersch & Zisserman,2017) CPC v1 (van den Oord et al., 2018) Rotation (Kolesnikov et al., 2019)27.6 31.5 36.2 39.6 1 48.7 55.448.3 53.1 59.2 62.5 69.3 73.6
CMC (Tian et al., 2019) Local Aggregation (Zhuang et al., 2019)60.1 60.282.8 1 =
BigBiGAN (Donahue & Simonyan,2019) AMDIM (Bachman et al., 2019) CPC v2 (ours) 65.961.3 68.181.9
" + }, + { + "category_id": 1, + "poly": [ + 300, + 903, + 1402, + 903, + 1402, + 996, + 300, + 996 + ], + "score": 0.97 + }, + { + "category_id": 6, + "poly": [ + 299, + 248, + 1402, + 248, + 1402, + 343, + 299, + 343 + ], + "score": 0.936 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.772 + }, + { + "category_id": 7, + "poly": [ + 369, + 782, + 534, + 782, + 534, + 814, + 369, + 814 + ], + "score": 0.463 + }, + { + "category_id": 13, + "poly": [ + 1023, + 1677, + 1142, + 1677, + 1142, + 1711, + 1023, + 1711 + ], + "score": 0.93, + "latex": "\\{ 2 , \\ldots , 5 \\}" + }, + { + "category_id": 13, + "poly": [ + 695, + 1677, + 765, + 1677, + 765, + 1711, + 695, + 1711 + ], + "score": 0.93, + "latex": "\\{ 2 , 3 \\}" + }, + { + "category_id": 13, + "poly": [ + 463, + 2003, + 539, + 2003, + 539, + 2034, + 463, + 2034 + ], + "score": 0.9, + "latex": "( + 1 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1279, + 1525, + 1360, + 1525, + 1360, + 1554, + 1279, + 1554 + ], + "score": 0.9, + "latex": "+ 1 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 608, + 251, + 644, + 251, + 644, + 283, + 608, + 283 + ], + "score": 0.89, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 813, + 965, + 886, + 965, + 886, + 994, + 813, + 994 + ], + "score": 0.88, + "latex": "+ 1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 1973, + 357, + 1973, + 357, + 2002, + 296, + 2002 + ], + "score": 0.87, + "latex": "+3 \\%" + }, + { + "category_id": 13, + "poly": [ + 913, + 1260, + 972, + 1260, + 972, + 1289, + 913, + 1289 + ], + "score": 0.87, + "latex": "+ 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1233, + 1495, + 1294, + 1495, + 1294, + 1524, + 1233, + 1524 + ], + "score": 0.87, + "latex": "42 \\%" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1291, + 1212, + 1291, + 1212, + 1321, + 1129, + 1321 + ], + "score": 0.87, + "latex": "( + 2 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 811, + 249, + 883, + 249, + 883, + 279, + 811, + 279 + ], + "score": 0.86, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 1112, + 1943, + 1173, + 1943, + 1173, + 1972, + 1112, + 1972 + ], + "score": 0.86, + "latex": "( + 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 306, + 1708, + 366, + 1708, + 366, + 1737, + 306, + 1737 + ], + "score": 0.86, + "latex": "+ 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 493, + 1556, + 547, + 1556, + 547, + 1585, + 493, + 1585 + ], + "score": 0.85, + "latex": "{ \\bf - 1 \\% }" + }, + { + "category_id": 13, + "poly": [ + 551, + 539, + 790, + 539, + 790, + 568, + 551, + 568 + ], + "score": 0.84, + "latex": "\\mathrm { M S } + \\mathrm { E x } + \\mathrm { R P } + \\mathrm { C o l }" + }, + { + "category_id": 13, + "poly": [ + 1339, + 934, + 1399, + 934, + 1399, + 965, + 1339, + 965 + ], + "score": 0.84, + "latex": "+5 \\%" + }, + { + "category_id": 13, + "poly": [ + 640, + 1678, + 659, + 1678, + 659, + 1705, + 640, + 1705 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 517, + 509, + 577, + 509, + 577, + 538, + 517, + 538 + ], + "score": 0.26, + "latex": "\\mathbf { \\tau } ( \\mathbf { C o l } )" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 246.0, + 607.0, + 246.0, + 607.0, + 286.0, + 293.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 246.0, + 810.0, + 246.0, + 810.0, + 286.0, + 645.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 246.0, + 1404.0, + 246.0, + 1404.0, + 286.0, + 884.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 281.0, + 1406.0, + 281.0, + 1406.0, + 315.0, + 295.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 311.0, + 1269.0, + 311.0, + 1269.0, + 345.0, + 295.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 779.0, + 538.0, + 779.0, + 538.0, + 820.0, + 368.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1371.0, + 1405.0, + 1371.0, + 1405.0, + 1408.0, + 294.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1403.0, + 1404.0, + 1403.0, + 1404.0, + 1438.0, + 295.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1430.0, + 1405.0, + 1430.0, + 1405.0, + 1472.0, + 292.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1464.0, + 1405.0, + 1464.0, + 1405.0, + 1498.0, + 294.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1491.0, + 1232.0, + 1491.0, + 1232.0, + 1531.0, + 291.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1491.0, + 1405.0, + 1491.0, + 1405.0, + 1531.0, + 1295.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1524.0, + 1278.0, + 1524.0, + 1278.0, + 1558.0, + 294.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1524.0, + 1405.0, + 1524.0, + 1405.0, + 1558.0, + 1361.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1556.0, + 492.0, + 1556.0, + 492.0, + 1590.0, + 295.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1556.0, + 1406.0, + 1556.0, + 1406.0, + 1590.0, + 548.0, + 1590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1586.0, + 1404.0, + 1586.0, + 1404.0, + 1619.0, + 292.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1618.0, + 1404.0, + 1618.0, + 1404.0, + 1652.0, + 295.0, + 1652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1645.0, + 1405.0, + 1645.0, + 1405.0, + 1683.0, + 292.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1677.0, + 639.0, + 1677.0, + 639.0, + 1715.0, + 292.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1677.0, + 694.0, + 1677.0, + 694.0, + 1715.0, + 660.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1677.0, + 1022.0, + 1677.0, + 1022.0, + 1715.0, + 766.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1677.0, + 1404.0, + 1677.0, + 1404.0, + 1715.0, + 1143.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 305.0, + 1708.0, + 305.0, + 1742.0, + 295.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1708.0, + 613.0, + 1708.0, + 613.0, + 1742.0, + 367.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1047.0, + 1404.0, + 1047.0, + 1404.0, + 1080.0, + 296.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1076.0, + 1406.0, + 1076.0, + 1406.0, + 1113.0, + 293.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1105.0, + 1402.0, + 1105.0, + 1402.0, + 1142.0, + 294.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1138.0, + 1405.0, + 1138.0, + 1405.0, + 1173.0, + 293.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1169.0, + 1405.0, + 1169.0, + 1405.0, + 1202.0, + 294.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 1405.0, + 1198.0, + 1405.0, + 1236.0, + 293.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1229.0, + 1404.0, + 1229.0, + 1404.0, + 1266.0, + 294.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1257.0, + 912.0, + 1257.0, + 912.0, + 1296.0, + 293.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1257.0, + 1405.0, + 1257.0, + 1405.0, + 1296.0, + 973.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1286.0, + 1128.0, + 1286.0, + 1128.0, + 1329.0, + 292.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 1286.0, + 1338.0, + 1286.0, + 1338.0, + 1329.0, + 1213.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1787.0, + 1404.0, + 1787.0, + 1404.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1403.0, + 1820.0, + 1403.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1407.0, + 1850.0, + 1407.0, + 1887.0, + 293.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1879.0, + 1404.0, + 1879.0, + 1404.0, + 1918.0, + 292.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1111.0, + 1941.0, + 1111.0, + 1977.0, + 293.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1977.0, + 1174.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 295.0, + 1970.0, + 295.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1970.0, + 1407.0, + 1970.0, + 1407.0, + 2010.0, + 358.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 462.0, + 2000.0, + 462.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 2000.0, + 553.0, + 2000.0, + 553.0, + 2037.0, + 540.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 902.0, + 1406.0, + 902.0, + 1406.0, + 940.0, + 295.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 933.0, + 1338.0, + 933.0, + 1338.0, + 970.0, + 294.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 964.0, + 812.0, + 964.0, + 812.0, + 1001.0, + 295.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 964.0, + 1168.0, + 964.0, + 1168.0, + 1001.0, + 887.0, + 1001.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 230, + 1404, + 230, + 1404, + 535, + 297, + 535 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 636, + 1404, + 636, + 1404, + 852, + 298, + 852 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 891, + 824, + 891, + 824, + 1651, + 298, + 1651 + ], + "score": 0.979 + }, + { + "category_id": 3, + "poly": [ + 856, + 900, + 1380, + 900, + 1380, + 1448, + 856, + 1448 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 1790, + 1403, + 1790, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 847, + 1462, + 1401, + 1462, + 1401, + 1767, + 847, + 1767 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 298, + 1669, + 821, + 1669, + 821, + 1789, + 298, + 1789 + ], + "score": 0.945 + }, + { + "category_id": 0, + "poly": [ + 303, + 579, + 776, + 579, + 776, + 609, + 303, + 609 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 857, + 76, + 857, + 104, + 299, + 104 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.635 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.253 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 856, + 76, + 856, + 104, + 300, + 104 + ], + "score": 0.155 + }, + { + "category_id": 13, + "poly": [ + 364, + 1103, + 519, + 1103, + 519, + 1136, + 364, + 1136 + ], + "score": 0.92, + "latex": "\\begin{array} { r l r } { z } & { { } = } & { f _ { \\theta } ( { \\pmb x } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 388, + 1469, + 540, + 1469, + 540, + 1499, + 388, + 1499 + ], + "score": 0.9, + "latex": "8 0 . 6 \\% / 9 5 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1530, + 452, + 1530, + 452, + 1559, + 297, + 1559 + ], + "score": 0.89, + "latex": "7 8 . 0 \\% / 9 3 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 382, + 699, + 415, + 699, + 415, + 729, + 382, + 729 + ], + "score": 0.88, + "latex": "\\mathbb { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1086, + 262, + 1140, + 262, + 1140, + 291, + 1086, + 291 + ], + "score": 0.87, + "latex": "17 \\%" + }, + { + "category_id": 13, + "poly": [ + 634, + 699, + 701, + 699, + 701, + 728, + 634, + 728 + ], + "score": 0.87, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 322, + 1378, + 379, + 1378, + 379, + 1407, + 322, + 1407 + ], + "score": 0.87, + "latex": "2 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 904, + 352, + 979, + 352, + 979, + 382, + 904, + 382 + ], + "score": 0.87, + "latex": "6 2 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1081, + 790, + 1156, + 790, + 1156, + 819, + 1081, + 819 + ], + "score": 0.86, + "latex": "4 4 . 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 764, + 821, + 839, + 821, + 839, + 850, + 764, + 850 + ], + "score": 0.86, + "latex": "9 3 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 346, + 352, + 422, + 352, + 422, + 382, + 346, + 382 + ], + "score": 0.86, + "latex": "6 6 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 454, + 1561, + 608, + 1561, + 608, + 1590, + 454, + 1590 + ], + "score": 0.86, + "latex": "7 9 . 9 \\% / 9 5 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 747, + 1317, + 824, + 1317, + 824, + 1346, + 747, + 1346 + ], + "score": 0.85, + "latex": "7 2 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 613, + 1348, + 655, + 1348, + 655, + 1377, + 613, + 1377 + ], + "score": 0.85, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1249, + 475, + 1302, + 475, + 1302, + 504, + 1249, + 504 + ], + "score": 0.85, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 415, + 822, + 455, + 822, + 455, + 849, + 415, + 849 + ], + "score": 0.84, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 557, + 699, + 597, + 699, + 597, + 727, + 557, + 727 + ], + "score": 0.84, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 803, + 262, + 879, + 262, + 879, + 292, + 803, + 292 + ], + "score": 0.83, + "latex": "6 5 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 664, + 1378, + 719, + 1378, + 719, + 1407, + 664, + 1407 + ], + "score": 0.83, + "latex": "65 \\%" + }, + { + "category_id": 13, + "poly": [ + 1194, + 1420, + 1257, + 1420, + 1257, + 1448, + 1194, + 1448 + ], + "score": 0.78, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 428, + 1139, + 451, + 1139, + 451, + 1162, + 428, + 1162 + ], + "score": 0.74, + "latex": "_ { \\textbf { \\em x } }" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 896.0, + 952.0, + 896.0, + 952.0, + 951.0, + 871.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 925.0, + 1249.0, + 925.0, + 1249.0, + 948.0, + 1222.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1341.0, + 949.0, + 1362.0, + 949.0, + 1362.0, + 964.0, + 1341.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1004.0, + 943.0, + 1004.0, + 943.0, + 1048.0, + 876.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 995.0, + 1232.0, + 995.0, + 1232.0, + 1019.0, + 1206.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1046.0, + 1060.0, + 1046.0, + 1060.0, + 1071.0, + 1033.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1067.0, + 1137.0, + 1067.0, + 1137.0, + 1090.0, + 1111.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1081.0, + 1157.0, + 1081.0, + 1157.0, + 1101.0, + 1134.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1104.0, + 941.0, + 1104.0, + 941.0, + 1145.0, + 876.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1288.0, + 1157.0, + 1310.0, + 1157.0, + 1310.0, + 1178.0, + 1288.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1180.0, + 1056.0, + 1180.0, + 1056.0, + 1205.0, + 1031.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1201.0, + 946.0, + 1201.0, + 946.0, + 1247.0, + 887.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1223.0, + 1147.0, + 1223.0, + 1147.0, + 1247.0, + 1122.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1245.0, + 1257.0, + 1245.0, + 1257.0, + 1260.0, + 1241.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1284.0, + 1366.0, + 1284.0, + 1366.0, + 1329.0, + 1154.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1313.0, + 1374.0, + 1313.0, + 1374.0, + 1354.0, + 1155.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1347.0, + 981.0, + 1347.0, + 981.0, + 1372.0, + 953.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1337.0, + 1376.0, + 1337.0, + 1376.0, + 1382.0, + 1154.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1381.0, + 990.0, + 1381.0, + 990.0, + 1423.0, + 923.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1382.0, + 1072.0, + 1382.0, + 1072.0, + 1422.0, + 1016.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1383.0, + 1165.0, + 1383.0, + 1165.0, + 1421.0, + 1098.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1382.0, + 1252.0, + 1382.0, + 1252.0, + 1421.0, + 1187.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1378.0, + 1343.0, + 1378.0, + 1343.0, + 1424.0, + 1272.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1411.0, + 1193.0, + 1411.0, + 1193.0, + 1455.0, + 955.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 1411.0, + 1359.0, + 1411.0, + 1359.0, + 1455.0, + 1258.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1462.0, + 1404.0, + 1462.0, + 1404.0, + 1494.0, + 845.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1492.0, + 1404.0, + 1492.0, + 1404.0, + 1524.0, + 844.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1523.0, + 1403.0, + 1523.0, + 1403.0, + 1553.0, + 845.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1554.0, + 1403.0, + 1554.0, + 1403.0, + 1584.0, + 845.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1616.0, + 844.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1614.0, + 1405.0, + 1614.0, + 1405.0, + 1648.0, + 844.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1645.0, + 1406.0, + 1645.0, + 1406.0, + 1679.0, + 844.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1675.0, + 1405.0, + 1675.0, + 1405.0, + 1708.0, + 843.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1739.0, + 843.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 1737.0, + 1316.0, + 1737.0, + 1316.0, + 1769.0, + 844.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 578.0, + 782.0, + 578.0, + 782.0, + 612.0, + 297.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 802.0, + 260.0, + 802.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 260.0, + 1085.0, + 260.0, + 1085.0, + 298.0, + 880.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 298.0, + 1141.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 288.0, + 1405.0, + 288.0, + 1405.0, + 329.0, + 291.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 319.0, + 1404.0, + 319.0, + 1404.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 352.0, + 345.0, + 352.0, + 345.0, + 388.0, + 295.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 352.0, + 903.0, + 352.0, + 903.0, + 388.0, + 423.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 388.0, + 980.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 382.0, + 1404.0, + 382.0, + 1404.0, + 418.0, + 295.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 412.0, + 1405.0, + 412.0, + 1405.0, + 450.0, + 294.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 443.0, + 1404.0, + 443.0, + 1404.0, + 479.0, + 295.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 472.0, + 1248.0, + 472.0, + 1248.0, + 510.0, + 292.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 472.0, + 1407.0, + 472.0, + 1407.0, + 510.0, + 1303.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 503.0, + 947.0, + 503.0, + 947.0, + 538.0, + 295.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 638.0, + 1404.0, + 638.0, + 1404.0, + 672.0, + 294.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 666.0, + 1406.0, + 666.0, + 1406.0, + 704.0, + 294.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 698.0, + 381.0, + 698.0, + 381.0, + 733.0, + 294.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 698.0, + 556.0, + 698.0, + 556.0, + 733.0, + 416.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 698.0, + 633.0, + 698.0, + 633.0, + 733.0, + 598.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 698.0, + 1405.0, + 698.0, + 1405.0, + 733.0, + 702.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 727.0, + 1406.0, + 727.0, + 1406.0, + 764.0, + 293.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 758.0, + 1405.0, + 758.0, + 1405.0, + 795.0, + 293.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 790.0, + 1080.0, + 790.0, + 1080.0, + 824.0, + 296.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 790.0, + 1405.0, + 790.0, + 1405.0, + 824.0, + 1157.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 821.0, + 414.0, + 821.0, + 414.0, + 855.0, + 294.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 821.0, + 763.0, + 821.0, + 763.0, + 855.0, + 456.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 821.0, + 1404.0, + 821.0, + 1404.0, + 855.0, + 840.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 889.0, + 825.0, + 889.0, + 825.0, + 924.0, + 298.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 920.0, + 826.0, + 920.0, + 826.0, + 953.0, + 296.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 950.0, + 826.0, + 950.0, + 826.0, + 984.0, + 296.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 981.0, + 827.0, + 981.0, + 827.0, + 1016.0, + 295.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1009.0, + 827.0, + 1009.0, + 827.0, + 1048.0, + 293.0, + 1048.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1040.0, + 827.0, + 1040.0, + 827.0, + 1077.0, + 294.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1073.0, + 826.0, + 1073.0, + 826.0, + 1107.0, + 295.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1103.0, + 363.0, + 1103.0, + 363.0, + 1138.0, + 295.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1103.0, + 826.0, + 1103.0, + 826.0, + 1138.0, + 520.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1134.0, + 427.0, + 1134.0, + 427.0, + 1167.0, + 293.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 1134.0, + 828.0, + 1134.0, + 828.0, + 1167.0, + 452.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1163.0, + 827.0, + 1163.0, + 827.0, + 1198.0, + 295.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1193.0, + 827.0, + 1193.0, + 827.0, + 1226.0, + 295.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1225.0, + 827.0, + 1225.0, + 827.0, + 1259.0, + 295.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1255.0, + 826.0, + 1255.0, + 826.0, + 1289.0, + 294.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1287.0, + 827.0, + 1287.0, + 827.0, + 1321.0, + 294.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1315.0, + 746.0, + 1315.0, + 746.0, + 1350.0, + 295.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1347.0, + 612.0, + 1347.0, + 612.0, + 1380.0, + 294.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1347.0, + 826.0, + 1347.0, + 826.0, + 1380.0, + 656.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1376.0, + 321.0, + 1376.0, + 321.0, + 1409.0, + 294.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1376.0, + 663.0, + 1376.0, + 663.0, + 1409.0, + 380.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1376.0, + 825.0, + 1376.0, + 825.0, + 1409.0, + 720.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1408.0, + 826.0, + 1408.0, + 826.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1438.0, + 826.0, + 1438.0, + 826.0, + 1469.0, + 296.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 387.0, + 1467.0, + 387.0, + 1502.0, + 294.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1467.0, + 826.0, + 1467.0, + 826.0, + 1502.0, + 541.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1501.0, + 826.0, + 1501.0, + 826.0, + 1530.0, + 293.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 1527.0, + 826.0, + 1527.0, + 826.0, + 1563.0, + 453.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1558.0, + 453.0, + 1558.0, + 453.0, + 1595.0, + 295.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 1558.0, + 826.0, + 1558.0, + 826.0, + 1595.0, + 609.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1587.0, + 826.0, + 1587.0, + 826.0, + 1624.0, + 295.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1618.0, + 603.0, + 1618.0, + 603.0, + 1655.0, + 294.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1790.0, + 1405.0, + 1790.0, + 1405.0, + 1824.0, + 294.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1822.0, + 1404.0, + 1822.0, + 1404.0, + 1852.0, + 296.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1945.0, + 292.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1975.0, + 293.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1669.0, + 824.0, + 1669.0, + 824.0, + 1700.0, + 296.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 826.0, + 1698.0, + 826.0, + 1731.0, + 294.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 825.0, + 1728.0, + 825.0, + 1761.0, + 294.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1759.0, + 826.0, + 1759.0, + 826.0, + 1791.0, + 293.0, + 1791.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1389, + 1404, + 1389, + 1404, + 1754, + 297, + 1754 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1769, + 1403, + 1769, + 1403, + 1956, + 297, + 1956 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1170, + 1404, + 1170, + 1404, + 1355, + 298, + 1355 + ], + "score": 0.967 + }, + { + "category_id": 2, + "poly": [ + 298, + 1978, + 1404, + 1978, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.937 + }, + { + "category_id": 5, + "poly": [ + 427, + 504, + 1268, + 504, + 1268, + 1073, + 427, + 1073 + ], + "score": 0.91, + "html": "
Labeled data 1% 10% 100% Method Top-5 accuracy
Supervised baseline 44.1 82.1 93.9 Methods using label-propagation:
Pseudolabeling (Zhai et al., 2019) 51.6 82.4 VAT + Entropy Minimization (Zhai et al., 2019) 47.0 83.4 Unsup. Data Augmentation (Xie et al., 2019) 88.5
Rotation+ VAT+Ent.Min. (Zhai et al.,2019) 91.2 95.0 =
Methods using representation learning only:
Instance Discrimination (Wu et al., 2018) 39.2 77.4
Rotation (Zhai et al., 2019) 57.5 86.4
+ResNet trained on BigBiGAN (fixed) 55.2 78.8 87.0
+ResNet trained on AMDIM (fixed) 67.4 85.8 92.2
+ResNet trained on CPC v2 (fixed) 72.3 89.1 94.4 +ResNet trained on CPC v2 (fine-tuned) 72.9 89.5 95.2
" + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.8 + }, + { + "category_id": 6, + "poly": [ + 297, + 247, + 1404, + 247, + 1404, + 494, + 297, + 494 + ], + "score": 0.62 + }, + { + "category_id": 1, + "poly": [ + 297, + 247, + 1404, + 247, + 1404, + 494, + 297, + 494 + ], + "score": 0.578 + }, + { + "category_id": 13, + "poly": [ + 1140, + 1893, + 1271, + 1893, + 1271, + 1923, + 1140, + 1923 + ], + "score": 0.9, + "latex": "7 \\times 7 \\times 2 5 6 0" + }, + { + "category_id": 13, + "poly": [ + 637, + 1893, + 767, + 1893, + 767, + 1922, + 637, + 1922 + ], + "score": 0.9, + "latex": "7 \\times 7 \\times 8 1 9 2" + }, + { + "category_id": 13, + "poly": [ + 603, + 251, + 638, + 251, + 638, + 283, + 603, + 283 + ], + "score": 0.89, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 776, + 281, + 830, + 281, + 830, + 310, + 776, + 310 + ], + "score": 0.88, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 523, + 1802, + 599, + 1802, + 599, + 1831, + 523, + 1831 + ], + "score": 0.86, + "latex": "6 8 . 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1326, + 1693, + 1401, + 1693, + 1401, + 1723, + 1326, + 1723 + ], + "score": 0.85, + "latex": "6 5 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1694, + 372, + 1694, + 372, + 1723, + 297, + 1723 + ], + "score": 0.85, + "latex": "5 7 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 798, + 250, + 843, + 250, + 843, + 278, + 798, + 278 + ], + "score": 0.85, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 957, + 1663, + 998, + 1663, + 998, + 1692, + 957, + 1692 + ], + "score": 0.85, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 883, + 250, + 941, + 250, + 941, + 279, + 883, + 279 + ], + "score": 0.85, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 697, + 281, + 738, + 281, + 738, + 310, + 697, + 310 + ], + "score": 0.84, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 392, + 1802, + 468, + 1802, + 468, + 1832, + 392, + 1832 + ], + "score": 0.83, + "latex": "6 1 . 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 1218, + 1694, + 1292, + 1694, + 1292, + 1723, + 1218, + 1723 + ], + "score": 0.82, + "latex": "5 5 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1145, + 1200, + 1386, + 1200, + 1386, + 1234, + 1145, + 1234 + ], + "score": 0.78, + "latex": "R ^ { 2 } = 0 . 1 7 , p = 0 . 3 6" + }, + { + "category_id": 13, + "poly": [ + 1042, + 403, + 1060, + 403, + 1060, + 435, + 1042, + 435 + ], + "score": 0.76, + "latex": "\\dagger" + }, + { + "category_id": 13, + "poly": [ + 505, + 725, + 528, + 725, + 528, + 750, + 505, + 750 + ], + "score": 0.71, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 550, + 785, + 573, + 785, + 573, + 810, + 550, + 810 + ], + "score": 0.36, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 550, + 783, + 689, + 783, + 689, + 812, + 550, + 812 + ], + "score": 0.34, + "latex": "+ \\mathrm { V A T } + \\mathrm { E n }" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2011.0, + 330.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 884.0, + 2005.0, + 884.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 249.0, + 602.0, + 249.0, + 602.0, + 283.0, + 295.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 249.0, + 797.0, + 249.0, + 797.0, + 283.0, + 639.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 249.0, + 882.0, + 249.0, + 882.0, + 283.0, + 844.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 249.0, + 1404.0, + 249.0, + 1404.0, + 283.0, + 942.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 281.0, + 696.0, + 281.0, + 696.0, + 315.0, + 295.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 281.0, + 775.0, + 281.0, + 775.0, + 315.0, + 739.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 281.0, + 1404.0, + 281.0, + 1404.0, + 315.0, + 831.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 308.0, + 1404.0, + 308.0, + 1404.0, + 348.0, + 294.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 342.0, + 1405.0, + 342.0, + 1405.0, + 375.0, + 296.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 372.0, + 1404.0, + 372.0, + 1404.0, + 405.0, + 295.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 402.0, + 1041.0, + 402.0, + 1041.0, + 436.0, + 295.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 402.0, + 1405.0, + 402.0, + 1405.0, + 436.0, + 1061.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 433.0, + 1405.0, + 433.0, + 1405.0, + 467.0, + 294.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 578.0, + 461.0, + 578.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1390.0, + 1405.0, + 1390.0, + 1405.0, + 1424.0, + 296.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1419.0, + 1406.0, + 1419.0, + 1406.0, + 1455.0, + 291.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1450.0, + 1402.0, + 1450.0, + 1402.0, + 1484.0, + 294.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1481.0, + 1404.0, + 1481.0, + 1404.0, + 1515.0, + 295.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1511.0, + 1405.0, + 1511.0, + 1405.0, + 1545.0, + 295.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1539.0, + 1405.0, + 1539.0, + 1405.0, + 1578.0, + 292.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1570.0, + 1405.0, + 1570.0, + 1405.0, + 1607.0, + 292.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1638.0, + 294.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 1404.0, + 1633.0, + 1404.0, + 1667.0, + 294.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1661.0, + 956.0, + 1661.0, + 956.0, + 1695.0, + 295.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 1661.0, + 1402.0, + 1661.0, + 1402.0, + 1695.0, + 999.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1690.0, + 296.0, + 1690.0, + 296.0, + 1730.0, + 291.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1690.0, + 1217.0, + 1690.0, + 1217.0, + 1730.0, + 373.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 1690.0, + 1325.0, + 1690.0, + 1325.0, + 1730.0, + 1293.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1690.0, + 1405.0, + 1690.0, + 1405.0, + 1730.0, + 1402.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1722.0, + 423.0, + 1722.0, + 423.0, + 1758.0, + 294.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 1405.0, + 1768.0, + 1405.0, + 1806.0, + 294.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1800.0, + 391.0, + 1800.0, + 391.0, + 1837.0, + 292.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1800.0, + 522.0, + 1800.0, + 522.0, + 1837.0, + 469.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1800.0, + 1407.0, + 1800.0, + 1407.0, + 1837.0, + 600.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1829.0, + 1407.0, + 1829.0, + 1407.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1863.0, + 1408.0, + 1863.0, + 1408.0, + 1898.0, + 295.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1889.0, + 636.0, + 1889.0, + 636.0, + 1928.0, + 292.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1889.0, + 1139.0, + 1889.0, + 1139.0, + 1928.0, + 768.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1889.0, + 1405.0, + 1889.0, + 1405.0, + 1928.0, + 1272.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1923.0, + 1405.0, + 1923.0, + 1405.0, + 1959.0, + 293.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1171.0, + 1406.0, + 1171.0, + 1406.0, + 1206.0, + 292.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1200.0, + 1144.0, + 1200.0, + 1144.0, + 1237.0, + 292.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1387.0, + 1200.0, + 1405.0, + 1200.0, + 1405.0, + 1237.0, + 1387.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1232.0, + 1402.0, + 1232.0, + 1402.0, + 1268.0, + 294.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1260.0, + 1404.0, + 1260.0, + 1404.0, + 1298.0, + 293.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1290.0, + 1406.0, + 1290.0, + 1406.0, + 1331.0, + 291.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1321.0, + 628.0, + 1321.0, + 628.0, + 1359.0, + 295.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 249.0, + 602.0, + 249.0, + 602.0, + 283.0, + 295.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 249.0, + 797.0, + 249.0, + 797.0, + 283.0, + 639.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 249.0, + 882.0, + 249.0, + 882.0, + 283.0, + 844.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 249.0, + 1404.0, + 249.0, + 1404.0, + 283.0, + 942.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 281.0, + 696.0, + 281.0, + 696.0, + 315.0, + 295.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 281.0, + 775.0, + 281.0, + 775.0, + 315.0, + 739.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 281.0, + 1404.0, + 281.0, + 1404.0, + 315.0, + 831.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 308.0, + 1404.0, + 308.0, + 1404.0, + 348.0, + 294.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 342.0, + 1405.0, + 342.0, + 1405.0, + 375.0, + 296.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 372.0, + 1404.0, + 372.0, + 1404.0, + 405.0, + 295.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 402.0, + 1041.0, + 402.0, + 1041.0, + 436.0, + 295.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 402.0, + 1405.0, + 402.0, + 1405.0, + 436.0, + 1061.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 433.0, + 1405.0, + 433.0, + 1405.0, + 467.0, + 294.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 578.0, + 461.0, + 578.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 639, + 1402, + 639, + 1402, + 1004, + 299, + 1004 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 1113, + 1404, + 1113, + 1404, + 1631, + 298, + 1631 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1759, + 1403, + 1759, + 1403, + 2034, + 297, + 2034 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 230, + 1403, + 230, + 1403, + 596, + 298, + 596 + ], + "score": 0.982 + }, + { + "category_id": 0, + "poly": [ + 300, + 1684, + 529, + 1684, + 529, + 1719, + 300, + 1719 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 300, + 1052, + 1159, + 1052, + 1159, + 1083, + 300, + 1083 + ], + "score": 0.821 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 857, + 76, + 857, + 104, + 299, + 104 + ], + "score": 0.819 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.79 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 856, + 76, + 856, + 104, + 300, + 104 + ], + "score": 0.316 + }, + { + "category_id": 13, + "poly": [ + 404, + 1235, + 439, + 1235, + 439, + 1268, + 404, + 1268 + ], + "score": 0.9, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 584, + 231, + 660, + 231, + 660, + 260, + 584, + 260 + ], + "score": 0.89, + "latex": "7 2 . 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 1144, + 1540, + 1172, + 1540, + 1172, + 1570, + 1144, + 1570 + ], + "score": 0.89, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 791, + 383, + 867, + 383, + 867, + 412, + 791, + 412 + ], + "score": 0.88, + "latex": "5 5 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 475, + 373, + 475, + 373, + 504, + 297, + 504 + ], + "score": 0.88, + "latex": "6 7 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 905, + 853, + 982, + 853, + 982, + 883, + 905, + 883 + ], + "score": 0.88, + "latex": "8 9 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 986, + 414, + 1044, + 414, + 1044, + 444, + 986, + 444 + ], + "score": 0.87, + "latex": "( + 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 298, + 414, + 389, + 414, + 389, + 445, + 298, + 445 + ], + "score": 0.87, + "latex": "( 5 7 . 5 \\% )" + }, + { + "category_id": 13, + "poly": [ + 538, + 853, + 592, + 853, + 592, + 882, + 538, + 882 + ], + "score": 0.87, + "latex": "10 \\%" + }, + { + "category_id": 13, + "poly": [ + 833, + 1479, + 909, + 1479, + 909, + 1508, + 833, + 1508 + ], + "score": 0.87, + "latex": "7 0 . 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 1570, + 372, + 1570, + 372, + 1599, + 296, + 1599 + ], + "score": 0.87, + "latex": "7 2 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1448, + 373, + 1448, + 373, + 1478, + 297, + 1478 + ], + "score": 0.86, + "latex": "7 0 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 403, + 1510, + 431, + 1510, + 431, + 1540, + 403, + 1540 + ], + "score": 0.86, + "latex": "f _ { \\theta }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1387, + 352, + 1387, + 352, + 1416, + 296, + 1416 + ], + "score": 0.86, + "latex": "65 \\%" + }, + { + "category_id": 13, + "poly": [ + 1326, + 822, + 1401, + 822, + 1401, + 852, + 1326, + 852 + ], + "score": 0.86, + "latex": "8 2 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 905, + 231, + 945, + 231, + 945, + 259, + 905, + 259 + ], + "score": 0.86, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 326, + 1418, + 401, + 1418, + 401, + 1447, + 326, + 1447 + ], + "score": 0.86, + "latex": "6 7 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1105, + 823, + 1179, + 823, + 1179, + 852, + 1105, + 852 + ], + "score": 0.86, + "latex": "8 5 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 736, + 1570, + 777, + 1570, + 777, + 1599, + 736, + 1599 + ], + "score": 0.84, + "latex": "2 \\%" + }, + { + "category_id": 13, + "poly": [ + 364, + 231, + 439, + 231, + 439, + 261, + 364, + 261 + ], + "score": 0.84, + "latex": "7 2 . 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1143, + 353, + 1184, + 353, + 1184, + 382, + 1143, + 382 + ], + "score": 0.84, + "latex": "1 \\%" + }, + { + "category_id": 13, + "poly": [ + 1191, + 823, + 1265, + 823, + 1265, + 852, + 1191, + 852 + ], + "score": 0.83, + "latex": "8 3 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1266, + 884, + 1343, + 884, + 1343, + 914, + 1266, + 914 + ], + "score": 0.8, + "latex": "9 1 . 2 \\%" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1683.0, + 535.0, + 1683.0, + 535.0, + 1726.0, + 292.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1050.0, + 1165.0, + 1050.0, + 1165.0, + 1088.0, + 294.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 639.0, + 1407.0, + 639.0, + 1407.0, + 676.0, + 294.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 671.0, + 1406.0, + 671.0, + 1406.0, + 704.0, + 293.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 702.0, + 1402.0, + 702.0, + 1402.0, + 736.0, + 295.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 731.0, + 1402.0, + 731.0, + 1402.0, + 765.0, + 295.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 759.0, + 1404.0, + 759.0, + 1404.0, + 799.0, + 294.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 790.0, + 1406.0, + 790.0, + 1406.0, + 826.0, + 294.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 823.0, + 1104.0, + 823.0, + 1104.0, + 857.0, + 295.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 823.0, + 1190.0, + 823.0, + 1190.0, + 857.0, + 1180.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 823.0, + 1325.0, + 823.0, + 1325.0, + 857.0, + 1266.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 851.0, + 537.0, + 851.0, + 537.0, + 889.0, + 293.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 851.0, + 904.0, + 851.0, + 904.0, + 889.0, + 593.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 851.0, + 1406.0, + 851.0, + 1406.0, + 889.0, + 983.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 883.0, + 1265.0, + 883.0, + 1265.0, + 921.0, + 293.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 883.0, + 1406.0, + 883.0, + 1406.0, + 921.0, + 1344.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 915.0, + 1406.0, + 915.0, + 1406.0, + 949.0, + 294.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 944.0, + 1407.0, + 944.0, + 1407.0, + 981.0, + 293.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 976.0, + 721.0, + 976.0, + 721.0, + 1007.0, + 297.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1112.0, + 1405.0, + 1112.0, + 1405.0, + 1148.0, + 296.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1143.0, + 1406.0, + 1143.0, + 1406.0, + 1179.0, + 293.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1171.0, + 1406.0, + 1171.0, + 1406.0, + 1210.0, + 292.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1206.0, + 1402.0, + 1206.0, + 1402.0, + 1238.0, + 294.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1233.0, + 403.0, + 1233.0, + 403.0, + 1269.0, + 294.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1233.0, + 1405.0, + 1233.0, + 1405.0, + 1269.0, + 440.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1264.0, + 1405.0, + 1264.0, + 1405.0, + 1299.0, + 293.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1295.0, + 1406.0, + 1295.0, + 1406.0, + 1331.0, + 292.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1324.0, + 1406.0, + 1324.0, + 1406.0, + 1360.0, + 293.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1355.0, + 1408.0, + 1355.0, + 1408.0, + 1395.0, + 291.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1384.0, + 295.0, + 1384.0, + 295.0, + 1426.0, + 291.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1384.0, + 1406.0, + 1384.0, + 1406.0, + 1426.0, + 353.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1416.0, + 325.0, + 1416.0, + 325.0, + 1452.0, + 293.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1416.0, + 1405.0, + 1416.0, + 1405.0, + 1452.0, + 402.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1444.0, + 296.0, + 1444.0, + 296.0, + 1485.0, + 293.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 1444.0, + 1406.0, + 1444.0, + 1406.0, + 1485.0, + 374.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1479.0, + 832.0, + 1479.0, + 832.0, + 1514.0, + 293.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1479.0, + 1406.0, + 1479.0, + 1406.0, + 1514.0, + 910.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1507.0, + 402.0, + 1507.0, + 402.0, + 1543.0, + 293.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1507.0, + 1405.0, + 1507.0, + 1405.0, + 1543.0, + 432.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1538.0, + 1143.0, + 1538.0, + 1143.0, + 1574.0, + 294.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1173.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1574.0, + 1173.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1566.0, + 735.0, + 1566.0, + 735.0, + 1605.0, + 373.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1566.0, + 1405.0, + 1566.0, + 1405.0, + 1605.0, + 778.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1600.0, + 1365.0, + 1600.0, + 1365.0, + 1636.0, + 294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1757.0, + 1405.0, + 1757.0, + 1405.0, + 1795.0, + 295.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1790.0, + 1405.0, + 1790.0, + 1405.0, + 1824.0, + 293.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1856.0, + 292.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1916.0, + 292.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1403.0, + 1911.0, + 1403.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1407.0, + 1969.0, + 1407.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 947.0, + 2001.0, + 947.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 363.0, + 231.0, + 363.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 231.0, + 583.0, + 231.0, + 583.0, + 265.0, + 440.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 231.0, + 904.0, + 231.0, + 904.0, + 265.0, + 661.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 231.0, + 1403.0, + 231.0, + 1403.0, + 265.0, + 946.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 1403.0, + 289.0, + 1403.0, + 328.0, + 293.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 320.0, + 1407.0, + 320.0, + 1407.0, + 358.0, + 292.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 351.0, + 1142.0, + 351.0, + 1142.0, + 388.0, + 293.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 351.0, + 1405.0, + 351.0, + 1405.0, + 388.0, + 1185.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 381.0, + 790.0, + 381.0, + 790.0, + 418.0, + 293.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 381.0, + 1405.0, + 381.0, + 1405.0, + 418.0, + 868.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 414.0, + 985.0, + 414.0, + 985.0, + 448.0, + 390.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 414.0, + 1403.0, + 414.0, + 1403.0, + 448.0, + 1045.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 443.0, + 1405.0, + 443.0, + 1405.0, + 477.0, + 294.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 474.0, + 296.0, + 474.0, + 296.0, + 509.0, + 293.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 474.0, + 1405.0, + 474.0, + 1405.0, + 509.0, + 374.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 1405.0, + 505.0, + 1405.0, + 539.0, + 294.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 536.0, + 1405.0, + 536.0, + 1405.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 566.0, + 1211.0, + 566.0, + 1211.0, + 600.0, + 294.0, + 600.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1072, + 1405, + 1072, + 1405, + 1498, + 297, + 1498 + ], + "score": 0.982 + }, + { + "category_id": 0, + "poly": [ + 299, + 1546, + 488, + 1546, + 488, + 1578, + 299, + 1578 + ], + "score": 0.863 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 293, + 1596, + 1352, + 1596, + 1352, + 1628, + 293, + 1628 + ], + "score": 0.734 + }, + { + "category_id": 6, + "poly": [ + 296, + 248, + 1405, + 248, + 1405, + 463, + 296, + 463 + ], + "score": 0.606 + }, + { + "category_id": 1, + "poly": [ + 296, + 248, + 1405, + 248, + 1405, + 463, + 296, + 463 + ], + "score": 0.584 + }, + { + "category_id": 1, + "poly": [ + 296, + 1646, + 1401, + 1646, + 1401, + 1710, + 296, + 1710 + ], + "score": 0.564 + }, + { + "category_id": 1, + "poly": [ + 299, + 1971, + 1399, + 1971, + 1399, + 2033, + 299, + 2033 + ], + "score": 0.561 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.56 + }, + { + "category_id": 5, + "poly": [ + 409, + 474, + 1289, + 474, + 1289, + 1006, + 409, + 1006 + ], + "score": 0.527, + "html": "
MethodmAP
Transfer from labeled data: Supervised - ResNet-15274.7
Transfer from unlabeled data: Exemplar (Ex) (Dosovitskiy et al., 2014) Motion Segmentation (MS) (Pathak et al.,2016)60.9
Colorization (Col) (Zhang et al.,2016) Relative Position (RP) (Doersch et al.,2015)61.1 65.5
Combination of Ex + MS + Col +RP (Doersch & Zisserman,2017)66.8 70.5
Instance Discrimination (Wu et al., 2018)65.4
Deep Cluster (Caron et al., 2018) Deeper Cluster (Caron et al., 2019)65.9
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1890, + 1400, + 1890, + 1400, + 1953, + 297, + 1953 + ], + "score": 0.497 + }, + { + "category_id": 1, + "poly": [ + 297, + 1727, + 1401, + 1727, + 1401, + 1791, + 297, + 1791 + ], + "score": 0.493 + }, + { + "category_id": 1, + "poly": [ + 296, + 1808, + 1400, + 1808, + 1400, + 1872, + 296, + 1872 + ], + "score": 0.489 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.345 + }, + { + "category_id": 13, + "poly": [ + 607, + 783, + 843, + 783, + 843, + 812, + 607, + 812 + ], + "score": 0.9, + "latex": "\\operatorname { E x } + \\mathbf { M S } + \\mathbf { C o l } + \\mathbf { R P }" + }, + { + "category_id": 13, + "poly": [ + 564, + 250, + 600, + 250, + 600, + 283, + 564, + 283 + ], + "score": 0.88, + "latex": "h _ { \\psi }" + }, + { + "category_id": 13, + "poly": [ + 755, + 249, + 827, + 249, + 827, + 279, + 755, + 279 + ], + "score": 0.86, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 1292, + 403, + 1311, + 403, + 1311, + 435, + 1292, + 435 + ], + "score": 0.73, + "latex": "\\dagger" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1543.0, + 491.0, + 1543.0, + 491.0, + 1583.0, + 296.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 248.0, + 563.0, + 248.0, + 563.0, + 283.0, + 295.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 248.0, + 754.0, + 248.0, + 754.0, + 283.0, + 601.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 248.0, + 1403.0, + 248.0, + 1403.0, + 283.0, + 828.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 280.0, + 1405.0, + 280.0, + 1405.0, + 314.0, + 293.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 310.0, + 1405.0, + 310.0, + 1405.0, + 344.0, + 294.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 343.0, + 1403.0, + 343.0, + 1403.0, + 374.0, + 295.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 371.0, + 1405.0, + 371.0, + 1405.0, + 406.0, + 294.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1291.0, + 400.0, + 1291.0, + 437.0, + 294.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 437.0, + 1312.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 434.0, + 694.0, + 434.0, + 694.0, + 465.0, + 296.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1070.0, + 1405.0, + 1070.0, + 1405.0, + 1108.0, + 294.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1105.0, + 1403.0, + 1105.0, + 1403.0, + 1137.0, + 295.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 1405.0, + 1131.0, + 1405.0, + 1169.0, + 294.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1165.0, + 1403.0, + 1165.0, + 1403.0, + 1196.0, + 296.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1196.0, + 1403.0, + 1196.0, + 1403.0, + 1228.0, + 295.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1222.0, + 1403.0, + 1222.0, + 1403.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1252.0, + 1406.0, + 1252.0, + 1406.0, + 1293.0, + 291.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1285.0, + 1406.0, + 1285.0, + 1406.0, + 1320.0, + 291.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1352.0, + 295.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1347.0, + 1405.0, + 1347.0, + 1405.0, + 1382.0, + 295.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1377.0, + 1405.0, + 1377.0, + 1405.0, + 1413.0, + 295.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1404.0, + 1406.0, + 1404.0, + 1406.0, + 1446.0, + 292.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1438.0, + 1406.0, + 1438.0, + 1406.0, + 1473.0, + 295.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 659.0, + 1467.0, + 659.0, + 1504.0, + 294.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1591.0, + 1355.0, + 1591.0, + 1355.0, + 1634.0, + 294.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 248.0, + 563.0, + 248.0, + 563.0, + 283.0, + 295.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 248.0, + 754.0, + 248.0, + 754.0, + 283.0, + 601.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 248.0, + 1403.0, + 248.0, + 1403.0, + 283.0, + 828.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 280.0, + 1405.0, + 280.0, + 1405.0, + 314.0, + 293.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 310.0, + 1405.0, + 310.0, + 1405.0, + 344.0, + 294.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 343.0, + 1403.0, + 343.0, + 1403.0, + 374.0, + 295.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 371.0, + 1405.0, + 371.0, + 1405.0, + 406.0, + 294.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 1291.0, + 400.0, + 1291.0, + 437.0, + 294.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 400.0, + 1405.0, + 400.0, + 1405.0, + 437.0, + 1312.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 434.0, + 694.0, + 434.0, + 694.0, + 465.0, + 296.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1646.0, + 1405.0, + 1646.0, + 1405.0, + 1682.0, + 295.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1677.0, + 1069.0, + 1677.0, + 1069.0, + 1713.0, + 323.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 2002.0, + 483.0, + 2002.0, + 483.0, + 2035.0, + 324.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 1404.0, + 1888.0, + 1404.0, + 1928.0, + 293.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1920.0, + 1153.0, + 1920.0, + 1153.0, + 1956.0, + 323.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1724.0, + 1405.0, + 1724.0, + 1405.0, + 1765.0, + 292.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1758.0, + 1013.0, + 1758.0, + 1013.0, + 1792.0, + 323.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1805.0, + 1405.0, + 1805.0, + 1405.0, + 1845.0, + 292.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1833.0, + 400.0, + 1833.0, + 400.0, + 1876.0, + 319.0, + 1876.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1506, + 1403, + 1506, + 1403, + 1573, + 295, + 1573 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 296, + 228, + 1400, + 228, + 1400, + 323, + 296, + 323 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.875 + }, + { + "category_id": 1, + "poly": [ + 293, + 809, + 1402, + 809, + 1402, + 873, + 293, + 873 + ], + "score": 0.861 + }, + { + "category_id": 1, + "poly": [ + 300, + 1272, + 1403, + 1272, + 1403, + 1369, + 300, + 1369 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 291, + 722, + 1402, + 722, + 1402, + 788, + 291, + 788 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 298, + 1389, + 1403, + 1389, + 1403, + 1484, + 298, + 1484 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 298, + 518, + 1401, + 518, + 1401, + 584, + 298, + 584 + ], + "score": 0.847 + }, + { + "category_id": 1, + "poly": [ + 296, + 432, + 1399, + 432, + 1399, + 497, + 296, + 497 + ], + "score": 0.844 + }, + { + "category_id": 1, + "poly": [ + 296, + 346, + 1397, + 346, + 1397, + 411, + 296, + 411 + ], + "score": 0.842 + }, + { + "category_id": 1, + "poly": [ + 293, + 1710, + 1404, + 1710, + 1404, + 1776, + 293, + 1776 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 294, + 1187, + 1402, + 1187, + 1402, + 1252, + 294, + 1252 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 297, + 1592, + 1405, + 1592, + 1405, + 1687, + 297, + 1687 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 296, + 605, + 1407, + 605, + 1407, + 700, + 296, + 700 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 303, + 1012, + 1400, + 1012, + 1400, + 1078, + 303, + 1078 + ], + "score": 0.829 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.825 + }, + { + "category_id": 1, + "poly": [ + 299, + 1099, + 1402, + 1099, + 1402, + 1164, + 299, + 1164 + ], + "score": 0.809 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1398, + 1972, + 1398, + 2035, + 297, + 2035 + ], + "score": 0.799 + }, + { + "category_id": 1, + "poly": [ + 293, + 1798, + 1402, + 1798, + 1402, + 1863, + 293, + 1863 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 298, + 895, + 1405, + 895, + 1405, + 992, + 298, + 992 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 294, + 1884, + 1402, + 1884, + 1402, + 1948, + 294, + 1948 + ], + "score": 0.762 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1506.0, + 1403.0, + 1506.0, + 1403.0, + 1542.0, + 297.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1538.0, + 1286.0, + 1538.0, + 1286.0, + 1575.0, + 322.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 292.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 299.0, + 323.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 290.0, + 520.0, + 290.0, + 520.0, + 325.0, + 322.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 805.0, + 1405.0, + 805.0, + 1405.0, + 849.0, + 292.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 840.0, + 611.0, + 840.0, + 611.0, + 876.0, + 322.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1270.0, + 1404.0, + 1270.0, + 1404.0, + 1311.0, + 293.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1304.0, + 1404.0, + 1304.0, + 1404.0, + 1340.0, + 321.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1337.0, + 984.0, + 1337.0, + 984.0, + 1368.0, + 323.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 719.0, + 1405.0, + 719.0, + 1405.0, + 762.0, + 292.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 755.0, + 610.0, + 755.0, + 610.0, + 786.0, + 323.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1388.0, + 1403.0, + 1388.0, + 1403.0, + 1429.0, + 295.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1418.0, + 1404.0, + 1418.0, + 1404.0, + 1461.0, + 321.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1447.0, + 518.0, + 1447.0, + 518.0, + 1487.0, + 323.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 517.0, + 1405.0, + 517.0, + 1405.0, + 556.0, + 294.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 550.0, + 1204.0, + 550.0, + 1204.0, + 585.0, + 323.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 431.0, + 1404.0, + 431.0, + 1404.0, + 471.0, + 292.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 463.0, + 1277.0, + 463.0, + 1277.0, + 499.0, + 322.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 344.0, + 1399.0, + 344.0, + 1399.0, + 384.0, + 294.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 379.0, + 1047.0, + 379.0, + 1047.0, + 412.0, + 321.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 1404.0, + 1708.0, + 1404.0, + 1748.0, + 295.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1742.0, + 780.0, + 1742.0, + 780.0, + 1778.0, + 324.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1186.0, + 1405.0, + 1186.0, + 1405.0, + 1226.0, + 296.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1218.0, + 1096.0, + 1218.0, + 1096.0, + 1253.0, + 322.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1594.0, + 1405.0, + 1594.0, + 1405.0, + 1629.0, + 296.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1626.0, + 1405.0, + 1626.0, + 1405.0, + 1658.0, + 321.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1657.0, + 519.0, + 1657.0, + 519.0, + 1687.0, + 324.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 602.0, + 1403.0, + 602.0, + 1403.0, + 645.0, + 294.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 634.0, + 1405.0, + 634.0, + 1405.0, + 676.0, + 320.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 668.0, + 534.0, + 668.0, + 534.0, + 699.0, + 324.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1013.0, + 1403.0, + 1013.0, + 1403.0, + 1049.0, + 297.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1044.0, + 1136.0, + 1044.0, + 1136.0, + 1079.0, + 323.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1098.0, + 1403.0, + 1098.0, + 1403.0, + 1139.0, + 294.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1132.0, + 1083.0, + 1132.0, + 1083.0, + 1165.0, + 322.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1403.0, + 1969.0, + 1403.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 966.0, + 2002.0, + 966.0, + 2035.0, + 322.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1796.0, + 1404.0, + 1796.0, + 1404.0, + 1838.0, + 293.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1829.0, + 1145.0, + 1829.0, + 1145.0, + 1863.0, + 321.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 897.0, + 1401.0, + 897.0, + 1401.0, + 931.0, + 296.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 929.0, + 1405.0, + 929.0, + 1405.0, + 963.0, + 322.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 957.0, + 908.0, + 957.0, + 908.0, + 994.0, + 323.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1920.0, + 294.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1915.0, + 472.0, + 1915.0, + 472.0, + 1948.0, + 323.0, + 1948.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.826 + }, + { + "category_id": 1, + "poly": [ + 296, + 677, + 1379, + 677, + 1379, + 713, + 296, + 713 + ], + "score": 0.817 + }, + { + "category_id": 1, + "poly": [ + 299, + 844, + 1399, + 844, + 1399, + 910, + 299, + 910 + ], + "score": 0.816 + }, + { + "category_id": 1, + "poly": [ + 297, + 731, + 1405, + 731, + 1405, + 825, + 297, + 825 + ], + "score": 0.809 + }, + { + "category_id": 1, + "poly": [ + 298, + 1127, + 1401, + 1127, + 1401, + 1191, + 298, + 1191 + ], + "score": 0.804 + }, + { + "category_id": 1, + "poly": [ + 292, + 1689, + 1403, + 1689, + 1403, + 1754, + 292, + 1754 + ], + "score": 0.773 + }, + { + "category_id": 1, + "poly": [ + 298, + 594, + 1395, + 594, + 1395, + 659, + 298, + 659 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 303, + 1491, + 1401, + 1491, + 1401, + 1586, + 303, + 1586 + ], + "score": 0.764 + }, + { + "category_id": 1, + "poly": [ + 293, + 928, + 1399, + 928, + 1399, + 964, + 293, + 964 + ], + "score": 0.76 + }, + { + "category_id": 1, + "poly": [ + 297, + 980, + 1404, + 980, + 1404, + 1107, + 297, + 1107 + ], + "score": 0.744 + }, + { + "category_id": 1, + "poly": [ + 293, + 1605, + 1400, + 1605, + 1400, + 1670, + 293, + 1670 + ], + "score": 0.743 + }, + { + "category_id": 1, + "poly": [ + 296, + 1407, + 1399, + 1407, + 1399, + 1474, + 296, + 1474 + ], + "score": 0.742 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1403, + 1941, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.735 + }, + { + "category_id": 1, + "poly": [ + 297, + 1323, + 1399, + 1323, + 1399, + 1390, + 297, + 1390 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 292, + 1773, + 1400, + 1773, + 1400, + 1838, + 292, + 1838 + ], + "score": 0.73 + }, + { + "category_id": 1, + "poly": [ + 294, + 1857, + 1401, + 1857, + 1401, + 1921, + 294, + 1921 + ], + "score": 0.71 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1400, + 229, + 1400, + 323, + 296, + 323 + ], + "score": 0.709 + }, + { + "category_id": 1, + "poly": [ + 294, + 510, + 1401, + 510, + 1401, + 575, + 294, + 575 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 297, + 426, + 1401, + 426, + 1401, + 491, + 297, + 491 + ], + "score": 0.682 + }, + { + "category_id": 1, + "poly": [ + 297, + 1209, + 1405, + 1209, + 1405, + 1304, + 297, + 1304 + ], + "score": 0.674 + }, + { + "category_id": 1, + "poly": [ + 296, + 343, + 1401, + 343, + 1401, + 408, + 296, + 408 + ], + "score": 0.67 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 674.0, + 1385.0, + 674.0, + 1385.0, + 718.0, + 295.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 841.0, + 1404.0, + 841.0, + 1404.0, + 884.0, + 294.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 873.0, + 398.0, + 873.0, + 398.0, + 912.0, + 319.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 1405.0, + 728.0, + 1405.0, + 770.0, + 294.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 760.0, + 1405.0, + 760.0, + 1405.0, + 801.0, + 319.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 790.0, + 423.0, + 790.0, + 423.0, + 827.0, + 319.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1120.0, + 1407.0, + 1120.0, + 1407.0, + 1169.0, + 291.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1159.0, + 887.0, + 1159.0, + 887.0, + 1190.0, + 324.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1687.0, + 1406.0, + 1687.0, + 1406.0, + 1728.0, + 291.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1721.0, + 690.0, + 1721.0, + 690.0, + 1754.0, + 325.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 594.0, + 1400.0, + 594.0, + 1400.0, + 630.0, + 295.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 627.0, + 860.0, + 627.0, + 860.0, + 660.0, + 323.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1491.0, + 1405.0, + 1491.0, + 1405.0, + 1528.0, + 295.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1524.0, + 1405.0, + 1524.0, + 1405.0, + 1558.0, + 322.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1551.0, + 399.0, + 1551.0, + 399.0, + 1587.0, + 320.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 928.0, + 1403.0, + 928.0, + 1403.0, + 967.0, + 295.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 981.0, + 1405.0, + 981.0, + 1405.0, + 1020.0, + 294.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1053.0, + 320.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1042.0, + 1405.0, + 1042.0, + 1405.0, + 1083.0, + 320.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1074.0, + 888.0, + 1074.0, + 888.0, + 1111.0, + 322.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1607.0, + 1402.0, + 1607.0, + 1402.0, + 1643.0, + 295.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1638.0, + 1141.0, + 1638.0, + 1141.0, + 1671.0, + 322.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1404.0, + 1404.0, + 1404.0, + 1404.0, + 1447.0, + 292.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1439.0, + 1242.0, + 1439.0, + 1242.0, + 1476.0, + 320.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1407.0, + 1939.0, + 1407.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 322.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 2002.0, + 520.0, + 2002.0, + 520.0, + 2035.0, + 319.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1360.0, + 296.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1355.0, + 1318.0, + 1355.0, + 1318.0, + 1392.0, + 322.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1769.0, + 1403.0, + 1769.0, + 1403.0, + 1812.0, + 294.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1804.0, + 984.0, + 1804.0, + 984.0, + 1839.0, + 323.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1857.0, + 1402.0, + 1857.0, + 1402.0, + 1892.0, + 296.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1889.0, + 1295.0, + 1889.0, + 1295.0, + 1921.0, + 323.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 258.0, + 1404.0, + 258.0, + 1404.0, + 300.0, + 320.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 292.0, + 533.0, + 292.0, + 533.0, + 322.0, + 324.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 508.0, + 1403.0, + 508.0, + 1403.0, + 548.0, + 293.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 543.0, + 933.0, + 543.0, + 933.0, + 576.0, + 321.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 422.0, + 1406.0, + 422.0, + 1406.0, + 468.0, + 292.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 459.0, + 1215.0, + 459.0, + 1215.0, + 492.0, + 322.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1208.0, + 1406.0, + 1208.0, + 1406.0, + 1248.0, + 292.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1241.0, + 1405.0, + 1241.0, + 1405.0, + 1280.0, + 323.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1272.0, + 783.0, + 1272.0, + 783.0, + 1306.0, + 324.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 343.0, + 1403.0, + 343.0, + 1403.0, + 379.0, + 296.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 374.0, + 1405.0, + 374.0, + 1405.0, + 410.0, + 323.0, + 410.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 303, + 229, + 1399, + 229, + 1399, + 324, + 303, + 324 + ], + "score": 0.77 + }, + { + "category_id": 1, + "poly": [ + 295, + 1355, + 1398, + 1355, + 1398, + 1421, + 295, + 1421 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 291, + 1271, + 1399, + 1271, + 1399, + 1337, + 291, + 1337 + ], + "score": 0.687 + }, + { + "category_id": 1, + "poly": [ + 295, + 827, + 1400, + 827, + 1400, + 890, + 295, + 890 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 298, + 343, + 1404, + 343, + 1404, + 465, + 298, + 465 + ], + "score": 0.674 + }, + { + "category_id": 1, + "poly": [ + 295, + 992, + 1400, + 992, + 1400, + 1056, + 295, + 1056 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 291, + 1189, + 1402, + 1189, + 1402, + 1255, + 291, + 1255 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 297, + 1438, + 1399, + 1438, + 1399, + 1504, + 297, + 1504 + ], + "score": 0.657 + }, + { + "category_id": 1, + "poly": [ + 295, + 1075, + 1400, + 1075, + 1400, + 1171, + 295, + 1171 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 295, + 909, + 1400, + 909, + 1400, + 974, + 295, + 974 + ], + "score": 0.651 + }, + { + "category_id": 1, + "poly": [ + 294, + 1521, + 1399, + 1521, + 1399, + 1587, + 294, + 1587 + ], + "score": 0.601 + }, + { + "category_id": 1, + "poly": [ + 296, + 1604, + 1401, + 1604, + 1401, + 1670, + 296, + 1670 + ], + "score": 0.558 + }, + { + "category_id": 1, + "poly": [ + 297, + 599, + 1403, + 599, + 1403, + 695, + 297, + 695 + ], + "score": 0.554 + }, + { + "category_id": 1, + "poly": [ + 301, + 486, + 1401, + 486, + 1401, + 581, + 301, + 581 + ], + "score": 0.388 + }, + { + "category_id": 1, + "poly": [ + 299, + 713, + 1400, + 713, + 1400, + 808, + 299, + 808 + ], + "score": 0.159 + }, + { + "category_id": 13, + "poly": [ + 1088, + 1354, + 1142, + 1354, + 1142, + 1385, + 1088, + 1385 + ], + "score": 0.87, + "latex": "S ^ { 4 } L" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 268.0, + 297.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 296.0, + 324.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 291.0, + 1303.0, + 291.0, + 1303.0, + 329.0, + 323.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1350.0, + 1087.0, + 1350.0, + 1087.0, + 1393.0, + 293.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1350.0, + 1403.0, + 1350.0, + 1403.0, + 1393.0, + 1143.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1388.0, + 1008.0, + 1388.0, + 1008.0, + 1420.0, + 321.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1270.0, + 1402.0, + 1270.0, + 1402.0, + 1309.0, + 294.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1303.0, + 1195.0, + 1303.0, + 1195.0, + 1338.0, + 323.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 824.0, + 1405.0, + 824.0, + 1405.0, + 865.0, + 294.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 859.0, + 609.0, + 859.0, + 609.0, + 889.0, + 321.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 339.0, + 1404.0, + 339.0, + 1404.0, + 382.0, + 292.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 375.0, + 1404.0, + 375.0, + 1404.0, + 407.0, + 321.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 402.0, + 1406.0, + 402.0, + 1406.0, + 440.0, + 320.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 431.0, + 398.0, + 431.0, + 398.0, + 470.0, + 320.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 991.0, + 1404.0, + 991.0, + 1404.0, + 1030.0, + 294.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1024.0, + 500.0, + 1024.0, + 500.0, + 1056.0, + 323.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1190.0, + 1405.0, + 1190.0, + 1405.0, + 1226.0, + 298.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1222.0, + 1046.0, + 1222.0, + 1046.0, + 1254.0, + 324.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1435.0, + 1403.0, + 1435.0, + 1403.0, + 1478.0, + 292.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1471.0, + 938.0, + 1471.0, + 938.0, + 1505.0, + 322.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1073.0, + 1404.0, + 1073.0, + 1404.0, + 1114.0, + 293.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1107.0, + 1406.0, + 1107.0, + 1406.0, + 1144.0, + 320.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1139.0, + 857.0, + 1139.0, + 857.0, + 1172.0, + 322.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 911.0, + 1402.0, + 911.0, + 1402.0, + 947.0, + 296.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 940.0, + 915.0, + 940.0, + 915.0, + 975.0, + 322.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1520.0, + 1401.0, + 1520.0, + 1401.0, + 1557.0, + 295.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1553.0, + 1271.0, + 1553.0, + 1271.0, + 1589.0, + 322.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1599.0, + 1405.0, + 1599.0, + 1405.0, + 1645.0, + 295.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1634.0, + 1029.0, + 1634.0, + 1029.0, + 1669.0, + 321.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 597.0, + 1404.0, + 597.0, + 1404.0, + 638.0, + 295.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 631.0, + 1402.0, + 631.0, + 1402.0, + 664.0, + 324.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 660.0, + 1228.0, + 660.0, + 1228.0, + 696.0, + 324.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 486.0, + 1405.0, + 486.0, + 1405.0, + 524.0, + 295.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 521.0, + 1402.0, + 521.0, + 1402.0, + 552.0, + 323.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 548.0, + 810.0, + 548.0, + 810.0, + 582.0, + 322.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 714.0, + 1404.0, + 714.0, + 1404.0, + 748.0, + 295.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 743.0, + 1405.0, + 743.0, + 1405.0, + 780.0, + 322.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 774.0, + 612.0, + 774.0, + 612.0, + 807.0, + 321.0, + 807.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1004, + 1405, + 1004, + 1405, + 1129, + 297, + 1129 + ], + "score": 0.967 + }, + { + "category_id": 5, + "poly": [ + 315, + 710, + 1388, + 710, + 1388, + 875, + 315, + 875 + ], + "score": 0.956, + "html": "
Labeled data Method1%2%5%10% Top-5 accuracy20%50%100%
Supervised trained on pixels44.159.975.282.187.991.893.8
ReseNet trained on CPC v2 (fine-tuned)72.979.686.189.591.493.695.2
" + }, + { + "category_id": 5, + "poly": [ + 315, + 425, + 1388, + 425, + 1388, + 589, + 315, + 589 + ], + "score": 0.947, + "html": "
Labeled data Method1%2%5%10% Top-1 accuracy20%50%100%
Supervised trained on pixels23.134.850.260.067.674.478.0
ReseNet trained on CPC v2 (fine-tuned)46.354.764.169.773.077.380.6
" + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.908 + }, + { + "category_id": 6, + "poly": [ + 486, + 385, + 1211, + 385, + 1211, + 419, + 486, + 419 + ], + "score": 0.901 + }, + { + "category_id": 6, + "poly": [ + 400, + 671, + 1297, + 671, + 1297, + 706, + 400, + 706 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 301, + 946, + 712, + 946, + 712, + 979, + 301, + 979 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.863 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 509, + 228, + 509, + 261, + 300, + 261 + ], + "score": 0.704 + }, + { + "category_id": 2, + "poly": [ + 294, + 1183, + 1367, + 1183, + 1367, + 1988, + 294, + 1988 + ], + "score": 0.539 + }, + { + "category_id": 0, + "poly": [ + 299, + 296, + 644, + 296, + 644, + 329, + 299, + 329 + ], + "score": 0.532 + }, + { + "category_id": 5, + "poly": [ + 294, + 1183, + 1367, + 1183, + 1367, + 1988, + 294, + 1988 + ], + "score": 0.217, + "html": "
def CPC(latents,target_dim=64,emb_scale=0.1, steps_to_ignore=2,steps_to_predict=3):
# latents:[B,H,W,D] loss = 0.0
context = pixelCNN(latents)
targets = Conv2D(output_channels=target_dim,
kernel_shape=(l,1))(latents)
batch_dim, col_dim, row_dim = targets.shape[:-1]
targets = reshape(target,[-1,target_dim])
col_dim_i = col_dim - i - 1for i in range(steps_to_ignore, steps_to_predict):
total_elements = batch_dim * col_dim_i * row_dim
preds_i = Conv2D(output_channels=target_dim,
kernel_shape=(1,1))(context)
preds_i = preds_i[:,:-(i+l),:,:] * emb_scale
preds_i = reshape(preds_i,[-1,target_dim])
logits = matmul(preds_i,targets,transpose_b=True)
b = range(total_elements)/ (col_dim_i * row_dim)
col = range(total_elements)
labels = b * col_dim * im_row_dim + (i+l) * row_dim + col
loss += softmax_cross_entropy_with_logits(logits,labels)
return loss
" + }, + { + "category_id": 1, + "poly": [ + 300, + 228, + 509, + 228, + 509, + 261, + 300, + 261 + ], + "score": 0.193 + }, + { + "category_id": 13, + "poly": [ + 379, + 1066, + 558, + 1066, + 558, + 1100, + 379, + 1100 + ], + "score": 0.94, + "latex": "z _ { i , j } = f _ { \\theta } ( \\pmb { x } _ { i , j } )" + }, + { + "category_id": 13, + "poly": [ + 606, + 1065, + 799, + 1065, + 799, + 1098, + 606, + 1098 + ], + "score": 0.92, + "latex": "i , j \\in \\{ 1 , . . . , 7 \\}" + }, + { + "category_id": 13, + "poly": [ + 1196, + 1097, + 1394, + 1097, + 1394, + 1126, + 1196, + 1126 + ], + "score": 0.91, + "latex": "B \\times 7 \\times 7 \\times 4 0 9 6" + }, + { + "category_id": 13, + "poly": [ + 756, + 1098, + 804, + 1098, + 804, + 1130, + 756, + 1130 + ], + "score": 0.9, + "latex": "\\{ { \\pmb x } \\}" + }, + { + "category_id": 13, + "poly": [ + 491, + 1314, + 523, + 1314, + 523, + 1338, + 491, + 1338 + ], + "score": 0.84, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 824, + 1407, + 854, + 1407, + 854, + 1429, + 824, + 1429 + ], + "score": 0.81, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 643, + 1099, + 668, + 1099, + 668, + 1124, + 643, + 1124 + ], + "score": 0.8, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 492, + 1345, + 522, + 1345, + 522, + 1368, + 492, + 1368 + ], + "score": 0.8, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 640, + 1527, + 672, + 1527, + 672, + 1551, + 640, + 1551 + ], + "score": 0.79, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 847, + 1376, + 868, + 1376, + 868, + 1398, + 847, + 1398 + ], + "score": 0.78, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 491, + 1437, + 522, + 1437, + 522, + 1459, + 491, + 1459 + ], + "score": 0.78, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 526, + 1649, + 556, + 1649, + 556, + 1672, + 526, + 1672 + ], + "score": 0.77, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 525, + 1679, + 555, + 1679, + 555, + 1704, + 525, + 1704 + ], + "score": 0.76, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 508, + 1741, + 538, + 1741, + 538, + 1765, + 508, + 1765 + ], + "score": 0.74, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 879, + 1620, + 901, + 1620, + 901, + 1641, + 879, + 1641 + ], + "score": 0.73, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 558, + 1496, + 589, + 1496, + 589, + 1521, + 558, + 1521 + ], + "score": 0.73, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 747, + 1191, + 797, + 1191, + 797, + 1218, + 747, + 1218 + ], + "score": 0.72, + "latex": "= 6 4" + }, + { + "category_id": 13, + "poly": [ + 680, + 1222, + 715, + 1222, + 715, + 1248, + 680, + 1248 + ], + "score": 0.72, + "latex": "= 2" + }, + { + "category_id": 13, + "poly": [ + 507, + 1862, + 538, + 1862, + 538, + 1886, + 507, + 1886 + ], + "score": 0.71, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 526, + 1587, + 556, + 1587, + 556, + 1611, + 526, + 1611 + ], + "score": 0.7, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 458, + 1832, + 489, + 1832, + 489, + 1856, + 458, + 1856 + ], + "score": 0.7, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 478, + 1921, + 522, + 1921, + 522, + 1947, + 478, + 1947 + ], + "score": 0.69, + "latex": "+ =" + }, + { + "category_id": 13, + "poly": [ + 1012, + 1222, + 1048, + 1222, + 1048, + 1247, + 1012, + 1247 + ], + "score": 0.67, + "latex": "= 3" + }, + { + "category_id": 13, + "poly": [ + 845, + 1529, + 867, + 1529, + 867, + 1551, + 845, + 1551 + ], + "score": 0.63, + "latex": "^ { \\star }" + }, + { + "category_id": 13, + "poly": [ + 994, + 1651, + 1016, + 1651, + 1016, + 1673, + 994, + 1673 + ], + "score": 0.62, + "latex": "\\star" + }, + { + "category_id": 13, + "poly": [ + 579, + 1863, + 601, + 1863, + 601, + 1886, + 579, + 1886 + ], + "score": 0.6, + "latex": "\\star" + }, + { + "category_id": 13, + "poly": [ + 1044, + 1802, + 1066, + 1802, + 1066, + 1825, + 1044, + 1825 + ], + "score": 0.6, + "latex": "\\star" + }, + { + "category_id": 13, + "poly": [ + 1093, + 1863, + 1117, + 1863, + 1117, + 1886, + 1093, + 1886 + ], + "score": 0.58, + "latex": "\\star" + }, + { + "category_id": 13, + "poly": [ + 745, + 1863, + 767, + 1863, + 767, + 1886, + 745, + 1886 + ], + "score": 0.57, + "latex": "\\star" + }, + { + "category_id": 13, + "poly": [ + 895, + 1346, + 914, + 1346, + 914, + 1367, + 895, + 1367 + ], + "score": 0.56, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1145, + 1742, + 1163, + 1742, + 1163, + 1763, + 1145, + 1763 + ], + "score": 0.55, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 930, + 1589, + 948, + 1589, + 948, + 1611, + 930, + 1611 + ], + "score": 0.54, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1044, + 1528, + 1066, + 1528, + 1066, + 1551, + 1044, + 1551 + ], + "score": 0.54, + "latex": "\\star" + }, + { + "category_id": 13, + "poly": [ + 960, + 1861, + 985, + 1861, + 985, + 1886, + 960, + 1886 + ], + "score": 0.5, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 443, + 1282, + 529, + 1282, + 529, + 1309, + 443, + 1309 + ], + "score": 0.47, + "latex": "= ~ 0 . 0" + }, + { + "category_id": 13, + "poly": [ + 426, + 1801, + 457, + 1801, + 457, + 1824, + 426, + 1824 + ], + "score": 0.36, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 979, + 1191, + 1045, + 1191, + 1045, + 1217, + 979, + 1217 + ], + "score": 0.33, + "latex": "= 0 \\mathrm { ~ . ~ } 1" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 378.0, + 1214.0, + 378.0, + 1214.0, + 427.0, + 486.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 666.0, + 1299.0, + 666.0, + 1299.0, + 711.0, + 397.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 946.0, + 715.0, + 946.0, + 715.0, + 982.0, + 296.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 514.0, + 221.0, + 514.0, + 271.0, + 294.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1186.0, + 746.0, + 1186.0, + 746.0, + 1223.0, + 295.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 1186.0, + 978.0, + 1186.0, + 978.0, + 1223.0, + 798.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1186.0, + 1069.0, + 1186.0, + 1069.0, + 1223.0, + 1046.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 1220.0, + 679.0, + 1220.0, + 679.0, + 1251.0, + 430.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1220.0, + 1011.0, + 1220.0, + 1011.0, + 1251.0, + 716.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1220.0, + 1082.0, + 1220.0, + 1082.0, + 1251.0, + 1049.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1247.0, + 751.0, + 1247.0, + 751.0, + 1284.0, + 361.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1279.0, + 442.0, + 1279.0, + 442.0, + 1311.0, + 363.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1279.0, + 537.0, + 1279.0, + 537.0, + 1311.0, + 530.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1309.0, + 490.0, + 1309.0, + 490.0, + 1343.0, + 363.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1309.0, + 815.0, + 1309.0, + 815.0, + 1343.0, + 524.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1341.0, + 491.0, + 1341.0, + 491.0, + 1375.0, + 363.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1341.0, + 894.0, + 1341.0, + 894.0, + 1375.0, + 523.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1341.0, + 1097.0, + 1341.0, + 1097.0, + 1375.0, + 915.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1370.0, + 846.0, + 1370.0, + 846.0, + 1404.0, + 644.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1370.0, + 1129.0, + 1370.0, + 1129.0, + 1404.0, + 869.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1400.0, + 823.0, + 1400.0, + 823.0, + 1437.0, + 361.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 1400.0, + 1163.0, + 1400.0, + 1163.0, + 1437.0, + 855.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1434.0, + 490.0, + 1434.0, + 490.0, + 1465.0, + 364.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1434.0, + 1079.0, + 1434.0, + 1079.0, + 1465.0, + 523.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1463.0, + 1197.0, + 1463.0, + 1197.0, + 1497.0, + 361.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1493.0, + 557.0, + 1493.0, + 557.0, + 1524.0, + 396.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1493.0, + 854.0, + 1493.0, + 854.0, + 1524.0, + 590.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1523.0, + 639.0, + 1523.0, + 639.0, + 1557.0, + 397.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1523.0, + 844.0, + 1523.0, + 844.0, + 1557.0, + 673.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1523.0, + 1043.0, + 1523.0, + 1043.0, + 1557.0, + 868.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1523.0, + 1200.0, + 1523.0, + 1200.0, + 1557.0, + 1067.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1584.0, + 525.0, + 1584.0, + 525.0, + 1619.0, + 394.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1584.0, + 929.0, + 1584.0, + 929.0, + 1619.0, + 557.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1584.0, + 1129.0, + 1584.0, + 1129.0, + 1619.0, + 949.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1616.0, + 878.0, + 1616.0, + 878.0, + 1647.0, + 678.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1616.0, + 1160.0, + 1616.0, + 1160.0, + 1647.0, + 902.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1646.0, + 525.0, + 1646.0, + 525.0, + 1678.0, + 392.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1646.0, + 993.0, + 1646.0, + 993.0, + 1678.0, + 557.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1646.0, + 1184.0, + 1646.0, + 1184.0, + 1678.0, + 1017.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1677.0, + 524.0, + 1677.0, + 524.0, + 1710.0, + 392.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1677.0, + 1129.0, + 1677.0, + 1129.0, + 1710.0, + 556.0, + 1710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1737.0, + 507.0, + 1737.0, + 507.0, + 1771.0, + 397.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1737.0, + 1144.0, + 1737.0, + 1144.0, + 1771.0, + 539.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1737.0, + 1243.0, + 1737.0, + 1243.0, + 1771.0, + 1164.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1795.0, + 425.0, + 1795.0, + 425.0, + 1829.0, + 393.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1795.0, + 1043.0, + 1795.0, + 1043.0, + 1829.0, + 458.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1795.0, + 1208.0, + 1795.0, + 1208.0, + 1829.0, + 1067.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1829.0, + 457.0, + 1829.0, + 457.0, + 1860.0, + 396.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 1829.0, + 849.0, + 1829.0, + 849.0, + 1860.0, + 490.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1857.0, + 506.0, + 1857.0, + 506.0, + 1892.0, + 396.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 539.0, + 1857.0, + 578.0, + 1857.0, + 578.0, + 1892.0, + 539.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1857.0, + 744.0, + 1857.0, + 744.0, + 1892.0, + 602.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1857.0, + 959.0, + 1857.0, + 959.0, + 1892.0, + 768.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1857.0, + 1092.0, + 1857.0, + 1092.0, + 1892.0, + 986.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1857.0, + 1347.0, + 1857.0, + 1347.0, + 1892.0, + 1118.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1916.0, + 477.0, + 1916.0, + 477.0, + 1953.0, + 393.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1916.0, + 1344.0, + 1916.0, + 1344.0, + 1953.0, + 523.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1951.0, + 554.0, + 1951.0, + 554.0, + 1983.0, + 361.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 295.0, + 644.0, + 295.0, + 644.0, + 331.0, + 297.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1004.0, + 1403.0, + 1004.0, + 1403.0, + 1037.0, + 295.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1036.0, + 1406.0, + 1036.0, + 1406.0, + 1069.0, + 295.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1065.0, + 378.0, + 1065.0, + 378.0, + 1102.0, + 292.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1065.0, + 605.0, + 1065.0, + 605.0, + 1102.0, + 559.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1065.0, + 1405.0, + 1065.0, + 1405.0, + 1102.0, + 800.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1096.0, + 642.0, + 1096.0, + 642.0, + 1129.0, + 294.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1096.0, + 755.0, + 1096.0, + 755.0, + 1129.0, + 669.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 1096.0, + 1195.0, + 1096.0, + 1195.0, + 1129.0, + 805.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1096.0, + 1403.0, + 1096.0, + 1403.0, + 1129.0, + 1395.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 514.0, + 221.0, + 514.0, + 271.0, + 294.0, + 271.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 303, + 1811, + 1404, + 1811, + 1404, + 1904, + 303, + 1904 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 372, + 1517, + 1402, + 1517, + 1402, + 1787, + 372, + 1787 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 300, + 264, + 1130, + 264, + 1130, + 878, + 300, + 878 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 294, + 1458, + 1277, + 1458, + 1277, + 1491, + 294, + 1491 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.85 + }, + { + "category_id": 0, + "poly": [ + 300, + 930, + 669, + 930, + 669, + 960, + 300, + 960 + ], + "score": 0.837 + }, + { + "category_id": 1, + "poly": [ + 371, + 975, + 1406, + 975, + 1406, + 1366, + 371, + 1366 + ], + "score": 0.78 + }, + { + "category_id": 1, + "poly": [ + 298, + 1403, + 962, + 1403, + 962, + 1433, + 298, + 1433 + ], + "score": 0.509 + }, + { + "category_id": 0, + "poly": [ + 298, + 1403, + 962, + 1403, + 962, + 1433, + 298, + 1433 + ], + "score": 0.471 + }, + { + "category_id": 0, + "poly": [ + 300, + 264, + 665, + 264, + 665, + 292, + 300, + 292 + ], + "score": 0.254 + }, + { + "category_id": 13, + "poly": [ + 601, + 1721, + 661, + 1721, + 661, + 1751, + 601, + 1751 + ], + "score": 0.9, + "latex": "1 0 ^ { - 5 }" + }, + { + "category_id": 13, + "poly": [ + 696, + 1721, + 756, + 1721, + 756, + 1751, + 696, + 1751 + ], + "score": 0.89, + "latex": "1 0 ^ { - 2 }" + }, + { + "category_id": 13, + "poly": [ + 975, + 1813, + 1007, + 1813, + 1007, + 1842, + 975, + 1842 + ], + "score": 0.88, + "latex": "\\mathbb { D } _ { l }" + }, + { + "category_id": 13, + "poly": [ + 511, + 1753, + 610, + 1753, + 610, + 1787, + 511, + 1787 + ], + "score": 0.77, + "latex": "\\{ 1 6 , 3 2 \\}" + }, + { + "category_id": 13, + "poly": [ + 780, + 632, + 802, + 632, + 802, + 653, + 780, + 653 + ], + "score": 0.77, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 780, + 541, + 802, + 541, + 802, + 562, + 780, + 562 + ], + "score": 0.74, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 507, + 357, + 539, + 357, + 539, + 381, + 507, + 381 + ], + "score": 0.74, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 442, + 326, + 473, + 326, + 473, + 350, + 442, + 350 + ], + "score": 0.71, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 964, + 1690, + 1143, + 1690, + 1143, + 1725, + 964, + 1725 + ], + "score": 0.7, + "latex": "\\{ 0 . 0 5 , 0 . 1 , 0 . 2 \\}" + }, + { + "category_id": 13, + "poly": [ + 780, + 754, + 802, + 754, + 802, + 775, + 780, + 775 + ], + "score": 0.7, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 830, + 725, + 848, + 725, + 848, + 745, + 830, + 745 + ], + "score": 0.69, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1300, + 976, + 1402, + 976, + 1402, + 1006, + 1300, + 1006 + ], + "score": 0.68, + "latex": "2 4 0 \\times 2 4 0" + }, + { + "category_id": 13, + "poly": [ + 780, + 450, + 802, + 450, + 802, + 472, + 780, + 472 + ], + "score": 0.66, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 647, + 662, + 679, + 662, + 679, + 684, + 647, + 684 + ], + "score": 0.65, + "latex": "= ^ { \\parallel }" + }, + { + "category_id": 13, + "poly": [ + 815, + 977, + 889, + 977, + 889, + 1006, + 815, + 1006 + ], + "score": 0.64, + "latex": "8 0 \\mathrm { x } 8 0" + }, + { + "category_id": 13, + "poly": [ + 424, + 418, + 456, + 418, + 456, + 442, + 424, + 442 + ], + "score": 0.63, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1170, + 1201, + 1271, + 1201, + 1271, + 1230, + 1170, + 1230 + ], + "score": 0.62, + "latex": "2 4 0 \\times 2 4 0" + }, + { + "category_id": 13, + "poly": [ + 443, + 815, + 471, + 815, + 471, + 837, + 443, + 837 + ], + "score": 0.59, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 802, + 1550, + 851, + 1550, + 851, + 1578, + 802, + 1578 + ], + "score": 0.42, + "latex": "\\mathbf { \\hat { v } } _ { \\mathbf { \\Delta } } 2 \\mathbf { \\hat { \\mu } }" + }, + { + "category_id": 13, + "poly": [ + 828, + 507, + 899, + 507, + 899, + 535, + 828, + 535 + ], + "score": 0.42, + "latex": "= 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 829, + 415, + 898, + 415, + 898, + 443, + 829, + 443 + ], + "score": 0.4, + "latex": "= 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 476, + 786, + 506, + 786, + 506, + 807, + 476, + 807 + ], + "score": 0.38, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 829, + 601, + 850, + 601, + 850, + 624, + 829, + 624 + ], + "score": 0.34, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1148, + 977, + 1221, + 977, + 1221, + 1005, + 1148, + 1005 + ], + "score": 0.33, + "latex": "3 2 \\mathrm { x } 3 2 " + }, + { + "category_id": 13, + "poly": [ + 426, + 693, + 456, + 693, + 456, + 715, + 426, + 715 + ], + "score": 0.31, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 828, + 599, + 898, + 599, + 898, + 626, + 828, + 626 + ], + "score": 0.29, + "latex": "= 2 5 6" + }, + { + "category_id": 13, + "poly": [ + 426, + 722, + 457, + 722, + 457, + 747, + 426, + 747 + ], + "score": 0.28, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 927.0, + 672.0, + 927.0, + 672.0, + 964.0, + 295.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1402.0, + 964.0, + 1402.0, + 964.0, + 1436.0, + 296.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 262.0, + 667.0, + 262.0, + 667.0, + 294.0, + 297.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1806.0, + 974.0, + 1806.0, + 974.0, + 1850.0, + 294.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1806.0, + 1405.0, + 1806.0, + 1405.0, + 1850.0, + 1008.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1839.0, + 1405.0, + 1839.0, + 1405.0, + 1876.0, + 294.0, + 1876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1873.0, + 464.0, + 1873.0, + 464.0, + 1905.0, + 297.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1517.0, + 1406.0, + 1517.0, + 1406.0, + 1551.0, + 370.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1549.0, + 801.0, + 1549.0, + 801.0, + 1581.0, + 395.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1549.0, + 1403.0, + 1549.0, + 1403.0, + 1581.0, + 852.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1579.0, + 1404.0, + 1579.0, + 1404.0, + 1612.0, + 393.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1610.0, + 1191.0, + 1610.0, + 1191.0, + 1643.0, + 395.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1649.0, + 1396.0, + 1649.0, + 1396.0, + 1687.0, + 370.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1690.0, + 963.0, + 1690.0, + 963.0, + 1728.0, + 370.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1690.0, + 1404.0, + 1690.0, + 1404.0, + 1728.0, + 1144.0, + 1728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1718.0, + 600.0, + 1718.0, + 600.0, + 1761.0, + 392.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1718.0, + 695.0, + 1718.0, + 695.0, + 1761.0, + 662.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1718.0, + 1407.0, + 1718.0, + 1407.0, + 1761.0, + 757.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1750.0, + 510.0, + 1750.0, + 510.0, + 1790.0, + 394.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1750.0, + 623.0, + 1750.0, + 623.0, + 1790.0, + 611.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 262.0, + 668.0, + 262.0, + 668.0, + 293.0, + 297.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 292.0, + 749.0, + 292.0, + 749.0, + 325.0, + 361.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 325.0, + 441.0, + 325.0, + 441.0, + 354.0, + 364.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 474.0, + 325.0, + 605.0, + 325.0, + 605.0, + 354.0, + 474.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 354.0, + 506.0, + 354.0, + 506.0, + 385.0, + 362.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 354.0, + 781.0, + 354.0, + 781.0, + 385.0, + 540.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 384.0, + 424.0, + 384.0, + 424.0, + 415.0, + 364.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 382.0, + 668.0, + 382.0, + 668.0, + 417.0, + 458.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 415.0, + 423.0, + 415.0, + 423.0, + 445.0, + 393.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 415.0, + 828.0, + 415.0, + 828.0, + 445.0, + 457.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 415.0, + 914.0, + 415.0, + 914.0, + 445.0, + 899.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 444.0, + 779.0, + 444.0, + 779.0, + 478.0, + 576.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 444.0, + 1012.0, + 444.0, + 1012.0, + 478.0, + 803.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 476.0, + 583.0, + 476.0, + 583.0, + 508.0, + 395.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 507.0, + 827.0, + 507.0, + 827.0, + 538.0, + 394.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 507.0, + 914.0, + 507.0, + 914.0, + 538.0, + 900.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 537.0, + 779.0, + 537.0, + 779.0, + 568.0, + 577.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 537.0, + 964.0, + 537.0, + 964.0, + 568.0, + 803.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 566.0, + 1129.0, + 566.0, + 1129.0, + 600.0, + 394.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 596.0, + 827.0, + 596.0, + 827.0, + 631.0, + 395.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 596.0, + 916.0, + 596.0, + 916.0, + 631.0, + 899.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 625.0, + 779.0, + 625.0, + 779.0, + 661.0, + 576.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 625.0, + 914.0, + 625.0, + 914.0, + 661.0, + 803.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 658.0, + 646.0, + 658.0, + 646.0, + 691.0, + 579.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 658.0, + 847.0, + 658.0, + 847.0, + 691.0, + 680.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 689.0, + 425.0, + 689.0, + 425.0, + 719.0, + 394.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 689.0, + 580.0, + 689.0, + 580.0, + 719.0, + 457.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 719.0, + 425.0, + 719.0, + 425.0, + 753.0, + 395.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 719.0, + 829.0, + 719.0, + 829.0, + 753.0, + 458.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 719.0, + 996.0, + 719.0, + 996.0, + 753.0, + 849.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 748.0, + 779.0, + 748.0, + 779.0, + 782.0, + 577.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 748.0, + 963.0, + 748.0, + 963.0, + 782.0, + 803.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 782.0, + 475.0, + 782.0, + 475.0, + 810.0, + 395.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 782.0, + 653.0, + 782.0, + 653.0, + 810.0, + 507.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 810.0, + 442.0, + 810.0, + 442.0, + 842.0, + 362.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 810.0, + 648.0, + 810.0, + 648.0, + 842.0, + 472.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 840.0, + 554.0, + 840.0, + 554.0, + 872.0, + 363.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1454.0, + 1281.0, + 1454.0, + 1281.0, + 1497.0, + 293.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 975.0, + 814.0, + 975.0, + 814.0, + 1011.0, + 375.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 975.0, + 1147.0, + 975.0, + 1147.0, + 1011.0, + 890.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 975.0, + 1299.0, + 975.0, + 1299.0, + 1011.0, + 1222.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1005.0, + 1406.0, + 1005.0, + 1406.0, + 1042.0, + 392.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1074.0, + 393.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1067.0, + 1404.0, + 1067.0, + 1404.0, + 1102.0, + 393.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1135.0, + 392.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1128.0, + 1405.0, + 1128.0, + 1405.0, + 1165.0, + 393.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1161.0, + 892.0, + 1161.0, + 892.0, + 1194.0, + 394.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1196.0, + 1169.0, + 1196.0, + 1169.0, + 1239.0, + 373.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1196.0, + 1406.0, + 1196.0, + 1406.0, + 1239.0, + 1272.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1232.0, + 1405.0, + 1232.0, + 1405.0, + 1268.0, + 394.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1260.0, + 600.0, + 1260.0, + 600.0, + 1301.0, + 392.0, + 1301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1301.0, + 1405.0, + 1301.0, + 1405.0, + 1339.0, + 370.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1333.0, + 1350.0, + 1333.0, + 1350.0, + 1369.0, + 393.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1402.0, + 964.0, + 1402.0, + 964.0, + 1436.0, + 296.0, + 1436.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ.md b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ.md new file mode 100644 index 0000000000000000000000000000000000000000..8a9552fa318e9287260436ce4695c37959ca364e --- /dev/null +++ b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ.md @@ -0,0 +1,232 @@ +# LENGTH-ADAPTIVE TRANSFORMER: TRAIN ONCE WITH LENGTH DROP, USE ANYTIME WITH SEARCH + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Although transformers have achieved impressive accuracies in various tasks in natural language processing, they often come with a prohibitive computational cost, that prevents their use in scenarios with limited computational resources for inference. This need for computational efficiency in inference has been addressed by for instance PoWER-BERT (Goyal et al., 2020) which gradually decreases the length of a sequence as it is passed through layers. These approaches however often assume that the target computational complexity is known in advance at the time of training. This implies that a separate model must be trained for each inference scenario with its distinct computational budget. In this paper, we extend PoWER-BERT to address this issue of inefficiency and redundancy. The proposed extension enables us to train a large-scale transformer, called Length-Adaptive Transformer, once and uses it for various inference scenarios without re-training it. To do so, we train a transformer with LengthDrop, a structural variant of dropout, which stochastically determines the length of a sequence at each layer. We then use a multi-objective evolutionary search to find a length configuration that maximizes the accuracy and minimizes the computational complexity under any given computational budget. Additionally, we significantly extend the applicability of PoWER-BERT beyond sequence-level classification into token-level classification such as span-based question-answering, by introducing the idea of Drop-and-Restore. With Drop-and-Restore, word-vectors are dropped temporarily in intermediate layers and restored at the last layer if necessary. We empirically verify the utility of the proposed approach by demonstrating the superior accuracyefficiency trade-off under various setups, including SQuAD 1.1, MNLI-m, and SST-2. Upon publication, the code to reproduce our work will be open-sourced. + +# 1 INTRODUCTION + +Pretrained language models (Peters et al., 2018; Devlin et al., 2018; Radford et al., 2019; Yang et al., 2019) have achieved notable improvements in various natural language processing (NLP) tasks. Most of them rely on transformers (Vaswani et al., 2017), and the number of model parameters ranges from hundreds of millions to billions (Shoeybi et al., 2019; Raffel et al., 2019; Kaplan et al., 2020; Brown et al., 2020). Despite this high accuracy, excessive computational overhead during inference, both in terms of time and memory, has hindered its use in real applications. This level of excessive computation has further raised the concern over energy consumption as well (Schwartz et al., 2019; Strubell et al., 2019). + +Recent studies have attempted at addressing these concerns regarding large-scale transformers’ computational and energy efficiency (see $\ S 6$ for a more extensive discussion.) Among these, we focus on PoWER-BERT (Goyal et al., 2020) which progressively reduces sequence length by eliminating word-vectors based on the attention values as passing layers. PoWER-BERT establishes the superiority of accuracy-time trade-off over earlier approaches (Sanh et al., 2019; Sun et al., 2019; Michel et al., 2019). It however requires us to train a separate model for each efficiency constraint. In this paper, we thus develop a framework based on PoWER-BERT such that we can train a single model that can be adapted in the inference time to meet any given efficiency target. + +In order to train a transformer to cope with a diverse set of computational budgets in the inference time, we propose to train one while reducing the sequence length with a random proportion at each layer. We refer to this procedure as LengthDrop which was motivated by the nested dropout (Rippel et al., 2014). We can extract sub-models of shared weights with any length configuration without requiring extra post-processing nor additional finetuning. + +Once a transformer is trained with the proposed LengthDrop, we search for the length configuration that maximizes the accuracy given a computational budget. Because this search is combinatorial and has multiple objectives (accuracy and efficiency), we use an evolutionary search algorithm, which further allows us to obtain a full Pareto frontier of accuracy-efficiency trade-off of each model. + +It is not trivial to find an optimal length configuration given the inference-time computational budget, although it is extremely important in order to deploy these large-scale transformers in practice. In this work, we propose to use evolutionary search to find a length configuration that maximizes the accuracy within a given computational budget. We can further compute the Pareto frontier of accuracy-efficiency trade-off to obtain a sequence of length configurations with varying efficiency profiles. + +PoWER-BERT, which forms the foundation of the proposed two-stage procedure, is only applicable to sequence-level classification, because by design it eliminates some of the word vectors at each layer. In other words, it cannot be used for token-level tasks such as span-based question answering (Rajpurkar et al., 2016), because these tasks require hidden representations of the entire input sequence at the final layer. We thus propose to extend PoWER-BERT with a novel Drop-and-Restore process (§3.3), which eliminates this inherent limitation. Word vectors are dropped and set aside, rather than eliminated, in intermediate layers to maintain the saving of computational cost, as was with the original PoWER-BERT. These set-aside vectors are then restored at the final hidden layer and provided as an input to a subsequent task-specific layer, which is unlike the original PoWER-BERT. + +The main contributions of this work are two-fold. First, we introduce LengthDrop, a structured variant of dropout for training a single Length-Adaptive Transformer model that allows us to automatically derive multiple sub-models with different length configurations in the inference time using evolutionary search, without requiring any re-training. Second, we design Drop-and-Restore process that makes PoWER-BERT applicable beyond classification, which enables PoWER-BERT to be applicable to a wider range of NLP tasks such as span-based question answering. We empirically verify Length-Adaptive Transformer works quite well using the variants of BERT on a diverse set of NLP tasks, including SQuAD 1.1 (Rajpurkar et al., 2016) and two sequence-level classification tasks in GLUE benchmark (Wang et al., 2018). Our experiments reveal that the proposed approach grants us a fine-grained control of computational efficiency and a superior accuracy-efficiency trade-off in the inference time, compared to existing approaches. + +# 2 BACKGROUND: TRANSFORMERS AND POWER-BERT + +Before we describe our main approach, we review some of the building blocks in this section. In particular, we review transformers, which are a standard backbone used in natural language processing these days, and PoWER-BERT, which was recently proposed as an effective way to train a large-scale, but highly efficient transformer for sequence-level classification. + +# 2.1 TRANSFORMERS AND BERT + +A transformer is a particular neural network that has been designed to work with a variable-length sequence input and is implemented as a stack of self-attention and fully-connected layers (Vaswani et al., 2017). Here, we give a brief overview of the transformer which is the basic building block of the proposed approach. + +Each token $x _ { t }$ in a sequence of tokens ${ \boldsymbol x } = ( x _ { 1 } , \dots , x _ { N } )$ , representing input text, is first turned into a continuous vector $h _ { t } ^ { 0 } \in \mathbb { R } ^ { H }$ which is the sum of the token and position embedding vectors. This sequence is fed into the first transformer layer which returns another sequence of the same length $\dot { h ^ { 1 } } \in \mathbb { R } ^ { N \times H }$ . We repeat this procedure $L$ times, for a transformer with $L$ layers, to obtain $h ^ { L } = ( h _ { 1 } ^ { L } , \ldots , h _ { N } ^ { L } )$ . We refer to each vector in the hidden sequence at each layer as a word vector to emphasize that there exists a correspondence between each such vector and one of the input words. + +Although the transformer was first introduced for the problem of machine translation, Devlin et al. +(2018) demonstrated that the transformer can be trained and used as a masked language model. + +More specifically, Devlin et al. (2018) showed that the transformer-based masked language model, called BERT, learns a universally useful parameter set that can be finetuned for any downstream task including sequence-level and token-level classification. + +In the case of sequence-level classification, a softmax classifier is attached to the word vector $h _ { 1 } ^ { L }$ associated with the special token [CLS], and the entire network, including the softmax classifier and BERT, is finetuned. For token-level classification, we use each $h _ { t } ^ { L }$ as the final hidden reprsentation of the associated $t$ -th word in the input sequence. This strategy of pretraining followed by finetuning, often referred to as transfer learning, has recently become a dominant approach to classification in natural language processing. + +# 2.2 POWER-BERT + +PoWER-BERT keeps only the topmost $l _ { j }$ word vectors at each layer $j$ by eliminating redundant ones based on the significance score, which is the total amount of attention imposed by a word on the other words (Goyal et al., 2020). $l _ { j }$ is the hyper-parameter determines how many vectors to keep at layer $j$ PoWER-BERT has the same model parameters with BERT, but the extraction layers are interspersed after the self-attention layer in every transformer blocks (Vaswani et al., 2017). + +PoWER-BERT reduces inference time successfully, achieving better accuracy-time trade-off than DistilBERT (Sanh et al., 2019), BERT-PKD (Sun et al., 2019), and Head-Prune (Michel et al., 2019). Despite the original intention of maximizing the inference efficiency with the minimal loss in accuracy, it is possible to set up PoWER-BERT to be both more efficient and more accurate compared to the original BERT, which was observed but largely overlooked by Goyal et al. (2020). + +Training a PoWER-BERT model consists of three steps: (1) finetuning, (2) length configuration search, and (3) re-training. The finetuning step is just like the standard finetuning step of BERT given a target task. A length configuration is a sequence of retention parameters $\left( l _ { 1 } , \cdots l _ { L } \right)$ , each of which corresponds to the number of word vectors that are kept at each layer. These retention parameters are learned along with all the other parameters to minimize the original task loss together with an extra term that approximately measures the number of retained word vectors across layers. In the re-training step, PoWER-BERT is finetuned with the length configuration fixed to its learned one. + +For each computational budget, we must train a separate model going through all three steps described above. Moreover, the length configuration search step above is only approximate, as it relies on relaxation of retention parameters which are inherently discrete. This leads to the lack of guaranteed correlation between the success of this stage and true run-time. Even worse, it is a delicate act to tune the length configuration given a target computational budget, because trade-off is implicitly made via a regularization coefficient. Furthermore, PoWER-BERT has an inherent limitation in that it only applies to sequence-level classification because it eliminates word vectors in intermediate layers. + +# 3 LENGTH-ADAPTIVE TRANSFORMER + +In this section, we explain our proposed framework which results in a transformer that reduces the length of a sequence at each layer with an arbitrary rate. We call such a resulting transformer a Length-Adaptive Transformer. We train Length-Adaptive Transformer with LengthDrop which randomly samples the number of hidden vectors to be dropped at each layer with the goal of making the final model robust to such drop in the inference time. Once the model is trained, we search for the optimal trade-off between accuracy and efficiency using multi-objective evolutionary search, which allows us to use the model for any given computational budget without finetuning nor re-training. At the end of this section, we describe Drop-and-Restore process as a way to greatly increase the applicability of PoWER-BERT which forms a building block of the proposed framework. + +In short, we train a Length-Adaptive Transformer once with LengthDrop and Drop-and-Restore, and use it with an automatically determined length configuration for inference with any target computational budget, on both sequence-level and token-level tasks. + +# 3.1 LENGTHDROP + +Earlier approaches to efficient inference with transformers have focused on a scenario where the target computational budget for inference is known in advance Sanh et al. (2019); Goyal et al. (2020). This greatly increases the cost of deploying transformers, as it requires us to train a separate transformer for each scenario. Instead, we propose to train one model that could be used for a diverse set of target computational budgets without retraining. + +LengthDrop randomly generates a length configuration by sequentially sampling a sequence length $l _ { i + 1 }$ at the $( i + 1 )$ -th layer based on the previous layer’s sequence length $l _ { i }$ , following the uniform distribution $\mathcal { U } ( \lceil ( 1 - p ) l _ { i } \rceil , l _ { i } )$ , where $l _ { 0 }$ is set to the length of the input sequence, and $p$ is the LengthDrop probability. This sequential sampling results in a length configuration $( l _ { 1 } , \cdots , l _ { L } )$ . Length-Adaptive Transformer can be thought of as consisting of a full model and many sub-models corresponding to different length configuration, similarly to a neural network trained with dropout (Srivastava et al., 2014). + +LayerDrop From the perspective of each word vector, the proposed LengthDrop could be thought of as skipping the layers between when it was set aside and the final layer where it was restored. The word vector however does not have any information based on which it can determine whether it would be dropped at any particular layer. In our preliminary experiments, we found that this greatly hinders optimization. We address this issue by using LayerDrop (Fan et al., 2019) which skips each layer of a transformer uniformly at random. The LayerDrop encourages each word vector to be agnostic to skipping any number of layers between when it is dropped and when it is restored, just like dropout (Srivastava et al., 2014) prevents hidden neurons from co-adapting with each other by randomly dropping them. + +Sandwich Rule and Inplace Distillation We observed that standard supervised training with LengthDrop does not work well in the preliminary experiments. We instead borrow a pair of training techniques developed by Yu & Huang (2019) which are sandwich rule and inplace distillation, for better optimization as well as final generalization. At each update, we update the full model without LengthDrop as usual to minimize the supervised loss function. We simultaneously update $n _ { s }$ randomly-sampled sub-models (which are called sandwiches) and the smallest-possible submodel, which corresponds to keeping only $\lceil ( 1 - p ) l _ { i } \rceil$ word vectors at each layer $i$ , using knowledge distillation (Hinton et al., 2015) from the full model. Here, sub-models mean models with length reduction. They are trained to their prediction close to the full model’s prediction (inplace distillation). + +# 3.2 EVOLUTIONARY SEARCH OF LENGTH CONFIGURATIONS + +After training a Length-Adaptive Transformer with LengthDrop, we search for appropriate length configurations for possible target computational budgets that will be given at inference time. The length configuration determines the model performance in terms of both accuracy and efficiency. In order to search for the optimal length configuration, we propose to use evolutionary search, similarly to Cai et al. (2019) and Wang et al. (2020). This procedure is efficient, as it only requires a single pass through the relatively small validation set for each length configuration, unlike re-training for a new computational budget which requires multiple passes through a significantly larger training set for each budget. + +We initialize the population with constant-ratio configurations. Each configuration is created by $l _ { i + 1 } = \lceil ( 1 - r ) \bar { l _ { i } } \rceil$ for each layer $i$ with $r$ so that the amount of computation within the initial population is uniformly distributed between those of the smallest and full models. At each iteration, we evolve the population to consist only of configurations lie on a newly updated efficiencyaccuracy Pareto frontier by mutation and cross-over. Mutation alters an original length configuration $( l _ { 1 } , \cdots , l _ { L } )$ to $( l _ { 1 } ^ { \prime } , \cdots , l _ { L } ^ { \prime } )$ by sampling $l _ { i } ^ { \prime }$ from the uniform distribution $\bar { \mathcal { U } } ( l _ { i - 1 } ^ { \prime } , l _ { i + 1 } )$ with the probability $p _ { m }$ or keeping the original length ${ \bar { l } } _ { i } ^ { \prime } = l _ { i }$ , sweeping the layers from $i = 1$ to $i = L$ . Crossover takes two length configurations and averages the lengths at each layer. Both of these operations are performed while ensuring the monotonicity of the lengths over the layers. We repeat this iteration $G$ times, while maintaining $n _ { m }$ mutated configurations and $n _ { c }$ crossover’d configurations. Repeating this procedure pushes the Pareto frontier further to identify the best trade-off between two objectives, efficiency and accuracy, without requiring any continuous relaxation of length configurations nor using a proxy objective function. + +![](images/320388afaf5c944756a7b3f52eddc7583af0fdd53b75f8ac7bf8c3063dd11f2f.jpg) +Figure 1: Illustration of (a) word-vector elimination process in PoWER-BERT (Goyal et al., 2020) and (b) Drop-and-Restore process in Length-Adaptive Transformer. Yellow box and blue boxes imply the output of embedding layer and transformer layers, respectively. Green boxes mean vectors dropped in lower layers and restored at the last layer. Red box is the task-specific layer. Though word-vectors in the middle could be eliminated (or dropped), remaining vectors are left-aligned for the better illustration. In this case, the number of transformer layers is four. + +# 3.3 DROP-AND-RESTORE PROCESS + +The applicability of the PoWER-BERT, based on which our main contribution above was made, is limited to sequence-level classification, because it eliminates word vectors at each layer. In addition to our main contribution above, we thus propose to extend the PoWER-BERT so that it is applicable to token-level classification, such as span-based question-answering. Our proposal, to which we refer as Drop-and-Restore, does not eliminate word vectors at each layer according to the length configuration but instead sets them aside until the final hidden layer. At the final hidden layer, these word vectors are brought back to form the full hidden sequence, as illustrated graphically in Fig. 1. + +# 4 EXPERIMENT SETUP + +Datasets We test the proposed approach on both sequence-level and token-level tasks, the latter of which could not have been done with the original PoWER-BERT unless for the proposed Dropand-Restore. We use MNLI-m and SST-2 from GLUE benchmark (Wang et al., 2018), as was done to test PoWER-BERT earlier, for sequence-level classification. We choose them because consistent accuracy scores from standard training on them due to their sufficiently large training set imply that they are reliable to verify our approach. We use SQuAD 1.1 (Rajpurkar et al., 2016) for token-level classification. + +Evaluation metrics We use the number of floating operations (FLOPs) as a main metric to measure the inference efficiency given any length configuration, as it is agnostic to the choice of underlying hardware, unlike other alternatives such as hardware-aware latency (Wang et al., 2020) or energy consumption (Henderson et al., 2020). We later demonstrate that FLOPs and wall-clock time on GPU and CPU correlate well with the proposed approach, which is not necessarily the case for other approaches, such as unstructured weight pruning (Han et al., 2015; See et al., 2016). + +Pretrained transformers Since BERT was introduced by Devlin et al. (2018), it has become a standard practice to start from a pretrained (masked) language model and finetune it for each downstream task. We follow the same strategy in this paper and test two pretrained transformerbased language models; BERTBASE (Devlin et al., 2018) and DistilBERT (Sanh et al., 2019), which allows us to demonstrate that the usefulness and applicability of our approach are not tied to any specific architectural choice, such as the number of layers and the maximum length of input sequence. Although we focus on BERT-based masked language models here, the proposed approach is readily applicable to any transformer-based models. + +Learning We train a Length-Adaptive Transformer with LengthDrop probability and LayerDrop probability both set to 0.2. We use $n _ { s } = 2$ randomly sampled intermediate sub-models in addition to the full model and smallest model for applying the sandwich learning rule. + +We start finetuning the pretrained transformer without Drop-and-Restore first, just as Goyal et al. (2020) did with PoWER-BERT. We then continue finetuning it for another five epochs with Drop-andRestore. This is unlike the recommended three epochs by Devlin et al. (2018), as learning progresses slower due to a higher level of stochasticity introduced by LengthDrop and LayerDrop. We use the batch size of 32, the learning rate of $5 e - 5$ for SQuAD v1.1 and $2 e - 5$ for MNLI-m and SST, and the maximum sequence length of 384 for SQuAD v1.1 and 128 for MNLI-m and SST. + +![](images/3ea54ad6f6a4d01240155dd50bd06714bd078c67453a78cfd7ce1b9404ef27c2.jpg) +Figure 2: Pareto curves of F1 score and FLOPs on SQuAD 1.1 (Rajpurkar et al., 2016). We apply the proposed method to $\mathbf { B E R T _ { B a s e } }$ (solid lines) and DistilBERT (dotted lines). For each model, we draw three curves using (1) standard finetuned transformer with constant-rate length reduction, (2) Length-Adaptive Transformer with constant-rate length reduction, and (3) Length-Adaptive Transformer with length configurations obtained from the evolutionary search. + +![](images/4608176aa7c7dba0b4748567c1c64abe798b711bebf826c4b323894a4bfb2b12.jpg) +Figure 3: Correlation between FLOPs and latency with different length configurations. + +Search We run up to $G = 3 0$ iterations of evolutionary search, using $n _ { m } = 3 0$ mutated configurations with mutation probability $p _ { m } = 0 . 5$ and $n _ { c } = 3 0$ crossover’d configurations, to find the Pareto frontier of accuracy and efficiency. + +# 5 RESULTS AND ANALYSIS + +Efficiency-accuracy trade-off We use SQuAD 1.1 to examine the effect of the proposed approach on the efficiency-accuracy trade-off. When the underlying classifier was not trained with LengthDrop, as proposed in this paper, the accuracy drops even more dramatically as more word vectors are dropped at each layer. The difference between standard transformer and Length-Adaptive Transformer is stark in Fig. 2. This verifies the importance of training a transformer in a way that makes it malleable for inference-time re-configuration. + +When the model was trained with the proposed LengthDrop, we notice the efficacy of the proposed approach of using evolutionary search to find the optimal trade-off between inference efficiency and accuracy. The trade-off curve from the proposed search strategy has a larger area-under-curve (AUC) than when constant-rate length reduction was used to meet a target computational budget. It demonstrates the importance of using both LengthDrop and evolutionary search. + +We make a minor observation that the proposed approach ends up with a significantly higher accuracy than DistillBERT when enough computational budget is allowed for inference $( \log { \mathrm { F L O P s } } > 1 0 )$ . This makes our approach desirable in a wide array of scenarios, as it does not require any additional pretraining stage, as does DistilBERT. With a severe constraint on the computational budget, the proposed approach could be used on DistilBERT to significantly improve the efficiency without compromising the accuracy. + +Maximizing inference efficiency We consider all three tasks, SQuAD 1.1, MNLI-m and SST-2, and investigate how much efficiency can be gained by the proposed approach with minimal sacrifice of accuracy. First, we look at how much efficiency could be gained without losing on the accuracy. That is, we use the length configuration that maximizes the inference efficiency (i.e., minimize the FLOPs) while ensuring that the accuracy is above or same as the accuracy of the standard approach without any drop of word vectors. The results are presented in the rows marked with Length-Adaptive† from Table 1. For example, in the case of $\mathbf { B E R T _ { B a s e } }$ , the proposed approach reduce FLOPs by more than half across all three tasks. + +Table 1: Comparison results of standard Transformer and LengthAdaptive Transformer. Among length configurations on the Pareto frontier of Length-Adaptive Transformer, we pick two representative points: Length-Adaptive? and Length-Adaptive† as the most efficient one while having the highest accuracy and the accuracy higher than (or equal to) standard Transformer, respectively. + +
ModelSQuAD 1.1MNLI-mSST-2
Pretrained TransformerMethodF1FLOPsAccFLOPsAccFLOPs
BERTBaseStandard88.51.00x84.41.00x92.81.00x
Length-Adaptive*89.60.89x85.00.58x93.10.36x
Length-Adaptivet88.70.45x84.40.35x92.80.35x
Standard85.81.00x80.91.00x90.61.00x
DistilBERT1Length-Adaptive*86.30.81x81.50.56x92.00.55x
Length-Adaptive†85.90.59x81.30.54x91.70.54x
+ +![](images/e4a6ecfc1cdcda5b3130f2509429e44b544d994935e6b44f36893f811295f752.jpg) +Figure 4: Example of area under Pareto curve as the evolutionary search of lenth configurations proceeds. + +From Fig. 2, we have observed that the proposed Length-Adaptive Transformer generalize better than the standard, base model in some cases. We thus try to maximize both the inference efficiency and accuracy, in order to see whether it is possible for the proposed algorithm to find a length configuration that both maximizes inference efficiency and improves accuracy. We present the results in the rows marked with Length-Adaptive? from Table 1. For all cases, Length-Adaptive Transformer achieves higher accuracy than a standard transformer does while reducing FLOPs significantly. Although it is not apparent from the table, tor MNLI-m and SST-2, the accuracy of the smallest sub-model is already greater than or equal to that of a standard transformer. + +FLOPs vs. Latency As has been discussed in recent literature (see, e.g., (Li et al., 2020; Chin et al., 2020)), FLOPs is not a perfect indicator of the real latency measured in wall-clock time, as the latter is affected by the combination of hardware choice and network architecture. To understand the real-world impact of the proposed approach, we study the relationship between FLOPs, obtained by the proposed procedure, and wall-clock time measured on both CPU and GPU by measuring them while varying length configurations. As shown in Fig. 3, FLOPs and latency exhibit near-linear correlation on GPU, when the minibatch size is $\geq 1 6$ , and regardless of the minibatch size, on CPU. In other words, the reduction in FLOPs with the proposed approach directly implies the reduction in wall-clock time. + +Convergence of search Although the proposed approach is efficient in that it requires only one round of training, it needs a separate search stage for each target budget. It is important for evolutionary search to converge quickly in the number of forward sweeps of a validation set. As exemplified in Fig. 4, evolutionary search converges after about fifteen iterations. + +# 6 RELATED WORK + +The main purpose of the proposed algorithm is to improve the inference efficiency of a large-scale transformer. This goal has been pursued from various directions, and in this section, we provide a brief overview of these earlier, and some concurrent, attempts in the context of the proposed approach. + +Weight pruning Weight pruning (Han et al., 2015) focuses on reducing the number of parameters which directly reflects the memory footprint of a model and indirectly correlates with inference speed. However, their actual speed-up in runtime is usually not significant, especially while executing a model with parallel computation using GPU devices (Tang et al., 2018; Li et al., 2020). + +Adaptive architecture There are three major axes along which computation can be reduced in a neural network; (1) input size/length, (2) network depth and (3) network width. The proposed approach, based on PoWER-BERT, adaptively reduces the input length as the input sequence is processed by the transformer layers. In our knowledge, Goyal et al. (2020) is the first work in this direction for transformers. More recently, Funnel-Transformer (Dai et al., 2020) and multi-scale transformer language models (Subramanian et al., 2020) also successfully reduce sequence length in the middle and rescale to full length for the final computation. However, their inference complexity is fixed unlike PoWER-BERT because they are not designed for the control of efficiency. + +LayerDrop (Fan et al., 2019) drops random layers during the training to be robust to pruning inspired by Huang et al. (2016). Word-level adaptive depth in Elbayad et al. (2019) might seemingly resemble with length reduction, but word vectors reached the maximal layer are used for self-attention computation without updating themselves. Escaping a network early (Teerapittayanon et al., 2016; Huang et al., 2017) based on the confidence of the prediction (Xin et al., 2020; Schwartz et al., 2020; Liu et al., 2020) also offers a control over accuracy-efficiency trade-off, but it is difficult to tune a threshold for a desired computational budget because of the example-wise adaptive computation. + +DynaBERT (Hou et al., 2020) can run at adaptive width (the number of attention heads and intermediate hidden dimension) and depth. Hardware-aware Transformers (Wang et al., 2020) construct a design space with arbitrary encoder-decoder attention and heterogeneous layers in terms of different numbers of layers, attention heads, hidden dimension, and embedding dimension. + +Structured dropout A major innovation we introduce over the existing PoWER-BERT is the use of stochastic, structured regularization to make a transformer robust to the choice of length configuration in the inference time. Rippel et al. (2014) proposes a nested dropout to learn ordered representations. Similar to LengthDrop, it samples an index form a prior distribution and drops all units having a larger index than sampled one. + +Search There have been a series of attempts at finding the optimal network configuration by solving a combinatorial optimization problem. In computer vision, Once-for-All (Cai et al., 2019) use an evolutionary search (Real et al., 2019) to find a better configuration in dimensions of depth, width, kernel size, and resolution given computational budget. Similarly but differently, our evolutionary search is mutli-objective to find length configurations on the Pareto accuracy-efficiency frontier to cope with any possible computational budgets. Moreover, we only change the sequence length of hidden vectors instead of architectural model size like dimensions. + +# 7 CONCLUSION AND FUTURE WORK + +In this work, we propose a new framework for training a transformer once and using it for efficient inference under any computational budget. With the help of training with LengthDrop and Drop-andRestore process followed by the evolutionary search, our proposed Length-Adaptive Transformer allows any given transformer models to be used with any inference-time computational budget for both sequence-level and token-level classification tasks. Our experiments, on SQuAD 1.1, MNLI-m and SST-2, have revealed that the proposed algorithmic framework significantly pushes a better Pareto frontier on the trade-off between inference efficiency and accuracy. Furthermore, we have observed that the proposed Length-Adaptive Transformer could achieve up to $3 \mathbf { x }$ speed-up over the standard transformer without sacrificing accuracy, both in terms of FLOPs and wallclock time. + +Although our approach finds an optimal length configuration of a trained classifier per computational budget, it leaves open a question whether the proposed approach could be further extended to support per-instance length configuration by for instance training a small, auxiliary neural network for each computational budget. Yet another aspect we have not investigated in this paper is the applicability of the proposed approach to sequence generation, such as machine translation. We leave both of these research directions for the future. + +Our approach is effective, as we have shown in this paper, and also quite simple to implement on top of existing language models. We will release our implementation, which is based on HuggingFace’s Transformers library (Wolf et al., 2019), publicly and plan to adapt it for a broader set of transformerbased models and downstream tasks. + +# REFERENCES + +Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. + +Han Cai, Chuang Gan, and Song Han. Once for all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019. + +Ting-Wu Chin, Ruizhou Ding, Cha Zhang, and Diana Marculescu. Towards efficient model compression via learned global ranking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1518–1528, 2020. + +Zihang Dai, Guokun Lai, Yiming Yang, and Quoc V Le. Funnel-transformer: Filtering out sequential redundancy for efficient language processing. arXiv preprint arXiv:2006.03236, 2020. + +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. + +Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv preprint arXiv:1910.10073, 2019. + +Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. + +Saurabh Goyal, Anamitra Roy Choudhary, Venkatesan Chakaravarthy, Saurabh ManishRaje, Yogish Sabharwal, and Ashish Verma. Power-bert: Accelerating bert inference for classification tasks. arXiv preprint arXiv:2001.08950, 2020. + +Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. + +Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and Joelle Pineau. Towards the systematic reporting of the energy and carbon footprints of machine learning. arXiv preprint arXiv:2002.05651, 2020. + +Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. + +Lu Hou, Lifeng Shang, Xin Jiang, and Qun Liu. Dynabert: Dynamic bert with adaptive width and depth. arXiv preprint arXiv:2004.04037, 2020. + +Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016. + +Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. + +Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. + +Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E Gonzalez. Train large, then compress: Rethinking model size for efficient training and inference of transformers. arXiv preprint arXiv:2002.11794, 2020. + +Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Haotang Deng, and Qi Ju. Fastbert: a self-distilling bert with adaptive inference time. arXiv preprint arXiv:2004.02178, 2020. + +Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems, pp. 14014–14024, 2019. + +Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. arXiv preprint arXiv:1802.05365, 2018. + +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019. + +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019. + +Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: $^ { 1 0 0 , 0 0 0 + }$ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250, 2016. + +Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the aaai conference on artificial intelligence, volume 33, pp. 4780–4789, 2019. + +Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested dropout. In International Conference on Machine Learning, pp. 1746–1754, 2014. + +Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. + +Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. Green ai. arXiv preprint arXiv:1907.10597, 2019. + +Roy Schwartz, Gabi Stanovsky, Swabha Swayamdipta, Jesse Dodge, and Noah A Smith. The right tool for the job: Matching model and instance complexities. arXiv preprint arXiv:2004.07453, 2020. + +Abigail See, Minh-Thang Luong, and Christopher D Manning. Compression of neural machine translation models via pruning. arXiv preprint arXiv:1606.09274, 2016. + +Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using gpu model parallelism. arXiv preprint arXiv:1909.08053, 2019. + +Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014. + +Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in nlp. arXiv preprint arXiv:1906.02243, 2019. + +Sandeep Subramanian, Ronan Collobert, Marc’Aurelio Ranzato, and Y-Lan Boureau. Multi-scale transformer language models. arXiv preprint arXiv:2005.00581, 2020. + +Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355, 2019. + +Raphael Tang, Ashutosh Adhikari, and Jimmy Lin. Flops as a direct optimization objective for learning sparse neural networks. arXiv preprint arXiv:1811.03060, 2018. + +Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pp. 2464–2469. IEEE, 2016. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017. + +Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. + +Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han. Hat: Hardware-aware transformers for efficient natural language processing. arXiv preprint arXiv:2005.14187, 2020. + +Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art ´ natural language processing. arXiv preprint arXiv:1910.03771, 2019. + +Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. Deebert: Dynamic early exiting for accelerating bert inference. arXiv preprint arXiv:2004.12993, 2020. + +Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pp. 5754–5764, 2019. + +Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1803–1811, 2019. \ No newline at end of file diff --git a/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_content_list.json b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a2d84d854deee88724ebaadd6e1ab1505f6f2dd --- /dev/null +++ b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_content_list.json @@ -0,0 +1,1314 @@ +[ + { + "type": "text", + "text": "LENGTH-ADAPTIVE TRANSFORMER: TRAIN ONCE WITH LENGTH DROP, USE ANYTIME WITH SEARCH ", + "text_level": 1, + "bbox": [ + 174, + 98, + 823, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 174, + 398, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 238, + 544, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although transformers have achieved impressive accuracies in various tasks in natural language processing, they often come with a prohibitive computational cost, that prevents their use in scenarios with limited computational resources for inference. This need for computational efficiency in inference has been addressed by for instance PoWER-BERT (Goyal et al., 2020) which gradually decreases the length of a sequence as it is passed through layers. These approaches however often assume that the target computational complexity is known in advance at the time of training. This implies that a separate model must be trained for each inference scenario with its distinct computational budget. In this paper, we extend PoWER-BERT to address this issue of inefficiency and redundancy. The proposed extension enables us to train a large-scale transformer, called Length-Adaptive Transformer, once and uses it for various inference scenarios without re-training it. To do so, we train a transformer with LengthDrop, a structural variant of dropout, which stochastically determines the length of a sequence at each layer. We then use a multi-objective evolutionary search to find a length configuration that maximizes the accuracy and minimizes the computational complexity under any given computational budget. Additionally, we significantly extend the applicability of PoWER-BERT beyond sequence-level classification into token-level classification such as span-based question-answering, by introducing the idea of Drop-and-Restore. With Drop-and-Restore, word-vectors are dropped temporarily in intermediate layers and restored at the last layer if necessary. We empirically verify the utility of the proposed approach by demonstrating the superior accuracyefficiency trade-off under various setups, including SQuAD 1.1, MNLI-m, and SST-2. Upon publication, the code to reproduce our work will be open-sourced. ", + "bbox": [ + 232, + 270, + 766, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 627, + 336, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pretrained language models (Peters et al., 2018; Devlin et al., 2018; Radford et al., 2019; Yang et al., 2019) have achieved notable improvements in various natural language processing (NLP) tasks. Most of them rely on transformers (Vaswani et al., 2017), and the number of model parameters ranges from hundreds of millions to billions (Shoeybi et al., 2019; Raffel et al., 2019; Kaplan et al., 2020; Brown et al., 2020). Despite this high accuracy, excessive computational overhead during inference, both in terms of time and memory, has hindered its use in real applications. This level of excessive computation has further raised the concern over energy consumption as well (Schwartz et al., 2019; Strubell et al., 2019). ", + "bbox": [ + 174, + 659, + 826, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies have attempted at addressing these concerns regarding large-scale transformers’ computational and energy efficiency (see $\\ S 6$ for a more extensive discussion.) Among these, we focus on PoWER-BERT (Goyal et al., 2020) which progressively reduces sequence length by eliminating word-vectors based on the attention values as passing layers. PoWER-BERT establishes the superiority of accuracy-time trade-off over earlier approaches (Sanh et al., 2019; Sun et al., 2019; Michel et al., 2019). It however requires us to train a separate model for each efficiency constraint. In this paper, we thus develop a framework based on PoWER-BERT such that we can train a single model that can be adapted in the inference time to meet any given efficiency target. ", + "bbox": [ + 174, + 777, + 825, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to train a transformer to cope with a diverse set of computational budgets in the inference time, we propose to train one while reducing the sequence length with a random proportion at each layer. We refer to this procedure as LengthDrop which was motivated by the nested dropout (Rippel et al., 2014). We can extract sub-models of shared weights with any length configuration without requiring extra post-processing nor additional finetuning. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once a transformer is trained with the proposed LengthDrop, we search for the length configuration that maximizes the accuracy given a computational budget. Because this search is combinatorial and has multiple objectives (accuracy and efficiency), we use an evolutionary search algorithm, which further allows us to obtain a full Pareto frontier of accuracy-efficiency trade-off of each model. ", + "bbox": [ + 174, + 152, + 823, + 208 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is not trivial to find an optimal length configuration given the inference-time computational budget, although it is extremely important in order to deploy these large-scale transformers in practice. In this work, we propose to use evolutionary search to find a length configuration that maximizes the accuracy within a given computational budget. We can further compute the Pareto frontier of accuracy-efficiency trade-off to obtain a sequence of length configurations with varying efficiency profiles. ", + "bbox": [ + 174, + 215, + 825, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PoWER-BERT, which forms the foundation of the proposed two-stage procedure, is only applicable to sequence-level classification, because by design it eliminates some of the word vectors at each layer. In other words, it cannot be used for token-level tasks such as span-based question answering (Rajpurkar et al., 2016), because these tasks require hidden representations of the entire input sequence at the final layer. We thus propose to extend PoWER-BERT with a novel Drop-and-Restore process (§3.3), which eliminates this inherent limitation. Word vectors are dropped and set aside, rather than eliminated, in intermediate layers to maintain the saving of computational cost, as was with the original PoWER-BERT. These set-aside vectors are then restored at the final hidden layer and provided as an input to a subsequent task-specific layer, which is unlike the original PoWER-BERT. ", + "bbox": [ + 174, + 306, + 825, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are two-fold. First, we introduce LengthDrop, a structured variant of dropout for training a single Length-Adaptive Transformer model that allows us to automatically derive multiple sub-models with different length configurations in the inference time using evolutionary search, without requiring any re-training. Second, we design Drop-and-Restore process that makes PoWER-BERT applicable beyond classification, which enables PoWER-BERT to be applicable to a wider range of NLP tasks such as span-based question answering. We empirically verify Length-Adaptive Transformer works quite well using the variants of BERT on a diverse set of NLP tasks, including SQuAD 1.1 (Rajpurkar et al., 2016) and two sequence-level classification tasks in GLUE benchmark (Wang et al., 2018). Our experiments reveal that the proposed approach grants us a fine-grained control of computational efficiency and a superior accuracy-efficiency trade-off in the inference time, compared to existing approaches. ", + "bbox": [ + 174, + 438, + 825, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 BACKGROUND: TRANSFORMERS AND POWER-BERT ", + "text_level": 1, + "bbox": [ + 174, + 611, + 651, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before we describe our main approach, we review some of the building blocks in this section. In particular, we review transformers, which are a standard backbone used in natural language processing these days, and PoWER-BERT, which was recently proposed as an effective way to train a large-scale, but highly efficient transformer for sequence-level classification. ", + "bbox": [ + 174, + 643, + 825, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 TRANSFORMERS AND BERT ", + "text_level": 1, + "bbox": [ + 176, + 715, + 413, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A transformer is a particular neural network that has been designed to work with a variable-length sequence input and is implemented as a stack of self-attention and fully-connected layers (Vaswani et al., 2017). Here, we give a brief overview of the transformer which is the basic building block of the proposed approach. ", + "bbox": [ + 174, + 742, + 825, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Each token $x _ { t }$ in a sequence of tokens ${ \\boldsymbol x } = ( x _ { 1 } , \\dots , x _ { N } )$ , representing input text, is first turned into a continuous vector $h _ { t } ^ { 0 } \\in \\mathbb { R } ^ { H }$ which is the sum of the token and position embedding vectors. This sequence is fed into the first transformer layer which returns another sequence of the same length $\\dot { h ^ { 1 } } \\in \\mathbb { R } ^ { N \\times H }$ . We repeat this procedure $L$ times, for a transformer with $L$ layers, to obtain $h ^ { L } = ( h _ { 1 } ^ { L } , \\ldots , h _ { N } ^ { L } )$ . We refer to each vector in the hidden sequence at each layer as a word vector to emphasize that there exists a correspondence between each such vector and one of the input words. ", + "bbox": [ + 174, + 804, + 825, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although the transformer was first introduced for the problem of machine translation, Devlin et al. \n(2018) demonstrated that the transformer can be trained and used as a masked language model. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More specifically, Devlin et al. (2018) showed that the transformer-based masked language model, called BERT, learns a universally useful parameter set that can be finetuned for any downstream task including sequence-level and token-level classification. ", + "bbox": [ + 176, + 103, + 823, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the case of sequence-level classification, a softmax classifier is attached to the word vector $h _ { 1 } ^ { L }$ associated with the special token [CLS], and the entire network, including the softmax classifier and BERT, is finetuned. For token-level classification, we use each $h _ { t } ^ { L }$ as the final hidden reprsentation of the associated $t$ -th word in the input sequence. This strategy of pretraining followed by finetuning, often referred to as transfer learning, has recently become a dominant approach to classification in natural language processing. ", + "bbox": [ + 174, + 152, + 825, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 POWER-BERT ", + "text_level": 1, + "bbox": [ + 174, + 273, + 323, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PoWER-BERT keeps only the topmost $l _ { j }$ word vectors at each layer $j$ by eliminating redundant ones based on the significance score, which is the total amount of attention imposed by a word on the other words (Goyal et al., 2020). $l _ { j }$ is the hyper-parameter determines how many vectors to keep at layer $j$ PoWER-BERT has the same model parameters with BERT, but the extraction layers are interspersed after the self-attention layer in every transformer blocks (Vaswani et al., 2017). ", + "bbox": [ + 174, + 308, + 825, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PoWER-BERT reduces inference time successfully, achieving better accuracy-time trade-off than DistilBERT (Sanh et al., 2019), BERT-PKD (Sun et al., 2019), and Head-Prune (Michel et al., 2019). Despite the original intention of maximizing the inference efficiency with the minimal loss in accuracy, it is possible to set up PoWER-BERT to be both more efficient and more accurate compared to the original BERT, which was observed but largely overlooked by Goyal et al. (2020). ", + "bbox": [ + 174, + 385, + 825, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training a PoWER-BERT model consists of three steps: (1) finetuning, (2) length configuration search, and (3) re-training. The finetuning step is just like the standard finetuning step of BERT given a target task. A length configuration is a sequence of retention parameters $\\left( l _ { 1 } , \\cdots l _ { L } \\right)$ , each of which corresponds to the number of word vectors that are kept at each layer. These retention parameters are learned along with all the other parameters to minimize the original task loss together with an extra term that approximately measures the number of retained word vectors across layers. In the re-training step, PoWER-BERT is finetuned with the length configuration fixed to its learned one. ", + "bbox": [ + 174, + 462, + 825, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each computational budget, we must train a separate model going through all three steps described above. Moreover, the length configuration search step above is only approximate, as it relies on relaxation of retention parameters which are inherently discrete. This leads to the lack of guaranteed correlation between the success of this stage and true run-time. Even worse, it is a delicate act to tune the length configuration given a target computational budget, because trade-off is implicitly made via a regularization coefficient. Furthermore, PoWER-BERT has an inherent limitation in that it only applies to sequence-level classification because it eliminates word vectors in intermediate layers. ", + "bbox": [ + 174, + 566, + 825, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 LENGTH-ADAPTIVE TRANSFORMER ", + "text_level": 1, + "bbox": [ + 174, + 705, + 504, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we explain our proposed framework which results in a transformer that reduces the length of a sequence at each layer with an arbitrary rate. We call such a resulting transformer a Length-Adaptive Transformer. We train Length-Adaptive Transformer with LengthDrop which randomly samples the number of hidden vectors to be dropped at each layer with the goal of making the final model robust to such drop in the inference time. Once the model is trained, we search for the optimal trade-off between accuracy and efficiency using multi-objective evolutionary search, which allows us to use the model for any given computational budget without finetuning nor re-training. At the end of this section, we describe Drop-and-Restore process as a way to greatly increase the applicability of PoWER-BERT which forms a building block of the proposed framework. ", + "bbox": [ + 174, + 748, + 825, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In short, we train a Length-Adaptive Transformer once with LengthDrop and Drop-and-Restore, and use it with an automatically determined length configuration for inference with any target computational budget, on both sequence-level and token-level tasks. ", + "bbox": [ + 176, + 882, + 823, + 922 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 LENGTHDROP ", + "text_level": 1, + "bbox": [ + 174, + 103, + 313, + 117 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Earlier approaches to efficient inference with transformers have focused on a scenario where the target computational budget for inference is known in advance Sanh et al. (2019); Goyal et al. (2020). This greatly increases the cost of deploying transformers, as it requires us to train a separate transformer for each scenario. Instead, we propose to train one model that could be used for a diverse set of target computational budgets without retraining. ", + "bbox": [ + 174, + 128, + 825, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LengthDrop randomly generates a length configuration by sequentially sampling a sequence length $l _ { i + 1 }$ at the $( i + 1 )$ -th layer based on the previous layer’s sequence length $l _ { i }$ , following the uniform distribution $\\mathcal { U } ( \\lceil ( 1 - p ) l _ { i } \\rceil , l _ { i } )$ , where $l _ { 0 }$ is set to the length of the input sequence, and $p$ is the LengthDrop probability. This sequential sampling results in a length configuration $( l _ { 1 } , \\cdots , l _ { L } )$ . Length-Adaptive Transformer can be thought of as consisting of a full model and many sub-models corresponding to different length configuration, similarly to a neural network trained with dropout (Srivastava et al., 2014). ", + "bbox": [ + 173, + 207, + 825, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LayerDrop From the perspective of each word vector, the proposed LengthDrop could be thought of as skipping the layers between when it was set aside and the final layer where it was restored. The word vector however does not have any information based on which it can determine whether it would be dropped at any particular layer. In our preliminary experiments, we found that this greatly hinders optimization. We address this issue by using LayerDrop (Fan et al., 2019) which skips each layer of a transformer uniformly at random. The LayerDrop encourages each word vector to be agnostic to skipping any number of layers between when it is dropped and when it is restored, just like dropout (Srivastava et al., 2014) prevents hidden neurons from co-adapting with each other by randomly dropping them. ", + "bbox": [ + 173, + 318, + 825, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sandwich Rule and Inplace Distillation We observed that standard supervised training with LengthDrop does not work well in the preliminary experiments. We instead borrow a pair of training techniques developed by Yu & Huang (2019) which are sandwich rule and inplace distillation, for better optimization as well as final generalization. At each update, we update the full model without LengthDrop as usual to minimize the supervised loss function. We simultaneously update $n _ { s }$ randomly-sampled sub-models (which are called sandwiches) and the smallest-possible submodel, which corresponds to keeping only $\\lceil ( 1 - p ) l _ { i } \\rceil$ word vectors at each layer $i$ , using knowledge distillation (Hinton et al., 2015) from the full model. Here, sub-models mean models with length reduction. They are trained to their prediction close to the full model’s prediction (inplace distillation). ", + "bbox": [ + 174, + 457, + 825, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 EVOLUTIONARY SEARCH OF LENGTH CONFIGURATIONS ", + "text_level": 1, + "bbox": [ + 176, + 598, + 604, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After training a Length-Adaptive Transformer with LengthDrop, we search for appropriate length configurations for possible target computational budgets that will be given at inference time. The length configuration determines the model performance in terms of both accuracy and efficiency. In order to search for the optimal length configuration, we propose to use evolutionary search, similarly to Cai et al. (2019) and Wang et al. (2020). This procedure is efficient, as it only requires a single pass through the relatively small validation set for each length configuration, unlike re-training for a new computational budget which requires multiple passes through a significantly larger training set for each budget. ", + "bbox": [ + 174, + 625, + 825, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We initialize the population with constant-ratio configurations. Each configuration is created by $l _ { i + 1 } = \\lceil ( 1 - r ) \\bar { l _ { i } } \\rceil$ for each layer $i$ with $r$ so that the amount of computation within the initial population is uniformly distributed between those of the smallest and full models. At each iteration, we evolve the population to consist only of configurations lie on a newly updated efficiencyaccuracy Pareto frontier by mutation and cross-over. Mutation alters an original length configuration $( l _ { 1 } , \\cdots , l _ { L } )$ to $( l _ { 1 } ^ { \\prime } , \\cdots , l _ { L } ^ { \\prime } )$ by sampling $l _ { i } ^ { \\prime }$ from the uniform distribution $\\bar { \\mathcal { U } } ( l _ { i - 1 } ^ { \\prime } , l _ { i + 1 } )$ with the probability $p _ { m }$ or keeping the original length ${ \\bar { l } } _ { i } ^ { \\prime } = l _ { i }$ , sweeping the layers from $i = 1$ to $i = L$ . Crossover takes two length configurations and averages the lengths at each layer. Both of these operations are performed while ensuring the monotonicity of the lengths over the layers. We repeat this iteration $G$ times, while maintaining $n _ { m }$ mutated configurations and $n _ { c }$ crossover’d configurations. Repeating this procedure pushes the Pareto frontier further to identify the best trade-off between two objectives, efficiency and accuracy, without requiring any continuous relaxation of length configurations nor using a proxy objective function. ", + "bbox": [ + 173, + 743, + 826, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/320388afaf5c944756a7b3f52eddc7583af0fdd53b75f8ac7bf8c3063dd11f2f.jpg", + "image_caption": [ + "Figure 1: Illustration of (a) word-vector elimination process in PoWER-BERT (Goyal et al., 2020) and (b) Drop-and-Restore process in Length-Adaptive Transformer. Yellow box and blue boxes imply the output of embedding layer and transformer layers, respectively. Green boxes mean vectors dropped in lower layers and restored at the last layer. Red box is the task-specific layer. Though word-vectors in the middle could be eliminated (or dropped), remaining vectors are left-aligned for the better illustration. In this case, the number of transformer layers is four. " + ], + "image_footnote": [], + "bbox": [ + 194, + 111, + 803, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 DROP-AND-RESTORE PROCESS ", + "text_level": 1, + "bbox": [ + 176, + 314, + 429, + 328 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The applicability of the PoWER-BERT, based on which our main contribution above was made, is limited to sequence-level classification, because it eliminates word vectors at each layer. In addition to our main contribution above, we thus propose to extend the PoWER-BERT so that it is applicable to token-level classification, such as span-based question-answering. Our proposal, to which we refer as Drop-and-Restore, does not eliminate word vectors at each layer according to the length configuration but instead sets them aside until the final hidden layer. At the final hidden layer, these word vectors are brought back to form the full hidden sequence, as illustrated graphically in Fig. 1. ", + "bbox": [ + 174, + 340, + 825, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENT SETUP ", + "text_level": 1, + "bbox": [ + 176, + 459, + 377, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We test the proposed approach on both sequence-level and token-level tasks, the latter of which could not have been done with the original PoWER-BERT unless for the proposed Dropand-Restore. We use MNLI-m and SST-2 from GLUE benchmark (Wang et al., 2018), as was done to test PoWER-BERT earlier, for sequence-level classification. We choose them because consistent accuracy scores from standard training on them due to their sufficiently large training set imply that they are reliable to verify our approach. We use SQuAD 1.1 (Rajpurkar et al., 2016) for token-level classification. ", + "bbox": [ + 174, + 491, + 825, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics We use the number of floating operations (FLOPs) as a main metric to measure the inference efficiency given any length configuration, as it is agnostic to the choice of underlying hardware, unlike other alternatives such as hardware-aware latency (Wang et al., 2020) or energy consumption (Henderson et al., 2020). We later demonstrate that FLOPs and wall-clock time on GPU and CPU correlate well with the proposed approach, which is not necessarily the case for other approaches, such as unstructured weight pruning (Han et al., 2015; See et al., 2016). ", + "bbox": [ + 174, + 604, + 825, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pretrained transformers Since BERT was introduced by Devlin et al. (2018), it has become a standard practice to start from a pretrained (masked) language model and finetune it for each downstream task. We follow the same strategy in this paper and test two pretrained transformerbased language models; BERTBASE (Devlin et al., 2018) and DistilBERT (Sanh et al., 2019), which allows us to demonstrate that the usefulness and applicability of our approach are not tied to any specific architectural choice, such as the number of layers and the maximum length of input sequence. Although we focus on BERT-based masked language models here, the proposed approach is readily applicable to any transformer-based models. ", + "bbox": [ + 174, + 704, + 825, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning We train a Length-Adaptive Transformer with LengthDrop probability and LayerDrop probability both set to 0.2. We use $n _ { s } = 2$ randomly sampled intermediate sub-models in addition to the full model and smallest model for applying the sandwich learning rule. ", + "bbox": [ + 174, + 833, + 825, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start finetuning the pretrained transformer without Drop-and-Restore first, just as Goyal et al. (2020) did with PoWER-BERT. We then continue finetuning it for another five epochs with Drop-andRestore. This is unlike the recommended three epochs by Devlin et al. (2018), as learning progresses slower due to a higher level of stochasticity introduced by LengthDrop and LayerDrop. We use the batch size of 32, the learning rate of $5 e - 5$ for SQuAD v1.1 and $2 e - 5$ for MNLI-m and SST, and the maximum sequence length of 384 for SQuAD v1.1 and 128 for MNLI-m and SST. ", + "bbox": [ + 176, + 882, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ea54ad6f6a4d01240155dd50bd06714bd078c67453a78cfd7ce1b9404ef27c2.jpg", + "image_caption": [ + "Figure 2: Pareto curves of F1 score and FLOPs on SQuAD 1.1 (Rajpurkar et al., 2016). We apply the proposed method to $\\mathbf { B E R T _ { B a s e } }$ (solid lines) and DistilBERT (dotted lines). For each model, we draw three curves using (1) standard finetuned transformer with constant-rate length reduction, (2) Length-Adaptive Transformer with constant-rate length reduction, and (3) Length-Adaptive Transformer with length configurations obtained from the evolutionary search. " + ], + "image_footnote": [], + "bbox": [ + 181, + 99, + 638, + 349 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4608176aa7c7dba0b4748567c1c64abe798b711bebf826c4b323894a4bfb2b12.jpg", + "image_caption": [ + "Figure 3: Correlation between FLOPs and latency with different length configurations. " + ], + "image_footnote": [], + "bbox": [ + 666, + 102, + 812, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 473, + 823, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Search We run up to $G = 3 0$ iterations of evolutionary search, using $n _ { m } = 3 0$ mutated configurations with mutation probability $p _ { m } = 0 . 5$ and $n _ { c } = 3 0$ crossover’d configurations, to find the Pareto frontier of accuracy and efficiency. ", + "bbox": [ + 174, + 532, + 825, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 RESULTS AND ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 595, + 413, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficiency-accuracy trade-off We use SQuAD 1.1 to examine the effect of the proposed approach on the efficiency-accuracy trade-off. When the underlying classifier was not trained with LengthDrop, as proposed in this paper, the accuracy drops even more dramatically as more word vectors are dropped at each layer. The difference between standard transformer and Length-Adaptive Transformer is stark in Fig. 2. This verifies the importance of training a transformer in a way that makes it malleable for inference-time re-configuration. ", + "bbox": [ + 174, + 627, + 825, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When the model was trained with the proposed LengthDrop, we notice the efficacy of the proposed approach of using evolutionary search to find the optimal trade-off between inference efficiency and accuracy. The trade-off curve from the proposed search strategy has a larger area-under-curve (AUC) than when constant-rate length reduction was used to meet a target computational budget. It demonstrates the importance of using both LengthDrop and evolutionary search. ", + "bbox": [ + 174, + 718, + 825, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We make a minor observation that the proposed approach ends up with a significantly higher accuracy than DistillBERT when enough computational budget is allowed for inference $( \\log { \\mathrm { F L O P s } } > 1 0 )$ . This makes our approach desirable in a wide array of scenarios, as it does not require any additional pretraining stage, as does DistilBERT. With a severe constraint on the computational budget, the proposed approach could be used on DistilBERT to significantly improve the efficiency without compromising the accuracy. ", + "bbox": [ + 174, + 795, + 825, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Maximizing inference efficiency We consider all three tasks, SQuAD 1.1, MNLI-m and SST-2, and investigate how much efficiency can be gained by the proposed approach with minimal sacrifice of accuracy. First, we look at how much efficiency could be gained without losing on the accuracy. That is, we use the length configuration that maximizes the inference efficiency (i.e., minimize the FLOPs) while ensuring that the accuracy is above or same as the accuracy of the standard approach without any drop of word vectors. The results are presented in the rows marked with Length-Adaptive† from Table 1. For example, in the case of $\\mathbf { B E R T _ { B a s e } }$ , the proposed approach reduce FLOPs by more than half across all three tasks. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0829b1fbd9636baa6152ad5a45e5f7dbf08c6a7333b2fd362453dfa90284833.jpg", + "table_caption": [ + "Table 1: Comparison results of standard Transformer and LengthAdaptive Transformer. Among length configurations on the Pareto frontier of Length-Adaptive Transformer, we pick two representative points: Length-Adaptive? and Length-Adaptive† as the most efficient one while having the highest accuracy and the accuracy higher than (or equal to) standard Transformer, respectively. " + ], + "table_footnote": [], + "table_body": "
ModelSQuAD 1.1MNLI-mSST-2
Pretrained TransformerMethodF1FLOPsAccFLOPsAccFLOPs
BERTBaseStandard88.51.00x84.41.00x92.81.00x
Length-Adaptive*89.60.89x85.00.58x93.10.36x
Length-Adaptivet88.70.45x84.40.35x92.80.35x
Standard85.81.00x80.91.00x90.61.00x
DistilBERT1Length-Adaptive*86.30.81x81.50.56x92.00.55x
Length-Adaptive†85.90.59x81.30.54x91.70.54x
", + "bbox": [ + 183, + 101, + 643, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4a6ecfc1cdcda5b3130f2509429e44b544d994935e6b44f36893f811295f752.jpg", + "image_caption": [ + "Figure 4: Example of area under Pareto curve as the evolutionary search of lenth configurations proceeds. " + ], + "image_footnote": [], + "bbox": [ + 669, + 108, + 813, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 344, + 826, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Fig. 2, we have observed that the proposed Length-Adaptive Transformer generalize better than the standard, base model in some cases. We thus try to maximize both the inference efficiency and accuracy, in order to see whether it is possible for the proposed algorithm to find a length configuration that both maximizes inference efficiency and improves accuracy. We present the results in the rows marked with Length-Adaptive? from Table 1. For all cases, Length-Adaptive Transformer achieves higher accuracy than a standard transformer does while reducing FLOPs significantly. Although it is not apparent from the table, tor MNLI-m and SST-2, the accuracy of the smallest sub-model is already greater than or equal to that of a standard transformer. ", + "bbox": [ + 173, + 435, + 825, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FLOPs vs. Latency As has been discussed in recent literature (see, e.g., (Li et al., 2020; Chin et al., 2020)), FLOPs is not a perfect indicator of the real latency measured in wall-clock time, as the latter is affected by the combination of hardware choice and network architecture. To understand the real-world impact of the proposed approach, we study the relationship between FLOPs, obtained by the proposed procedure, and wall-clock time measured on both CPU and GPU by measuring them while varying length configurations. As shown in Fig. 3, FLOPs and latency exhibit near-linear correlation on GPU, when the minibatch size is $\\geq 1 6$ , and regardless of the minibatch size, on CPU. In other words, the reduction in FLOPs with the proposed approach directly implies the reduction in wall-clock time. ", + "bbox": [ + 174, + 561, + 825, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Convergence of search Although the proposed approach is efficient in that it requires only one round of training, it needs a separate search stage for each target budget. It is important for evolutionary search to converge quickly in the number of forward sweeps of a validation set. As exemplified in Fig. 4, evolutionary search converges after about fifteen iterations. ", + "bbox": [ + 174, + 703, + 825, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 779, + 344, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The main purpose of the proposed algorithm is to improve the inference efficiency of a large-scale transformer. This goal has been pursued from various directions, and in this section, we provide a brief overview of these earlier, and some concurrent, attempts in the context of the proposed approach. ", + "bbox": [ + 174, + 810, + 825, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Weight pruning Weight pruning (Han et al., 2015) focuses on reducing the number of parameters which directly reflects the memory footprint of a model and indirectly correlates with inference speed. However, their actual speed-up in runtime is usually not significant, especially while executing a model with parallel computation using GPU devices (Tang et al., 2018; Li et al., 2020). ", + "bbox": [ + 176, + 867, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive architecture There are three major axes along which computation can be reduced in a neural network; (1) input size/length, (2) network depth and (3) network width. The proposed approach, based on PoWER-BERT, adaptively reduces the input length as the input sequence is processed by the transformer layers. In our knowledge, Goyal et al. (2020) is the first work in this direction for transformers. More recently, Funnel-Transformer (Dai et al., 2020) and multi-scale transformer language models (Subramanian et al., 2020) also successfully reduce sequence length in the middle and rescale to full length for the final computation. However, their inference complexity is fixed unlike PoWER-BERT because they are not designed for the control of efficiency. ", + "bbox": [ + 174, + 103, + 825, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LayerDrop (Fan et al., 2019) drops random layers during the training to be robust to pruning inspired by Huang et al. (2016). Word-level adaptive depth in Elbayad et al. (2019) might seemingly resemble with length reduction, but word vectors reached the maximal layer are used for self-attention computation without updating themselves. Escaping a network early (Teerapittayanon et al., 2016; Huang et al., 2017) based on the confidence of the prediction (Xin et al., 2020; Schwartz et al., 2020; Liu et al., 2020) also offers a control over accuracy-efficiency trade-off, but it is difficult to tune a threshold for a desired computational budget because of the example-wise adaptive computation. ", + "bbox": [ + 174, + 222, + 825, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DynaBERT (Hou et al., 2020) can run at adaptive width (the number of attention heads and intermediate hidden dimension) and depth. Hardware-aware Transformers (Wang et al., 2020) construct a design space with arbitrary encoder-decoder attention and heterogeneous layers in terms of different numbers of layers, attention heads, hidden dimension, and embedding dimension. ", + "bbox": [ + 174, + 327, + 825, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Structured dropout A major innovation we introduce over the existing PoWER-BERT is the use of stochastic, structured regularization to make a transformer robust to the choice of length configuration in the inference time. Rippel et al. (2014) proposes a nested dropout to learn ordered representations. Similar to LengthDrop, it samples an index form a prior distribution and drops all units having a larger index than sampled one. ", + "bbox": [ + 174, + 397, + 825, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Search There have been a series of attempts at finding the optimal network configuration by solving a combinatorial optimization problem. In computer vision, Once-for-All (Cai et al., 2019) use an evolutionary search (Real et al., 2019) to find a better configuration in dimensions of depth, width, kernel size, and resolution given computational budget. Similarly but differently, our evolutionary search is mutli-objective to find length configurations on the Pareto accuracy-efficiency frontier to cope with any possible computational budgets. Moreover, we only change the sequence length of hidden vectors instead of architectural model size like dimensions. ", + "bbox": [ + 174, + 482, + 825, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7 CONCLUSION AND FUTURE WORK ", + "text_level": 1, + "bbox": [ + 176, + 601, + 495, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a new framework for training a transformer once and using it for efficient inference under any computational budget. With the help of training with LengthDrop and Drop-andRestore process followed by the evolutionary search, our proposed Length-Adaptive Transformer allows any given transformer models to be used with any inference-time computational budget for both sequence-level and token-level classification tasks. Our experiments, on SQuAD 1.1, MNLI-m and SST-2, have revealed that the proposed algorithmic framework significantly pushes a better Pareto frontier on the trade-off between inference efficiency and accuracy. Furthermore, we have observed that the proposed Length-Adaptive Transformer could achieve up to $3 \\mathbf { x }$ speed-up over the standard transformer without sacrificing accuracy, both in terms of FLOPs and wallclock time. ", + "bbox": [ + 174, + 632, + 825, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our approach finds an optimal length configuration of a trained classifier per computational budget, it leaves open a question whether the proposed approach could be further extended to support per-instance length configuration by for instance training a small, auxiliary neural network for each computational budget. Yet another aspect we have not investigated in this paper is the applicability of the proposed approach to sequence generation, such as machine translation. We leave both of these research directions for the future. ", + "bbox": [ + 174, + 765, + 825, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach is effective, as we have shown in this paper, and also quite simple to implement on top of existing language models. We will release our implementation, which is based on HuggingFace’s Transformers library (Wolf et al., 2019), publicly and plan to adapt it for a broader set of transformerbased models and downstream tasks. ", + "bbox": [ + 176, + 854, + 825, + 910 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 287, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. ", + "bbox": [ + 176, + 126, + 825, + 169 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Han Cai, Chuang Gan, and Song Han. Once for all: Train one network and specialize it for efficient deployment. arXiv preprint arXiv:1908.09791, 2019. ", + "bbox": [ + 171, + 179, + 823, + 208 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ting-Wu Chin, Ruizhou Ding, Cha Zhang, and Diana Marculescu. Towards efficient model compression via learned global ranking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1518–1528, 2020. ", + "bbox": [ + 174, + 218, + 823, + 262 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zihang Dai, Guokun Lai, Yiming Yang, and Quoc V Le. Funnel-transformer: Filtering out sequential redundancy for efficient language processing. arXiv preprint arXiv:2006.03236, 2020. ", + "bbox": [ + 173, + 272, + 823, + 303 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. ", + "bbox": [ + 173, + 311, + 823, + 342 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv preprint arXiv:1910.10073, 2019. ", + "bbox": [ + 173, + 351, + 823, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556, 2019. ", + "bbox": [ + 169, + 390, + 825, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Saurabh Goyal, Anamitra Roy Choudhary, Venkatesan Chakaravarthy, Saurabh ManishRaje, Yogish Sabharwal, and Ashish Verma. Power-bert: Accelerating bert inference for classification tasks. arXiv preprint arXiv:2001.08950, 2020. ", + "bbox": [ + 176, + 430, + 823, + 473 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. ", + "bbox": [ + 173, + 483, + 821, + 513 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and Joelle Pineau. Towards the systematic reporting of the energy and carbon footprints of machine learning. arXiv preprint arXiv:2002.05651, 2020. ", + "bbox": [ + 174, + 522, + 823, + 566 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. ", + "bbox": [ + 171, + 577, + 823, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Lu Hou, Lifeng Shang, Xin Jiang, and Qun Liu. Dynabert: Dynamic bert with adaptive width and depth. arXiv preprint arXiv:2004.04037, 2020. ", + "bbox": [ + 171, + 616, + 825, + 645 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016. ", + "bbox": [ + 171, + 655, + 823, + 685 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Q Weinberger. Multi-scale dense networks for resource efficient image classification. arXiv preprint arXiv:1703.09844, 2017. ", + "bbox": [ + 173, + 695, + 825, + 738 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. ", + "bbox": [ + 171, + 748, + 826, + 791 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E Gonzalez. Train large, then compress: Rethinking model size for efficient training and inference of transformers. arXiv preprint arXiv:2002.11794, 2020. ", + "bbox": [ + 173, + 803, + 823, + 844 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Haotang Deng, and Qi Ju. Fastbert: a self-distilling bert with adaptive inference time. arXiv preprint arXiv:2004.02178, 2020. ", + "bbox": [ + 171, + 856, + 823, + 885 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems, pp. 14014–14024, 2019. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. arXiv preprint arXiv:1802.05365, 2018. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019. ", + "bbox": [ + 173, + 154, + 823, + 184 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683, 2019. ", + "bbox": [ + 178, + 191, + 823, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: $^ { 1 0 0 , 0 0 0 + }$ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250, 2016. ", + "bbox": [ + 171, + 242, + 825, + 271 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image classifier architecture search. In Proceedings of the aaai conference on artificial intelligence, volume 33, pp. 4780–4789, 2019. ", + "bbox": [ + 176, + 279, + 825, + 321 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested dropout. In International Conference on Machine Learning, pp. 1746–1754, 2014. ", + "bbox": [ + 173, + 329, + 823, + 359 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. ", + "bbox": [ + 171, + 367, + 823, + 396 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. Green ai. arXiv preprint arXiv:1907.10597, 2019. ", + "bbox": [ + 171, + 404, + 823, + 433 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Roy Schwartz, Gabi Stanovsky, Swabha Swayamdipta, Jesse Dodge, and Noah A Smith. The right tool for the job: Matching model and instance complexities. arXiv preprint arXiv:2004.07453, 2020. ", + "bbox": [ + 174, + 440, + 823, + 483 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Abigail See, Minh-Thang Luong, and Christopher D Manning. Compression of neural machine translation models via pruning. arXiv preprint arXiv:1606.09274, 2016. ", + "bbox": [ + 171, + 492, + 823, + 521 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using gpu model parallelism. arXiv preprint arXiv:1909.08053, 2019. ", + "bbox": [ + 178, + 529, + 821, + 571 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014. ", + "bbox": [ + 176, + 579, + 823, + 623 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep learning in nlp. arXiv preprint arXiv:1906.02243, 2019. ", + "bbox": [ + 169, + 631, + 823, + 660 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sandeep Subramanian, Ronan Collobert, Marc’Aurelio Ranzato, and Y-Lan Boureau. Multi-scale transformer language models. arXiv preprint arXiv:2005.00581, 2020. ", + "bbox": [ + 173, + 667, + 823, + 696 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. Patient knowledge distillation for bert model compression. arXiv preprint arXiv:1908.09355, 2019. ", + "bbox": [ + 173, + 704, + 823, + 734 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Raphael Tang, Ashutosh Adhikari, and Jimmy Lin. Flops as a direct optimization objective for learning sparse neural networks. arXiv preprint arXiv:1811.03060, 2018. ", + "bbox": [ + 171, + 742, + 823, + 771 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR), pp. 2464–2469. IEEE, 2016. ", + "bbox": [ + 176, + 780, + 823, + 821 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998–6008, 2017. ", + "bbox": [ + 174, + 830, + 823, + 873 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. ", + "bbox": [ + 174, + 882, + 823, + 922 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han. Hat: Hardware-aware transformers for efficient natural language processing. arXiv preprint arXiv:2005.14187, 2020. ", + "bbox": [ + 174, + 103, + 825, + 145 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art ´ natural language processing. arXiv preprint arXiv:1910.03771, 2019. ", + "bbox": [ + 174, + 155, + 823, + 198 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. Deebert: Dynamic early exiting for accelerating bert inference. arXiv preprint arXiv:2004.12993, 2020. ", + "bbox": [ + 171, + 205, + 825, + 234 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pp. 5754–5764, 2019. ", + "bbox": [ + 174, + 244, + 823, + 287 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1803–1811, 2019. ", + "bbox": [ + 174, + 295, + 823, + 324 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_middle.json b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..b45587731e2f902f49fb6ba6a8beff1c16ab8e59 --- /dev/null +++ b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_middle.json @@ -0,0 +1,28224 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 504, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 505, + 96 + ], + "score": 1.0, + "content": "LENGTH-ADAPTIVE TRANSFORMER: TRAIN ONCE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 485, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 485, + 117 + ], + "score": 1.0, + "content": "WITH LENGTH DROP, USE ANYTIME WITH SEARCH", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 138, + 244, + 159 + ], + "lines": [ + { + "bbox": [ + 113, + 138, + 201, + 150 + ], + "spans": [ + { + "bbox": [ + 113, + 138, + 201, + 150 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 148, + 245, + 161 + ], + "spans": [ + { + "bbox": [ + 112, + 148, + 245, + 161 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 189, + 333, + 201 + ], + "lines": [ + { + "bbox": [ + 276, + 187, + 336, + 203 + ], + "spans": [ + { + "bbox": [ + 276, + 187, + 336, + 203 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 142, + 214, + 469, + 477 + ], + "lines": [ + { + "bbox": [ + 142, + 214, + 469, + 226 + ], + "spans": [ + { + "bbox": [ + 142, + 214, + 469, + 226 + ], + "score": 1.0, + "content": "Although transformers have achieved impressive accuracies in various tasks in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 225, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 469, + 236 + ], + "score": 1.0, + "content": "natural language processing, they often come with a prohibitive computational", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 236, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 470, + 247 + ], + "score": 1.0, + "content": "cost, that prevents their use in scenarios with limited computational resources for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "score": 1.0, + "content": "inference. This need for computational efficiency in inference has been addressed", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 469, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 269 + ], + "score": 1.0, + "content": "by for instance PoWER-BERT (Goyal et al., 2020) which gradually decreases the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 280 + ], + "score": 1.0, + "content": "length of a sequence as it is passed through layers. These approaches however", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 470, + 291 + ], + "score": 1.0, + "content": "often assume that the target computational complexity is known in advance at", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "the time of training. This implies that a separate model must be trained for each", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 302, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 469, + 313 + ], + "score": 1.0, + "content": "inference scenario with its distinct computational budget. In this paper, we extend", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "score": 1.0, + "content": "PoWER-BERT to address this issue of inefficiency and redundancy. The proposed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "extension enables us to train a large-scale transformer, called Length-Adaptive", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "score": 1.0, + "content": "Transformer, once and uses it for various inference scenarios without re-training", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 140, + 344, + 470, + 357 + ], + "spans": [ + { + "bbox": [ + 140, + 344, + 470, + 357 + ], + "score": 1.0, + "content": "it. To do so, we train a transformer with LengthDrop, a structural variant of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 470, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 470, + 368 + ], + "score": 1.0, + "content": "dropout, which stochastically determines the length of a sequence at each layer.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "score": 1.0, + "content": "We then use a multi-objective evolutionary search to find a length configuration", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "score": 1.0, + "content": "that maximizes the accuracy and minimizes the computational complexity under", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "score": 1.0, + "content": "any given computational budget. Additionally, we significantly extend the appli-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 400, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 142, + 400, + 470, + 412 + ], + "score": 1.0, + "content": "cability of PoWER-BERT beyond sequence-level classification into token-level", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 411, + 470, + 424 + ], + "spans": [ + { + "bbox": [ + 141, + 411, + 470, + 424 + ], + "score": 1.0, + "content": "classification such as span-based question-answering, by introducing the idea of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "score": 1.0, + "content": "Drop-and-Restore. With Drop-and-Restore, word-vectors are dropped temporarily", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 432, + 469, + 446 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 446 + ], + "score": 1.0, + "content": "in intermediate layers and restored at the last layer if necessary. We empirically", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 443, + 470, + 457 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 470, + 457 + ], + "score": 1.0, + "content": "verify the utility of the proposed approach by demonstrating the superior accuracy-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 455, + 470, + 466 + ], + "spans": [ + { + "bbox": [ + 142, + 455, + 470, + 466 + ], + "score": 1.0, + "content": "efficiency trade-off under various setups, including SQuAD 1.1, MNLI-m, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 465, + 462, + 478 + ], + "spans": [ + { + "bbox": [ + 142, + 465, + 462, + 478 + ], + "score": 1.0, + "content": "SST-2. Upon publication, the code to reproduce our work will be open-sourced.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 206, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 506, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "score": 1.0, + "content": "Pretrained language models (Peters et al., 2018; Devlin et al., 2018; Radford et al., 2019; Yang et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "2019) have achieved notable improvements in various natural language processing (NLP) tasks. Most", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "score": 1.0, + "content": "of them rely on transformers (Vaswani et al., 2017), and the number of model parameters ranges", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "score": 1.0, + "content": "from hundreds of millions to billions (Shoeybi et al., 2019; Raffel et al., 2019; Kaplan et al., 2020;", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 565, + 507, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 565, + 507, + 579 + ], + "score": 1.0, + "content": "Brown et al., 2020). Despite this high accuracy, excessive computational overhead during inference,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "score": 1.0, + "content": "both in terms of time and memory, has hindered its use in real applications. This level of excessive", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "score": 1.0, + "content": "computation has further raised the concern over energy consumption as well (Schwartz et al., 2019;", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 599, + 193, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 193, + 611 + ], + "score": 1.0, + "content": "Strubell et al., 2019).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 507, + 630 + ], + "score": 1.0, + "content": "Recent studies have attempted at addressing these concerns regarding large-scale transformers’", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 277, + 639 + ], + "score": 1.0, + "content": "computational and energy efficiency (see", + "type": "text" + }, + { + "bbox": [ + 278, + 627, + 289, + 639 + ], + "score": 0.62, + "content": "\\ S 6", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "for a more extensive discussion.) Among these, we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "focus on PoWER-BERT (Goyal et al., 2020) which progressively reduces sequence length by", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "eliminating word-vectors based on the attention values as passing layers. PoWER-BERT establishes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "the superiority of accuracy-time trade-off over earlier approaches (Sanh et al., 2019; Sun et al., 2019;", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "score": 1.0, + "content": "Michel et al., 2019). It however requires us to train a separate model for each efficiency constraint.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "In this paper, we thus develop a framework based on PoWER-BERT such that we can train a single", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 439, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 439, + 706 + ], + "score": 1.0, + "content": "model that can be adapted in the inference time to meet any given efficiency target.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "In order to train a transformer to cope with a diverse set of computational budgets in the inference", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "score": 1.0, + "content": "time, we propose to train one while reducing the sequence length with a random proportion at each", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 504, + 117 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 505, + 96 + ], + "score": 1.0, + "content": "LENGTH-ADAPTIVE TRANSFORMER: TRAIN ONCE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 485, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 485, + 117 + ], + "score": 1.0, + "content": "WITH LENGTH DROP, USE ANYTIME WITH SEARCH", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 138, + 244, + 159 + ], + "lines": [ + { + "bbox": [ + 113, + 138, + 201, + 150 + ], + "spans": [ + { + "bbox": [ + 113, + 138, + 201, + 150 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 148, + 245, + 161 + ], + "spans": [ + { + "bbox": [ + 112, + 148, + 245, + 161 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 138, + 245, + 161 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 189, + 333, + 201 + ], + "lines": [ + { + "bbox": [ + 276, + 187, + 336, + 203 + ], + "spans": [ + { + "bbox": [ + 276, + 187, + 336, + 203 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 142, + 214, + 469, + 477 + ], + "lines": [ + { + "bbox": [ + 142, + 214, + 469, + 226 + ], + "spans": [ + { + "bbox": [ + 142, + 214, + 469, + 226 + ], + "score": 1.0, + "content": "Although transformers have achieved impressive accuracies in various tasks in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 225, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 469, + 236 + ], + "score": 1.0, + "content": "natural language processing, they often come with a prohibitive computational", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 236, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 470, + 247 + ], + "score": 1.0, + "content": "cost, that prevents their use in scenarios with limited computational resources for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "score": 1.0, + "content": "inference. This need for computational efficiency in inference has been addressed", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 469, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 469, + 269 + ], + "score": 1.0, + "content": "by for instance PoWER-BERT (Goyal et al., 2020) which gradually decreases the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 280 + ], + "score": 1.0, + "content": "length of a sequence as it is passed through layers. These approaches however", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 279, + 470, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 470, + 291 + ], + "score": 1.0, + "content": "often assume that the target computational complexity is known in advance at", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "spans": [ + { + "bbox": [ + 142, + 290, + 469, + 302 + ], + "score": 1.0, + "content": "the time of training. This implies that a separate model must be trained for each", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 302, + 469, + 313 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 469, + 313 + ], + "score": 1.0, + "content": "inference scenario with its distinct computational budget. In this paper, we extend", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "score": 1.0, + "content": "PoWER-BERT to address this issue of inefficiency and redundancy. The proposed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "extension enables us to train a large-scale transformer, called Length-Adaptive", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 470, + 347 + ], + "score": 1.0, + "content": "Transformer, once and uses it for various inference scenarios without re-training", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 140, + 344, + 470, + 357 + ], + "spans": [ + { + "bbox": [ + 140, + 344, + 470, + 357 + ], + "score": 1.0, + "content": "it. To do so, we train a transformer with LengthDrop, a structural variant of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 470, + 368 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 470, + 368 + ], + "score": 1.0, + "content": "dropout, which stochastically determines the length of a sequence at each layer.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 469, + 379 + ], + "score": 1.0, + "content": "We then use a multi-objective evolutionary search to find a length configuration", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "score": 1.0, + "content": "that maximizes the accuracy and minimizes the computational complexity under", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "score": 1.0, + "content": "any given computational budget. Additionally, we significantly extend the appli-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 400, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 142, + 400, + 470, + 412 + ], + "score": 1.0, + "content": "cability of PoWER-BERT beyond sequence-level classification into token-level", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 411, + 470, + 424 + ], + "spans": [ + { + "bbox": [ + 141, + 411, + 470, + 424 + ], + "score": 1.0, + "content": "classification such as span-based question-answering, by introducing the idea of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 421, + 470, + 434 + ], + "score": 1.0, + "content": "Drop-and-Restore. With Drop-and-Restore, word-vectors are dropped temporarily", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 432, + 469, + 446 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 446 + ], + "score": 1.0, + "content": "in intermediate layers and restored at the last layer if necessary. We empirically", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 443, + 470, + 457 + ], + "spans": [ + { + "bbox": [ + 141, + 443, + 470, + 457 + ], + "score": 1.0, + "content": "verify the utility of the proposed approach by demonstrating the superior accuracy-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 142, + 455, + 470, + 466 + ], + "spans": [ + { + "bbox": [ + 142, + 455, + 470, + 466 + ], + "score": 1.0, + "content": "efficiency trade-off under various setups, including SQuAD 1.1, MNLI-m, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 142, + 465, + 462, + 478 + ], + "spans": [ + { + "bbox": [ + 142, + 465, + 462, + 478 + ], + "score": 1.0, + "content": "SST-2. Upon publication, the code to reproduce our work will be open-sourced.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 16.5, + "bbox_fs": [ + 140, + 214, + 470, + 478 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 497, + 206, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 208, + 513 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 522, + 506, + 610 + ], + "lines": [ + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 506, + 535 + ], + "score": 1.0, + "content": "Pretrained language models (Peters et al., 2018; Devlin et al., 2018; Radford et al., 2019; Yang et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 506, + 547 + ], + "score": 1.0, + "content": "2019) have achieved notable improvements in various natural language processing (NLP) tasks. Most", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 506, + 558 + ], + "score": 1.0, + "content": "of them rely on transformers (Vaswani et al., 2017), and the number of model parameters ranges", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 506, + 567 + ], + "score": 1.0, + "content": "from hundreds of millions to billions (Shoeybi et al., 2019; Raffel et al., 2019; Kaplan et al., 2020;", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 565, + 507, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 565, + 507, + 579 + ], + "score": 1.0, + "content": "Brown et al., 2020). Despite this high accuracy, excessive computational overhead during inference,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 589 + ], + "score": 1.0, + "content": "both in terms of time and memory, has hindered its use in real applications. This level of excessive", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 506, + 600 + ], + "score": 1.0, + "content": "computation has further raised the concern over energy consumption as well (Schwartz et al., 2019;", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 599, + 193, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 193, + 611 + ], + "score": 1.0, + "content": "Strubell et al., 2019).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 521, + 507, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 507, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 507, + 630 + ], + "score": 1.0, + "content": "Recent studies have attempted at addressing these concerns regarding large-scale transformers’", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 277, + 639 + ], + "score": 1.0, + "content": "computational and energy efficiency (see", + "type": "text" + }, + { + "bbox": [ + 278, + 627, + 289, + 639 + ], + "score": 0.62, + "content": "\\ S 6", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "for a more extensive discussion.) Among these, we", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "score": 1.0, + "content": "focus on PoWER-BERT (Goyal et al., 2020) which progressively reduces sequence length by", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "eliminating word-vectors based on the attention values as passing layers. PoWER-BERT establishes", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 506, + 672 + ], + "score": 1.0, + "content": "the superiority of accuracy-time trade-off over earlier approaches (Sanh et al., 2019; Sun et al., 2019;", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "score": 1.0, + "content": "Michel et al., 2019). It however requires us to train a separate model for each efficiency constraint.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "In this paper, we thus develop a framework based on PoWER-BERT such that we can train a single", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 692, + 439, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 439, + 706 + ], + "score": 1.0, + "content": "model that can be adapted in the inference time to meet any given efficiency target.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 615, + 507, + 706 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "In order to train a transformer to cope with a diverse set of computational budgets in the inference", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 504, + 732 + ], + "score": 1.0, + "content": "time, we propose to train one while reducing the sequence length with a random proportion at each", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "layer. We refer to this procedure as LengthDrop which was motivated by the nested dropout (Rippel", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "et al., 2014). We can extract sub-models of shared weights with any length configuration without", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 336, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 336, + 118 + ], + "score": 1.0, + "content": "requiring extra post-processing nor additional finetuning.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "layer. We refer to this procedure as LengthDrop which was motivated by the nested dropout (Rippel", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "et al., 2014). We can extract sub-models of shared weights with any length configuration without", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 336, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 336, + 118 + ], + "score": 1.0, + "content": "requiring extra post-processing nor additional finetuning.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "score": 1.0, + "content": "Once a transformer is trained with the proposed LengthDrop, we search for the length configuration", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "that maximizes the accuracy given a computational budget. Because this search is combinatorial and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "has multiple objectives (accuracy and efficiency), we use an evolutionary search algorithm, which", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 486, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 486, + 166 + ], + "score": 1.0, + "content": "further allows us to obtain a full Pareto frontier of accuracy-efficiency trade-off of each model.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "It is not trivial to find an optimal length configuration given the inference-time computational budget,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 182, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 195 + ], + "score": 1.0, + "content": "although it is extremely important in order to deploy these large-scale transformers in practice. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 193, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 506, + 206 + ], + "score": 1.0, + "content": "this work, we propose to use evolutionary search to find a length configuration that maximizes", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "the accuracy within a given computational budget. We can further compute the Pareto frontier of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 214, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 505, + 229 + ], + "score": 1.0, + "content": "accuracy-efficiency trade-off to obtain a sequence of length configurations with varying efficiency", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 142, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 142, + 239 + ], + "score": 1.0, + "content": "profiles.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "PoWER-BERT, which forms the foundation of the proposed two-stage procedure, is only applicable", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "to sequence-level classification, because by design it eliminates some of the word vectors at each", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "layer. In other words, it cannot be used for token-level tasks such as span-based question answering", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "score": 1.0, + "content": "(Rajpurkar et al., 2016), because these tasks require hidden representations of the entire input sequence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "at the final layer. We thus propose to extend PoWER-BERT with a novel Drop-and-Restore process", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "(§3.3), which eliminates this inherent limitation. Word vectors are dropped and set aside, rather", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "than eliminated, in intermediate layers to maintain the saving of computational cost, as was with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "the original PoWER-BERT. These set-aside vectors are then restored at the final hidden layer and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "provided as an input to a subsequent task-specific layer, which is unlike the original PoWER-BERT.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 347, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "The main contributions of this work are two-fold. First, we introduce LengthDrop, a structured", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "variant of dropout for training a single Length-Adaptive Transformer model that allows us to au-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "tomatically derive multiple sub-models with different length configurations in the inference time", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "using evolutionary search, without requiring any re-training. Second, we design Drop-and-Restore", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "score": 1.0, + "content": "process that makes PoWER-BERT applicable beyond classification, which enables PoWER-BERT to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "be applicable to a wider range of NLP tasks such as span-based question answering. We empirically", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "verify Length-Adaptive Transformer works quite well using the variants of BERT on a diverse set of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "NLP tasks, including SQuAD 1.1 (Rajpurkar et al., 2016) and two sequence-level classification tasks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "in GLUE benchmark (Wang et al., 2018). Our experiments reveal that the proposed approach grants", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "us a fine-grained control of computational efficiency and a superior accuracy-efficiency trade-off in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 458, + 319, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 319, + 469 + ], + "score": 1.0, + "content": "the inference time, compared to existing approaches.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 484, + 399, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 401, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 401, + 499 + ], + "score": 1.0, + "content": "2 BACKGROUND: TRANSFORMERS AND POWER-BERT", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "Before we describe our main approach, we review some of the building blocks in this section. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "particular, we review transformers, which are a standard backbone used in natural language processing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 531, + 507, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 507, + 545 + ], + "score": 1.0, + "content": "these days, and PoWER-BERT, which was recently proposed as an effective way to train a large-scale,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 543, + 366, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 366, + 555 + ], + "score": 1.0, + "content": "but highly efficient transformer for sequence-level classification.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 253, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 254, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 254, + 580 + ], + "score": 1.0, + "content": "2.1 TRANSFORMERS AND BERT", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 504, + 599 + ], + "score": 1.0, + "content": "A transformer is a particular neural network that has been designed to work with a variable-length", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "sequence input and is implemented as a stack of self-attention and fully-connected layers (Vaswani", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "et al., 2017). Here, we give a brief overview of the transformer which is the basic building block of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 622, + 201, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 201, + 633 + ], + "score": 1.0, + "content": "the proposed approach.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 155, + 651 + ], + "score": 1.0, + "content": "Each token", + "type": "text" + }, + { + "bbox": [ + 155, + 640, + 166, + 649 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 637, + 268, + 651 + ], + "score": 1.0, + "content": "in a sequence of tokens", + "type": "text" + }, + { + "bbox": [ + 269, + 638, + 344, + 650 + ], + "score": 0.91, + "content": "{ \\boldsymbol x } = ( x _ { 1 } , \\dots , x _ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 637, + 505, + 651 + ], + "score": 1.0, + "content": ", representing input text, is first turned", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 646, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 208, + 662 + ], + "score": 1.0, + "content": "into a continuous vector", + "type": "text" + }, + { + "bbox": [ + 208, + 648, + 246, + 660 + ], + "score": 0.93, + "content": "h _ { t } ^ { 0 } \\in \\mathbb { R } ^ { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 646, + 507, + 662 + ], + "score": 1.0, + "content": "which is the sum of the token and position embedding vectors.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "This sequence is fed into the first transformer layer which returns another sequence of the same", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 667, + 507, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 667, + 135, + 685 + ], + "score": 1.0, + "content": "length", + "type": "text" + }, + { + "bbox": [ + 135, + 670, + 187, + 681 + ], + "score": 0.91, + "content": "\\dot { h ^ { 1 } } \\in \\mathbb { R } ^ { N \\times H }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 667, + 297, + 685 + ], + "score": 1.0, + "content": ". We repeat this procedure", + "type": "text" + }, + { + "bbox": [ + 297, + 671, + 306, + 681 + ], + "score": 0.81, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 667, + 427, + 685 + ], + "score": 1.0, + "content": "times, for a transformer with", + "type": "text" + }, + { + "bbox": [ + 427, + 671, + 435, + 681 + ], + "score": 0.75, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 667, + 507, + 685 + ], + "score": 1.0, + "content": "layers, to obtain", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 107, + 679, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 107, + 681, + 188, + 694 + ], + "score": 0.89, + "content": "h ^ { L } = ( h _ { 1 } ^ { L } , \\ldots , h _ { N } ^ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 679, + 506, + 696 + ], + "score": 1.0, + "content": ". We refer to each vector in the hidden sequence at each layer as a word vector to", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 692, + 504, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 504, + 705 + ], + "score": 1.0, + "content": "emphasize that there exists a correspondence between each such vector and one of the input words.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "Although the transformer was first introduced for the problem of machine translation, Devlin et al.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "(2018) demonstrated that the transformer can be trained and used as a masked language model.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 118 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 134 + ], + "score": 1.0, + "content": "Once a transformer is trained with the proposed LengthDrop, we search for the length configuration", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "that maximizes the accuracy given a computational budget. Because this search is combinatorial and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "has multiple objectives (accuracy and efficiency), we use an evolutionary search algorithm, which", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 486, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 486, + 166 + ], + "score": 1.0, + "content": "further allows us to obtain a full Pareto frontier of accuracy-efficiency trade-off of each model.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 121, + 506, + 166 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 184 + ], + "score": 1.0, + "content": "It is not trivial to find an optimal length configuration given the inference-time computational budget,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 182, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 195 + ], + "score": 1.0, + "content": "although it is extremely important in order to deploy these large-scale transformers in practice. In", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 193, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 506, + 206 + ], + "score": 1.0, + "content": "this work, we propose to use evolutionary search to find a length configuration that maximizes", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "the accuracy within a given computational budget. We can further compute the Pareto frontier of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 214, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 104, + 214, + 505, + 229 + ], + "score": 1.0, + "content": "accuracy-efficiency trade-off to obtain a sequence of length configurations with varying efficiency", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 142, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 142, + 239 + ], + "score": 1.0, + "content": "profiles.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 104, + 170, + 506, + 239 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 342 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 256 + ], + "score": 1.0, + "content": "PoWER-BERT, which forms the foundation of the proposed two-stage procedure, is only applicable", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "to sequence-level classification, because by design it eliminates some of the word vectors at each", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "layer. In other words, it cannot be used for token-level tasks such as span-based question answering", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "score": 1.0, + "content": "(Rajpurkar et al., 2016), because these tasks require hidden representations of the entire input sequence", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "at the final layer. We thus propose to extend PoWER-BERT with a novel Drop-and-Restore process", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 505, + 310 + ], + "score": 1.0, + "content": "(§3.3), which eliminates this inherent limitation. Word vectors are dropped and set aside, rather", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "than eliminated, in intermediate layers to maintain the saving of computational cost, as was with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "the original PoWER-BERT. These set-aside vectors are then restored at the final hidden layer and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "provided as an input to a subsequent task-specific layer, which is unlike the original PoWER-BERT.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 243, + 506, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 347, + 505, + 468 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 360 + ], + "score": 1.0, + "content": "The main contributions of this work are two-fold. First, we introduce LengthDrop, a structured", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 359, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 506, + 371 + ], + "score": 1.0, + "content": "variant of dropout for training a single Length-Adaptive Transformer model that allows us to au-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "tomatically derive multiple sub-models with different length configurations in the inference time", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "score": 1.0, + "content": "using evolutionary search, without requiring any re-training. Second, we design Drop-and-Restore", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 405 + ], + "score": 1.0, + "content": "process that makes PoWER-BERT applicable beyond classification, which enables PoWER-BERT to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "be applicable to a wider range of NLP tasks such as span-based question answering. We empirically", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "verify Length-Adaptive Transformer works quite well using the variants of BERT on a diverse set of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "NLP tasks, including SQuAD 1.1 (Rajpurkar et al., 2016) and two sequence-level classification tasks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "in GLUE benchmark (Wang et al., 2018). Our experiments reveal that the proposed approach grants", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "us a fine-grained control of computational efficiency and a superior accuracy-efficiency trade-off in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 458, + 319, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 319, + 469 + ], + "score": 1.0, + "content": "the inference time, compared to existing approaches.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 347, + 506, + 469 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 484, + 399, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 401, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 401, + 499 + ], + "score": 1.0, + "content": "2 BACKGROUND: TRANSFORMERS AND POWER-BERT", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 505, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "Before we describe our main approach, we review some of the building blocks in this section. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "particular, we review transformers, which are a standard backbone used in natural language processing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 531, + 507, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 507, + 545 + ], + "score": 1.0, + "content": "these days, and PoWER-BERT, which was recently proposed as an effective way to train a large-scale,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 543, + 366, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 366, + 555 + ], + "score": 1.0, + "content": "but highly efficient transformer for sequence-level classification.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 510, + 507, + 555 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 253, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 254, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 254, + 580 + ], + "score": 1.0, + "content": "2.1 TRANSFORMERS AND BERT", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 588, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 504, + 599 + ], + "score": 1.0, + "content": "A transformer is a particular neural network that has been designed to work with a variable-length", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "sequence input and is implemented as a stack of self-attention and fully-connected layers (Vaswani", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 623 + ], + "score": 1.0, + "content": "et al., 2017). Here, we give a brief overview of the transformer which is the basic building block of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 622, + 201, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 201, + 633 + ], + "score": 1.0, + "content": "the proposed approach.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 588, + 506, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 155, + 651 + ], + "score": 1.0, + "content": "Each token", + "type": "text" + }, + { + "bbox": [ + 155, + 640, + 166, + 649 + ], + "score": 0.84, + "content": "x _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 637, + 268, + 651 + ], + "score": 1.0, + "content": "in a sequence of tokens", + "type": "text" + }, + { + "bbox": [ + 269, + 638, + 344, + 650 + ], + "score": 0.91, + "content": "{ \\boldsymbol x } = ( x _ { 1 } , \\dots , x _ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 637, + 505, + 651 + ], + "score": 1.0, + "content": ", representing input text, is first turned", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 646, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 208, + 662 + ], + "score": 1.0, + "content": "into a continuous vector", + "type": "text" + }, + { + "bbox": [ + 208, + 648, + 246, + 660 + ], + "score": 0.93, + "content": "h _ { t } ^ { 0 } \\in \\mathbb { R } ^ { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 646, + 507, + 662 + ], + "score": 1.0, + "content": "which is the sum of the token and position embedding vectors.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "This sequence is fed into the first transformer layer which returns another sequence of the same", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 104, + 667, + 507, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 667, + 135, + 685 + ], + "score": 1.0, + "content": "length", + "type": "text" + }, + { + "bbox": [ + 135, + 670, + 187, + 681 + ], + "score": 0.91, + "content": "\\dot { h ^ { 1 } } \\in \\mathbb { R } ^ { N \\times H }", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 667, + 297, + 685 + ], + "score": 1.0, + "content": ". We repeat this procedure", + "type": "text" + }, + { + "bbox": [ + 297, + 671, + 306, + 681 + ], + "score": 0.81, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 667, + 427, + 685 + ], + "score": 1.0, + "content": "times, for a transformer with", + "type": "text" + }, + { + "bbox": [ + 427, + 671, + 435, + 681 + ], + "score": 0.75, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 667, + 507, + 685 + ], + "score": 1.0, + "content": "layers, to obtain", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 107, + 679, + 506, + 696 + ], + "spans": [ + { + "bbox": [ + 107, + 681, + 188, + 694 + ], + "score": 0.89, + "content": "h ^ { L } = ( h _ { 1 } ^ { L } , \\ldots , h _ { N } ^ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 679, + 506, + 696 + ], + "score": 1.0, + "content": ". We refer to each vector in the hidden sequence at each layer as a word vector to", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 692, + 504, + 705 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 504, + 705 + ], + "score": 1.0, + "content": "emphasize that there exists a correspondence between each such vector and one of the input words.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 104, + 637, + 507, + 705 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 506, + 721 + ], + "score": 1.0, + "content": "Although the transformer was first introduced for the problem of machine translation, Devlin et al.", + "type": "text" + } + ], + "index": 49, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "(2018) demonstrated that the transformer can be trained and used as a masked language model.", + "type": "text" + } + ], + "index": 50, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 49.5, + "bbox_fs": [ + 106, + 710, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "More specifically, Devlin et al. (2018) showed that the transformer-based masked language model,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "called BERT, learns a universally useful parameter set that can be finetuned for any downstream task", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 329, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 329, + 117 + ], + "score": 1.0, + "content": "including sequence-level and token-level classification.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 504, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 491, + 134 + ], + "score": 1.0, + "content": "In the case of sequence-level classification, a softmax classifier is attached to the word vector", + "type": "text" + }, + { + "bbox": [ + 491, + 120, + 504, + 133 + ], + "score": 0.89, + "content": "h _ { 1 } ^ { L }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "associated with the special token [CLS], and the entire network, including the softmax classifier and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 353, + 156 + ], + "score": 1.0, + "content": "BERT, is finetuned. For token-level classification, we use each", + "type": "text" + }, + { + "bbox": [ + 353, + 143, + 366, + 155 + ], + "score": 0.91, + "content": "h _ { t } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "as the final hidden reprsentation of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 165, + 168 + ], + "score": 1.0, + "content": "the associated", + "type": "text" + }, + { + "bbox": [ + 165, + 155, + 170, + 164 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 152, + 507, + 168 + ], + "score": 1.0, + "content": "-th word in the input sequence. This strategy of pretraining followed by finetuning,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "often referred to as transfer learning, has recently become a dominant approach to classification in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 222, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 222, + 189 + ], + "score": 1.0, + "content": "natural language processing.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 107, + 217, + 198, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 199, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 199, + 231 + ], + "score": 1.0, + "content": "2.2 POWER-BERT", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 244, + 505, + 300 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 261, + 258 + ], + "score": 1.0, + "content": "PoWER-BERT keeps only the topmost", + "type": "text" + }, + { + "bbox": [ + 262, + 245, + 270, + 257 + ], + "score": 0.87, + "content": "l _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 244, + 376, + 258 + ], + "score": 1.0, + "content": "word vectors at each layer", + "type": "text" + }, + { + "bbox": [ + 376, + 246, + 383, + 256 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "by eliminating redundant ones", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 256, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 505, + 268 + ], + "score": 1.0, + "content": "based on the significance score, which is the total amount of attention imposed by a word on the other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 266, + 503, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 214, + 280 + ], + "score": 1.0, + "content": "words (Goyal et al., 2020).", + "type": "text" + }, + { + "bbox": [ + 214, + 267, + 223, + 279 + ], + "score": 0.86, + "content": "l _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 266, + 497, + 280 + ], + "score": 1.0, + "content": "is the hyper-parameter determines how many vectors to keep at layer", + "type": "text" + }, + { + "bbox": [ + 497, + 268, + 503, + 279 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "score": 1.0, + "content": "PoWER-BERT has the same model parameters with BERT, but the extraction layers are interspersed", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 289, + 423, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 423, + 301 + ], + "score": 1.0, + "content": "after the self-attention layer in every transformer blocks (Vaswani et al., 2017).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 305, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "PoWER-BERT reduces inference time successfully, achieving better accuracy-time trade-off than", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 317, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 506, + 329 + ], + "score": 1.0, + "content": "DistilBERT (Sanh et al., 2019), BERT-PKD (Sun et al., 2019), and Head-Prune (Michel et al.,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "2019). Despite the original intention of maximizing the inference efficiency with the minimal loss in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "accuracy, it is possible to set up PoWER-BERT to be both more efficient and more accurate compared", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 350, + 461, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 461, + 362 + ], + "score": 1.0, + "content": "to the original BERT, which was observed but largely overlooked by Goyal et al. (2020).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Training a PoWER-BERT model consists of three steps: (1) finetuning, (2) length configuration", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "search, and (3) re-training. The finetuning step is just like the standard finetuning step of BERT given", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 401, + 401 + ], + "score": 1.0, + "content": "a target task. A length configuration is a sequence of retention parameters", + "type": "text" + }, + { + "bbox": [ + 401, + 389, + 443, + 401 + ], + "score": 0.91, + "content": "\\left( l _ { 1 } , \\cdots l _ { L } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 389, + 505, + 401 + ], + "score": 1.0, + "content": ", each of which", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "corresponds to the number of word vectors that are kept at each layer. These retention parameters", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "are learned along with all the other parameters to minimize the original task loss together with an", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "extra term that approximately measures the number of retained word vectors across layers. In the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 433, + 498, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 498, + 444 + ], + "score": 1.0, + "content": "re-training step, PoWER-BERT is finetuned with the length configuration fixed to its learned one.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 449, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "For each computational budget, we must train a separate model going through all three steps described", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "above. Moreover, the length configuration search step above is only approximate, as it relies on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "relaxation of retention parameters which are inherently discrete. This leads to the lack of guaranteed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 483, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 494 + ], + "score": 1.0, + "content": "correlation between the success of this stage and true run-time. Even worse, it is a delicate act to tune", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 493, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 505 + ], + "score": 1.0, + "content": "the length configuration given a target computational budget, because trade-off is implicitly made via", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "a regularization coefficient. Furthermore, PoWER-BERT has an inherent limitation in that it only", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 515, + 495, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 495, + 528 + ], + "score": 1.0, + "content": "applies to sequence-level classification because it eliminates word vectors in intermediate layers.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 559, + 309, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 311, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 311, + 574 + ], + "score": 1.0, + "content": "3 LENGTH-ADAPTIVE TRANSFORMER", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "score": 1.0, + "content": "In this section, we explain our proposed framework which results in a transformer that reduces", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "the length of a sequence at each layer with an arbitrary rate. We call such a resulting transformer", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "a Length-Adaptive Transformer. We train Length-Adaptive Transformer with LengthDrop which", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 625, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 625, + 506, + 641 + ], + "score": 1.0, + "content": "randomly samples the number of hidden vectors to be dropped at each layer with the goal of making", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "the final model robust to such drop in the inference time. Once the model is trained, we search for the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "optimal trade-off between accuracy and efficiency using multi-objective evolutionary search, which", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 659, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 674 + ], + "score": 1.0, + "content": "allows us to use the model for any given computational budget without finetuning nor re-training.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "At the end of this section, we describe Drop-and-Restore process as a way to greatly increase the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 681, + 465, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 465, + 695 + ], + "score": 1.0, + "content": "applicability of PoWER-BERT which forms a building block of the proposed framework.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "In short, we train a Length-Adaptive Transformer once with LengthDrop and Drop-and-Restore,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "and use it with an automatically determined length configuration for inference with any target", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 381, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 381, + 732 + ], + "score": 1.0, + "content": "computational budget, on both sequence-level and token-level tasks.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "More specifically, Devlin et al. (2018) showed that the transformer-based masked language model,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "called BERT, learns a universally useful parameter set that can be finetuned for any downstream task", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 329, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 329, + 117 + ], + "score": 1.0, + "content": "including sequence-level and token-level classification.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 120, + 504, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 491, + 134 + ], + "score": 1.0, + "content": "In the case of sequence-level classification, a softmax classifier is attached to the word vector", + "type": "text" + }, + { + "bbox": [ + 491, + 120, + 504, + 133 + ], + "score": 0.89, + "content": "h _ { 1 } ^ { L }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "associated with the special token [CLS], and the entire network, including the softmax classifier and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 353, + 156 + ], + "score": 1.0, + "content": "BERT, is finetuned. For token-level classification, we use each", + "type": "text" + }, + { + "bbox": [ + 353, + 143, + 366, + 155 + ], + "score": 0.91, + "content": "h _ { t } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "as the final hidden reprsentation of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 152, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 165, + 168 + ], + "score": 1.0, + "content": "the associated", + "type": "text" + }, + { + "bbox": [ + 165, + 155, + 170, + 164 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 152, + 507, + 168 + ], + "score": 1.0, + "content": "-th word in the input sequence. This strategy of pretraining followed by finetuning,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "often referred to as transfer learning, has recently become a dominant approach to classification in", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 176, + 222, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 222, + 189 + ], + "score": 1.0, + "content": "natural language processing.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 120, + 507, + 189 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 217, + 198, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 217, + 199, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 199, + 231 + ], + "score": 1.0, + "content": "2.2 POWER-BERT", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 244, + 505, + 300 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 261, + 258 + ], + "score": 1.0, + "content": "PoWER-BERT keeps only the topmost", + "type": "text" + }, + { + "bbox": [ + 262, + 245, + 270, + 257 + ], + "score": 0.87, + "content": "l _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 244, + 376, + 258 + ], + "score": 1.0, + "content": "word vectors at each layer", + "type": "text" + }, + { + "bbox": [ + 376, + 246, + 383, + 256 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "by eliminating redundant ones", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 256, + 505, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 256, + 505, + 268 + ], + "score": 1.0, + "content": "based on the significance score, which is the total amount of attention imposed by a word on the other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 266, + 503, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 214, + 280 + ], + "score": 1.0, + "content": "words (Goyal et al., 2020).", + "type": "text" + }, + { + "bbox": [ + 214, + 267, + 223, + 279 + ], + "score": 0.86, + "content": "l _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 266, + 497, + 280 + ], + "score": 1.0, + "content": "is the hyper-parameter determines how many vectors to keep at layer", + "type": "text" + }, + { + "bbox": [ + 497, + 268, + 503, + 279 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 291 + ], + "score": 1.0, + "content": "PoWER-BERT has the same model parameters with BERT, but the extraction layers are interspersed", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 289, + 423, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 423, + 301 + ], + "score": 1.0, + "content": "after the self-attention layer in every transformer blocks (Vaswani et al., 2017).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 244, + 506, + 301 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 305, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "PoWER-BERT reduces inference time successfully, achieving better accuracy-time trade-off than", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 317, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 506, + 329 + ], + "score": 1.0, + "content": "DistilBERT (Sanh et al., 2019), BERT-PKD (Sun et al., 2019), and Head-Prune (Michel et al.,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "2019). Despite the original intention of maximizing the inference efficiency with the minimal loss in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "accuracy, it is possible to set up PoWER-BERT to be both more efficient and more accurate compared", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 350, + 461, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 461, + 362 + ], + "score": 1.0, + "content": "to the original BERT, which was observed but largely overlooked by Goyal et al. (2020).", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 305, + 506, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 443 + ], + "lines": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "Training a PoWER-BERT model consists of three steps: (1) finetuning, (2) length configuration", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "search, and (3) re-training. The finetuning step is just like the standard finetuning step of BERT given", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 401, + 401 + ], + "score": 1.0, + "content": "a target task. A length configuration is a sequence of retention parameters", + "type": "text" + }, + { + "bbox": [ + 401, + 389, + 443, + 401 + ], + "score": 0.91, + "content": "\\left( l _ { 1 } , \\cdots l _ { L } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 389, + 505, + 401 + ], + "score": 1.0, + "content": ", each of which", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 413 + ], + "score": 1.0, + "content": "corresponds to the number of word vectors that are kept at each layer. These retention parameters", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "are learned along with all the other parameters to minimize the original task loss together with an", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "extra term that approximately measures the number of retained word vectors across layers. In the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 433, + 498, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 498, + 444 + ], + "score": 1.0, + "content": "re-training step, PoWER-BERT is finetuned with the length configuration fixed to its learned one.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 366, + 506, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 449, + 505, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 506, + 462 + ], + "score": 1.0, + "content": "For each computational budget, we must train a separate model going through all three steps described", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "above. Moreover, the length configuration search step above is only approximate, as it relies on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "relaxation of retention parameters which are inherently discrete. This leads to the lack of guaranteed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 483, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 494 + ], + "score": 1.0, + "content": "correlation between the success of this stage and true run-time. Even worse, it is a delicate act to tune", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 493, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 505 + ], + "score": 1.0, + "content": "the length configuration given a target computational budget, because trade-off is implicitly made via", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 516 + ], + "score": 1.0, + "content": "a regularization coefficient. Furthermore, PoWER-BERT has an inherent limitation in that it only", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 515, + 495, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 495, + 528 + ], + "score": 1.0, + "content": "applies to sequence-level classification because it eliminates word vectors in intermediate layers.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 449, + 506, + 528 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 559, + 309, + 572 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 311, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 311, + 574 + ], + "score": 1.0, + "content": "3 LENGTH-ADAPTIVE TRANSFORMER", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "score": 1.0, + "content": "In this section, we explain our proposed framework which results in a transformer that reduces", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "the length of a sequence at each layer with an arbitrary rate. We call such a resulting transformer", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "a Length-Adaptive Transformer. We train Length-Adaptive Transformer with LengthDrop which", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 625, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 625, + 506, + 641 + ], + "score": 1.0, + "content": "randomly samples the number of hidden vectors to be dropped at each layer with the goal of making", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "the final model robust to such drop in the inference time. Once the model is trained, we search for the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "optimal trade-off between accuracy and efficiency using multi-objective evolutionary search, which", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 659, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 674 + ], + "score": 1.0, + "content": "allows us to use the model for any given computational budget without finetuning nor re-training.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "At the end of this section, we describe Drop-and-Restore process as a way to greatly increase the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 681, + 465, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 465, + 695 + ], + "score": 1.0, + "content": "applicability of PoWER-BERT which forms a building block of the proposed framework.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 593, + 506, + 695 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "In short, we train a Length-Adaptive Transformer once with LengthDrop and Drop-and-Restore,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "and use it with an automatically determined length configuration for inference with any target", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 721, + 381, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 381, + 732 + ], + "score": 1.0, + "content": "computational budget, on both sequence-level and token-level tasks.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 192, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 194, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 194, + 95 + ], + "score": 1.0, + "content": "3.1 LENGTHDROP", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 1.0, + "content": "Earlier approaches to efficient inference with transformers have focused on a scenario where the target", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "computational budget for inference is known in advance Sanh et al. (2019); Goyal et al. (2020). This", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "greatly increases the cost of deploying transformers, as it requires us to train a separate transformer", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "for each scenario. Instead, we propose to train one model that could be used for a diverse set of target", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 147, + 275, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 275, + 160 + ], + "score": 1.0, + "content": "computational budgets without retraining.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 505, + 241 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "score": 1.0, + "content": "LengthDrop randomly generates a length configuration by sequentially sampling a sequence length", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 107, + 175, + 124, + 186 + ], + "score": 0.91, + "content": "l _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 174, + 150, + 188 + ], + "score": 1.0, + "content": "at the", + "type": "text" + }, + { + "bbox": [ + 151, + 175, + 179, + 186 + ], + "score": 0.91, + "content": "( i + 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 174, + 402, + 188 + ], + "score": 1.0, + "content": "-th layer based on the previous layer’s sequence length", + "type": "text" + }, + { + "bbox": [ + 402, + 175, + 409, + 186 + ], + "score": 0.86, + "content": "l _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 174, + 506, + 188 + ], + "score": 1.0, + "content": ", following the uniform", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 185, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 156, + 199 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 157, + 186, + 230, + 198 + ], + "score": 0.92, + "content": "\\mathcal { U } ( \\lceil ( 1 - p ) l _ { i } \\rceil , l _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 185, + 263, + 199 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 186, + 273, + 197 + ], + "score": 0.85, + "content": "l _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 185, + 471, + 199 + ], + "score": 1.0, + "content": "is set to the length of the input sequence, and", + "type": "text" + }, + { + "bbox": [ + 471, + 187, + 478, + 197 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 185, + 506, + 199 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 454, + 210 + ], + "score": 1.0, + "content": "LengthDrop probability. This sequential sampling results in a length configuration", + "type": "text" + }, + { + "bbox": [ + 454, + 197, + 503, + 209 + ], + "score": 0.91, + "content": "( l _ { 1 } , \\cdots , l _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 196, + 506, + 210 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 506, + 220 + ], + "score": 1.0, + "content": "Length-Adaptive Transformer can be thought of as consisting of a full model and many sub-models", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "score": 1.0, + "content": "corresponding to different length configuration, similarly to a neural network trained with dropout", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 230, + 205, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 205, + 242 + ], + "score": 1.0, + "content": "(Srivastava et al., 2014).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "LayerDrop From the perspective of each word vector, the proposed LengthDrop could be thought", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "of as skipping the layers between when it was set aside and the final layer where it was restored.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "The word vector however does not have any information based on which it can determine whether it", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "would be dropped at any particular layer. In our preliminary experiments, we found that this greatly", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "hinders optimization. We address this issue by using LayerDrop (Fan et al., 2019) which skips each", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "layer of a transformer uniformly at random. The LayerDrop encourages each word vector to be", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "agnostic to skipping any number of layers between when it is dropped and when it is restored, just", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "like dropout (Srivastava et al., 2014) prevents hidden neurons from co-adapting with each other by", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 340, + 210, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 210, + 354 + ], + "score": 1.0, + "content": "randomly dropping them.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "Sandwich Rule and Inplace Distillation We observed that standard supervised training with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "score": 1.0, + "content": "LengthDrop does not work well in the preliminary experiments. We instead borrow a pair of training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "techniques developed by Yu & Huang (2019) which are sandwich rule and inplace distillation,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "for better optimization as well as final generalization. At each update, we update the full model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "without LengthDrop as usual to minimize the supervised loss function. We simultaneously update", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 107, + 420, + 118, + 429 + ], + "score": 0.83, + "content": "n _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "randomly-sampled sub-models (which are called sandwiches) and the smallest-possible sub-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 276, + 442 + ], + "score": 1.0, + "content": "model, which corresponds to keeping only", + "type": "text" + }, + { + "bbox": [ + 276, + 429, + 321, + 441 + ], + "score": 0.92, + "content": "\\lceil ( 1 - p ) l _ { i } \\rceil", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 429, + 428, + 442 + ], + "score": 1.0, + "content": "word vectors at each layer", + "type": "text" + }, + { + "bbox": [ + 428, + 430, + 433, + 439 + ], + "score": 0.69, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 429, + 505, + 442 + ], + "score": 1.0, + "content": ", using knowledge", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "distillation (Hinton et al., 2015) from the full model. Here, sub-models mean models with length", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "reduction. They are trained to their prediction close to the full model’s prediction (inplace distillation).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + }, + { + "type": "title", + "bbox": [ + 108, + 474, + 370, + 486 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 371, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 371, + 487 + ], + "score": 1.0, + "content": "3.2 EVOLUTIONARY SEARCH OF LENGTH CONFIGURATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "After training a Length-Adaptive Transformer with LengthDrop, we search for appropriate length", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "configurations for possible target computational budgets that will be given at inference time. The", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "length configuration determines the model performance in terms of both accuracy and efficiency. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "order to search for the optimal length configuration, we propose to use evolutionary search, similarly", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "to Cai et al. (2019) and Wang et al. (2020). This procedure is efficient, as it only requires a single", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 550, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 104, + 550, + 507, + 564 + ], + "score": 1.0, + "content": "pass through the relatively small validation set for each length configuration, unlike re-training for a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "score": 1.0, + "content": "new computational budget which requires multiple passes through a significantly larger training set", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 572, + 173, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 173, + 585 + ], + "score": 1.0, + "content": "for each budget.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 589, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "We initialize the population with constant-ratio configurations. Each configuration is created by", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 187, + 612 + ], + "score": 0.92, + "content": "l _ { i + 1 } = \\lceil ( 1 - r ) \\bar { l _ { i } } \\rceil", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 599, + 250, + 613 + ], + "score": 1.0, + "content": "for each layer", + "type": "text" + }, + { + "bbox": [ + 250, + 601, + 255, + 610 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 599, + 278, + 613 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 279, + 603, + 285, + 610 + ], + "score": 0.7, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "so that the amount of computation within the initial", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 610, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 507, + 624 + ], + "score": 1.0, + "content": "population is uniformly distributed between those of the smallest and full models. At each itera-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 620, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 620, + 507, + 636 + ], + "score": 1.0, + "content": "tion, we evolve the population to consist only of configurations lie on a newly updated efficiency-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "accuracy Pareto frontier by mutation and cross-over. Mutation alters an original length configuration", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 643, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 154, + 656 + ], + "score": 0.91, + "content": "( l _ { 1 } , \\cdots , l _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 643, + 166, + 658 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 167, + 644, + 214, + 656 + ], + "score": 0.91, + "content": "( l _ { 1 } ^ { \\prime } , \\cdots , l _ { L } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 643, + 267, + 658 + ], + "score": 1.0, + "content": "by sampling", + "type": "text" + }, + { + "bbox": [ + 267, + 644, + 275, + 655 + ], + "score": 0.84, + "content": "l _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 643, + 393, + 658 + ], + "score": 1.0, + "content": "from the uniform distribution", + "type": "text" + }, + { + "bbox": [ + 394, + 644, + 447, + 656 + ], + "score": 0.91, + "content": "\\bar { \\mathcal { U } } ( l _ { i - 1 } ^ { \\prime } , l _ { i + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 643, + 507, + 658 + ], + "score": 1.0, + "content": "with the prob-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 134, + 668 + ], + "score": 1.0, + "content": "ability", + "type": "text" + }, + { + "bbox": [ + 134, + 657, + 147, + 667 + ], + "score": 0.85, + "content": "p _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 654, + 267, + 668 + ], + "score": 1.0, + "content": "or keeping the original length", + "type": "text" + }, + { + "bbox": [ + 268, + 655, + 294, + 667 + ], + "score": 0.9, + "content": "{ \\bar { l } } _ { i } ^ { \\prime } = l _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 654, + 400, + 668 + ], + "score": 1.0, + "content": ", sweeping the layers from", + "type": "text" + }, + { + "bbox": [ + 400, + 656, + 423, + 665 + ], + "score": 0.88, + "content": "i = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 654, + 434, + 668 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 434, + 655, + 459, + 665 + ], + "score": 0.9, + "content": "i = L", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 654, + 506, + 668 + ], + "score": 1.0, + "content": ". Crossover", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "takes two length configurations and averages the lengths at each layer. Both of these operations are", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 495, + 689 + ], + "score": 1.0, + "content": "performed while ensuring the monotonicity of the lengths over the layers. We repeat this iteration", + "type": "text" + }, + { + "bbox": [ + 495, + 677, + 504, + 687 + ], + "score": 0.79, + "content": "G", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 209, + 701 + ], + "score": 1.0, + "content": "times, while maintaining", + "type": "text" + }, + { + "bbox": [ + 209, + 689, + 223, + 699 + ], + "score": 0.83, + "content": "n _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 687, + 337, + 701 + ], + "score": 1.0, + "content": "mutated configurations and", + "type": "text" + }, + { + "bbox": [ + 338, + 689, + 348, + 699 + ], + "score": 0.79, + "content": "n _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "crossover’d configurations. Repeating", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "this procedure pushes the Pareto frontier further to identify the best trade-off between two objectives,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "efficiency and accuracy, without requiring any continuous relaxation of length configurations nor", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 240, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 240, + 733 + ], + "score": 1.0, + "content": "using a proxy objective function.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 46 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 192, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 194, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 194, + 95 + ], + "score": 1.0, + "content": "3.1 LENGTHDROP", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 158 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 115 + ], + "score": 1.0, + "content": "Earlier approaches to efficient inference with transformers have focused on a scenario where the target", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "computational budget for inference is known in advance Sanh et al. (2019); Goyal et al. (2020). This", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 506, + 137 + ], + "score": 1.0, + "content": "greatly increases the cost of deploying transformers, as it requires us to train a separate transformer", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "for each scenario. Instead, we propose to train one model that could be used for a diverse set of target", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 147, + 275, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 275, + 160 + ], + "score": 1.0, + "content": "computational budgets without retraining.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 103, + 506, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 505, + 241 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 177 + ], + "score": 1.0, + "content": "LengthDrop randomly generates a length configuration by sequentially sampling a sequence length", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 174, + 506, + 188 + ], + "spans": [ + { + "bbox": [ + 107, + 175, + 124, + 186 + ], + "score": 0.91, + "content": "l _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 124, + 174, + 150, + 188 + ], + "score": 1.0, + "content": "at the", + "type": "text" + }, + { + "bbox": [ + 151, + 175, + 179, + 186 + ], + "score": 0.91, + "content": "( i + 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 174, + 402, + 188 + ], + "score": 1.0, + "content": "-th layer based on the previous layer’s sequence length", + "type": "text" + }, + { + "bbox": [ + 402, + 175, + 409, + 186 + ], + "score": 0.86, + "content": "l _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 174, + 506, + 188 + ], + "score": 1.0, + "content": ", following the uniform", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 185, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 156, + 199 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 157, + 186, + 230, + 198 + ], + "score": 0.92, + "content": "\\mathcal { U } ( \\lceil ( 1 - p ) l _ { i } \\rceil , l _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 185, + 263, + 199 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 264, + 186, + 273, + 197 + ], + "score": 0.85, + "content": "l _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 185, + 471, + 199 + ], + "score": 1.0, + "content": "is set to the length of the input sequence, and", + "type": "text" + }, + { + "bbox": [ + 471, + 187, + 478, + 197 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 185, + 506, + 199 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 454, + 210 + ], + "score": 1.0, + "content": "LengthDrop probability. This sequential sampling results in a length configuration", + "type": "text" + }, + { + "bbox": [ + 454, + 197, + 503, + 209 + ], + "score": 0.91, + "content": "( l _ { 1 } , \\cdots , l _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 196, + 506, + 210 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 207, + 506, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 506, + 220 + ], + "score": 1.0, + "content": "Length-Adaptive Transformer can be thought of as consisting of a full model and many sub-models", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 232 + ], + "score": 1.0, + "content": "corresponding to different length configuration, similarly to a neural network trained with dropout", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 230, + 205, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 205, + 242 + ], + "score": 1.0, + "content": "(Srivastava et al., 2014).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 163, + 506, + 242 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 252, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 265 + ], + "score": 1.0, + "content": "LayerDrop From the perspective of each word vector, the proposed LengthDrop could be thought", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "of as skipping the layers between when it was set aside and the final layer where it was restored.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "The word vector however does not have any information based on which it can determine whether it", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "would be dropped at any particular layer. In our preliminary experiments, we found that this greatly", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 505, + 309 + ], + "score": 1.0, + "content": "hinders optimization. We address this issue by using LayerDrop (Fan et al., 2019) which skips each", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 506, + 320 + ], + "score": 1.0, + "content": "layer of a transformer uniformly at random. The LayerDrop encourages each word vector to be", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "agnostic to skipping any number of layers between when it is dropped and when it is restored, just", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 342 + ], + "score": 1.0, + "content": "like dropout (Srivastava et al., 2014) prevents hidden neurons from co-adapting with each other by", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 340, + 210, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 210, + 354 + ], + "score": 1.0, + "content": "randomly dropping them.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 252, + 506, + 354 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "Sandwich Rule and Inplace Distillation We observed that standard supervised training with", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 506, + 388 + ], + "score": 1.0, + "content": "LengthDrop does not work well in the preliminary experiments. We instead borrow a pair of training", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 398 + ], + "score": 1.0, + "content": "techniques developed by Yu & Huang (2019) which are sandwich rule and inplace distillation,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 506, + 408 + ], + "score": 1.0, + "content": "for better optimization as well as final generalization. At each update, we update the full model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "without LengthDrop as usual to minimize the supervised loss function. We simultaneously update", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 107, + 417, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 107, + 420, + 118, + 429 + ], + "score": 0.83, + "content": "n _ { s }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 417, + 506, + 431 + ], + "score": 1.0, + "content": "randomly-sampled sub-models (which are called sandwiches) and the smallest-possible sub-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 429, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 276, + 442 + ], + "score": 1.0, + "content": "model, which corresponds to keeping only", + "type": "text" + }, + { + "bbox": [ + 276, + 429, + 321, + 441 + ], + "score": 0.92, + "content": "\\lceil ( 1 - p ) l _ { i } \\rceil", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 429, + 428, + 442 + ], + "score": 1.0, + "content": "word vectors at each layer", + "type": "text" + }, + { + "bbox": [ + 428, + 430, + 433, + 439 + ], + "score": 0.69, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 429, + 505, + 442 + ], + "score": 1.0, + "content": ", using knowledge", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 452 + ], + "score": 1.0, + "content": "distillation (Hinton et al., 2015) from the full model. Here, sub-models mean models with length", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "reduction. They are trained to their prediction close to the full model’s prediction (inplace distillation).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 362, + 506, + 464 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 474, + 370, + 486 + ], + "lines": [ + { + "bbox": [ + 106, + 474, + 371, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 371, + 487 + ], + "score": 1.0, + "content": "3.2 EVOLUTIONARY SEARCH OF LENGTH CONFIGURATIONS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 505, + 583 + ], + "lines": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 509 + ], + "score": 1.0, + "content": "After training a Length-Adaptive Transformer with LengthDrop, we search for appropriate length", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "configurations for possible target computational budgets that will be given at inference time. The", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "length configuration determines the model performance in terms of both accuracy and efficiency. In", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "order to search for the optimal length configuration, we propose to use evolutionary search, similarly", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "to Cai et al. (2019) and Wang et al. (2020). This procedure is efficient, as it only requires a single", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 550, + 507, + 564 + ], + "spans": [ + { + "bbox": [ + 104, + 550, + 507, + 564 + ], + "score": 1.0, + "content": "pass through the relatively small validation set for each length configuration, unlike re-training for a", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 575 + ], + "score": 1.0, + "content": "new computational budget which requires multiple passes through a significantly larger training set", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 572, + 173, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 173, + 585 + ], + "score": 1.0, + "content": "for each budget.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 495, + 507, + 585 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 589, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "We initialize the population with constant-ratio configurations. Each configuration is created by", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 187, + 612 + ], + "score": 0.92, + "content": "l _ { i + 1 } = \\lceil ( 1 - r ) \\bar { l _ { i } } \\rceil", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 599, + 250, + 613 + ], + "score": 1.0, + "content": "for each layer", + "type": "text" + }, + { + "bbox": [ + 250, + 601, + 255, + 610 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 599, + 278, + 613 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 279, + 603, + 285, + 610 + ], + "score": 0.7, + "content": "r", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "so that the amount of computation within the initial", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 610, + 507, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 610, + 507, + 624 + ], + "score": 1.0, + "content": "population is uniformly distributed between those of the smallest and full models. At each itera-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 620, + 507, + 636 + ], + "spans": [ + { + "bbox": [ + 104, + 620, + 507, + 636 + ], + "score": 1.0, + "content": "tion, we evolve the population to consist only of configurations lie on a newly updated efficiency-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "accuracy Pareto frontier by mutation and cross-over. Mutation alters an original length configuration", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 643, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 154, + 656 + ], + "score": 0.91, + "content": "( l _ { 1 } , \\cdots , l _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 643, + 166, + 658 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 167, + 644, + 214, + 656 + ], + "score": 0.91, + "content": "( l _ { 1 } ^ { \\prime } , \\cdots , l _ { L } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 643, + 267, + 658 + ], + "score": 1.0, + "content": "by sampling", + "type": "text" + }, + { + "bbox": [ + 267, + 644, + 275, + 655 + ], + "score": 0.84, + "content": "l _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 643, + 393, + 658 + ], + "score": 1.0, + "content": "from the uniform distribution", + "type": "text" + }, + { + "bbox": [ + 394, + 644, + 447, + 656 + ], + "score": 0.91, + "content": "\\bar { \\mathcal { U } } ( l _ { i - 1 } ^ { \\prime } , l _ { i + 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 643, + 507, + 658 + ], + "score": 1.0, + "content": "with the prob-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 134, + 668 + ], + "score": 1.0, + "content": "ability", + "type": "text" + }, + { + "bbox": [ + 134, + 657, + 147, + 667 + ], + "score": 0.85, + "content": "p _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 654, + 267, + 668 + ], + "score": 1.0, + "content": "or keeping the original length", + "type": "text" + }, + { + "bbox": [ + 268, + 655, + 294, + 667 + ], + "score": 0.9, + "content": "{ \\bar { l } } _ { i } ^ { \\prime } = l _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 654, + 400, + 668 + ], + "score": 1.0, + "content": ", sweeping the layers from", + "type": "text" + }, + { + "bbox": [ + 400, + 656, + 423, + 665 + ], + "score": 0.88, + "content": "i = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 654, + 434, + 668 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 434, + 655, + 459, + 665 + ], + "score": 0.9, + "content": "i = L", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 654, + 506, + 668 + ], + "score": 1.0, + "content": ". Crossover", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "takes two length configurations and averages the lengths at each layer. Both of these operations are", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 495, + 689 + ], + "score": 1.0, + "content": "performed while ensuring the monotonicity of the lengths over the layers. We repeat this iteration", + "type": "text" + }, + { + "bbox": [ + 495, + 677, + 504, + 687 + ], + "score": 0.79, + "content": "G", + "type": "inline_equation" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 209, + 701 + ], + "score": 1.0, + "content": "times, while maintaining", + "type": "text" + }, + { + "bbox": [ + 209, + 689, + 223, + 699 + ], + "score": 0.83, + "content": "n _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 687, + 337, + 701 + ], + "score": 1.0, + "content": "mutated configurations and", + "type": "text" + }, + { + "bbox": [ + 338, + 689, + 348, + 699 + ], + "score": 0.79, + "content": "n _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "crossover’d configurations. Repeating", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "this procedure pushes the Pareto frontier further to identify the best trade-off between two objectives,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "efficiency and accuracy, without requiring any continuous relaxation of length configurations nor", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 721, + 240, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 240, + 733 + ], + "score": 1.0, + "content": "using a proxy objective function.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 46, + "bbox_fs": [ + 104, + 588, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 88, + 492, + 157 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 88, + 492, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 88, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 119, + 88, + 492, + 157 + ], + "score": 0.963, + "type": "image", + "image_path": "320388afaf5c944756a7b3f52eddc7583af0fdd53b75f8ac7bf8c3063dd11f2f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 88, + 492, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 111.0, + 492, + 134.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 134.0, + 492, + 157.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 165, + 505, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "Figure 1: Illustration of (a) word-vector elimination process in PoWER-BERT (Goyal et al., 2020)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "and (b) Drop-and-Restore process in Length-Adaptive Transformer. Yellow box and blue boxes", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 189, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 200 + ], + "score": 1.0, + "content": "imply the output of embedding layer and transformer layers, respectively. Green boxes mean vectors", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "score": 1.0, + "content": "dropped in lower layers and restored at the last layer. Red box is the task-specific layer. Though", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "word-vectors in the middle could be eliminated (or dropped), remaining vectors are left-aligned for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 220, + 408, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 408, + 234 + ], + "score": 1.0, + "content": "the better illustration. In this case, the number of transformer layers is four.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 108, + 249, + 263, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 265, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 265, + 262 + ], + "score": 1.0, + "content": "3.3 DROP-AND-RESTORE PROCESS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "The applicability of the PoWER-BERT, based on which our main contribution above was made, is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "limited to sequence-level classification, because it eliminates word vectors at each layer. In addition", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 504, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 504, + 304 + ], + "score": 1.0, + "content": "to our main contribution above, we thus propose to extend the PoWER-BERT so that it is applicable", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "to token-level classification, such as span-based question-answering. Our proposal, to which we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "refer as Drop-and-Restore, does not eliminate word vectors at each layer according to the length", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 326, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 505, + 337 + ], + "score": 1.0, + "content": "configuration but instead sets them aside until the final hidden layer. At the final hidden layer, these", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "score": 1.0, + "content": "word vectors are brought back to form the full hidden sequence, as illustrated graphically in Fig. 1.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 364, + 231, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 232, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 232, + 378 + ], + "score": 1.0, + "content": "4 EXPERIMENT SETUP", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 506, + 402 + ], + "score": 1.0, + "content": "Datasets We test the proposed approach on both sequence-level and token-level tasks, the latter", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 507, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 507, + 414 + ], + "score": 1.0, + "content": "of which could not have been done with the original PoWER-BERT unless for the proposed Drop-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "and-Restore. We use MNLI-m and SST-2 from GLUE benchmark (Wang et al., 2018), as was done", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "to test PoWER-BERT earlier, for sequence-level classification. We choose them because consistent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "accuracy scores from standard training on them due to their sufficiently large training set imply that", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "they are reliable to verify our approach. We use SQuAD 1.1 (Rajpurkar et al., 2016) for token-level", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 455, + 164, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 164, + 467 + ], + "score": 1.0, + "content": "classification.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "score": 1.0, + "content": "Evaluation metrics We use the number of floating operations (FLOPs) as a main metric to measure", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "score": 1.0, + "content": "the inference efficiency given any length configuration, as it is agnostic to the choice of underlying", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 499, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 516 + ], + "score": 1.0, + "content": "hardware, unlike other alternatives such as hardware-aware latency (Wang et al., 2020) or energy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "consumption (Henderson et al., 2020). We later demonstrate that FLOPs and wall-clock time on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "GPU and CPU correlate well with the proposed approach, which is not necessarily the case for other", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 535, + 444, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 444, + 547 + ], + "score": 1.0, + "content": "approaches, such as unstructured weight pruning (Han et al., 2015; See et al., 2016).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Pretrained transformers Since BERT was introduced by Devlin et al. (2018), it has become", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "a standard practice to start from a pretrained (masked) language model and finetune it for each", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "downstream task. We follow the same strategy in this paper and test two pretrained transformer-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "based language models; BERTBASE (Devlin et al., 2018) and DistilBERT (Sanh et al., 2019), which", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 601, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 616 + ], + "score": 1.0, + "content": "allows us to demonstrate that the usefulness and applicability of our approach are not tied to any", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "score": 1.0, + "content": "specific architectural choice, such as the number of layers and the maximum length of input sequence.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "Although we focus on BERT-based masked language models here, the proposed approach is readily", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 636, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 286, + 647 + ], + "score": 1.0, + "content": "applicable to any transformer-based models.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 660, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "Learning We train a Length-Adaptive Transformer with LengthDrop probability and LayerDrop", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 244, + 683 + ], + "score": 1.0, + "content": "probability both set to 0.2. We use", + "type": "text" + }, + { + "bbox": [ + 245, + 671, + 275, + 682 + ], + "score": 0.9, + "content": "n _ { s } = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "randomly sampled intermediate sub-models in addition to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 681, + 406, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 406, + 694 + ], + "score": 1.0, + "content": "the full model and smallest model for applying the sandwich learning rule.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "We start finetuning the pretrained transformer without Drop-and-Restore first, just as Goyal et al.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "(2020) did with PoWER-BERT. We then continue finetuning it for another five epochs with Drop-and-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "Restore. This is unlike the recommended three epochs by Devlin et al. (2018), as learning progresses", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 88, + 492, + 157 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 88, + 492, + 157 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 88, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 119, + 88, + 492, + 157 + ], + "score": 0.963, + "type": "image", + "image_path": "320388afaf5c944756a7b3f52eddc7583af0fdd53b75f8ac7bf8c3063dd11f2f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 88, + 492, + 111.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 111.0, + 492, + 134.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 134.0, + 492, + 157.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 165, + 505, + 232 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "Figure 1: Illustration of (a) word-vector elimination process in PoWER-BERT (Goyal et al., 2020)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "and (b) Drop-and-Restore process in Length-Adaptive Transformer. Yellow box and blue boxes", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 189, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 200 + ], + "score": 1.0, + "content": "imply the output of embedding layer and transformer layers, respectively. Green boxes mean vectors", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 212 + ], + "score": 1.0, + "content": "dropped in lower layers and restored at the last layer. Red box is the task-specific layer. Though", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 506, + 223 + ], + "score": 1.0, + "content": "word-vectors in the middle could be eliminated (or dropped), remaining vectors are left-aligned for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 220, + 408, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 408, + 234 + ], + "score": 1.0, + "content": "the better illustration. In this case, the number of transformer layers is four.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 108, + 249, + 263, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 248, + 265, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 265, + 262 + ], + "score": 1.0, + "content": "3.3 DROP-AND-RESTORE PROCESS", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 270, + 505, + 347 + ], + "lines": [ + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 506, + 283 + ], + "score": 1.0, + "content": "The applicability of the PoWER-BERT, based on which our main contribution above was made, is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 294 + ], + "score": 1.0, + "content": "limited to sequence-level classification, because it eliminates word vectors at each layer. In addition", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 293, + 504, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 504, + 304 + ], + "score": 1.0, + "content": "to our main contribution above, we thus propose to extend the PoWER-BERT so that it is applicable", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "to token-level classification, such as span-based question-answering. Our proposal, to which we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "refer as Drop-and-Restore, does not eliminate word vectors at each layer according to the length", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 326, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 505, + 337 + ], + "score": 1.0, + "content": "configuration but instead sets them aside until the final hidden layer. At the final hidden layer, these", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "score": 1.0, + "content": "word vectors are brought back to form the full hidden sequence, as illustrated graphically in Fig. 1.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 271, + 506, + 349 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 364, + 231, + 376 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 232, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 232, + 378 + ], + "score": 1.0, + "content": "4 EXPERIMENT SETUP", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 389, + 505, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 506, + 402 + ], + "score": 1.0, + "content": "Datasets We test the proposed approach on both sequence-level and token-level tasks, the latter", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 507, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 507, + 414 + ], + "score": 1.0, + "content": "of which could not have been done with the original PoWER-BERT unless for the proposed Drop-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "and-Restore. We use MNLI-m and SST-2 from GLUE benchmark (Wang et al., 2018), as was done", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "to test PoWER-BERT earlier, for sequence-level classification. We choose them because consistent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 446 + ], + "score": 1.0, + "content": "accuracy scores from standard training on them due to their sufficiently large training set imply that", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "they are reliable to verify our approach. We use SQuAD 1.1 (Rajpurkar et al., 2016) for token-level", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 455, + 164, + 467 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 164, + 467 + ], + "score": 1.0, + "content": "classification.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 389, + 507, + 467 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "score": 1.0, + "content": "Evaluation metrics We use the number of floating operations (FLOPs) as a main metric to measure", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 506, + 505 + ], + "score": 1.0, + "content": "the inference efficiency given any length configuration, as it is agnostic to the choice of underlying", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 499, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 506, + 516 + ], + "score": 1.0, + "content": "hardware, unlike other alternatives such as hardware-aware latency (Wang et al., 2020) or energy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "consumption (Henderson et al., 2020). We later demonstrate that FLOPs and wall-clock time on", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 536 + ], + "score": 1.0, + "content": "GPU and CPU correlate well with the proposed approach, which is not necessarily the case for other", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 535, + 444, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 444, + 547 + ], + "score": 1.0, + "content": "approaches, such as unstructured weight pruning (Han et al., 2015; See et al., 2016).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 478, + 506, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 558, + 505, + 646 + ], + "lines": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "Pretrained transformers Since BERT was introduced by Devlin et al. (2018), it has become", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "a standard practice to start from a pretrained (masked) language model and finetune it for each", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 594 + ], + "score": 1.0, + "content": "downstream task. We follow the same strategy in this paper and test two pretrained transformer-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "based language models; BERTBASE (Devlin et al., 2018) and DistilBERT (Sanh et al., 2019), which", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 601, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 616 + ], + "score": 1.0, + "content": "allows us to demonstrate that the usefulness and applicability of our approach are not tied to any", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 627 + ], + "score": 1.0, + "content": "specific architectural choice, such as the number of layers and the maximum length of input sequence.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "score": 1.0, + "content": "Although we focus on BERT-based masked language models here, the proposed approach is readily", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 636, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 286, + 647 + ], + "score": 1.0, + "content": "applicable to any transformer-based models.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 559, + 506, + 647 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 660, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 673 + ], + "score": 1.0, + "content": "Learning We train a Length-Adaptive Transformer with LengthDrop probability and LayerDrop", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 671, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 244, + 683 + ], + "score": 1.0, + "content": "probability both set to 0.2. We use", + "type": "text" + }, + { + "bbox": [ + 245, + 671, + 275, + 682 + ], + "score": 0.9, + "content": "n _ { s } = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 671, + 505, + 683 + ], + "score": 1.0, + "content": "randomly sampled intermediate sub-models in addition to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 681, + 406, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 406, + 694 + ], + "score": 1.0, + "content": "the full model and smallest model for applying the sandwich learning rule.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 659, + 505, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "We start finetuning the pretrained transformer without Drop-and-Restore first, just as Goyal et al.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "(2020) did with PoWER-BERT. We then continue finetuning it for another five epochs with Drop-and-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "Restore. This is unlike the recommended three epochs by Devlin et al. (2018), as learning progresses", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "slower due to a higher level of stochasticity introduced by LengthDrop and LayerDrop. We use the", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "score": 1.0, + "content": "batch size of 32, the learning rate of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 253, + 387, + 281, + 397 + ], + "score": 0.85, + "content": "5 e - 5", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 281, + 387, + 367, + 399 + ], + "score": 1.0, + "content": "for SQuAD v1.1 and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 367, + 387, + 396, + 397 + ], + "score": 0.84, + "content": "2 e - 5", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 396, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "for MNLI-m and SST, and", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 398, + 453, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 453, + 410 + ], + "score": 1.0, + "content": "the maximum sequence length of 384 for SQuAD v1.1 and 128 for MNLI-m and SST.", + "type": "text", + "cross_page": true + } + ], + "index": 17 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 699, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 79, + 391, + 277 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 79, + 391, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 79, + 391, + 277 + ], + "spans": [ + { + "bbox": [ + 111, + 79, + 391, + 277 + ], + "score": 0.935, + "type": "image", + "image_path": "3ea54ad6f6a4d01240155dd50bd06714bd078c67453a78cfd7ce1b9404ef27c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 79, + 391, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 145.0, + 391, + 211.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 211.0, + 391, + 277.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 285, + 396, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 285, + 397, + 298 + ], + "spans": [ + { + "bbox": [ + 107, + 285, + 397, + 298 + ], + "score": 1.0, + "content": "Figure 2: Pareto curves of F1 score and FLOPs on SQuAD 1.1 (Rajpurkar", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 297, + 396, + 309 + ], + "spans": [ + { + "bbox": [ + 107, + 297, + 291, + 309 + ], + "score": 1.0, + "content": "et al., 2016). We apply the proposed method to", + "type": "text" + }, + { + "bbox": [ + 291, + 297, + 331, + 308 + ], + "score": 0.77, + "content": "\\mathbf { B E R T _ { B a s e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 297, + 396, + 309 + ], + "score": 1.0, + "content": "(solid lines) and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 306, + 397, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 397, + 321 + ], + "score": 1.0, + "content": "DistilBERT (dotted lines). For each model, we draw three curves using", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 318, + 397, + 332 + ], + "spans": [ + { + "bbox": [ + 107, + 318, + 397, + 332 + ], + "score": 1.0, + "content": "(1) standard finetuned transformer with constant-rate length reduction, (2)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 330, + 397, + 342 + ], + "spans": [ + { + "bbox": [ + 107, + 330, + 397, + 342 + ], + "score": 1.0, + "content": "Length-Adaptive Transformer with constant-rate length reduction, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 341, + 396, + 353 + ], + "spans": [ + { + "bbox": [ + 107, + 341, + 396, + 353 + ], + "score": 1.0, + "content": "(3) Length-Adaptive Transformer with length configurations obtained", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 352, + 227, + 364 + ], + "spans": [ + { + "bbox": [ + 108, + 352, + 227, + 364 + ], + "score": 1.0, + "content": "from the evolutionary search.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "image", + "bbox": [ + 408, + 81, + 497, + 297 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 408, + 81, + 497, + 297 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 408, + 81, + 497, + 297 + ], + "spans": [ + { + "bbox": [ + 408, + 81, + 497, + 297 + ], + "score": 0.934, + "type": "image", + "image_path": "4608176aa7c7dba0b4748567c1c64abe798b711bebf826c4b323894a4bfb2b12.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 408, + 81, + 497, + 297 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 408, + 305, + 498, + 350 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 407, + 305, + 498, + 317 + ], + "spans": [ + { + "bbox": [ + 407, + 305, + 498, + 317 + ], + "score": 1.0, + "content": "Figure 3: Correlation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 408, + 316, + 498, + 328 + ], + "spans": [ + { + "bbox": [ + 408, + 316, + 498, + 328 + ], + "score": 1.0, + "content": "between FLOPs and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 408, + 327, + 498, + 339 + ], + "spans": [ + { + "bbox": [ + 408, + 327, + 498, + 339 + ], + "score": 1.0, + "content": "latency with different", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 408, + 338, + 498, + 351 + ], + "spans": [ + { + "bbox": [ + 408, + 338, + 498, + 351 + ], + "score": 1.0, + "content": "length configurations.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + } + ], + "index": 8.75 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 504, + 409 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "slower due to a higher level of stochasticity introduced by LengthDrop and LayerDrop. We use the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "score": 1.0, + "content": "batch size of 32, the learning rate of", + "type": "text" + }, + { + "bbox": [ + 253, + 387, + 281, + 397 + ], + "score": 0.85, + "content": "5 e - 5", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 387, + 367, + 399 + ], + "score": 1.0, + "content": "for SQuAD v1.1 and", + "type": "text" + }, + { + "bbox": [ + 367, + 387, + 396, + 397 + ], + "score": 0.84, + "content": "2 e - 5", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "for MNLI-m and SST, and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 398, + 453, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 453, + 410 + ], + "score": 1.0, + "content": "the maximum sequence length of 384 for SQuAD v1.1 and 128 for MNLI-m and SST.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 422, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 199, + 435 + ], + "score": 1.0, + "content": "Search We run up to", + "type": "text" + }, + { + "bbox": [ + 200, + 422, + 232, + 432 + ], + "score": 0.9, + "content": "G = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 421, + 389, + 435 + ], + "score": 1.0, + "content": "iterations of evolutionary search, using", + "type": "text" + }, + { + "bbox": [ + 389, + 423, + 428, + 433 + ], + "score": 0.9, + "content": "n _ { m } = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 421, + 507, + 435 + ], + "score": 1.0, + "content": "mutated configura-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 231, + 446 + ], + "score": 1.0, + "content": "tions with mutation probability", + "type": "text" + }, + { + "bbox": [ + 232, + 433, + 272, + 445 + ], + "score": 0.9, + "content": "p _ { m } = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 433, + 289, + 446 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 433, + 325, + 444 + ], + "score": 0.91, + "content": "n _ { c } = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "crossover’d configurations, to find the Pareto", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 443, + 248, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 248, + 457 + ], + "score": 1.0, + "content": "frontier of accuracy and efficiency.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 472, + 253, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 255, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 255, + 487 + ], + "score": 1.0, + "content": "5 RESULTS AND ANALYSIS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "Efficiency-accuracy trade-off We use SQuAD 1.1 to examine the effect of the proposed approach", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 508, + 507, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 507, + 523 + ], + "score": 1.0, + "content": "on the efficiency-accuracy trade-off. When the underlying classifier was not trained with LengthDrop,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "as proposed in this paper, the accuracy drops even more dramatically as more word vectors are dropped", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "at each layer. The difference between standard transformer and Length-Adaptive Transformer is stark", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "score": 1.0, + "content": "in Fig. 2. This verifies the importance of training a transformer in a way that makes it malleable for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 553, + 236, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 236, + 565 + ], + "score": 1.0, + "content": "inference-time re-configuration.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 625 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "When the model was trained with the proposed LengthDrop, we notice the efficacy of the proposed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 579, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 595 + ], + "score": 1.0, + "content": "approach of using evolutionary search to find the optimal trade-off between inference efficiency", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "and accuracy. The trade-off curve from the proposed search strategy has a larger area-under-curve", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "(AUC) than when constant-rate length reduction was used to meet a target computational budget. It", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 613, + 429, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 429, + 627 + ], + "score": 1.0, + "content": "demonstrates the importance of using both LengthDrop and evolutionary search.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 630, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 629, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 644 + ], + "score": 1.0, + "content": "We make a minor observation that the proposed approach ends up with a significantly higher accuracy", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 429, + 654 + ], + "score": 1.0, + "content": "than DistillBERT when enough computational budget is allowed for inference", + "type": "text" + }, + { + "bbox": [ + 430, + 641, + 502, + 653 + ], + "score": 0.8, + "content": "( \\log { \\mathrm { F L O P s } } > 1 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 641, + 506, + 654 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "score": 1.0, + "content": "This makes our approach desirable in a wide array of scenarios, as it does not require any additional", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "pretraining stage, as does DistilBERT. With a severe constraint on the computational budget, the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "score": 1.0, + "content": "proposed approach could be used on DistilBERT to significantly improve the efficiency without", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 685, + 221, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 221, + 699 + ], + "score": 1.0, + "content": "compromising the accuracy.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Maximizing inference efficiency We consider all three tasks, SQuAD 1.1, MNLI-m and SST-2,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "and investigate how much efficiency can be gained by the proposed approach with minimal sacrifice of", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 79, + 391, + 277 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 79, + 391, + 277 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 79, + 391, + 277 + ], + "spans": [ + { + "bbox": [ + 111, + 79, + 391, + 277 + ], + "score": 0.935, + "type": "image", + "image_path": "3ea54ad6f6a4d01240155dd50bd06714bd078c67453a78cfd7ce1b9404ef27c2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 79, + 391, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 145.0, + 391, + 211.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 211.0, + 391, + 277.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 285, + 396, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 285, + 397, + 298 + ], + "spans": [ + { + "bbox": [ + 107, + 285, + 397, + 298 + ], + "score": 1.0, + "content": "Figure 2: Pareto curves of F1 score and FLOPs on SQuAD 1.1 (Rajpurkar", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 297, + 396, + 309 + ], + "spans": [ + { + "bbox": [ + 107, + 297, + 291, + 309 + ], + "score": 1.0, + "content": "et al., 2016). We apply the proposed method to", + "type": "text" + }, + { + "bbox": [ + 291, + 297, + 331, + 308 + ], + "score": 0.77, + "content": "\\mathbf { B E R T _ { B a s e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 297, + 396, + 309 + ], + "score": 1.0, + "content": "(solid lines) and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 306, + 397, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 397, + 321 + ], + "score": 1.0, + "content": "DistilBERT (dotted lines). For each model, we draw three curves using", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 318, + 397, + 332 + ], + "spans": [ + { + "bbox": [ + 107, + 318, + 397, + 332 + ], + "score": 1.0, + "content": "(1) standard finetuned transformer with constant-rate length reduction, (2)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 330, + 397, + 342 + ], + "spans": [ + { + "bbox": [ + 107, + 330, + 397, + 342 + ], + "score": 1.0, + "content": "Length-Adaptive Transformer with constant-rate length reduction, and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 341, + 396, + 353 + ], + "spans": [ + { + "bbox": [ + 107, + 341, + 396, + 353 + ], + "score": 1.0, + "content": "(3) Length-Adaptive Transformer with length configurations obtained", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 352, + 227, + 364 + ], + "spans": [ + { + "bbox": [ + 108, + 352, + 227, + 364 + ], + "score": 1.0, + "content": "from the evolutionary search.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7 + } + ], + "index": 4.0 + }, + { + "type": "image", + "bbox": [ + 408, + 81, + 497, + 297 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 408, + 81, + 497, + 297 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 408, + 81, + 497, + 297 + ], + "spans": [ + { + "bbox": [ + 408, + 81, + 497, + 297 + ], + "score": 0.934, + "type": "image", + "image_path": "4608176aa7c7dba0b4748567c1c64abe798b711bebf826c4b323894a4bfb2b12.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 408, + 81, + 497, + 297 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 408, + 305, + 498, + 350 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 407, + 305, + 498, + 317 + ], + "spans": [ + { + "bbox": [ + 407, + 305, + 498, + 317 + ], + "score": 1.0, + "content": "Figure 3: Correlation", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 408, + 316, + 498, + 328 + ], + "spans": [ + { + "bbox": [ + 408, + 316, + 498, + 328 + ], + "score": 1.0, + "content": "between FLOPs and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 408, + 327, + 498, + 339 + ], + "spans": [ + { + "bbox": [ + 408, + 327, + 498, + 339 + ], + "score": 1.0, + "content": "latency with different", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 408, + 338, + 498, + 351 + ], + "spans": [ + { + "bbox": [ + 408, + 338, + 498, + 351 + ], + "score": 1.0, + "content": "length configurations.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + } + ], + "index": 8.75 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 504, + 409 + ], + "lines": [], + "index": 16, + "bbox_fs": [ + 105, + 375, + 505, + 410 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 422, + 505, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 199, + 435 + ], + "score": 1.0, + "content": "Search We run up to", + "type": "text" + }, + { + "bbox": [ + 200, + 422, + 232, + 432 + ], + "score": 0.9, + "content": "G = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 421, + 389, + 435 + ], + "score": 1.0, + "content": "iterations of evolutionary search, using", + "type": "text" + }, + { + "bbox": [ + 389, + 423, + 428, + 433 + ], + "score": 0.9, + "content": "n _ { m } = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 421, + 507, + 435 + ], + "score": 1.0, + "content": "mutated configura-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 433, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 231, + 446 + ], + "score": 1.0, + "content": "tions with mutation probability", + "type": "text" + }, + { + "bbox": [ + 232, + 433, + 272, + 445 + ], + "score": 0.9, + "content": "p _ { m } = 0 . 5", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 433, + 289, + 446 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 433, + 325, + 444 + ], + "score": 0.91, + "content": "n _ { c } = 3 0", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 433, + 505, + 446 + ], + "score": 1.0, + "content": "crossover’d configurations, to find the Pareto", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 443, + 248, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 248, + 457 + ], + "score": 1.0, + "content": "frontier of accuracy and efficiency.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 421, + 507, + 457 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 472, + 253, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 255, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 255, + 487 + ], + "score": 1.0, + "content": "5 RESULTS AND ANALYSIS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 505, + 510 + ], + "score": 1.0, + "content": "Efficiency-accuracy trade-off We use SQuAD 1.1 to examine the effect of the proposed approach", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 508, + 507, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 507, + 523 + ], + "score": 1.0, + "content": "on the efficiency-accuracy trade-off. When the underlying classifier was not trained with LengthDrop,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "as proposed in this paper, the accuracy drops even more dramatically as more word vectors are dropped", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "at each layer. The difference between standard transformer and Length-Adaptive Transformer is stark", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 555 + ], + "score": 1.0, + "content": "in Fig. 2. This verifies the importance of training a transformer in a way that makes it malleable for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 553, + 236, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 236, + 565 + ], + "score": 1.0, + "content": "inference-time re-configuration.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 498, + 507, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 625 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 583 + ], + "score": 1.0, + "content": "When the model was trained with the proposed LengthDrop, we notice the efficacy of the proposed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 579, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 595 + ], + "score": 1.0, + "content": "approach of using evolutionary search to find the optimal trade-off between inference efficiency", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 605 + ], + "score": 1.0, + "content": "and accuracy. The trade-off curve from the proposed search strategy has a larger area-under-curve", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "(AUC) than when constant-rate length reduction was used to meet a target computational budget. It", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 613, + 429, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 429, + 627 + ], + "score": 1.0, + "content": "demonstrates the importance of using both LengthDrop and evolutionary search.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 569, + 506, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 630, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 629, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 644 + ], + "score": 1.0, + "content": "We make a minor observation that the proposed approach ends up with a significantly higher accuracy", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 641, + 506, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 429, + 654 + ], + "score": 1.0, + "content": "than DistillBERT when enough computational budget is allowed for inference", + "type": "text" + }, + { + "bbox": [ + 430, + 641, + 502, + 653 + ], + "score": 0.8, + "content": "( \\log { \\mathrm { F L O P s } } > 1 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 641, + 506, + 654 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 506, + 665 + ], + "score": 1.0, + "content": "This makes our approach desirable in a wide array of scenarios, as it does not require any additional", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 662, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 104, + 662, + 506, + 676 + ], + "score": 1.0, + "content": "pretraining stage, as does DistilBERT. With a severe constraint on the computational budget, the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 674, + 506, + 687 + ], + "score": 1.0, + "content": "proposed approach could be used on DistilBERT to significantly improve the efficiency without", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 685, + 221, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 221, + 699 + ], + "score": 1.0, + "content": "compromising the accuracy.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 629, + 506, + 699 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Maximizing inference efficiency We consider all three tasks, SQuAD 1.1, MNLI-m and SST-2,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "and investigate how much efficiency can be gained by the proposed approach with minimal sacrifice of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "accuracy. First, we look at how much efficiency could be gained without losing on the accuracy. That", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 297 + ], + "score": 1.0, + "content": "is, we use the length configuration that maximizes the inference efficiency (i.e., minimize the FLOPs)", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "score": 1.0, + "content": "while ensuring that the accuracy is above or same as the accuracy of the standard approach without", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "any drop of word vectors. The results are presented in the rows marked with Length-Adaptive† from", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 252, + 331 + ], + "score": 1.0, + "content": "Table 1. For example, in the case of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 252, + 318, + 292, + 329 + ], + "score": 0.62, + "content": "\\mathbf { B E R T _ { B a s e } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 292, + 317, + 506, + 331 + ], + "score": 1.0, + "content": ", the proposed approach reduce FLOPs by more than", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 329, + 211, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 211, + 340 + ], + "score": 1.0, + "content": "half across all three tasks.", + "type": "text", + "cross_page": true + } + ], + "index": 22 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 112, + 80, + 394, + 190 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 112, + 80, + 394, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 80, + 394, + 190 + ], + "spans": [ + { + "bbox": [ + 112, + 80, + 394, + 190 + ], + "score": 0.981, + "html": "
ModelSQuAD 1.1MNLI-mSST-2
Pretrained TransformerMethodF1FLOPsAccFLOPsAccFLOPs
BERTBaseStandard88.51.00x84.41.00x92.81.00x
Length-Adaptive*89.60.89x85.00.58x93.10.36x
Length-Adaptivet88.70.45x84.40.35x92.80.35x
Standard85.81.00x80.91.00x90.61.00x
DistilBERT1Length-Adaptive*86.30.81x81.50.56x92.00.55x
Length-Adaptive†85.90.59x81.30.54x91.70.54x
", + "type": "table", + "image_path": "f0829b1fbd9636baa6152ad5a45e5f7dbf08c6a7333b2fd362453dfa90284833.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 80, + 394, + 116.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 116.66666666666666, + 394, + 153.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 153.33333333333331, + 394, + 189.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 109, + 198, + 398, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 197, + 398, + 210 + ], + "spans": [ + { + "bbox": [ + 108, + 197, + 398, + 210 + ], + "score": 1.0, + "content": "Table 1: Comparison results of standard Transformer and Length-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 208, + 398, + 221 + ], + "spans": [ + { + "bbox": [ + 109, + 208, + 398, + 221 + ], + "score": 1.0, + "content": "Adaptive Transformer. Among length configurations on the Pareto fron-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 219, + 399, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 219, + 399, + 232 + ], + "score": 1.0, + "content": "tier of Length-Adaptive Transformer, we pick two representative points:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 231, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 108, + 231, + 398, + 242 + ], + "score": 1.0, + "content": "Length-Adaptive? and Length-Adaptive† as the most efficient one while", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 241, + 398, + 255 + ], + "spans": [ + { + "bbox": [ + 108, + 241, + 398, + 255 + ], + "score": 1.0, + "content": "having the highest accuracy and the accuracy higher than (or equal to)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 252, + 252, + 266 + ], + "spans": [ + { + "bbox": [ + 108, + 252, + 252, + 266 + ], + "score": 1.0, + "content": "standard Transformer, respectively.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 410, + 86, + 498, + 182 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 410, + 86, + 498, + 182 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 410, + 86, + 498, + 182 + ], + "spans": [ + { + "bbox": [ + 410, + 86, + 498, + 182 + ], + "score": 0.961, + "type": "image", + "image_path": "e4a6ecfc1cdcda5b3130f2509429e44b544d994935e6b44f36893f811295f752.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 410, + 86, + 498, + 134.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 410, + 134.0, + 498, + 182.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 409, + 191, + 500, + 257 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 408, + 188, + 502, + 203 + ], + "spans": [ + { + "bbox": [ + 408, + 188, + 502, + 203 + ], + "score": 1.0, + "content": "Figure 4: Exam-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 408, + 201, + 501, + 213 + ], + "spans": [ + { + "bbox": [ + 408, + 201, + 501, + 213 + ], + "score": 1.0, + "content": "ple of area under", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 409, + 213, + 499, + 223 + ], + "spans": [ + { + "bbox": [ + 409, + 213, + 499, + 223 + ], + "score": 1.0, + "content": "Pareto curve as the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 409, + 223, + 500, + 235 + ], + "spans": [ + { + "bbox": [ + 409, + 223, + 500, + 235 + ], + "score": 1.0, + "content": "evolutionary search of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 408, + 233, + 500, + 246 + ], + "spans": [ + { + "bbox": [ + 408, + 233, + 500, + 246 + ], + "score": 1.0, + "content": "lenth configurations", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 408, + 246, + 450, + 257 + ], + "spans": [ + { + "bbox": [ + 408, + 246, + 450, + 257 + ], + "score": 1.0, + "content": "proceeds.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 273, + 506, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 286 + ], + "score": 1.0, + "content": "accuracy. First, we look at how much efficiency could be gained without losing on the accuracy. That", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 285, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 297 + ], + "score": 1.0, + "content": "is, we use the length configuration that maximizes the inference efficiency (i.e., minimize the FLOPs)", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 506, + 308 + ], + "score": 1.0, + "content": "while ensuring that the accuracy is above or same as the accuracy of the standard approach without", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "any drop of word vectors. The results are presented in the rows marked with Length-Adaptive† from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 252, + 331 + ], + "score": 1.0, + "content": "Table 1. For example, in the case of", + "type": "text" + }, + { + "bbox": [ + 252, + 318, + 292, + 329 + ], + "score": 0.62, + "content": "\\mathbf { B E R T _ { B a s e } }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 317, + 506, + 331 + ], + "score": 1.0, + "content": ", the proposed approach reduce FLOPs by more than", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 329, + 211, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 211, + 340 + ], + "score": 1.0, + "content": "half across all three tasks.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 504, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 504, + 358 + ], + "score": 1.0, + "content": "From Fig. 2, we have observed that the proposed Length-Adaptive Transformer generalize better than", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "the standard, base model in some cases. We thus try to maximize both the inference efficiency and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "accuracy, in order to see whether it is possible for the proposed algorithm to find a length configuration", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 378, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 391 + ], + "score": 1.0, + "content": "that both maximizes inference efficiency and improves accuracy. We present the results in the rows", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "marked with Length-Adaptive? from Table 1. For all cases, Length-Adaptive Transformer achieves", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "higher accuracy than a standard transformer does while reducing FLOPs significantly. Although it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "is not apparent from the table, tor MNLI-m and SST-2, the accuracy of the smallest sub-model is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 423, + 356, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 356, + 435 + ], + "score": 1.0, + "content": "already greater than or equal to that of a standard transformer.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 445, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "FLOPs vs. Latency As has been discussed in recent literature (see, e.g., (Li et al., 2020; Chin", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "et al., 2020)), FLOPs is not a perfect indicator of the real latency measured in wall-clock time, as the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 468, + 504, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 504, + 478 + ], + "score": 1.0, + "content": "latter is affected by the combination of hardware choice and network architecture. To understand", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 492 + ], + "score": 1.0, + "content": "the real-world impact of the proposed approach, we study the relationship between FLOPs, obtained", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 488, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 504 + ], + "score": 1.0, + "content": "by the proposed procedure, and wall-clock time measured on both CPU and GPU by measuring", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "score": 1.0, + "content": "them while varying length configurations. As shown in Fig. 3, FLOPs and latency exhibit near-linear", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 297, + 524 + ], + "score": 1.0, + "content": "correlation on GPU, when the minibatch size is", + "type": "text" + }, + { + "bbox": [ + 297, + 512, + 319, + 523 + ], + "score": 0.87, + "content": "\\geq 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 512, + 506, + 524 + ], + "score": 1.0, + "content": ", and regardless of the minibatch size, on CPU.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "In other words, the reduction in FLOPs with the proposed approach directly implies the reduction in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 532, + 173, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 173, + 546 + ], + "score": 1.0, + "content": "wall-clock time.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 557, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 570 + ], + "score": 1.0, + "content": "Convergence of search Although the proposed approach is efficient in that it requires only one", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "round of training, it needs a separate search stage for each target budget. It is important for evolution-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "ary search to converge quickly in the number of forward sweeps of a validation set. As exemplified in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 591, + 372, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 372, + 602 + ], + "score": 1.0, + "content": "Fig. 4, evolutionary search converges after about fifteen iterations.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "title", + "bbox": [ + 108, + 617, + 211, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 213, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 213, + 632 + ], + "score": 1.0, + "content": "6 RELATED WORK", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 641, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 655 + ], + "score": 1.0, + "content": "The main purpose of the proposed algorithm is to improve the inference efficiency of a large-scale", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "transformer. This goal has been pursued from various directions, and in this section, we provide a", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 506, + 677 + ], + "score": 1.0, + "content": "brief overview of these earlier, and some concurrent, attempts in the context of the proposed approach.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "Weight pruning Weight pruning (Han et al., 2015) focuses on reducing the number of parameters", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "which directly reflects the memory footprint of a model and indirectly correlates with inference speed.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "However, their actual speed-up in runtime is usually not significant, especially while executing a", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 721, + 456, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 456, + 732 + ], + "score": 1.0, + "content": "model with parallel computation using GPU devices (Tang et al., 2018; Li et al., 2020).", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 112, + 80, + 394, + 190 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 112, + 80, + 394, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 80, + 394, + 190 + ], + "spans": [ + { + "bbox": [ + 112, + 80, + 394, + 190 + ], + "score": 0.981, + "html": "
ModelSQuAD 1.1MNLI-mSST-2
Pretrained TransformerMethodF1FLOPsAccFLOPsAccFLOPs
BERTBaseStandard88.51.00x84.41.00x92.81.00x
Length-Adaptive*89.60.89x85.00.58x93.10.36x
Length-Adaptivet88.70.45x84.40.35x92.80.35x
Standard85.81.00x80.91.00x90.61.00x
DistilBERT1Length-Adaptive*86.30.81x81.50.56x92.00.55x
Length-Adaptive†85.90.59x81.30.54x91.70.54x
", + "type": "table", + "image_path": "f0829b1fbd9636baa6152ad5a45e5f7dbf08c6a7333b2fd362453dfa90284833.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 112, + 80, + 394, + 116.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 112, + 116.66666666666666, + 394, + 153.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 153.33333333333331, + 394, + 189.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 109, + 198, + 398, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 197, + 398, + 210 + ], + "spans": [ + { + "bbox": [ + 108, + 197, + 398, + 210 + ], + "score": 1.0, + "content": "Table 1: Comparison results of standard Transformer and Length-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 208, + 398, + 221 + ], + "spans": [ + { + "bbox": [ + 109, + 208, + 398, + 221 + ], + "score": 1.0, + "content": "Adaptive Transformer. Among length configurations on the Pareto fron-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 219, + 399, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 219, + 399, + 232 + ], + "score": 1.0, + "content": "tier of Length-Adaptive Transformer, we pick two representative points:", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 231, + 398, + 242 + ], + "spans": [ + { + "bbox": [ + 108, + 231, + 398, + 242 + ], + "score": 1.0, + "content": "Length-Adaptive? and Length-Adaptive† as the most efficient one while", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 241, + 398, + 255 + ], + "spans": [ + { + "bbox": [ + 108, + 241, + 398, + 255 + ], + "score": 1.0, + "content": "having the highest accuracy and the accuracy higher than (or equal to)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 108, + 252, + 252, + 266 + ], + "spans": [ + { + "bbox": [ + 108, + 252, + 252, + 266 + ], + "score": 1.0, + "content": "standard Transformer, respectively.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "image", + "bbox": [ + 410, + 86, + 498, + 182 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 410, + 86, + 498, + 182 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 410, + 86, + 498, + 182 + ], + "spans": [ + { + "bbox": [ + 410, + 86, + 498, + 182 + ], + "score": 0.961, + "type": "image", + "image_path": "e4a6ecfc1cdcda5b3130f2509429e44b544d994935e6b44f36893f811295f752.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 410, + 86, + 498, + 134.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 410, + 134.0, + 498, + 182.0 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 409, + 191, + 500, + 257 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 408, + 188, + 502, + 203 + ], + "spans": [ + { + "bbox": [ + 408, + 188, + 502, + 203 + ], + "score": 1.0, + "content": "Figure 4: Exam-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 408, + 201, + 501, + 213 + ], + "spans": [ + { + "bbox": [ + 408, + 201, + 501, + 213 + ], + "score": 1.0, + "content": "ple of area under", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 409, + 213, + 499, + 223 + ], + "spans": [ + { + "bbox": [ + 409, + 213, + 499, + 223 + ], + "score": 1.0, + "content": "Pareto curve as the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 409, + 223, + 500, + 235 + ], + "spans": [ + { + "bbox": [ + 409, + 223, + 500, + 235 + ], + "score": 1.0, + "content": "evolutionary search of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 408, + 233, + 500, + 246 + ], + "spans": [ + { + "bbox": [ + 408, + 233, + 500, + 246 + ], + "score": 1.0, + "content": "lenth configurations", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 408, + 246, + 450, + 257 + ], + "spans": [ + { + "bbox": [ + 408, + 246, + 450, + 257 + ], + "score": 1.0, + "content": "proceeds.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 273, + 506, + 340 + ], + "lines": [], + "index": 19.5, + "bbox_fs": [ + 105, + 274, + 506, + 340 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 504, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 504, + 358 + ], + "score": 1.0, + "content": "From Fig. 2, we have observed that the proposed Length-Adaptive Transformer generalize better than", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 369 + ], + "score": 1.0, + "content": "the standard, base model in some cases. We thus try to maximize both the inference efficiency and", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "accuracy, in order to see whether it is possible for the proposed algorithm to find a length configuration", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 378, + 505, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 505, + 391 + ], + "score": 1.0, + "content": "that both maximizes inference efficiency and improves accuracy. We present the results in the rows", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "marked with Length-Adaptive? from Table 1. For all cases, Length-Adaptive Transformer achieves", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "higher accuracy than a standard transformer does while reducing FLOPs significantly. Although it", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 424 + ], + "score": 1.0, + "content": "is not apparent from the table, tor MNLI-m and SST-2, the accuracy of the smallest sub-model is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 423, + 356, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 356, + 435 + ], + "score": 1.0, + "content": "already greater than or equal to that of a standard transformer.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 346, + 506, + 435 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 445, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 458 + ], + "score": 1.0, + "content": "FLOPs vs. Latency As has been discussed in recent literature (see, e.g., (Li et al., 2020; Chin", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 469 + ], + "score": 1.0, + "content": "et al., 2020)), FLOPs is not a perfect indicator of the real latency measured in wall-clock time, as the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 468, + 504, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 504, + 478 + ], + "score": 1.0, + "content": "latter is affected by the combination of hardware choice and network architecture. To understand", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 492 + ], + "score": 1.0, + "content": "the real-world impact of the proposed approach, we study the relationship between FLOPs, obtained", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 488, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 504 + ], + "score": 1.0, + "content": "by the proposed procedure, and wall-clock time measured on both CPU and GPU by measuring", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 512 + ], + "score": 1.0, + "content": "them while varying length configurations. As shown in Fig. 3, FLOPs and latency exhibit near-linear", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 512, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 297, + 524 + ], + "score": 1.0, + "content": "correlation on GPU, when the minibatch size is", + "type": "text" + }, + { + "bbox": [ + 297, + 512, + 319, + 523 + ], + "score": 0.87, + "content": "\\geq 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 512, + 506, + 524 + ], + "score": 1.0, + "content": ", and regardless of the minibatch size, on CPU.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 536 + ], + "score": 1.0, + "content": "In other words, the reduction in FLOPs with the proposed approach directly implies the reduction in", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 532, + 173, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 173, + 546 + ], + "score": 1.0, + "content": "wall-clock time.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 445, + 506, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 557, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 556, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 570 + ], + "score": 1.0, + "content": "Convergence of search Although the proposed approach is efficient in that it requires only one", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "round of training, it needs a separate search stage for each target budget. It is important for evolution-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "ary search to converge quickly in the number of forward sweeps of a validation set. As exemplified in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 591, + 372, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 372, + 602 + ], + "score": 1.0, + "content": "Fig. 4, evolutionary search converges after about fifteen iterations.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 556, + 506, + 602 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 617, + 211, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 213, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 213, + 632 + ], + "score": 1.0, + "content": "6 RELATED WORK", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 642, + 505, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 641, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 655 + ], + "score": 1.0, + "content": "The main purpose of the proposed algorithm is to improve the inference efficiency of a large-scale", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 505, + 666 + ], + "score": 1.0, + "content": "transformer. This goal has been pursued from various directions, and in this section, we provide a", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 506, + 677 + ], + "score": 1.0, + "content": "brief overview of these earlier, and some concurrent, attempts in the context of the proposed approach.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 641, + 506, + 677 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "Weight pruning Weight pruning (Han et al., 2015) focuses on reducing the number of parameters", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "which directly reflects the memory footprint of a model and indirectly correlates with inference speed.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "However, their actual speed-up in runtime is usually not significant, especially while executing a", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 721, + 456, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 456, + 732 + ], + "score": 1.0, + "content": "model with parallel computation using GPU devices (Tang et al., 2018; Li et al., 2020).", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 687, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 170 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Adaptive architecture There are three major axes along which computation can be reduced in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "a neural network; (1) input size/length, (2) network depth and (3) network width. The proposed", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 117 + ], + "score": 1.0, + "content": "approach, based on PoWER-BERT, adaptively reduces the input length as the input sequence is", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "processed by the transformer layers. In our knowledge, Goyal et al. (2020) is the first work in this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "direction for transformers. More recently, Funnel-Transformer (Dai et al., 2020) and multi-scale", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "score": 1.0, + "content": "transformer language models (Subramanian et al., 2020) also successfully reduce sequence length in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "the middle and rescale to full length for the final computation. However, their inference complexity is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 455, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 455, + 173 + ], + "score": 1.0, + "content": "fixed unlike PoWER-BERT because they are not designed for the control of efficiency.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 176, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "LayerDrop (Fan et al., 2019) drops random layers during the training to be robust to pruning", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 504, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 504, + 200 + ], + "score": 1.0, + "content": "inspired by Huang et al. (2016). Word-level adaptive depth in Elbayad et al. (2019) might seemingly", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "resemble with length reduction, but word vectors reached the maximal layer are used for self-attention", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 210, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 506, + 221 + ], + "score": 1.0, + "content": "computation without updating themselves. Escaping a network early (Teerapittayanon et al., 2016;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "Huang et al., 2017) based on the confidence of the prediction (Xin et al., 2020; Schwartz et al., 2020;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "Liu et al., 2020) also offers a control over accuracy-efficiency trade-off, but it is difficult to tune a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 495, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 495, + 254 + ], + "score": 1.0, + "content": "threshold for a desired computational budget because of the example-wise adaptive computation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 507, + 272 + ], + "score": 1.0, + "content": "DynaBERT (Hou et al., 2020) can run at adaptive width (the number of attention heads and interme-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "diate hidden dimension) and depth. Hardware-aware Transformers (Wang et al., 2020) construct a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "score": 1.0, + "content": "design space with arbitrary encoder-decoder attention and heterogeneous layers in terms of different", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 434, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 434, + 304 + ], + "score": 1.0, + "content": "numbers of layers, attention heads, hidden dimension, and embedding dimension.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Structured dropout A major innovation we introduce over the existing PoWER-BERT is the use of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "stochastic, structured regularization to make a transformer robust to the choice of length configuration", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 336, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 507, + 351 + ], + "score": 1.0, + "content": "in the inference time. Rippel et al. (2014) proposes a nested dropout to learn ordered representations.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 346, + 507, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 507, + 362 + ], + "score": 1.0, + "content": "Similar to LengthDrop, it samples an index form a prior distribution and drops all units having a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 359, + 231, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 231, + 373 + ], + "score": 1.0, + "content": "larger index than sampled one.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "score": 1.0, + "content": "Search There have been a series of attempts at finding the optimal network configuration by solving", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 394, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 406 + ], + "score": 1.0, + "content": "a combinatorial optimization problem. In computer vision, Once-for-All (Cai et al., 2019) use an", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 405, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 506, + 416 + ], + "score": 1.0, + "content": "evolutionary search (Real et al., 2019) to find a better configuration in dimensions of depth, width,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 414, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 430 + ], + "score": 1.0, + "content": "kernel size, and resolution given computational budget. Similarly but differently, our evolutionary", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "search is mutli-objective to find length configurations on the Pareto accuracy-efficiency frontier to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "cope with any possible computational budgets. Moreover, we only change the sequence length of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 374, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 374, + 461 + ], + "score": 1.0, + "content": "hidden vectors instead of architectural model size like dimensions.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 108, + 476, + 303, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 304, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 304, + 491 + ], + "score": 1.0, + "content": "7 CONCLUSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 501, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "In this work, we propose a new framework for training a transformer once and using it for efficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "score": 1.0, + "content": "inference under any computational budget. With the help of training with LengthDrop and Drop-and-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Restore process followed by the evolutionary search, our proposed Length-Adaptive Transformer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "allows any given transformer models to be used with any inference-time computational budget for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "both sequence-level and token-level classification tasks. Our experiments, on SQuAD 1.1, MNLI-m", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "and SST-2, have revealed that the proposed algorithmic framework significantly pushes a better Pareto", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 566, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 580 + ], + "score": 1.0, + "content": "frontier on the trade-off between inference efficiency and accuracy. Furthermore, we have observed", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 382, + 591 + ], + "score": 1.0, + "content": "that the proposed Length-Adaptive Transformer could achieve up to", + "type": "text" + }, + { + "bbox": [ + 382, + 578, + 394, + 588 + ], + "score": 0.55, + "content": "3 \\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "speed-up over the standard", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 589, + 449, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 449, + 601 + ], + "score": 1.0, + "content": "transformer without sacrificing accuracy, both in terms of FLOPs and wallclock time.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 606, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "Although our approach finds an optimal length configuration of a trained classifier per computational", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "budget, it leaves open a question whether the proposed approach could be further extended to support", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "per-instance length configuration by for instance training a small, auxiliary neural network for each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "computational budget. Yet another aspect we have not investigated in this paper is the applicability of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "the proposed approach to sequence generation, such as machine translation. We leave both of these", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 660, + 242, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 242, + 673 + ], + "score": 1.0, + "content": "research directions for the future.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 505, + 721 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "score": 1.0, + "content": "Our approach is effective, as we have shown in this paper, and also quite simple to implement on top", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "of existing language models. We will release our implementation, which is based on HuggingFace’s", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 507, + 712 + ], + "score": 1.0, + "content": "Transformers library (Wolf et al., 2019), publicly and plan to adapt it for a broader set of transformer-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 711, + 255, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 255, + 721 + ], + "score": 1.0, + "content": "based models and downstream tasks.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 170 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Adaptive architecture There are three major axes along which computation can be reduced in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "a neural network; (1) input size/length, (2) network depth and (3) network width. The proposed", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 117 + ], + "score": 1.0, + "content": "approach, based on PoWER-BERT, adaptively reduces the input length as the input sequence is", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "processed by the transformer layers. In our knowledge, Goyal et al. (2020) is the first work in this", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "direction for transformers. More recently, Funnel-Transformer (Dai et al., 2020) and multi-scale", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "score": 1.0, + "content": "transformer language models (Subramanian et al., 2020) also successfully reduce sequence length in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "the middle and rescale to full length for the final computation. However, their inference complexity is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 455, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 455, + 173 + ], + "score": 1.0, + "content": "fixed unlike PoWER-BERT because they are not designed for the control of efficiency.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 82, + 506, + 173 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 176, + 505, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 190 + ], + "score": 1.0, + "content": "LayerDrop (Fan et al., 2019) drops random layers during the training to be robust to pruning", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 504, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 504, + 200 + ], + "score": 1.0, + "content": "inspired by Huang et al. (2016). Word-level adaptive depth in Elbayad et al. (2019) might seemingly", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "resemble with length reduction, but word vectors reached the maximal layer are used for self-attention", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 210, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 506, + 221 + ], + "score": 1.0, + "content": "computation without updating themselves. Escaping a network early (Teerapittayanon et al., 2016;", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "Huang et al., 2017) based on the confidence of the prediction (Xin et al., 2020; Schwartz et al., 2020;", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "Liu et al., 2020) also offers a control over accuracy-efficiency trade-off, but it is difficult to tune a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 242, + 495, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 495, + 254 + ], + "score": 1.0, + "content": "threshold for a desired computational budget because of the example-wise adaptive computation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 175, + 506, + 254 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 259, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 258, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 507, + 272 + ], + "score": 1.0, + "content": "DynaBERT (Hou et al., 2020) can run at adaptive width (the number of attention heads and interme-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "score": 1.0, + "content": "diate hidden dimension) and depth. Hardware-aware Transformers (Wang et al., 2020) construct a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "score": 1.0, + "content": "design space with arbitrary encoder-decoder attention and heterogeneous layers in terms of different", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 434, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 434, + 304 + ], + "score": 1.0, + "content": "numbers of layers, attention heads, hidden dimension, and embedding dimension.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 258, + 507, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 315, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 327 + ], + "score": 1.0, + "content": "Structured dropout A major innovation we introduce over the existing PoWER-BERT is the use of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "stochastic, structured regularization to make a transformer robust to the choice of length configuration", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 336, + 507, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 507, + 351 + ], + "score": 1.0, + "content": "in the inference time. Rippel et al. (2014) proposes a nested dropout to learn ordered representations.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 346, + 507, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 507, + 362 + ], + "score": 1.0, + "content": "Similar to LengthDrop, it samples an index form a prior distribution and drops all units having a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 359, + 231, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 231, + 373 + ], + "score": 1.0, + "content": "larger index than sampled one.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 315, + 507, + 373 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 382, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 506, + 397 + ], + "score": 1.0, + "content": "Search There have been a series of attempts at finding the optimal network configuration by solving", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 394, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 406 + ], + "score": 1.0, + "content": "a combinatorial optimization problem. In computer vision, Once-for-All (Cai et al., 2019) use an", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 405, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 506, + 416 + ], + "score": 1.0, + "content": "evolutionary search (Real et al., 2019) to find a better configuration in dimensions of depth, width,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 414, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 430 + ], + "score": 1.0, + "content": "kernel size, and resolution given computational budget. Similarly but differently, our evolutionary", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "search is mutli-objective to find length configurations on the Pareto accuracy-efficiency frontier to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "cope with any possible computational budgets. Moreover, we only change the sequence length of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 374, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 374, + 461 + ], + "score": 1.0, + "content": "hidden vectors instead of architectural model size like dimensions.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 381, + 506, + 461 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 476, + 303, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 474, + 304, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 304, + 491 + ], + "score": 1.0, + "content": "7 CONCLUSION AND FUTURE WORK", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 501, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 506, + 514 + ], + "score": 1.0, + "content": "In this work, we propose a new framework for training a transformer once and using it for efficient", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 506, + 525 + ], + "score": 1.0, + "content": "inference under any computational budget. With the help of training with LengthDrop and Drop-and-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Restore process followed by the evolutionary search, our proposed Length-Adaptive Transformer", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "allows any given transformer models to be used with any inference-time computational budget for", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "both sequence-level and token-level classification tasks. Our experiments, on SQuAD 1.1, MNLI-m", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "and SST-2, have revealed that the proposed algorithmic framework significantly pushes a better Pareto", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 566, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 580 + ], + "score": 1.0, + "content": "frontier on the trade-off between inference efficiency and accuracy. Furthermore, we have observed", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 382, + 591 + ], + "score": 1.0, + "content": "that the proposed Length-Adaptive Transformer could achieve up to", + "type": "text" + }, + { + "bbox": [ + 382, + 578, + 394, + 588 + ], + "score": 0.55, + "content": "3 \\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 577, + 506, + 591 + ], + "score": 1.0, + "content": "speed-up over the standard", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 589, + 449, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 449, + 601 + ], + "score": 1.0, + "content": "transformer without sacrificing accuracy, both in terms of FLOPs and wallclock time.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 500, + 506, + 601 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 606, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "Although our approach finds an optimal length configuration of a trained classifier per computational", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "budget, it leaves open a question whether the proposed approach could be further extended to support", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "per-instance length configuration by for instance training a small, auxiliary neural network for each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 506, + 651 + ], + "score": 1.0, + "content": "computational budget. Yet another aspect we have not investigated in this paper is the applicability of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "the proposed approach to sequence generation, such as machine translation. We leave both of these", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 660, + 242, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 242, + 673 + ], + "score": 1.0, + "content": "research directions for the future.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 605, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 677, + 505, + 721 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 691 + ], + "score": 1.0, + "content": "Our approach is effective, as we have shown in this paper, and also quite simple to implement on top", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "of existing language models. We will release our implementation, which is based on HuggingFace’s", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 507, + 712 + ], + "score": 1.0, + "content": "Transformers library (Wolf et al., 2019), publicly and plan to adapt it for a broader set of transformer-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 711, + 255, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 255, + 721 + ], + "score": 1.0, + "content": "based models and downstream tasks.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 676, + 507, + 721 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 505, + 134 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 506, + 113 + ], + "score": 1.0, + "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "score": 1.0, + "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 353, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 353, + 135 + ], + "score": 1.0, + "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "Han Cai, Chuang Gan, and Song Han. Once for all: Train one network and specialize it for efficient", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 154, + 331, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 331, + 166 + ], + "score": 1.0, + "content": "deployment. arXiv preprint arXiv:1908.09791, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 173, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "score": 1.0, + "content": "Ting-Wu Chin, Ruizhou Ding, Cha Zhang, and Diana Marculescu. Towards efficient model compres-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 185, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 116, + 185, + 505, + 197 + ], + "score": 1.0, + "content": "sion via learned global ranking. In Proceedings of the IEEE/CVF Conference on Computer Vision", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 196, + 308, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 308, + 209 + ], + "score": 1.0, + "content": "and Pattern Recognition, pp. 1518–1528, 2020.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "score": 1.0, + "content": "Zihang Dai, Guokun Lai, Yiming Yang, and Quoc V Le. Funnel-transformer: Filtering out sequential", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 227, + 464, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 227, + 464, + 240 + ], + "score": 1.0, + "content": "redundancy for efficient language processing. arXiv preprint arXiv:2006.03236, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 247, + 504, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 258, + 500, + 272 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 500, + 272 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 504, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "score": 1.0, + "content": "Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 290, + 254, + 302 + ], + "spans": [ + { + "bbox": [ + 115, + 290, + 254, + 302 + ], + "score": 1.0, + "content": "preprint arXiv:1910.10073, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 104, + 309, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 321, + 358, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 358, + 334 + ], + "score": 1.0, + "content": "structured dropout. arXiv preprint arXiv:1909.11556, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 108, + 341, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Saurabh Goyal, Anamitra Roy Choudhary, Venkatesan Chakaravarthy, Saurabh ManishRaje, Yogish", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 353, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 117, + 353, + 506, + 365 + ], + "score": 1.0, + "content": "Sabharwal, and Ashish Verma. Power-bert: Accelerating bert inference for classification tasks.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 363, + 279, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 279, + 376 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2001.08950, 2020.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 503, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 395, + 503, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 395, + 503, + 407 + ], + "score": 1.0, + "content": "with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 504, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and Joelle Pineau.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "Towards the systematic reporting of the energy and carbon footprints of machine learning. arXiv", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 438, + 254, + 449 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 254, + 449 + ], + "score": 1.0, + "content": "preprint arXiv:2002.05651, 2020.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 105, + 457, + 504, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 469, + 254, + 480 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 254, + 480 + ], + "score": 1.0, + "content": "preprint arXiv:1503.02531, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 105, + 488, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "Lu Hou, Lifeng Shang, Xin Jiang, and Qun Liu. Dynabert: Dynamic bert with adaptive width and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 500, + 306, + 512 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 306, + 512 + ], + "score": 1.0, + "content": "depth. arXiv preprint arXiv:2004.04037, 2020.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 105, + 519, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 530, + 484, + 544 + ], + "spans": [ + { + "bbox": [ + 115, + 530, + 484, + 544 + ], + "score": 1.0, + "content": "stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 551, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 563 + ], + "score": 1.0, + "content": "Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Q Wein-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 115, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "berger. Multi-scale dense networks for resource efficient image classification. arXiv preprint", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 573, + 220, + 584 + ], + "spans": [ + { + "bbox": [ + 116, + 573, + 220, + 584 + ], + "score": 1.0, + "content": "arXiv:1703.09844, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 105, + 593, + 506, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "score": 1.0, + "content": "Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 616, + 279, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 279, + 627 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2001.08361, 2020.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 506, + 649 + ], + "score": 1.0, + "content": "Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E Gon-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 647, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 116, + 647, + 506, + 659 + ], + "score": 1.0, + "content": "zalez. Train large, then compress: Rethinking model size for efficient training and inference of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 658, + 335, + 669 + ], + "spans": [ + { + "bbox": [ + 116, + 658, + 335, + 669 + ], + "score": 1.0, + "content": "transformers. arXiv preprint arXiv:2002.11794, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Haotang Deng, and Qi Ju. Fastbert: a self-distilling", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 689, + 416, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 689, + 416, + 701 + ], + "score": 1.0, + "content": "bert with adaptive inference time. arXiv preprint arXiv:2004.02178, 2020.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 389, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 389, + 732 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 14014–14024, 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 505, + 134 + ], + "lines": [ + { + "bbox": [ + 106, + 100, + 506, + 113 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 506, + 113 + ], + "score": 1.0, + "content": "Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 125 + ], + "score": 1.0, + "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 353, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 353, + 135 + ], + "score": 1.0, + "content": "few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 100, + 506, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 156 + ], + "score": 1.0, + "content": "Han Cai, Chuang Gan, and Song Han. Once for all: Train one network and specialize it for efficient", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 154, + 331, + 166 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 331, + 166 + ], + "score": 1.0, + "content": "deployment. arXiv preprint arXiv:1908.09791, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 143, + 505, + 166 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 173, + 504, + 208 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 506, + 187 + ], + "score": 1.0, + "content": "Ting-Wu Chin, Ruizhou Ding, Cha Zhang, and Diana Marculescu. Towards efficient model compres-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 185, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 116, + 185, + 505, + 197 + ], + "score": 1.0, + "content": "sion via learned global ranking. In Proceedings of the IEEE/CVF Conference on Computer Vision", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 196, + 308, + 209 + ], + "spans": [ + { + "bbox": [ + 115, + 196, + 308, + 209 + ], + "score": 1.0, + "content": "and Pattern Recognition, pp. 1518–1528, 2020.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 173, + 506, + 209 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "score": 1.0, + "content": "Zihang Dai, Guokun Lai, Yiming Yang, and Quoc V Le. Funnel-transformer: Filtering out sequential", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 227, + 464, + 240 + ], + "spans": [ + { + "bbox": [ + 115, + 227, + 464, + 240 + ], + "score": 1.0, + "content": "redundancy for efficient language processing. arXiv preprint arXiv:2006.03236, 2020.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 216, + 505, + 240 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 247, + 504, + 271 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 261 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 258, + 500, + 272 + ], + "spans": [ + { + "bbox": [ + 116, + 258, + 500, + 272 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 246, + 505, + 272 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 504, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "score": 1.0, + "content": "Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive transformer. arXiv", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 290, + 254, + 302 + ], + "spans": [ + { + "bbox": [ + 115, + 290, + 254, + 302 + ], + "score": 1.0, + "content": "preprint arXiv:1910.10073, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 279, + 505, + 302 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 309, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "Angela Fan, Edouard Grave, and Armand Joulin. Reducing transformer depth on demand with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 321, + 358, + 334 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 358, + 334 + ], + "score": 1.0, + "content": "structured dropout. arXiv preprint arXiv:1909.11556, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 310, + 505, + 334 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 341, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "Saurabh Goyal, Anamitra Roy Choudhary, Venkatesan Chakaravarthy, Saurabh ManishRaje, Yogish", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 353, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 117, + 353, + 506, + 365 + ], + "score": 1.0, + "content": "Sabharwal, and Ashish Verma. Power-bert: Accelerating bert inference for classification tasks.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 363, + 279, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 363, + 279, + 376 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2001.08950, 2020.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 342, + 506, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 503, + 407 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 395, + 503, + 407 + ], + "spans": [ + { + "bbox": [ + 115, + 395, + 503, + 407 + ], + "score": 1.0, + "content": "with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 384, + 505, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 504, + 449 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 428 + ], + "score": 1.0, + "content": "Peter Henderson, Jieru Hu, Joshua Romoff, Emma Brunskill, Dan Jurafsky, and Joelle Pineau.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 115, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "Towards the systematic reporting of the energy and carbon footprints of machine learning. arXiv", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 438, + 254, + 449 + ], + "spans": [ + { + "bbox": [ + 115, + 438, + 254, + 449 + ], + "score": 1.0, + "content": "preprint arXiv:2002.05651, 2020.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 414, + 505, + 449 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 457, + 504, + 480 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 469, + 254, + 480 + ], + "spans": [ + { + "bbox": [ + 115, + 469, + 254, + 480 + ], + "score": 1.0, + "content": "preprint arXiv:1503.02531, 2015.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 457, + 505, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 488, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "Lu Hou, Lifeng Shang, Xin Jiang, and Qun Liu. Dynabert: Dynamic bert with adaptive width and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 500, + 306, + 512 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 306, + 512 + ], + "score": 1.0, + "content": "depth. arXiv preprint arXiv:2004.04037, 2020.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 488, + 506, + 512 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 519, + 504, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 530, + 484, + 544 + ], + "spans": [ + { + "bbox": [ + 115, + 530, + 484, + 544 + ], + "score": 1.0, + "content": "stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 520, + 505, + 544 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 551, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 563 + ], + "score": 1.0, + "content": "Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Q Wein-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 115, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "berger. Multi-scale dense networks for resource efficient image classification. arXiv preprint", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 573, + 220, + 584 + ], + "spans": [ + { + "bbox": [ + 116, + 573, + 220, + 584 + ], + "score": 1.0, + "content": "arXiv:1703.09844, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 551, + 506, + 584 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 593, + 506, + 627 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "spans": [ + { + "bbox": [ + 116, + 605, + 507, + 617 + ], + "score": 1.0, + "content": "Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 616, + 279, + 627 + ], + "spans": [ + { + "bbox": [ + 115, + 616, + 279, + 627 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2001.08361, 2020.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 594, + 507, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 504, + 669 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 506, + 649 + ], + "score": 1.0, + "content": "Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E Gon-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 647, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 116, + 647, + 506, + 659 + ], + "score": 1.0, + "content": "zalez. Train large, then compress: Rethinking model size for efficient training and inference of", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 658, + 335, + 669 + ], + "spans": [ + { + "bbox": [ + 116, + 658, + 335, + 669 + ], + "score": 1.0, + "content": "transformers. arXiv preprint arXiv:2002.11794, 2020.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 636, + 506, + 669 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Haotang Deng, and Qi Ju. Fastbert: a self-distilling", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 689, + 416, + 701 + ], + "spans": [ + { + "bbox": [ + 115, + 689, + 416, + 701 + ], + "score": 1.0, + "content": "bert with adaptive inference time. arXiv preprint arXiv:2004.02178, 2020.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 677, + 505, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 389, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 389, + 732 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems, pp. 14014–14024, 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "Luke Zettlemoyer. Deep contextualized word representations. arXiv preprint arXiv:1802.05365,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 143, + 117 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 137 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 406, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 406, + 147 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 109, + 152, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 166 + ], + "score": 1.0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 174, + 331, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 331, + 186 + ], + "score": 1.0, + "content": "transformer. arXiv preprint arXiv:1910.10683, 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 105, + 192, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 412, + 205 + ], + "score": 1.0, + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad:", + "type": "text" + }, + { + "bbox": [ + 412, + 193, + 451, + 204 + ], + "score": 0.73, + "content": "^ { 1 0 0 , 0 0 0 + }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "questions for", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 410, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 410, + 216 + ], + "score": 1.0, + "content": "machine comprehension of text. arXiv preprint arXiv:1606.05250, 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 108, + 221, + 505, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 236 + ], + "score": 1.0, + "content": "Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "classifier architecture search. In Proceedings of the aaai conference on artificial intelligence,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 244, + 253, + 256 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 253, + 256 + ], + "score": 1.0, + "content": "volume 33, pp. 4780–4789, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 107, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 107, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 273, + 448, + 285 + ], + "spans": [ + { + "bbox": [ + 116, + 273, + 448, + 285 + ], + "score": 1.0, + "content": "dropout. In International Conference on Machine Learning, pp. 1746–1754, 2014.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 107, + 292, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 107, + 292, + 505, + 303 + ], + "score": 1.0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 302, + 443, + 315 + ], + "spans": [ + { + "bbox": [ + 115, + 302, + 443, + 315 + ], + "score": 1.0, + "content": "bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 105, + 320, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 334 + ], + "score": 1.0, + "content": "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. Green ai. arXiv preprint", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 331, + 219, + 344 + ], + "spans": [ + { + "bbox": [ + 115, + 331, + 219, + 344 + ], + "score": 1.0, + "content": "arXiv:1907.10597, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 505, + 362 + ], + "score": 1.0, + "content": "Roy Schwartz, Gabi Stanovsky, Swabha Swayamdipta, Jesse Dodge, and Noah A Smith. The right", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 360, + 507, + 374 + ], + "spans": [ + { + "bbox": [ + 114, + 360, + 507, + 374 + ], + "score": 1.0, + "content": "tool for the job: Matching model and instance complexities. arXiv preprint arXiv:2004.07453,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 370, + 142, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 142, + 385 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 105, + 390, + 504, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "Abigail See, Minh-Thang Luong, and Christopher D Manning. Compression of neural machine", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 401, + 406, + 414 + ], + "spans": [ + { + "bbox": [ + 115, + 401, + 406, + 414 + ], + "score": 1.0, + "content": "translation models via pruning. arXiv preprint arXiv:1606.09274, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 109, + 419, + 503, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "Catanzaro. Megatron-lm: Training multi-billion parameter language models using gpu model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 442, + 330, + 454 + ], + "spans": [ + { + "bbox": [ + 115, + 442, + 330, + 454 + ], + "score": 1.0, + "content": "parallelism. arXiv preprint arXiv:1909.08053, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 108, + 459, + 504, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 482, + 291, + 495 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 291, + 495 + ], + "score": 1.0, + "content": "learning research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 104, + 500, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 515 + ], + "score": 1.0, + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 511, + 343, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 343, + 524 + ], + "score": 1.0, + "content": "learning in nlp. arXiv preprint arXiv:1906.02243, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 107, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 107, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Sandeep Subramanian, Ronan Collobert, Marc’Aurelio Ranzato, and Y-Lan Boureau. Multi-scale", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 540, + 401, + 553 + ], + "spans": [ + { + "bbox": [ + 115, + 540, + 401, + 553 + ], + "score": 1.0, + "content": "transformer language models. arXiv preprint arXiv:2005.00581, 2020.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 504, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 504, + 571 + ], + "score": 1.0, + "content": "Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. Patient knowledge distillation for bert model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 570, + 335, + 582 + ], + "spans": [ + { + "bbox": [ + 115, + 570, + 335, + 582 + ], + "score": 1.0, + "content": "compression. arXiv preprint arXiv:1908.09355, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 105, + 588, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Raphael Tang, Ashutosh Adhikari, and Jimmy Lin. Flops as a direct optimization objective for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 599, + 411, + 613 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 411, + 613 + ], + "score": 1.0, + "content": "learning sparse neural networks. arXiv preprint arXiv:1811.03060, 2018.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 108, + 618, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 632 + ], + "score": 1.0, + "content": "Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 639, + 317, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 317, + 653 + ], + "score": 1.0, + "content": "Recognition (ICPR), pp. 2464–2469. IEEE, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 681, + 287, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 287, + 692 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue:", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 721, + 219, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 219, + 731 + ], + "score": 1.0, + "content": "arXiv:1804.07461, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "Luke Zettlemoyer. Deep contextualized word representations. arXiv preprint arXiv:1802.05365,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 143, + 117 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 82, + 506, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 122, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 506, + 137 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 406, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 406, + 147 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI Blog, 1(8):9, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 121, + 506, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 152, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 150, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 150, + 505, + 166 + ], + "score": 1.0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 174, + 331, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 331, + 186 + ], + "score": 1.0, + "content": "transformer. arXiv preprint arXiv:1910.10683, 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 150, + 505, + 186 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 192, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 412, + 205 + ], + "score": 1.0, + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad:", + "type": "text" + }, + { + "bbox": [ + 412, + 193, + 451, + 204 + ], + "score": 0.73, + "content": "^ { 1 0 0 , 0 0 0 + }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "questions for", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 410, + 216 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 410, + 216 + ], + "score": 1.0, + "content": "machine comprehension of text. arXiv preprint arXiv:1606.05250, 2016.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 192, + 505, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 221, + 505, + 255 + ], + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 236 + ], + "score": 1.0, + "content": "Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 232, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 506, + 246 + ], + "score": 1.0, + "content": "classifier architecture search. In Proceedings of the aaai conference on artificial intelligence,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 244, + 253, + 256 + ], + "spans": [ + { + "bbox": [ + 115, + 244, + 253, + 256 + ], + "score": 1.0, + "content": "volume 33, pp. 4780–4789, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 219, + 506, + 256 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 261, + 504, + 285 + ], + "lines": [ + { + "bbox": [ + 107, + 262, + 505, + 275 + ], + "spans": [ + { + "bbox": [ + 107, + 262, + 505, + 275 + ], + "score": 1.0, + "content": "Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 273, + 448, + 285 + ], + "spans": [ + { + "bbox": [ + 116, + 273, + 448, + 285 + ], + "score": 1.0, + "content": "dropout. In International Conference on Machine Learning, pp. 1746–1754, 2014.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 107, + 262, + 505, + 285 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 291, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 107, + 292, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 107, + 292, + 505, + 303 + ], + "score": 1.0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 302, + 443, + 315 + ], + "spans": [ + { + "bbox": [ + 115, + 302, + 443, + 315 + ], + "score": 1.0, + "content": "bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 107, + 292, + 505, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 320, + 504, + 343 + ], + "lines": [ + { + "bbox": [ + 105, + 318, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 334 + ], + "score": 1.0, + "content": "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. Green ai. arXiv preprint", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 331, + 219, + 344 + ], + "spans": [ + { + "bbox": [ + 115, + 331, + 219, + 344 + ], + "score": 1.0, + "content": "arXiv:1907.10597, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 318, + 505, + 344 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 349, + 504, + 383 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 505, + 362 + ], + "score": 1.0, + "content": "Roy Schwartz, Gabi Stanovsky, Swabha Swayamdipta, Jesse Dodge, and Noah A Smith. The right", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 114, + 360, + 507, + 374 + ], + "spans": [ + { + "bbox": [ + 114, + 360, + 507, + 374 + ], + "score": 1.0, + "content": "tool for the job: Matching model and instance complexities. arXiv preprint arXiv:2004.07453,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 370, + 142, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 370, + 142, + 385 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 350, + 507, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 390, + 504, + 413 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "Abigail See, Minh-Thang Luong, and Christopher D Manning. Compression of neural machine", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 401, + 406, + 414 + ], + "spans": [ + { + "bbox": [ + 115, + 401, + 406, + 414 + ], + "score": 1.0, + "content": "translation models via pruning. arXiv preprint arXiv:1606.09274, 2016.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 390, + 505, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 419, + 503, + 453 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 430, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 116, + 430, + 505, + 444 + ], + "score": 1.0, + "content": "Catanzaro. Megatron-lm: Training multi-billion parameter language models using gpu model", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 442, + 330, + 454 + ], + "spans": [ + { + "bbox": [ + 115, + 442, + 330, + 454 + ], + "score": 1.0, + "content": "parallelism. arXiv preprint arXiv:1909.08053, 2019.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 419, + 505, + 454 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 459, + 504, + 494 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 472, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 116, + 472, + 505, + 484 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 482, + 291, + 495 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 291, + 495 + ], + "score": 1.0, + "content": "learning research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 459, + 506, + 495 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 500, + 504, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 515 + ], + "score": 1.0, + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. Energy and policy considerations for deep", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 511, + 343, + 524 + ], + "spans": [ + { + "bbox": [ + 115, + 511, + 343, + 524 + ], + "score": 1.0, + "content": "learning in nlp. arXiv preprint arXiv:1906.02243, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 498, + 505, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 504, + 552 + ], + "lines": [ + { + "bbox": [ + 107, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 107, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "Sandeep Subramanian, Ronan Collobert, Marc’Aurelio Ranzato, and Y-Lan Boureau. Multi-scale", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 540, + 401, + 553 + ], + "spans": [ + { + "bbox": [ + 115, + 540, + 401, + 553 + ], + "score": 1.0, + "content": "transformer language models. arXiv preprint arXiv:2005.00581, 2020.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 107, + 530, + 505, + 553 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 558, + 504, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 504, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 504, + 571 + ], + "score": 1.0, + "content": "Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. Patient knowledge distillation for bert model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 570, + 335, + 582 + ], + "spans": [ + { + "bbox": [ + 115, + 570, + 335, + 582 + ], + "score": 1.0, + "content": "compression. arXiv preprint arXiv:1908.09355, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 558, + 504, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 588, + 504, + 611 + ], + "lines": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Raphael Tang, Ashutosh Adhikari, and Jimmy Lin. Flops as a direct optimization objective for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 599, + 411, + 613 + ], + "spans": [ + { + "bbox": [ + 116, + 599, + 411, + 613 + ], + "score": 1.0, + "content": "learning sparse neural networks. arXiv preprint arXiv:1811.03060, 2018.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 588, + 505, + 613 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 618, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 632 + ], + "score": 1.0, + "content": "Surat Teerapittayanon, Bradley McDanel, and Hsiang-Tsung Kung. Branchynet: Fast inference", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 639, + 317, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 317, + 653 + ], + "score": 1.0, + "content": "Recognition (ICPR), pp. 2464–2469. IEEE, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 617, + 506, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 504, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 681, + 287, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 287, + 692 + ], + "score": 1.0, + "content": "processing systems, pp. 5998–6008, 2017.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 658, + 505, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue:", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 721, + 219, + 731 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 219, + 731 + ], + "score": 1.0, + "content": "arXiv:1804.07461, 2018.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45, + "bbox_fs": [ + 106, + 699, + 506, + 731 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Hat: Hardware-aware transformers for efficient natural language processing. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "score": 1.0, + "content": "arXiv:2005.14187, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 504, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 396, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 396, + 158 + ], + "score": 1.0, + "content": "natural language processing. arXiv preprint arXiv:1910.03771, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. Deebert: Dynamic early exiting for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 390, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 390, + 187 + ], + "score": 1.0, + "content": "accelerating bert inference. arXiv preprint arXiv:2004.12993, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 506, + 218 + ], + "score": 1.0, + "content": "Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 336, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 336, + 228 + ], + "score": 1.0, + "content": "information processing systems, pp. 5754–5764, 2019.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 504, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 248 + ], + "score": 1.0, + "content": "Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE International Conference on Computer Vision, pp. 1803–1811, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan, and Song Han.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "Hat: Hardware-aware transformers for efficient natural language processing. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 219, + 115 + ], + "score": 1.0, + "content": "arXiv:2005.14187, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 506, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 504, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 505, + 146 + ], + "score": 1.0, + "content": "Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 396, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 396, + 158 + ], + "score": 1.0, + "content": "natural language processing. arXiv preprint arXiv:1910.03771, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 122, + 506, + 158 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 163, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. Deebert: Dynamic early exiting for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 390, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 390, + 187 + ], + "score": 1.0, + "content": "accelerating bert inference. arXiv preprint arXiv:2004.12993, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 164, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 506, + 207 + ], + "score": 1.0, + "content": "Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 506, + 218 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 506, + 218 + ], + "score": 1.0, + "content": "Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 336, + 228 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 336, + 228 + ], + "score": 1.0, + "content": "information processing systems, pp. 5754–5764, 2019.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 193, + 506, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 234, + 504, + 257 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 506, + 248 + ], + "score": 1.0, + "content": "Jiahui Yu and Thomas S Huang. Universally slimmable networks and improved training techniques.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 245, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 506, + 258 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE International Conference on Computer Vision, pp. 1803–1811, 2019.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 234, + 506, + 258 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_model.json b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c1b7f31b51e54a8734ba2a32236f80950875a702 --- /dev/null +++ b/parse/train/tqc8n6oHCtZ/tqc8n6oHCtZ_model.json @@ -0,0 +1,12984 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 397, + 596, + 1303, + 596, + 1303, + 1326, + 397, + 1326 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1712, + 1404, + 1712, + 1404, + 1957, + 298, + 1957 + ], + "score": 0.981 + }, + { + "category_id": 0, + "poly": [ + 299, + 219, + 1402, + 219, + 1402, + 325, + 299, + 325 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 298, + 1451, + 1406, + 1451, + 1406, + 1696, + 298, + 1696 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1398, + 1972, + 1398, + 2034, + 297, + 2034 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 313, + 384, + 680, + 384, + 680, + 444, + 313, + 444 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 302, + 1382, + 573, + 1382, + 573, + 1417, + 302, + 1417 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 773, + 525, + 927, + 525, + 927, + 560, + 773, + 560 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.76 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 856, + 2088, + 856, + 2112, + 841, + 2112 + ], + "score": 0.729 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.131 + }, + { + "category_id": 13, + "poly": [ + 773, + 1743, + 805, + 1743, + 805, + 1775, + 773, + 1775 + ], + "score": 0.62, + "latex": "\\ S 6" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 221.0, + 1403.0, + 221.0, + 1403.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 275.0, + 1349.0, + 275.0, + 1349.0, + 327.0, + 295.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1379.0, + 579.0, + 1379.0, + 579.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 522.0, + 934.0, + 522.0, + 934.0, + 565.0, + 768.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 857.0, + 2088.0, + 857.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 596.0, + 1304.0, + 596.0, + 1304.0, + 628.0, + 397.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 626.0, + 1304.0, + 626.0, + 1304.0, + 658.0, + 393.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 656.0, + 1306.0, + 656.0, + 1306.0, + 688.0, + 393.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 685.0, + 1304.0, + 685.0, + 1304.0, + 719.0, + 393.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 715.0, + 1304.0, + 715.0, + 1304.0, + 749.0, + 393.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 746.0, + 1306.0, + 746.0, + 1306.0, + 780.0, + 392.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 776.0, + 1307.0, + 776.0, + 1307.0, + 811.0, + 394.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 808.0, + 1305.0, + 808.0, + 1305.0, + 841.0, + 395.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 839.0, + 1305.0, + 839.0, + 1305.0, + 871.0, + 395.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 867.0, + 1305.0, + 867.0, + 1305.0, + 903.0, + 393.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 898.0, + 1304.0, + 898.0, + 1304.0, + 933.0, + 393.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 927.0, + 1306.0, + 927.0, + 1306.0, + 966.0, + 392.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 957.0, + 1307.0, + 957.0, + 1307.0, + 994.0, + 391.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 989.0, + 1307.0, + 989.0, + 1307.0, + 1023.0, + 393.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1018.0, + 1305.0, + 1018.0, + 1305.0, + 1055.0, + 392.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1050.0, + 1307.0, + 1050.0, + 1307.0, + 1085.0, + 393.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1083.0, + 1307.0, + 1083.0, + 1307.0, + 1115.0, + 394.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1112.0, + 1306.0, + 1112.0, + 1306.0, + 1145.0, + 395.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1143.0, + 1307.0, + 1143.0, + 1307.0, + 1178.0, + 394.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1172.0, + 1306.0, + 1172.0, + 1306.0, + 1208.0, + 392.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1202.0, + 1305.0, + 1202.0, + 1305.0, + 1240.0, + 392.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1231.0, + 1308.0, + 1231.0, + 1308.0, + 1272.0, + 393.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1265.0, + 1306.0, + 1265.0, + 1306.0, + 1297.0, + 396.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1294.0, + 1285.0, + 1294.0, + 1285.0, + 1329.0, + 395.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1710.0, + 1409.0, + 1710.0, + 1409.0, + 1751.0, + 292.0, + 1751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 772.0, + 1743.0, + 772.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1777.0, + 806.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1811.0, + 292.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1802.0, + 1407.0, + 1802.0, + 1407.0, + 1838.0, + 292.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1836.0, + 1408.0, + 1836.0, + 1408.0, + 1869.0, + 294.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1862.0, + 1409.0, + 1862.0, + 1409.0, + 1902.0, + 292.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1893.0, + 1406.0, + 1893.0, + 1406.0, + 1932.0, + 292.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1924.0, + 1222.0, + 1924.0, + 1222.0, + 1963.0, + 292.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1449.0, + 1408.0, + 1449.0, + 1408.0, + 1488.0, + 292.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1479.0, + 1408.0, + 1479.0, + 1408.0, + 1520.0, + 292.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1510.0, + 1407.0, + 1510.0, + 1407.0, + 1550.0, + 292.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1544.0, + 1407.0, + 1544.0, + 1407.0, + 1577.0, + 295.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1571.0, + 1409.0, + 1571.0, + 1409.0, + 1611.0, + 291.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1605.0, + 1406.0, + 1605.0, + 1406.0, + 1638.0, + 295.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1636.0, + 1408.0, + 1636.0, + 1408.0, + 1669.0, + 295.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1666.0, + 537.0, + 1666.0, + 537.0, + 1699.0, + 295.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1403.0, + 1970.0, + 1403.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2004.0, + 1402.0, + 2004.0, + 1402.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 385.0, + 560.0, + 385.0, + 560.0, + 417.0, + 315.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 413.0, + 681.0, + 413.0, + 681.0, + 448.0, + 312.0, + 448.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 965, + 1405, + 965, + 1405, + 1302, + 298, + 1302 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 476, + 1404, + 476, + 1404, + 659, + 298, + 659 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 676, + 1404, + 676, + 1404, + 951, + 298, + 951 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1772, + 1403, + 1772, + 1403, + 1957, + 299, + 1957 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1417, + 1405, + 1417, + 1405, + 1539, + 299, + 1539 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1634, + 1403, + 1634, + 1403, + 1758, + 299, + 1758 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 338, + 1402, + 338, + 1402, + 460, + 298, + 460 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1401, + 229, + 1401, + 323, + 298, + 323 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1402, + 1972, + 1402, + 2034, + 297, + 2034 + ], + "score": 0.941 + }, + { + "category_id": 0, + "poly": [ + 301, + 1576, + 703, + 1576, + 703, + 1608, + 301, + 1608 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 299, + 1347, + 1111, + 1347, + 1111, + 1383, + 299, + 1383 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.794 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.693 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.187 + }, + { + "category_id": 13, + "poly": [ + 579, + 1802, + 686, + 1802, + 686, + 1836, + 579, + 1836 + ], + "score": 0.93, + "latex": "h _ { t } ^ { 0 } \\in \\mathbb { R } ^ { H }" + }, + { + "category_id": 13, + "poly": [ + 748, + 1773, + 958, + 1773, + 958, + 1807, + 748, + 1807 + ], + "score": 0.91, + "latex": "{ \\boldsymbol x } = ( x _ { 1 } , \\dots , x _ { N } )" + }, + { + "category_id": 13, + "poly": [ + 376, + 1862, + 522, + 1862, + 522, + 1892, + 376, + 1892 + ], + "score": 0.91, + "latex": "\\dot { h ^ { 1 } } \\in \\mathbb { R } ^ { N \\times H }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1893, + 523, + 1893, + 523, + 1928, + 298, + 1928 + ], + "score": 0.89, + "latex": "h ^ { L } = ( h _ { 1 } ^ { L } , \\ldots , h _ { N } ^ { L } )" + }, + { + "category_id": 13, + "poly": [ + 433, + 1779, + 462, + 1779, + 462, + 1804, + 433, + 1804 + ], + "score": 0.84, + "latex": "x _ { t }" + }, + { + "category_id": 13, + "poly": [ + 827, + 1866, + 850, + 1866, + 850, + 1892, + 827, + 1892 + ], + "score": 0.81, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 1188, + 1866, + 1211, + 1866, + 1211, + 1892, + 1188, + 1892 + ], + "score": 0.75, + "latex": "L" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1576.0, + 706.0, + 1576.0, + 706.0, + 1612.0, + 295.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1345.0, + 1115.0, + 1345.0, + 1115.0, + 1388.0, + 292.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 966.0, + 1405.0, + 966.0, + 1405.0, + 1001.0, + 295.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 998.0, + 1408.0, + 998.0, + 1408.0, + 1033.0, + 295.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1028.0, + 1405.0, + 1028.0, + 1405.0, + 1063.0, + 295.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1059.0, + 1406.0, + 1059.0, + 1406.0, + 1094.0, + 293.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1086.0, + 1406.0, + 1086.0, + 1406.0, + 1125.0, + 292.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1120.0, + 1405.0, + 1120.0, + 1405.0, + 1155.0, + 295.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1150.0, + 1406.0, + 1150.0, + 1406.0, + 1184.0, + 295.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1214.0, + 293.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1209.0, + 1403.0, + 1209.0, + 1403.0, + 1246.0, + 293.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 1406.0, + 1240.0, + 1406.0, + 1279.0, + 293.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1274.0, + 887.0, + 1274.0, + 887.0, + 1305.0, + 296.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 474.0, + 1408.0, + 474.0, + 1408.0, + 512.0, + 292.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 506.0, + 1406.0, + 506.0, + 1406.0, + 542.0, + 293.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 537.0, + 1406.0, + 537.0, + 1406.0, + 573.0, + 294.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 569.0, + 1406.0, + 569.0, + 1406.0, + 604.0, + 294.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 597.0, + 1405.0, + 597.0, + 1405.0, + 637.0, + 291.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 629.0, + 395.0, + 629.0, + 395.0, + 664.0, + 294.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 675.0, + 1405.0, + 675.0, + 1405.0, + 712.0, + 293.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 708.0, + 1404.0, + 708.0, + 1404.0, + 741.0, + 294.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 734.0, + 1406.0, + 734.0, + 1406.0, + 777.0, + 292.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 806.0, + 293.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 795.0, + 1405.0, + 795.0, + 1405.0, + 834.0, + 292.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 829.0, + 1405.0, + 829.0, + 1405.0, + 862.0, + 296.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 858.0, + 1405.0, + 858.0, + 1405.0, + 894.0, + 294.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 890.0, + 1404.0, + 890.0, + 1404.0, + 923.0, + 296.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 919.0, + 1407.0, + 919.0, + 1407.0, + 955.0, + 293.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1770.0, + 432.0, + 1770.0, + 432.0, + 1809.0, + 293.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1770.0, + 747.0, + 1770.0, + 747.0, + 1809.0, + 463.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 1770.0, + 1405.0, + 1770.0, + 1405.0, + 1809.0, + 959.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1797.0, + 578.0, + 1797.0, + 578.0, + 1839.0, + 292.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1797.0, + 1410.0, + 1797.0, + 1410.0, + 1839.0, + 687.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1831.0, + 1405.0, + 1831.0, + 1405.0, + 1870.0, + 293.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1853.0, + 375.0, + 1853.0, + 375.0, + 1904.0, + 290.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1853.0, + 826.0, + 1853.0, + 826.0, + 1904.0, + 523.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1853.0, + 1187.0, + 1853.0, + 1187.0, + 1904.0, + 851.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1853.0, + 1409.0, + 1853.0, + 1409.0, + 1904.0, + 1212.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1887.0, + 297.0, + 1887.0, + 297.0, + 1936.0, + 292.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1887.0, + 1408.0, + 1887.0, + 1408.0, + 1936.0, + 524.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1924.0, + 1400.0, + 1924.0, + 1400.0, + 1960.0, + 295.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1418.0, + 1403.0, + 1418.0, + 1403.0, + 1450.0, + 295.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1446.0, + 1405.0, + 1446.0, + 1405.0, + 1485.0, + 293.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1476.0, + 1410.0, + 1476.0, + 1410.0, + 1514.0, + 293.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1509.0, + 1017.0, + 1509.0, + 1017.0, + 1542.0, + 295.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1636.0, + 1402.0, + 1636.0, + 1402.0, + 1666.0, + 295.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1666.0, + 1404.0, + 1666.0, + 1404.0, + 1699.0, + 293.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1695.0, + 1407.0, + 1695.0, + 1407.0, + 1731.0, + 293.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1728.0, + 560.0, + 1728.0, + 560.0, + 1761.0, + 295.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 338.0, + 1406.0, + 338.0, + 1406.0, + 374.0, + 294.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 368.0, + 1406.0, + 368.0, + 1406.0, + 404.0, + 293.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 399.0, + 1406.0, + 399.0, + 1406.0, + 434.0, + 293.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 430.0, + 1350.0, + 430.0, + 1350.0, + 463.0, + 294.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 295.0, + 294.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 292.0, + 936.0, + 292.0, + 936.0, + 329.0, + 293.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2004.0, + 296.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1408.0, + 2002.0, + 1408.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1649, + 1404, + 1649, + 1404, + 1926, + 298, + 1926 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1018, + 1403, + 1018, + 1403, + 1233, + 298, + 1233 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1249, + 1404, + 1249, + 1404, + 1463, + 298, + 1463 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 337, + 1404, + 337, + 1404, + 523, + 299, + 523 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 680, + 1404, + 680, + 1404, + 834, + 299, + 834 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 849, + 1403, + 849, + 1403, + 1003, + 299, + 1003 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1402, + 1942, + 1402, + 2033, + 300, + 2033 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 1402, + 229, + 1402, + 322, + 300, + 322 + ], + "score": 0.962 + }, + { + "category_id": 0, + "poly": [ + 299, + 1554, + 861, + 1554, + 861, + 1589, + 299, + 1589 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 299, + 605, + 550, + 605, + 550, + 636, + 299, + 636 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.732 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2112, + 842, + 2112 + ], + "score": 0.57 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.433 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 853, + 76, + 853, + 104, + 299, + 104 + ], + "score": 0.15 + }, + { + "category_id": 13, + "poly": [ + 1115, + 1081, + 1232, + 1081, + 1232, + 1114, + 1115, + 1114 + ], + "score": 0.91, + "latex": "\\left( l _ { 1 } , \\cdots l _ { L } \\right)" + }, + { + "category_id": 13, + "poly": [ + 982, + 398, + 1018, + 398, + 1018, + 432, + 982, + 432 + ], + "score": 0.91, + "latex": "h _ { t } ^ { L }" + }, + { + "category_id": 13, + "poly": [ + 1365, + 336, + 1400, + 336, + 1400, + 372, + 1365, + 372 + ], + "score": 0.89, + "latex": "h _ { 1 } ^ { L }" + }, + { + "category_id": 13, + "poly": [ + 728, + 683, + 751, + 683, + 751, + 715, + 728, + 715 + ], + "score": 0.87, + "latex": "l _ { j }" + }, + { + "category_id": 13, + "poly": [ + 597, + 744, + 620, + 744, + 620, + 776, + 597, + 776 + ], + "score": 0.86, + "latex": "l _ { j }" + }, + { + "category_id": 13, + "poly": [ + 1047, + 684, + 1064, + 684, + 1064, + 713, + 1047, + 713 + ], + "score": 0.83, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1382, + 745, + 1398, + 745, + 1398, + 775, + 1382, + 775 + ], + "score": 0.8, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 461, + 432, + 474, + 432, + 474, + 457, + 461, + 457 + ], + "score": 0.73, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1551.0, + 866.0, + 1551.0, + 866.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 603.0, + 555.0, + 603.0, + 555.0, + 642.0, + 294.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1649.0, + 1405.0, + 1649.0, + 1405.0, + 1687.0, + 293.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1683.0, + 1405.0, + 1683.0, + 1405.0, + 1716.0, + 296.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1748.0, + 293.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1738.0, + 1406.0, + 1738.0, + 1406.0, + 1781.0, + 291.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1770.0, + 1406.0, + 1770.0, + 1406.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1840.0, + 294.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1831.0, + 1406.0, + 1831.0, + 1406.0, + 1873.0, + 292.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1862.0, + 1406.0, + 1862.0, + 1406.0, + 1901.0, + 293.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1892.0, + 1294.0, + 1892.0, + 1294.0, + 1931.0, + 292.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1018.0, + 1404.0, + 1018.0, + 1404.0, + 1057.0, + 293.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1050.0, + 1404.0, + 1050.0, + 1404.0, + 1085.0, + 294.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1081.0, + 1114.0, + 1081.0, + 1114.0, + 1116.0, + 294.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1081.0, + 1404.0, + 1081.0, + 1404.0, + 1116.0, + 1233.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1110.0, + 1407.0, + 1110.0, + 1407.0, + 1148.0, + 293.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1141.0, + 1405.0, + 1141.0, + 1405.0, + 1176.0, + 294.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1172.0, + 1405.0, + 1172.0, + 1405.0, + 1205.0, + 293.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1203.0, + 1386.0, + 1203.0, + 1386.0, + 1236.0, + 292.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1248.0, + 1406.0, + 1248.0, + 1406.0, + 1286.0, + 293.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1279.0, + 1405.0, + 1279.0, + 1405.0, + 1314.0, + 293.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 1405.0, + 1309.0, + 1405.0, + 1342.0, + 293.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1343.0, + 1404.0, + 1343.0, + 1404.0, + 1373.0, + 296.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1371.0, + 1406.0, + 1371.0, + 1406.0, + 1405.0, + 294.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1401.0, + 1404.0, + 1401.0, + 1404.0, + 1436.0, + 293.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1431.0, + 1377.0, + 1431.0, + 1377.0, + 1468.0, + 293.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 334.0, + 1364.0, + 334.0, + 1364.0, + 374.0, + 293.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 334.0, + 1404.0, + 334.0, + 1404.0, + 374.0, + 1401.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1405.0, + 369.0, + 1405.0, + 401.0, + 294.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 398.0, + 981.0, + 398.0, + 981.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 398.0, + 1406.0, + 398.0, + 1406.0, + 434.0, + 1019.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 424.0, + 460.0, + 424.0, + 460.0, + 468.0, + 293.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 424.0, + 1409.0, + 424.0, + 1409.0, + 468.0, + 475.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 458.0, + 1405.0, + 458.0, + 1405.0, + 496.0, + 294.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 491.0, + 618.0, + 491.0, + 618.0, + 526.0, + 296.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 679.0, + 727.0, + 679.0, + 727.0, + 717.0, + 293.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 679.0, + 1046.0, + 679.0, + 1046.0, + 717.0, + 752.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 679.0, + 1406.0, + 679.0, + 1406.0, + 717.0, + 1065.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 713.0, + 1405.0, + 713.0, + 1405.0, + 746.0, + 297.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 741.0, + 596.0, + 741.0, + 596.0, + 780.0, + 293.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 741.0, + 1381.0, + 741.0, + 1381.0, + 780.0, + 621.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 741.0, + 1409.0, + 741.0, + 1409.0, + 780.0, + 1399.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 773.0, + 1406.0, + 773.0, + 1406.0, + 810.0, + 294.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 805.0, + 1175.0, + 805.0, + 1175.0, + 838.0, + 295.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 849.0, + 1404.0, + 849.0, + 1404.0, + 885.0, + 294.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 881.0, + 1407.0, + 881.0, + 1407.0, + 914.0, + 295.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 913.0, + 1405.0, + 913.0, + 1405.0, + 946.0, + 297.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 941.0, + 1405.0, + 941.0, + 1405.0, + 976.0, + 293.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 974.0, + 1281.0, + 974.0, + 1281.0, + 1007.0, + 295.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1407.0, + 1939.0, + 1407.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1059.0, + 2003.0, + 1059.0, + 2035.0, + 295.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1407.0, + 227.0, + 1407.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 295.0, + 295.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 291.0, + 915.0, + 291.0, + 915.0, + 325.0, + 295.0, + 325.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 702, + 1404, + 702, + 1404, + 978, + 297, + 978 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1637, + 1406, + 1637, + 1406, + 2034, + 297, + 2034 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1376, + 1404, + 1376, + 1404, + 1622, + 298, + 1622 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1008, + 1405, + 1008, + 1405, + 1285, + 298, + 1285 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 456, + 1405, + 456, + 1405, + 670, + 297, + 670 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 286, + 1403, + 286, + 1403, + 441, + 298, + 441 + ], + "score": 0.977 + }, + { + "category_id": 0, + "poly": [ + 299, + 229, + 535, + 229, + 535, + 261, + 299, + 261 + ], + "score": 0.885 + }, + { + "category_id": 0, + "poly": [ + 300, + 1319, + 1030, + 1319, + 1030, + 1351, + 300, + 1351 + ], + "score": 0.841 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 854, + 75, + 854, + 104, + 299, + 104 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 857, + 2089, + 857, + 2111, + 840, + 2111 + ], + "score": 0.765 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 853, + 76, + 853, + 104, + 300, + 104 + ], + "score": 0.164 + }, + { + "category_id": 13, + "poly": [ + 769, + 1192, + 893, + 1192, + 893, + 1226, + 769, + 1226 + ], + "score": 0.92, + "latex": "\\lceil ( 1 - p ) l _ { i } \\rceil" + }, + { + "category_id": 13, + "poly": [ + 437, + 517, + 640, + 517, + 640, + 551, + 437, + 551 + ], + "score": 0.92, + "latex": "\\mathcal { U } ( \\lceil ( 1 - p ) l _ { i } \\rceil , l _ { i } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1667, + 520, + 1667, + 520, + 1702, + 297, + 1702 + ], + "score": 0.92, + "latex": "l _ { i + 1 } = \\lceil ( 1 - r ) \\bar { l _ { i } } \\rceil" + }, + { + "category_id": 13, + "poly": [ + 464, + 1789, + 597, + 1789, + 597, + 1824, + 464, + 1824 + ], + "score": 0.91, + "latex": "( l _ { 1 } ^ { \\prime } , \\cdots , l _ { L } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 487, + 345, + 487, + 345, + 519, + 298, + 519 + ], + "score": 0.91, + "latex": "l _ { i + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1263, + 548, + 1398, + 548, + 1398, + 581, + 1263, + 581 + ], + "score": 0.91, + "latex": "( l _ { 1 } , \\cdots , l _ { L } )" + }, + { + "category_id": 13, + "poly": [ + 1095, + 1789, + 1242, + 1789, + 1242, + 1824, + 1095, + 1824 + ], + "score": 0.91, + "latex": "\\bar { \\mathcal { U } } ( l _ { i - 1 } ^ { \\prime } , l _ { i + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 420, + 487, + 499, + 487, + 499, + 519, + 420, + 519 + ], + "score": 0.91, + "latex": "( i + 1 )" + }, + { + "category_id": 13, + "poly": [ + 295, + 1789, + 430, + 1789, + 430, + 1823, + 295, + 1823 + ], + "score": 0.91, + "latex": "( l _ { 1 } , \\cdots , l _ { L } )" + }, + { + "category_id": 13, + "poly": [ + 745, + 1821, + 819, + 1821, + 819, + 1854, + 745, + 1854 + ], + "score": 0.9, + "latex": "{ \\bar { l } } _ { i } ^ { \\prime } = l _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1822, + 1276, + 1822, + 1276, + 1848, + 1208, + 1848 + ], + "score": 0.9, + "latex": "i = L" + }, + { + "category_id": 13, + "poly": [ + 1113, + 1824, + 1176, + 1824, + 1176, + 1848, + 1113, + 1848 + ], + "score": 0.88, + "latex": "i = 1" + }, + { + "category_id": 13, + "poly": [ + 1118, + 488, + 1138, + 488, + 1138, + 517, + 1118, + 517 + ], + "score": 0.86, + "latex": "l _ { i }" + }, + { + "category_id": 13, + "poly": [ + 734, + 519, + 759, + 519, + 759, + 548, + 734, + 548 + ], + "score": 0.85, + "latex": "l _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1826, + 411, + 1826, + 411, + 1853, + 374, + 1853 + ], + "score": 0.85, + "latex": "p _ { m }" + }, + { + "category_id": 13, + "poly": [ + 744, + 1790, + 764, + 1790, + 764, + 1822, + 744, + 1822 + ], + "score": 0.84, + "latex": "l _ { i } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1167, + 329, + 1167, + 329, + 1193, + 298, + 1193 + ], + "score": 0.83, + "latex": "n _ { s }" + }, + { + "category_id": 13, + "poly": [ + 582, + 1916, + 622, + 1916, + 622, + 1943, + 582, + 1943 + ], + "score": 0.83, + "latex": "n _ { m }" + }, + { + "category_id": 13, + "poly": [ + 1376, + 1883, + 1401, + 1883, + 1401, + 1909, + 1376, + 1909 + ], + "score": 0.79, + "latex": "G" + }, + { + "category_id": 13, + "poly": [ + 939, + 1916, + 969, + 1916, + 969, + 1943, + 939, + 1943 + ], + "score": 0.79, + "latex": "n _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1310, + 522, + 1328, + 522, + 1328, + 549, + 1310, + 549 + ], + "score": 0.77, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 696, + 1671, + 710, + 1671, + 710, + 1696, + 696, + 1696 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 775, + 1675, + 792, + 1675, + 792, + 1696, + 775, + 1696 + ], + "score": 0.7, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 1190, + 1195, + 1203, + 1195, + 1203, + 1221, + 1190, + 1221 + ], + "score": 0.69, + "latex": "i" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 539.0, + 228.0, + 539.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1319.0, + 1033.0, + 1319.0, + 1033.0, + 1355.0, + 295.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 702.0, + 1405.0, + 702.0, + 1405.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 734.0, + 1406.0, + 734.0, + 1406.0, + 767.0, + 295.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 762.0, + 1407.0, + 762.0, + 1407.0, + 797.0, + 294.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 792.0, + 1405.0, + 792.0, + 1405.0, + 830.0, + 292.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 824.0, + 1405.0, + 824.0, + 1405.0, + 861.0, + 294.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 854.0, + 1406.0, + 854.0, + 1406.0, + 891.0, + 294.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 886.0, + 1406.0, + 886.0, + 1406.0, + 922.0, + 294.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 913.0, + 1404.0, + 913.0, + 1404.0, + 952.0, + 294.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 947.0, + 585.0, + 947.0, + 585.0, + 984.0, + 294.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1635.0, + 1404.0, + 1635.0, + 1404.0, + 1674.0, + 294.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1666.0, + 296.0, + 1666.0, + 296.0, + 1703.0, + 291.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 1666.0, + 695.0, + 1666.0, + 695.0, + 1703.0, + 521.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1666.0, + 774.0, + 1666.0, + 774.0, + 1703.0, + 711.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1666.0, + 1407.0, + 1666.0, + 1407.0, + 1703.0, + 793.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1697.0, + 1409.0, + 1697.0, + 1409.0, + 1735.0, + 291.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1724.0, + 1409.0, + 1724.0, + 1409.0, + 1767.0, + 291.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1759.0, + 1406.0, + 1759.0, + 1406.0, + 1795.0, + 294.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1788.0, + 463.0, + 1788.0, + 463.0, + 1828.0, + 431.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1788.0, + 743.0, + 1788.0, + 743.0, + 1828.0, + 598.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1788.0, + 1094.0, + 1788.0, + 1094.0, + 1828.0, + 765.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1243.0, + 1788.0, + 1409.0, + 1788.0, + 1409.0, + 1828.0, + 1243.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1817.0, + 373.0, + 1817.0, + 373.0, + 1858.0, + 292.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1817.0, + 744.0, + 1817.0, + 744.0, + 1858.0, + 412.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1817.0, + 1112.0, + 1817.0, + 1112.0, + 1858.0, + 820.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1817.0, + 1207.0, + 1817.0, + 1207.0, + 1858.0, + 1177.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1817.0, + 1407.0, + 1817.0, + 1407.0, + 1858.0, + 1277.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1404.0, + 1850.0, + 1404.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1375.0, + 1882.0, + 1375.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1882.0, + 1406.0, + 1882.0, + 1406.0, + 1916.0, + 1402.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 581.0, + 1909.0, + 581.0, + 1949.0, + 294.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 1909.0, + 938.0, + 1909.0, + 938.0, + 1949.0, + 623.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1909.0, + 1406.0, + 1909.0, + 1406.0, + 1949.0, + 970.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1408.0, + 1942.0, + 1408.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1970.0, + 1406.0, + 1970.0, + 1406.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2005.0, + 668.0, + 2005.0, + 668.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1376.0, + 1405.0, + 1376.0, + 1405.0, + 1414.0, + 293.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1408.0, + 1405.0, + 1408.0, + 1405.0, + 1440.0, + 293.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 1405.0, + 1438.0, + 1405.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1505.0, + 293.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1498.0, + 1405.0, + 1498.0, + 1405.0, + 1536.0, + 292.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1528.0, + 1409.0, + 1528.0, + 1409.0, + 1567.0, + 291.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1559.0, + 1405.0, + 1559.0, + 1405.0, + 1598.0, + 292.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1589.0, + 483.0, + 1589.0, + 483.0, + 1627.0, + 294.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1006.0, + 1405.0, + 1006.0, + 1405.0, + 1044.0, + 293.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1080.0, + 292.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1069.0, + 1408.0, + 1069.0, + 1408.0, + 1107.0, + 292.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1098.0, + 1406.0, + 1098.0, + 1406.0, + 1136.0, + 292.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1131.0, + 1405.0, + 1131.0, + 1405.0, + 1166.0, + 293.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1160.0, + 297.0, + 1160.0, + 297.0, + 1198.0, + 291.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1160.0, + 1408.0, + 1160.0, + 1408.0, + 1198.0, + 330.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1192.0, + 768.0, + 1192.0, + 768.0, + 1229.0, + 293.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1192.0, + 1189.0, + 1192.0, + 1189.0, + 1229.0, + 894.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1192.0, + 1405.0, + 1192.0, + 1405.0, + 1229.0, + 1204.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1220.0, + 1405.0, + 1220.0, + 1405.0, + 1258.0, + 293.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1253.0, + 1407.0, + 1253.0, + 1407.0, + 1289.0, + 295.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 455.0, + 1403.0, + 455.0, + 1403.0, + 493.0, + 294.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 486.0, + 419.0, + 486.0, + 419.0, + 524.0, + 346.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 486.0, + 1117.0, + 486.0, + 1117.0, + 524.0, + 500.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 486.0, + 1406.0, + 486.0, + 1406.0, + 524.0, + 1139.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 516.0, + 436.0, + 516.0, + 436.0, + 554.0, + 294.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 516.0, + 733.0, + 516.0, + 733.0, + 554.0, + 641.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 516.0, + 1309.0, + 516.0, + 1309.0, + 554.0, + 760.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 516.0, + 1406.0, + 516.0, + 1406.0, + 554.0, + 1329.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 547.0, + 1262.0, + 547.0, + 1262.0, + 585.0, + 294.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 547.0, + 1408.0, + 547.0, + 1408.0, + 585.0, + 1399.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 577.0, + 1406.0, + 577.0, + 1406.0, + 613.0, + 292.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 608.0, + 1406.0, + 608.0, + 1406.0, + 646.0, + 294.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 639.0, + 572.0, + 639.0, + 572.0, + 673.0, + 295.0, + 673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 288.0, + 1405.0, + 288.0, + 1405.0, + 321.0, + 296.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 317.0, + 1405.0, + 317.0, + 1405.0, + 353.0, + 293.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 349.0, + 1407.0, + 349.0, + 1407.0, + 383.0, + 293.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 378.0, + 1405.0, + 378.0, + 1405.0, + 415.0, + 294.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 409.0, + 766.0, + 409.0, + 766.0, + 447.0, + 295.0, + 447.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1552, + 1405, + 1552, + 1405, + 1797, + 298, + 1797 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 751, + 1404, + 751, + 1404, + 966, + 298, + 966 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1332, + 1403, + 1332, + 1403, + 1518, + 298, + 1518 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1083, + 1405, + 1083, + 1405, + 1297, + 298, + 1297 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1834, + 1404, + 1834, + 1404, + 1926, + 299, + 1926 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1404, + 1942, + 1404, + 2034, + 300, + 2034 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 331, + 247, + 1367, + 247, + 1367, + 438, + 331, + 438 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 296, + 461, + 1405, + 461, + 1405, + 646, + 296, + 646 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 301, + 1012, + 643, + 1012, + 643, + 1047, + 301, + 1047 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 302, + 694, + 733, + 694, + 733, + 724, + 302, + 724 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 854, + 76, + 854, + 104, + 298, + 104 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.711 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 852, + 76, + 852, + 104, + 299, + 104 + ], + "score": 0.113 + }, + { + "category_id": 1, + "poly": [ + 296, + 461, + 1405, + 461, + 1405, + 646, + 296, + 646 + ], + "score": 0.112 + }, + { + "category_id": 13, + "poly": [ + 681, + 1866, + 764, + 1866, + 764, + 1895, + 681, + 1895 + ], + "score": 0.9, + "latex": "n _ { s } = 2" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 288.0, + 727.0, + 288.0, + 727.0, + 386.0, + 596.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 273.0, + 1275.0, + 273.0, + 1275.0, + 358.0, + 935.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 325.0, + 693.0, + 325.0, + 693.0, + 337.0, + 666.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 321.0, + 1140.0, + 321.0, + 1140.0, + 378.0, + 969.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 324.0, + 1256.0, + 324.0, + 1256.0, + 370.0, + 1141.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 409.0, + 640.0, + 409.0, + 640.0, + 435.0, + 519.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 407.0, + 1235.0, + 407.0, + 1235.0, + 440.0, + 1006.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 313.5, + 1066.0, + 313.5, + 1066.0, + 331.5, + 1030.0, + 331.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 462.0, + 1407.0, + 462.0, + 1407.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 492.0, + 1406.0, + 492.0, + 1406.0, + 528.0, + 295.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 526.0, + 1405.0, + 526.0, + 1405.0, + 557.0, + 295.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 553.0, + 1405.0, + 553.0, + 1405.0, + 589.0, + 296.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 584.0, + 1406.0, + 584.0, + 1406.0, + 620.0, + 295.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 612.0, + 1136.0, + 612.0, + 1136.0, + 650.0, + 293.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1011.0, + 646.0, + 1011.0, + 646.0, + 1051.0, + 293.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 690.0, + 738.0, + 690.0, + 738.0, + 730.0, + 293.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1554.0, + 1405.0, + 1554.0, + 1405.0, + 1587.0, + 295.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1618.0, + 292.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1612.0, + 1407.0, + 1612.0, + 1407.0, + 1651.0, + 293.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1643.0, + 1406.0, + 1643.0, + 1406.0, + 1680.0, + 293.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1672.0, + 1405.0, + 1672.0, + 1405.0, + 1713.0, + 292.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1704.0, + 1408.0, + 1704.0, + 1408.0, + 1742.0, + 293.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1736.0, + 1403.0, + 1736.0, + 1403.0, + 1772.0, + 293.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 795.0, + 1768.0, + 795.0, + 1799.0, + 294.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 753.0, + 1406.0, + 753.0, + 1406.0, + 787.0, + 296.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 781.0, + 1406.0, + 781.0, + 1406.0, + 817.0, + 293.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 815.0, + 1401.0, + 815.0, + 1401.0, + 846.0, + 296.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 844.0, + 1405.0, + 844.0, + 1405.0, + 878.0, + 294.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 872.0, + 1404.0, + 872.0, + 1404.0, + 906.0, + 294.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 906.0, + 1404.0, + 906.0, + 1404.0, + 937.0, + 296.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 936.0, + 1400.0, + 936.0, + 1400.0, + 970.0, + 294.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1330.0, + 1405.0, + 1330.0, + 1405.0, + 1369.0, + 293.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1360.0, + 1407.0, + 1360.0, + 1407.0, + 1403.0, + 292.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1388.0, + 1407.0, + 1388.0, + 1407.0, + 1434.0, + 292.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1424.0, + 1407.0, + 1424.0, + 1407.0, + 1461.0, + 292.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1453.0, + 1407.0, + 1453.0, + 1407.0, + 1490.0, + 293.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1488.0, + 1236.0, + 1488.0, + 1236.0, + 1521.0, + 293.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1083.0, + 1407.0, + 1083.0, + 1407.0, + 1118.0, + 295.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1109.0, + 1409.0, + 1109.0, + 1409.0, + 1152.0, + 292.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1143.0, + 1405.0, + 1143.0, + 1405.0, + 1178.0, + 295.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1173.0, + 1406.0, + 1173.0, + 1406.0, + 1210.0, + 293.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1206.0, + 1405.0, + 1206.0, + 1405.0, + 1239.0, + 293.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1234.0, + 1407.0, + 1234.0, + 1407.0, + 1272.0, + 293.0, + 1272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1265.0, + 456.0, + 1265.0, + 456.0, + 1299.0, + 295.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1831.0, + 1404.0, + 1831.0, + 1404.0, + 1872.0, + 293.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1864.0, + 680.0, + 1864.0, + 680.0, + 1898.0, + 294.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 1864.0, + 1404.0, + 1864.0, + 1404.0, + 1898.0, + 765.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1893.0, + 1129.0, + 1893.0, + 1129.0, + 1929.0, + 294.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 1408.0, + 1942.0, + 1408.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1408.0, + 1971.0, + 1408.0, + 2006.0, + 295.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1404.0, + 2000.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 462.0, + 1407.0, + 462.0, + 1407.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 492.0, + 1406.0, + 492.0, + 1406.0, + 528.0, + 295.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 526.0, + 1405.0, + 526.0, + 1405.0, + 557.0, + 295.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 553.0, + 1405.0, + 553.0, + 1405.0, + 589.0, + 296.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 584.0, + 1406.0, + 584.0, + 1406.0, + 620.0, + 295.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 612.0, + 1136.0, + 612.0, + 1136.0, + 650.0, + 293.0, + 650.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1752, + 1405, + 1752, + 1405, + 1936, + 298, + 1936 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1383, + 1404, + 1383, + 1404, + 1568, + 298, + 1568 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1583, + 1403, + 1583, + 1403, + 1737, + 298, + 1737 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1173, + 1405, + 1173, + 1405, + 1267, + 299, + 1267 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 301, + 794, + 1100, + 794, + 1100, + 1010, + 301, + 1010 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 297, + 1044, + 1400, + 1044, + 1400, + 1137, + 297, + 1137 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 1135, + 849, + 1385, + 849, + 1385, + 973, + 1135, + 973 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1399, + 1971, + 1399, + 2035, + 298, + 2035 + ], + "score": 0.945 + }, + { + "category_id": 3, + "poly": [ + 310, + 220, + 1088, + 220, + 1088, + 771, + 310, + 771 + ], + "score": 0.935 + }, + { + "category_id": 3, + "poly": [ + 1135, + 227, + 1383, + 227, + 1383, + 825, + 1135, + 825 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 302, + 1313, + 705, + 1313, + 705, + 1348, + 302, + 1348 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 104, + 298, + 104 + ], + "score": 0.851 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.78 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 852, + 76, + 852, + 104, + 299, + 104 + ], + "score": 0.214 + }, + { + "category_id": 13, + "poly": [ + 806, + 1205, + 903, + 1205, + 903, + 1235, + 806, + 1235 + ], + "score": 0.91, + "latex": "n _ { c } = 3 0" + }, + { + "category_id": 13, + "poly": [ + 645, + 1205, + 756, + 1205, + 756, + 1237, + 645, + 1237 + ], + "score": 0.9, + "latex": "p _ { m } = 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 556, + 1174, + 646, + 1174, + 646, + 1202, + 556, + 1202 + ], + "score": 0.9, + "latex": "G = 3 0" + }, + { + "category_id": 13, + "poly": [ + 1083, + 1175, + 1190, + 1175, + 1190, + 1205, + 1083, + 1205 + ], + "score": 0.9, + "latex": "n _ { m } = 3 0" + }, + { + "category_id": 13, + "poly": [ + 703, + 1076, + 781, + 1076, + 781, + 1105, + 703, + 1105 + ], + "score": 0.85, + "latex": "5 e - 5" + }, + { + "category_id": 13, + "poly": [ + 1021, + 1076, + 1100, + 1076, + 1100, + 1105, + 1021, + 1105 + ], + "score": 0.84, + "latex": "2 e - 5" + }, + { + "category_id": 13, + "poly": [ + 1195, + 1783, + 1395, + 1783, + 1395, + 1814, + 1195, + 1814 + ], + "score": 0.8, + "latex": "( \\log { \\mathrm { F L O P s } } > 1 0 )" + }, + { + "category_id": 13, + "poly": [ + 810, + 827, + 920, + 827, + 920, + 858, + 810, + 858 + ], + "score": 0.77, + "latex": "\\mathbf { B E R T _ { B a s e } }" + }, + { + "category_id": 13, + "poly": [ + 768, + 256, + 791, + 256, + 791, + 278, + 768, + 278 + ], + "score": 0.47, + "latex": "\\cdot" + }, + { + "category_id": 13, + "poly": [ + 464, + 229, + 488, + 229, + 488, + 250, + 464, + 250 + ], + "score": 0.39, + "latex": "0" + }, + { + "category_id": 13, + "poly": [ + 769, + 227, + 792, + 227, + 792, + 249, + 769, + 249 + ], + "score": 0.26, + "latex": "\\bigcirc" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 793.0, + 1103.0, + 793.0, + 1103.0, + 830.0, + 299.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 825.0, + 809.0, + 825.0, + 809.0, + 861.0, + 298.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 825.0, + 1102.0, + 825.0, + 1102.0, + 861.0, + 921.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 852.0, + 1104.0, + 852.0, + 1104.0, + 893.0, + 297.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 886.0, + 1103.0, + 886.0, + 1103.0, + 923.0, + 298.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 918.0, + 1103.0, + 918.0, + 1103.0, + 951.0, + 298.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 948.0, + 1102.0, + 948.0, + 1102.0, + 982.0, + 299.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 980.0, + 631.0, + 980.0, + 631.0, + 1012.0, + 301.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 848.0, + 1386.0, + 848.0, + 1386.0, + 881.0, + 1133.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 880.0, + 1385.0, + 880.0, + 1385.0, + 912.0, + 1134.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 911.0, + 1385.0, + 911.0, + 1385.0, + 943.0, + 1134.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 941.0, + 1384.0, + 941.0, + 1384.0, + 976.0, + 1134.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 231.0, + 329.0, + 231.0, + 329.0, + 243.0, + 317.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 223.0, + 453.0, + 223.0, + 453.0, + 254.0, + 333.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 224.0, + 758.0, + 224.0, + 758.0, + 253.0, + 489.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 224.0, + 1092.0, + 224.0, + 1092.0, + 252.0, + 793.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 254.0, + 452.0, + 254.0, + 452.0, + 281.0, + 323.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 253.0, + 758.0, + 253.0, + 758.0, + 282.0, + 473.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 253.0, + 767.0, + 253.0, + 767.0, + 281.0, + 764.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 253.0, + 1084.0, + 253.0, + 1084.0, + 281.0, + 792.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 293.0, + 413.0, + 293.0, + 413.0, + 326.0, + 352.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 364.0, + 412.0, + 364.0, + 412.0, + 397.0, + 352.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 435.0, + 409.0, + 435.0, + 409.0, + 467.0, + 352.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 485.0, + 342.0, + 485.0, + 342.0, + 521.0, + 310.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 509.0, + 392.0, + 509.0, + 392.0, + 539.0, + 353.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 578.0, + 410.0, + 578.0, + 410.0, + 610.0, + 352.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 650.0, + 406.0, + 650.0, + 406.0, + 683.0, + 352.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 724.0, + 456.0, + 724.0, + 456.0, + 747.0, + 436.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 723.0, + 506.0, + 723.0, + 506.0, + 748.0, + 477.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 720.0, + 822.0, + 720.0, + 822.0, + 749.0, + 789.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 721.0, + 1005.0, + 721.0, + 1005.0, + 750.0, + 972.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 737.0, + 809.0, + 737.0, + 809.0, + 779.0, + 664.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 575.5, + 516.0, + 575.5, + 516.0, + 598.5, + 511.0, + 598.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 620.0, + 516.0, + 620.0, + 516.0, + 628.0, + 460.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 651.5, + 465.0, + 651.5, + 465.0, + 672.0, + 461.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 230.0, + 1256.0, + 230.0, + 1256.0, + 258.0, + 1196.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 230.0, + 1343.0, + 230.0, + 1343.0, + 257.0, + 1283.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 259.0, + 1183.0, + 259.0, + 1183.0, + 271.0, + 1171.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 254.0, + 1253.0, + 254.0, + 1253.0, + 274.0, + 1189.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 259.0, + 1281.0, + 259.0, + 1281.0, + 271.0, + 1267.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 251.0, + 1350.0, + 251.0, + 1350.0, + 276.0, + 1283.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 281.0, + 1182.0, + 281.0, + 1182.0, + 300.0, + 1155.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 287.0, + 1192.0, + 287.0, + 1192.0, + 294.0, + 1182.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 294.0, + 1159.0, + 294.0, + 1159.0, + 433.0, + 1137.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 315.0, + 1179.0, + 315.0, + 1179.0, + 335.0, + 1156.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 350.0, + 1178.0, + 350.0, + 1178.0, + 368.0, + 1156.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 385.0, + 1179.0, + 385.0, + 1179.0, + 403.0, + 1163.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 419.0, + 1179.0, + 419.0, + 1179.0, + 435.0, + 1163.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 435.0, + 1379.0, + 435.0, + 1379.0, + 463.0, + 1187.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 455.0, + 1321.0, + 455.0, + 1321.0, + 484.0, + 1244.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 484.0, + 1306.0, + 484.0, + 1306.0, + 517.0, + 1212.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 542.0, + 1255.0, + 542.0, + 1255.0, + 565.0, + 1196.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 542.0, + 1344.0, + 542.0, + 1344.0, + 566.0, + 1266.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 566.0, + 1186.0, + 566.0, + 1186.0, + 582.0, + 1170.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 562.0, + 1255.0, + 562.0, + 1255.0, + 584.0, + 1188.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 568.0, + 1282.0, + 568.0, + 1282.0, + 580.0, + 1268.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 558.0, + 1350.0, + 558.0, + 1350.0, + 586.0, + 1284.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 583.0, + 1191.0, + 583.0, + 1191.0, + 607.0, + 1150.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 592.0, + 1202.0, + 592.0, + 1202.0, + 599.0, + 1189.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 600.0, + 1163.0, + 600.0, + 1163.0, + 741.0, + 1138.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1151.0, + 613.0, + 1189.0, + 613.0, + 1189.0, + 634.0, + 1151.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 643.0, + 1188.0, + 643.0, + 1188.0, + 662.0, + 1159.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 669.0, + 1189.0, + 669.0, + 1189.0, + 690.0, + 1157.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 696.0, + 1189.0, + 696.0, + 1189.0, + 717.0, + 1157.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 727.0, + 1192.0, + 727.0, + 1192.0, + 742.0, + 1174.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 727.0, + 1374.0, + 727.0, + 1374.0, + 741.0, + 1354.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 743.0, + 1376.0, + 743.0, + 1376.0, + 765.0, + 1199.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 765.0, + 1325.0, + 765.0, + 1325.0, + 790.0, + 1251.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 794.0, + 1306.0, + 794.0, + 1306.0, + 824.0, + 1211.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.75, + 377.5, + 1367.75, + 377.5, + 1367.75, + 393.5, + 1301.75, + 393.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 642.0, + 1387.0, + 642.0, + 1387.0, + 669.5, + 1190.0, + 669.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1310.0, + 709.0, + 1310.0, + 709.0, + 1354.0, + 293.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1749.0, + 1406.0, + 1749.0, + 1406.0, + 1790.0, + 293.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1781.0, + 1194.0, + 1781.0, + 1194.0, + 1817.0, + 293.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1781.0, + 1407.0, + 1781.0, + 1407.0, + 1817.0, + 1396.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1811.0, + 1406.0, + 1811.0, + 1406.0, + 1848.0, + 293.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1841.0, + 1406.0, + 1841.0, + 1406.0, + 1879.0, + 291.0, + 1879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1873.0, + 1406.0, + 1873.0, + 1406.0, + 1911.0, + 292.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1903.0, + 614.0, + 1903.0, + 614.0, + 1943.0, + 293.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1385.0, + 1404.0, + 1385.0, + 1404.0, + 1417.0, + 296.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1412.0, + 1409.0, + 1412.0, + 1409.0, + 1453.0, + 292.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1446.0, + 1406.0, + 1446.0, + 1406.0, + 1481.0, + 292.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1474.0, + 1406.0, + 1474.0, + 1406.0, + 1510.0, + 293.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1503.0, + 1408.0, + 1503.0, + 1408.0, + 1543.0, + 292.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1538.0, + 656.0, + 1538.0, + 656.0, + 1570.0, + 296.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1405.0, + 1581.0, + 1405.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1610.0, + 1407.0, + 1610.0, + 1407.0, + 1653.0, + 292.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1644.0, + 1405.0, + 1644.0, + 1405.0, + 1681.0, + 293.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1676.0, + 1405.0, + 1676.0, + 1405.0, + 1709.0, + 296.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1705.0, + 1192.0, + 1705.0, + 1192.0, + 1742.0, + 294.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1170.0, + 555.0, + 1170.0, + 555.0, + 1209.0, + 293.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1170.0, + 1082.0, + 1170.0, + 1082.0, + 1209.0, + 647.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1191.0, + 1170.0, + 1409.0, + 1170.0, + 1409.0, + 1209.0, + 1191.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1205.0, + 644.0, + 1205.0, + 644.0, + 1239.0, + 295.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1205.0, + 805.0, + 1205.0, + 805.0, + 1239.0, + 757.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1205.0, + 1403.0, + 1205.0, + 1403.0, + 1239.0, + 904.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1232.0, + 689.0, + 1232.0, + 689.0, + 1271.0, + 294.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1043.0, + 1405.0, + 1043.0, + 1405.0, + 1081.0, + 293.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1076.0, + 702.0, + 1076.0, + 702.0, + 1110.0, + 295.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1076.0, + 1020.0, + 1076.0, + 1020.0, + 1110.0, + 782.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1076.0, + 1405.0, + 1076.0, + 1405.0, + 1110.0, + 1101.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1106.0, + 1259.0, + 1106.0, + 1259.0, + 1139.0, + 295.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1404.0, + 2002.0, + 1404.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 312, + 223, + 1096, + 223, + 1096, + 529, + 312, + 529 + ], + "score": 0.981, + "html": "
ModelSQuAD 1.1MNLI-mSST-2
Pretrained TransformerMethodF1FLOPsAccFLOPsAccFLOPs
BERTBaseStandard88.51.00x84.41.00x92.81.00x
Length-Adaptive*89.60.89x85.00.58x93.10.36x
Length-Adaptivet88.70.45x84.40.35x92.80.35x
Standard85.81.00x80.91.00x90.61.00x
DistilBERT1Length-Adaptive*86.30.81x81.50.56x92.00.55x
Length-Adaptive†85.90.59x81.30.54x91.70.54x
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1238, + 1405, + 1238, + 1405, + 1514, + 298, + 1514 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 960, + 1404, + 960, + 1404, + 1206, + 297, + 1206 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 761, + 1406, + 761, + 1406, + 945, + 298, + 945 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1548, + 1404, + 1548, + 1404, + 1672, + 299, + 1672 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 1910, + 1403, + 1910, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1784, + 1403, + 1784, + 1403, + 1877, + 299, + 1877 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 1141, + 240, + 1385, + 240, + 1385, + 507, + 1141, + 507 + ], + "score": 0.961 + }, + { + "category_id": 4, + "poly": [ + 1138, + 531, + 1389, + 531, + 1389, + 714, + 1138, + 714 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 300, + 1715, + 587, + 1715, + 587, + 1751, + 300, + 1751 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 854, + 76, + 854, + 104, + 298, + 104 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 857, + 2087, + 857, + 2111, + 841, + 2111 + ], + "score": 0.742 + }, + { + "category_id": 6, + "poly": [ + 305, + 550, + 1106, + 550, + 1106, + 735, + 305, + 735 + ], + "score": 0.684 + }, + { + "category_id": 1, + "poly": [ + 305, + 550, + 1106, + 550, + 1106, + 735, + 305, + 735 + ], + "score": 0.236 + }, + { + "category_id": 13, + "poly": [ + 826, + 1423, + 887, + 1423, + 887, + 1453, + 826, + 1453 + ], + "score": 0.87, + "latex": "\\geq 1 6" + }, + { + "category_id": 13, + "poly": [ + 702, + 885, + 812, + 885, + 812, + 916, + 702, + 916 + ], + "score": 0.62, + "latex": "\\mathbf { B E R T _ { B a s e } }" + }, + { + "category_id": 13, + "poly": [ + 772, + 886, + 812, + 886, + 812, + 916, + 772, + 916 + ], + "score": 0.3, + "latex": "_ \\mathrm { B a s e } ^ { \\prime }" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 245.0, + 1208.0, + 245.0, + 1208.0, + 267.0, + 1161.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 284.0, + 1195.0, + 284.0, + 1195.0, + 307.0, + 1160.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 324.0, + 1196.0, + 324.0, + 1196.0, + 346.0, + 1160.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 340.0, + 1164.0, + 340.0, + 1164.0, + 381.0, + 1142.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 359.0, + 1235.0, + 359.0, + 1235.0, + 372.0, + 1224.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 362.0, + 1196.0, + 362.0, + 1196.0, + 387.0, + 1161.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 403.0, + 1197.0, + 403.0, + 1197.0, + 425.0, + 1160.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 415.0, + 1216.0, + 415.0, + 1216.0, + 426.0, + 1207.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 433.0, + 1212.0, + 433.0, + 1212.0, + 448.0, + 1201.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 441.0, + 1196.0, + 441.0, + 1196.0, + 463.0, + 1160.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1370.0, + 453.0, + 1381.0, + 453.0, + 1381.0, + 468.0, + 1370.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 469.0, + 1214.0, + 469.0, + 1214.0, + 486.0, + 1199.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 467.0, + 1274.0, + 467.0, + 1274.0, + 487.0, + 1251.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 467.0, + 1329.0, + 467.0, + 1329.0, + 487.0, + 1306.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 466.0, + 1384.0, + 466.0, + 1384.0, + 487.0, + 1363.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 485.0, + 1318.0, + 485.0, + 1318.0, + 509.0, + 1260.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 332.5, + 1260.0, + 332.5, + 1260.0, + 343.5, + 1226.0, + 343.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 386.5, + 1231.0, + 386.5, + 1231.0, + 394.5, + 1207.0, + 394.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.75, + 446.0, + 1214.75, + 446.0, + 1214.75, + 466.0, + 1190.75, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 524.0, + 1396.0, + 524.0, + 1396.0, + 565.0, + 1136.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 560.0, + 1392.0, + 560.0, + 1392.0, + 592.0, + 1134.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 592.0, + 1387.0, + 592.0, + 1387.0, + 621.0, + 1137.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 622.0, + 1390.0, + 622.0, + 1390.0, + 653.0, + 1137.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 648.0, + 1389.0, + 648.0, + 1389.0, + 685.0, + 1134.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 685.0, + 1251.0, + 685.0, + 1251.0, + 714.0, + 1135.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1713.0, + 593.0, + 1713.0, + 593.0, + 1756.0, + 293.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 548.0, + 1108.0, + 548.0, + 1108.0, + 585.0, + 302.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 580.0, + 1108.0, + 580.0, + 1108.0, + 615.0, + 303.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 611.0, + 1109.0, + 611.0, + 1109.0, + 646.0, + 302.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 643.0, + 1106.0, + 643.0, + 1106.0, + 674.0, + 302.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 672.0, + 1108.0, + 672.0, + 1108.0, + 709.0, + 302.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 702.0, + 701.0, + 702.0, + 701.0, + 740.0, + 302.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1274.0, + 295.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1269.0, + 1405.0, + 1269.0, + 1405.0, + 1304.0, + 292.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1302.0, + 1402.0, + 1302.0, + 1402.0, + 1330.0, + 296.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1330.0, + 1405.0, + 1330.0, + 1405.0, + 1367.0, + 295.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1356.0, + 1406.0, + 1356.0, + 1406.0, + 1401.0, + 292.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1392.0, + 1405.0, + 1392.0, + 1405.0, + 1424.0, + 296.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1423.0, + 825.0, + 1423.0, + 825.0, + 1456.0, + 296.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1423.0, + 1407.0, + 1423.0, + 1407.0, + 1456.0, + 888.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1450.0, + 1406.0, + 1450.0, + 1406.0, + 1490.0, + 292.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 482.0, + 1480.0, + 482.0, + 1517.0, + 294.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 962.0, + 1402.0, + 962.0, + 1402.0, + 995.0, + 295.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 992.0, + 1404.0, + 992.0, + 1404.0, + 1025.0, + 295.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1023.0, + 1404.0, + 1023.0, + 1404.0, + 1057.0, + 295.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1052.0, + 1405.0, + 1052.0, + 1405.0, + 1087.0, + 294.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1083.0, + 1404.0, + 1083.0, + 1404.0, + 1117.0, + 295.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1114.0, + 1406.0, + 1114.0, + 1406.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1144.0, + 1405.0, + 1144.0, + 1405.0, + 1178.0, + 292.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 989.0, + 1176.0, + 989.0, + 1209.0, + 295.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 762.0, + 1406.0, + 762.0, + 1406.0, + 796.0, + 293.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 792.0, + 1407.0, + 792.0, + 1407.0, + 826.0, + 292.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 823.0, + 1406.0, + 823.0, + 1406.0, + 858.0, + 295.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 852.0, + 1404.0, + 852.0, + 1404.0, + 888.0, + 293.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 882.0, + 701.0, + 882.0, + 701.0, + 921.0, + 292.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 882.0, + 1406.0, + 882.0, + 1406.0, + 921.0, + 813.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 914.0, + 588.0, + 914.0, + 588.0, + 947.0, + 295.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1546.0, + 1405.0, + 1546.0, + 1405.0, + 1584.0, + 294.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1581.0, + 1408.0, + 1581.0, + 1408.0, + 1614.0, + 295.0, + 1614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1610.0, + 1405.0, + 1610.0, + 1405.0, + 1644.0, + 293.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1643.0, + 1036.0, + 1643.0, + 1036.0, + 1673.0, + 294.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1408.0, + 1940.0, + 1408.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1408.0, + 1971.0, + 1408.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1268.0, + 2004.0, + 1268.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1783.0, + 1405.0, + 1783.0, + 1405.0, + 1820.0, + 293.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1816.0, + 1405.0, + 1816.0, + 1405.0, + 1850.0, + 295.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1845.0, + 1407.0, + 1845.0, + 1407.0, + 1881.0, + 294.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 548.0, + 1108.0, + 548.0, + 1108.0, + 585.0, + 302.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 580.0, + 1108.0, + 580.0, + 1108.0, + 615.0, + 303.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 611.0, + 1109.0, + 611.0, + 1109.0, + 646.0, + 302.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 643.0, + 1106.0, + 643.0, + 1106.0, + 674.0, + 302.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 672.0, + 1108.0, + 672.0, + 1108.0, + 709.0, + 302.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 702.0, + 701.0, + 702.0, + 701.0, + 740.0, + 302.0, + 740.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1063, + 1405, + 1063, + 1405, + 1278, + 298, + 1278 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 474, + 298, + 474 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 490, + 1404, + 490, + 1404, + 705, + 298, + 705 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1392, + 1404, + 1392, + 1404, + 1669, + 298, + 1669 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1684, + 1403, + 1684, + 1403, + 1867, + 298, + 1867 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 877, + 1405, + 877, + 1405, + 1030, + 298, + 1030 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 1883, + 1403, + 1883, + 1403, + 2005, + 300, + 2005 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 720, + 1404, + 720, + 1404, + 842, + 298, + 842 + ], + "score": 0.975 + }, + { + "category_id": 0, + "poly": [ + 300, + 1324, + 842, + 1324, + 842, + 1359, + 300, + 1359 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.787 + }, + { + "category_id": 13, + "poly": [ + 1063, + 1608, + 1095, + 1608, + 1095, + 1635, + 1063, + 1635 + ], + "score": 0.55, + "latex": "3 \\mathbf { x }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1318.0, + 847.0, + 1318.0, + 847.0, + 1365.0, + 292.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1060.0, + 1406.0, + 1060.0, + 1406.0, + 1103.0, + 292.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1095.0, + 1406.0, + 1095.0, + 1406.0, + 1130.0, + 293.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1127.0, + 1406.0, + 1127.0, + 1406.0, + 1158.0, + 296.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1152.0, + 1406.0, + 1152.0, + 1406.0, + 1195.0, + 292.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1186.0, + 1405.0, + 1186.0, + 1405.0, + 1221.0, + 295.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1217.0, + 1407.0, + 1217.0, + 1407.0, + 1255.0, + 293.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1246.0, + 1041.0, + 1246.0, + 1041.0, + 1283.0, + 293.0, + 1283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 295.0, + 292.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 292.0, + 1408.0, + 292.0, + 1408.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 322.0, + 1407.0, + 322.0, + 1407.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 353.0, + 1405.0, + 353.0, + 1405.0, + 386.0, + 296.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 379.0, + 1405.0, + 379.0, + 1405.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 440.0, + 1264.0, + 440.0, + 1264.0, + 481.0, + 292.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 487.0, + 1404.0, + 487.0, + 1404.0, + 529.0, + 293.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 519.0, + 1402.0, + 519.0, + 1402.0, + 557.0, + 293.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 552.0, + 1405.0, + 552.0, + 1405.0, + 586.0, + 294.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 585.0, + 1406.0, + 585.0, + 1406.0, + 616.0, + 296.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 612.0, + 1408.0, + 612.0, + 1408.0, + 647.0, + 294.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 643.0, + 1408.0, + 643.0, + 1408.0, + 677.0, + 294.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 673.0, + 1376.0, + 673.0, + 1376.0, + 708.0, + 295.0, + 708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1390.0, + 1406.0, + 1390.0, + 1406.0, + 1428.0, + 293.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1422.0, + 1408.0, + 1422.0, + 1408.0, + 1461.0, + 293.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1453.0, + 1405.0, + 1453.0, + 1405.0, + 1490.0, + 294.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1482.0, + 1405.0, + 1482.0, + 1405.0, + 1520.0, + 293.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1515.0, + 1405.0, + 1515.0, + 1405.0, + 1551.0, + 294.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1545.0, + 1406.0, + 1545.0, + 1406.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1404.0, + 1574.0, + 1404.0, + 1612.0, + 293.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1062.0, + 1605.0, + 1062.0, + 1642.0, + 294.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1605.0, + 1406.0, + 1605.0, + 1406.0, + 1642.0, + 1096.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1637.0, + 1248.0, + 1637.0, + 1248.0, + 1672.0, + 293.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1682.0, + 1405.0, + 1682.0, + 1405.0, + 1719.0, + 293.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1716.0, + 1405.0, + 1716.0, + 1405.0, + 1748.0, + 296.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1746.0, + 1405.0, + 1746.0, + 1405.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1776.0, + 1407.0, + 1776.0, + 1407.0, + 1810.0, + 293.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1805.0, + 1405.0, + 1805.0, + 1405.0, + 1840.0, + 293.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1835.0, + 674.0, + 1835.0, + 674.0, + 1872.0, + 293.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 877.0, + 1405.0, + 877.0, + 1405.0, + 910.0, + 296.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 906.0, + 1405.0, + 906.0, + 1405.0, + 943.0, + 295.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 936.0, + 1410.0, + 936.0, + 1410.0, + 975.0, + 293.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 963.0, + 1409.0, + 963.0, + 1409.0, + 1008.0, + 292.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 998.0, + 644.0, + 998.0, + 644.0, + 1037.0, + 292.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1920.0, + 294.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1943.0, + 1409.0, + 1943.0, + 1409.0, + 1980.0, + 293.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1976.0, + 710.0, + 1976.0, + 710.0, + 2005.0, + 296.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 719.0, + 1409.0, + 719.0, + 1409.0, + 756.0, + 293.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 750.0, + 1406.0, + 750.0, + 1406.0, + 784.0, + 293.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 783.0, + 1402.0, + 783.0, + 1402.0, + 816.0, + 296.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 811.0, + 1207.0, + 811.0, + 1207.0, + 846.0, + 293.0, + 846.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 0, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 261, + 300, + 261 + ], + "score": 0.809 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.776 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.662 + }, + { + "category_id": 1, + "poly": [ + 292, + 1270, + 1402, + 1270, + 1402, + 1334, + 292, + 1334 + ], + "score": 0.66 + }, + { + "category_id": 1, + "poly": [ + 292, + 1444, + 1402, + 1444, + 1402, + 1510, + 292, + 1510 + ], + "score": 0.629 + }, + { + "category_id": 1, + "poly": [ + 300, + 280, + 1403, + 280, + 1403, + 374, + 300, + 374 + ], + "score": 0.629 + }, + { + "category_id": 1, + "poly": [ + 293, + 1357, + 1403, + 1357, + 1403, + 1422, + 293, + 1422 + ], + "score": 0.624 + }, + { + "category_id": 1, + "poly": [ + 295, + 774, + 1402, + 774, + 1402, + 839, + 295, + 839 + ], + "score": 0.608 + }, + { + "category_id": 1, + "poly": [ + 298, + 1152, + 1401, + 1152, + 1401, + 1248, + 298, + 1248 + ], + "score": 0.577 + }, + { + "category_id": 1, + "poly": [ + 295, + 1066, + 1399, + 1066, + 1399, + 1131, + 295, + 1131 + ], + "score": 0.57 + }, + { + "category_id": 1, + "poly": [ + 290, + 861, + 1403, + 861, + 1403, + 927, + 290, + 927 + ], + "score": 0.565 + }, + { + "category_id": 1, + "poly": [ + 296, + 600, + 1402, + 600, + 1402, + 667, + 296, + 667 + ], + "score": 0.559 + }, + { + "category_id": 1, + "poly": [ + 297, + 1767, + 1402, + 1767, + 1402, + 1861, + 297, + 1861 + ], + "score": 0.546 + }, + { + "category_id": 1, + "poly": [ + 295, + 687, + 1402, + 687, + 1402, + 754, + 295, + 754 + ], + "score": 0.545 + }, + { + "category_id": 1, + "poly": [ + 300, + 948, + 1402, + 948, + 1402, + 1043, + 300, + 1043 + ], + "score": 0.543 + }, + { + "category_id": 1, + "poly": [ + 292, + 1884, + 1402, + 1884, + 1402, + 1948, + 292, + 1948 + ], + "score": 0.501 + }, + { + "category_id": 1, + "poly": [ + 293, + 396, + 1402, + 396, + 1402, + 461, + 293, + 461 + ], + "score": 0.489 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1399, + 1971, + 1399, + 2035, + 297, + 2035 + ], + "score": 0.486 + }, + { + "category_id": 1, + "poly": [ + 298, + 483, + 1402, + 483, + 1402, + 579, + 298, + 579 + ], + "score": 0.477 + }, + { + "category_id": 1, + "poly": [ + 296, + 1531, + 1404, + 1531, + 1404, + 1626, + 296, + 1626 + ], + "score": 0.3 + }, + { + "category_id": 1, + "poly": [ + 294, + 1648, + 1408, + 1648, + 1408, + 1742, + 294, + 1742 + ], + "score": 0.263 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.222 + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 490.0, + 229.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1271.0, + 1403.0, + 1271.0, + 1403.0, + 1307.0, + 297.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1303.0, + 707.0, + 1303.0, + 707.0, + 1335.0, + 320.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1447.0, + 1403.0, + 1447.0, + 1403.0, + 1480.0, + 297.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1474.0, + 1347.0, + 1474.0, + 1347.0, + 1512.0, + 321.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 278.0, + 1407.0, + 278.0, + 1407.0, + 315.0, + 295.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 309.0, + 1405.0, + 309.0, + 1405.0, + 348.0, + 321.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 342.0, + 981.0, + 342.0, + 981.0, + 376.0, + 324.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1356.0, + 1406.0, + 1356.0, + 1406.0, + 1396.0, + 293.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1390.0, + 852.0, + 1390.0, + 852.0, + 1423.0, + 322.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 776.0, + 1404.0, + 776.0, + 1404.0, + 812.0, + 294.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 808.0, + 706.0, + 808.0, + 706.0, + 841.0, + 320.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1151.0, + 1405.0, + 1151.0, + 1405.0, + 1189.0, + 293.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1184.0, + 1403.0, + 1184.0, + 1403.0, + 1220.0, + 322.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1217.0, + 706.0, + 1217.0, + 706.0, + 1248.0, + 321.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1068.0, + 1403.0, + 1068.0, + 1403.0, + 1104.0, + 296.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1099.0, + 1398.0, + 1099.0, + 1398.0, + 1132.0, + 321.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 862.0, + 1404.0, + 862.0, + 1404.0, + 898.0, + 297.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 894.0, + 997.0, + 894.0, + 997.0, + 928.0, + 321.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 601.0, + 1404.0, + 601.0, + 1404.0, + 637.0, + 295.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 632.0, + 1289.0, + 632.0, + 1289.0, + 669.0, + 321.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1767.0, + 1406.0, + 1767.0, + 1406.0, + 1804.0, + 295.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1799.0, + 1406.0, + 1799.0, + 1406.0, + 1833.0, + 324.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1830.0, + 931.0, + 1830.0, + 931.0, + 1861.0, + 324.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 685.0, + 1404.0, + 685.0, + 1404.0, + 726.0, + 294.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 719.0, + 1390.0, + 719.0, + 1390.0, + 756.0, + 323.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 950.0, + 1404.0, + 950.0, + 1404.0, + 985.0, + 296.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 983.0, + 1407.0, + 983.0, + 1407.0, + 1014.0, + 325.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1011.0, + 775.0, + 1011.0, + 775.0, + 1045.0, + 321.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1921.0, + 295.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1915.0, + 1157.0, + 1915.0, + 1157.0, + 1949.0, + 321.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 398.0, + 1404.0, + 398.0, + 1404.0, + 434.0, + 295.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 429.0, + 922.0, + 429.0, + 922.0, + 462.0, + 324.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1967.0, + 1405.0, + 1967.0, + 1405.0, + 2009.0, + 292.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2002.0, + 1082.0, + 2002.0, + 1082.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 483.0, + 1407.0, + 483.0, + 1407.0, + 521.0, + 296.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 515.0, + 1403.0, + 515.0, + 1403.0, + 549.0, + 323.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 546.0, + 857.0, + 546.0, + 857.0, + 581.0, + 322.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1532.0, + 1406.0, + 1532.0, + 1406.0, + 1566.0, + 296.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1563.0, + 1405.0, + 1563.0, + 1405.0, + 1598.0, + 322.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1594.0, + 612.0, + 1594.0, + 612.0, + 1624.0, + 323.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1684.0, + 296.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1682.0, + 1409.0, + 1682.0, + 1409.0, + 1716.0, + 324.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1712.0, + 775.0, + 1712.0, + 775.0, + 1743.0, + 322.0, + 1743.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.831 + }, + { + "category_id": 1, + "poly": [ + 299, + 972, + 1402, + 972, + 1402, + 1066, + 299, + 1066 + ], + "score": 0.572 + }, + { + "category_id": 1, + "poly": [ + 294, + 890, + 1402, + 890, + 1402, + 955, + 294, + 955 + ], + "score": 0.572 + }, + { + "category_id": 1, + "poly": [ + 294, + 809, + 1400, + 809, + 1400, + 874, + 294, + 874 + ], + "score": 0.548 + }, + { + "category_id": 1, + "poly": [ + 296, + 727, + 1402, + 727, + 1402, + 794, + 296, + 794 + ], + "score": 0.489 + }, + { + "category_id": 1, + "poly": [ + 293, + 1084, + 1402, + 1084, + 1402, + 1149, + 293, + 1149 + ], + "score": 0.486 + }, + { + "category_id": 1, + "poly": [ + 292, + 1634, + 1402, + 1634, + 1402, + 1699, + 292, + 1699 + ], + "score": 0.485 + }, + { + "category_id": 1, + "poly": [ + 295, + 1471, + 1401, + 1471, + 1401, + 1536, + 295, + 1536 + ], + "score": 0.477 + }, + { + "category_id": 1, + "poly": [ + 296, + 340, + 1402, + 340, + 1402, + 406, + 296, + 406 + ], + "score": 0.474 + }, + { + "category_id": 1, + "poly": [ + 296, + 1552, + 1400, + 1552, + 1400, + 1618, + 296, + 1618 + ], + "score": 0.471 + }, + { + "category_id": 1, + "poly": [ + 301, + 615, + 1403, + 615, + 1403, + 711, + 301, + 711 + ], + "score": 0.468 + }, + { + "category_id": 1, + "poly": [ + 302, + 1277, + 1401, + 1277, + 1401, + 1373, + 302, + 1373 + ], + "score": 0.463 + }, + { + "category_id": 1, + "poly": [ + 303, + 1165, + 1399, + 1165, + 1399, + 1261, + 303, + 1261 + ], + "score": 0.462 + }, + { + "category_id": 1, + "poly": [ + 303, + 423, + 1400, + 423, + 1400, + 517, + 303, + 517 + ], + "score": 0.449 + }, + { + "category_id": 1, + "poly": [ + 293, + 534, + 1404, + 534, + 1404, + 599, + 293, + 599 + ], + "score": 0.448 + }, + { + "category_id": 1, + "poly": [ + 300, + 229, + 1401, + 229, + 1401, + 323, + 300, + 323 + ], + "score": 0.447 + }, + { + "category_id": 1, + "poly": [ + 300, + 1717, + 1401, + 1717, + 1401, + 1811, + 300, + 1811 + ], + "score": 0.438 + }, + { + "category_id": 1, + "poly": [ + 290, + 1389, + 1402, + 1389, + 1402, + 1455, + 290, + 1455 + ], + "score": 0.431 + }, + { + "category_id": 1, + "poly": [ + 299, + 1829, + 1402, + 1829, + 1402, + 1923, + 299, + 1923 + ], + "score": 0.43 + }, + { + "category_id": 1, + "poly": [ + 299, + 1942, + 1401, + 1942, + 1401, + 2033, + 299, + 2033 + ], + "score": 0.289 + }, + { + "category_id": 13, + "poly": [ + 1146, + 537, + 1254, + 537, + 1254, + 567, + 1146, + 567 + ], + "score": 0.73, + "latex": "^ { 1 0 0 , 0 0 0 + }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 974.0, + 1404.0, + 974.0, + 1404.0, + 1008.0, + 295.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1001.0, + 1409.0, + 1001.0, + 1409.0, + 1040.0, + 319.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1029.0, + 397.0, + 1029.0, + 397.0, + 1070.0, + 320.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 886.0, + 1405.0, + 886.0, + 1405.0, + 929.0, + 293.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 921.0, + 611.0, + 921.0, + 611.0, + 957.0, + 322.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 812.0, + 1405.0, + 812.0, + 1405.0, + 844.0, + 298.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 840.0, + 1231.0, + 840.0, + 1231.0, + 875.0, + 321.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 728.0, + 1404.0, + 728.0, + 1404.0, + 764.0, + 298.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 759.0, + 1247.0, + 759.0, + 1247.0, + 794.0, + 323.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1086.0, + 1404.0, + 1086.0, + 1404.0, + 1122.0, + 296.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1115.0, + 1130.0, + 1115.0, + 1130.0, + 1150.0, + 321.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1636.0, + 1405.0, + 1636.0, + 1405.0, + 1672.0, + 295.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1664.0, + 1143.0, + 1664.0, + 1143.0, + 1704.0, + 323.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1474.0, + 1403.0, + 1474.0, + 1403.0, + 1506.0, + 298.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1502.0, + 1114.0, + 1502.0, + 1114.0, + 1537.0, + 322.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 337.0, + 1406.0, + 337.0, + 1406.0, + 382.0, + 294.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 373.0, + 1129.0, + 373.0, + 1129.0, + 409.0, + 324.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1551.0, + 1402.0, + 1551.0, + 1402.0, + 1588.0, + 296.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1585.0, + 932.0, + 1585.0, + 932.0, + 1619.0, + 321.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 610.0, + 1405.0, + 610.0, + 1405.0, + 657.0, + 292.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 647.0, + 1408.0, + 647.0, + 1408.0, + 684.0, + 322.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 678.0, + 704.0, + 678.0, + 704.0, + 712.0, + 322.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1277.0, + 1407.0, + 1277.0, + 1407.0, + 1315.0, + 296.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1312.0, + 1404.0, + 1312.0, + 1404.0, + 1346.0, + 323.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1341.0, + 810.0, + 1341.0, + 810.0, + 1375.0, + 323.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1165.0, + 1403.0, + 1165.0, + 1403.0, + 1202.0, + 295.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1197.0, + 1405.0, + 1197.0, + 1405.0, + 1235.0, + 323.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1229.0, + 917.0, + 1229.0, + 917.0, + 1262.0, + 321.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 419.0, + 1405.0, + 419.0, + 1405.0, + 462.0, + 294.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 453.0, + 1405.0, + 453.0, + 1405.0, + 489.0, + 322.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 485.0, + 920.0, + 485.0, + 920.0, + 517.0, + 323.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 536.0, + 1145.0, + 536.0, + 1145.0, + 572.0, + 295.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 536.0, + 1405.0, + 536.0, + 1405.0, + 572.0, + 1255.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 567.0, + 1140.0, + 567.0, + 1140.0, + 600.0, + 321.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 261.0, + 1407.0, + 261.0, + 1407.0, + 295.0, + 321.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 288.0, + 398.0, + 288.0, + 398.0, + 326.0, + 320.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1715.0, + 1406.0, + 1715.0, + 1406.0, + 1756.0, + 294.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1747.0, + 1406.0, + 1747.0, + 1406.0, + 1785.0, + 321.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1777.0, + 882.0, + 1777.0, + 882.0, + 1815.0, + 321.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1386.0, + 1405.0, + 1386.0, + 1405.0, + 1431.0, + 292.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1422.0, + 955.0, + 1422.0, + 955.0, + 1457.0, + 322.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1829.0, + 1404.0, + 1829.0, + 1404.0, + 1864.0, + 294.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1859.0, + 1404.0, + 1859.0, + 1404.0, + 1894.0, + 323.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1892.0, + 798.0, + 1892.0, + 798.0, + 1923.0, + 320.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1976.0, + 297.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 320.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 609.0, + 2003.0, + 609.0, + 2032.0, + 323.0, + 2032.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 652, + 1402, + 652, + 1402, + 716, + 299, + 716 + ], + "score": 0.828 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 861, + 2088, + 861, + 2113, + 835, + 2113 + ], + "score": 0.825 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.824 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1404, + 229, + 1404, + 322, + 299, + 322 + ], + "score": 0.673 + }, + { + "category_id": 1, + "poly": [ + 298, + 539, + 1401, + 539, + 1401, + 634, + 298, + 634 + ], + "score": 0.64 + }, + { + "category_id": 1, + "poly": [ + 293, + 455, + 1403, + 455, + 1403, + 519, + 293, + 519 + ], + "score": 0.583 + }, + { + "category_id": 1, + "poly": [ + 299, + 343, + 1402, + 343, + 1402, + 437, + 299, + 437 + ], + "score": 0.581 + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 867.0, + 2085.0, + 867.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 651.0, + 1407.0, + 651.0, + 1407.0, + 689.0, + 294.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 683.0, + 1406.0, + 683.0, + 1406.0, + 719.0, + 322.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1406.0, + 227.0, + 1406.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 297.0, + 322.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 292.0, + 609.0, + 292.0, + 609.0, + 321.0, + 322.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 537.0, + 1406.0, + 537.0, + 1406.0, + 577.0, + 293.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 569.0, + 1406.0, + 569.0, + 1406.0, + 607.0, + 322.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 600.0, + 936.0, + 600.0, + 936.0, + 636.0, + 323.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 456.0, + 1404.0, + 456.0, + 1404.0, + 492.0, + 295.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 487.0, + 1085.0, + 487.0, + 1085.0, + 520.0, + 322.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 340.0, + 1407.0, + 340.0, + 1407.0, + 380.0, + 293.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 377.0, + 1404.0, + 377.0, + 1404.0, + 407.0, + 324.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 403.0, + 1100.0, + 403.0, + 1100.0, + 440.0, + 323.0, + 440.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/vlm/dev/3R3Pz5i0tye/0.png b/vlm/dev/3R3Pz5i0tye/0.png new file mode 100644 index 0000000000000000000000000000000000000000..da4199ae29818795c17a555a10c4a0bf207e75b9 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fedea5ebf224830e980138ab2805408ec9c2919c021ed65485ee4232476a7ac +size 531424 diff --git a/vlm/dev/3R3Pz5i0tye/1.png b/vlm/dev/3R3Pz5i0tye/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a2e331d644299032c9f686d203656ccf49ca44d8 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5af5c40541bae7458bc7a6c84d7b49d747336d9c7b1c136bfe6feafc39cb6ddc +size 736550 diff --git a/vlm/dev/3R3Pz5i0tye/10.png b/vlm/dev/3R3Pz5i0tye/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2615244ed01eae5abbb68a4d34439a1f271fba30 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea44a07953166ad1d7d53e848c5f2376863ccb735a150bdbb6e2d058526c6f38 +size 555625 diff --git a/vlm/dev/3R3Pz5i0tye/11.png b/vlm/dev/3R3Pz5i0tye/11.png new file mode 100644 index 0000000000000000000000000000000000000000..6fb52f42d3a2970a98a3263d8d74b6bddde95dc1 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05ef0c283885073235f393638c38fc82cc740942318e912693dc3129c1f62b32 +size 592261 diff --git a/vlm/dev/3R3Pz5i0tye/12.png b/vlm/dev/3R3Pz5i0tye/12.png new file mode 100644 index 0000000000000000000000000000000000000000..920591b1f9751d010a4ef64bf352ea8673cf19b4 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c3e47d6bbad42b548bfcc3bb92408386a6ce44ac93b096d29699d204d2dea95 +size 563703 diff --git a/vlm/dev/3R3Pz5i0tye/13.png b/vlm/dev/3R3Pz5i0tye/13.png new file mode 100644 index 0000000000000000000000000000000000000000..9de3cfd57a572eca9fcbccc5a46cd1d28b1cf6c4 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:788886d16ad08cc425fdb53b47ce7de91e25efa200805886f7f65f602b1d774d +size 187529 diff --git a/vlm/dev/3R3Pz5i0tye/2.png b/vlm/dev/3R3Pz5i0tye/2.png new file mode 100644 index 0000000000000000000000000000000000000000..8173b2238e207fd6a93e97a31c2245382017b49b --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ed759c58fd3c0f934574cac5ef5daca8838bc22628092d29e03d8643232c666 +size 790832 diff --git a/vlm/dev/3R3Pz5i0tye/3.png b/vlm/dev/3R3Pz5i0tye/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b07c4cdaab0943bcbfc5c9c46b1977488272fce1 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:130ca474d2f54d025c5053482d929c4ffe7579eb97b3f689692b0aac14051443 +size 848761 diff --git a/vlm/dev/3R3Pz5i0tye/4.png b/vlm/dev/3R3Pz5i0tye/4.png new file mode 100644 index 0000000000000000000000000000000000000000..2fdb1669bcd9a5cd15b4d395197d71848d7ccd80 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7883dff7602e331313cd4a9cf77e27960a1cbf8132438969588c73055a1e2d8a +size 588336 diff --git a/vlm/dev/3R3Pz5i0tye/5.png b/vlm/dev/3R3Pz5i0tye/5.png new file mode 100644 index 0000000000000000000000000000000000000000..97cef2361f441df01d7492118b8fbf34c481f5c5 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9700d6b7eaec132f5a94ec2e171a31751c193c45cc7cd9c91a6af5588ac270ee +size 583404 diff --git a/vlm/dev/3R3Pz5i0tye/6.png b/vlm/dev/3R3Pz5i0tye/6.png new file mode 100644 index 0000000000000000000000000000000000000000..93c6e0891dfadb3077270a446fc7f54dae60d575 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6634af8ec27c75bfbd1de2bc1547523597bd5e83a5257ff6620f30387fe57ea6 +size 573472 diff --git a/vlm/dev/3R3Pz5i0tye/7.png b/vlm/dev/3R3Pz5i0tye/7.png new file mode 100644 index 0000000000000000000000000000000000000000..26761c8bc14271d6a6f21609b02cbe369298b50c --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9337bb4cb139f8b19fef175dab27a40f5fcf34a3c7e1f391390d752e7c7d6769 +size 806976 diff --git a/vlm/dev/3R3Pz5i0tye/8.png b/vlm/dev/3R3Pz5i0tye/8.png new file mode 100644 index 0000000000000000000000000000000000000000..d5596759d8471d95e331314132f206d5c5dd1be0 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0a972f625b62d1e720f52153197d6c62bcb131602975b00442cd9477ba84eb5 +size 527650 diff --git a/vlm/dev/3R3Pz5i0tye/9.png b/vlm/dev/3R3Pz5i0tye/9.png new file mode 100644 index 0000000000000000000000000000000000000000..32d70295e23521cbde5d9eaf1ad30de72961dea3 --- /dev/null +++ b/vlm/dev/3R3Pz5i0tye/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b940293ca7dd6ae051a82ce0d59e918ed19a67ab5a9deb7293ada6ad7b2776db +size 575275 diff --git a/vlm/dev/401LFvBGIb/0.png b/vlm/dev/401LFvBGIb/0.png new file mode 100644 index 0000000000000000000000000000000000000000..bbebabde0f46238abb1d2a66a0c5161d57f32d6f --- /dev/null +++ b/vlm/dev/401LFvBGIb/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92e8ef8ccdf626f7de2962a7f2d353c7f5d174b7bfadfcb871572103bb579996 +size 436579 diff --git a/vlm/dev/401LFvBGIb/1.png b/vlm/dev/401LFvBGIb/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0bd0d8842ebbc5eefb4e7e4098d5497131d09430 --- /dev/null +++ b/vlm/dev/401LFvBGIb/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c05f53f9b3230745adef239e7909271b8659bd10a9f02d2c335c9752c36a529 +size 490493 diff --git a/vlm/dev/401LFvBGIb/2.png b/vlm/dev/401LFvBGIb/2.png new file mode 100644 index 0000000000000000000000000000000000000000..6b1d25d36a8f8d3f201390385ceb4e31f49b8cd7 --- /dev/null +++ b/vlm/dev/401LFvBGIb/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:060f02880e069b1148ca3f72286189f965f38497ce6b5fd5910c9f1fc890ef26 +size 422767 diff --git a/vlm/dev/401LFvBGIb/3.png b/vlm/dev/401LFvBGIb/3.png new file mode 100644 index 0000000000000000000000000000000000000000..8ad103a6f9a7a0c51fb60cd4dd2bbfc77ff53384 --- /dev/null +++ b/vlm/dev/401LFvBGIb/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00f9671aa6cfede300d117ef74e91ac4ac81cb9624a892918c623cccc22c36e9 +size 405437 diff --git a/vlm/dev/401LFvBGIb/4.png b/vlm/dev/401LFvBGIb/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ee50b388360faf3c6f2f2747ebbcd23386e7ba51 --- /dev/null +++ b/vlm/dev/401LFvBGIb/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccec683eb69f8af84206e2b487b1c9657eaed9fd0fe46b547ec6abd9c9ab7ecd +size 526713 diff --git a/vlm/dev/401LFvBGIb/5.png b/vlm/dev/401LFvBGIb/5.png new file mode 100644 index 0000000000000000000000000000000000000000..ee54403c290d28b2af2fb95047fd31fee03e912e --- /dev/null +++ b/vlm/dev/401LFvBGIb/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d0fcaf14dcb61f1ae029973bbe8129553195a12a73dd4a41cf9fbd25ad18caa +size 60370 diff --git a/vlm/dev/401LFvBGIb/6.png b/vlm/dev/401LFvBGIb/6.png new file mode 100644 index 0000000000000000000000000000000000000000..b162ffb63bfe39dd8b62d782f5ee1bd2d64765bd --- /dev/null +++ b/vlm/dev/401LFvBGIb/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02da432c19b7c3827ea22ac5b70441d19436eccc0b83259953adbca395baab55 +size 516135 diff --git a/vlm/dev/401LFvBGIb/7.png b/vlm/dev/401LFvBGIb/7.png new file mode 100644 index 0000000000000000000000000000000000000000..06c31dfee9d1d08502bdc858d9b0908fd137d1b1 --- /dev/null +++ b/vlm/dev/401LFvBGIb/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d618261683ad6d755572cf8440b16bfa8798914bd62ceab113add6f7d35e982 +size 499375 diff --git a/vlm/dev/401LFvBGIb/8.png b/vlm/dev/401LFvBGIb/8.png new file mode 100644 index 0000000000000000000000000000000000000000..9ea5d2db6b37cf034277ba523328c066e5543626 --- /dev/null +++ b/vlm/dev/401LFvBGIb/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:baa92d69ad1c5eb41909f8581e47c20f1e7898a7d0bd45254c01e3bb9100cf23 +size 479219 diff --git a/vlm/dev/6u6N8WWwYSM/0.png b/vlm/dev/6u6N8WWwYSM/0.png new file mode 100644 index 0000000000000000000000000000000000000000..001bc8ae588e64e4db51df0d0a2f2705ce692ccc --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43b167799284084da274ca211c50ec0f0d4b3c66fae6c613b6f7082b36d95a12 +size 603380 diff --git a/vlm/dev/6u6N8WWwYSM/1.png b/vlm/dev/6u6N8WWwYSM/1.png new file mode 100644 index 0000000000000000000000000000000000000000..06655d169301db7fcc1765b82d6a9a4e0f7bb748 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6063abddd0db2e2d53aff682ccc02d810d98a728a56bb39451d511e072f0acb9 +size 614714 diff --git a/vlm/dev/6u6N8WWwYSM/10.png b/vlm/dev/6u6N8WWwYSM/10.png new file mode 100644 index 0000000000000000000000000000000000000000..470bc0c6b69fcf9944ba055218ba3e8bb92c4051 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:536e8c2341aded40def944cba8491aa3b94e166926638ce65a8193e3c8321cec +size 567768 diff --git a/vlm/dev/6u6N8WWwYSM/11.png b/vlm/dev/6u6N8WWwYSM/11.png new file mode 100644 index 0000000000000000000000000000000000000000..3b52acf4e4415b483eb969edb3d80f35dc3250f2 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b30724b7ba3dbd929d71d18a131ce610abf6da3e463a25456c4c7256770d7b6 +size 383052 diff --git a/vlm/dev/6u6N8WWwYSM/12.png b/vlm/dev/6u6N8WWwYSM/12.png new file mode 100644 index 0000000000000000000000000000000000000000..29d8a35b0cd46c71b5feec3abe25f676ad37f6c4 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bced22b9be5e198a4920a6f8373e0a56ca153efb211fdc5e18a40d574dd77a8 +size 527264 diff --git a/vlm/dev/6u6N8WWwYSM/13.png b/vlm/dev/6u6N8WWwYSM/13.png new file mode 100644 index 0000000000000000000000000000000000000000..d97f6d03f2b1b89d650dedd04e65f03034f0f6f9 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d5832f058e645823efd5a6143f004c25b3374934a5de74ee308c59eb4883071 +size 410630 diff --git a/vlm/dev/6u6N8WWwYSM/14.png b/vlm/dev/6u6N8WWwYSM/14.png new file mode 100644 index 0000000000000000000000000000000000000000..c8b27cca8ab09ac2cb459434bb5c49150f7c7036 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42ad05681c507232a4b03584d5347747fb8b0181dd19f609e974108104a44b92 +size 1444192 diff --git a/vlm/dev/6u6N8WWwYSM/15.png b/vlm/dev/6u6N8WWwYSM/15.png new file mode 100644 index 0000000000000000000000000000000000000000..78f4c0fca552966007b1540a31d19610496998c6 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7467d4e321dc78ba49d4223588c1310b069a228ff8606cc01e0a341b4e5f42d6 +size 846964 diff --git a/vlm/dev/6u6N8WWwYSM/16.png b/vlm/dev/6u6N8WWwYSM/16.png new file mode 100644 index 0000000000000000000000000000000000000000..4ef114dc8573b6f5c7c8ad6ea9cea5604bc87463 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f31dd06d1e13f2628d0b4b4b41257266aa004dfa6ca3cf0a62da61e9c5e3a5e +size 370776 diff --git a/vlm/dev/6u6N8WWwYSM/2.png b/vlm/dev/6u6N8WWwYSM/2.png new file mode 100644 index 0000000000000000000000000000000000000000..f645914287a90b5b04ad639356570d03b73c627e --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dc66075ca03a04e4efa36faddc8d9b12ee85411d3b99ba33ee759eab1e09872 +size 536052 diff --git a/vlm/dev/6u6N8WWwYSM/3.png b/vlm/dev/6u6N8WWwYSM/3.png new file mode 100644 index 0000000000000000000000000000000000000000..0630d990b8b7c50047442caa2f5ec3bf8cfa8b71 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a42eace5dbcecfcc05baac9c84c390a6101368abdf1b63a8197cb41a6c0de58 +size 710785 diff --git a/vlm/dev/6u6N8WWwYSM/4.png b/vlm/dev/6u6N8WWwYSM/4.png new file mode 100644 index 0000000000000000000000000000000000000000..f1d820d161fb8accc61092dd13ca34f266afe9a5 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f3cd0b1964b06417f4b119e7a2e711a20269793cfd79dc4e6800e588da26f22 +size 574177 diff --git a/vlm/dev/6u6N8WWwYSM/5.png b/vlm/dev/6u6N8WWwYSM/5.png new file mode 100644 index 0000000000000000000000000000000000000000..12269f22debb9f48ad9f23910c039abc48b616d4 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b373cedc0d4e1860059196441ef464ced36c3467e43ddd9e820c704ccc229527 +size 610126 diff --git a/vlm/dev/6u6N8WWwYSM/6.png b/vlm/dev/6u6N8WWwYSM/6.png new file mode 100644 index 0000000000000000000000000000000000000000..86812de949a1a353f0329e7ab713b2e9c2cac862 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:707a2b25a054d185f496de43699743133e4d2f66322d60696058f2cca855ca31 +size 1069859 diff --git a/vlm/dev/6u6N8WWwYSM/7.png b/vlm/dev/6u6N8WWwYSM/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c3460d78fdc32397d6b6f5d4cd7fc5705fd9cd4d --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f1b4b4d9a76f45fb3582a7700e7c2643245835e42799e767cb443c8c6bcdd90 +size 1042816 diff --git a/vlm/dev/6u6N8WWwYSM/8.png b/vlm/dev/6u6N8WWwYSM/8.png new file mode 100644 index 0000000000000000000000000000000000000000..0268ce57b4726e64ddcd9bb0ac59ff4caebac1e6 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8968d049dbfa4ad7956c7f7eac8a37daff0eafab2cf7dc901754742a7efb7b6f +size 517975 diff --git a/vlm/dev/6u6N8WWwYSM/9.png b/vlm/dev/6u6N8WWwYSM/9.png new file mode 100644 index 0000000000000000000000000000000000000000..6d75a25b457ad1968ca4e2fd6bf710f16986e842 --- /dev/null +++ b/vlm/dev/6u6N8WWwYSM/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59702a20824f4cbf4410f1bb882c5a92a05d2b22cc3aca8d802d58089ab01240 +size 496540 diff --git a/vlm/dev/8aHzds2uUyB/13.png b/vlm/dev/8aHzds2uUyB/13.png new file mode 100644 index 0000000000000000000000000000000000000000..987ac794d381b369c22024c1f828bc4fa8daf750 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9005e95fa68a9a8d621a239f9878dd998ada3178fabb3a9502b63f9402285ad2 +size 548515 diff --git a/vlm/dev/8aHzds2uUyB/16.png b/vlm/dev/8aHzds2uUyB/16.png new file mode 100644 index 0000000000000000000000000000000000000000..b90843300b03cc5f3936d1d7733fb72dadb43bc5 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bed3a2672986738f8d79606a7946f72c5549ba286a174ba2cd134ba0f52fd38 +size 150985 diff --git a/vlm/dev/8aHzds2uUyB/18.png b/vlm/dev/8aHzds2uUyB/18.png new file mode 100644 index 0000000000000000000000000000000000000000..c688d1a95d144c3c05d3af6d333a885ad055ea45 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffa011ab078f8f43433d6500fba3ff8c465e844250f1e59545553eeca32d221b +size 507984 diff --git a/vlm/dev/8aHzds2uUyB/19.png b/vlm/dev/8aHzds2uUyB/19.png new file mode 100644 index 0000000000000000000000000000000000000000..07db9320a789243f0b00094f76340e7828103828 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c913e64018da73ffe7d61abc77114d7479b4f9f14f78bdb8b614e17ae547cfc +size 369083 diff --git a/vlm/dev/8aHzds2uUyB/26.png b/vlm/dev/8aHzds2uUyB/26.png new file mode 100644 index 0000000000000000000000000000000000000000..3d1eed8131e7590067b42b65e81d504ab3081b3f --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48abbf261dc8dfb8c4007b1780589f35ee3607054bd80f23c79329afaaf7b342 +size 116550 diff --git a/vlm/dev/8aHzds2uUyB/32.png b/vlm/dev/8aHzds2uUyB/32.png new file mode 100644 index 0000000000000000000000000000000000000000..d66592b440a9b5b48a906adfa3acc1b3ec540cd6 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/32.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28a9f1617f9eef9fc8fc60dbe86412a431d81d3a0d67f9d710ed8426f9b60dcc +size 556368 diff --git a/vlm/dev/8aHzds2uUyB/34.png b/vlm/dev/8aHzds2uUyB/34.png new file mode 100644 index 0000000000000000000000000000000000000000..f07cd66fb5568230f7983e7e212b60281dbb0f52 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/34.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bd5df9cea7950736ba3cf278aba603b4df9f6c3cefff8b223853ca2adb41ae1 +size 340162 diff --git a/vlm/dev/8aHzds2uUyB/35.png b/vlm/dev/8aHzds2uUyB/35.png new file mode 100644 index 0000000000000000000000000000000000000000..0eb4be599c37916be261a85367cdab2a0e0cca17 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/35.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81bf3b32e917892b4c6dd64be60789fc5afed688d6e4f7814f62735618dc6720 +size 247633 diff --git a/vlm/dev/8aHzds2uUyB/39.png b/vlm/dev/8aHzds2uUyB/39.png new file mode 100644 index 0000000000000000000000000000000000000000..0c0281c98a2f0c14d3401fdef7559cd3cb5500bd --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/39.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ef271fa201ce247bb8ada49e7449c8e9f254d2f35503db298480cad2535ea0f +size 687586 diff --git a/vlm/dev/8aHzds2uUyB/4.png b/vlm/dev/8aHzds2uUyB/4.png new file mode 100644 index 0000000000000000000000000000000000000000..447d63b9d56b904e259eb821cb699ef1ed5511cf --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adf4bd2229862eb9d31090cfe5f0380e6f1d7fc2b86dbc7fcdc5f5514a53f8aa +size 472127 diff --git a/vlm/dev/8aHzds2uUyB/44.png b/vlm/dev/8aHzds2uUyB/44.png new file mode 100644 index 0000000000000000000000000000000000000000..aa9b8fb6eb8e8c1c9d50161b5c093a150bf45bc7 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/44.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70eac9a9e96e841f50f3e9ca511023f1862ab454f1f6913a801f10aa1d62483b +size 283400 diff --git a/vlm/dev/8aHzds2uUyB/47.png b/vlm/dev/8aHzds2uUyB/47.png new file mode 100644 index 0000000000000000000000000000000000000000..9e73c2d7843f0b7c9d197817df3ff5911b4e8c34 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/47.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24e796f900315ad21e9e4e2fd80d06bba7a9e7fde41909a6fa67eeaac62c8cbe +size 596296 diff --git a/vlm/dev/8aHzds2uUyB/48.png b/vlm/dev/8aHzds2uUyB/48.png new file mode 100644 index 0000000000000000000000000000000000000000..13caeb2af41c3281dda93b298ef70b98c2f92c09 --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/48.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccbe1276129521cecd270e388f30b144edb7321d4a895f18a0489b7ad28208aa +size 258648 diff --git a/vlm/dev/8aHzds2uUyB/50.png b/vlm/dev/8aHzds2uUyB/50.png new file mode 100644 index 0000000000000000000000000000000000000000..6c79115445ab0a5b1ec3c188ed9ce49a56cf9d5c --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/50.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c7da367de9ac96a56e7bc8175764d59e07dd59687d85b0b9b6c04151018eefa +size 619377 diff --git a/vlm/dev/8aHzds2uUyB/51.png b/vlm/dev/8aHzds2uUyB/51.png new file mode 100644 index 0000000000000000000000000000000000000000..a90fc426b2e5f5061a1bd0ce5ea9cb53fa189efb --- /dev/null +++ b/vlm/dev/8aHzds2uUyB/51.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abfc7aaa835255a9cbcb7c3796286e9b67660d721a8d9f09608d56168181a97b +size 465250 diff --git a/vlm/dev/9h3KsOVXhLZ/0.png b/vlm/dev/9h3KsOVXhLZ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..3225ed07297c37d155873d74071ed789cc7a6fc4 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90859135256a87eb68d8a3593c5c74deeb3d9afeed984e4966f62a754a3f65fb +size 449702 diff --git a/vlm/dev/9h3KsOVXhLZ/1.png b/vlm/dev/9h3KsOVXhLZ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..f7d4bca08325c610a9ee96099e54d2ac17d7cf7c --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c2cd82b597b070216dc3b22b9d7308b356ddcf5f70dbb9df1d6e4b4a4b23d7d +size 496069 diff --git a/vlm/dev/9h3KsOVXhLZ/10.png b/vlm/dev/9h3KsOVXhLZ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..da2f9b9bd3a49fa0e28117f1acf41fe7e9b0c567 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcd8d4697a82f072d335a821d954ebaadf2b4d4f30a8d73dfe505675f16789fe +size 436567 diff --git a/vlm/dev/9h3KsOVXhLZ/11.png b/vlm/dev/9h3KsOVXhLZ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..3e84448ee4c076c5e44b30bbdc02f2d6f119ddd7 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d4dde70efe5ad669568dee3e45b14f1a8c8f2367e4a05360c6f1f8a599634a9 +size 336185 diff --git a/vlm/dev/9h3KsOVXhLZ/2.png b/vlm/dev/9h3KsOVXhLZ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..67c9bdfef3818e3e7f8cd35eb53e0fc4dfa6cc47 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf9e5dac57dcadb854514944fd9c8bd5a864064127f0ec3dad697303d4f53192 +size 618328 diff --git a/vlm/dev/9h3KsOVXhLZ/3.png b/vlm/dev/9h3KsOVXhLZ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..450b474dae35e15b214679b323dfd4057834b8b0 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2724a0bb12c839271393066d164e08c68eb2a16302ce84c3e480189ea97e95e5 +size 501991 diff --git a/vlm/dev/9h3KsOVXhLZ/4.png b/vlm/dev/9h3KsOVXhLZ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..eeacc5c08132d6ec91362ba82f4fc231207408e9 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2d2e207d70b971938425e742cb49a4705d4c15d024733b0ec45d735fbcfc5c1 +size 487446 diff --git a/vlm/dev/9h3KsOVXhLZ/5.png b/vlm/dev/9h3KsOVXhLZ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7cbba15eb2b972b2093207ba3042738015b1c5de --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:295e5ed8729a69fac8dc87b3a33b65d1f7c8603333d040f4041194e746745e89 +size 561667 diff --git a/vlm/dev/9h3KsOVXhLZ/6.png b/vlm/dev/9h3KsOVXhLZ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..bab0f12f0222feb207612c84a997e782602a82b7 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71f5966411bd0daad2dc62220f61df14d19ac3174daf506f55565898ee644077 +size 531859 diff --git a/vlm/dev/9h3KsOVXhLZ/7.png b/vlm/dev/9h3KsOVXhLZ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e213a48748c23ffa3cdcc30cd178cda62fcbc83a --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b07297f49d98e47fc15e3e8d67e600d93875f315e3de006f29b6446e90998a5f +size 550814 diff --git a/vlm/dev/9h3KsOVXhLZ/8.png b/vlm/dev/9h3KsOVXhLZ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..5e93ed517c2c395f2c78b17ac322b76db2891afc --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64895223491af1f87c1f6e97f877f0a4bb76776965b30b8452dca790a5a2cdf0 +size 513206 diff --git a/vlm/dev/9h3KsOVXhLZ/9.png b/vlm/dev/9h3KsOVXhLZ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..072f82f9030c7c54dc3fcf438bc350a520c8f2c8 --- /dev/null +++ b/vlm/dev/9h3KsOVXhLZ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25adb67dda9ba5ea5f620492d0633a86187b2ba68f5f16fde602733dbfad6c8d +size 493096 diff --git a/vlm/dev/CQsmMYmlP5T/10.png b/vlm/dev/CQsmMYmlP5T/10.png new file mode 100644 index 0000000000000000000000000000000000000000..21578523b22e21a94f6b7417d3f5d8e406ab07d0 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35819ffa9ae703ffa71402e59071ec605ac526d019ff52d3f03e1e6dc13f8faf +size 610548 diff --git a/vlm/dev/CQsmMYmlP5T/11.png b/vlm/dev/CQsmMYmlP5T/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5abbdea0a393708b76ff387957650f5fdb2f4d1a --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56c82145bf213d03c21d410b585299604cc13354bb1cf781834b5084db32bdb2 +size 655937 diff --git a/vlm/dev/CQsmMYmlP5T/14.png b/vlm/dev/CQsmMYmlP5T/14.png new file mode 100644 index 0000000000000000000000000000000000000000..25ed7180b0a72974605ee584d79ee8f77096f99a --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bba4d3ec785e9e74b88934440555329fde76dc70bbd88781688fb836b85c8336 +size 638539 diff --git a/vlm/dev/CQsmMYmlP5T/15.png b/vlm/dev/CQsmMYmlP5T/15.png new file mode 100644 index 0000000000000000000000000000000000000000..0b66fcd594d698896bc614a8e8a2ff6347e5182c --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73ca758339e512f39c0062d68e2ef0a00c3348db6bf28fbbaca11a6b62eb33c9 +size 620567 diff --git a/vlm/dev/CQsmMYmlP5T/17.png b/vlm/dev/CQsmMYmlP5T/17.png new file mode 100644 index 0000000000000000000000000000000000000000..98dc03b4d49c22aa0e40943f0e35be27a947395b --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4cbe71a13695158cb11d27a039e9401900345fd608dee7f75be2363c33ba2a3 +size 302311 diff --git a/vlm/dev/CQsmMYmlP5T/2.png b/vlm/dev/CQsmMYmlP5T/2.png new file mode 100644 index 0000000000000000000000000000000000000000..1844180b15de0898c02d61da0d7432022685f877 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cde9b5c060ec876fd55f9292e1d0329f8d0c5d272661d63bf7bb880204dd00a +size 606787 diff --git a/vlm/dev/CQsmMYmlP5T/21.png b/vlm/dev/CQsmMYmlP5T/21.png new file mode 100644 index 0000000000000000000000000000000000000000..31124c4d9a2bda54e2018e422020e70ffc0c1136 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c511ab04b0b0d26521a50218ea420641b604f8a7caa42e4cb12747aca145606 +size 580702 diff --git a/vlm/dev/CQsmMYmlP5T/23.png b/vlm/dev/CQsmMYmlP5T/23.png new file mode 100644 index 0000000000000000000000000000000000000000..0bc3f997e2e95d1c6b2289375c3962a59a550a03 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b83f47dee508ee927b7b2a12e9e313aabf3c100c53dd9e752dbc0106454bd381 +size 417234 diff --git a/vlm/dev/CQsmMYmlP5T/24.png b/vlm/dev/CQsmMYmlP5T/24.png new file mode 100644 index 0000000000000000000000000000000000000000..cbb2e59473ae5a27d497993ca32b64ecb76b08b2 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4805c8fac72db46829c6118cbfee16a2c25a7c2be72907f2cc734caf097f5c5 +size 371732 diff --git a/vlm/dev/CQsmMYmlP5T/3.png b/vlm/dev/CQsmMYmlP5T/3.png new file mode 100644 index 0000000000000000000000000000000000000000..0891a2739b2bf8bcf5499304c307dc52c4aaa1f4 --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af048b5c8062b25f9f5bc056d26926723f02f2886e3679a14351da65621de23d +size 531296 diff --git a/vlm/dev/CQsmMYmlP5T/6.png b/vlm/dev/CQsmMYmlP5T/6.png new file mode 100644 index 0000000000000000000000000000000000000000..1fe72ae14a415b8b5adacebb618b96ee6160214a --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d6f54099aeb38e728bbd2083258bb07948001c739d6e2616313ed268491f279 +size 529393 diff --git a/vlm/dev/CQsmMYmlP5T/9.png b/vlm/dev/CQsmMYmlP5T/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1f7dba41e351d9940c65f61c34ee2fc53e47c21b --- /dev/null +++ b/vlm/dev/CQsmMYmlP5T/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d6ad1e18956861ce4278b624f9734389c16393cc089a73252ac2f151604d825 +size 245193 diff --git a/vlm/dev/ECvgmYVyeUz/0.png b/vlm/dev/ECvgmYVyeUz/0.png new file mode 100644 index 0000000000000000000000000000000000000000..aeab4c10938db00afef99d699df09a4c4731450b --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b662c2ee43cdba9be38b0e61565817b4311494446f10507e31b1027db7d7b95 +size 541485 diff --git a/vlm/dev/ECvgmYVyeUz/1.png b/vlm/dev/ECvgmYVyeUz/1.png new file mode 100644 index 0000000000000000000000000000000000000000..2925aca7f888a50da379a9638098f1fc37a1637a --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d86a50b8d0be55a3435c765cf9c063a3e3d1d2aeaf218d9c9b48407b566be67a +size 786275 diff --git a/vlm/dev/ECvgmYVyeUz/10.png b/vlm/dev/ECvgmYVyeUz/10.png new file mode 100644 index 0000000000000000000000000000000000000000..0ae95e7a416f73edc1052f93d5badc25efad61fa --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afcc54def8c01c6d6771572c0dbda472cadf88ea11c874c6fd5d0928326546f4 +size 312068 diff --git a/vlm/dev/ECvgmYVyeUz/11.png b/vlm/dev/ECvgmYVyeUz/11.png new file mode 100644 index 0000000000000000000000000000000000000000..10bd02b502be429ee92582474be1202f0aae3139 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ce2ab37056d7e9e3fa4d55765c06b9446ae275022045711b790e1c97f62ebf6 +size 379459 diff --git a/vlm/dev/ECvgmYVyeUz/12.png b/vlm/dev/ECvgmYVyeUz/12.png new file mode 100644 index 0000000000000000000000000000000000000000..98faed2878e22038a984d6a5eb1397000b6eaa34 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6195107bc31420bd1219f8acb67afbe6e235e2da81caa823cda0554fac9a376f +size 391063 diff --git a/vlm/dev/ECvgmYVyeUz/13.png b/vlm/dev/ECvgmYVyeUz/13.png new file mode 100644 index 0000000000000000000000000000000000000000..0f151c024ec1ed1f9c069593d48863c032af44ee --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4555484cc1bd5cf506c1c953237d0f08e3e0b06b7a68f492475569e93b5e700a +size 362238 diff --git a/vlm/dev/ECvgmYVyeUz/14.png b/vlm/dev/ECvgmYVyeUz/14.png new file mode 100644 index 0000000000000000000000000000000000000000..c32f9ea689bdeffe4c42d389424b2b50b172b4ab --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:419995c241b121d417dc35fed94bf1d7e6e172a8430ba2971fd51381768b8cd8 +size 416745 diff --git a/vlm/dev/ECvgmYVyeUz/15.png b/vlm/dev/ECvgmYVyeUz/15.png new file mode 100644 index 0000000000000000000000000000000000000000..cabf0776997f85f563bdc0d2ea166414beff9d58 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63369deb14ebe62fd003542a36cb9cac5a9b26755c3344f497aa5cad1cb362c2 +size 486448 diff --git a/vlm/dev/ECvgmYVyeUz/16.png b/vlm/dev/ECvgmYVyeUz/16.png new file mode 100644 index 0000000000000000000000000000000000000000..fa070f5f2b57ccf1e9593ef681bc35daaac153aa --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:568fa64aeb2314da8b43dca87e45506ecaa731ca9519b51ee7842e2a291dad95 +size 416673 diff --git a/vlm/dev/ECvgmYVyeUz/17.png b/vlm/dev/ECvgmYVyeUz/17.png new file mode 100644 index 0000000000000000000000000000000000000000..579a05603e8abbb8c62d44c44b8de55cf0eb3d72 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2cbeb8f45674351019628024b436bf25747479e35056a33d474672f915eda16 +size 376249 diff --git a/vlm/dev/ECvgmYVyeUz/18.png b/vlm/dev/ECvgmYVyeUz/18.png new file mode 100644 index 0000000000000000000000000000000000000000..96d1a2521fed0058cb8a01a892dbb09ca05a5662 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6546f0decee590e029c5d75efd3327f0faa9dc77a0e37a377b5090d288dd9657 +size 671849 diff --git a/vlm/dev/ECvgmYVyeUz/19.png b/vlm/dev/ECvgmYVyeUz/19.png new file mode 100644 index 0000000000000000000000000000000000000000..2fd214b057235a3952996facf518ef514aebb491 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcd3216392c1e85e04f8b2625b1c3714f5f3537511dd3ff114f209fee90f7395 +size 541819 diff --git a/vlm/dev/ECvgmYVyeUz/2.png b/vlm/dev/ECvgmYVyeUz/2.png new file mode 100644 index 0000000000000000000000000000000000000000..4ba1eb92fcbfeaf45fa1fc5a9a0f222f827d19e4 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3813508d72ff943ce1a5b529885290475d2235fad9295085f30ea41299cb2539 +size 601821 diff --git a/vlm/dev/ECvgmYVyeUz/20.png b/vlm/dev/ECvgmYVyeUz/20.png new file mode 100644 index 0000000000000000000000000000000000000000..ff2ac862d4666f444567f374de00d659b0c77ce1 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fed1c1032df484ec795756ba4be7f2fab1aa310fe07bba4c618cfad9415889f4 +size 453403 diff --git a/vlm/dev/ECvgmYVyeUz/3.png b/vlm/dev/ECvgmYVyeUz/3.png new file mode 100644 index 0000000000000000000000000000000000000000..f5f4dae66e0aae7428edc4c289ae6958f6391e3d --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62416afdb478464e8d8b1d41748f63f7b98cff15e7ded046641325c28321aff4 +size 585343 diff --git a/vlm/dev/ECvgmYVyeUz/4.png b/vlm/dev/ECvgmYVyeUz/4.png new file mode 100644 index 0000000000000000000000000000000000000000..112e881be50b9bfde1a4c19b03990232f8301cd5 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8052adb273fc9bb53b3a508dc283c89d21da3f231384f8663fa734932c094441 +size 557715 diff --git a/vlm/dev/ECvgmYVyeUz/5.png b/vlm/dev/ECvgmYVyeUz/5.png new file mode 100644 index 0000000000000000000000000000000000000000..e1ee5e086b0dc2829b543c4e9f24b290b040010a --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:daed9a4408b04bfbe2d664ef85572fec2fa0555bd1bad08130f470f310b446dc +size 624785 diff --git a/vlm/dev/ECvgmYVyeUz/6.png b/vlm/dev/ECvgmYVyeUz/6.png new file mode 100644 index 0000000000000000000000000000000000000000..01cb55bc247d1b94fdd9838138ecdce2244706ca --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d06c20a415dc9b871f8f26a6308b78932d70eac7106d7bdb2827012573a74cb +size 645501 diff --git a/vlm/dev/ECvgmYVyeUz/7.png b/vlm/dev/ECvgmYVyeUz/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8cd543f1249c04640ab9899df5a5b67ecf9ba076 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc0740d7c4d2b0c354d05dd06606795b8b024717a73508f8a26a1f0040142fb1 +size 587168 diff --git a/vlm/dev/ECvgmYVyeUz/8.png b/vlm/dev/ECvgmYVyeUz/8.png new file mode 100644 index 0000000000000000000000000000000000000000..24c4fba446fc5768834d3975731bab0726f04745 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2d30c3a248bd27b5f49510d6b63fd025502f54726efef9a267df8a186730049 +size 534900 diff --git a/vlm/dev/ECvgmYVyeUz/9.png b/vlm/dev/ECvgmYVyeUz/9.png new file mode 100644 index 0000000000000000000000000000000000000000..61565121789f48cc0b49edcfd07b9e82bf10bd96 --- /dev/null +++ b/vlm/dev/ECvgmYVyeUz/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ff541557f5e299868e2df0b7d3a6c4f3982720d4c7e51ae9210b4461bb02935 +size 539942 diff --git a/vlm/dev/FWMQYjFso-a/0.png b/vlm/dev/FWMQYjFso-a/0.png new file mode 100644 index 0000000000000000000000000000000000000000..abefb67170c83ec627d19739a5b9e379e2e49296 --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4d9a6764f65c394e855d58c0402a99f45e523f9b3fba9ef653caae4d65ff0a6 +size 440329 diff --git a/vlm/dev/FWMQYjFso-a/1.png b/vlm/dev/FWMQYjFso-a/1.png new file mode 100644 index 0000000000000000000000000000000000000000..c4c06967f8433e28abd40a87a0fcfe47d0ce8c6b --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfc32508c4e2b5697cce9759dd0c0b816f69a0779203493a0f7f774c2ca277f2 +size 646764 diff --git a/vlm/dev/FWMQYjFso-a/10.png b/vlm/dev/FWMQYjFso-a/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2c921bc2b6d264147298b3dda99756c88b893c63 --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d94add13ce58458c107ddd8b37c13d9a12580416fa4d4138f32a0b98e388a71 +size 507290 diff --git a/vlm/dev/FWMQYjFso-a/11.png b/vlm/dev/FWMQYjFso-a/11.png new file mode 100644 index 0000000000000000000000000000000000000000..e02dedfe4cbe966fd61b0d004dc79a373e2b379a --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:165c998f78dfe925513b8bba3d7a53b9752c2f5b32c1c059e19ddefdaf0f1ca2 +size 541789 diff --git a/vlm/dev/FWMQYjFso-a/12.png b/vlm/dev/FWMQYjFso-a/12.png new file mode 100644 index 0000000000000000000000000000000000000000..3c4130fe775a19bfe6687637a4697c45d6e382ce --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28834385d1f278d3248cf539feb85b6e44b1c6de141499b8282f72bb03ed3a20 +size 405088 diff --git a/vlm/dev/FWMQYjFso-a/13.png b/vlm/dev/FWMQYjFso-a/13.png new file mode 100644 index 0000000000000000000000000000000000000000..b3dbeb11554096d955bb4baf7d86595dbcf19a24 --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35c93f0edf0535c00cd7c668d90734a6c6972fcdbc719db9de71b88aff04dc58 +size 336551 diff --git a/vlm/dev/FWMQYjFso-a/2.png b/vlm/dev/FWMQYjFso-a/2.png new file mode 100644 index 0000000000000000000000000000000000000000..bb8a55ecf081b6c409694efe7b7214a24a959db4 --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:230af9f088abd6688174924f2ee2b20a834fbf3e374bcd96fbf84ee95613c790 +size 549574 diff --git a/vlm/dev/FWMQYjFso-a/3.png b/vlm/dev/FWMQYjFso-a/3.png new file mode 100644 index 0000000000000000000000000000000000000000..7f791045131513a3e414547ef20e1e3069027235 --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a970052857b544d27f0404758aa4a97095fc2fefe12bdc7553188c3397ac7ccc +size 541998 diff --git a/vlm/dev/FWMQYjFso-a/4.png b/vlm/dev/FWMQYjFso-a/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4d7577d2f4b8c5e0de834bd1fe7f40284ce034b8 --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:891d3b466912d93633f12d81d0acf4f7aa470ff6b7c45b6a6451d2e332c15b3c +size 683304 diff --git a/vlm/dev/FWMQYjFso-a/5.png b/vlm/dev/FWMQYjFso-a/5.png new file mode 100644 index 0000000000000000000000000000000000000000..526ac5b50e4901093dbc2c4bc8a34508796177eb --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edad8ae87dd0f68c9b95249c00e9efe15ccc40666c65ca5eb7d6f9758ebcd33f +size 486648 diff --git a/vlm/dev/FWMQYjFso-a/6.png b/vlm/dev/FWMQYjFso-a/6.png new file mode 100644 index 0000000000000000000000000000000000000000..06b84b610f05e69bdc3792fcba4d84141b173ed9 --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd95663ab08dc724f83abcbe423091a8783f443c2aa12da091176e2c15cd61da +size 587757 diff --git a/vlm/dev/FWMQYjFso-a/7.png b/vlm/dev/FWMQYjFso-a/7.png new file mode 100644 index 0000000000000000000000000000000000000000..cbfdd6f7313b23d41bb7a17b9316a76497f9f0ab --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5b6090d1107cfc54d54e505a199c49c6c003dd29be811eb2c7e6e3cc89c8310 +size 625130 diff --git a/vlm/dev/FWMQYjFso-a/8.png b/vlm/dev/FWMQYjFso-a/8.png new file mode 100644 index 0000000000000000000000000000000000000000..de1d488e1e7db42aab7c9fa342ab0c31c81c0bff --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf8ff7952bcc124cea291578fdaaefba63704b3f3770bcf4bdd68dc5ea7765fb +size 580563 diff --git a/vlm/dev/FWMQYjFso-a/9.png b/vlm/dev/FWMQYjFso-a/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e5c6f09ef23432aae54f5060fcb23c07806bbfea --- /dev/null +++ b/vlm/dev/FWMQYjFso-a/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ef49741f7f5fc50eb6c83074fa5bc95deea7f4c9551358b9368d82569fd1089 +size 1161303 diff --git a/vlm/dev/IajGRJuM7D3/0.png b/vlm/dev/IajGRJuM7D3/0.png new file mode 100644 index 0000000000000000000000000000000000000000..2ba21a929fd1d157414df7eae39a55ec1bbea5bb --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83611cb6fc7169290003b79b97ec23225e17409f52a36864395d28b1402971c3 +size 600052 diff --git a/vlm/dev/IajGRJuM7D3/1.png b/vlm/dev/IajGRJuM7D3/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a12086d5b7276903806dd3f65079f881fe94e410 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fed235bf4490a13d98a42599dc3a20b79dd9e24f49bb831f4e283067776d687 +size 693711 diff --git a/vlm/dev/IajGRJuM7D3/10.png b/vlm/dev/IajGRJuM7D3/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d7475763b007f87758d03a3dc2c3ea1ad4f3b4e1 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92b19f525b6bc0151b7c8ad2609a3001f8d21c2a202d8408bcb56ab96d2d3292 +size 581991 diff --git a/vlm/dev/IajGRJuM7D3/11.png b/vlm/dev/IajGRJuM7D3/11.png new file mode 100644 index 0000000000000000000000000000000000000000..b8d4c5a34dd21a86db95dbdc57bb2221026ef52e --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a381a991e898f0e3c4a667f05aed97c700cad61b950c7594813a90acc83be23 +size 553847 diff --git a/vlm/dev/IajGRJuM7D3/12.png b/vlm/dev/IajGRJuM7D3/12.png new file mode 100644 index 0000000000000000000000000000000000000000..aa6eb136cd315061a61ed5c134c12d15de134e03 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30cef0515869531b449d33669b3ac433d0b278edb8b767e4a42ef734bc1405ae +size 567473 diff --git a/vlm/dev/IajGRJuM7D3/13.png b/vlm/dev/IajGRJuM7D3/13.png new file mode 100644 index 0000000000000000000000000000000000000000..2955a006322275cb0a1d0d1dc8682f187e5eec0e --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e0833c6027b2f8a02fddad2c4f25c9e63df6b134bc5ac9182ef32e9c48f5408 +size 554449 diff --git a/vlm/dev/IajGRJuM7D3/14.png b/vlm/dev/IajGRJuM7D3/14.png new file mode 100644 index 0000000000000000000000000000000000000000..71069ab1dc56f3c6eee6bc123a2c2c9f667ecfa0 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e40613561ea3c653a6b5d679567f89012a8909b97fe95d14d4231004d137c6af +size 488855 diff --git a/vlm/dev/IajGRJuM7D3/15.png b/vlm/dev/IajGRJuM7D3/15.png new file mode 100644 index 0000000000000000000000000000000000000000..f8350db72d65f035463c5615523b1a808f30fb11 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:385c754ec45a9720282c74ef2e92d3e5aebd696f3617b182e8587209284156be +size 294626 diff --git a/vlm/dev/IajGRJuM7D3/16.png b/vlm/dev/IajGRJuM7D3/16.png new file mode 100644 index 0000000000000000000000000000000000000000..f3d941a8c1df478a7498994a53eab4affee3ad99 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b93bfef7f61ad90257363564d9e4ba002b6af1e0732f748e12ceb0dbec86a279 +size 381209 diff --git a/vlm/dev/IajGRJuM7D3/17.png b/vlm/dev/IajGRJuM7D3/17.png new file mode 100644 index 0000000000000000000000000000000000000000..fa55c9cf136e938ca9eca376fda5587d49e9b6e4 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37105d61f7584c2fcdbb5ba9504168843ef6d1d336650f339d55330404b13948 +size 354503 diff --git a/vlm/dev/IajGRJuM7D3/18.png b/vlm/dev/IajGRJuM7D3/18.png new file mode 100644 index 0000000000000000000000000000000000000000..825848ce4fded95fb2203df965c55f541f9e504f --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9381d95537b9a383a17d8ab14aecaec4855bfff8b070ae1dbe0d4a98c0e4daf1 +size 426787 diff --git a/vlm/dev/IajGRJuM7D3/19.png b/vlm/dev/IajGRJuM7D3/19.png new file mode 100644 index 0000000000000000000000000000000000000000..34fd27707b2fbae496b4449330dfcc8a20f9409c --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:470895fb0796ebc1fad6747c7f44b2ce5504753d7c3b163625c7aa097c8afebc +size 426926 diff --git a/vlm/dev/IajGRJuM7D3/2.png b/vlm/dev/IajGRJuM7D3/2.png new file mode 100644 index 0000000000000000000000000000000000000000..d06a1885183136c287d92ddd5b006aa845f0770f --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d3720b4b9cb9b16fa88d581a18f649ff98b06fd509f3886eeab9b189268c11f +size 617321 diff --git a/vlm/dev/IajGRJuM7D3/20.png b/vlm/dev/IajGRJuM7D3/20.png new file mode 100644 index 0000000000000000000000000000000000000000..1e38a5ce66c7de9fa659c940faebeab868741aa9 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8e997fb8ffada540d51ca63a9d21dafc57433949aacfeef3c86b8ebdbc8bed7 +size 406493 diff --git a/vlm/dev/IajGRJuM7D3/21.png b/vlm/dev/IajGRJuM7D3/21.png new file mode 100644 index 0000000000000000000000000000000000000000..cc0be26f44b7392441db86c5f80064c0b2182b5a --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f68da1e169ae6f1a0e1d04095dba1bc0b2ae4c4dd9216011acaf69d94cd4cdd8 +size 320898 diff --git a/vlm/dev/IajGRJuM7D3/22.png b/vlm/dev/IajGRJuM7D3/22.png new file mode 100644 index 0000000000000000000000000000000000000000..bf94713d448ae084f15ebc3a8bf9e92efba2251c --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50c5e9561a1d9613eba1bf8e93743556779c49a340d65e536d72bc6dd3a2b2ce +size 383011 diff --git a/vlm/dev/IajGRJuM7D3/23.png b/vlm/dev/IajGRJuM7D3/23.png new file mode 100644 index 0000000000000000000000000000000000000000..8b05f504e467a15b42d5beebc42d90b9ced34cdb --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9854c7e0c89fb83fbd277b02a1307cd5a8189df37770a727079646cb7e266922 +size 358972 diff --git a/vlm/dev/IajGRJuM7D3/24.png b/vlm/dev/IajGRJuM7D3/24.png new file mode 100644 index 0000000000000000000000000000000000000000..9f7f656a00246afee7f3ab6ea8ce8f5e79cb8414 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dd7a396be3e4af30f3e4699f8e09e90310a1823b01ba19e9b02dc7313c19fbe +size 285974 diff --git a/vlm/dev/IajGRJuM7D3/25.png b/vlm/dev/IajGRJuM7D3/25.png new file mode 100644 index 0000000000000000000000000000000000000000..93e745c6a27b5421bd189545e48769d26082a9b7 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6c59d1f49dfd406d32bee9f6d4a77fe4fdb4913214bb7fd94dd5e7edaf59bcd +size 344642 diff --git a/vlm/dev/IajGRJuM7D3/26.png b/vlm/dev/IajGRJuM7D3/26.png new file mode 100644 index 0000000000000000000000000000000000000000..f6ed43e59d1a905e44dbac367243f326c384813e --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06fa3220b10e988900ad0515bbea6fb167d342561a694a09d4c8712d44e0ff81 +size 486161 diff --git a/vlm/dev/IajGRJuM7D3/27.png b/vlm/dev/IajGRJuM7D3/27.png new file mode 100644 index 0000000000000000000000000000000000000000..53ed886ad803ead09628048b14019eba5197fa1c --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c39dc5e7b8d849292acdf4973fd2e73bca328cae602c2e6e6d0a9e1fd3aa6fb +size 427426 diff --git a/vlm/dev/IajGRJuM7D3/28.png b/vlm/dev/IajGRJuM7D3/28.png new file mode 100644 index 0000000000000000000000000000000000000000..3249edb4327f8740497cb87ffa2e35514d556d16 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9e1345a373ecc7ba74546e944804f9cdda274a8964e09cc06b216e641945130 +size 476608 diff --git a/vlm/dev/IajGRJuM7D3/29.png b/vlm/dev/IajGRJuM7D3/29.png new file mode 100644 index 0000000000000000000000000000000000000000..4faf15df900aecb9e34ee449348a894391a56ec2 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:daface36d016db49995d1808bb4c798f8c57b17e37510b211058ce1be39ea903 +size 508026 diff --git a/vlm/dev/IajGRJuM7D3/3.png b/vlm/dev/IajGRJuM7D3/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b441726fdd152e805e050f4ddd6154d89ecb5a7e --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b87482823781167ce03dd6824bf50dae5d1c987462d63b80a8d5e2094556084c +size 570996 diff --git a/vlm/dev/IajGRJuM7D3/30.png b/vlm/dev/IajGRJuM7D3/30.png new file mode 100644 index 0000000000000000000000000000000000000000..26cf196d5d2bba5b9b8881ba138097fc13a07bd6 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/30.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01bee97dc8932cef0391ba27842a3e40e332289ae042ba2b87a831f542b9c9ae +size 419187 diff --git a/vlm/dev/IajGRJuM7D3/4.png b/vlm/dev/IajGRJuM7D3/4.png new file mode 100644 index 0000000000000000000000000000000000000000..2b42d209fe230767d9ef0742f1911290d3a8320c --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:117783ad6aaba86b158766e5128d149a047bd2c46b1992951fcd85575fed357c +size 574904 diff --git a/vlm/dev/IajGRJuM7D3/5.png b/vlm/dev/IajGRJuM7D3/5.png new file mode 100644 index 0000000000000000000000000000000000000000..8c589e3cd61e876dbb158c74f72d9431cb09f30d --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e143eb1c588d3ea2835acbc7dcc1ce9c01c2daa51f6e40acaaf5382bab0441c +size 597306 diff --git a/vlm/dev/IajGRJuM7D3/6.png b/vlm/dev/IajGRJuM7D3/6.png new file mode 100644 index 0000000000000000000000000000000000000000..1db4b5d720e16b623a88ce8e699a32bef7675a0a --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddfe18ff965ed79a00e4ba2bae977c860898333169c85c75f42cae0ae46db77b +size 649525 diff --git a/vlm/dev/IajGRJuM7D3/7.png b/vlm/dev/IajGRJuM7D3/7.png new file mode 100644 index 0000000000000000000000000000000000000000..7c258ac8519dbc1b12b8184668f92f7a9f7d96ac --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:044af5ae57e6f21be6d35371d8788c6af096972dc6b7d313ea0e8d80acb3cabf +size 642479 diff --git a/vlm/dev/IajGRJuM7D3/8.png b/vlm/dev/IajGRJuM7D3/8.png new file mode 100644 index 0000000000000000000000000000000000000000..9b0f2669ed3e169264126a2b5111ee7a43a82b18 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:916cc27c564642d4a7f9d5ac1b72b24f9ccba9006a618ec3d9f0e73468430673 +size 659528 diff --git a/vlm/dev/IajGRJuM7D3/9.png b/vlm/dev/IajGRJuM7D3/9.png new file mode 100644 index 0000000000000000000000000000000000000000..b1aeadb3fd63c4a47d4ba4d0bea5c5c760518fc1 --- /dev/null +++ b/vlm/dev/IajGRJuM7D3/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c1fd753692ceb78466df32b8a81835aeebe4321cc530639387d53dc66ea4721 +size 530266 diff --git a/vlm/dev/Ih2bG6h1r4S/0.png b/vlm/dev/Ih2bG6h1r4S/0.png new file mode 100644 index 0000000000000000000000000000000000000000..d80be410cabc2963e438ef05d8c253b27b0e7eec --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0857829102f18b483e2508b9f77521e9aa1fe86aabe0b5d8ed125cf7cdc8794d +size 336465 diff --git a/vlm/dev/Ih2bG6h1r4S/1.png b/vlm/dev/Ih2bG6h1r4S/1.png new file mode 100644 index 0000000000000000000000000000000000000000..7971993fb62050d67ab559035fb356b03ac95bce --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6e28fa95a2645a44fc15de9b8ff9f50b976dfbf478abb824267af1ddfc43650 +size 606293 diff --git a/vlm/dev/Ih2bG6h1r4S/10.png b/vlm/dev/Ih2bG6h1r4S/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c6f5e5938cd207ad0561e6d76b59b9678a35f329 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a286a8d8e7a50da9e67c1292546a358d514bbe78bf3d90e51919c923b8096810 +size 519502 diff --git a/vlm/dev/Ih2bG6h1r4S/11.png b/vlm/dev/Ih2bG6h1r4S/11.png new file mode 100644 index 0000000000000000000000000000000000000000..cd662acaeb7a6542daa8d8c586148edb4dce833b --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d04cc246c9469e1c3b66cf20802de84a10fc061e9bd69c05b5a6d79e51a15a98 +size 207911 diff --git a/vlm/dev/Ih2bG6h1r4S/2.png b/vlm/dev/Ih2bG6h1r4S/2.png new file mode 100644 index 0000000000000000000000000000000000000000..803e84e89c54563e22faecfa6e4e862d15fcd469 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0dd2a424fd020c2a6fee036d3001a1a0511a4a4a3dac587b5467bbb52276956 +size 408104 diff --git a/vlm/dev/Ih2bG6h1r4S/3.png b/vlm/dev/Ih2bG6h1r4S/3.png new file mode 100644 index 0000000000000000000000000000000000000000..ad22a5ccdc9bf54096ea4a699a7e5a334d2063e3 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d76efbe5872dd4306c11abfdb6f93e8dab6caa7b03c2e5aac2074d2ee638c43 +size 506349 diff --git a/vlm/dev/Ih2bG6h1r4S/4.png b/vlm/dev/Ih2bG6h1r4S/4.png new file mode 100644 index 0000000000000000000000000000000000000000..fd6d7d856c4a779c7cc6c0dce4d000670240cc41 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16780f8fa1ecd041be8454115885642b8c00c5b64ce0d388406f2ae9bfb9412e +size 408747 diff --git a/vlm/dev/Ih2bG6h1r4S/5.png b/vlm/dev/Ih2bG6h1r4S/5.png new file mode 100644 index 0000000000000000000000000000000000000000..ab205fe04c51bd8df68aa9dfbc87292667ae6836 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00f45d14ae4cdfb3c0ece6bc8f63aeebcd47a86ec8bfb545b7f6a42beb7f7ff6 +size 479704 diff --git a/vlm/dev/Ih2bG6h1r4S/6.png b/vlm/dev/Ih2bG6h1r4S/6.png new file mode 100644 index 0000000000000000000000000000000000000000..c150b4638013943929e46b933da6161928db0367 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac211e7cea8f8d4cfeea62939002d89ddf49584a961bf4912126cf2fc15289d7 +size 578781 diff --git a/vlm/dev/Ih2bG6h1r4S/7.png b/vlm/dev/Ih2bG6h1r4S/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c262e1902d48b597a055a4bff251f12df361d355 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6453d498bc87e50d734ae96ab8fece6750407e4fe5230425f2907e7cf4b01c94 +size 614723 diff --git a/vlm/dev/Ih2bG6h1r4S/8.png b/vlm/dev/Ih2bG6h1r4S/8.png new file mode 100644 index 0000000000000000000000000000000000000000..fdf3021e82ef08e53961ae8cf2c5175e4f4a34d7 --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efbff181b538d91d7a8c648192e3ea7f9950aea98fc519308a1b71dab90d27e2 +size 391883 diff --git a/vlm/dev/Ih2bG6h1r4S/9.png b/vlm/dev/Ih2bG6h1r4S/9.png new file mode 100644 index 0000000000000000000000000000000000000000..de11ad71aec05c3ed6631516323acd7e1a43205e --- /dev/null +++ b/vlm/dev/Ih2bG6h1r4S/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3a4ab840657ea0a3caccb0ab8a520a047711a7e0e77e9273139ed1ab8fd2905 +size 574441 diff --git a/vlm/dev/JavFPcsscd5/0.png b/vlm/dev/JavFPcsscd5/0.png new file mode 100644 index 0000000000000000000000000000000000000000..5f517aeb9e8fc06945017b1c80673fcf9a822f94 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed352a4681750645ca403e531336a6d575c8d5f6aaeba906c4addfdf8d0e5cf2 +size 456215 diff --git a/vlm/dev/JavFPcsscd5/1.png b/vlm/dev/JavFPcsscd5/1.png new file mode 100644 index 0000000000000000000000000000000000000000..416a492d20f915a15da91a69e527881aeed1e3dc --- /dev/null +++ b/vlm/dev/JavFPcsscd5/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f99d0f24ca02b4c08d0a9cff418a95786c270d3efaa42036389205f4650ce01 +size 596842 diff --git a/vlm/dev/JavFPcsscd5/10.png b/vlm/dev/JavFPcsscd5/10.png new file mode 100644 index 0000000000000000000000000000000000000000..cdecc1a8a2f027198c119b15c4a75d41383050d4 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05b5ac02854de9a26e8ac8b4ef822660bc8f9707eb139c8629116493d3dc2a2d +size 547361 diff --git a/vlm/dev/JavFPcsscd5/11.png b/vlm/dev/JavFPcsscd5/11.png new file mode 100644 index 0000000000000000000000000000000000000000..364cf5bc815b56b5bd4a119bbe07e62340bd9e97 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e44720711589e5da95b0387e9d7acc0a216e13f4dc5b4045ad208ce95c7e105e +size 538893 diff --git a/vlm/dev/JavFPcsscd5/12.png b/vlm/dev/JavFPcsscd5/12.png new file mode 100644 index 0000000000000000000000000000000000000000..6093b26d33f32f89dccb83eaf9f3b6cb0f3b3e14 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be2a26d7f8fdea2e2b55f8683e409b29cab0a3c8ec8dcf3d23498e5cdcc427c2 +size 495415 diff --git a/vlm/dev/JavFPcsscd5/13.png b/vlm/dev/JavFPcsscd5/13.png new file mode 100644 index 0000000000000000000000000000000000000000..a1edf1543710ba90df116e72d878ae66612e24bd --- /dev/null +++ b/vlm/dev/JavFPcsscd5/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:586ccba71b239f20cf165fc14a728d157e8f037d461b1340510e809256140425 +size 105309 diff --git a/vlm/dev/JavFPcsscd5/2.png b/vlm/dev/JavFPcsscd5/2.png new file mode 100644 index 0000000000000000000000000000000000000000..5ef3e1d8d043a6fe7e4df1ffcd930f1f64178706 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f7472268e5e74b51c4b00e6eff412493f71813bddafb512c6bcd0caa45a17c4 +size 598286 diff --git a/vlm/dev/JavFPcsscd5/3.png b/vlm/dev/JavFPcsscd5/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b682052d16fca0f3bea450bc388cfc57a4f2caf8 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:872f560492d5a2201a0380b011e84adae2f34587017a28b1c4db4bdc736677fb +size 642361 diff --git a/vlm/dev/JavFPcsscd5/4.png b/vlm/dev/JavFPcsscd5/4.png new file mode 100644 index 0000000000000000000000000000000000000000..bcffde0f3ebc5385a21bedf25dfb1ad93e6139de --- /dev/null +++ b/vlm/dev/JavFPcsscd5/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41e07627423c9cf6529892ed202ab42931b22ad851b2b60c7a2d46bf6c6028b3 +size 573324 diff --git a/vlm/dev/JavFPcsscd5/5.png b/vlm/dev/JavFPcsscd5/5.png new file mode 100644 index 0000000000000000000000000000000000000000..00ff73118436bba3909b56821dbdf929cc0edea7 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3923b5fbb87dfcfeb018a37a0bbe45cc6cc57139cd52ce1202b5122d91ef687 +size 583762 diff --git a/vlm/dev/JavFPcsscd5/6.png b/vlm/dev/JavFPcsscd5/6.png new file mode 100644 index 0000000000000000000000000000000000000000..036a7cda2001af4709a5e7e70ba0f138b241d0ae --- /dev/null +++ b/vlm/dev/JavFPcsscd5/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc2d429aa3fdd8aa4bc708fc045a841095c07d4595a4c64816ccd2b7a7aad0b +size 1015066 diff --git a/vlm/dev/JavFPcsscd5/7.png b/vlm/dev/JavFPcsscd5/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8c41c3c9fc3e99339a70edb359836a60a1fe59af --- /dev/null +++ b/vlm/dev/JavFPcsscd5/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ff858c95ec6c0b813ff0f299a53a55a8f8bc4d317527c1c1734aac756188596 +size 715602 diff --git a/vlm/dev/JavFPcsscd5/8.png b/vlm/dev/JavFPcsscd5/8.png new file mode 100644 index 0000000000000000000000000000000000000000..6f7bc52c63d9c6ec17e8b01035a5062b72945bba --- /dev/null +++ b/vlm/dev/JavFPcsscd5/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a66cc5724e744dc5dab1af97477309d23bfb65a098b3522528821246b29335b2 +size 549906 diff --git a/vlm/dev/JavFPcsscd5/9.png b/vlm/dev/JavFPcsscd5/9.png new file mode 100644 index 0000000000000000000000000000000000000000..19e7328f257e7c0149d68623c9b8406eca2a5674 --- /dev/null +++ b/vlm/dev/JavFPcsscd5/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c06c13d500e700624f199ba04e3d830a5e85166e778b04ff6bdefeb6d40ea3d5 +size 490309 diff --git a/vlm/dev/RKiWwhocuiU/0.png b/vlm/dev/RKiWwhocuiU/0.png new file mode 100644 index 0000000000000000000000000000000000000000..015a6305af83f14bceb15e44d6a0ca4897a5bd70 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0e5a7ab5f437296c6ccf80e5a336f83f0a858037d646991e6be68c4fdb6392e +size 509098 diff --git a/vlm/dev/RKiWwhocuiU/1.png b/vlm/dev/RKiWwhocuiU/1.png new file mode 100644 index 0000000000000000000000000000000000000000..7c958da0ecaf86a2e7c2d2fe39ab44eaa00a3015 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea7dd9fbb7215567030015a51ad4f1c9cb976cc6224b209a073afef0c836a93c +size 762171 diff --git a/vlm/dev/RKiWwhocuiU/10.png b/vlm/dev/RKiWwhocuiU/10.png new file mode 100644 index 0000000000000000000000000000000000000000..24f53ab49e0d7332d6c8733e84ac8dc668a046b7 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46e97b3477f52e4ee8ed722b54517a7402ab29efa5ffe15b07c410633da72186 +size 555942 diff --git a/vlm/dev/RKiWwhocuiU/11.png b/vlm/dev/RKiWwhocuiU/11.png new file mode 100644 index 0000000000000000000000000000000000000000..e2c0862dc9478993840718f9763cd8be5d8644cc --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fe4f7268c74643fe5a6be394462e17f0aff245a20a5f6ac9ca9de7be927dcaf +size 544634 diff --git a/vlm/dev/RKiWwhocuiU/12.png b/vlm/dev/RKiWwhocuiU/12.png new file mode 100644 index 0000000000000000000000000000000000000000..a5c985e81615621a756f5bc9f8fb7ae31569b7fa --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a67e6c82268fef8d80b552ef0a2e6ddfc58b613b1191db7824a5d756601ebf04 +size 595503 diff --git a/vlm/dev/RKiWwhocuiU/13.png b/vlm/dev/RKiWwhocuiU/13.png new file mode 100644 index 0000000000000000000000000000000000000000..22c994181046ae4fe8ecab48b1864dca45604307 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:485756b9936dc0956ea93f129a7f2fe0a5f1e2d208944f27bbc4df95abd3c820 +size 549305 diff --git a/vlm/dev/RKiWwhocuiU/14.png b/vlm/dev/RKiWwhocuiU/14.png new file mode 100644 index 0000000000000000000000000000000000000000..4abdaba2fd1316ee15fdedc20b66814479fcb6ee --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95ef69b574cf1de9860fc2149356e9bbb2911673671a879a96e96255c57278da +size 496889 diff --git a/vlm/dev/RKiWwhocuiU/15.png b/vlm/dev/RKiWwhocuiU/15.png new file mode 100644 index 0000000000000000000000000000000000000000..562a6100d4f28b7b7b3d00a516ec70f912939af7 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e478e548c28915c84bd8cf2aca859b92502fa168f8d01846324d1f12650beb7 +size 503642 diff --git a/vlm/dev/RKiWwhocuiU/16.png b/vlm/dev/RKiWwhocuiU/16.png new file mode 100644 index 0000000000000000000000000000000000000000..8f992b4a8e73fc932482badc2db346f733623529 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:560b626ae0844459aee18893a869985c791b2bb65b10d2f9137f5826dd8fbf60 +size 551371 diff --git a/vlm/dev/RKiWwhocuiU/17.png b/vlm/dev/RKiWwhocuiU/17.png new file mode 100644 index 0000000000000000000000000000000000000000..ec25cdb27fdaa49c93310c81d5c0d3585f627e07 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:191322a463410e6ffb870342d740048a651ab58fa181253ba5ebd38a4f2aa984 +size 341387 diff --git a/vlm/dev/RKiWwhocuiU/18.png b/vlm/dev/RKiWwhocuiU/18.png new file mode 100644 index 0000000000000000000000000000000000000000..a193791e44b9fd6888a64525faa426724cb1e875 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e24aad0f42aa2db2ac33ea72288c716f2cab0ce85cf200d5960866bfa8d2ceb9 +size 511103 diff --git a/vlm/dev/RKiWwhocuiU/19.png b/vlm/dev/RKiWwhocuiU/19.png new file mode 100644 index 0000000000000000000000000000000000000000..52c502c98f8e47eec7f90cf4cf008ed3c5932a19 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fa1dc7d913c104a32fdd89fced86af0c0252341ca8a104e3dc52aa6309fad58 +size 427152 diff --git a/vlm/dev/RKiWwhocuiU/2.png b/vlm/dev/RKiWwhocuiU/2.png new file mode 100644 index 0000000000000000000000000000000000000000..8854fa589bf63ae8bc921240220024d33e6fe5e3 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45899dca1c24d3c28ff0368f448b6c3bc54246ccf8aa30aee6e404d002538191 +size 569587 diff --git a/vlm/dev/RKiWwhocuiU/20.png b/vlm/dev/RKiWwhocuiU/20.png new file mode 100644 index 0000000000000000000000000000000000000000..b4fccc5079b5e032b596c5b75f24af5844c729b2 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:330803247822d2265268d2402d65523044689aa8dcf7f0abc54fe309b0be621a +size 36863 diff --git a/vlm/dev/RKiWwhocuiU/3.png b/vlm/dev/RKiWwhocuiU/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e93e7b4c15252bc6791cf9c8bef09ffe33b8629c --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb8fd6af6377b50e49bc9fb1f982c1c38ab525042c4933dadc1b9db3777712b3 +size 584301 diff --git a/vlm/dev/RKiWwhocuiU/4.png b/vlm/dev/RKiWwhocuiU/4.png new file mode 100644 index 0000000000000000000000000000000000000000..87c3dbeef772d4ad86c85dc1d8b8f28dd792906d --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5eb62177875ffb7895e8941113fc08921d7ad7f88ec45bd1ba78df387249b5c +size 419174 diff --git a/vlm/dev/RKiWwhocuiU/5.png b/vlm/dev/RKiWwhocuiU/5.png new file mode 100644 index 0000000000000000000000000000000000000000..66c2f70ee3996d8830288de9ac07ec54406e8cc3 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a01541b6d88ce302a334d38100cd45f224e9c1094647d2f9938bb2cba4ad02b +size 573163 diff --git a/vlm/dev/RKiWwhocuiU/6.png b/vlm/dev/RKiWwhocuiU/6.png new file mode 100644 index 0000000000000000000000000000000000000000..eba7c75092949e89c900a80bcf909131b8ec16b2 --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cd04b6b266cb1dec1df87e203b4d85b68c740bc5a0d923361e38b5d36816b07 +size 619181 diff --git a/vlm/dev/RKiWwhocuiU/7.png b/vlm/dev/RKiWwhocuiU/7.png new file mode 100644 index 0000000000000000000000000000000000000000..da29a3ea35cb9ce57e59c4ff84da13a287234adb --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5945331539f97b33f46c382911328bbdb0a3ea111112954b2580409db5f96cf6 +size 601064 diff --git a/vlm/dev/RKiWwhocuiU/8.png b/vlm/dev/RKiWwhocuiU/8.png new file mode 100644 index 0000000000000000000000000000000000000000..2919d52e96ecc1d2fa30dc5cd2bb749eba7509be --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d010e7fe757f35ae48432c24a4e94df5ba37e9fa2ac331c0025c765a6ebcca8 +size 612861 diff --git a/vlm/dev/RKiWwhocuiU/9.png b/vlm/dev/RKiWwhocuiU/9.png new file mode 100644 index 0000000000000000000000000000000000000000..499c5fd83c909ec3cc499b72d551cdf3ffafb67c --- /dev/null +++ b/vlm/dev/RKiWwhocuiU/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d6150bbc6f81f1b035e4f3b58dcb33f7a9f3422b48aa94ed5309e7b4cdb05bf +size 515503 diff --git a/vlm/dev/RecZ9nB9Q4/16.png b/vlm/dev/RecZ9nB9Q4/16.png new file mode 100644 index 0000000000000000000000000000000000000000..5a044adb97395da21c8192876c5e9286685a44cb --- /dev/null +++ b/vlm/dev/RecZ9nB9Q4/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e75a6ea8fd0e0b0a946805dfdab6f2a563b02484f063904c0f2cb4997d55a55a +size 208671 diff --git a/vlm/dev/RecZ9nB9Q4/19.png b/vlm/dev/RecZ9nB9Q4/19.png new file mode 100644 index 0000000000000000000000000000000000000000..505ecbe636691c6a638f584a8b4734a76b50c2df --- /dev/null +++ b/vlm/dev/RecZ9nB9Q4/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8c9b9c257f93c5125289146d6de76959dbc735b6175aba972ab9821b044b265 +size 497853 diff --git a/vlm/dev/RecZ9nB9Q4/32.png b/vlm/dev/RecZ9nB9Q4/32.png new file mode 100644 index 0000000000000000000000000000000000000000..43a32f5df8dbb5446278acd057853d4325459ee2 --- /dev/null +++ b/vlm/dev/RecZ9nB9Q4/32.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6f54fb44d794a26135f55f07f5fe77200f03665e696dce45090ad4510264b86 +size 319167 diff --git a/vlm/dev/S5wmbQc1We/0.png b/vlm/dev/S5wmbQc1We/0.png new file mode 100644 index 0000000000000000000000000000000000000000..4ea14404c87a96b9b7efd0badb90ad196b090e4c --- /dev/null +++ b/vlm/dev/S5wmbQc1We/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84d0c0dfec799f03bd5b87dae1b97324604697ebaefe93111c3b42b7acb36bb2 +size 442401 diff --git a/vlm/dev/S5wmbQc1We/1.png b/vlm/dev/S5wmbQc1We/1.png new file mode 100644 index 0000000000000000000000000000000000000000..420425a00a043cc4e1d9bc80acad813121d72521 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b670568ff1342dcb037243ba11d66d4835503fd5932eab7410cca3cd91ea345b +size 486479 diff --git a/vlm/dev/S5wmbQc1We/10.png b/vlm/dev/S5wmbQc1We/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c54efb2fecd3a6c77301bd498ad584e5d45658c2 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55dbea1d6cbb204c0eb53622fb1730db3cd3f2cc044326a178fdcda7b7e6d751 +size 633782 diff --git a/vlm/dev/S5wmbQc1We/11.png b/vlm/dev/S5wmbQc1We/11.png new file mode 100644 index 0000000000000000000000000000000000000000..43a892b5e1d75af7df6634bbb2146cd94cf55159 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90334a6437edd2ba804d63e31ffb9520ade4f31bef3f4e4ec2f1b2e2075911d2 +size 340544 diff --git a/vlm/dev/S5wmbQc1We/12.png b/vlm/dev/S5wmbQc1We/12.png new file mode 100644 index 0000000000000000000000000000000000000000..7dff7445294a98e758727002426e9b75e3694e86 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:312af12105454ca5ebcc1c683487c12f37ae8a2a38ec6228bb1d7687e57cf709 +size 355799 diff --git a/vlm/dev/S5wmbQc1We/13.png b/vlm/dev/S5wmbQc1We/13.png new file mode 100644 index 0000000000000000000000000000000000000000..1d67f2d40bc7ba1b17b219846f469b02a606902c --- /dev/null +++ b/vlm/dev/S5wmbQc1We/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5f45c32088825573f67d9dfbd3da1e5c38914f1386bff72b19c74359adfdb20 +size 450712 diff --git a/vlm/dev/S5wmbQc1We/14.png b/vlm/dev/S5wmbQc1We/14.png new file mode 100644 index 0000000000000000000000000000000000000000..49b90d23f8dbad75b7926152b7889bbc68cff8d8 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1544d06eaa2391af81ee9424e84892ee842d142decab807dfebd8705f52d7853 +size 722289 diff --git a/vlm/dev/S5wmbQc1We/15.png b/vlm/dev/S5wmbQc1We/15.png new file mode 100644 index 0000000000000000000000000000000000000000..4774b102f560778a99139f9e781ab90db88e5317 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c46b467d816128d65328314e760b7ac87808b3acab17c563e35afa6098ee05b2 +size 656929 diff --git a/vlm/dev/S5wmbQc1We/16.png b/vlm/dev/S5wmbQc1We/16.png new file mode 100644 index 0000000000000000000000000000000000000000..8c8cefa287a2a412fd8139ea8872bbe1c5086297 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05ec8c996f10cd8563306020fcf65fd2bb327f022e0d499336b0d1919535641e +size 609569 diff --git a/vlm/dev/S5wmbQc1We/17.png b/vlm/dev/S5wmbQc1We/17.png new file mode 100644 index 0000000000000000000000000000000000000000..9cc5503ebc0081cd4e139716eaf6ce1f8f0cb9c1 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3da8250a0d344302d523873cdc95b424f728710ea0a745d98a0fbb272ba67008 +size 444115 diff --git a/vlm/dev/S5wmbQc1We/18.png b/vlm/dev/S5wmbQc1We/18.png new file mode 100644 index 0000000000000000000000000000000000000000..119920ff7e5de805d773b8a163b92b9b355f86bf --- /dev/null +++ b/vlm/dev/S5wmbQc1We/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c49276898a30a1903d316de4c3b1ce9cfa7de2201776f925250a535dd21e79d3 +size 761710 diff --git a/vlm/dev/S5wmbQc1We/19.png b/vlm/dev/S5wmbQc1We/19.png new file mode 100644 index 0000000000000000000000000000000000000000..8f2caf7995506c07adfe44ca2b669c69fdd28be7 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1e3c78aadb5132fba4e77a71d4878dcb16867b6ca2c9634d7150664a990e936 +size 411952 diff --git a/vlm/dev/S5wmbQc1We/2.png b/vlm/dev/S5wmbQc1We/2.png new file mode 100644 index 0000000000000000000000000000000000000000..dca605bf04adca87a0183fd6f8fd974940394e89 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffdea256f4e1228d944568d64027d0624a48a6413d738d41376cc83696d7717a +size 498871 diff --git a/vlm/dev/S5wmbQc1We/20.png b/vlm/dev/S5wmbQc1We/20.png new file mode 100644 index 0000000000000000000000000000000000000000..76bf889b0a104d7ccce8cbd0a98ad7772ef84737 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d664e30b48cda6be743b56427a5ea46a1ff51aac389e6453869069cbdce0b66 +size 606897 diff --git a/vlm/dev/S5wmbQc1We/21.png b/vlm/dev/S5wmbQc1We/21.png new file mode 100644 index 0000000000000000000000000000000000000000..b3c516c9c00d472519568151f5de61c24bc8032e --- /dev/null +++ b/vlm/dev/S5wmbQc1We/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba28eee704cd990cc29251ff2dca4f9514751ab8617ead9a816244b80d41bde2 +size 469185 diff --git a/vlm/dev/S5wmbQc1We/22.png b/vlm/dev/S5wmbQc1We/22.png new file mode 100644 index 0000000000000000000000000000000000000000..0ca24348265a20737bf011eb3d0128640176a767 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58f39fc97b529ff83bb87b8aff4b36c4d3d72746942cf917798b47d82006653d +size 678154 diff --git a/vlm/dev/S5wmbQc1We/23.png b/vlm/dev/S5wmbQc1We/23.png new file mode 100644 index 0000000000000000000000000000000000000000..86f93befbee8d449dd87337548fc154e0085737b --- /dev/null +++ b/vlm/dev/S5wmbQc1We/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2197cc2abf6e3f9ec4eca98dcca782bffeb542924cda89615977d187c0ad95c0 +size 1379131 diff --git a/vlm/dev/S5wmbQc1We/24.png b/vlm/dev/S5wmbQc1We/24.png new file mode 100644 index 0000000000000000000000000000000000000000..43e2270eccf6ebb322328fc0497dd607f067f0a3 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf1bc56f6a425c361800b11f09a837837435cf918a1eebccd200dfaf3172c1d2 +size 484040 diff --git a/vlm/dev/S5wmbQc1We/25.png b/vlm/dev/S5wmbQc1We/25.png new file mode 100644 index 0000000000000000000000000000000000000000..7df11aa55110175c30a3f8849eb91046aa327773 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab2ee7e379a7f5f48fa0611cd3a6e3b4125f52d1c60c65f7a3fb0fa9878617d5 +size 524990 diff --git a/vlm/dev/S5wmbQc1We/26.png b/vlm/dev/S5wmbQc1We/26.png new file mode 100644 index 0000000000000000000000000000000000000000..3457809c1d4227f7a988f03afea2aac64d685e48 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dea0a3b36eb27fe8626f80cad9ed6f3b0ebd35db80b7764d962be2c50e1709e4 +size 507334 diff --git a/vlm/dev/S5wmbQc1We/27.png b/vlm/dev/S5wmbQc1We/27.png new file mode 100644 index 0000000000000000000000000000000000000000..a0ad6b5ff9cbc62ab3f8469b53b4218b4a14124e --- /dev/null +++ b/vlm/dev/S5wmbQc1We/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c13ea515626853a263db0341a90a74732a05445c5e822496d03a8f4946c57c0 +size 67814 diff --git a/vlm/dev/S5wmbQc1We/3.png b/vlm/dev/S5wmbQc1We/3.png new file mode 100644 index 0000000000000000000000000000000000000000..d8ffba3d6652a4be3ab864b89a4a9307e8d14fd1 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:803ac46b45bfb68a30babbc623c4d3d721664a6cc01cb24fd0530a390c186cd6 +size 632446 diff --git a/vlm/dev/S5wmbQc1We/4.png b/vlm/dev/S5wmbQc1We/4.png new file mode 100644 index 0000000000000000000000000000000000000000..a59a14753968ee2c34728a5bb3de2cde26246d9e --- /dev/null +++ b/vlm/dev/S5wmbQc1We/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3429e3d45bd6d1ef1a6655997b52854a041faa5f748301136c73de3aa188ba7e +size 777776 diff --git a/vlm/dev/S5wmbQc1We/5.png b/vlm/dev/S5wmbQc1We/5.png new file mode 100644 index 0000000000000000000000000000000000000000..ac861d0f7676e2445b842917e55025e094e83e70 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f01e048421aed0a7b3826af318a3b6e7b000ff856b3cf63e7f1836f55adf177 +size 762453 diff --git a/vlm/dev/S5wmbQc1We/6.png b/vlm/dev/S5wmbQc1We/6.png new file mode 100644 index 0000000000000000000000000000000000000000..b908879f169c508a7835f89a1b468d0b17fe5e65 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bf81c1f05cdeb82171911e8b4eaf4344f6bb25dadf1412993631ce0beec4d80 +size 494789 diff --git a/vlm/dev/S5wmbQc1We/7.png b/vlm/dev/S5wmbQc1We/7.png new file mode 100644 index 0000000000000000000000000000000000000000..837be7ae0ca307819830903873275262bfbd38f7 --- /dev/null +++ b/vlm/dev/S5wmbQc1We/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bb28b8ef2ab0ceebc8ab2c0fa664331c01e3d51920cd1d4c8fcb3e4402dc8ff +size 518497 diff --git a/vlm/dev/S5wmbQc1We/8.png b/vlm/dev/S5wmbQc1We/8.png new file mode 100644 index 0000000000000000000000000000000000000000..3e4725ad1494cb38ea31a0c87983aecfc94866fe --- /dev/null +++ b/vlm/dev/S5wmbQc1We/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ea243293d03d421382f310cd3f6b89e5fef4e6a901384d8b3ac5a798e9cfed9 +size 787008 diff --git a/vlm/dev/S5wmbQc1We/9.png b/vlm/dev/S5wmbQc1We/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1b7bd9bab03f64c9135af2e24ab2ab0814a2a56f --- /dev/null +++ b/vlm/dev/S5wmbQc1We/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1e931c86e8c92043f033f50d126515d2a084efefb35bf39c4408ed09252ac3d +size 529101 diff --git a/vlm/dev/S9GpoS2TmN/0.png b/vlm/dev/S9GpoS2TmN/0.png new file mode 100644 index 0000000000000000000000000000000000000000..e31a8e5b62f755e8aad2a1a3be3c8e6862e8df93 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94d547225f3e2ca1e4933f754aa63fa8de53500055ffcd8724ff491196f922e9 +size 505600 diff --git a/vlm/dev/S9GpoS2TmN/1.png b/vlm/dev/S9GpoS2TmN/1.png new file mode 100644 index 0000000000000000000000000000000000000000..07365c3e0bf477ec2b650aec10453bd6f580abc1 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8ad8873ccdf3483e14598a6e19ae13d02091304f0f98b30d8a8ab73c3efe968 +size 583566 diff --git a/vlm/dev/S9GpoS2TmN/10.png b/vlm/dev/S9GpoS2TmN/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c27aac7826dfcc8716a1c9ca512199e1428872fe --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99c585b09113f8bc0a455482769d492bdcb1665a4a50f82a06486278c50579e2 +size 552981 diff --git a/vlm/dev/S9GpoS2TmN/11.png b/vlm/dev/S9GpoS2TmN/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ff2ccf9bcc309f9002b7acc659dfee8e589c67ab --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4c92310f5bdf372160260b25d414efe6667939e922398f7a1b2dc61d81cc4ad +size 571623 diff --git a/vlm/dev/S9GpoS2TmN/12.png b/vlm/dev/S9GpoS2TmN/12.png new file mode 100644 index 0000000000000000000000000000000000000000..ca7527e5e46f9854b1b3bbd12ed2c13c9f8dfe11 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0d1afaf1db4b7adc1b8c6b25ffa25910e5a06b37e9651cfec6da8026dd0c6cc +size 547658 diff --git a/vlm/dev/S9GpoS2TmN/13.png b/vlm/dev/S9GpoS2TmN/13.png new file mode 100644 index 0000000000000000000000000000000000000000..d8c6219eaa44e89c538f8d3ec20767d2e4cd1bef --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:499658f9714c8a20f13684a6449b93f5e3c0110be18a9fd708523a6cc82fdf74 +size 494225 diff --git a/vlm/dev/S9GpoS2TmN/14.png b/vlm/dev/S9GpoS2TmN/14.png new file mode 100644 index 0000000000000000000000000000000000000000..334e5da889b09e6cc488bc12682a9523ae05667f --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:662f3aab4b605b44b50a2af26031233d1211292c36c66cc51766a983c00e7608 +size 175981 diff --git a/vlm/dev/S9GpoS2TmN/15.png b/vlm/dev/S9GpoS2TmN/15.png new file mode 100644 index 0000000000000000000000000000000000000000..bb6d5fdea6aed2abe6fcafc00d67957a32c56d40 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:965d2367498ed1d43787fcb80dc0e9e62584ff6e0c247e3bc12eece947ada421 +size 456225 diff --git a/vlm/dev/S9GpoS2TmN/16.png b/vlm/dev/S9GpoS2TmN/16.png new file mode 100644 index 0000000000000000000000000000000000000000..96603dd98ebd5e3f0a93cde4f9cf63ee2e44fea2 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a18e7829b52074e4f99ad066bfb372c33db43a12b50da0e62a16a1b72636e692 +size 1251660 diff --git a/vlm/dev/S9GpoS2TmN/17.png b/vlm/dev/S9GpoS2TmN/17.png new file mode 100644 index 0000000000000000000000000000000000000000..2c62525812a8568bbab68ee78aa340729cb38328 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73481a0ad2764dd08be5423e639af8aaa0ce28310c26f14b839518857d4e2e91 +size 524658 diff --git a/vlm/dev/S9GpoS2TmN/18.png b/vlm/dev/S9GpoS2TmN/18.png new file mode 100644 index 0000000000000000000000000000000000000000..b83043c1db7197f678d087fbafca6d182d7a6eb8 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4063e88cb9d287944ca8d3c78d048f5ede388a2cad7e06a8aa27f9d437082298 +size 321961 diff --git a/vlm/dev/S9GpoS2TmN/19.png b/vlm/dev/S9GpoS2TmN/19.png new file mode 100644 index 0000000000000000000000000000000000000000..b0c52fa34a7fb3fdd2aa0d208c8305e8439cafa1 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4039103cb7a320720d4dae2352ad39a1e49eb5b122bec23b3af7c4279a0a236c +size 504993 diff --git a/vlm/dev/S9GpoS2TmN/2.png b/vlm/dev/S9GpoS2TmN/2.png new file mode 100644 index 0000000000000000000000000000000000000000..72ce6bc6a78bd16da3be879694963c3a57f625c9 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e70da6af50b1d35fd0163229fa9ddffcf7507cf7c273a1c8176639bb3230fc5 +size 610975 diff --git a/vlm/dev/S9GpoS2TmN/20.png b/vlm/dev/S9GpoS2TmN/20.png new file mode 100644 index 0000000000000000000000000000000000000000..454b1e9db6201bbef68eade034ca6928fa48ecb8 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbf5090d544812e4ea8900974f29c02f0f37b72969df2edb87c48e9ed4b2eb83 +size 265866 diff --git a/vlm/dev/S9GpoS2TmN/21.png b/vlm/dev/S9GpoS2TmN/21.png new file mode 100644 index 0000000000000000000000000000000000000000..ae89d52cb07d1844afc325e3ba497e643772d3e4 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91e7150fbf8385bd9ea49aec58ae46567f209d608203397281bbad7190d8c631 +size 478473 diff --git a/vlm/dev/S9GpoS2TmN/22.png b/vlm/dev/S9GpoS2TmN/22.png new file mode 100644 index 0000000000000000000000000000000000000000..eeddab47abe156f694b45c2e64ab50446e24dd1f --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cb3daeb7252f1b4467e71ebb724fc50522d7645ec1d30c925cf38ded049ec21 +size 290586 diff --git a/vlm/dev/S9GpoS2TmN/23.png b/vlm/dev/S9GpoS2TmN/23.png new file mode 100644 index 0000000000000000000000000000000000000000..052ed8c69eb70002b776468077f1bc455b4df080 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdbaaa9ac00c40c5bcc4755bd7dbe49c60b7f56f2a886dd4751fb9ca22bc3722 +size 334145 diff --git a/vlm/dev/S9GpoS2TmN/24.png b/vlm/dev/S9GpoS2TmN/24.png new file mode 100644 index 0000000000000000000000000000000000000000..7712a7088d82756bbcd84a7a630ce5f144c62981 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eca09f48cfcbb9794b69c4052bf18f4dc93c90801a718b1b22bfed5175e19d2 +size 402600 diff --git a/vlm/dev/S9GpoS2TmN/25.png b/vlm/dev/S9GpoS2TmN/25.png new file mode 100644 index 0000000000000000000000000000000000000000..0474686a0ffaa56a6d6c19c64100234c4b978816 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a13559e628dc710cf68d8dc0ca0635c2931c41c2b743bce8af01f3afc4c86db +size 427143 diff --git a/vlm/dev/S9GpoS2TmN/26.png b/vlm/dev/S9GpoS2TmN/26.png new file mode 100644 index 0000000000000000000000000000000000000000..1b83a06b9074a70653539bbb80a96cddb17fd4dd --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efe1912aa0add3c772639f1a4eee5e40c9abeebb6729181505006f3eae7d44e9 +size 412362 diff --git a/vlm/dev/S9GpoS2TmN/27.png b/vlm/dev/S9GpoS2TmN/27.png new file mode 100644 index 0000000000000000000000000000000000000000..bdda84ac7ef67ed58748fd43fadb327ad1e0aae8 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c6bfcf963d257ae5221a4b8873a3a3e728b6d165212cfa77c1b151349d2f7b +size 627175 diff --git a/vlm/dev/S9GpoS2TmN/28.png b/vlm/dev/S9GpoS2TmN/28.png new file mode 100644 index 0000000000000000000000000000000000000000..0b12a75ebbddb8dfbe0cf4219a318ea49e50befc --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8f726c28a7af219dfe64bb2f1965d61874f76a2559e5e65a6c9947298eb3b4f +size 488497 diff --git a/vlm/dev/S9GpoS2TmN/29.png b/vlm/dev/S9GpoS2TmN/29.png new file mode 100644 index 0000000000000000000000000000000000000000..9caa0d80c5d91b036de7fd592b5f3331ea83ae3c --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afa619588f5ab97b07e897848ba30447e744fd54fcf262788a7663cc68b2a4a9 +size 428489 diff --git a/vlm/dev/S9GpoS2TmN/3.png b/vlm/dev/S9GpoS2TmN/3.png new file mode 100644 index 0000000000000000000000000000000000000000..615d90f7b845f1b45ad4cb63a12b0747db70073d --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:559a13845cfd4543cce5e38ab13b2d365a68ac7b9b7838d32991a29de99ba286 +size 633408 diff --git a/vlm/dev/S9GpoS2TmN/30.png b/vlm/dev/S9GpoS2TmN/30.png new file mode 100644 index 0000000000000000000000000000000000000000..a2de0d666772b9e08f6de2aeea53c439d055305d --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/30.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:447038a1b64c98f36ca821406f3d0a47013b18bbed4dc9ea74cbe1dea6b34c5b +size 254230 diff --git a/vlm/dev/S9GpoS2TmN/31.png b/vlm/dev/S9GpoS2TmN/31.png new file mode 100644 index 0000000000000000000000000000000000000000..4e6632c49843b00a0fb12615708851624e2bd2aa --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/31.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e80d30e074f78b79f676699332d2998f71d6d9dcdcb73e90513be551fde33d05 +size 352221 diff --git a/vlm/dev/S9GpoS2TmN/32.png b/vlm/dev/S9GpoS2TmN/32.png new file mode 100644 index 0000000000000000000000000000000000000000..23027d4cc307010f8d0e5bf54a739f5ba2b86e14 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/32.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1753c8aa8562df9a4fa90884746a6b7a6919e39b2186c1e7a0e9c8ad29b66500 +size 224910 diff --git a/vlm/dev/S9GpoS2TmN/33.png b/vlm/dev/S9GpoS2TmN/33.png new file mode 100644 index 0000000000000000000000000000000000000000..fc81845787416f95ef58ae140da4039ffce8095e --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/33.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c40847d528d0ba7687c4020bd51f6f72de761f71f1a6dc366fd74bb82515506 +size 445363 diff --git a/vlm/dev/S9GpoS2TmN/34.png b/vlm/dev/S9GpoS2TmN/34.png new file mode 100644 index 0000000000000000000000000000000000000000..9e583628eb1226b73a6f8148cc70cec71cdc8297 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/34.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:936bafd9ba01363cf79ae7dd0de56451cd4a5f264a84d3011625f856b44d678b +size 276633 diff --git a/vlm/dev/S9GpoS2TmN/4.png b/vlm/dev/S9GpoS2TmN/4.png new file mode 100644 index 0000000000000000000000000000000000000000..0bb41e330de39a89275484dba47600e2c0f481c4 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47e0641906c753a23ee5439f671ffcf65731ad1f8ce226adf7890e619f820b27 +size 566456 diff --git a/vlm/dev/S9GpoS2TmN/5.png b/vlm/dev/S9GpoS2TmN/5.png new file mode 100644 index 0000000000000000000000000000000000000000..b1626fd68a47cdd9596af5d940ffc9d56cd2e230 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d00ba50bc89b70795f9c51876706bbcc79fdffc50a36a8ca4d775bd81e408e1 +size 562659 diff --git a/vlm/dev/S9GpoS2TmN/6.png b/vlm/dev/S9GpoS2TmN/6.png new file mode 100644 index 0000000000000000000000000000000000000000..9a1771742e84309f8ee09bc45dc06db2599f5d34 --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:979035efdf6ac8da62eb875ac641b36f176c38b0a3a02e0642a163b3923ddf60 +size 553726 diff --git a/vlm/dev/S9GpoS2TmN/7.png b/vlm/dev/S9GpoS2TmN/7.png new file mode 100644 index 0000000000000000000000000000000000000000..da27c7d7b543640d6105c3c766fdfee99c29371f --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adbb0baf06d7d6ebb88af70488e797d48d8a48a3bacd81a4f49816bda012b1e8 +size 563557 diff --git a/vlm/dev/S9GpoS2TmN/8.png b/vlm/dev/S9GpoS2TmN/8.png new file mode 100644 index 0000000000000000000000000000000000000000..c8ad03070b6ca555a539ff63115e7e8898dda91a --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f728b666e573490149b064a89d983d6634cee0927dbf1390dd4d8423b8cb530 +size 449659 diff --git a/vlm/dev/S9GpoS2TmN/9.png b/vlm/dev/S9GpoS2TmN/9.png new file mode 100644 index 0000000000000000000000000000000000000000..2a29cbf5aa920bd66c4331256a61cf6329c7908b --- /dev/null +++ b/vlm/dev/S9GpoS2TmN/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb437eafc6c8e4821597d421f8f1d2e2a0dab94421fad31233902191f89f1d19 +size 117602 diff --git a/vlm/dev/SMa9EAovKMC/0.png b/vlm/dev/SMa9EAovKMC/0.png new file mode 100644 index 0000000000000000000000000000000000000000..4dc65eff40469efcb449ccd819b3d359c3f3eb0f --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2c0fee7381e2139d30db7798798f09fad019f3e345384d7defad0413e7f80d2 +size 527958 diff --git a/vlm/dev/SMa9EAovKMC/1.png b/vlm/dev/SMa9EAovKMC/1.png new file mode 100644 index 0000000000000000000000000000000000000000..165cb7ce3ce65129a448762d67643ba5792b813e --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:529e5743a7430da7e099a42a902a08c1b4158b2542fa75d77337f424101ba824 +size 592786 diff --git a/vlm/dev/SMa9EAovKMC/10.png b/vlm/dev/SMa9EAovKMC/10.png new file mode 100644 index 0000000000000000000000000000000000000000..dc703ef929bc08febc90e08b8a1a3645e6812101 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f84701fd94894e0f693da5ebddc4e54c22dc3c1f61e713939d0a080749453ad4 +size 606931 diff --git a/vlm/dev/SMa9EAovKMC/11.png b/vlm/dev/SMa9EAovKMC/11.png new file mode 100644 index 0000000000000000000000000000000000000000..7e20b079fda8ae8e9586455f77e78449af37fbd1 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:103bb5896cd482d8097a1bbe34d48c070865668b16be8eb2bd949167aa000a3d +size 622996 diff --git a/vlm/dev/SMa9EAovKMC/12.png b/vlm/dev/SMa9EAovKMC/12.png new file mode 100644 index 0000000000000000000000000000000000000000..1a37d87f7a5f95e75fce05ec124937e8e132bd67 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fea0af6734346f2a3fe60ad62ac1c97adf3b2cbff86ba82be4569b43c4d2ae1 +size 298111 diff --git a/vlm/dev/SMa9EAovKMC/13.png b/vlm/dev/SMa9EAovKMC/13.png new file mode 100644 index 0000000000000000000000000000000000000000..64f5d62f1ddbeb7897997726cab37a5b40572973 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:577132acdc0dcf0f20410eb5500f149632a87fa517886c334fe4c013f6f8c52f +size 488660 diff --git a/vlm/dev/SMa9EAovKMC/14.png b/vlm/dev/SMa9EAovKMC/14.png new file mode 100644 index 0000000000000000000000000000000000000000..6e8cf525b3f5c7a4a305703e347b7940fee70c04 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:889df542dfb4be9b79b0843665b029734c2b68b0b5ac50b7df8ec36d5afc587a +size 148976 diff --git a/vlm/dev/SMa9EAovKMC/15.png b/vlm/dev/SMa9EAovKMC/15.png new file mode 100644 index 0000000000000000000000000000000000000000..3d8c894ffcbf87b42dcedc3b6d55e02642047f36 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edbf4175568c2810239436e5f0d851a008081b8de0e2679a532f9dda579a4ee6 +size 203276 diff --git a/vlm/dev/SMa9EAovKMC/16.png b/vlm/dev/SMa9EAovKMC/16.png new file mode 100644 index 0000000000000000000000000000000000000000..a8d6040e69d7f284870729cdbef631139d212b6f --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:526c7c6de7214f982d19aa5f1103c6b382fdb27e604893e73957ea35ce53db47 +size 303865 diff --git a/vlm/dev/SMa9EAovKMC/17.png b/vlm/dev/SMa9EAovKMC/17.png new file mode 100644 index 0000000000000000000000000000000000000000..eb8bb7ce98a461b719e5f0bdbc54d38ed9a779d0 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34be60830986089508725709f7542f2072a298a73121e99c99bbdc56aefbef80 +size 496984 diff --git a/vlm/dev/SMa9EAovKMC/18.png b/vlm/dev/SMa9EAovKMC/18.png new file mode 100644 index 0000000000000000000000000000000000000000..61a6e6a3ed1c2fb2bd27d6d33a00e9e9af47a3f9 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd98e7f823eb5182cd23030bf61462fbce1f3174d2d2ad58217e83995fe91cb4 +size 524938 diff --git a/vlm/dev/SMa9EAovKMC/2.png b/vlm/dev/SMa9EAovKMC/2.png new file mode 100644 index 0000000000000000000000000000000000000000..662deebfaca6baa594ac69df1b7368589c7bb8b0 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b55200984487b178080b7839190fd312807734fecdb2bed365d9eb7d2a41a93a +size 617050 diff --git a/vlm/dev/SMa9EAovKMC/3.png b/vlm/dev/SMa9EAovKMC/3.png new file mode 100644 index 0000000000000000000000000000000000000000..80c6f8fc5be4bed6200ca3ea2ab695219dd72be4 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14e31594eadeb12a0a6d62f61a3e0d2a29d4bd3d7c88af16a74a8ebc0aee9aa7 +size 517828 diff --git a/vlm/dev/SMa9EAovKMC/4.png b/vlm/dev/SMa9EAovKMC/4.png new file mode 100644 index 0000000000000000000000000000000000000000..21461c5a6ef789383ab5aa35064ea449511ff7ff --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdc59b620be69c872f6a9e2884eb400490cbc8e104c52a192e8789beb350e6c3 +size 646783 diff --git a/vlm/dev/SMa9EAovKMC/5.png b/vlm/dev/SMa9EAovKMC/5.png new file mode 100644 index 0000000000000000000000000000000000000000..c49bab4ab21c2b10a130828186b84331942aaf09 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3cf63f71e51f2f31d0c0be04c4af18d39511ab9a2581661ed3b3e2f948cf3b1 +size 552498 diff --git a/vlm/dev/SMa9EAovKMC/6.png b/vlm/dev/SMa9EAovKMC/6.png new file mode 100644 index 0000000000000000000000000000000000000000..814589b55ec929923fb4c515cdfbb5cb473fcfc1 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5345c3b144ae5c20377413893095932a36fecbcec931df24be20c0ff7fb69ab8 +size 604114 diff --git a/vlm/dev/SMa9EAovKMC/7.png b/vlm/dev/SMa9EAovKMC/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c844dd91ef28f813783015c383fbbc344b6ec95e --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d243c010d950068f5e36df9b88412b8212275a382f6fe043d25f2a85a2a3805 +size 465620 diff --git a/vlm/dev/SMa9EAovKMC/8.png b/vlm/dev/SMa9EAovKMC/8.png new file mode 100644 index 0000000000000000000000000000000000000000..ca3d08222a88aefa6163830de89d550c8a38ca08 --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7c56324687d2febc07343634e0463e12cdf6b48278fd16069cd53cf76bdb065 +size 565572 diff --git a/vlm/dev/SMa9EAovKMC/9.png b/vlm/dev/SMa9EAovKMC/9.png new file mode 100644 index 0000000000000000000000000000000000000000..90e4453f579a48836be9b9411eec7c80823387be --- /dev/null +++ b/vlm/dev/SMa9EAovKMC/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d453dd3a7da2942ecec668832d5c999c621bef9917035af1fcf5f0d3905afad1 +size 599350 diff --git a/vlm/dev/SVBR6xBaMl/0.png b/vlm/dev/SVBR6xBaMl/0.png new file mode 100644 index 0000000000000000000000000000000000000000..e2f5ff974397db3fa5056aa452311a9b8706561e --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4bbd85ac91902a16d2087392bc9f0c5c00b50ebe9b4404e8d6c6894ac7fba1b +size 456886 diff --git a/vlm/dev/SVBR6xBaMl/1.png b/vlm/dev/SVBR6xBaMl/1.png new file mode 100644 index 0000000000000000000000000000000000000000..9a974419464a37097ac5460b8ebb5e2b7bc6f61d --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa8bf14fc9700411bedbe10cf184d2d2a80e6a899d36ded6317199a2fa2f8ab6 +size 661697 diff --git a/vlm/dev/SVBR6xBaMl/10.png b/vlm/dev/SVBR6xBaMl/10.png new file mode 100644 index 0000000000000000000000000000000000000000..9a297597ca78dd8f007d487779b7856d339481c6 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44d5a7332e75ecaa73801dc65fee8760b5d2517015ff85e2a205cd23b5d710c6 +size 571047 diff --git a/vlm/dev/SVBR6xBaMl/11.png b/vlm/dev/SVBR6xBaMl/11.png new file mode 100644 index 0000000000000000000000000000000000000000..e64fe7bdb105f07d966b4baaee65708fc1349a7c --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c7e241583d237344dbda43487633a58cd1b8984074887a3f962d394bdccfc71 +size 582470 diff --git a/vlm/dev/SVBR6xBaMl/12.png b/vlm/dev/SVBR6xBaMl/12.png new file mode 100644 index 0000000000000000000000000000000000000000..1ec902ee02bc7142423fd0ace02eb5a8075c1c09 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:786bb4518a236ed724e9faf563858638cc7733972794a9ee1d48badc28e48fe1 +size 590039 diff --git a/vlm/dev/SVBR6xBaMl/13.png b/vlm/dev/SVBR6xBaMl/13.png new file mode 100644 index 0000000000000000000000000000000000000000..26f5911aefc5866c744af26d1802e411984e412c --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:777d031427700735dcb52e0f06f7a8c5ca07a8fe31ed1eda497ad86ebaf54c3d +size 493227 diff --git a/vlm/dev/SVBR6xBaMl/14.png b/vlm/dev/SVBR6xBaMl/14.png new file mode 100644 index 0000000000000000000000000000000000000000..7ec0ac48439fe60e09e583612bfc6c083ecc311a --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:089ef7f77852543ed9ad98bd84ac714436be4f4798250ebdd6896c5fb34dfba3 +size 389354 diff --git a/vlm/dev/SVBR6xBaMl/15.png b/vlm/dev/SVBR6xBaMl/15.png new file mode 100644 index 0000000000000000000000000000000000000000..6df851be2d0da046c6d7ff2a7515b5604b31f220 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa091fea8e105bf054824304c60eef278b581d53cc7ef6d4c1c918a1de94830e +size 226317 diff --git a/vlm/dev/SVBR6xBaMl/16.png b/vlm/dev/SVBR6xBaMl/16.png new file mode 100644 index 0000000000000000000000000000000000000000..5b9dee69425bd1e13601f1ab243875bc90b9c19c --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d481dce2339ae0f6fba10889305a9093635e46129eab894ff86d900e9668a1ae +size 203791 diff --git a/vlm/dev/SVBR6xBaMl/17.png b/vlm/dev/SVBR6xBaMl/17.png new file mode 100644 index 0000000000000000000000000000000000000000..79f8466e50bafcc8fb078362ef79e29b713ffeab --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a825d0f6a75ddfeb5445100239825dd0e143bd6fefe9b3f492d6fba7285bb0e +size 232003 diff --git a/vlm/dev/SVBR6xBaMl/18.png b/vlm/dev/SVBR6xBaMl/18.png new file mode 100644 index 0000000000000000000000000000000000000000..26ea1459022b6e09d3b62209dffb002843d5b5cf --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6182b32340fce428cf8b84b8cf8b9463d66edeefcca00e4348e338e959f9193 +size 483288 diff --git a/vlm/dev/SVBR6xBaMl/19.png b/vlm/dev/SVBR6xBaMl/19.png new file mode 100644 index 0000000000000000000000000000000000000000..fb00e00b2fd7f3df78ee43796dec7d0c09adbb92 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:364535f43765c2b4f65166d34707a2585ac67f0c6cc8f621f9cb8d9b84a3867d +size 401356 diff --git a/vlm/dev/SVBR6xBaMl/2.png b/vlm/dev/SVBR6xBaMl/2.png new file mode 100644 index 0000000000000000000000000000000000000000..71032d184deabe3ddeadd5df0df0e8285bd1ca0c --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:067d77b5edf7e0851fa21914887ac869fa9319e19fb4da44e8867ce3e17e92f7 +size 627719 diff --git a/vlm/dev/SVBR6xBaMl/20.png b/vlm/dev/SVBR6xBaMl/20.png new file mode 100644 index 0000000000000000000000000000000000000000..21f33a3e8c8ca3f100beba28c9143f019d0f83f0 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62b831368b8e604e711a15c9be2fde2d34b441b1746dcaaa1a4a415ca22c5c08 +size 87149 diff --git a/vlm/dev/SVBR6xBaMl/3.png b/vlm/dev/SVBR6xBaMl/3.png new file mode 100644 index 0000000000000000000000000000000000000000..89b3022ca6d922bb70b91bb47d5bd70e34d6baac --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18c70b482f62a853335c69d6917f5aa65b2cbde2ea1d5a4e830e9c431af70e94 +size 680230 diff --git a/vlm/dev/SVBR6xBaMl/4.png b/vlm/dev/SVBR6xBaMl/4.png new file mode 100644 index 0000000000000000000000000000000000000000..eb807a566c3574fe38bca8377224d22e350b3b35 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8ad38f8ad5901d5888d2ce9d77981a49f22326cbefa1422cc248a992ed47a21 +size 556489 diff --git a/vlm/dev/SVBR6xBaMl/5.png b/vlm/dev/SVBR6xBaMl/5.png new file mode 100644 index 0000000000000000000000000000000000000000..d72b75b7ee782408d3b7b6aef35e493672e4d2b7 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7ba40d152f345689e49d942d1120e1c0c53db902bba40c9e8c2ba4b71630c1f +size 600102 diff --git a/vlm/dev/SVBR6xBaMl/6.png b/vlm/dev/SVBR6xBaMl/6.png new file mode 100644 index 0000000000000000000000000000000000000000..e2469ba1edd430a995b0e921c0d3a3fe5087db18 --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f44d422f57c3dc5052c1f4bcbd3c6c67e956b2f09a52e1066af833360dd7f7a +size 621859 diff --git a/vlm/dev/SVBR6xBaMl/7.png b/vlm/dev/SVBR6xBaMl/7.png new file mode 100644 index 0000000000000000000000000000000000000000..1136b1428252b03d51086e168475914169e539cd --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5f705a0b9bcb823f2de4d3640c0096cf7b77e87bc53577617f9bfc85be2b4ba +size 249169 diff --git a/vlm/dev/SVBR6xBaMl/8.png b/vlm/dev/SVBR6xBaMl/8.png new file mode 100644 index 0000000000000000000000000000000000000000..2dd4e5dafe5ba7c5970f37a34eb06aab8bc55d5a --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fb127a286ed6d349ea6816894df720508c5db2e997db27fb4ae6ca7536e108d +size 521680 diff --git a/vlm/dev/SVBR6xBaMl/9.png b/vlm/dev/SVBR6xBaMl/9.png new file mode 100644 index 0000000000000000000000000000000000000000..d778913abf769ef322b45d24ef0a0b5b4c33118a --- /dev/null +++ b/vlm/dev/SVBR6xBaMl/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:658e925987371e6c0557dadb051c29a034c0c64d0b94b14580a86465d6e8efec +size 479621 diff --git a/vlm/dev/T0GpzBQ1Fg6/0.png b/vlm/dev/T0GpzBQ1Fg6/0.png new file mode 100644 index 0000000000000000000000000000000000000000..fe18740911792b0c1e72e496413fdc4e20a6a620 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caa4ab4bddbe63982e6c5e73e771d290ed5cd4bf450fa6037b69bf7943a4a2ef +size 546622 diff --git a/vlm/dev/T0GpzBQ1Fg6/1.png b/vlm/dev/T0GpzBQ1Fg6/1.png new file mode 100644 index 0000000000000000000000000000000000000000..bafdb6a39651e256b3320172af6bc676253820b7 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa27e88790e4c69552fe0acc531721b684b2d23a00234564b06aa76b6421e282 +size 457731 diff --git a/vlm/dev/T0GpzBQ1Fg6/10.png b/vlm/dev/T0GpzBQ1Fg6/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c016343415312882aee24b999f36d0c700ee8402 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:779676c2780e99ad1c0a52101f6e412e06a32a55daf7ec2a7d182b69a523bd8f +size 547267 diff --git a/vlm/dev/T0GpzBQ1Fg6/11.png b/vlm/dev/T0GpzBQ1Fg6/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a770babf2c2dd4ac9e32f2abde0c1a8ea1a3d4b3 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02080f2b85f0790905337829fed1d4eba60d87d751fd7b2cc60a8e419af462e7 +size 521147 diff --git a/vlm/dev/T0GpzBQ1Fg6/12.png b/vlm/dev/T0GpzBQ1Fg6/12.png new file mode 100644 index 0000000000000000000000000000000000000000..a04a1901641b25a05e32ee940f752de422982331 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3473f7c2a862b258eb2e360ab142fbd5cbc0db1011eb2d97c201f358b8d1e49c +size 502253 diff --git a/vlm/dev/T0GpzBQ1Fg6/13.png b/vlm/dev/T0GpzBQ1Fg6/13.png new file mode 100644 index 0000000000000000000000000000000000000000..84710bbf04c46570d3e8d065f663a3138afe4d03 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a266e16a166973fd0f3770bed8a38b15e620d3f47e66be3d80dedcb3bb28f3db +size 333475 diff --git a/vlm/dev/T0GpzBQ1Fg6/14.png b/vlm/dev/T0GpzBQ1Fg6/14.png new file mode 100644 index 0000000000000000000000000000000000000000..73eeaa5a5da0b3824599bcf5192b2174f42d209c --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ed2d02a2cba28b8158251573c4c9b569ea1f346efd4bb30bf3f83b5515e1ae7 +size 472611 diff --git a/vlm/dev/T0GpzBQ1Fg6/15.png b/vlm/dev/T0GpzBQ1Fg6/15.png new file mode 100644 index 0000000000000000000000000000000000000000..5fc9722b0dbb09c43ae5c1dfed11d6889264b3c7 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de495c5a0115bc48ac7430885dd8f316197978c46f958fe545855e1bf6335e78 +size 403560 diff --git a/vlm/dev/T0GpzBQ1Fg6/16.png b/vlm/dev/T0GpzBQ1Fg6/16.png new file mode 100644 index 0000000000000000000000000000000000000000..0ba2b3816e915a9ec9057cf0d22da6c60729f033 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5369388e293faeaa91d08bbd19ff88f845d79bed0e2f8e32da37324317ae736 +size 471649 diff --git a/vlm/dev/T0GpzBQ1Fg6/17.png b/vlm/dev/T0GpzBQ1Fg6/17.png new file mode 100644 index 0000000000000000000000000000000000000000..12edc3ac33bdab5ce4ae3baca667d1e75bf40664 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f01fcaf6cac4a7add9bf64edc1894a9a92f9e08eb2505955926fa5d276f7abc +size 358152 diff --git a/vlm/dev/T0GpzBQ1Fg6/18.png b/vlm/dev/T0GpzBQ1Fg6/18.png new file mode 100644 index 0000000000000000000000000000000000000000..59cbe83ac54b9c190759c67454d10dead9ec3ede --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01199ed0cc23b1388e21627a1a3d685a3baa9123ff25152374fa378b814f6f4d +size 460683 diff --git a/vlm/dev/T0GpzBQ1Fg6/19.png b/vlm/dev/T0GpzBQ1Fg6/19.png new file mode 100644 index 0000000000000000000000000000000000000000..f17074030ee85139cef0b03f10599b438b1c3442 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a875abc193a2e0cb1455c7237bc684d83b985b1444c9212aba1ac29aa53d8d56 +size 224241 diff --git a/vlm/dev/T0GpzBQ1Fg6/2.png b/vlm/dev/T0GpzBQ1Fg6/2.png new file mode 100644 index 0000000000000000000000000000000000000000..9385c9fb5787a5fcd64f81686b960cf816ff0fd9 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd4bff22047403c90d33633bfdec39b0d53b343072246aadf8ad889a946034fd +size 539819 diff --git a/vlm/dev/T0GpzBQ1Fg6/3.png b/vlm/dev/T0GpzBQ1Fg6/3.png new file mode 100644 index 0000000000000000000000000000000000000000..7a9c351302883a99ac5d4b4f5c3fcb0ef86a9477 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c3623350b060a72db881fd912143f4d096bd899facfa72d4bd7f6a31739f9f1 +size 585618 diff --git a/vlm/dev/T0GpzBQ1Fg6/4.png b/vlm/dev/T0GpzBQ1Fg6/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4cf29a5c0b4153d9965986897926174a762a9125 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96af9b3649b2f8bdd6a68dc21be3064662791403aa0a7539a19533ba36baea11 +size 707574 diff --git a/vlm/dev/T0GpzBQ1Fg6/5.png b/vlm/dev/T0GpzBQ1Fg6/5.png new file mode 100644 index 0000000000000000000000000000000000000000..d41d611f2b9b4c7b30791170f4523dcfe612bdfc --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8f98a88d46cdd7d83023f8d7edd9dfcb16a107bcc9c027d951d7880955c32fb +size 637787 diff --git a/vlm/dev/T0GpzBQ1Fg6/6.png b/vlm/dev/T0GpzBQ1Fg6/6.png new file mode 100644 index 0000000000000000000000000000000000000000..07849a56ad2653435495d0855099fd2f4c19c5a0 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:423a2de88aa534cac1b78097cc4dd57e8dd402ee16a8a6911a79f824298c9a58 +size 636973 diff --git a/vlm/dev/T0GpzBQ1Fg6/7.png b/vlm/dev/T0GpzBQ1Fg6/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f4f762cf7a44a9ebcdc04411e81b174b934b7e6e --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:598a03de3ed9390ac6dbba34d1ff25697a139d7fb1aee1c997042fb2c0fe0e26 +size 549607 diff --git a/vlm/dev/T0GpzBQ1Fg6/8.png b/vlm/dev/T0GpzBQ1Fg6/8.png new file mode 100644 index 0000000000000000000000000000000000000000..d1fde3f5cda5db76521a2b0bb8f54914f9d99be9 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ff5c5db2415eee47b508f469e27337943017bf3a7c55e4ca09e641816760df2 +size 687401 diff --git a/vlm/dev/T0GpzBQ1Fg6/9.png b/vlm/dev/T0GpzBQ1Fg6/9.png new file mode 100644 index 0000000000000000000000000000000000000000..6d7e5f409dff22f4beaede81cccacf6fdf945176 --- /dev/null +++ b/vlm/dev/T0GpzBQ1Fg6/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3191d0fe587547a14ff8a4b1593ba99588e89eb3310fb3fb0cf5ec85b78dbc76 +size 534952 diff --git a/vlm/dev/TVHS5Y4dNvM/0.png b/vlm/dev/TVHS5Y4dNvM/0.png new file mode 100644 index 0000000000000000000000000000000000000000..78ee523abe3e127f2a2231b6540267dba4845753 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d16069e32fdcb9e599cc40100a4bac9df34c87021306da8efe629c33d202c48b +size 521303 diff --git a/vlm/dev/TVHS5Y4dNvM/1.png b/vlm/dev/TVHS5Y4dNvM/1.png new file mode 100644 index 0000000000000000000000000000000000000000..4a321ebe9ed009c228cacf3ad683a8d9eee4f10f --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66756fc7f61d345a3a3c504a5bc975b64892ec2adf5acba469010daeda1d1591 +size 554276 diff --git a/vlm/dev/TVHS5Y4dNvM/10.png b/vlm/dev/TVHS5Y4dNvM/10.png new file mode 100644 index 0000000000000000000000000000000000000000..76189b1afb948242a2e456556c16b6702a41e095 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e06aa84ec977a9ce2048b246a798672f2c6e2c95dfb43869eb3c5d01b1b621a7 +size 617692 diff --git a/vlm/dev/TVHS5Y4dNvM/11.png b/vlm/dev/TVHS5Y4dNvM/11.png new file mode 100644 index 0000000000000000000000000000000000000000..40ac69d025bad8344fcfdb25f7dd7c59114ff7b0 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cff62268a43e112b4f8e3fe41733f5c65bca13baede94827b3a9848caa10656b +size 336863 diff --git a/vlm/dev/TVHS5Y4dNvM/12.png b/vlm/dev/TVHS5Y4dNvM/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b14002ac6a13d3bec726c8979e1571e5acc98725 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c61ad9fc103eb9ea0cd8dbd3a5c729bb2babd4debb2a13ecad34c577d09fc0fb +size 2091281 diff --git a/vlm/dev/TVHS5Y4dNvM/13.png b/vlm/dev/TVHS5Y4dNvM/13.png new file mode 100644 index 0000000000000000000000000000000000000000..209eb65dd35552a9a7c9267b305e8dc5f0dcbbd2 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b40cb5fd22dbcff2f489cda21110106e27d840883f05257d0e531d88ec5ce183 +size 823576 diff --git a/vlm/dev/TVHS5Y4dNvM/14.png b/vlm/dev/TVHS5Y4dNvM/14.png new file mode 100644 index 0000000000000000000000000000000000000000..c83434c8031de2befbbbe7c7b2daf04fb92c2342 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d982c6990db1a1ea6ec41e732322083f30d929cdd6904866e30f59f3ed77066 +size 307727 diff --git a/vlm/dev/TVHS5Y4dNvM/2.png b/vlm/dev/TVHS5Y4dNvM/2.png new file mode 100644 index 0000000000000000000000000000000000000000..d1996d988fece7590fba5dc02f292b05083faffa --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36ab3b4cc029d5661a4979e012ea3b5786cd13921606271273e209b380ffa741 +size 598300 diff --git a/vlm/dev/TVHS5Y4dNvM/3.png b/vlm/dev/TVHS5Y4dNvM/3.png new file mode 100644 index 0000000000000000000000000000000000000000..1696734af0193bbf56180bf232dd89a3c7366d45 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc00acac0750b24a1f559b102a996ef233977df6185ec188103746d1d997cd0e +size 644661 diff --git a/vlm/dev/TVHS5Y4dNvM/4.png b/vlm/dev/TVHS5Y4dNvM/4.png new file mode 100644 index 0000000000000000000000000000000000000000..0be3e548118cf5014ec8a0140df6fc39130b4a11 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50c8c70856bdf76838e83b64c1d9b36a631a586ad7dce0eaeedd710aa1fe1065 +size 582545 diff --git a/vlm/dev/TVHS5Y4dNvM/5.png b/vlm/dev/TVHS5Y4dNvM/5.png new file mode 100644 index 0000000000000000000000000000000000000000..c8ebb22c1fedfb2edc95a9a77aca833e7dcab27c --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ca4a6f589412f3969a55be4323682a1573cb89019a5b7a37479b92a78796fc7 +size 556654 diff --git a/vlm/dev/TVHS5Y4dNvM/6.png b/vlm/dev/TVHS5Y4dNvM/6.png new file mode 100644 index 0000000000000000000000000000000000000000..edb8f0bc001f77bd69638be47dfaa839df306e5b --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13201960b0722b447cb6c238f0ea48c4e670d4c4eb2d6c176f08e9354db88afd +size 239972 diff --git a/vlm/dev/TVHS5Y4dNvM/7.png b/vlm/dev/TVHS5Y4dNvM/7.png new file mode 100644 index 0000000000000000000000000000000000000000..a8f1ffefe0ba301b39346e4d996b1141997f5771 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff0a962ae94e521afae7c1d0e6a72e86e4849738c0d8faa16bd5b6ebd771ffce +size 495838 diff --git a/vlm/dev/TVHS5Y4dNvM/8.png b/vlm/dev/TVHS5Y4dNvM/8.png new file mode 100644 index 0000000000000000000000000000000000000000..02dc34523521fa25e5e3ba42a61ae4d80b14c8ec --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5e43c3d7ee67e22cd013f0e7122840f76c9409acce6905805425ec2218486e1 +size 679519 diff --git a/vlm/dev/TVHS5Y4dNvM/9.png b/vlm/dev/TVHS5Y4dNvM/9.png new file mode 100644 index 0000000000000000000000000000000000000000..2a3f08cdd793001e9c0344a65c5f613f07429996 --- /dev/null +++ b/vlm/dev/TVHS5Y4dNvM/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9509e4a11ad96cdae945ebd7430bc56b32087051463c24030da625a428aa7632 +size 495197 diff --git a/vlm/dev/UaXD4Al3mdb/0.png b/vlm/dev/UaXD4Al3mdb/0.png new file mode 100644 index 0000000000000000000000000000000000000000..78b435ec314b8b511fff91f00b8ff0cd80f7e891 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61712f8ee4bf904c796f64f781bc3471485426036dc9f725eb7b2525516b31f2 +size 504518 diff --git a/vlm/dev/UaXD4Al3mdb/1.png b/vlm/dev/UaXD4Al3mdb/1.png new file mode 100644 index 0000000000000000000000000000000000000000..68bb9423c84e4eec229618f6e9dfb5b87f315a1c --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e3058caee58b1a5253c829e2104da2a868479f09fe9abc4019391d7efaa4947 +size 708495 diff --git a/vlm/dev/UaXD4Al3mdb/10.png b/vlm/dev/UaXD4Al3mdb/10.png new file mode 100644 index 0000000000000000000000000000000000000000..c24617567a9538130f887bb59e3c8317830b3bfd --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2925e959a46a6678b9c330783e2cb7d646ae79d029486e03cb78099668b5de5b +size 624329 diff --git a/vlm/dev/UaXD4Al3mdb/11.png b/vlm/dev/UaXD4Al3mdb/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ec16f071bc4f892c107d91eb79f5178ae9bdf2f5 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d899ea3c7b3a57f7206d325b27fe0414b0ef217141db129e88429be912ef06f5 +size 632609 diff --git a/vlm/dev/UaXD4Al3mdb/12.png b/vlm/dev/UaXD4Al3mdb/12.png new file mode 100644 index 0000000000000000000000000000000000000000..65ce84a783104f788c389a75873afbb85a932fc6 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c847453fcfe1c9777fe0a512e4a928af9c6b8e6abe43dd02f1d5172e405e573c +size 341463 diff --git a/vlm/dev/UaXD4Al3mdb/2.png b/vlm/dev/UaXD4Al3mdb/2.png new file mode 100644 index 0000000000000000000000000000000000000000..e5e59f7b012cb685d9ffc319b8c390f00d14e2a1 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d23f15328eecc505956d06f62eeaefb813950fd795aea9156ca19bf279b9cf93 +size 1630063 diff --git a/vlm/dev/UaXD4Al3mdb/3.png b/vlm/dev/UaXD4Al3mdb/3.png new file mode 100644 index 0000000000000000000000000000000000000000..0ef4420638c38aa5082b375b8823f8ba7e91ac59 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b251b2cd360affb5c50535c767b6ee0e2b7a48104a83a27bdd327f73a9281a8f +size 724774 diff --git a/vlm/dev/UaXD4Al3mdb/4.png b/vlm/dev/UaXD4Al3mdb/4.png new file mode 100644 index 0000000000000000000000000000000000000000..d9912c8cd42574ef7141cc158380a1446371cbe6 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b908e4a12bc3b80907204a1167a722803e2c170838034383f8f7127ca541730 +size 535791 diff --git a/vlm/dev/UaXD4Al3mdb/5.png b/vlm/dev/UaXD4Al3mdb/5.png new file mode 100644 index 0000000000000000000000000000000000000000..dff66788a068e3795290ee8524899a979d321953 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23c2b771422d00eaecdb38e7c93331142a4ac37234fc77b9a19df617ed72a7b7 +size 553355 diff --git a/vlm/dev/UaXD4Al3mdb/6.png b/vlm/dev/UaXD4Al3mdb/6.png new file mode 100644 index 0000000000000000000000000000000000000000..e0b26a4c75dc66e93a80b694e6e6d2ac2216fcdf --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f62858e7e56caea1ef688a4f008124e2fe9aea2941b52c81c6d606e7ae7c14e +size 583900 diff --git a/vlm/dev/UaXD4Al3mdb/7.png b/vlm/dev/UaXD4Al3mdb/7.png new file mode 100644 index 0000000000000000000000000000000000000000..69b1454c3dd1128454a38af542ba56dc5d25d349 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27d08e7dab9e8a6e98f8a76a5b5b38374fd69394d3edf3031248c39a13f9c8b9 +size 601144 diff --git a/vlm/dev/UaXD4Al3mdb/8.png b/vlm/dev/UaXD4Al3mdb/8.png new file mode 100644 index 0000000000000000000000000000000000000000..a06cce917f548debf1a101feb126ebfaee8f342b --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b782296eeccab062973c1bfd39435d4265192b79e1b4ccfad13075d05f150112 +size 530793 diff --git a/vlm/dev/UaXD4Al3mdb/9.png b/vlm/dev/UaXD4Al3mdb/9.png new file mode 100644 index 0000000000000000000000000000000000000000..481530e44237dad577de1db9d30383f650131755 --- /dev/null +++ b/vlm/dev/UaXD4Al3mdb/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4105fa067a8fc37a001c887a789b630b1eea484683a62ccf957f2398d270c6f +size 631866 diff --git a/vlm/dev/UmvSlP-PyV/0.png b/vlm/dev/UmvSlP-PyV/0.png new file mode 100644 index 0000000000000000000000000000000000000000..d20a4da4629c37d15264043450aa931e035ad270 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7402ca687d764d704c242755dadb284701e83ac8c26ea34da6621e3ad3f8445 +size 352657 diff --git a/vlm/dev/UmvSlP-PyV/1.png b/vlm/dev/UmvSlP-PyV/1.png new file mode 100644 index 0000000000000000000000000000000000000000..4375916afec36ab9bb23b0f7675ede800681983d --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02575d0a4b7d2bcbfeaf0113bba3f23b5d14d8c76b4b77ff5a20b77ea2c4a26f +size 688882 diff --git a/vlm/dev/UmvSlP-PyV/10.png b/vlm/dev/UmvSlP-PyV/10.png new file mode 100644 index 0000000000000000000000000000000000000000..e272fe478094cbb3ef40416ba0e943b914dfbbea --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18d287ed93e8fc677e6f4127d9fbd8a7ea47fa133d70c5be7c01c9fe468c40fd +size 612066 diff --git a/vlm/dev/UmvSlP-PyV/11.png b/vlm/dev/UmvSlP-PyV/11.png new file mode 100644 index 0000000000000000000000000000000000000000..2221afbaa7a5e94adda19d499e14186ed2ffade4 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dfa0b36ba90759c95febc40df6f4c1bfb4854e964aa4d98ea17de19d60e9767 +size 596064 diff --git a/vlm/dev/UmvSlP-PyV/12.png b/vlm/dev/UmvSlP-PyV/12.png new file mode 100644 index 0000000000000000000000000000000000000000..8192e24cf519e4da9772b938d76d7aaec9a1f122 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dd0d59bc0f01881c0925180a15ad6a0741180c38894f64de4f95e851f06b92b +size 331411 diff --git a/vlm/dev/UmvSlP-PyV/13.png b/vlm/dev/UmvSlP-PyV/13.png new file mode 100644 index 0000000000000000000000000000000000000000..c41b9e943b26f0efbcde72475a63a36e5e6d945f --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:555a264f11beb7d71b958eacff4dc384fdf1f07f8205f7ef316c80e62b0cea38 +size 384753 diff --git a/vlm/dev/UmvSlP-PyV/2.png b/vlm/dev/UmvSlP-PyV/2.png new file mode 100644 index 0000000000000000000000000000000000000000..b8c37deeb990f6169390b924f2c6e507436ef962 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d80a81b504b9a7cef85499f04debdcad425d3ff255ada391a7a8a07ee9c0cc4 +size 559681 diff --git a/vlm/dev/UmvSlP-PyV/3.png b/vlm/dev/UmvSlP-PyV/3.png new file mode 100644 index 0000000000000000000000000000000000000000..f53f9c02e2fd6f9def2909667730f60339c67bd4 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a444ba227ce09f43c8388d717e6bdbf6bbe2d3cbe4f6d6ed1ee0fda89aa86e4 +size 600265 diff --git a/vlm/dev/UmvSlP-PyV/4.png b/vlm/dev/UmvSlP-PyV/4.png new file mode 100644 index 0000000000000000000000000000000000000000..d45372ead782670666c785b604cbb6d3c51173d0 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da5ec66dd20bdc0272c74caddd0561aed4f7e51d674ecf69b2133bf06ab62e5a +size 629361 diff --git a/vlm/dev/UmvSlP-PyV/5.png b/vlm/dev/UmvSlP-PyV/5.png new file mode 100644 index 0000000000000000000000000000000000000000..8f63032dd6c82a49193383d3c14e0f8ff2e433f2 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b648ef4d51701527c34e0ee43e66680d28c15ab7970c5573ec0d0737d8924fb5 +size 682351 diff --git a/vlm/dev/UmvSlP-PyV/6.png b/vlm/dev/UmvSlP-PyV/6.png new file mode 100644 index 0000000000000000000000000000000000000000..73be2393b48cd911a8552adf44586e7cd2a16992 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cb477f7457279f62cf67b2b350fdb88c16da43d4d8f64d8bca93a6bb593c376 +size 745611 diff --git a/vlm/dev/UmvSlP-PyV/7.png b/vlm/dev/UmvSlP-PyV/7.png new file mode 100644 index 0000000000000000000000000000000000000000..123f5ce7171c625f660841c319bb48428b60a3d7 --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cc735b19a41b94eed691a2502c64629fd8ef3ee8a9451f044f3edf9c8db9976 +size 623923 diff --git a/vlm/dev/UmvSlP-PyV/8.png b/vlm/dev/UmvSlP-PyV/8.png new file mode 100644 index 0000000000000000000000000000000000000000..88b6220d21354d25b30d9cd5e3aabe1eb7e2263f --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfd0be84bb0d74b0a5971cacf5218e4bc03a6d5fb539c5799cc6f1350c081616 +size 860256 diff --git a/vlm/dev/UmvSlP-PyV/9.png b/vlm/dev/UmvSlP-PyV/9.png new file mode 100644 index 0000000000000000000000000000000000000000..2110a6371f58c2e8526dcb5e98c7d2e61359352c --- /dev/null +++ b/vlm/dev/UmvSlP-PyV/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be63bc7037614edee125aa100e8df1251687f2e223710749876f27c5673f0793 +size 581648 diff --git a/vlm/dev/Uynr3iPhksa/0.png b/vlm/dev/Uynr3iPhksa/0.png new file mode 100644 index 0000000000000000000000000000000000000000..71e76808b0e8936671558f7e8da1f9746080bb1b --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31ae81c89202467f341f17356e492d752ce31090e31b6b4ab1aed6961e009ab5 +size 442328 diff --git a/vlm/dev/Uynr3iPhksa/1.png b/vlm/dev/Uynr3iPhksa/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0ab4a53ebb1b39bc5a2788bb7dca02861d291fe9 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89a8a49aba049d4f98365813fa5c454648f8663514efb1b8fef71df0ddf3defd +size 584959 diff --git a/vlm/dev/Uynr3iPhksa/10.png b/vlm/dev/Uynr3iPhksa/10.png new file mode 100644 index 0000000000000000000000000000000000000000..54c6eb772e2c0a7cd5a6065f88e76732e70a20f2 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eee727312c6937d28474ebfd818bcfb310cada514b24ad0d9c40a016316ee5ee +size 539339 diff --git a/vlm/dev/Uynr3iPhksa/11.png b/vlm/dev/Uynr3iPhksa/11.png new file mode 100644 index 0000000000000000000000000000000000000000..b84b755e2fed09b9968efa223caedfd4de43134a --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6de57b1cc01286c00dc5bddd380205c8dc030c33a56e5c8f6bb63abb0efcb711 +size 372940 diff --git a/vlm/dev/Uynr3iPhksa/12.png b/vlm/dev/Uynr3iPhksa/12.png new file mode 100644 index 0000000000000000000000000000000000000000..a239677224b14ccb791e7765f2b2b9ef1e54bccd --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22661994a386eb8bb6ae2255f323ff0260b7d83da70546e7b277d9b983a97c9a +size 473290 diff --git a/vlm/dev/Uynr3iPhksa/2.png b/vlm/dev/Uynr3iPhksa/2.png new file mode 100644 index 0000000000000000000000000000000000000000..733d64f4ca8f291e68fc74d9faca8958e18d3725 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ef1212059c24356d137f5b6d329238f9ab1cfd979a1d7748ea6fb2d293bd4e5 +size 584129 diff --git a/vlm/dev/Uynr3iPhksa/3.png b/vlm/dev/Uynr3iPhksa/3.png new file mode 100644 index 0000000000000000000000000000000000000000..180b161aa61cc883e9d7040849754207ad40a4a2 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:355848bf1ecdd7ac851fb7d4efd435100456f4b6fe4659245fb7b7cdcecb90ac +size 564206 diff --git a/vlm/dev/Uynr3iPhksa/4.png b/vlm/dev/Uynr3iPhksa/4.png new file mode 100644 index 0000000000000000000000000000000000000000..6a92facc6e0a31ed12ee0ac9765e0f5a4a3104e4 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87d1f05423d19af570bdfb2544b3687736f088793ee75c3fa6b1ffc1a11d1eca +size 583373 diff --git a/vlm/dev/Uynr3iPhksa/5.png b/vlm/dev/Uynr3iPhksa/5.png new file mode 100644 index 0000000000000000000000000000000000000000..f11fb79ba739db6c42ce3ec0406e5ea290f3687b --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bc8187ac9831a10f4fe8dd7106b5ff6b5394d8e761ab1c908ed38c31602c474 +size 577897 diff --git a/vlm/dev/Uynr3iPhksa/6.png b/vlm/dev/Uynr3iPhksa/6.png new file mode 100644 index 0000000000000000000000000000000000000000..2c75bcde14cf59b050cf62256b7734d244e1a429 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e996d39cebe77e0dea90701c91290ca1fcc0240d539f0b07ad9832190bf31c45 +size 602946 diff --git a/vlm/dev/Uynr3iPhksa/7.png b/vlm/dev/Uynr3iPhksa/7.png new file mode 100644 index 0000000000000000000000000000000000000000..52ceb74316d511f104f4b5726f34670ff4ab46e1 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55784aa96c877cc20991bf0478a561b6afe35ebeedfc86dfe2a7280e3c890820 +size 588796 diff --git a/vlm/dev/Uynr3iPhksa/8.png b/vlm/dev/Uynr3iPhksa/8.png new file mode 100644 index 0000000000000000000000000000000000000000..97e99f575f501f2d92cf26e1ee1e6237b37d57e1 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e79c99aaa8e9cb77b3b64712dac53ced0fa2bca04799522280d9a89d7d4b1b0 +size 617107 diff --git a/vlm/dev/Uynr3iPhksa/9.png b/vlm/dev/Uynr3iPhksa/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0ecf8b41fdca614be5c880e6529aaac61ccd6fa9 --- /dev/null +++ b/vlm/dev/Uynr3iPhksa/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5588f0540d96f0269ada9a0556dac1c80f2942c1c01a8e2f6f8c94d8baf0098 +size 559664 diff --git a/vlm/dev/ZBESeIUB5k/0.png b/vlm/dev/ZBESeIUB5k/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6cbbf489ed9eadc648bfe4844ddb015a7853825b --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6e9aca6140468bda3b6f6980946836127234c308d907dc6c7245aea8c8874df +size 483200 diff --git a/vlm/dev/ZBESeIUB5k/1.png b/vlm/dev/ZBESeIUB5k/1.png new file mode 100644 index 0000000000000000000000000000000000000000..669410ee905b0c2e12dbdfbaf8cc36cb9039b9c3 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de8f44a22a2091907140fe8c6cde420b33125b054f9fcdaad900e50b992b9c7e +size 567586 diff --git a/vlm/dev/ZBESeIUB5k/10.png b/vlm/dev/ZBESeIUB5k/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2021ee29c05e837473b0d8f276eaef07015c7c49 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e96627126646b9fadb401d9c4fb7734a1238ae6029a85c7ac7b00917dfd58a1a +size 629815 diff --git a/vlm/dev/ZBESeIUB5k/11.png b/vlm/dev/ZBESeIUB5k/11.png new file mode 100644 index 0000000000000000000000000000000000000000..c46d5ec84aa9639a47db21f631f6d3bea6cdda70 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:362b2631c6aaf942d3c8353a98037d7571b79f3f839415369d39626826ecda45 +size 625618 diff --git a/vlm/dev/ZBESeIUB5k/12.png b/vlm/dev/ZBESeIUB5k/12.png new file mode 100644 index 0000000000000000000000000000000000000000..ce061fe3d2bc8ee90fd65b7c893731cbdf16c3f3 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4441dac197ac02304b418a10e2b1188134611925b820a414da1aab5de95d57 +size 571622 diff --git a/vlm/dev/ZBESeIUB5k/13.png b/vlm/dev/ZBESeIUB5k/13.png new file mode 100644 index 0000000000000000000000000000000000000000..f47413bd9876d7b3d9261b53957cf07f3e857712 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:858f067c429b1d4da14d0ceaa7c01416b05eb668d3fdfd66b5c083b537e8cf12 +size 582759 diff --git a/vlm/dev/ZBESeIUB5k/14.png b/vlm/dev/ZBESeIUB5k/14.png new file mode 100644 index 0000000000000000000000000000000000000000..2d76c8356d2d58aeaf70783daaa9d3f1a43f4f1b --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7181f995d433beb51c9e1c27429e2633a28f627c6e884e8c29f44a27fab5ed22 +size 607714 diff --git a/vlm/dev/ZBESeIUB5k/15.png b/vlm/dev/ZBESeIUB5k/15.png new file mode 100644 index 0000000000000000000000000000000000000000..267be600c51545220fa0a186e661ca2ca0d1a916 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1b0605a49c5550e5d14883f094992eca861450528fa0542f7971285bcb40dc8 +size 581172 diff --git a/vlm/dev/ZBESeIUB5k/16.png b/vlm/dev/ZBESeIUB5k/16.png new file mode 100644 index 0000000000000000000000000000000000000000..4009957bf418b7b84baabe0281722a95a7dc026f --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6acba8307ed80cb3a011ad0e39af9c44dd4629b7337a2edeae622c6f3b31b5e1 +size 490063 diff --git a/vlm/dev/ZBESeIUB5k/17.png b/vlm/dev/ZBESeIUB5k/17.png new file mode 100644 index 0000000000000000000000000000000000000000..a15b3e982c211e7c1fed1b7dfa6e37c9acf3b450 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cd3c668a22e0d75237cbadb2b49095b56ada7230aa6c722d07155b9909579e6 +size 465117 diff --git a/vlm/dev/ZBESeIUB5k/18.png b/vlm/dev/ZBESeIUB5k/18.png new file mode 100644 index 0000000000000000000000000000000000000000..f1204d54bcd89892021c49852d1344f118bd51a1 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65ebbda7995991d49c9120ae95815ad76fa51826af4054d842e1a81da518249c +size 487415 diff --git a/vlm/dev/ZBESeIUB5k/19.png b/vlm/dev/ZBESeIUB5k/19.png new file mode 100644 index 0000000000000000000000000000000000000000..13beffc04cb2adf3025ac725b6545730d9811f59 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:555160e7da6709b5e6cd9c90318e6023d8a96855ae7659e167048c93bb31529a +size 586355 diff --git a/vlm/dev/ZBESeIUB5k/2.png b/vlm/dev/ZBESeIUB5k/2.png new file mode 100644 index 0000000000000000000000000000000000000000..529268ec4d146971679a3b4b5cceaa5ebee22faa --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:405e8ba963255d305b1e87d49f8879dd59e25e4eed6d4727abb36c18bd685ab3 +size 634825 diff --git a/vlm/dev/ZBESeIUB5k/20.png b/vlm/dev/ZBESeIUB5k/20.png new file mode 100644 index 0000000000000000000000000000000000000000..b566a0ff65e64ad337c2a824e0d67b76b52d9357 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6c9f8970a2b00a15fda2d324d008e9482a314efb361022c395adcc00ee8a639 +size 552533 diff --git a/vlm/dev/ZBESeIUB5k/21.png b/vlm/dev/ZBESeIUB5k/21.png new file mode 100644 index 0000000000000000000000000000000000000000..7ffb05b0ce71cf8522ed70d5b680d53627b21843 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57d8d20e06f7450a3131cca65d3e0f455cf833763e85f6b95a8eb1ee7de89e7a +size 579019 diff --git a/vlm/dev/ZBESeIUB5k/22.png b/vlm/dev/ZBESeIUB5k/22.png new file mode 100644 index 0000000000000000000000000000000000000000..9d2aea3fd8c64b9830d659d35d810f206bc2c5d9 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bfefabf453c9951480da51d4735ecd259bdefcda196990848ca5a826128b413 +size 447495 diff --git a/vlm/dev/ZBESeIUB5k/23.png b/vlm/dev/ZBESeIUB5k/23.png new file mode 100644 index 0000000000000000000000000000000000000000..1d71170d4436a5cf919b832a6fe97865ad640a18 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a5f5e00443a80a99a67832ac52b03dede9daaf1cff6eba9b3a111305d14307a +size 276482 diff --git a/vlm/dev/ZBESeIUB5k/3.png b/vlm/dev/ZBESeIUB5k/3.png new file mode 100644 index 0000000000000000000000000000000000000000..0f09b9ce5440ff355f1b97bc238754702b95b12f --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1481bd244923e4fe6914570ba96b8c5b0b684d180289db97fc8a42ee9bc27f3e +size 601813 diff --git a/vlm/dev/ZBESeIUB5k/4.png b/vlm/dev/ZBESeIUB5k/4.png new file mode 100644 index 0000000000000000000000000000000000000000..cf95e345196085de81687a9096245d3f0dc26549 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:062a0069d4e5a434d0fe32edc72cdbaf648866fa857f142d111cbe5739bd78cc +size 590082 diff --git a/vlm/dev/ZBESeIUB5k/5.png b/vlm/dev/ZBESeIUB5k/5.png new file mode 100644 index 0000000000000000000000000000000000000000..92df9ad228782680d6f5527ecbcf9949a47a7953 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38fa47da237f2507dddcfba28078b845284b09e2a51e2de55d4788757b0eb3bc +size 587668 diff --git a/vlm/dev/ZBESeIUB5k/6.png b/vlm/dev/ZBESeIUB5k/6.png new file mode 100644 index 0000000000000000000000000000000000000000..06fee40cfdb64f668a0f11bb0b46cf9875be8fdd --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac376fe446c335d2ad964a9bdaa9cb74a9b3331dc4454ccb55240cdda45fad40 +size 544341 diff --git a/vlm/dev/ZBESeIUB5k/7.png b/vlm/dev/ZBESeIUB5k/7.png new file mode 100644 index 0000000000000000000000000000000000000000..d1db18bacc718d3851aec841d3e76e570ae5884d --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95b77d0be2dc2028233a5d9a97ddcf70bf222d7f797b611c4cd4cebc8334dec8 +size 613380 diff --git a/vlm/dev/ZBESeIUB5k/8.png b/vlm/dev/ZBESeIUB5k/8.png new file mode 100644 index 0000000000000000000000000000000000000000..14f07e5925704a5242afb6a61f046bb5e0a33ba5 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:104a0714a3e07aae6326c477529d9a6697051c292ff95115228cbcb00522da93 +size 561304 diff --git a/vlm/dev/ZBESeIUB5k/9.png b/vlm/dev/ZBESeIUB5k/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e007010875043d05050befec24f9183d4b4969f5 --- /dev/null +++ b/vlm/dev/ZBESeIUB5k/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a77c590638346b8d16dad0c144b803a5efa86151a933d9868644dd6aff21eb15 +size 536206 diff --git a/vlm/dev/a6NvoZ5DLoe/0.png b/vlm/dev/a6NvoZ5DLoe/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c85169bea1d6bebbb893b303b2356b7a1bd19b28 --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c56b212c551f5adcae5fdfc596e779a7f762b777d3761622a1bba2ba3539929 +size 500876 diff --git a/vlm/dev/a6NvoZ5DLoe/1.png b/vlm/dev/a6NvoZ5DLoe/1.png new file mode 100644 index 0000000000000000000000000000000000000000..363e13bbc8a969e8685bf255cb1420655b700e38 --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e475c79a090962fd6ec5f485421728f133ff9952fa264e637f2d06f65421828 +size 557200 diff --git a/vlm/dev/a6NvoZ5DLoe/10.png b/vlm/dev/a6NvoZ5DLoe/10.png new file mode 100644 index 0000000000000000000000000000000000000000..04b455994107261fc57a56b2aa18672b0f1c9e5f --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38321077c3bfa94204a3c59213ac6e07147f57c7d8eb1d1a95e926bfe7b789d1 +size 553137 diff --git a/vlm/dev/a6NvoZ5DLoe/11.png b/vlm/dev/a6NvoZ5DLoe/11.png new file mode 100644 index 0000000000000000000000000000000000000000..1d58372fa239c047c98bddc76c9d1a9ab88da208 --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14404df222fb467174c47aa438d91a0f8d053b27353d8be9c0feb1e8dda2f0b4 +size 491518 diff --git a/vlm/dev/a6NvoZ5DLoe/12.png b/vlm/dev/a6NvoZ5DLoe/12.png new file mode 100644 index 0000000000000000000000000000000000000000..5730571e423d1d281cef37bc65c92dcab6b60aee --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc4696c090d41c9fe23051683428ee524d4ea371ca2371d259d4da0146c56604 +size 358275 diff --git a/vlm/dev/a6NvoZ5DLoe/13.png b/vlm/dev/a6NvoZ5DLoe/13.png new file mode 100644 index 0000000000000000000000000000000000000000..2873d0d56ef5322819204c5c8018e9ca44a85c17 --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b63990b4391de9be57382a1a27df36ce54ebea60ac18084adf2bd72e8213d0bc +size 175290 diff --git a/vlm/dev/a6NvoZ5DLoe/2.png b/vlm/dev/a6NvoZ5DLoe/2.png new file mode 100644 index 0000000000000000000000000000000000000000..8f92cb8c24704fa5b7edae46a9edea6d52e87c9c --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84ca66fbf4a70881754ba1dfe8ade7df71f8e71e47152caa577092f28690e57b +size 596126 diff --git a/vlm/dev/a6NvoZ5DLoe/3.png b/vlm/dev/a6NvoZ5DLoe/3.png new file mode 100644 index 0000000000000000000000000000000000000000..4cf976d8afc2cf2a26519dee09f42a8ca187eafe --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f491f4ae6ffb9d1057ea4e4abb41ca67c306065863bccfac257c27c2e80b0fa +size 489716 diff --git a/vlm/dev/a6NvoZ5DLoe/4.png b/vlm/dev/a6NvoZ5DLoe/4.png new file mode 100644 index 0000000000000000000000000000000000000000..1c2e0ae02686e5d59bc3266873b01a784766c195 --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbd1f8dbeb685df10b8c5936e2e7b8acbd71c86baa7aed1168ebfdb3c6e71725 +size 548583 diff --git a/vlm/dev/a6NvoZ5DLoe/5.png b/vlm/dev/a6NvoZ5DLoe/5.png new file mode 100644 index 0000000000000000000000000000000000000000..1aea81881b8948658f9c8d38c4c2bf9a781eca83 --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:187529f31859b8b21beeddf927c046b967ecd7ba64e01dc0c453b5a3c8c1b715 +size 521582 diff --git a/vlm/dev/a6NvoZ5DLoe/6.png b/vlm/dev/a6NvoZ5DLoe/6.png new file mode 100644 index 0000000000000000000000000000000000000000..4bda26338568fc8af38d663b4b4d5018c89c21be --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a01fff6a822764230af0cc36f71a1cd6794fbc9390c736b4c21c433de3cfe101 +size 518974 diff --git a/vlm/dev/a6NvoZ5DLoe/7.png b/vlm/dev/a6NvoZ5DLoe/7.png new file mode 100644 index 0000000000000000000000000000000000000000..7148d6400698068a087752c28a8388cd0aff3f2e --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a47739c3b544252f1b501151eaab71f2cd42c7a312b13c418b778d9815341544 +size 566885 diff --git a/vlm/dev/a6NvoZ5DLoe/8.png b/vlm/dev/a6NvoZ5DLoe/8.png new file mode 100644 index 0000000000000000000000000000000000000000..2e2903d571d55cabc649d0dcd08e231abb690cad --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06fad3cf641931cf2c2a6f0c09a4ff3fc82f7d855184e50447505f238cfaedcf +size 571488 diff --git a/vlm/dev/a6NvoZ5DLoe/9.png b/vlm/dev/a6NvoZ5DLoe/9.png new file mode 100644 index 0000000000000000000000000000000000000000..320a10b36fcfcf5e8e0e65f98184fdbeacc3dc9a --- /dev/null +++ b/vlm/dev/a6NvoZ5DLoe/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d93634d6084d2a55457b740daaf8b90fd3f1b8245fd4868f1c9f385c45e07d99 +size 536436 diff --git a/vlm/dev/aa8KsqfTPa/0.png b/vlm/dev/aa8KsqfTPa/0.png new file mode 100644 index 0000000000000000000000000000000000000000..166deeccb9a5638294ab3cf0f4617b397cae9d70 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5be6f8c1f3ef6947fbcd4c3ce4bfb31e5470b4325425bd7fd2d1616b3ae0b7d1 +size 480633 diff --git a/vlm/dev/aa8KsqfTPa/1.png b/vlm/dev/aa8KsqfTPa/1.png new file mode 100644 index 0000000000000000000000000000000000000000..6ae45ea3be4f12b3ac27b2910a4e9be195bd3ef6 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd5616005bef713d9ae1b13a3423456b0a21adcccb2315c3a1f7e0674e0e3e65 +size 591900 diff --git a/vlm/dev/aa8KsqfTPa/10.png b/vlm/dev/aa8KsqfTPa/10.png new file mode 100644 index 0000000000000000000000000000000000000000..3d03e69ed585e7a1927670822bbee44e5fa9f715 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1774d1a59b224cca7aead1b63675503ee9c96c48e96ad23d726a601ec295abf0 +size 513064 diff --git a/vlm/dev/aa8KsqfTPa/11.png b/vlm/dev/aa8KsqfTPa/11.png new file mode 100644 index 0000000000000000000000000000000000000000..40e83569b63e3dfa814c693690f8e904c622add2 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c42fe8bc7ae224a8030273dbd303296848f1ebafae341d2c8e1dc98b42af022 +size 535394 diff --git a/vlm/dev/aa8KsqfTPa/12.png b/vlm/dev/aa8KsqfTPa/12.png new file mode 100644 index 0000000000000000000000000000000000000000..0d55851b10438e35ee77a16d13b3f4ceca65485c --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a678896f15ee266965da22ffacacdb78ebc40636a695ebc5466f6bb79a36d554 +size 233437 diff --git a/vlm/dev/aa8KsqfTPa/13.png b/vlm/dev/aa8KsqfTPa/13.png new file mode 100644 index 0000000000000000000000000000000000000000..1aa8b6f5b59592f2a7f33bd7db8a62d86c967a21 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ceda1f14273b7f1430181231e6feb9acbbc60610505a7e0598cd5d231799407 +size 205361 diff --git a/vlm/dev/aa8KsqfTPa/14.png b/vlm/dev/aa8KsqfTPa/14.png new file mode 100644 index 0000000000000000000000000000000000000000..c6ffbf85f96b3f96ab11706c037f86252e5ca4f9 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9223434cccbbbfa243abb0a670ad05294e56ccb59d6f490d0441eb33eaaf8d35 +size 91814 diff --git a/vlm/dev/aa8KsqfTPa/2.png b/vlm/dev/aa8KsqfTPa/2.png new file mode 100644 index 0000000000000000000000000000000000000000..17535259e93da17164f0d4b2c58da160b4506ea9 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7546a3de5b65f05464f91d57920af753149da2d2641dc129c454b9dd577f85a +size 609910 diff --git a/vlm/dev/aa8KsqfTPa/3.png b/vlm/dev/aa8KsqfTPa/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b1f7034f5bbae7ebcefebba0a1e6147154fddb93 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a08b399c49d7ac72ab054b7b1966835efb0736220fc4410f8fb12308a53af8d +size 512006 diff --git a/vlm/dev/aa8KsqfTPa/4.png b/vlm/dev/aa8KsqfTPa/4.png new file mode 100644 index 0000000000000000000000000000000000000000..c4e02058c3a9b8bfe302427cee955155bd7a12b2 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f30c3038e83eb987fb4d8fec5111be9566b8c841e956c3ba0cfa8394b2e1ad2 +size 533117 diff --git a/vlm/dev/aa8KsqfTPa/5.png b/vlm/dev/aa8KsqfTPa/5.png new file mode 100644 index 0000000000000000000000000000000000000000..b01c3ef256e72d22c61fba977b5e1e071bb41421 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f893ab665f3f2da9d23a88e1856676936734fdc1fd906846b76b468b2805cbc +size 479759 diff --git a/vlm/dev/aa8KsqfTPa/6.png b/vlm/dev/aa8KsqfTPa/6.png new file mode 100644 index 0000000000000000000000000000000000000000..e0ff267d9ebd5dc49ba613775d6c8076c558c65e --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed9188a0144ed0a6020e982d50c28bca7df2261be229419e52082f1247fee29b +size 511408 diff --git a/vlm/dev/aa8KsqfTPa/7.png b/vlm/dev/aa8KsqfTPa/7.png new file mode 100644 index 0000000000000000000000000000000000000000..34a1f2266a2dd0106981f3b517f1e0e346e04e66 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e158b4d1ea9ae914fbf38a737ddb41626d2ccbe974ef2df5b0459b09c03adf1 +size 517577 diff --git a/vlm/dev/aa8KsqfTPa/8.png b/vlm/dev/aa8KsqfTPa/8.png new file mode 100644 index 0000000000000000000000000000000000000000..2a4ef521d1dc9713e3a03ad94922343869476756 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf63c85061e2907de78f371c820b6732708581623171c149345d4c205c64534f +size 639912 diff --git a/vlm/dev/aa8KsqfTPa/9.png b/vlm/dev/aa8KsqfTPa/9.png new file mode 100644 index 0000000000000000000000000000000000000000..91e0cd37773666aa9e45f72504b56e528ecc1c20 --- /dev/null +++ b/vlm/dev/aa8KsqfTPa/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d22f5de6a58435cdd8ec6043258f1666a0abd68cd7487c946d5ebbc187b70a79 +size 545272 diff --git a/vlm/dev/bdHkMjBJG_w/0.png b/vlm/dev/bdHkMjBJG_w/0.png new file mode 100644 index 0000000000000000000000000000000000000000..bf564c02fb421573140b310b3ac59e52a0f43814 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:665b166190c6bbc23a0516a1193097de2820c09523d545f4fa5c989e073361ae +size 654296 diff --git a/vlm/dev/bdHkMjBJG_w/1.png b/vlm/dev/bdHkMjBJG_w/1.png new file mode 100644 index 0000000000000000000000000000000000000000..1c5309028b6de8e0298667af29c59f189c19567a --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:274a80b31bf945d0562e07166821d6f5d6cedffc08f1bb2497d1adfdc2ae255f +size 625502 diff --git a/vlm/dev/bdHkMjBJG_w/10.png b/vlm/dev/bdHkMjBJG_w/10.png new file mode 100644 index 0000000000000000000000000000000000000000..958b8f6accf7a06a4b44f6f8ff410c9416ee4512 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86fa3e6571d2fe5dd39d35ff8c47ced48d3ab73481cb68899cd341cf2640d060 +size 506952 diff --git a/vlm/dev/bdHkMjBJG_w/11.png b/vlm/dev/bdHkMjBJG_w/11.png new file mode 100644 index 0000000000000000000000000000000000000000..cd5a88b22613e5b9e03ea6c107481ddc335453e5 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50cf3c8b8ed1727caba013c6de58e14750338d69458754c5ec5e4d213c616ac2 +size 521534 diff --git a/vlm/dev/bdHkMjBJG_w/12.png b/vlm/dev/bdHkMjBJG_w/12.png new file mode 100644 index 0000000000000000000000000000000000000000..3b471fd6f7a0fd477847ccedf937ffd88bf208b0 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:262792b9b99c3f6b0b72be04ed03f6107547e05255e4823f7c5b451e9645dd75 +size 544219 diff --git a/vlm/dev/bdHkMjBJG_w/13.png b/vlm/dev/bdHkMjBJG_w/13.png new file mode 100644 index 0000000000000000000000000000000000000000..073f7610eeb6fd7e249b516e1b762abb049ef239 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d271dee3d937aa3befbcebacc2cd388ef29536390605f488e982d00ed7e7d2f +size 523015 diff --git a/vlm/dev/bdHkMjBJG_w/14.png b/vlm/dev/bdHkMjBJG_w/14.png new file mode 100644 index 0000000000000000000000000000000000000000..4e8bd1e3a524bd889b1b8d469562d836f70ea077 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eac9252f85aa32db1c49ca49f64203ed664834cfd2499c938d1b5bacbe7a3fc7 +size 549562 diff --git a/vlm/dev/bdHkMjBJG_w/15.png b/vlm/dev/bdHkMjBJG_w/15.png new file mode 100644 index 0000000000000000000000000000000000000000..c54255779401d6d0ab5cc3fdb546c9eb4031e730 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:217cafe99c966ba76a044f2eebf2666c50dead5b0d3f65a09b20b7062463fc3f +size 582159 diff --git a/vlm/dev/bdHkMjBJG_w/16.png b/vlm/dev/bdHkMjBJG_w/16.png new file mode 100644 index 0000000000000000000000000000000000000000..aa673f38f607b8b638bb7c0ba178fba449b2d9a7 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6be93581521095b9891c6222df8bc24c1d19c497302ca9b87608344328ba733f +size 537924 diff --git a/vlm/dev/bdHkMjBJG_w/17.png b/vlm/dev/bdHkMjBJG_w/17.png new file mode 100644 index 0000000000000000000000000000000000000000..f00a483cf012d35fd4598dd0ab6f427e8dc3e600 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:433fd90704ce8cda286e671b168ba923bf4e921f4647f1177d3d9452dc558a0f +size 497396 diff --git a/vlm/dev/bdHkMjBJG_w/18.png b/vlm/dev/bdHkMjBJG_w/18.png new file mode 100644 index 0000000000000000000000000000000000000000..4546a50cc636b30507e0e560b5113ca5c8e98012 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4648afa552d3caa338c094317de791b408f682a019799ee50949a50a952df141 +size 786103 diff --git a/vlm/dev/bdHkMjBJG_w/19.png b/vlm/dev/bdHkMjBJG_w/19.png new file mode 100644 index 0000000000000000000000000000000000000000..e1744d6e8568ae85adbd000c9b3b5a81aad6f620 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:930177ac13992a75f90dd7a4a1533bda0ca4d56a1a7beea61287f8545d4a7b38 +size 554015 diff --git a/vlm/dev/bdHkMjBJG_w/2.png b/vlm/dev/bdHkMjBJG_w/2.png new file mode 100644 index 0000000000000000000000000000000000000000..fd502f33083045695c4ac30f958d40a922335f1d --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f453c340752470f979f53f397f32b16d3d68f522f94dfd1f33ec07693b0f3b1 +size 625263 diff --git a/vlm/dev/bdHkMjBJG_w/20.png b/vlm/dev/bdHkMjBJG_w/20.png new file mode 100644 index 0000000000000000000000000000000000000000..3e40038470a3b6d6e8f5c91fe0107f36dcc10677 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3714b3c770ad0f7a3f4023f6e8aa3d25227ca38b6a5c8274d26a616385f88b1f +size 526881 diff --git a/vlm/dev/bdHkMjBJG_w/21.png b/vlm/dev/bdHkMjBJG_w/21.png new file mode 100644 index 0000000000000000000000000000000000000000..7449709b6f223cc5303b4a3fd0a2d25c35fe148d --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e08ed6ed475012663605289956010be0555c2aab20f2d8f3954223b3e4ad782f +size 448202 diff --git a/vlm/dev/bdHkMjBJG_w/22.png b/vlm/dev/bdHkMjBJG_w/22.png new file mode 100644 index 0000000000000000000000000000000000000000..86daa1a1bfb111b8fff258119d9dcdd5cb2bdaa9 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c46e01e309b94bebed22f61590a762014e23e84251c7413c666acbad0e2845e7 +size 499842 diff --git a/vlm/dev/bdHkMjBJG_w/23.png b/vlm/dev/bdHkMjBJG_w/23.png new file mode 100644 index 0000000000000000000000000000000000000000..a98ad995a618329e62a07f51f149d01b8397eea1 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:537b310ad398c0d0129d8163830410f3f0f54cda355dbca0da5facdf759cf757 +size 427543 diff --git a/vlm/dev/bdHkMjBJG_w/24.png b/vlm/dev/bdHkMjBJG_w/24.png new file mode 100644 index 0000000000000000000000000000000000000000..d1ed52294a0f4d7ce9a02bf25352780dd5929497 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf5d5a8daf95cac1a816c6d54abd18eac4794a3126c332b09ab8de38ca6a303b +size 351471 diff --git a/vlm/dev/bdHkMjBJG_w/25.png b/vlm/dev/bdHkMjBJG_w/25.png new file mode 100644 index 0000000000000000000000000000000000000000..b76a7d115c28e3f8e8cc927bd8e66eaba175ec9b --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00a7bbe551cf3cdc6b5064fdbd76abcd351dd2c4fb70ca44cc73bdb43e569476 +size 385643 diff --git a/vlm/dev/bdHkMjBJG_w/26.png b/vlm/dev/bdHkMjBJG_w/26.png new file mode 100644 index 0000000000000000000000000000000000000000..bb0823da870b52b00867cc0dec6430410a0684ed --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad2d74c2510c5fbb0da60c67e48c6a7fcaa6908218785b929dae657c2080a598 +size 435023 diff --git a/vlm/dev/bdHkMjBJG_w/27.png b/vlm/dev/bdHkMjBJG_w/27.png new file mode 100644 index 0000000000000000000000000000000000000000..5bbd82412c7d2c5dc3e285f2c66cf24bde138b13 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a9976ea8b824bb6f83c91bd4c44e485ba4d630f9268fb5718a3e6b9a962fc6f +size 402083 diff --git a/vlm/dev/bdHkMjBJG_w/28.png b/vlm/dev/bdHkMjBJG_w/28.png new file mode 100644 index 0000000000000000000000000000000000000000..a4de93feb9ef9b53593e3c028c50c8a374231b92 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fcb6de872fe0d806ccab2cfe0ddf0a56f462d28e98b871f17e90d919e237fc3 +size 871697 diff --git a/vlm/dev/bdHkMjBJG_w/29.png b/vlm/dev/bdHkMjBJG_w/29.png new file mode 100644 index 0000000000000000000000000000000000000000..bfc3454d4322bd1d503156cfa22d5259dc162b58 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf1772a536b3a95196593d80040544f5de376a672c4afac46142ce2a94cd57e9 +size 941778 diff --git a/vlm/dev/bdHkMjBJG_w/3.png b/vlm/dev/bdHkMjBJG_w/3.png new file mode 100644 index 0000000000000000000000000000000000000000..4b7cb8c4348293ae615848cd85fc1439ca1f38a8 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:763ba1b8d23217c1e38ec14934e48dfd18b50ff47cd76c9339d21093c4ba4951 +size 690088 diff --git a/vlm/dev/bdHkMjBJG_w/30.png b/vlm/dev/bdHkMjBJG_w/30.png new file mode 100644 index 0000000000000000000000000000000000000000..7dcf2a75e1e49a613d5ec61f882dd8b78fe0c4fb --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/30.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5f38b1c3e69762f215a0528fc0b9ddc27f86e48cd90aff1c5ad3ca4cc68aa73 +size 895710 diff --git a/vlm/dev/bdHkMjBJG_w/31.png b/vlm/dev/bdHkMjBJG_w/31.png new file mode 100644 index 0000000000000000000000000000000000000000..8078e2c1c58d2dff2e7e621e689f4d1126d74d26 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/31.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf60c4c6284f7bd27b83cff32d09f8c6065ac16bc4fbb97c39dd0cc317080820 +size 602046 diff --git a/vlm/dev/bdHkMjBJG_w/4.png b/vlm/dev/bdHkMjBJG_w/4.png new file mode 100644 index 0000000000000000000000000000000000000000..eda96c8106459786025bded53bf25b8eb51e25c8 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e05ace25c3e17d53d27ff793f9c96b802454ec336af2f9b4506d97629bf2fd6 +size 820778 diff --git a/vlm/dev/bdHkMjBJG_w/5.png b/vlm/dev/bdHkMjBJG_w/5.png new file mode 100644 index 0000000000000000000000000000000000000000..9b8df702ebab828f0aa149c3af20a7617fbda71e --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ca20c3c49c30ad2dc624595c369cb0ec6b6a710685df1fc5d1e979e898bc595 +size 1155287 diff --git a/vlm/dev/bdHkMjBJG_w/6.png b/vlm/dev/bdHkMjBJG_w/6.png new file mode 100644 index 0000000000000000000000000000000000000000..3e89a87d5125e74033b34f4c0024b9e7e7c9eb15 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a286cd9fb2d6cb63338770946261fdad933685e97eadc9f88b2d49b25d698e85 +size 629421 diff --git a/vlm/dev/bdHkMjBJG_w/7.png b/vlm/dev/bdHkMjBJG_w/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c5c18e0da4825c238859d2ebef24a06c8977cb8d --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de10b38aa413fcf144743e59f9c1c77171f10b8575d8fe1db4887098b31a126e +size 637477 diff --git a/vlm/dev/bdHkMjBJG_w/8.png b/vlm/dev/bdHkMjBJG_w/8.png new file mode 100644 index 0000000000000000000000000000000000000000..2c44ac7d8ce4bc5c612fc9f90efebdf5791cf8d0 --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:baf0fbb69a5abba19ce2fc4fd3689e3efc5fdde8902affe3facec870dd5e38e2 +size 534772 diff --git a/vlm/dev/bdHkMjBJG_w/9.png b/vlm/dev/bdHkMjBJG_w/9.png new file mode 100644 index 0000000000000000000000000000000000000000..65c621e0fc4b8b0033355e68bb3317d77121713b --- /dev/null +++ b/vlm/dev/bdHkMjBJG_w/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee8118436a3fa8ed8b982975303b9d229d2dc4781ae6040c83e9f58319ab7530 +size 537119 diff --git a/vlm/dev/c5Inzw6giM/0.png b/vlm/dev/c5Inzw6giM/0.png new file mode 100644 index 0000000000000000000000000000000000000000..5980bdd046a061f7acf0e9377ca60df90f83e57c --- /dev/null +++ b/vlm/dev/c5Inzw6giM/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fcc3ac70fd15f36b69d6326bbf5ea5b9eb82e806e03f411cb7e2d0c608776be +size 424895 diff --git a/vlm/dev/c5Inzw6giM/1.png b/vlm/dev/c5Inzw6giM/1.png new file mode 100644 index 0000000000000000000000000000000000000000..05b6871fb4308bee361be59e81fcfc9f88c25973 --- /dev/null +++ b/vlm/dev/c5Inzw6giM/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c49814ee63abc199367255037e277195655c3b29d46f825798be274981b8f842 +size 510764 diff --git a/vlm/dev/c5Inzw6giM/2.png b/vlm/dev/c5Inzw6giM/2.png new file mode 100644 index 0000000000000000000000000000000000000000..eb7163a1c2652da7525c36a901ef7a2721863c72 --- /dev/null +++ b/vlm/dev/c5Inzw6giM/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7cb56fdbb6ea2b6899f6c3bbfe2b242cfc97feaa5b99f949353402795d3b6ad +size 443000 diff --git a/vlm/dev/c5Inzw6giM/3.png b/vlm/dev/c5Inzw6giM/3.png new file mode 100644 index 0000000000000000000000000000000000000000..f7747cd5c6b6fa13638736044d71667b43a9c811 --- /dev/null +++ b/vlm/dev/c5Inzw6giM/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c459d1592cad3ff877ce2a49984c612698d847aa0b15348cf3de89528110e72 +size 319896 diff --git a/vlm/dev/c5Inzw6giM/4.png b/vlm/dev/c5Inzw6giM/4.png new file mode 100644 index 0000000000000000000000000000000000000000..8a72dbbd32add99c6315236252761aaf16118273 --- /dev/null +++ b/vlm/dev/c5Inzw6giM/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68b5782623bb38290bcd0d90b69fc699165c7febe8192a09406dd9ca8d13b702 +size 480297 diff --git a/vlm/dev/c5Inzw6giM/5.png b/vlm/dev/c5Inzw6giM/5.png new file mode 100644 index 0000000000000000000000000000000000000000..384b1fe699a15853d7c3591bb22ddedfcf202bea --- /dev/null +++ b/vlm/dev/c5Inzw6giM/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e91ae52c83e4481572f8790c6e9c87b0d9d04beb6ba1a465cfb5990edd944579 +size 373558 diff --git a/vlm/dev/c5Inzw6giM/6.png b/vlm/dev/c5Inzw6giM/6.png new file mode 100644 index 0000000000000000000000000000000000000000..6468d6e3a01ff6d30a734ed081d8ea3056d9d395 --- /dev/null +++ b/vlm/dev/c5Inzw6giM/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f05f3d4a7c92fe7de11fd316251363013f6968c4097fe1783829deaab2407ada +size 341484 diff --git a/vlm/dev/c5Inzw6giM/7.png b/vlm/dev/c5Inzw6giM/7.png new file mode 100644 index 0000000000000000000000000000000000000000..46144d7f0be4733324b578dbc6fb3a4a7325c8c4 --- /dev/null +++ b/vlm/dev/c5Inzw6giM/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:441b2871d9ca679de63d17b2bd820577b42b5948db92e76834bbc7661d68dc17 +size 344836 diff --git a/vlm/dev/c5Inzw6giM/8.png b/vlm/dev/c5Inzw6giM/8.png new file mode 100644 index 0000000000000000000000000000000000000000..731dee8e71be72ef21d94003404541bcfee5edad --- /dev/null +++ b/vlm/dev/c5Inzw6giM/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcdb88b4ab6754eb83bda7f8d65970138fbaa7e34d78d31812d3858b2d643d8a +size 513051 diff --git a/vlm/dev/c5Inzw6giM/9.png b/vlm/dev/c5Inzw6giM/9.png new file mode 100644 index 0000000000000000000000000000000000000000..114ffc3c2f76603ccabae83b431a38d318c853e6 --- /dev/null +++ b/vlm/dev/c5Inzw6giM/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:283ce2cc52921a2efde80eb835235b4c4146e724a0f601d073633a2535efaae3 +size 512762 diff --git a/vlm/dev/gUL6zYN4Uaf/0.png b/vlm/dev/gUL6zYN4Uaf/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6f88809d2538eb45366ab231507f48a8fbc26f5b --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bcf0b296be9810ef0728261ba7566220bccc4a7fecd087fca0847f4c60ba74c +size 526846 diff --git a/vlm/dev/gUL6zYN4Uaf/1.png b/vlm/dev/gUL6zYN4Uaf/1.png new file mode 100644 index 0000000000000000000000000000000000000000..5e85d16e3864f385248409804a4018a889410318 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4dcb054ffd60be39d81822833d913acf4f588d3cf254b77072c5e4db22d70cc +size 590066 diff --git a/vlm/dev/gUL6zYN4Uaf/11.png b/vlm/dev/gUL6zYN4Uaf/11.png new file mode 100644 index 0000000000000000000000000000000000000000..0c30566fb2ce2a5f89134aa98e4185a593df11da --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f53912b2f827782073c9e728c67ead641ec7287ace3b826406470e1c1e67fea8 +size 606528 diff --git a/vlm/dev/gUL6zYN4Uaf/12.png b/vlm/dev/gUL6zYN4Uaf/12.png new file mode 100644 index 0000000000000000000000000000000000000000..e3c808093ec8754368e5d94e806ed3b188a10ad7 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81c6707d694f2a8383c4679b8f743557e053029a88a275872af761a99f6f3ac6 +size 631964 diff --git a/vlm/dev/gUL6zYN4Uaf/13.png b/vlm/dev/gUL6zYN4Uaf/13.png new file mode 100644 index 0000000000000000000000000000000000000000..e3f7399009f0c9e40990ce44090d054bcf196823 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c90264f50e176e47f2c125d8129456e8023556110021320411d91bbafaaa07bf +size 611033 diff --git a/vlm/dev/gUL6zYN4Uaf/14.png b/vlm/dev/gUL6zYN4Uaf/14.png new file mode 100644 index 0000000000000000000000000000000000000000..9c14add5fc95eb8b4458401ee4f1f5fc4c8df8b6 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b477c1b2c0aa7d89cd35d7af27f3a8154188e1419557a35232f7305fee813a61 +size 602033 diff --git a/vlm/dev/gUL6zYN4Uaf/15.png b/vlm/dev/gUL6zYN4Uaf/15.png new file mode 100644 index 0000000000000000000000000000000000000000..bb908b7393244b30cee7c04d985310e78add6dbb --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f3dd9f1e9113947f8e02a73625f7da6f65f23d490ff953dc4fef8625f77efd8 +size 662269 diff --git a/vlm/dev/gUL6zYN4Uaf/16.png b/vlm/dev/gUL6zYN4Uaf/16.png new file mode 100644 index 0000000000000000000000000000000000000000..b0f2866f80142dc2ef29900ffcb6718b05634044 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3557b6810b330402fab5777591de23d17bdc581ca265b325c44e0bd6a3541d1 +size 602667 diff --git a/vlm/dev/gUL6zYN4Uaf/17.png b/vlm/dev/gUL6zYN4Uaf/17.png new file mode 100644 index 0000000000000000000000000000000000000000..c6c7d26f3e04b41a1e87671812d85205cacd7ff4 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91a39a613ca59359836676097a44a45907034844a02ff7cc9e4a7d4ccb7efe95 +size 624825 diff --git a/vlm/dev/gUL6zYN4Uaf/18.png b/vlm/dev/gUL6zYN4Uaf/18.png new file mode 100644 index 0000000000000000000000000000000000000000..dc3168d52fcf2c33869c2e7532f91514a67265e8 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e3320f1763c407e3500cc32b84da02508f590b172db2f0ed8e2b248f26c0b97 +size 616951 diff --git a/vlm/dev/gUL6zYN4Uaf/19.png b/vlm/dev/gUL6zYN4Uaf/19.png new file mode 100644 index 0000000000000000000000000000000000000000..e4204a4be441520bef6a748afbc3ade0dfa4294d --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29194fbebf041d610c2e5864408b19905e3692fbb54de8893d262c0e77079828 +size 460128 diff --git a/vlm/dev/gUL6zYN4Uaf/2.png b/vlm/dev/gUL6zYN4Uaf/2.png new file mode 100644 index 0000000000000000000000000000000000000000..58c78be7dcec817389e54dac0cdb1a8957b5b780 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:427ef3bdd6a950cf86d77c6ee9f6d71f8fc685ac01faba4815e6da16cdb4ea20 +size 671439 diff --git a/vlm/dev/gUL6zYN4Uaf/20.png b/vlm/dev/gUL6zYN4Uaf/20.png new file mode 100644 index 0000000000000000000000000000000000000000..557a8496b7cb491d9774d62606dc0e5ebbd46359 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57f9ff169a34b73de143f24cb3878b9cd698c92fc2a8a0f095d3fd3e275d4360 +size 341160 diff --git a/vlm/dev/gUL6zYN4Uaf/3.png b/vlm/dev/gUL6zYN4Uaf/3.png new file mode 100644 index 0000000000000000000000000000000000000000..5266249c25ae414a8cffb3a7bfdd6fa018f7d95a --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9ffa801412a5c1e7b209b570a296cf6837cfd3d7f60b513db93f951d446893f +size 647955 diff --git a/vlm/dev/gUL6zYN4Uaf/4.png b/vlm/dev/gUL6zYN4Uaf/4.png new file mode 100644 index 0000000000000000000000000000000000000000..98978740fcc4e88e499472884b58b4d792bc1f11 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3ac5733e31f2b159b19ffb8b5a693babfd0350f4d840c8fe1c472823e6a9c3a +size 644245 diff --git a/vlm/dev/gUL6zYN4Uaf/5.png b/vlm/dev/gUL6zYN4Uaf/5.png new file mode 100644 index 0000000000000000000000000000000000000000..39220ba00428bcfccfad24f1f65dd97ed2ad0af0 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f322ffc368b69fc3619adc2af3a367f7918cc97d80def013769e74cfd1a544d4 +size 635266 diff --git a/vlm/dev/gUL6zYN4Uaf/7.png b/vlm/dev/gUL6zYN4Uaf/7.png new file mode 100644 index 0000000000000000000000000000000000000000..45710b2fea3fdf2ccc79e7e38b2a2bf5b6b9a5f5 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:137e478eee42caa24086ad3c26abfe52d900ef82dd57b6e2332b483173523a72 +size 554742 diff --git a/vlm/dev/gUL6zYN4Uaf/8.png b/vlm/dev/gUL6zYN4Uaf/8.png new file mode 100644 index 0000000000000000000000000000000000000000..8d713f820a3fa53b4d03e860e15e394a501c0590 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:706ae0fd1a9f33a8386b28b70636491bce4b21fed890d4bebfc8b88e5104853e +size 621170 diff --git a/vlm/dev/gUL6zYN4Uaf/9.png b/vlm/dev/gUL6zYN4Uaf/9.png new file mode 100644 index 0000000000000000000000000000000000000000..197bf16cd9203234339c043ac23555eaf765c745 --- /dev/null +++ b/vlm/dev/gUL6zYN4Uaf/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b530fb1ac83e1edd60e5c765576ee49b13ea18c73d6a5a9cd8633e935523a1d +size 599166 diff --git a/vlm/dev/hfjbX1UKNx/0.png b/vlm/dev/hfjbX1UKNx/0.png new file mode 100644 index 0000000000000000000000000000000000000000..493082396e599b7e990d8c51ebeb6fe6acc4fdf5 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07d3efd03731f4b390562c1e8f620d3d2f698cbe2daf1ec7a0c8f025c5a0e81b +size 511655 diff --git a/vlm/dev/hfjbX1UKNx/1.png b/vlm/dev/hfjbX1UKNx/1.png new file mode 100644 index 0000000000000000000000000000000000000000..4e2cdfd32823940f30e248abbb96321e76cf3148 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02e9d42ab7155b73ef7266b1ffe1ba0125250de6b56cadb098f78ebd6f9af4dd +size 588257 diff --git a/vlm/dev/hfjbX1UKNx/10.png b/vlm/dev/hfjbX1UKNx/10.png new file mode 100644 index 0000000000000000000000000000000000000000..e6c909dfc2337721ef6c6f0f9ddfce2b1559e107 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c568f3ec1fe0f8c2911bd27428907b46d39e90db47bdb0ef0978793476c7186 +size 503262 diff --git a/vlm/dev/hfjbX1UKNx/11.png b/vlm/dev/hfjbX1UKNx/11.png new file mode 100644 index 0000000000000000000000000000000000000000..05677c404093fcb879dc9a112ca0e895ee3b251b --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:822ad89fc66069dc230fe23cc3ac79838aa16684b4f1803aaf777f09fb8d35b7 +size 374274 diff --git a/vlm/dev/hfjbX1UKNx/12.png b/vlm/dev/hfjbX1UKNx/12.png new file mode 100644 index 0000000000000000000000000000000000000000..783011655f2c9d61caa451ec44f7dbb655239d23 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18986a78bf1140707d9a2c0c78e5b2c47c6fa51f2ac6d20d40e59931caf1c54c +size 387426 diff --git a/vlm/dev/hfjbX1UKNx/13.png b/vlm/dev/hfjbX1UKNx/13.png new file mode 100644 index 0000000000000000000000000000000000000000..e1b59f80b898c4dbe497820e867afbc42006107b --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53281f15d949609047d975bdacc1ac1f488093d61697570d14445e16ba215630 +size 1699558 diff --git a/vlm/dev/hfjbX1UKNx/14.png b/vlm/dev/hfjbX1UKNx/14.png new file mode 100644 index 0000000000000000000000000000000000000000..a3c990e0d1a56b0e04b49cd4adb32a2850a95a65 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1d760f1c4cfe4d50fc40ce5c5caaf9da938944dd63ad6ed4fd51369e7b10af2 +size 1601420 diff --git a/vlm/dev/hfjbX1UKNx/15.png b/vlm/dev/hfjbX1UKNx/15.png new file mode 100644 index 0000000000000000000000000000000000000000..7c67ac6af720f47a7f06e3cb11812a09d94fa443 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d75dc95b1481bc4fa200c25b93ba18c719f89a87514a45c562391df20a43351b +size 792453 diff --git a/vlm/dev/hfjbX1UKNx/16.png b/vlm/dev/hfjbX1UKNx/16.png new file mode 100644 index 0000000000000000000000000000000000000000..d3f5d0f9ddb7417f971bf07d523d6c6ca2b9cf0b --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fcc106bcbdf2d0028f631780ddfe6dc27f834ac0c4aef364ee1541d4fa4175f +size 312262 diff --git a/vlm/dev/hfjbX1UKNx/17.png b/vlm/dev/hfjbX1UKNx/17.png new file mode 100644 index 0000000000000000000000000000000000000000..a00545e24d863cd0917f7f2b3ab544c7cdb01947 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcca20ebed1f30ce2512e776748d315525e917072c6475ade82db6872822f0d0 +size 848571 diff --git a/vlm/dev/hfjbX1UKNx/18.png b/vlm/dev/hfjbX1UKNx/18.png new file mode 100644 index 0000000000000000000000000000000000000000..61d6dd7ad1a026cde628fa1bcec48ccf78d9158c --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f96a4fa30a957bf9789053a897da52553cb9b60729d7c521786eaf832fb5f045 +size 447578 diff --git a/vlm/dev/hfjbX1UKNx/19.png b/vlm/dev/hfjbX1UKNx/19.png new file mode 100644 index 0000000000000000000000000000000000000000..5d120df043924768a7a3774d7659ee08ab81ec4b --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de4b799a5b91afc08c3321c1a29a627e70224ada41c57dca5a0eac71b931e64b +size 347208 diff --git a/vlm/dev/hfjbX1UKNx/2.png b/vlm/dev/hfjbX1UKNx/2.png new file mode 100644 index 0000000000000000000000000000000000000000..8c17c13e537b1f1cf4e161ae02352398184bc9b7 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0087ad4cdf243c6512dc320103153541516f9fc851bdbe6d877a35beaaf69726 +size 532648 diff --git a/vlm/dev/hfjbX1UKNx/20.png b/vlm/dev/hfjbX1UKNx/20.png new file mode 100644 index 0000000000000000000000000000000000000000..1937167431fd13cf8a8da06bd68a1c2c74b9f792 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1e1284acf0c5050f829ba5e98ceb353bb16528177f5fcd508bbcb70eb0f7992 +size 2222210 diff --git a/vlm/dev/hfjbX1UKNx/21.png b/vlm/dev/hfjbX1UKNx/21.png new file mode 100644 index 0000000000000000000000000000000000000000..508bc7a67bc128fd167d40be826ccb338644df30 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:623006ef0e9f2e7afd4827bc7c1cb6266e0d97d33471ba193fa9a5d9e0587b1a +size 977999 diff --git a/vlm/dev/hfjbX1UKNx/22.png b/vlm/dev/hfjbX1UKNx/22.png new file mode 100644 index 0000000000000000000000000000000000000000..be997ad621c3ea0bf5ab7581ccd5cecaed7aa817 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:819a3f85598e6497c67d366c020e24e9ded6f8f176303d7985b305e6323e8a65 +size 1335169 diff --git a/vlm/dev/hfjbX1UKNx/3.png b/vlm/dev/hfjbX1UKNx/3.png new file mode 100644 index 0000000000000000000000000000000000000000..884e44440aefc909d644877e7a904d35d469c1a5 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb909d11c08771527aff0197d99ab18c0ff5f8d2db475207f35684ff8d87e41a +size 410238 diff --git a/vlm/dev/hfjbX1UKNx/4.png b/vlm/dev/hfjbX1UKNx/4.png new file mode 100644 index 0000000000000000000000000000000000000000..edb4980b0fbec8d9cd52a45043b6d2719f39fa17 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bd14d4a5c96a7c22bebc20b6d1c0012003b11ebabb79afeb987f59e695b5d0c +size 514010 diff --git a/vlm/dev/hfjbX1UKNx/5.png b/vlm/dev/hfjbX1UKNx/5.png new file mode 100644 index 0000000000000000000000000000000000000000..6b7de49844f5e50959c841f9abb7774ef453d658 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13e24fb8a73115f8083f0921039cdb483ebe0b80fa2c97586a1df0cd8f00c1c1 +size 532493 diff --git a/vlm/dev/hfjbX1UKNx/6.png b/vlm/dev/hfjbX1UKNx/6.png new file mode 100644 index 0000000000000000000000000000000000000000..d733d038336d3ef36f3d81f14177942587d64ddd --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da5b0388e83b83462479a9ebc7222940d64bf01ead9b3bac15eca4a5aa5790eb +size 566837 diff --git a/vlm/dev/hfjbX1UKNx/7.png b/vlm/dev/hfjbX1UKNx/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e15a09e1a2a18fc6317063695bac96cbe117b645 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f219056cfbac5485a2a38cdd0e9ffc16607bce22dc6df4103ee0cb118b598ffa +size 795439 diff --git a/vlm/dev/hfjbX1UKNx/8.png b/vlm/dev/hfjbX1UKNx/8.png new file mode 100644 index 0000000000000000000000000000000000000000..44e962f60489e639b932a21768d093478de04aac --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d26f87ca047e47d5d311bfec44cbe862cd16b2217a2e8edd5bb12e22210f2258 +size 528251 diff --git a/vlm/dev/hfjbX1UKNx/9.png b/vlm/dev/hfjbX1UKNx/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1d8771a120e31e071c172bec2c55a22487384111 --- /dev/null +++ b/vlm/dev/hfjbX1UKNx/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:849a12da823e6e35b1f53cad4aa2ad356ccb87a64f32922b28df3c3e3efa8b25 +size 495119 diff --git a/vlm/dev/mTiHLHu3sP/0.png b/vlm/dev/mTiHLHu3sP/0.png new file mode 100644 index 0000000000000000000000000000000000000000..7134948b2502b37141a6413f442e9c77457bc7b8 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5924780e80a96f990abf96b94bdac86f55f9a9b8d394007eff1a8ddc776ff8fb +size 675381 diff --git a/vlm/dev/mTiHLHu3sP/1.png b/vlm/dev/mTiHLHu3sP/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0e66288f6830b453f25080f058ae878d54a70b73 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c86c4f0b4052f35f62d429137ee2edd6aac05e0bfc894c200c78c878ccf76cd0 +size 743017 diff --git a/vlm/dev/mTiHLHu3sP/10.png b/vlm/dev/mTiHLHu3sP/10.png new file mode 100644 index 0000000000000000000000000000000000000000..a4e50e655a109a7fbd7ea711975e57ad88e3fa00 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7d220604973c3f710025ab5333c29b6ed5dfd237c3e7d9312e00384acaa7be5 +size 964997 diff --git a/vlm/dev/mTiHLHu3sP/11.png b/vlm/dev/mTiHLHu3sP/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ae8a3bf9a800a279a07b38df45dde31b79454adf --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bade80f2ad5ef5040259af2b5137f28c280d8b4e66235b7f71dcffcabdc6bc09 +size 81450 diff --git a/vlm/dev/mTiHLHu3sP/12.png b/vlm/dev/mTiHLHu3sP/12.png new file mode 100644 index 0000000000000000000000000000000000000000..fe8d22d8e6237dfc8e51620d1c827d31964cfd3e --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d6afd1e88c795b9ab658d7d6582a2d43847df3989866f213663ed772267bea1 +size 635799 diff --git a/vlm/dev/mTiHLHu3sP/2.png b/vlm/dev/mTiHLHu3sP/2.png new file mode 100644 index 0000000000000000000000000000000000000000..4f667d506c85dfa8cac7ebf6c3816b62a1383d69 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28807de0ea3f9748ac6b3d8659a0c9a97525ff479da559d30cb8e2a49c7ab778 +size 696033 diff --git a/vlm/dev/mTiHLHu3sP/3.png b/vlm/dev/mTiHLHu3sP/3.png new file mode 100644 index 0000000000000000000000000000000000000000..61ff74159085a2fa5e6356aa0f47e475709db148 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cbaa52405d03744364852b2717b5f811a162977630d806671309475424f096b +size 761975 diff --git a/vlm/dev/mTiHLHu3sP/5.png b/vlm/dev/mTiHLHu3sP/5.png new file mode 100644 index 0000000000000000000000000000000000000000..2ffd0acc116cb4da6fe14f11c927fac02f76ca98 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f482cc9fee2d3581144f2fbf62a41b1f36bda8337b96213915a62630e08ad41e +size 742699 diff --git a/vlm/dev/mTiHLHu3sP/6.png b/vlm/dev/mTiHLHu3sP/6.png new file mode 100644 index 0000000000000000000000000000000000000000..31e55df6639a52c2aaeeeb3c44386ca57d888f2c --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48601503ed76613d5dcc5002d91ee380a0737fd199290175ac5621fdd6c13b43 +size 625888 diff --git a/vlm/dev/mTiHLHu3sP/7.png b/vlm/dev/mTiHLHu3sP/7.png new file mode 100644 index 0000000000000000000000000000000000000000..659bd8b635a53ee77d3ac50c9933529efde6b506 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:556645e44bea1495530ad7527faca392d0b7fdcafb97e8408972501945cfedef +size 763802 diff --git a/vlm/dev/mTiHLHu3sP/9.png b/vlm/dev/mTiHLHu3sP/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1cad0ac23aeb57195c19328382283e8b068baa71 --- /dev/null +++ b/vlm/dev/mTiHLHu3sP/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04964f83a395df8e646b7463b57494a1f297d9257f7e74f85b276aab8263856d +size 955692 diff --git a/vlm/dev/pAq8iDy00Oa/0.png b/vlm/dev/pAq8iDy00Oa/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c62a37bda319b78c8e58db528fdbaa1ed40ee0db --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0cac8c65d6cd503d7c15dbdf37f057b7706ed367b014a2a1af35a7ba270b482 +size 457040 diff --git a/vlm/dev/pAq8iDy00Oa/1.png b/vlm/dev/pAq8iDy00Oa/1.png new file mode 100644 index 0000000000000000000000000000000000000000..b31c0bdff0579c4ba43d2eaafb78880337c6b86d --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b679e16ba79b61e7e8fea84b93ec2c1f3b6384a58727643fc1013da66edc5274 +size 631026 diff --git a/vlm/dev/pAq8iDy00Oa/10.png b/vlm/dev/pAq8iDy00Oa/10.png new file mode 100644 index 0000000000000000000000000000000000000000..8550b131aba09e21a8b4bac5279bc97f37463f85 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaf9863a423520267ccafde861e5b542dcc3542670ec95bd2e5283e374829c45 +size 589878 diff --git a/vlm/dev/pAq8iDy00Oa/11.png b/vlm/dev/pAq8iDy00Oa/11.png new file mode 100644 index 0000000000000000000000000000000000000000..d88acb4ea599c86a120627dd4f81b1c98eefd233 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2993e3047621b7e6036e4553f8202b3b4e2a6a11fc1662d5750c2ef47a1ae04b +size 588219 diff --git a/vlm/dev/pAq8iDy00Oa/12.png b/vlm/dev/pAq8iDy00Oa/12.png new file mode 100644 index 0000000000000000000000000000000000000000..2f04eebd922fc5d27ae86ac18baf31e88f48327f --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:782917c864070b7e4c04cf3fad6de3d8d7c696cfdeaeee774ba382c9bdcc9fc2 +size 410666 diff --git a/vlm/dev/pAq8iDy00Oa/13.png b/vlm/dev/pAq8iDy00Oa/13.png new file mode 100644 index 0000000000000000000000000000000000000000..97ee4ca3e2a6ff1c653ddfb7ac02b2410f4d6d0d --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9254c9760d07be8d647884a6f60993dddbb3af75b5360a28d15e837e6a49a850 +size 485416 diff --git a/vlm/dev/pAq8iDy00Oa/14.png b/vlm/dev/pAq8iDy00Oa/14.png new file mode 100644 index 0000000000000000000000000000000000000000..caf38fde77bc1687fcb62286e58254bce4cd72ea --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f3b712c4cd33f21ac36e038d5a3d44b9dc87fd09b138569af3f1de09ad476ed +size 440324 diff --git a/vlm/dev/pAq8iDy00Oa/15.png b/vlm/dev/pAq8iDy00Oa/15.png new file mode 100644 index 0000000000000000000000000000000000000000..486181ab98723245cfabdc5261aa4403376453e3 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfbfc3d59180b10df1acec5029f0aca92d70a5fe8c1b3da6f9cf4300eccf35aa +size 551775 diff --git a/vlm/dev/pAq8iDy00Oa/16.png b/vlm/dev/pAq8iDy00Oa/16.png new file mode 100644 index 0000000000000000000000000000000000000000..f8633d26d6d09475bd7880e0322b02b39e3c9109 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e66d66d10de5323bca7bba85fc9685de67b8038bc7dc670e5677401731d123e4 +size 311061 diff --git a/vlm/dev/pAq8iDy00Oa/17.png b/vlm/dev/pAq8iDy00Oa/17.png new file mode 100644 index 0000000000000000000000000000000000000000..2b3851c75babdd268771c128658da839e3a16c8a --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a08cb28d35cd43620485655d4d0821b7fd2cd73da30f733fcddfc6518d069ba +size 1011582 diff --git a/vlm/dev/pAq8iDy00Oa/18.png b/vlm/dev/pAq8iDy00Oa/18.png new file mode 100644 index 0000000000000000000000000000000000000000..06b66d0df7e5c7e333b8d22fc52a41cbfeb7dd83 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6138243cf05e043db47f9235447d6e785d15f17a38f858a2a00a81a4d35ad621 +size 655289 diff --git a/vlm/dev/pAq8iDy00Oa/19.png b/vlm/dev/pAq8iDy00Oa/19.png new file mode 100644 index 0000000000000000000000000000000000000000..d8b5b47e01086274aad4248589d368fa10d39899 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83f53f4b5f73c3c0078fa6e16e635ba74102cc657dd54a7031c057e2f5812633 +size 473580 diff --git a/vlm/dev/pAq8iDy00Oa/2.png b/vlm/dev/pAq8iDy00Oa/2.png new file mode 100644 index 0000000000000000000000000000000000000000..0aefc3a2b4de2a747f9dc0fe06c83b744bdd2e97 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e015da04aed2ed7a6a9110f4a5328bb700090db3d27924827d44c7f6db4a89fd +size 537247 diff --git a/vlm/dev/pAq8iDy00Oa/20.png b/vlm/dev/pAq8iDy00Oa/20.png new file mode 100644 index 0000000000000000000000000000000000000000..36828d3d7156c4259a245153ceae12e9d814a505 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9917351bb054bcdd3115560b0743b60c06276572b6af21d10265c3591b85005c +size 284659 diff --git a/vlm/dev/pAq8iDy00Oa/3.png b/vlm/dev/pAq8iDy00Oa/3.png new file mode 100644 index 0000000000000000000000000000000000000000..cfb5cd702ba4855d94dfce4e5ce333cce09c64e6 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28ee412361894e222da27f3c90835b25a05d14d450363ca69b397db2ff2f703b +size 619601 diff --git a/vlm/dev/pAq8iDy00Oa/4.png b/vlm/dev/pAq8iDy00Oa/4.png new file mode 100644 index 0000000000000000000000000000000000000000..3a1533bb6ae6f948b59b8762d26e820bd5bb2b8f --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64a72af5c8bcbf282d5e1a1d63a6c50a87cda0c1c3a8d25080348630d918a7b9 +size 498849 diff --git a/vlm/dev/pAq8iDy00Oa/5.png b/vlm/dev/pAq8iDy00Oa/5.png new file mode 100644 index 0000000000000000000000000000000000000000..74c8dd9abd6ab1a5505d22d1a8973df4bca32849 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd0ab8a2d78257efe2717ec04f0dc60841071dc0e870aa24d97ea2730af8d4d9 +size 631388 diff --git a/vlm/dev/pAq8iDy00Oa/6.png b/vlm/dev/pAq8iDy00Oa/6.png new file mode 100644 index 0000000000000000000000000000000000000000..5d86dd2c9bc2dc06e425b8eb920deefbdb39975e --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e54208d5c0caccbc0b2d24e7514de976f274fd359aed678faf367ebae9c1869 +size 847513 diff --git a/vlm/dev/pAq8iDy00Oa/7.png b/vlm/dev/pAq8iDy00Oa/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f62afb8095bb9855c8fafbb75fde9c0f5d662c52 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1126dc335d6efaad9462867b7d343f5d0041bf9bb043fde0da97b820a8a6bd77 +size 568680 diff --git a/vlm/dev/pAq8iDy00Oa/8.png b/vlm/dev/pAq8iDy00Oa/8.png new file mode 100644 index 0000000000000000000000000000000000000000..6815591c0f21bb44b688f953254d4aa30f10c325 --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1736e20c315d3be3605f05253aa0669a5b7a57c9642d71f0b42f6efc97b81fd +size 563767 diff --git a/vlm/dev/pAq8iDy00Oa/9.png b/vlm/dev/pAq8iDy00Oa/9.png new file mode 100644 index 0000000000000000000000000000000000000000..f29e052231174240d4717bbf35e2f6159de05d9d --- /dev/null +++ b/vlm/dev/pAq8iDy00Oa/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8adf7a7bb4bb9a4db9d859ab4d745a1ab2cfc066bd4676d45dab79e2541531cd +size 631762 diff --git a/vlm/dev/uKiE0VIluA-/0.png b/vlm/dev/uKiE0VIluA-/0.png new file mode 100644 index 0000000000000000000000000000000000000000..229680cfac9ac55ed2ca006b4ac729a521941828 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4d74531bac8f34cdb50f13e18b17c08a96b1551ed0156f5f625959f921c3459 +size 542405 diff --git a/vlm/dev/uKiE0VIluA-/1.png b/vlm/dev/uKiE0VIluA-/1.png new file mode 100644 index 0000000000000000000000000000000000000000..fe23f38b1e878f6da4b4c7ef3768c272f718b758 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a3ff439c57acb233f81a235ee688290929f0c9d1904c88e3654479e873e16cc +size 594649 diff --git a/vlm/dev/uKiE0VIluA-/10.png b/vlm/dev/uKiE0VIluA-/10.png new file mode 100644 index 0000000000000000000000000000000000000000..6698ca7742b562972efa88eff354e93235fc9f97 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35806b4d471aa8730e38d901a3d186b9390a3af79b821335facf7bcf5bb78213 +size 584031 diff --git a/vlm/dev/uKiE0VIluA-/11.png b/vlm/dev/uKiE0VIluA-/11.png new file mode 100644 index 0000000000000000000000000000000000000000..b61f3e1246575d2a60964057b4216fd47cd8a627 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ab8f154d0f463ec46321840c267126c545a62a30a37ecf847b01f8223317bed +size 553595 diff --git a/vlm/dev/uKiE0VIluA-/12.png b/vlm/dev/uKiE0VIluA-/12.png new file mode 100644 index 0000000000000000000000000000000000000000..a087d57293187b47123fe20927774e88ab39ba99 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ced70629a06e2a3be91ea415710c6278868fe43f708b83f5e97a416b4c650208 +size 339171 diff --git a/vlm/dev/uKiE0VIluA-/13.png b/vlm/dev/uKiE0VIluA-/13.png new file mode 100644 index 0000000000000000000000000000000000000000..74afdb2fa0461da6bea1f7f98c65eea03482f5c5 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fe4bcc5e587b970213aa72cb4ae896f9c0fdd0e4de356f6ed8366c7a0d0780a +size 376727 diff --git a/vlm/dev/uKiE0VIluA-/14.png b/vlm/dev/uKiE0VIluA-/14.png new file mode 100644 index 0000000000000000000000000000000000000000..2407a9fbf89610b7eeb270a5a5a83c2894d117d3 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fff9b54468d5c373057390326955cbd4d7668343c6a12d0f0e6dbd0612148df5 +size 437359 diff --git a/vlm/dev/uKiE0VIluA-/15.png b/vlm/dev/uKiE0VIluA-/15.png new file mode 100644 index 0000000000000000000000000000000000000000..3896ed49057fbcc142bb6499faade3741c8d555e --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fcfdf33e6757cde2f1ccc616c54f20a3cebdc5a909804bac7b77878a092387d +size 479910 diff --git a/vlm/dev/uKiE0VIluA-/16.png b/vlm/dev/uKiE0VIluA-/16.png new file mode 100644 index 0000000000000000000000000000000000000000..3516d5463d99a8f23a4491af54a73c2889c418c6 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3e805c04b942b9233e350e1b244793af6de09338a29f2a4dd92d9021a9b455d +size 286117 diff --git a/vlm/dev/uKiE0VIluA-/17.png b/vlm/dev/uKiE0VIluA-/17.png new file mode 100644 index 0000000000000000000000000000000000000000..be87f49f24113f545e81780763c0fcf891dda663 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbefde5e5c3ad69ec9c67869b6ec1f7255bd9003509b528b442dec0d5348b363 +size 382839 diff --git a/vlm/dev/uKiE0VIluA-/18.png b/vlm/dev/uKiE0VIluA-/18.png new file mode 100644 index 0000000000000000000000000000000000000000..0ac092985018c8841beea58c52c522b967765c21 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f300bd6bcc5de766b553284417da508aff6ae4dc55a630a4623971019f34561b +size 558143 diff --git a/vlm/dev/uKiE0VIluA-/19.png b/vlm/dev/uKiE0VIluA-/19.png new file mode 100644 index 0000000000000000000000000000000000000000..22c2bbc9a122c16eb7f11a41680aa0470be58ffd --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a55631802f7552139c5a6dbee7786b97db0686f5bc7a0425ab6bcce7e59d0801 +size 594377 diff --git a/vlm/dev/uKiE0VIluA-/2.png b/vlm/dev/uKiE0VIluA-/2.png new file mode 100644 index 0000000000000000000000000000000000000000..43c42744f8b670b93676269d48918ebfef3dfbf4 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29eed802d8bc1639c333515888c28aeedb70237edefe0c5c0862b02b090a1697 +size 618786 diff --git a/vlm/dev/uKiE0VIluA-/20.png b/vlm/dev/uKiE0VIluA-/20.png new file mode 100644 index 0000000000000000000000000000000000000000..dbfa54885ecaaa963d53c4c62911e70e346d4518 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89a5d7e5216a744dd6cf0e5a5608a1b3fad35b4f896031fc4cf859d8cabde380 +size 504926 diff --git a/vlm/dev/uKiE0VIluA-/21.png b/vlm/dev/uKiE0VIluA-/21.png new file mode 100644 index 0000000000000000000000000000000000000000..151b71551146414d8a52f173691e0291e2f6e384 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63d792d039c5edd8174fbdefe4b968950f362f71df1dbf77c221b545cbb68a6b +size 399896 diff --git a/vlm/dev/uKiE0VIluA-/22.png b/vlm/dev/uKiE0VIluA-/22.png new file mode 100644 index 0000000000000000000000000000000000000000..2a450ab2c333b32c75483d5f82a8aa5be1feebac --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f3db5a98ff35a08efc29ab77e4401a99833d4cf15fb19b35c918ab29350959f3 +size 512869 diff --git a/vlm/dev/uKiE0VIluA-/23.png b/vlm/dev/uKiE0VIluA-/23.png new file mode 100644 index 0000000000000000000000000000000000000000..878013fd564227410dffb8cdbd86f3598c0f1139 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e2c479c9ffdfc9b4634a1e32f1a57e2d9132f3ce04069d4531a5ef511fe2f12 +size 485454 diff --git a/vlm/dev/uKiE0VIluA-/3.png b/vlm/dev/uKiE0VIluA-/3.png new file mode 100644 index 0000000000000000000000000000000000000000..7efc637813d62a446e2a995ba16a5ec4c623eb5b --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f277dfef5dc4290e4192e249c0bfb2f5eeaf4ca594972803261b5f5406bea9ff +size 652942 diff --git a/vlm/dev/uKiE0VIluA-/4.png b/vlm/dev/uKiE0VIluA-/4.png new file mode 100644 index 0000000000000000000000000000000000000000..e93637d9d62da70523dd92eb76b4ce55189ee612 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9526e5f026023037317804d6918f3b64d68d4eaa7b09db4db1a4aaab505d59b5 +size 616595 diff --git a/vlm/dev/uKiE0VIluA-/5.png b/vlm/dev/uKiE0VIluA-/5.png new file mode 100644 index 0000000000000000000000000000000000000000..8eba72e3b59757cd801a6c60070b2bf61f92c563 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27e272532eed8fedba6803bb920d32104da18e1cf9dfb96dc9420bfd1980d100 +size 637318 diff --git a/vlm/dev/uKiE0VIluA-/6.png b/vlm/dev/uKiE0VIluA-/6.png new file mode 100644 index 0000000000000000000000000000000000000000..047cb7e98a021dfb2b70303eb46c383c56d13c7f --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b69484417b789f4e6f4757270f6193d95e975c0d811a01faeab7a0fd084309b7 +size 609452 diff --git a/vlm/dev/uKiE0VIluA-/7.png b/vlm/dev/uKiE0VIluA-/7.png new file mode 100644 index 0000000000000000000000000000000000000000..0445b58eab883c42986515efd31f1c7b767cc409 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9369f579751d520e52b2665e0dbabdc077c3ed5fc05ec2ba59c3affb2fbf1bf +size 580814 diff --git a/vlm/dev/uKiE0VIluA-/8.png b/vlm/dev/uKiE0VIluA-/8.png new file mode 100644 index 0000000000000000000000000000000000000000..cfefdec9095b7670013a09f21a816a5b25b405e5 --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7af0322518c453d68e8733daef8fbfda164a5f0bdfaa5f4e06512bcb1a5c257c +size 663469 diff --git a/vlm/dev/uKiE0VIluA-/9.png b/vlm/dev/uKiE0VIluA-/9.png new file mode 100644 index 0000000000000000000000000000000000000000..b7886f7052d2f7c43ffd560bace912e98606b70f --- /dev/null +++ b/vlm/dev/uKiE0VIluA-/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13f122591bdeadd420ff6a62cb8c4988f1a6b2e0012f6959f42e87772af255d8 +size 545018 diff --git a/vlm/dev/vKBdabh_WV/10.png b/vlm/dev/vKBdabh_WV/10.png new file mode 100644 index 0000000000000000000000000000000000000000..76cbd32cec33d988fbca79f6763efcc36a51774d --- /dev/null +++ b/vlm/dev/vKBdabh_WV/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f1d29bd7a0aa37ef00788f50e8d1b4cc8ca707235ba6bc23f99e36480106789 +size 626191 diff --git a/vlm/dev/vKBdabh_WV/11.png b/vlm/dev/vKBdabh_WV/11.png new file mode 100644 index 0000000000000000000000000000000000000000..8b2469f0d5a81c1bdd1dc2bee493a9ca9f844970 --- /dev/null +++ b/vlm/dev/vKBdabh_WV/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86f5f1102056385f794cb8b0b7ff5c204d716554d75a0c5a875142f68548e069 +size 568420 diff --git a/vlm/dev/vKBdabh_WV/14.png b/vlm/dev/vKBdabh_WV/14.png new file mode 100644 index 0000000000000000000000000000000000000000..27be4b7181dd75ef87f1cdc86e02cf1f6f223862 --- /dev/null +++ b/vlm/dev/vKBdabh_WV/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:628041b37591b89ca1600c34da522a1bb56ed7d725b79346742ed0a5269c4af8 +size 489932 diff --git a/vlm/dev/vKBdabh_WV/17.png b/vlm/dev/vKBdabh_WV/17.png new file mode 100644 index 0000000000000000000000000000000000000000..e8597e543feb500af87852a731e412ddf0fc6420 --- /dev/null +++ b/vlm/dev/vKBdabh_WV/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a431440068dda7e7893ffe5507961f40330d69421882ced78485240ff7a5d40 +size 275564 diff --git a/vlm/dev/vKBdabh_WV/2.png b/vlm/dev/vKBdabh_WV/2.png new file mode 100644 index 0000000000000000000000000000000000000000..94374f22a2c71b6126e815234e0cab8629170f24 --- /dev/null +++ b/vlm/dev/vKBdabh_WV/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84da1a4069f0a67357f60a8e5273de056ba1e0aab9d4dbfcb8925c2ada2dcba0 +size 604604 diff --git a/vlm/dev/vKBdabh_WV/21.png b/vlm/dev/vKBdabh_WV/21.png new file mode 100644 index 0000000000000000000000000000000000000000..7d4e969f2535f5c76b96007296b3583b79ed0daa --- /dev/null +++ b/vlm/dev/vKBdabh_WV/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:179c3603ca2996deeb4d6f9ee138854a8ff4cf27565c1584b170a195bf027d88 +size 4012312 diff --git a/vlm/dev/vKBdabh_WV/3.png b/vlm/dev/vKBdabh_WV/3.png new file mode 100644 index 0000000000000000000000000000000000000000..4924683dcd8b90626a3a29a18880ef6e3fc917bb --- /dev/null +++ b/vlm/dev/vKBdabh_WV/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f87f8287a8124d62195a8f576081987a46edd07c686df7615a58e38a38304402 +size 507435 diff --git a/vlm/dev/vKBdabh_WV/6.png b/vlm/dev/vKBdabh_WV/6.png new file mode 100644 index 0000000000000000000000000000000000000000..67f7557e1dc5c3b808148344611ef4965cab5266 --- /dev/null +++ b/vlm/dev/vKBdabh_WV/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4620d4fe2dc4da773bab13acbd8643d6eefd5ef5c16a7ef27c8072da747a7c00 +size 579954 diff --git a/vlm/dev/vKBdabh_WV/9.png b/vlm/dev/vKBdabh_WV/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1f34db2d1a20a167f9228026a6f0aff250b77df4 --- /dev/null +++ b/vlm/dev/vKBdabh_WV/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac5b54d760a3fc47bef8f84232ef91590553db8c46cfd00f73fa215f97f9904e +size 547503 diff --git a/vlm/dev/xT5rDp5VqKO/0.png b/vlm/dev/xT5rDp5VqKO/0.png new file mode 100644 index 0000000000000000000000000000000000000000..f87d205d4ed8f10c75b79d8346d6c5e6a468f692 --- /dev/null +++ b/vlm/dev/xT5rDp5VqKO/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ff5e6b30f06b3638448488432b45870ea7be8b37a464344334ed24052462dfa +size 385272 diff --git a/vlm/dev/xT5rDp5VqKO/1.png b/vlm/dev/xT5rDp5VqKO/1.png new file mode 100644 index 0000000000000000000000000000000000000000..ba9c703d106b2376333e36e44c627d8d59f306a4 --- /dev/null +++ b/vlm/dev/xT5rDp5VqKO/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c98ce1e8bd6aae18911eb453fd3f6540f3200ea8a1b25ef2a63896bc22ede06d +size 459829 diff --git a/vlm/dev/xT5rDp5VqKO/2.png b/vlm/dev/xT5rDp5VqKO/2.png new file mode 100644 index 0000000000000000000000000000000000000000..458a10849a2d4318421e250287099c727474c93c --- /dev/null +++ b/vlm/dev/xT5rDp5VqKO/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3318830ddf9b1811ff1287d751716338aeb583ca51fc1e3e8a5b1d11517b5b83 +size 465499 diff --git a/vlm/dev/xT5rDp5VqKO/3.png b/vlm/dev/xT5rDp5VqKO/3.png new file mode 100644 index 0000000000000000000000000000000000000000..fee2f2f25e7da6f1100010d1c99d3d78711e2385 --- /dev/null +++ b/vlm/dev/xT5rDp5VqKO/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e792a30620e5595c7663711669fc986da9db8790cb21489bb344354d3d32742f +size 169626 diff --git a/vlm/dev/xT5rDp5VqKO/4.png b/vlm/dev/xT5rDp5VqKO/4.png new file mode 100644 index 0000000000000000000000000000000000000000..f89e4ee628a07c03b1e47a7b2ff4acba852e90ed --- /dev/null +++ b/vlm/dev/xT5rDp5VqKO/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:803109dd3f51a186b4fd64443c2d77ac20d1d7247c48a6637513b3c76d066167 +size 106451