Datasets:
Add files using upload-large-folder tool
Browse files- parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo.md +362 -0
- parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo_content_list.json +1593 -0
- parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo_middle.json +0 -0
- parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo_model.json +0 -0
- parse/train/k-0oq5eNjh/k-0oq5eNjh.md +234 -0
- parse/train/k-0oq5eNjh/k-0oq5eNjh_content_list.json +974 -0
- parse/train/k-0oq5eNjh/k-0oq5eNjh_middle.json +0 -0
- parse/train/k-0oq5eNjh/k-0oq5eNjh_model.json +0 -0
- parse/train/rJgzzJHtDB/rJgzzJHtDB.md +352 -0
- parse/train/rJgzzJHtDB/rJgzzJHtDB_content_list.json +1873 -0
- parse/train/rJgzzJHtDB/rJgzzJHtDB_middle.json +0 -0
- parse/train/rJgzzJHtDB/rJgzzJHtDB_model.json +0 -0
- parse/train/rJljdh4KDH/rJljdh4KDH.md +325 -0
- parse/train/rJljdh4KDH/rJljdh4KDH_content_list.json +1659 -0
- parse/train/rJljdh4KDH/rJljdh4KDH_middle.json +0 -0
- parse/train/rJljdh4KDH/rJljdh4KDH_model.json +0 -0
- parse/train/rJxGLlBtwH/rJxGLlBtwH.md +273 -0
- parse/train/rJxGLlBtwH/rJxGLlBtwH_content_list.json +1515 -0
- parse/train/rJxGLlBtwH/rJxGLlBtwH_middle.json +0 -0
- parse/train/rJxGLlBtwH/rJxGLlBtwH_model.json +0 -0
parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo.md
ADDED
|
@@ -0,0 +1,362 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Fast Policy Extragradient Methods for Competitive Games with Entropy Regularization
|
| 2 |
+
|
| 3 |
+
Shicong Cen Carnegie Mellon University shicongc@andrew.cmu.edu
|
| 4 |
+
|
| 5 |
+
Yuting Wei University of Pennsylvania ytwei@wharton.upenn.edu
|
| 6 |
+
|
| 7 |
+
Yuejie Chi Carnegie Mellon University yuejiechi@cmu.edu
|
| 8 |
+
|
| 9 |
+
# Abstract
|
| 10 |
+
|
| 11 |
+
This paper investigates the problem of computing the equilibrium of competitive games, which is often modeled as a constrained saddle-point optimization problem with probability simplex constraints. Despite recent efforts in understanding the last-iterate convergence of extragradient methods in the unconstrained setting, the theoretical underpinnings of these methods in the constrained settings, especially those using multiplicative updates, remain highly inadequate, even when the objective function is bilinear. Motivated by the algorithmic role of entropy regularization in single-agent reinforcement learning and game theory, we develop provably efficient extragradient methods to find the quantal response equilibrium (QRE)—which are solutions to zero-sum two-player matrix games with entropy regularization—at a linear rate. The proposed algorithms can be implemented in a decentralized manner, where each player executes symmetric and multiplicative updates iteratively using its own payoff without observing the opponent’s actions directly. In addition, by controlling the knob of entropy regularization, the proposed algorithms can locate an approximate Nash equilibrium of the unregularized matrix game at a sublinear rate without assuming the Nash equilibrium to be unique. Our methods also lead to efficient policy extragradient algorithms for solving entropy-regularized zero-sum Markov games at a linear rate. All of our convergence rates are nearly dimension-free, which are independent of the size of the state and action spaces up to logarithm factors, highlighting the positive role of entropy regularization for accelerating convergence.
|
| 12 |
+
|
| 13 |
+
# 1 Introduction
|
| 14 |
+
|
| 15 |
+
Finding the equilibrium of competitive games, which can be viewed as constrained saddle-point optimization problems with probability simplex constraints, lies at the heart of modern machine learning and decision making paradigms such as Generative Adversarial Networks (GANs) (Goodfellow et al., 2014), competitive reinforcement learning (RL) (Littman, 1994), game theory (Shapley, 1953), adversarial training (Mertikopoulos et al., 2018b), to name a few.
|
| 16 |
+
|
| 17 |
+
In this paper, we study one of the most basic forms of competitive games, namely two-player zero-sum games, in both the matrix setting and the Markov setting. Our goal is to find the equilibrium policies of both players in an independent and decentralized manner (Daskalakis et al., 2020; Wei et al., 2021a) with guaranteed last-iterate convergence. Namely, each player will execute symmetric and independent updates iteratively using its own payoff without observing the opponent’s actions directly, and the final policies of the iterative process should be a close approximation to the equilibrium up to any prescribed precision. This kind of algorithms is more advantageous and versatile especially in federated environments, as it requires neither prior coordination between the players like twotimescale algorithms, nor a central controller to collect and disseminate the policies of all the players, which are often unavailable due to privacy constraints.
|
| 18 |
+
|
| 19 |
+
# 1.1 Last-iterate convergence in competitive games
|
| 20 |
+
|
| 21 |
+
In recent years, there have been significant progresses in understanding the last-iterate convergence of simple iterative algorithms for unconstrained saddle-point optimization, where one is interested in bounding the sub-optimality of the last iterate of the algorithm, rather than say, the ergodic iterate — which is the average of all the iterations — that are commonly studied in the earlier literature. This shift of focus is motivated, for example, by the infeasibility of averaging large machine learning models in training GANs (Goodfellow et al., 2014). While vanilla Gradient Descent / Ascent (GDA) may diverge or cycle even for bilinear matrix games (Daskalakis et al., 2018), quite remarkably, small modifications lead to guaranteed last-iterate convergence to the equilibrium in a non-asymptotic fashion. A flurry of algorithms is proposed, including Optimistic Gradient Descent Ascent (OGDA) (Rakhlin and Sridharan, 2013; Daskalakis and Panageas, 2018b; Wei et al., 2021b), predictive updates (Yadav et al., 2017), implicit updates (Liang and Stokes, 2019), and more. Several unified analyses of these algorithms have been carried out (see, e.g. Mokhtari et al. (2020a); Liang and Stokes (2019) and references therein), where these methods in principle all make clever extrapolation of the local curvature in a predictive manner to accelerate convergence. With slight abuse of terminology, in this paper, we refer to this ensemble of algorithms as extragradient methods (Korpelevich, 1976; Tseng, 1995; Mertikopoulos et al., 2018a; Harker and Pang, 1990).
|
| 22 |
+
|
| 23 |
+
However, saddle-point optimization in the constrained setting, which includes competitive games as a special case, remains largely under-explored even for bilinear matrix games. While it is possible to reformulate constrained bilinear games to unconstrained ones using softmax parameterization of the probability simplex, this approach falls short of preserving the bilinear structure and convexconcave properties in the original problem, which are crucial to the convergence of gradient methods. Therefore, there is a strong necessity of understanding and developing improved extragradient methods in the constrained setting. Daskalakis and Panageas (2018a) proposed the optimistic variant of the multiplicative weight updates (MWU) method (Arora et al., 2012) – which is extremely natural and popular for optimizing over probability simplexes – called Optimistic Multiplicative Weight Updates (OMWU), and established the asymptotic last-iterate convergence of OMWU for matrix games. Very recently, Wei et al. (2021b) established non-asymptotic last-iterate convergences of OMWU. However, these last-iterate convergence results require the Nash equilibrium to be unique, and cannot be applied to problems with multiple Nash equilibria.
|
| 24 |
+
|
| 25 |
+
# 1.2 Our contributions
|
| 26 |
+
|
| 27 |
+
Motivated by the algorithmic role of entropy regularization in single-agent RL (Neu et al., 2017; Geist et al., 2019; Cen et al., 2020) as well as its wide use in game theory to account for imperfect and noisy information (McKelvey and Palfrey, 1995; Savas et al., 2019), we initiate the design and analysis of extragradient algorithms using multiplicative updates for finding the quantal response equilibrium (QRE), which are solutions to competitive games with entropy regularization (McKelvey and Palfrey, 1995). While finding QRE is of interest in its own right, by controlling the knob of entropy regularization, the QRE provides a close approximation to the Nash equilibrium (NE), and in turn acts as a smoothing scheme for finding the NE. Our contributions are summarized below.
|
| 28 |
+
|
| 29 |
+
• Near dimension-free last-iterate convergence to QRE of entropy-regularized matrix games. We propose two policy extragradient algorithms to solve entropy-regularized matrix games, namely the Predictive Update (PU) and OMWU methods, where both players execute symmetric and multiplicative updates without knowing the entire payoff matrix nor the opponent’s actions. Encouragingly, we show that the last iterate of the proposed algorithms converges to the unique QRE at a linear rate that is almost independent of the size of the action spaces. Roughly speaking, to find an $\epsilon$ -optimal QRE in terms of Kullback-Leibler (KL) divergence, it takes no more than $\begin{array} { r } { \widetilde O \left( \frac { 1 } { \eta \tau } \log \left( \frac { 1 } { \epsilon } \right) \right) } \end{array}$ iterations, where ${ \widetilde { O } } ( \cdot )$ hides logarithmic dependencies. Here, $\tau$ is the regularization parameter, and $\eta$ is the learning rate of both players. Maximizing the learning rate, the iteration complexity is bounded by $\widetilde { O } \left( ( 1 + \| A \| _ { \infty } / \tau ) \log ( 1 / \epsilon ) \right)$ , where $\left\| A \right\| _ { \infty } = \operatorname* { m a x } _ { i , j } \left| A _ { i , j } \right|$ is the $\ell _ { \infty }$ norm of the payoff matrix $A$ .
|
| 30 |
+
|
| 31 |
+
• Last-iterate convergence to $\epsilon$ -NE of unregularized matrix games without uniqueness assumption. The QRE provides an accurate approximation to the NE by setting the entropy regularization $\tau$ sufficiently small, therefore our result directly translates to finding a NE with last-iterate convergence guarantee. Roughly speaking, to find an $\epsilon$ -NE (Zhang et al., 2020, Definition 2.1), it takes no more than $\begin{array} { r } { \widetilde { O } \left( 1 + \frac { \| A \| _ { \infty } } { \epsilon } \right) } \end{array}$ iterations with optimized learning rates, which is again independent of the size of the action spaces up to logarithmic factors. Unlike prior literature (Daskalakis and Panageas, 2018a; Wei et al., 2021b), our last-iterate convergence guarantee does not require the NE to be unique.
|
| 32 |
+
|
| 33 |
+
Table 1: Comparisons of last-iterate convergence of the proposed entropy-regularized PU and OMWU methods with prior results for finding $\epsilon$ -QRE or $\mathrm { \epsilon - N E }$ of competitive matrix games. We note that the convergence rates of unregularized OMWU established in Wei et al. (2021b) are problemdependent, and scale at least polynomially on the size of the action spaces. Desirable features in the last two columns are highlighted in blue.
|
| 34 |
+
|
| 35 |
+
<table><tr><td rowspan=1 colspan=1>Equilibriumtype</td><td rowspan=1 colspan=1>Method</td><td rowspan=1 colspan=1>Convergence rate</td><td rowspan=1 colspan=1>Dimension-free</td><td rowspan=1 colspan=1>Requireunique NE</td></tr><tr><td rowspan=1 colspan=1>E-QRE</td><td rowspan=1 colspan=1>PU&OMWU(this work)</td><td rowspan=1 colspan=1>linear</td><td rowspan=1 colspan=1>yes</td><td rowspan=1 colspan=1>n/a</td></tr><tr><td rowspan=3 colspan=1>e-NE</td><td rowspan=1 colspan=1>OMWU(Daskalakis and Panageas, 2018a)</td><td rowspan=1 colspan=1>asymptotic</td><td rowspan=1 colspan=1>no</td><td rowspan=1 colspan=1>yes</td></tr><tr><td rowspan=1 colspan=1>OMWU(Wei et al., 2021b)</td><td rowspan=1 colspan=1>sublinear + linear</td><td rowspan=1 colspan=1>no</td><td rowspan=1 colspan=1>yes</td></tr><tr><td rowspan=1 colspan=1>PU&OMWU(this work)</td><td rowspan=1 colspan=1>sublinear</td><td rowspan=1 colspan=1>yes</td><td rowspan=1 colspan=1>no</td></tr></table>
|
| 36 |
+
|
| 37 |
+
• Extensions to two-player zero-sum Markov games. By connecting value iteration with matrix games, we propose a policy extragradient method for solving infinite-horizon discounted entropyregularized zero-sum Markov games, which finds an $\epsilon$ -optimal minimax soft Q-function—in terms of $\ell _ { \infty }$ error—in at most $\begin{array} { r } { \widetilde O \left( \frac { 1 } { \tau ( 1 - \gamma ) ^ { 2 } } \log ^ { 2 } \left( \frac { 1 } { \epsilon } \right) \right) } \end{array}$ iterations, where $\gamma \in ( 0 , 1 )$ is the discount factor.
|
| 38 |
+
|
| 39 |
+
To the best of our knowledge, our paper is the first that develops policy extragradient algorithms for solving entropy-regularized competitive games with multiplicative updates and dimension-free linear last-iterate convergence, and demonstrates entropy regularization as a smoothing technique to find $\mathrm { \epsilon - N E }$ without the uniqueness assumption. Table 1 provides detailed comparisons of the proposed methods with prior arts for solving matrix games. Our results highlight the positive role of entropy regularization for accelerating convergence and safeguarding against imperfect information in competitive games. We defer the complete proof of our results to Cen et al. (2021).
|
| 40 |
+
|
| 41 |
+
# 1.3 Related works
|
| 42 |
+
|
| 43 |
+
Our work lies at the intersection of saddle-point optimization, game theory, and reinforcement learning. In what follows, we discuss a few topics that are closely related to ours.
|
| 44 |
+
|
| 45 |
+
Unregularized matrix game. Freund and Schapire (1999) showed that many standard methods such as GDA and MWU have a converging average duality gap at the rate of $O ( 1 / \sqrt { T } )$ , which is improved to $O ( 1 / T )$ by considering optimistic variants of these methods, such as OGDA and OMWU (Rakhlin and Sridharan, 2013; Daskalakis et al., 2011; Syrgkanis et al., 2015). However, the last-iterate convergence of these methods are less understood until recently (Daskalakis and Panageas, 2018a; Wei et al., 2021b). In particular, under the assumption that the NE is unique for the unregularized matrix game, Daskalakis and Panageas (2018a) showed the asymptotic convergence of the last iterate of OMWU to the unique equilibrium, and Wei et al. (2021b) showed the last iterate of OMWU achieves a linear rate of convergence after an initial phase of sublinear convergence, however the rates therein can be highly pessimistic in terms of the problem dimension, while our rate for entropy-regularized OMWU is dimension-free up to logarithmic factors.
|
| 46 |
+
|
| 47 |
+
Saddle-point optimization. Considerable progress has been made towards understanding OGDA and extragradient (EG) methods in the unconstrained convex-concave saddle-point optimization with general objective functions (Mokhtari et al., 2020a,b; Nemirovski, 2004; Liang and Stokes, 2019). However, the last-iterate convergence of constrained convex-concave saddle-point optimization still lacks theoretical understanding in general and most works fall short of characterizing a finite-time convergence result. In particular, Mertikopoulos et al. (2018a) demonstrated the asymptotic lastiterate convergence of EG, and Hsieh et al. (2019) investigated similar questions for single-call EG algorithms. Lei et al. (2021) showed that OMWU converges to the equilibrium locally without an explicit rate. Wei et al. (2021b) showed that the last-iterate of OGDA converges linearly for strongly-convex strongly-concave constrained saddle-point optimization with an explicit rate.
|
| 48 |
+
|
| 49 |
+
Entropy regularization in RL and games. In single-agent RL, the role of entropy regularization as an algorithmic mechanism to encourage exploration and accelerate convergence has been investigated extensively (Neu et al., 2017; Geist et al., 2019; Mei et al., 2020; Cen et al., 2020; Lan, 2021; Zhan et al., 2021). Turning to the game setting, entropy regularization is used to account for imperfect information in the seminal work of McKelvey and Palfrey (1995) that introduced the QRE, and a few representative works on entropy and more general regularizations in games include Savas et al. (2019); Hofbauer and Sandholm (2002); Mertikopoulos and Sandholm (2016).
|
| 50 |
+
|
| 51 |
+
Zero-sum Markov games. There have been a significant recent interest in developing provably efficient self-play algorithms for Markov games, including model-based algorithms (Perolat et al., 2015; Zhang et al., 2020), value-based algorithms (Bai and Jin, 2020; Xie et al., 2020), and policybased algorithms (Daskalakis et al., 2020; Wei et al., 2021a; Zhao et al., 2021). The iteration complexities in prior works (Perolat et al., 2015; Daskalakis et al., 2020; Wei et al., 2021a; Zhao et al., 2021) depend on various notions of concentrability coefficient and therefore can scale quite pessimistically with the problem dimension. Our approach can be regarded as a policy-based algorithm to approximate value iteration, which can be implemented in a decentralized manner with symmetric and multiplicative updates from both players, and the iteration complexity is almost independent of the size of the state-action space.
|
| 52 |
+
|
| 53 |
+
Notation. We denote by $\Delta ( \mathcal { A } )$ the probability simplex over the set $\mathcal { A }$ . We overload the functions such as $\log ( \cdot )$ and $\exp ( \cdot )$ to take vector inputs with the understanding that the function is applied in an entrywise manner. For instance, given any vector $z = [ z _ { i } ] _ { 1 \leq i \leq n } \in \mathbb { R } ^ { n }$ , the notation $\exp ( z )$ denotes $\exp ( z ) : = [ \exp ( z _ { i } ) ] _ { 1 \leq i \leq n }$ ; other functions are defined analogously. Given two probability distributions $\mu$ and $\mu ^ { \prime }$ over $\mathcal { A }$ , the KL divergence from $\mu ^ { \prime }$ to $\mu$ is defined by $\begin{array} { r } { \mathsf { K L } ( \mu \parallel \mu ^ { \prime } ) : = \sum _ { a \in \mathcal { A } } \mu ( a ) \log \frac { \mu ( a ) } { \mu ^ { \prime } ( a ) } } \end{array}$ . Given a matrix $A$ , $\| A \| _ { \infty }$ is used to denote entrywise maximum norm, namely, $\left\| A \right\| _ { \infty } = \operatorname* { m a x } _ { i , j } \left| A _ { i , j } \right|$ . The all-one vector is denoted as 1.
|
| 54 |
+
|
| 55 |
+
# 2 Zero-sum matrix games with entropy regularization
|
| 56 |
+
|
| 57 |
+
We first consider a two-player zero-sum game with bilinear objective and probability simplex constraints, and demonstrate the positive role of entropy regularization in solving this problem. Throughout this paper, let $\mathcal { A } = \{ 1 , \dotsc , m \}$ and $\boldsymbol { B } = \{ 1 , \ldots , \bar { n } \}$ be the action spaces of each player.
|
| 58 |
+
|
| 59 |
+
# 2.1 Background and problem formulation
|
| 60 |
+
|
| 61 |
+
Zero-sum two-player matrix game. The focal point of this subsection is a constrained two-player zero-sum matrix game, which can be formulated as the following min-max problem (or saddle point optimization problem):
|
| 62 |
+
|
| 63 |
+
$$
|
| 64 |
+
\operatorname* { m a x } _ { \mu \in \Delta ( \mathcal { A } ) } \operatorname* { m i n } _ { \nu \in \Delta ( \mathcal { B } ) } f ( \mu , \nu ) : = \mu ^ { \top } A \nu ,
|
| 65 |
+
$$
|
| 66 |
+
|
| 67 |
+
where $\textbf { \textit { A } } \in \mathbb { R } ^ { m \times n }$ denotes the payoff matrix, $\mu ~ \in ~ \Delta ( { \mathcal { A } } )$ and $\nu ~ \in ~ \Delta ( B )$ stand for the mixed/randomized policies of each player, defined respectively as distributions over the probability simplex $\Delta ( \mathcal { A } )$ and $\Delta ( B )$ . A pair of policies $( \mu ^ { \star } , \nu ^ { \star } )$ is said to be a Nash equilibrium (NE) of (1) if $f ( \mu ^ { \star } , \nu ) \geq f ( \mu ^ { \star } , \nu ^ { \star } ) \geq f ( \mu , \nu ^ { \star } )$ for all $( \mu , \nu ) \in \Delta ( \mathcal { A } ) \times \Delta ( \mathcal { B } )$ .
|
| 68 |
+
|
| 69 |
+
Entropy-regularized zero-sum two-player matrix game. There is no shortage of scenarios where the payoff matrix $A$ might not be known perfectly. In an attempt to accommodate imperfect knowledge of $A$ , McKelvey and Palfrey (1995) proposed a seminal extension to the Nash equilibrium called the quantal response equilibrium $( Q R E )$ when the payoffs are perturbed by Gumbel-distributed noise. Formally, this amounts to solving the following matrix game with entropy regularization (Mertikopoulos and Sandholm, 2016):
|
| 70 |
+
|
| 71 |
+
$$
|
| 72 |
+
\operatorname* { m a x } _ { \mu \in \Delta ( \mathcal { A } ) } \operatorname* { m i n } _ { \nu \in \Delta ( \mathcal { B } ) } f _ { \tau } ( \mu , \nu ) : = \mu ^ { \top } A \nu + \tau \mathcal { H } ( \mu ) - \tau \mathcal { H } ( \nu ) ,
|
| 73 |
+
$$
|
| 74 |
+
|
| 75 |
+
where $\begin{array} { r } { \mathcal { H } ( \pi ) = - \sum _ { i } \pi _ { i } \log ( \pi _ { i } ) } \end{array}$ denotes the Shannon entropy of a distribution $\pi$ , and $\tau \geq 0$ is the regularization parameter. As is well known, the optimal solution $( \mu _ { \tau } ^ { \star } , \nu _ { \tau } ^ { \star } )$ to (2), dubbed as the QRE,
|
| 76 |
+
|
| 77 |
+
is unique whenever $\tau > 0$ (due to the presence of strong concavity/convexity), which satisfies the following fixed point equations:
|
| 78 |
+
|
| 79 |
+
$$
|
| 80 |
+
\begin{array} { r } { \biggr \{ \mu _ { \tau } ^ { \star } ( a ) = \frac { \exp ( [ A \nu _ { \tau } ^ { \star } ] a / \tau ) } { \sum _ { a = 1 } ^ { m } \exp ( [ A \nu _ { \tau } ^ { \star } ] a / \tau ) } \propto \exp ( [ A \nu _ { \tau } ^ { \star } ] a / \tau ) , \qquad \mathrm { f o r ~ a l l } \ a \in \mathcal { A } , } \\ { \nu _ { \tau } ^ { \star } ( b ) = \frac { \exp ( - [ A ^ { \top } \mu _ { \tau } ^ { \star } ] b / \tau ) } { \sum _ { b = 1 } ^ { n } \exp ( - [ A ^ { \top } \mu _ { \tau } ^ { \star } ] b / \tau ) } \propto \exp ( - [ A ^ { \top } \mu _ { \tau } ^ { \star } ] b / \tau ) , \quad \mathrm { f o r ~ a l l } \ b \in \mathcal { B } . } \end{array}
|
| 81 |
+
$$
|
| 82 |
+
|
| 83 |
+
Goal. We aim to efficiently compute the QRE of the entropy-regularized matrix game in a decentralized manner, and investigate how an efficient solver of QRE can be leveraged to find a NE of the unregularized matrix game (1). Namely, we only assume access to “first-order information” as opposed to full knowledge of the payoff matrix $A$ or the actions of the opponent. The information received by each player is formally described in the following sampling oracle.
|
| 84 |
+
|
| 85 |
+
Definition 1 (Sampling oracle for matrix games). For any policy pair $( \mu , \nu )$ and payoff matrix $A$ the sampling oracle returns the exact values of $\mu ^ { \top } A$ and $A \nu$ .
|
| 86 |
+
|
| 87 |
+
Additional notation. For notational convenience, we let $\zeta$ represent the concatenation of $\mu \in \mathbb { R } ^ { | \mathcal { A } | }$ and $\nu \in \mathbb { R } ^ { | B | }$ , namely, $\zeta = ( \mu , \nu )$ . The solution to (2), which is specified in (3), is denoted by $\zeta _ { \tau } ^ { \star } = ( \mu _ { \tau } ^ { \star } , \nu _ { \tau } ^ { \star } )$ . For any $\zeta = ( \mu , \nu )$ and $\zeta ^ { \prime } = ( \mu ^ { \prime } , \nu ^ { \prime } )$ , we shall often abuse the notation and let $\mathsf { K L } \big ( \zeta \| \zeta ^ { \prime } \big ) = \mathsf { K L } \big ( \mu \| \mu ^ { \prime } \big ) + \mathsf { K L } \big ( \nu \| \nu ^ { \prime } \big )$ . The duality gap of the entropy-regularized matrix game (2) at $\zeta = ( \mu , \nu )$ is defined as $\begin{array} { r } { \mathsf { D u a l G a p } _ { \tau } ( \zeta ) = \operatorname* { m a x } _ { \mu ^ { \prime } \in \Delta ( A ) } f _ { \tau } ( \mu ^ { \prime } , \nu ) - \operatorname* { m i n } _ { \nu ^ { \prime } \in \Delta ( B ) } f _ { \tau } ( \mu , \nu ^ { \prime } ) } \end{array}$ which is clearly nonnegative and $\mathsf { D u a l G a p } _ { \tau } ( \boldsymbol { \zeta } _ { \tau } ^ { \star } ) = 0$ . Similarly, let the optimality gap of the entropyregularized matrix game (2) at $\zeta = ( \mu , \nu )$ be $\mathsf { O p t G a p } ( \zeta ) = \left| f _ { \tau } ( \mu , \nu ) - f _ { \tau } ( \mu _ { \tau } ^ { \star } , \nu _ { \tau } ^ { \star } ) \right|$ .
|
| 88 |
+
|
| 89 |
+
# 2.2 Proposed extragradient methods: PU and OMWU
|
| 90 |
+
|
| 91 |
+
To begin, assume we are given a pair of policies $z _ { 1 } \in \Delta ( \mathcal { A } )$ , $z _ { 2 } \in \Delta ( B )$ employed by each player respectively. If we proceed with fictitious play, i.e. player 1 (resp. player 2) aims to optimize its own policy by assuming the opponent’s policy is fixed as $z _ { 2 }$ (resp. $z _ { 1 }$ ), the saddle-point optimization problem (2) is then decoupled into two independent min/max optimization problems:
|
| 92 |
+
|
| 93 |
+
$$
|
| 94 |
+
\operatorname* { m a x } _ { \mu \in \Delta ( A ) } \mu ^ { \top } A z _ { 2 } + \tau \mathcal { H } ( \mu ) - \tau \mathcal { H } ( z _ { 2 } ) \qquad \mathrm { a n d } \qquad \operatorname* { m i n } _ { \nu \in \Delta ( B ) } z _ { 1 } ^ { \top } A \nu + \tau \mathcal { H } ( z _ { 1 } ) - \tau \mathcal { H } ( \nu ) ,
|
| 95 |
+
$$
|
| 96 |
+
|
| 97 |
+
which are naturally solved via mirror descent / ascent with KL divergence. Specifically, one step of mirror descent / ascent takes the form
|
| 98 |
+
|
| 99 |
+
$$
|
| 100 |
+
\begin{array} { r } { \left\{ \begin{array} { l l } { \mu ^ { ( t + 1 ) } ( a ) \propto \mu ^ { ( t ) } ( a ) ^ { 1 - \eta \tau } \exp ( \eta [ A z _ { 2 } ] _ { a } ) , } & { \mathrm { f o r ~ a l l ~ } a \in \mathcal { A } , } \\ { \nu ^ { ( t + 1 ) } ( b ) \propto \nu ^ { ( t ) } ( b ) ^ { 1 - \eta \tau } \exp ( - \eta [ A ^ { \top } z _ { 1 } ] _ { b } ) , } & { \mathrm { f o r ~ a l l ~ } b \in \mathcal { B } , } \end{array} \right. } \end{array}
|
| 101 |
+
$$
|
| 102 |
+
|
| 103 |
+
where $\eta$ is the learning rate. The above update rule forms the basis of our algorithm design.
|
| 104 |
+
|
| 105 |
+
Motivation: a form of implicit updates with linear convergence. It turns out, if we could select the policy pair $( z _ { 1 } , z _ { 2 } ) = \hat { \zeta } ^ { ( t + 1 ) } : \stackrel { \bullet } { = } ( \mu ^ { ( t + 1 ) } , \nu ^ { ( t + 1 ) } )$ as the ones to be taken in the future, and call the resulting update rule as the Implicit Update (IU) method:
|
| 106 |
+
|
| 107 |
+
$$
|
| 108 |
+
\begin{array} { r } { \left\{ \mu ^ { ( t + 1 ) } ( a ) \propto \mu ^ { ( t ) } ( a ) ^ { 1 - \eta \tau } \exp ( \eta [ A \nu ^ { ( t + 1 ) } ] _ { a } ) , \right. \left. \mathrm { f o r ~ a l l } \ : a \in \mathcal { A } , \right. } \\ { \nu ^ { ( t + 1 ) } ( b ) \propto \nu ^ { ( t ) } ( b ) ^ { 1 - \eta \tau } \exp ( - \eta [ A ^ { \top } \mu ^ { ( t + 1 ) } ] _ { b } ) , \left. \mathrm { f o r ~ a l l } \ : b \in \mathcal { B } . \right. } \end{array}
|
| 109 |
+
$$
|
| 110 |
+
|
| 111 |
+
Though unrealistic — since it uses the future updates — it leads to a one-step convergence to the QRE when $\eta = 1 / \tau$ (see the optimality condition in (3)). Encouragingly, we have the following linear convergence guarantee of IU when adopting a general learning rate.
|
| 112 |
+
|
| 113 |
+
Proposition 1 (Linear convergence of IU). Assume $0 < \eta \leq 1 / \tau$ , then for all $t \geq 0$ , the iterates $\zeta ^ { ( t ) } : = ( \mu ^ { ( t ) } , \nu ^ { ( t ) } )$ of the $I U$ method in (5) satisfy $\mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \parallel \zeta ^ { ( t ) } \big ) \leq ( 1 - \eta \tau ) ^ { t } \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \parallel \zeta ^ { ( 0 ) } \big )$ .
|
| 114 |
+
|
| 115 |
+
In words, the IU method achieves an appealing linear rate of convergence that is independent of the problem dimension. Motivated by this observation, we seek to design algorithms where the policies $( z _ { 1 } , z _ { 2 } )$ employed in (4) serve as good predictions of $( \mu ^ { ( t + 1 ) } , \nu ^ { ( t + \bar { 1 ) } } )$ , such that the resulting algorithms are both practical and retain the appealing convergence rate of IU.
|
| 116 |
+
|
| 117 |
+
Proposed algorithms. We propose two extragradient algorithms for solving the entropy-regularized matrix game, namely the Predictive Update $( P U )$ method and the Optimistic Multiplicative Weights
|
| 118 |
+
|
| 119 |
+
# Algorithm 1: The PU method
|
| 120 |
+
|
| 121 |
+
# Algorithm 2: The OMWU method
|
| 122 |
+
|
| 123 |
+
1 initialization: $\mu ^ { ( 0 ) }$ , $\nu ^ { ( 0 ) }$
|
| 124 |
+
|
| 125 |
+
2 for $t = 0 , 1 , 2 , \cdots$ do
|
| 126 |
+
|
| 127 |
+
2 for $t = 0 , 1 , 2 , \cdots$ do
|
| 128 |
+
|
| 129 |
+
3 Update $\bar { \mu }$ and $\bar { \nu }$ according to
|
| 130 |
+
|
| 131 |
+
3
|
| 132 |
+
|
| 133 |
+
Update $\bar { \mu }$ and $\bar { \nu }$ according to
|
| 134 |
+
|
| 135 |
+
$$
|
| 136 |
+
\begin{array} { r } { \left\{ \bar { \mu } ^ { ( t + 1 ) } ( a ) \propto { \mu } ^ { ( t ) } ( a ) ^ { 1 - \eta \tau } \exp ( \eta [ A \nu ^ { ( t ) } ] _ { a } ) , \right. \qquad } \\ { \left. \bar { \nu } ^ { ( t + 1 ) } ( b ) \propto \nu ^ { ( t ) } ( b ) ^ { 1 - \eta \tau } \exp ( - \eta [ A ^ { \top } { \mu } ^ { ( t ) } ] _ { b } ) . \right. } \end{array}
|
| 137 |
+
$$
|
| 138 |
+
|
| 139 |
+
$$
|
| 140 |
+
\begin{array} { r } { \left\{ \bar { \mu } ^ { ( t + 1 ) } ( a ) \propto { \mu } ^ { ( t ) } ( a ) ^ { 1 - \eta \tau } \exp ( \eta [ A \bar { \nu } ^ { ( t ) } ] _ { a } ) , \right. \mathrm { ~ } } \\ { \left. \bar { \nu } ^ { ( t + 1 ) } ( b ) \propto \nu ^ { ( t ) } ( b ) ^ { 1 - \eta \tau } \exp ( - \eta [ A ^ { \top } \bar { \mu } ^ { ( t ) } ] _ { b } ) . \right. } \end{array}
|
| 141 |
+
$$
|
| 142 |
+
|
| 143 |
+
4 Update $\mu$ and $\nu$ according to
|
| 144 |
+
|
| 145 |
+
Update $\mu$ and $\nu$ according to
|
| 146 |
+
|
| 147 |
+
$$
|
| 148 |
+
\begin{array} { r } { \left\{ \begin{array} { l l } { \mu ^ { ( t + 1 ) } ( a ) \propto \mu ^ { ( t ) } ( a ) ^ { 1 - \eta \tau } \exp ( \eta [ A \bar { \nu } ^ { ( t + 1 ) } ] _ { a } ) , } \\ { \nu ^ { ( t + 1 ) } ( b ) \propto \nu ^ { ( t ) } ( b ) ^ { 1 - \eta \tau } \exp ( - \eta [ A ^ { \top } \bar { \mu } ^ { ( t + 1 ) } ] _ { b } ) . } \end{array} \right. } \end{array}
|
| 149 |
+
$$
|
| 150 |
+
|
| 151 |
+
$$
|
| 152 |
+
\begin{array} { r } { \left\{ \begin{array} { l l } { \mu ^ { ( t + 1 ) } ( a ) \propto \mu ^ { ( t ) } ( a ) ^ { 1 - \eta \tau } \exp ( \eta [ A \bar { \nu } ^ { ( t + 1 ) } ] _ { a } ) , } \\ { \nu ^ { ( t + 1 ) } ( b ) \propto \nu ^ { ( t ) } ( b ) ^ { 1 - \eta \tau } \exp ( - \eta [ A ^ { \top } \bar { \mu } ^ { ( t + 1 ) } ] _ { b } ) . } \end{array} \right. } \end{array}
|
| 153 |
+
$$
|
| 154 |
+
|
| 155 |
+
Update (OMWU) method, the latter adapted from Rakhlin and Sridharan (2013); Daskalakis et al. (2011). Detailed procedures can be found in Algorithm 1 and Algorithm 2, respectively. On a high level, both algorithms maintain two intertwined sequences $\{ ( \mu ^ { ( \bar { t } ) } , \nu ^ { ( t ) } ) \} _ { t \geq 0 }$ and $\{ ( \bar { \mu } ^ { ( \dot { t } ) } , \bar { \nu } ^ { ( t ) } ) \} _ { t \geq 0 }$ , and in each iteration $t = 0 , 1 , \ldots$ , proceed in two steps:
|
| 156 |
+
|
| 157 |
+
• The midpoint $( \bar { \mu } ^ { ( t + 1 ) } , \bar { \nu } ^ { ( t + 1 ) } )$ serves as a prediction of $( \mu ^ { ( t + 1 ) } , \nu ^ { ( t + 1 ) } )$ by running one step of mirror descent / ascent (cf. (4)) from either $( z _ { 1 } , z _ { 2 } ) = ( \mu ^ { ( t ) } , \nu ^ { ( t ) } )$ (for PU) or $( z _ { 1 } , z _ { 2 } ) = ( \bar { \mu } ^ { ( t ) } , \bar { \nu } ^ { ( t ) } )$ (for OMWU).
|
| 158 |
+
|
| 159 |
+
• The update of $( \boldsymbol { \mu } ^ { ( t + 1 ) } , \boldsymbol { \nu } ^ { ( t + 1 ) } )$ then mimics the implicit update (5) using the prediction $( \bar { \mu } ^ { ( t + 1 ) } , \bar { \nu } ^ { ( t + 1 ) } )$ obtained above.
|
| 160 |
+
|
| 161 |
+
When the proposed algorithms converge, both $( \mu ^ { ( t ) } , \nu ^ { ( t ) } )$ and $( \bar { \mu } ^ { ( t ) } , \bar { \nu } ^ { ( t ) } )$ converge to the same point. The two players are completely symmetric and adopt the same learning rate, and require only first-order information provided by the sampling oracle. While the two algorithms resemble each other in many aspects, a key difference lies in the query and use of the sampling oracle: in each iteration, OMWU makes a single call to the sampling oracle for gradient evaluation, while PU calls the sampling oracle twice. It is worth noting that, when $\tau = 0$ (i.e., no entropy regularization is enforced), the OMWU method in Algorithm 2 reduces to the method analyzed in Rakhlin and Sridharan (2013); Daskalakis and Panageas (2018a); Wei et al. (2021b) without entropy regularization.
|
| 162 |
+
|
| 163 |
+
Remark 1. It is worth highlighting that the proposed algorithms are different from Mertikopoulos et al. (2018a), as the extragradient is only applied to the bilinear term but not the entropy regularization term. This seemingly small, but important, difference leads to a more concise closed-form update rule and a cleaner analysis, as shall be seen momentarily.
|
| 164 |
+
|
| 165 |
+
# 2.3 Performance guarantees
|
| 166 |
+
|
| 167 |
+
We are now positioned to present our main theorem concerning the last-iterate convergence of PU and OMWU for solving (2).
|
| 168 |
+
|
| 169 |
+
Theorem 1 (Last-iterate convergence of PU and OMWU). Suppose that the learning rates $\eta = \eta _ { \mathsf { P U } }$ of $P U$ in Algorithm $I$ and $\eta = \eta$ OMWU of OMWU in Algorithm 2 satisfy
|
| 170 |
+
|
| 171 |
+
$$
|
| 172 |
+
0 < \eta _ { \mathsf { P U } } \leq \frac { 1 } { \tau + 2 \left\| A \right\| _ { \infty } } , a n d 0 < \eta _ { \mathsf { O M W U } } \leq \operatorname* { m i n } \left\{ \frac { 1 } { 2 \tau + 2 \left\| A \right\| _ { \infty } } , \frac { 1 } { 4 \left\| A \right\| _ { \infty } } \right\} .
|
| 173 |
+
$$
|
| 174 |
+
|
| 175 |
+
Then for any $t \geq 0$ , the iterates $\zeta ^ { ( t ) } = ( \mu ^ { ( t ) } , \nu ^ { ( t ) } )$ and $\bar { \zeta } ^ { ( t ) } = ( \bar { \mu } ^ { ( t ) } , \bar { \nu } ^ { ( t ) } )$ of $P U$ and OMWU achieve
|
| 176 |
+
|
| 177 |
+
# • Linear convergence of policies in KL divergence and entrywise log-ratios:
|
| 178 |
+
|
| 179 |
+
$$
|
| 180 |
+
\begin{array} { r l } & { \operatorname* { m a x } \left\{ \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \| \zeta ^ { ( t ) } \big ) , \frac { 1 } { 2 } \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \| \bar { \zeta } ^ { ( t + 1 ) } \big ) \right\} \leq ( 1 - \eta \tau ) ^ { t } \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \| \zeta ^ { ( 0 ) } \big ) , } \\ & { \left\| \log \frac { \zeta ^ { ( t ) } } { \zeta _ { \tau } ^ { \star } } \right\| _ { \infty } \leq 2 ( 1 - \eta \tau ) ^ { t } \left\| \log \frac { \zeta ^ { ( 0 ) } } { \zeta _ { \tau } ^ { \star } } \right\| _ { \infty } + \frac { 8 \| A \| _ { \infty } } { \tau } ( 1 - \eta \tau ) ^ { t / 2 } \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \| \zeta ^ { ( 0 ) } \big ) ^ { 1 / 2 } . } \end{array}
|
| 181 |
+
$$
|
| 182 |
+
|
| 183 |
+
# • Linear convergence of values in optimality and duality gaps:
|
| 184 |
+
|
| 185 |
+
$$
|
| 186 |
+
\begin{array} { r l } & { \mathsf { O p t G a p } _ { \tau } ( \bar { \zeta } ^ { ( t ) } ) \leq \eta ^ { - 1 } \cdot \frac { 1 } { 1 - ( \tau + \| A \| _ { \infty } ) \eta } \cdot \frac { ( 1 - \eta \tau ) ^ { t } } { 1 - ( 1 - \eta \tau ) ^ { t } } \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \| \zeta ^ { ( 0 ) } \big ) , } \\ & { \mathsf { D u a l G a p } _ { \tau } ( \bar { \zeta } ^ { ( t ) } ) \leq \left( \eta ^ { - 1 } + 2 \tau ^ { - 1 } \| A \| _ { \infty } ^ { 2 } \right) ( 1 - \eta \tau ) ^ { t - 1 } \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \| \zeta ^ { ( 0 ) } \big ) . } \end{array}
|
| 187 |
+
$$
|
| 188 |
+
|
| 189 |
+
Remark 2. Setting $\mu ^ { ( 0 ) }$ and $\nu ^ { ( 0 ) }$ to be uniform policies leads to a universal bound
|
| 190 |
+
|
| 191 |
+
$$
|
| 192 |
+
\begin{array} { r } { \mathsf { K L } \big ( \zeta _ { \tau } ^ { \star } \| \zeta ^ { ( 0 ) } \big ) = \log | \mathcal { A } | + \log | \mathcal { B } | - \mathcal { H } ( \mu _ { \tau } ^ { \star } ) - \mathcal { H } ( \nu _ { \tau } ^ { \star } ) \leq \log | \mathcal { A } | + \log | \mathcal { B } | . } \end{array}
|
| 193 |
+
$$
|
| 194 |
+
|
| 195 |
+
Remark 3. Similar results continue to hold even when the two players use different regularization parameters $\tau _ { \mu } , \tau _ { \nu } > 0$ in (2), as long as the regularization parameter $\tau$ is replaced by max $\{ \tau _ { \mu } , \tau _ { \nu } \}$ in the upper bounds of the learning rate, and the contraction parameter is replaced by $1 - \operatorname* { m i n } \{ \tau _ { \mu } , \tau _ { \nu } \} \eta$ .
|
| 196 |
+
|
| 197 |
+
Theorem 1 characterizes the convergence of the last-iterates $\zeta ^ { ( t ) }$ and $\bar { \zeta } ^ { ( t ) }$ of PU and OMWU as long as the learning rate lies within the specified ranges. While PU doubles the number of calls to the sampling oracle, it also allows roughly as large as twice the learning rate compared with OMWU (cf. (6)). Compared with the vast literature analyzing the average-iterate performance of variants of extragradient methods, our results contribute towards characterizing the last-iterate convergence of multiplicative update methods in the presence of entropy regularization and simplex constraints, which to the best of our knowledge, are the first of its kind. Several remarks are in order.
|
| 198 |
+
|
| 199 |
+
Linear convergence to QRE. To achieve an $\epsilon$ -accurate estimate of the QRE in terms of the KL divergence, the bound (7a) tells that it is sufficient to take
|
| 200 |
+
|
| 201 |
+
$$
|
| 202 |
+
\frac { 1 } { \eta \tau } \log \left( \frac { \log | \cal { A } | + \log | \cal { B } | } { \epsilon } \right)
|
| 203 |
+
$$
|
| 204 |
+
|
| 205 |
+
iterations using either PU or OMWU. Notably, this iteration complexity does not depend on any hidden constants and only depends double logarithmically on the cardinality of action spaces, which is almost dimension-free. Maximizing the learning rate, the iteration complexity is bounded by $( 1 + \| A \| _ { \infty } / \tau ) \log ( 1 / \epsilon )$ (modulo log factors), which only depends on the ratio $\| A \| _ { \infty } / \tau$ .
|
| 206 |
+
|
| 207 |
+
Entrywise error of the policy log-ratios. Both PU and OMWU enjoy strong entrywise guarantees in the sense we can guarantee the convergence of the $\ell _ { \infty }$ norm of the log-ratios between the learned policy pair and the QRE at the same dimension-free linear rate (cf. (7b)), which suggests the policy pair converges in a somewhat uniform manner across the entire action space.
|
| 208 |
+
|
| 209 |
+
Linear convergence of optimality and duality gaps. Our theorem also establishes the last-iterate convergence of the game values in terms of the optimality gap (cf. (7c)) and the duality gap (cf. (7d)) for both PU and OMWU. In particular, as will be seen, bounding the optimality gap of matrix games turns out to be the key enabler for generalizing our algorithms to Markov games, and bounding the duality gap allows to directly translate our results to finding a NE of unregularized matrix games.
|
| 210 |
+
|
| 211 |
+
Last-iterate convergence to approximate NE. The entropy-regularized matrix game can be thought as a smooth surrogate of the unregularized matrix game (1); in particular, it is possible to find an $\epsilon$ -NE by setting $\tau$ sufficiently small in (2). According to (Zhang et al., 2020, Definition 2.1), a policy pair $\bar { \zeta } = ( \mu , \bar { \nu } )$ is an $\epsilon$ -NE if it satisfies $\begin{array} { r } { \mathsf { D u a l G a p } ( \zeta ) : = \operatorname* { m a x } _ { \mu ^ { \prime } \in \Delta ( A ) } f ( \mu ^ { \prime } , \nu ) - \operatorname* { m i n } _ { \nu ^ { \prime } \in \Delta ( B ) } f ( \mu , \nu ^ { \prime } ) \leq \epsilon } \end{array}$ .
|
| 212 |
+
|
| 213 |
+
Observe that setting $\begin{array} { r } { \tau = \frac { \epsilon / 4 } { \log | \mathcal { A } | + \log | \mathcal { B } | } } \end{array}$ guarantees that $| f _ { \tau } ( \mu , \nu ) - f ( \mu , \nu ) | < \epsilon / 4$ uniformly over $( \mu , \nu ) \in \Delta ( \mathcal { A } ) \times \Delta ( \mathcal { B } )$ in view of the boundedness of the Shannon entropy $\mathcal { H } ( \cdot )$ . Theorem 7 (cf. (7d)) also ensures that our proposed algorithms find an approximate QRE $\bar { \zeta } ^ { ( T ) }$ such that $\mathsf { D u a l G a p } _ { \tau } \bigl ( \bar { \zeta } ^ { ( T ) } \bigr ) \leq$ $\epsilon / 2$ after taking $\begin{array} { r } { T = \widetilde { O } \left( \frac { 1 } { \eta \epsilon } \right) } \end{array}$ iterations, which is no more than $\begin{array} { r } { \widetilde { O } \left( 1 + \frac { \| A \| _ { \infty } } { \epsilon } \right) } \end{array}$ iterations with optimized learning rates. It follows immediately that
|
| 214 |
+
|
| 215 |
+
[ $\mathtt { \mathtt { M a l G a p } } ( \bar { \zeta } ^ { ( T ) } ) \le \mathtt { D u a l G a p } _ { \tau } ( \bar { \zeta } ^ { ( T ) } ) + \operatorname* { m a x } _ { \mu ^ { \prime } , \nu ^ { \prime } } \Big | f _ { \tau } ( \mu ^ { \prime } , \bar { \nu } ^ { ( T ) } ) - f _ { \tau } \big ( \bar { \mu } ^ { ( T ) } , \nu ^ { \prime } \big ) - \big ( f ( \mu ^ { \prime } , \bar { \nu } ^ { ( T ) } ) - f ( \bar { \mu } ^ { ( T ) } , \nu ^ { \prime } ) \big ) \Big | \le \epsilon ,$ and therefore $\bar { \zeta } ^ { ( T ) }$ is an $\mathrm { \epsilon - N E }$ . Intriguingly, unlike prior work (Daskalakis and Panageas, 2018a; Wei et al., 2021b) that analyzed the last-iterate convergence of OMWU in the unregularized setting $\mathit { \Pi } _ { \mathcal { T } } = 0 \mathit { \Pi } _ { \mathcal { c } }$ ), our last-iterate convergence does not require the NE of (1) to be unique.
|
| 216 |
+
|
| 217 |
+
Rationality. Another attractive feature of the algorithms developed above is being rational (as introduced in Bowling and Veloso (2001)) in the sense that the algorithm returns the best-response policy of one player when the opponent takes any fixed stationary policy. More specially, in terms of matrix games, when player 2 sticks to a stationary policy $\nu$ , the update of player 1 reduces to $\mu ^ { ( t + 1 ) } ( a ) \bar { \propto } \mu ^ { ( t ) } ( a ) ^ { 1 - \bar { \eta tau } } \exp ( \eta [ A \nu ] _ { a } )$ . In this case, Theorem 1 can be established in exactly the same fashion by restricting attention only to the updates of $\mu ^ { ( t ) }$ .
|
| 218 |
+
|
| 219 |
+

|
| 220 |
+
Figure 1: Performance illustration of the PU and OMWU methods for solving entropy-regularized matrix games with $| \mathcal { A } | = | \mathcal { B } | = 1 0 0$ , where the entries of the payoff matrix $A$ is generated independently from the uniform distribution on $[ - 1 , 1 ]$ . The learning rates are fixed as $\eta = 0 . 1$ . The left panel plots various error metrics of convergence w.r.t. the iteration count with $\tau = 0 . 0 1$ , while the right panel plots these error metrics at 1000-th iteration with different choices of $\tau$ .
|
| 221 |
+
|
| 222 |
+
No-regret learning of OMWU. Besides convergence to equilibria, in game-theoretical settings, it is often desirable to design and implement no-regret algorithms, which are capable of providing black-box guarantees over arbitrary sequences played by the opponent (Cesa-Bianchi and Lugosi, 2006; Rakhlin and Sridharan, 2013). Fortunately, it turns out that entropy regularization not only accelerates the convergence, but also enables no-regret learning somewhat “for free”: it encourages exploration by putting a positive mass on every action, therefore guards against adversaries. By using a properly chosen learning rate schedule, the proposed OMWU (Algorithm 2) can be further established as a no-regret algorithm; the details can be found in (Cen et al., 2021).
|
| 223 |
+
|
| 224 |
+
# 3 Zero-sum Markov games with entropy regularization
|
| 225 |
+
|
| 226 |
+
Leveraging the success of PU and OMWU in solving the entropy-regularized matrix games, this section extends our current analysis to solve the zero-sum two-player Markov game with entropy regularization, which is again formulated as finding the equilibrium of a saddle-point optimization problem. We start by introducing its basic setup, which will be followed by the proposed policy extragradient method with its theoretical guarantees.
|
| 227 |
+
|
| 228 |
+
# 3.1 Background and problem formulation
|
| 229 |
+
|
| 230 |
+
We consider a discounted Markov Game (MG) which is defined as $\mathcal { M } = \{ { S , A , B , P , r , \gamma } \}$ , with discrete state space $s$ , action spaces of two players $\mathcal { A }$ and $\boldsymbol { B }$ , transition probability $P$ , reward function $r : \mathcal { S } \times \mathcal { A } \times \mathcal { B } [ 0 , 1 ]$ and discount factor $\gamma \in [ 0 , 1 )$ . A policy $\mu : { \mathcal { S } } \Delta ( { \mathcal { A } } )$ (resp. $\nu : S \to \Delta ( B ) )$ defines how player 1 (resp. player 2) reacts to a given state $s$ , where the probability of taking action $a \in { \mathcal { A } }$ (resp. $b \in B ,$ ) is $\mu ( a | s )$ (resp. $\nu ( b | s ) )$ . The transition probability kernel $P : \mathcal { S } \times \mathcal { A } \times \mathcal { B } \Delta ( \mathcal { S } )$ defines the dynamics of the Markov game, where $P ( s ^ { \prime } | s , a , b )$ specifies the probability of transiting to state $s ^ { \prime }$ from state $s$ when the players take actions $a$ and $b$ respectively.
|
| 231 |
+
|
| 232 |
+
Motivated by entropy regularization in Markov decision processes (MDP) (Geist et al., 2019), we consider an entropy-regularized variant of MG, where the value function is defined as
|
| 233 |
+
|
| 234 |
+
$$
|
| 235 |
+
V _ { \tau } ^ { \mu , \nu } ( s ) : = \mathbb { E } \left[ \sum _ { t = 0 } ^ { \infty } \gamma ^ { t } \left( r ( s _ { t } , a _ { t } , b _ { t } ) - \tau \log \mu ( a _ { t } | s _ { t } ) + \tau \log \nu ( b _ { t } | s _ { t } ) \right) ~ \middle | s _ { 0 } = s \right] ,
|
| 236 |
+
$$
|
| 237 |
+
|
| 238 |
+
where the quantity $\tau \geq 0$ denotes the regularization parameter, and the expectation is evaluated over the randomness of the transition kernel as well as the policies. The regularized Q-function $Q _ { \tau } ^ { \mu , \nu }$ of a
|
| 239 |
+
|
| 240 |
+
policy pair $( \mu , \nu )$ is related to $V _ { \tau } ^ { \mu , \nu }$ as
|
| 241 |
+
|
| 242 |
+
$$
|
| 243 |
+
\begin{array} { r } { Q _ { \tau } ^ { \mu , \nu } ( s , a , b ) = r ( s , a , b ) + \gamma \mathbb { E } _ { s ^ { \prime } \sim P ( \cdot \mid s , a , b ) } \bigl [ V _ { \tau } ^ { \mu , \nu } \bigl ( s ^ { \prime } \bigr ) \bigr ] . } \end{array}
|
| 244 |
+
$$
|
| 245 |
+
|
| 246 |
+
We will call $V _ { \tau } ^ { \mu , \nu }$ and $Q _ { \tau } ^ { \mu , \nu }$ the soft value function and soft $Q$ -function, respectively. A policy pair $( \mu _ { \tau } ^ { \star } , \nu _ { \tau } ^ { \star } )$ is said to be the quantal response equilibrium (QRE) of the entropy-regularized MG, if its value attains the minimax value of the entropy-regularized MG over all states $s \in S$ , i.e.
|
| 247 |
+
|
| 248 |
+
$$
|
| 249 |
+
V _ { \tau } ^ { \star } ( s ) = \operatorname* { m a x } _ { \mu } \operatorname* { m i n } _ { \nu } V _ { \tau } ^ { \mu , \nu } ( s ) = \operatorname* { m i n } _ { \nu } \operatorname* { m a x } _ { \mu } V _ { \tau } ^ { \mu , \nu } ( s ) : = V _ { \tau } ^ { \mu _ { \tau } ^ { \star } , \nu _ { \tau } ^ { \star } } ( s ) ,
|
| 250 |
+
$$
|
| 251 |
+
|
| 252 |
+
where ${ \cal V } _ { \tau } ^ { \star }$ is called the optimal minimax soft value function, and similarly $Q _ { \tau } ^ { \star } : = Q _ { \tau } ^ { \mu _ { \tau } ^ { \star } , \nu _ { \tau } ^ { \star } }$ is called the optimal minimax soft Q-function.
|
| 253 |
+
|
| 254 |
+
Goal. Our goal is to find the QRE of the entropy-regularized MG in a decentralized manner where the players only observe its own reward without accessing the opponent’s actions. By setting the regularization parameter sufficiently small $\tau$ , this also allows us to find an approximate NE of the unregularized MG.
|
| 255 |
+
|
| 256 |
+
# 3.2 From value iteration to policy extragradient methods
|
| 257 |
+
|
| 258 |
+
Entropy-regularized value iteration. It is known that classical dynamic programming approaches such as value iteration can be extended to solve MG (Perolat et al., 2015), where each iteration amounts to solving a series of matrix games for each state. Similar to the single-agent case (Cen et al., 2020), we can extend these approaches to solve the entropy-regularized MG. Setting the stage, let us introduce the per-state $\mathbf { Q }$ -value matrix $Q ( s ) : = Q ( s , \cdot , \cdot ) \bar { \in \mathbb { R } ^ { | \mathcal { A } | \times | \mathcal { B } | } }$ for every $s \in S$ , where the element indexed by the action pair $( a , b )$ is $Q ( s , a , b )$ . Similarly, we define the per-state policies $\mu ( s ) : = \mu ( \cdot | s ) \in \Delta ( { \dot { A } } )$ and $\nu ( s ) : = \nu ( \cdot | s ) \in \Delta ( B )$ for both players.
|
| 259 |
+
|
| 260 |
+
In parallel to the original Bellman operator, we denote the soft Bellman operator $\mathcal { T } _ { \tau }$ as
|
| 261 |
+
|
| 262 |
+
$$
|
| 263 |
+
\mathcal { T } _ { \tau } ( Q ) ( s , a , b ) : = r ( s , a , b ) + \gamma \mathbb { E } _ { s ^ { \prime } \sim P ( \cdot \vert s , a , b ) } \left[ \operatorname* { m a x } _ { \mu ( s ^ { \prime } ) \in \Delta ( A ) } \operatorname* { m i n } _ { \nu ( s ^ { \prime } ) \in \Delta ( B ) } f _ { \tau } \left( Q ( s ^ { \prime } ) ; \mu ( s ^ { \prime } ) , \nu ( s ^ { \prime } ) \right) \right] ,
|
| 264 |
+
$$
|
| 265 |
+
|
| 266 |
+
where for each per-state Q-value matrix $Q ( s )$ , we introduce an entropy-regularized matrix game in the form of
|
| 267 |
+
|
| 268 |
+
$$
|
| 269 |
+
\displaystyle \operatorname* { m a x } _ { \mu \in \Delta ( A ) } \operatorname* { m i n } _ { \nu \in \Delta ( B ) } f _ { \tau } \big ( Q ( s ) ; \mu ( s ) , \nu ( s ) \big ) : = \mu ( s ) ^ { \top } Q ( s ) \nu ( s ) - \tau \mathcal { H } ( \mu ( s ) ) + \tau \mathcal { H } ( \nu ( s ) ) .
|
| 270 |
+
$$
|
| 271 |
+
|
| 272 |
+
The entropy-regularized value iteration then proceeds as
|
| 273 |
+
|
| 274 |
+
$$
|
| 275 |
+
Q ^ { ( t + 1 ) } = T _ { \tau } ( Q ^ { ( t ) } ) ,
|
| 276 |
+
$$
|
| 277 |
+
|
| 278 |
+
where $Q ^ { ( 0 ) }$ is an initialization. By definition, the optimal minimax soft Q-function obeys ${ \cal T } _ { \tau } ( Q _ { \tau } ^ { \star } ) =$ $Q _ { \tau } ^ { \star }$ and therefore corresponds to the fix point of the soft Bellman operator. Given the above entropyregularized value iteration, the following lemma states its iterates contract linearly to the optimal minimax soft Q-function at a rate of the discount factor $\gamma$ .
|
| 279 |
+
|
| 280 |
+
Proposition 2. The entropy-regularized value iteration (10) converges at a linear rate, i.e. $\parallel Q ^ { ( t ) } -$
|
| 281 |
+
$Q _ { \tau } ^ { \star } \| _ { \infty } \leq \gamma ^ { t } \| Q ^ { ( 0 ) } - Q _ { \tau } ^ { \star } \| _ { \infty }$ .
|
| 282 |
+
|
| 283 |
+
Approximate value iteration via policy extragradient methods. Proposition 2 suggests that the optimal minimax soft Q-function of the entropy-regularized MG can be found by solving a series of entropy-regularized matrix games induced by $\bar { \{ Q ^ { ( t ) } \} } _ { t \geq 0 }$ in (10), a task that can be accomplished by adopting the fast extragradient methods developed earlier. To proceed, we first define the following sampling oracle, which makes it rigorous that the proposed algorithm does not require access to the Q-function of the entire MG, but only its own single-agent Q-function when playing against the opponent’s policy.
|
| 284 |
+
|
| 285 |
+
Definition 2 (Sampling oracle for Markov games). Given any policy pair $\mu ( s ) , \nu ( s )$ and $Q$ -value matrix $Q ( s )$ for any $s \in S$ , the sampling oracle returns
|
| 286 |
+
|
| 287 |
+
$$
|
| 288 |
+
[ Q ( s ) \nu ( s ) ] _ { a } = \mathbb { E } _ { b \sim \nu ( s ) } \left[ Q ( s , a , b ) \right] , \qquad a n d \qquad [ Q ( s ) ^ { \top } \mu ( s ) ] _ { b } = \mathbb { E } _ { a \sim \mu ( s ) } \left[ Q ( s , a , b ) \right]
|
| 289 |
+
$$
|
| 290 |
+
|
| 291 |
+
for any $a \in { \mathcal { A } }$ and $b \in B$ .
|
| 292 |
+
|
| 293 |
+
# Algorithm 3: Policy Extragradient Method for Entropy-regularized Markov Game
|
| 294 |
+
|
| 295 |
+
1 initialization: $Q ^ { ( 0 ) } = 0$ .
|
| 296 |
+
2 for $t = 0 , 1 , 2 , \cdots , T _ { \mathrm { m a i n } } \mathrm { { \bf d o } }$
|
| 297 |
+
3 Let $Q ^ { ( t ) }$ denote $\begin{array} { r } { Q ^ { ( t ) } ( s , a , b ) = r ( s , a , b ) + \gamma \mathbb { E } _ { s ^ { \prime } \sim P ( \cdot | s , a , b ) } V ^ { ( t ) } ( s ^ { \prime } ) . } \end{array}$ (11)
|
| 298 |
+
4 Invoke PU (Algorithm 1) or OMWU (Algorithm 2) for $T _ { \mathrm { s u b } }$ iterations to solve the following entropy-regularized matrix game for every state $s$ , where the initialization is set as uniform distributions: $\operatorname* { m a x } _ { \mu ( s ) \in \Delta ( A ) } \operatorname* { m i n } _ { \nu ( s ) \in \Delta ( B ) } f _ { \tau } \big ( Q ^ { ( t ) } ( s ) ; \mu ( s ) , \nu ( s ) \big ) .$ Return the last iterate $\bar { \mu } ^ { ( t , T _ { \mathrm { s u b } } ) } ( s ) , \bar { \nu } ^ { ( t , T _ { \mathrm { s u b } } ) } ( s )$ .
|
| 299 |
+
5 Set $V ^ { ( t + 1 ) } ( s ) = f _ { \tau } \left( Q ^ { ( t ) } ( s ) ; \bar { \mu } ^ { ( t , \bar { T } _ { \mathrm { s u b } } ) } ( s ) , \bar { \nu } ^ { ( t , \bar { T } _ { \mathrm { s u b } } ) } ( s ) \right)$ .
|
| 300 |
+
|
| 301 |
+
Encouragingly, by judiciously setting the number of iterations in both the outer loop (for updating the Q-value matrices) and the inner loop (for updating the QRE of the corresponding Q-value matrix), we are guaranteed to find the QRE of the entropy-regularized MG in a small number of iterations without solving the iteration-varying matrix games exactly, as dictated by the following theorem.
|
| 302 |
+
|
| 303 |
+
Theorem 2. Assume $| { \mathcal { A } } | \geq | { \mathcal { B } } |$ and $\tau \leq 1$ . Setting $\begin{array} { r } { \eta = \frac { 1 - \gamma } { 2 ( 1 + \tau ( \log | \mathcal { A } | + 1 - \gamma ) ) } } \end{array}$ , the total iterations (namely, the product $T _ { \mathrm { m a i n } } \cdot T _ { \mathrm { s u b , } }$ ) required for Algorithm 3 to achieve $\left\| Q ^ { ( { \vec { T } _ { \operatorname* { m i n } } } ) } - Q _ { \tau } ^ { \star } \right\| _ { \infty } \leq \epsilon$ is at most $\begin{array} { r } { O \left( \frac { ( \log | \cal { A } | + 1 / \tau ) } { ( 1 - \gamma ) ^ { 2 } } \left( \log \frac { \log | \cal { A } | } { ( 1 - \gamma ) \epsilon } \right) ^ { 2 } \right) } \end{array}$ .
|
| 304 |
+
|
| 305 |
+
Theorem 2 ensures that within $\begin{array} { r } { \widetilde O \left( \frac { 1 } { \tau ( 1 - \gamma ) ^ { 2 } } \log ^ { 2 } \left( \frac { 1 } { \epsilon } \right) \right) } \end{array}$ iterations, Algorithm 3 finds a pair of policies whose value is close to the optimal minimax soft Q-function $Q _ { \tau } ^ { \star }$ in an entrywise manner to a prescribed accuracy $\epsilon$ . Remarkably, the iteration complexity is independent of the dimensions of the state space and the action space (up to log factors).
|
| 306 |
+
|
| 307 |
+
Remark 4 (Duality gap and solving the unregularized MG). Solving the entropy-regularized MG provides a viable strategy to find an $\epsilon$ -approximate NE of the unregularized MG, where the optimality of a policy pair is typically gauged by the duality gap. Fortunately, this can be achieved similar to the case of matrix games, and we refer interested readers to Cen et al. (2021) for details.
|
| 308 |
+
|
| 309 |
+
# 4 Conclusions
|
| 310 |
+
|
| 311 |
+
This paper develops provably efficient policy extragradient methods (PU and OMWU) for entropyregularized matrix games and Markov games, whose last iterates are guaranteed to converge linearly to the quantal response equilibrium at a linear rate. Encouragingly, the rate of convergence is independent of the dimension of the problem, i.e. the sizes of the space space and the action space. In addition, the last iterates of the proposed algorithms can also be used to locate Nash equilibria for the unregularized competitive games without assuming the uniqueness of the Nash equilibria by judiciously tuning the amount of regularization. This work opens up interesting opportunities for further investigations of policy extragradient methods for solving competitive games. For example, can we develop a two-time-scale policy extragradient algorithms for Markov games where the Qfunction is updated simultaneously with the policy but potentially at a different time scale, using samples, such as in an actor-critic algorithm (Konda and Tsitsiklis, 2000)?
|
| 312 |
+
|
| 313 |
+
# Acknowledgments and Disclosure of Funding
|
| 314 |
+
|
| 315 |
+
S. Cen and Y. Chi are supported in part by the grants ONR N00014-18-1-2142 and N00014-19-1- 2404, ARO W911NF-18-1-0303, NSF CCF-1901199, CCF-2007911 and CCF-2106778. Y. Wei is supported in part by the NSF grants CCF-2007911, DMS-2147546/2015447 and CCF-2106778.
|
| 316 |
+
|
| 317 |
+
References
|
| 318 |
+
Arora, S., Hazan, E., and Kale, S. (2012). The multiplicative weights update method: a meta-algorithm and applications. Theory of Computing, 8(1):121–164.
|
| 319 |
+
Bai, Y. and Jin, C. (2020). Provable self-play algorithms for competitive reinforcement learning. In International Conference on Machine Learning, pages 551–560. PMLR.
|
| 320 |
+
Bowling, M. and Veloso, M. (2001). Rational and convergent learning in stochastic games. In Proceedings of the 17th international joint conference on Artificial intelligence-Volume 2, pages 1021–1026.
|
| 321 |
+
Cen, S., Cheng, C., Chen, Y., Wei, Y., and Chi, Y. (2020). Fast global convergence of natural policy gradient methods with entropy regularization. arXiv preprint arXiv:2007.06558.
|
| 322 |
+
Cen, S., Wei, Y., and Chi, Y. (2021). Fast policy extragradient methods for competitive games with entropy regularization. arXiv preprint arXiv:2105.15186.
|
| 323 |
+
Cesa-Bianchi, N. and Lugosi, G. (2006). Prediction, learning, and games. Cambridge university press.
|
| 324 |
+
Daskalakis, C., Deckelbaum, A., and Kim, A. (2011). Near-optimal no-regret algorithms for zero-sum games. In Proceedings of the twenty-second annual ACM-SIAM symposium on Discrete Algorithms, pages 235–254. SIAM.
|
| 325 |
+
Daskalakis, C., Foster, D. J., and Golowich, N. (2020). Independent policy gradient methods for competitive reinforcement learning. In Advances in Neural Information Processing Systems, volume 33, pages 5527–5540.
|
| 326 |
+
Daskalakis, C., Ilyas, A., Syrgkanis, V., and Zeng, H. (2018). Training GANs with optimism. In International Conference on Learning Representations (ICLR 2018).
|
| 327 |
+
Daskalakis, C. and Panageas, I. (2018a). Last-iterate convergence: Zero-sum games and constrained min-max optimization. arXiv preprint arXiv:1807.04252.
|
| 328 |
+
Daskalakis, C. and Panageas, I. (2018b). The limit points of (optimistic) gradient descent in minmax optimization. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 9256–9266.
|
| 329 |
+
Freund, Y. and Schapire, R. E. (1999). Adaptive game playing using multiplicative weights. Games and Economic Behavior, 29(1-2):79–103.
|
| 330 |
+
Geist, M., Scherrer, B., and Pietquin, O. (2019). A theory of regularized Markov decision processes. In International Conference on Machine Learning, pages 2160–2169.
|
| 331 |
+
Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial networks. arXiv preprint arXiv:1406.2661.
|
| 332 |
+
Harker, P. T. and Pang, J.-S. (1990). Finite-dimensional variational inequality and nonlinear complementarity problems: a survey of theory, algorithms and applications. Mathematical programming, 48(1):161–220.
|
| 333 |
+
Hofbauer, J. and Sandholm, W. H. (2002). On the global convergence of stochastic fictitious play. Econometrica, 70(6):2265–2294.
|
| 334 |
+
Hsieh, Y.-G., Iutzeler, F., Malick, J., and Mertikopoulos, P. (2019). On the convergence of single-call stochastic extra-gradient methods. arXiv preprint arXiv:1908.08465.
|
| 335 |
+
Konda, V. R. and Tsitsiklis, J. N. (2000). Actor-critic algorithms. In Advances in neural information processing systems, pages 1008–1014. Citeseer.
|
| 336 |
+
Korpelevich, G. M. (1976). The extragradient method for finding saddle points and other problems. Matecon, 12:747–756.
|
| 337 |
+
Lan, G. (2021). Policy mirror descent for reinforcement learning: Linear convergence, new sampling complexity, and generalized problem classes. arXiv preprint arXiv:2102.00135.
|
| 338 |
+
Lei, Q., Nagarajan, S. G., Panageas, I., and Wang, X. (2021). Last iterate convergence in no-regret learning: constrained min-max optimization for convex-concave landscapes. In International Conference on Artificial Intelligence and Statistics, pages 1441–1449. PMLR.
|
| 339 |
+
Liang, T. and Stokes, J. (2019). Interaction matters: A note on non-asymptotic local convergence of generative adversarial networks. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 907–915. PMLR.
|
| 340 |
+
Littman, M. L. (1994). Markov games as a framework for multi-agent reinforcement learning. In Machine Learning Proceedings, pages 157–163. Elsevier.
|
| 341 |
+
McKelvey, R. D. and Palfrey, T. R. (1995). Quantal response equilibria for normal form games. Games and economic behavior, 10(1):6–38.
|
| 342 |
+
Mei, J., Xiao, C., Szepesvari, C., and Schuurmans, D. (2020). On the global convergence rates of softmax policy gradient methods. arXiv preprint arXiv:2005.06392.
|
| 343 |
+
Mertikopoulos, P., Lecouat, B., Zenati, H., Foo, C.-S., Chandrasekhar, V., and Piliouras, G. (2018a). Optimistic mirror descent in saddle-point problems: Going the extra (gradient) mile. In International Conference on Learning Representations.
|
| 344 |
+
Mertikopoulos, P., Papadimitriou, C., and Piliouras, G. (2018b). Cycles in adversarial regularized learning. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 2703–2717. SIAM.
|
| 345 |
+
Mertikopoulos, P. and Sandholm, W. H. (2016). Learning in games via reinforcement and regularization. Mathematics of Operations Research, 41(4):1297–1324.
|
| 346 |
+
Mokhtari, A., Ozdaglar, A., and Pattathil, S. (2020a). A unified analysis of extra-gradient and optimistic gradient methods for saddle point problems: Proximal point approach. In International Conference on Artificial Intelligence and Statistics, pages 1497–1507. PMLR.
|
| 347 |
+
Mokhtari, A., Ozdaglar, A. E., and Pattathil, S. (2020b). Convergence rate of $O ( 1 / k )$ for optimistic gradient and extragradient methods in smooth convex-concave saddle point problems. SIAM Journal on Optimization, 30(4):3230–3251.
|
| 348 |
+
Nemirovski, A. (2004). Prox-method with rate of convergence $O ( 1 / t )$ for variational inequalities with Lipschitz continuous monotone operators and smooth convex-concave saddle point problems. SIAM Journal on Optimization, 15(1):229–251.
|
| 349 |
+
Neu, G., Jonsson, A., and Gómez, V. (2017). A unified view of entropy-regularized Markov decision processes. arXiv preprint arXiv:1705.07798.
|
| 350 |
+
Perolat, J., Scherrer, B., Piot, B., and Pietquin, O. (2015). Approximate dynamic programming for two-player zero-sum Markov games. In International Conference on Machine Learning, pages 1321–1329. PMLR.
|
| 351 |
+
Rakhlin, A. and Sridharan, K. (2013). Optimization, learning, and games with predictable sequences. arXiv preprint arXiv:1311.1869.
|
| 352 |
+
Savas, Y., Ahmadi, M., Tanaka, T., and Topcu, U. (2019). Entropy-regularized stochastic games. In 2019 IEEE 58th Conference on Decision and Control (CDC), pages 5955–5962. IEEE.
|
| 353 |
+
Shapley, L. S. (1953). Stochastic games. Proceedings of the National Academy of Sciences, 39(10):1095–1100.
|
| 354 |
+
Syrgkanis, V., Agarwal, A., Luo, H., and Schapire, R. E. (2015). Fast convergence of regularized learning in games. In Proceedings of the 28th International Conference on Neural Information Processing Systems-Volume 2, pages 2989–2997.
|
| 355 |
+
Tseng, P. (1995). On linear convergence of iterative methods for the variational inequality problem. Journal of Computational and Applied Mathematics, 60(1-2):237–252.
|
| 356 |
+
Wei, C.-Y., Lee, C.-W., Zhang, M., and Luo, H. (2021a). Last-iterate convergence of decentralized optimistic gradient descent/ascent in infinite-horizon competitive Markov games. arXiv preprint arXiv:2102.04540.
|
| 357 |
+
Wei, C.-Y., Lee, C.-W., Zhang, M., and Luo, H. (2021b). Linear last-iterate convergence in constrained saddle-point optimization. In International Conference on Learning Representations (ICLR).
|
| 358 |
+
Xie, Q., Chen, Y., Wang, Z., and Yang, Z. (2020). Learning zero-sum simultaneous-move Markov games using function approximation and correlated equilibrium. In Conference on Learning Theory, pages 3674–3682. PMLR.
|
| 359 |
+
Yadav, A., Shah, S., Xu, Z., Jacobs, D., and Goldstein, T. (2017). Stabilizing adversarial nets with prediction methods. arXiv preprint arXiv:1705.07364.
|
| 360 |
+
Zhan, W., Cen, S., Huang, B., Chen, Y., Lee, J. D., and Chi, Y. (2021). Policy mirror descent for regularized reinforcement learning: A generalized framework with linear convergence. arXiv preprint arXiv:2105.11066.
|
| 361 |
+
Zhang, K., Kakade, S., Basar, T., and Yang, L. (2020). Model-based multi-agent RL in zero-sum Markov games with near-optimal sample complexity. Advances in Neural Information Processing Systems, 33.
|
| 362 |
+
Zhao, Y., Tian, Y., Lee, J. D., and Du, S. S. (2021). Provably efficient policy gradient methods for two-player zero-sum Markov games. arXiv preprint arXiv:2102.08903.
|
parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo_content_list.json
ADDED
|
@@ -0,0 +1,1593 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Fast Policy Extragradient Methods for Competitive Games with Entropy Regularization ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
189,
|
| 8 |
+
122,
|
| 9 |
+
808,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Shicong Cen Carnegie Mellon University shicongc@andrew.cmu.edu ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
187,
|
| 19 |
+
227,
|
| 20 |
+
385,
|
| 21 |
+
268
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Yuting Wei University of Pennsylvania ytwei@wharton.upenn.edu ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
406,
|
| 30 |
+
226,
|
| 31 |
+
604,
|
| 32 |
+
268
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Yuejie Chi Carnegie Mellon University yuejiechi@cmu.edu ",
|
| 39 |
+
"bbox": [
|
| 40 |
+
625,
|
| 41 |
+
226,
|
| 42 |
+
810,
|
| 43 |
+
268
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "Abstract ",
|
| 50 |
+
"text_level": 1,
|
| 51 |
+
"bbox": [
|
| 52 |
+
462,
|
| 53 |
+
304,
|
| 54 |
+
535,
|
| 55 |
+
320
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "This paper investigates the problem of computing the equilibrium of competitive games, which is often modeled as a constrained saddle-point optimization problem with probability simplex constraints. Despite recent efforts in understanding the last-iterate convergence of extragradient methods in the unconstrained setting, the theoretical underpinnings of these methods in the constrained settings, especially those using multiplicative updates, remain highly inadequate, even when the objective function is bilinear. Motivated by the algorithmic role of entropy regularization in single-agent reinforcement learning and game theory, we develop provably efficient extragradient methods to find the quantal response equilibrium (QRE)—which are solutions to zero-sum two-player matrix games with entropy regularization—at a linear rate. The proposed algorithms can be implemented in a decentralized manner, where each player executes symmetric and multiplicative updates iteratively using its own payoff without observing the opponent’s actions directly. In addition, by controlling the knob of entropy regularization, the proposed algorithms can locate an approximate Nash equilibrium of the unregularized matrix game at a sublinear rate without assuming the Nash equilibrium to be unique. Our methods also lead to efficient policy extragradient algorithms for solving entropy-regularized zero-sum Markov games at a linear rate. All of our convergence rates are nearly dimension-free, which are independent of the size of the state and action spaces up to logarithm factors, highlighting the positive role of entropy regularization for accelerating convergence. ",
|
| 62 |
+
"bbox": [
|
| 63 |
+
233,
|
| 64 |
+
338,
|
| 65 |
+
766,
|
| 66 |
+
627
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "1 Introduction ",
|
| 73 |
+
"text_level": 1,
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
660,
|
| 77 |
+
310,
|
| 78 |
+
676
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Finding the equilibrium of competitive games, which can be viewed as constrained saddle-point optimization problems with probability simplex constraints, lies at the heart of modern machine learning and decision making paradigms such as Generative Adversarial Networks (GANs) (Goodfellow et al., 2014), competitive reinforcement learning (RL) (Littman, 1994), game theory (Shapley, 1953), adversarial training (Mertikopoulos et al., 2018b), to name a few. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
688,
|
| 88 |
+
825,
|
| 89 |
+
757
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "text",
|
| 95 |
+
"text": "In this paper, we study one of the most basic forms of competitive games, namely two-player zero-sum games, in both the matrix setting and the Markov setting. Our goal is to find the equilibrium policies of both players in an independent and decentralized manner (Daskalakis et al., 2020; Wei et al., 2021a) with guaranteed last-iterate convergence. Namely, each player will execute symmetric and independent updates iteratively using its own payoff without observing the opponent’s actions directly, and the final policies of the iterative process should be a close approximation to the equilibrium up to any prescribed precision. This kind of algorithms is more advantageous and versatile especially in federated environments, as it requires neither prior coordination between the players like twotimescale algorithms, nor a central controller to collect and disseminate the policies of all the players, which are often unavailable due to privacy constraints. ",
|
| 96 |
+
"bbox": [
|
| 97 |
+
174,
|
| 98 |
+
763,
|
| 99 |
+
825,
|
| 100 |
+
902
|
| 101 |
+
],
|
| 102 |
+
"page_idx": 0
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"type": "text",
|
| 106 |
+
"text": "1.1 Last-iterate convergence in competitive games ",
|
| 107 |
+
"text_level": 1,
|
| 108 |
+
"bbox": [
|
| 109 |
+
178,
|
| 110 |
+
92,
|
| 111 |
+
531,
|
| 112 |
+
106
|
| 113 |
+
],
|
| 114 |
+
"page_idx": 1
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"type": "text",
|
| 118 |
+
"text": "In recent years, there have been significant progresses in understanding the last-iterate convergence of simple iterative algorithms for unconstrained saddle-point optimization, where one is interested in bounding the sub-optimality of the last iterate of the algorithm, rather than say, the ergodic iterate — which is the average of all the iterations — that are commonly studied in the earlier literature. This shift of focus is motivated, for example, by the infeasibility of averaging large machine learning models in training GANs (Goodfellow et al., 2014). While vanilla Gradient Descent / Ascent (GDA) may diverge or cycle even for bilinear matrix games (Daskalakis et al., 2018), quite remarkably, small modifications lead to guaranteed last-iterate convergence to the equilibrium in a non-asymptotic fashion. A flurry of algorithms is proposed, including Optimistic Gradient Descent Ascent (OGDA) (Rakhlin and Sridharan, 2013; Daskalakis and Panageas, 2018b; Wei et al., 2021b), predictive updates (Yadav et al., 2017), implicit updates (Liang and Stokes, 2019), and more. Several unified analyses of these algorithms have been carried out (see, e.g. Mokhtari et al. (2020a); Liang and Stokes (2019) and references therein), where these methods in principle all make clever extrapolation of the local curvature in a predictive manner to accelerate convergence. With slight abuse of terminology, in this paper, we refer to this ensemble of algorithms as extragradient methods (Korpelevich, 1976; Tseng, 1995; Mertikopoulos et al., 2018a; Harker and Pang, 1990). ",
|
| 119 |
+
"bbox": [
|
| 120 |
+
173,
|
| 121 |
+
113,
|
| 122 |
+
826,
|
| 123 |
+
334
|
| 124 |
+
],
|
| 125 |
+
"page_idx": 1
|
| 126 |
+
},
|
| 127 |
+
{
|
| 128 |
+
"type": "text",
|
| 129 |
+
"text": "However, saddle-point optimization in the constrained setting, which includes competitive games as a special case, remains largely under-explored even for bilinear matrix games. While it is possible to reformulate constrained bilinear games to unconstrained ones using softmax parameterization of the probability simplex, this approach falls short of preserving the bilinear structure and convexconcave properties in the original problem, which are crucial to the convergence of gradient methods. Therefore, there is a strong necessity of understanding and developing improved extragradient methods in the constrained setting. Daskalakis and Panageas (2018a) proposed the optimistic variant of the multiplicative weight updates (MWU) method (Arora et al., 2012) – which is extremely natural and popular for optimizing over probability simplexes – called Optimistic Multiplicative Weight Updates (OMWU), and established the asymptotic last-iterate convergence of OMWU for matrix games. Very recently, Wei et al. (2021b) established non-asymptotic last-iterate convergences of OMWU. However, these last-iterate convergence results require the Nash equilibrium to be unique, and cannot be applied to problems with multiple Nash equilibria. ",
|
| 130 |
+
"bbox": [
|
| 131 |
+
173,
|
| 132 |
+
340,
|
| 133 |
+
826,
|
| 134 |
+
520
|
| 135 |
+
],
|
| 136 |
+
"page_idx": 1
|
| 137 |
+
},
|
| 138 |
+
{
|
| 139 |
+
"type": "text",
|
| 140 |
+
"text": "1.2 Our contributions ",
|
| 141 |
+
"text_level": 1,
|
| 142 |
+
"bbox": [
|
| 143 |
+
176,
|
| 144 |
+
536,
|
| 145 |
+
338,
|
| 146 |
+
550
|
| 147 |
+
],
|
| 148 |
+
"page_idx": 1
|
| 149 |
+
},
|
| 150 |
+
{
|
| 151 |
+
"type": "text",
|
| 152 |
+
"text": "Motivated by the algorithmic role of entropy regularization in single-agent RL (Neu et al., 2017; Geist et al., 2019; Cen et al., 2020) as well as its wide use in game theory to account for imperfect and noisy information (McKelvey and Palfrey, 1995; Savas et al., 2019), we initiate the design and analysis of extragradient algorithms using multiplicative updates for finding the quantal response equilibrium (QRE), which are solutions to competitive games with entropy regularization (McKelvey and Palfrey, 1995). While finding QRE is of interest in its own right, by controlling the knob of entropy regularization, the QRE provides a close approximation to the Nash equilibrium (NE), and in turn acts as a smoothing scheme for finding the NE. Our contributions are summarized below. ",
|
| 153 |
+
"bbox": [
|
| 154 |
+
174,
|
| 155 |
+
558,
|
| 156 |
+
825,
|
| 157 |
+
669
|
| 158 |
+
],
|
| 159 |
+
"page_idx": 1
|
| 160 |
+
},
|
| 161 |
+
{
|
| 162 |
+
"type": "text",
|
| 163 |
+
"text": "• Near dimension-free last-iterate convergence to QRE of entropy-regularized matrix games. We propose two policy extragradient algorithms to solve entropy-regularized matrix games, namely the Predictive Update (PU) and OMWU methods, where both players execute symmetric and multiplicative updates without knowing the entire payoff matrix nor the opponent’s actions. Encouragingly, we show that the last iterate of the proposed algorithms converges to the unique QRE at a linear rate that is almost independent of the size of the action spaces. Roughly speaking, to find an $\\epsilon$ -optimal QRE in terms of Kullback-Leibler (KL) divergence, it takes no more than $\\begin{array} { r } { \\widetilde O \\left( \\frac { 1 } { \\eta \\tau } \\log \\left( \\frac { 1 } { \\epsilon } \\right) \\right) } \\end{array}$ iterations, where ${ \\widetilde { O } } ( \\cdot )$ hides logarithmic dependencies. Here, $\\tau$ is the regularization parameter, and $\\eta$ is the learning rate of both players. Maximizing the learning rate, the iteration complexity is bounded by $\\widetilde { O } \\left( ( 1 + \\| A \\| _ { \\infty } / \\tau ) \\log ( 1 / \\epsilon ) \\right)$ , where $\\left\\| A \\right\\| _ { \\infty } = \\operatorname* { m a x } _ { i , j } \\left| A _ { i , j } \\right|$ is the $\\ell _ { \\infty }$ norm of the payoff matrix $A$ . ",
|
| 164 |
+
"bbox": [
|
| 165 |
+
173,
|
| 166 |
+
675,
|
| 167 |
+
825,
|
| 168 |
+
830
|
| 169 |
+
],
|
| 170 |
+
"page_idx": 1
|
| 171 |
+
},
|
| 172 |
+
{
|
| 173 |
+
"type": "text",
|
| 174 |
+
"text": "• Last-iterate convergence to $\\epsilon$ -NE of unregularized matrix games without uniqueness assumption. The QRE provides an accurate approximation to the NE by setting the entropy regularization $\\tau$ sufficiently small, therefore our result directly translates to finding a NE with last-iterate convergence guarantee. Roughly speaking, to find an $\\epsilon$ -NE (Zhang et al., 2020, Definition 2.1), it takes no more than $\\begin{array} { r } { \\widetilde { O } \\left( 1 + \\frac { \\| A \\| _ { \\infty } } { \\epsilon } \\right) } \\end{array}$ iterations with optimized learning rates, which is again independent of the size of the action spaces up to logarithmic factors. Unlike prior literature (Daskalakis and Panageas, 2018a; Wei et al., 2021b), our last-iterate convergence guarantee does not require the NE to be unique. ",
|
| 175 |
+
"bbox": [
|
| 176 |
+
174,
|
| 177 |
+
835,
|
| 178 |
+
825,
|
| 179 |
+
915
|
| 180 |
+
],
|
| 181 |
+
"page_idx": 1
|
| 182 |
+
},
|
| 183 |
+
{
|
| 184 |
+
"type": "table",
|
| 185 |
+
"img_path": "images/7cd63fa4d3a31b21549b1c3cba203d1cfde83dc305268984dcc577e9d1c629f3.jpg",
|
| 186 |
+
"table_caption": [
|
| 187 |
+
"Table 1: Comparisons of last-iterate convergence of the proposed entropy-regularized PU and OMWU methods with prior results for finding $\\epsilon$ -QRE or $\\mathrm { \\epsilon - N E }$ of competitive matrix games. We note that the convergence rates of unregularized OMWU established in Wei et al. (2021b) are problemdependent, and scale at least polynomially on the size of the action spaces. Desirable features in the last two columns are highlighted in blue. "
|
| 188 |
+
],
|
| 189 |
+
"table_footnote": [],
|
| 190 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Equilibriumtype</td><td rowspan=1 colspan=1>Method</td><td rowspan=1 colspan=1>Convergence rate</td><td rowspan=1 colspan=1>Dimension-free</td><td rowspan=1 colspan=1>Requireunique NE</td></tr><tr><td rowspan=1 colspan=1>E-QRE</td><td rowspan=1 colspan=1>PU&OMWU(this work)</td><td rowspan=1 colspan=1>linear</td><td rowspan=1 colspan=1>yes</td><td rowspan=1 colspan=1>n/a</td></tr><tr><td rowspan=3 colspan=1>e-NE</td><td rowspan=1 colspan=1>OMWU(Daskalakis and Panageas, 2018a)</td><td rowspan=1 colspan=1>asymptotic</td><td rowspan=1 colspan=1>no</td><td rowspan=1 colspan=1>yes</td></tr><tr><td rowspan=1 colspan=1>OMWU(Wei et al., 2021b)</td><td rowspan=1 colspan=1>sublinear + linear</td><td rowspan=1 colspan=1>no</td><td rowspan=1 colspan=1>yes</td></tr><tr><td rowspan=1 colspan=1>PU&OMWU(this work)</td><td rowspan=1 colspan=1>sublinear</td><td rowspan=1 colspan=1>yes</td><td rowspan=1 colspan=1>no</td></tr></table>",
|
| 191 |
+
"bbox": [
|
| 192 |
+
176,
|
| 193 |
+
90,
|
| 194 |
+
825,
|
| 195 |
+
241
|
| 196 |
+
],
|
| 197 |
+
"page_idx": 2
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"type": "text",
|
| 201 |
+
"text": "",
|
| 202 |
+
"bbox": [
|
| 203 |
+
173,
|
| 204 |
+
357,
|
| 205 |
+
823,
|
| 206 |
+
386
|
| 207 |
+
],
|
| 208 |
+
"page_idx": 2
|
| 209 |
+
},
|
| 210 |
+
{
|
| 211 |
+
"type": "text",
|
| 212 |
+
"text": "• Extensions to two-player zero-sum Markov games. By connecting value iteration with matrix games, we propose a policy extragradient method for solving infinite-horizon discounted entropyregularized zero-sum Markov games, which finds an $\\epsilon$ -optimal minimax soft Q-function—in terms of $\\ell _ { \\infty }$ error—in at most $\\begin{array} { r } { \\widetilde O \\left( \\frac { 1 } { \\tau ( 1 - \\gamma ) ^ { 2 } } \\log ^ { 2 } \\left( \\frac { 1 } { \\epsilon } \\right) \\right) } \\end{array}$ iterations, where $\\gamma \\in ( 0 , 1 )$ is the discount factor. ",
|
| 213 |
+
"bbox": [
|
| 214 |
+
174,
|
| 215 |
+
392,
|
| 216 |
+
825,
|
| 217 |
+
457
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "To the best of our knowledge, our paper is the first that develops policy extragradient algorithms for solving entropy-regularized competitive games with multiplicative updates and dimension-free linear last-iterate convergence, and demonstrates entropy regularization as a smoothing technique to find $\\mathrm { \\epsilon - N E }$ without the uniqueness assumption. Table 1 provides detailed comparisons of the proposed methods with prior arts for solving matrix games. Our results highlight the positive role of entropy regularization for accelerating convergence and safeguarding against imperfect information in competitive games. We defer the complete proof of our results to Cen et al. (2021). ",
|
| 224 |
+
"bbox": [
|
| 225 |
+
174,
|
| 226 |
+
463,
|
| 227 |
+
825,
|
| 228 |
+
560
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "1.3 Related works ",
|
| 235 |
+
"text_level": 1,
|
| 236 |
+
"bbox": [
|
| 237 |
+
174,
|
| 238 |
+
582,
|
| 239 |
+
313,
|
| 240 |
+
595
|
| 241 |
+
],
|
| 242 |
+
"page_idx": 2
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"type": "text",
|
| 246 |
+
"text": "Our work lies at the intersection of saddle-point optimization, game theory, and reinforcement learning. In what follows, we discuss a few topics that are closely related to ours. ",
|
| 247 |
+
"bbox": [
|
| 248 |
+
174,
|
| 249 |
+
606,
|
| 250 |
+
823,
|
| 251 |
+
633
|
| 252 |
+
],
|
| 253 |
+
"page_idx": 2
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"type": "text",
|
| 257 |
+
"text": "Unregularized matrix game. Freund and Schapire (1999) showed that many standard methods such as GDA and MWU have a converging average duality gap at the rate of $O ( 1 / \\sqrt { T } )$ , which is improved to $O ( 1 / T )$ by considering optimistic variants of these methods, such as OGDA and OMWU (Rakhlin and Sridharan, 2013; Daskalakis et al., 2011; Syrgkanis et al., 2015). However, the last-iterate convergence of these methods are less understood until recently (Daskalakis and Panageas, 2018a; Wei et al., 2021b). In particular, under the assumption that the NE is unique for the unregularized matrix game, Daskalakis and Panageas (2018a) showed the asymptotic convergence of the last iterate of OMWU to the unique equilibrium, and Wei et al. (2021b) showed the last iterate of OMWU achieves a linear rate of convergence after an initial phase of sublinear convergence, however the rates therein can be highly pessimistic in terms of the problem dimension, while our rate for entropy-regularized OMWU is dimension-free up to logarithmic factors. ",
|
| 258 |
+
"bbox": [
|
| 259 |
+
174,
|
| 260 |
+
640,
|
| 261 |
+
825,
|
| 262 |
+
794
|
| 263 |
+
],
|
| 264 |
+
"page_idx": 2
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"type": "text",
|
| 268 |
+
"text": "Saddle-point optimization. Considerable progress has been made towards understanding OGDA and extragradient (EG) methods in the unconstrained convex-concave saddle-point optimization with general objective functions (Mokhtari et al., 2020a,b; Nemirovski, 2004; Liang and Stokes, 2019). However, the last-iterate convergence of constrained convex-concave saddle-point optimization still lacks theoretical understanding in general and most works fall short of characterizing a finite-time convergence result. In particular, Mertikopoulos et al. (2018a) demonstrated the asymptotic lastiterate convergence of EG, and Hsieh et al. (2019) investigated similar questions for single-call EG algorithms. Lei et al. (2021) showed that OMWU converges to the equilibrium locally without an explicit rate. Wei et al. (2021b) showed that the last-iterate of OGDA converges linearly for strongly-convex strongly-concave constrained saddle-point optimization with an explicit rate. ",
|
| 269 |
+
"bbox": [
|
| 270 |
+
174,
|
| 271 |
+
800,
|
| 272 |
+
825,
|
| 273 |
+
911
|
| 274 |
+
],
|
| 275 |
+
"page_idx": 2
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"type": "text",
|
| 279 |
+
"text": "",
|
| 280 |
+
"bbox": [
|
| 281 |
+
171,
|
| 282 |
+
90,
|
| 283 |
+
823,
|
| 284 |
+
119
|
| 285 |
+
],
|
| 286 |
+
"page_idx": 3
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"type": "text",
|
| 290 |
+
"text": "Entropy regularization in RL and games. In single-agent RL, the role of entropy regularization as an algorithmic mechanism to encourage exploration and accelerate convergence has been investigated extensively (Neu et al., 2017; Geist et al., 2019; Mei et al., 2020; Cen et al., 2020; Lan, 2021; Zhan et al., 2021). Turning to the game setting, entropy regularization is used to account for imperfect information in the seminal work of McKelvey and Palfrey (1995) that introduced the QRE, and a few representative works on entropy and more general regularizations in games include Savas et al. (2019); Hofbauer and Sandholm (2002); Mertikopoulos and Sandholm (2016). ",
|
| 291 |
+
"bbox": [
|
| 292 |
+
173,
|
| 293 |
+
126,
|
| 294 |
+
825,
|
| 295 |
+
223
|
| 296 |
+
],
|
| 297 |
+
"page_idx": 3
|
| 298 |
+
},
|
| 299 |
+
{
|
| 300 |
+
"type": "text",
|
| 301 |
+
"text": "Zero-sum Markov games. There have been a significant recent interest in developing provably efficient self-play algorithms for Markov games, including model-based algorithms (Perolat et al., 2015; Zhang et al., 2020), value-based algorithms (Bai and Jin, 2020; Xie et al., 2020), and policybased algorithms (Daskalakis et al., 2020; Wei et al., 2021a; Zhao et al., 2021). The iteration complexities in prior works (Perolat et al., 2015; Daskalakis et al., 2020; Wei et al., 2021a; Zhao et al., 2021) depend on various notions of concentrability coefficient and therefore can scale quite pessimistically with the problem dimension. Our approach can be regarded as a policy-based algorithm to approximate value iteration, which can be implemented in a decentralized manner with symmetric and multiplicative updates from both players, and the iteration complexity is almost independent of the size of the state-action space. ",
|
| 302 |
+
"bbox": [
|
| 303 |
+
173,
|
| 304 |
+
228,
|
| 305 |
+
825,
|
| 306 |
+
367
|
| 307 |
+
],
|
| 308 |
+
"page_idx": 3
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"type": "text",
|
| 312 |
+
"text": "Notation. We denote by $\\Delta ( \\mathcal { A } )$ the probability simplex over the set $\\mathcal { A }$ . We overload the functions such as $\\log ( \\cdot )$ and $\\exp ( \\cdot )$ to take vector inputs with the understanding that the function is applied in an entrywise manner. For instance, given any vector $z = [ z _ { i } ] _ { 1 \\leq i \\leq n } \\in \\mathbb { R } ^ { n }$ , the notation $\\exp ( z )$ denotes $\\exp ( z ) : = [ \\exp ( z _ { i } ) ] _ { 1 \\leq i \\leq n }$ ; other functions are defined analogously. Given two probability distributions $\\mu$ and $\\mu ^ { \\prime }$ over $\\mathcal { A }$ , the KL divergence from $\\mu ^ { \\prime }$ to $\\mu$ is defined by $\\begin{array} { r } { \\mathsf { K L } ( \\mu \\parallel \\mu ^ { \\prime } ) : = \\sum _ { a \\in \\mathcal { A } } \\mu ( a ) \\log \\frac { \\mu ( a ) } { \\mu ^ { \\prime } ( a ) } } \\end{array}$ . Given a matrix $A$ , $\\| A \\| _ { \\infty }$ is used to denote entrywise maximum norm, namely, $\\left\\| A \\right\\| _ { \\infty } = \\operatorname* { m a x } _ { i , j } \\left| A _ { i , j } \\right|$ . The all-one vector is denoted as 1. ",
|
| 313 |
+
"bbox": [
|
| 314 |
+
173,
|
| 315 |
+
372,
|
| 316 |
+
826,
|
| 317 |
+
479
|
| 318 |
+
],
|
| 319 |
+
"page_idx": 3
|
| 320 |
+
},
|
| 321 |
+
{
|
| 322 |
+
"type": "text",
|
| 323 |
+
"text": "2 Zero-sum matrix games with entropy regularization ",
|
| 324 |
+
"text_level": 1,
|
| 325 |
+
"bbox": [
|
| 326 |
+
174,
|
| 327 |
+
497,
|
| 328 |
+
638,
|
| 329 |
+
515
|
| 330 |
+
],
|
| 331 |
+
"page_idx": 3
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"type": "text",
|
| 335 |
+
"text": "We first consider a two-player zero-sum game with bilinear objective and probability simplex constraints, and demonstrate the positive role of entropy regularization in solving this problem. Throughout this paper, let $\\mathcal { A } = \\{ 1 , \\dotsc , m \\}$ and $\\boldsymbol { B } = \\{ 1 , \\ldots , \\bar { n } \\}$ be the action spaces of each player. ",
|
| 336 |
+
"bbox": [
|
| 337 |
+
174,
|
| 338 |
+
530,
|
| 339 |
+
826,
|
| 340 |
+
573
|
| 341 |
+
],
|
| 342 |
+
"page_idx": 3
|
| 343 |
+
},
|
| 344 |
+
{
|
| 345 |
+
"type": "text",
|
| 346 |
+
"text": "2.1 Background and problem formulation ",
|
| 347 |
+
"text_level": 1,
|
| 348 |
+
"bbox": [
|
| 349 |
+
176,
|
| 350 |
+
585,
|
| 351 |
+
478,
|
| 352 |
+
599
|
| 353 |
+
],
|
| 354 |
+
"page_idx": 3
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"type": "text",
|
| 358 |
+
"text": "Zero-sum two-player matrix game. The focal point of this subsection is a constrained two-player zero-sum matrix game, which can be formulated as the following min-max problem (or saddle point optimization problem): ",
|
| 359 |
+
"bbox": [
|
| 360 |
+
174,
|
| 361 |
+
606,
|
| 362 |
+
825,
|
| 363 |
+
648
|
| 364 |
+
],
|
| 365 |
+
"page_idx": 3
|
| 366 |
+
},
|
| 367 |
+
{
|
| 368 |
+
"type": "equation",
|
| 369 |
+
"img_path": "images/a84cc1579eae155c735ed137faef7ebb0f81235573bb5194136ab35bb171c683.jpg",
|
| 370 |
+
"text": "$$\n\\operatorname* { m a x } _ { \\mu \\in \\Delta ( \\mathcal { A } ) } \\operatorname* { m i n } _ { \\nu \\in \\Delta ( \\mathcal { B } ) } f ( \\mu , \\nu ) : = \\mu ^ { \\top } A \\nu ,\n$$",
|
| 371 |
+
"text_format": "latex",
|
| 372 |
+
"bbox": [
|
| 373 |
+
385,
|
| 374 |
+
656,
|
| 375 |
+
609,
|
| 376 |
+
684
|
| 377 |
+
],
|
| 378 |
+
"page_idx": 3
|
| 379 |
+
},
|
| 380 |
+
{
|
| 381 |
+
"type": "text",
|
| 382 |
+
"text": "where $\\textbf { \\textit { A } } \\in \\mathbb { R } ^ { m \\times n }$ denotes the payoff matrix, $\\mu ~ \\in ~ \\Delta ( { \\mathcal { A } } )$ and $\\nu ~ \\in ~ \\Delta ( B )$ stand for the mixed/randomized policies of each player, defined respectively as distributions over the probability simplex $\\Delta ( \\mathcal { A } )$ and $\\Delta ( B )$ . A pair of policies $( \\mu ^ { \\star } , \\nu ^ { \\star } )$ is said to be a Nash equilibrium (NE) of (1) if $f ( \\mu ^ { \\star } , \\nu ) \\geq f ( \\mu ^ { \\star } , \\nu ^ { \\star } ) \\geq f ( \\mu , \\nu ^ { \\star } )$ for all $( \\mu , \\nu ) \\in \\Delta ( \\mathcal { A } ) \\times \\Delta ( \\mathcal { B } )$ . ",
|
| 383 |
+
"bbox": [
|
| 384 |
+
173,
|
| 385 |
+
693,
|
| 386 |
+
825,
|
| 387 |
+
751
|
| 388 |
+
],
|
| 389 |
+
"page_idx": 3
|
| 390 |
+
},
|
| 391 |
+
{
|
| 392 |
+
"type": "text",
|
| 393 |
+
"text": "Entropy-regularized zero-sum two-player matrix game. There is no shortage of scenarios where the payoff matrix $A$ might not be known perfectly. In an attempt to accommodate imperfect knowledge of $A$ , McKelvey and Palfrey (1995) proposed a seminal extension to the Nash equilibrium called the quantal response equilibrium $( Q R E )$ when the payoffs are perturbed by Gumbel-distributed noise. Formally, this amounts to solving the following matrix game with entropy regularization (Mertikopoulos and Sandholm, 2016): ",
|
| 394 |
+
"bbox": [
|
| 395 |
+
173,
|
| 396 |
+
756,
|
| 397 |
+
825,
|
| 398 |
+
839
|
| 399 |
+
],
|
| 400 |
+
"page_idx": 3
|
| 401 |
+
},
|
| 402 |
+
{
|
| 403 |
+
"type": "equation",
|
| 404 |
+
"img_path": "images/b4d50912a83264c73c0021a71bd120be521f76f0af08da4df50bf194bda49736.jpg",
|
| 405 |
+
"text": "$$\n\\operatorname* { m a x } _ { \\mu \\in \\Delta ( \\mathcal { A } ) } \\operatorname* { m i n } _ { \\nu \\in \\Delta ( \\mathcal { B } ) } f _ { \\tau } ( \\mu , \\nu ) : = \\mu ^ { \\top } A \\nu + \\tau \\mathcal { H } ( \\mu ) - \\tau \\mathcal { H } ( \\nu ) ,\n$$",
|
| 406 |
+
"text_format": "latex",
|
| 407 |
+
"bbox": [
|
| 408 |
+
318,
|
| 409 |
+
847,
|
| 410 |
+
678,
|
| 411 |
+
875
|
| 412 |
+
],
|
| 413 |
+
"page_idx": 3
|
| 414 |
+
},
|
| 415 |
+
{
|
| 416 |
+
"type": "text",
|
| 417 |
+
"text": "where $\\begin{array} { r } { \\mathcal { H } ( \\pi ) = - \\sum _ { i } \\pi _ { i } \\log ( \\pi _ { i } ) } \\end{array}$ denotes the Shannon entropy of a distribution $\\pi$ , and $\\tau \\geq 0$ is the regularization parameter. As is well known, the optimal solution $( \\mu _ { \\tau } ^ { \\star } , \\nu _ { \\tau } ^ { \\star } )$ to (2), dubbed as the QRE, ",
|
| 418 |
+
"bbox": [
|
| 419 |
+
173,
|
| 420 |
+
882,
|
| 421 |
+
823,
|
| 422 |
+
912
|
| 423 |
+
],
|
| 424 |
+
"page_idx": 3
|
| 425 |
+
},
|
| 426 |
+
{
|
| 427 |
+
"type": "text",
|
| 428 |
+
"text": "is unique whenever $\\tau > 0$ (due to the presence of strong concavity/convexity), which satisfies the following fixed point equations: ",
|
| 429 |
+
"bbox": [
|
| 430 |
+
169,
|
| 431 |
+
90,
|
| 432 |
+
825,
|
| 433 |
+
119
|
| 434 |
+
],
|
| 435 |
+
"page_idx": 4
|
| 436 |
+
},
|
| 437 |
+
{
|
| 438 |
+
"type": "equation",
|
| 439 |
+
"img_path": "images/71555ba46498e3920949431f5517f0cfdf53213e02f54c9ada168880b5c7fd49.jpg",
|
| 440 |
+
"text": "$$\n\\begin{array} { r } { \\biggr \\{ \\mu _ { \\tau } ^ { \\star } ( a ) = \\frac { \\exp ( [ A \\nu _ { \\tau } ^ { \\star } ] a / \\tau ) } { \\sum _ { a = 1 } ^ { m } \\exp ( [ A \\nu _ { \\tau } ^ { \\star } ] a / \\tau ) } \\propto \\exp ( [ A \\nu _ { \\tau } ^ { \\star } ] a / \\tau ) , \\qquad \\mathrm { f o r ~ a l l } \\ a \\in \\mathcal { A } , } \\\\ { \\nu _ { \\tau } ^ { \\star } ( b ) = \\frac { \\exp ( - [ A ^ { \\top } \\mu _ { \\tau } ^ { \\star } ] b / \\tau ) } { \\sum _ { b = 1 } ^ { n } \\exp ( - [ A ^ { \\top } \\mu _ { \\tau } ^ { \\star } ] b / \\tau ) } \\propto \\exp ( - [ A ^ { \\top } \\mu _ { \\tau } ^ { \\star } ] b / \\tau ) , \\quad \\mathrm { f o r ~ a l l } \\ b \\in \\mathcal { B } . } \\end{array}\n$$",
|
| 441 |
+
"text_format": "latex",
|
| 442 |
+
"bbox": [
|
| 443 |
+
259,
|
| 444 |
+
123,
|
| 445 |
+
736,
|
| 446 |
+
178
|
| 447 |
+
],
|
| 448 |
+
"page_idx": 4
|
| 449 |
+
},
|
| 450 |
+
{
|
| 451 |
+
"type": "text",
|
| 452 |
+
"text": "Goal. We aim to efficiently compute the QRE of the entropy-regularized matrix game in a decentralized manner, and investigate how an efficient solver of QRE can be leveraged to find a NE of the unregularized matrix game (1). Namely, we only assume access to “first-order information” as opposed to full knowledge of the payoff matrix $A$ or the actions of the opponent. The information received by each player is formally described in the following sampling oracle. ",
|
| 453 |
+
"bbox": [
|
| 454 |
+
173,
|
| 455 |
+
186,
|
| 456 |
+
826,
|
| 457 |
+
257
|
| 458 |
+
],
|
| 459 |
+
"page_idx": 4
|
| 460 |
+
},
|
| 461 |
+
{
|
| 462 |
+
"type": "text",
|
| 463 |
+
"text": "Definition 1 (Sampling oracle for matrix games). For any policy pair $( \\mu , \\nu )$ and payoff matrix $A$ the sampling oracle returns the exact values of $\\mu ^ { \\top } A$ and $A \\nu$ . ",
|
| 464 |
+
"bbox": [
|
| 465 |
+
173,
|
| 466 |
+
260,
|
| 467 |
+
823,
|
| 468 |
+
290
|
| 469 |
+
],
|
| 470 |
+
"page_idx": 4
|
| 471 |
+
},
|
| 472 |
+
{
|
| 473 |
+
"type": "text",
|
| 474 |
+
"text": "Additional notation. For notational convenience, we let $\\zeta$ represent the concatenation of $\\mu \\in \\mathbb { R } ^ { | \\mathcal { A } | }$ and $\\nu \\in \\mathbb { R } ^ { | B | }$ , namely, $\\zeta = ( \\mu , \\nu )$ . The solution to (2), which is specified in (3), is denoted by $\\zeta _ { \\tau } ^ { \\star } = ( \\mu _ { \\tau } ^ { \\star } , \\nu _ { \\tau } ^ { \\star } )$ . For any $\\zeta = ( \\mu , \\nu )$ and $\\zeta ^ { \\prime } = ( \\mu ^ { \\prime } , \\nu ^ { \\prime } )$ , we shall often abuse the notation and let $\\mathsf { K L } \\big ( \\zeta \\| \\zeta ^ { \\prime } \\big ) = \\mathsf { K L } \\big ( \\mu \\| \\mu ^ { \\prime } \\big ) + \\mathsf { K L } \\big ( \\nu \\| \\nu ^ { \\prime } \\big )$ . The duality gap of the entropy-regularized matrix game (2) at $\\zeta = ( \\mu , \\nu )$ is defined as $\\begin{array} { r } { \\mathsf { D u a l G a p } _ { \\tau } ( \\zeta ) = \\operatorname* { m a x } _ { \\mu ^ { \\prime } \\in \\Delta ( A ) } f _ { \\tau } ( \\mu ^ { \\prime } , \\nu ) - \\operatorname* { m i n } _ { \\nu ^ { \\prime } \\in \\Delta ( B ) } f _ { \\tau } ( \\mu , \\nu ^ { \\prime } ) } \\end{array}$ which is clearly nonnegative and $\\mathsf { D u a l G a p } _ { \\tau } ( \\boldsymbol { \\zeta } _ { \\tau } ^ { \\star } ) = 0$ . Similarly, let the optimality gap of the entropyregularized matrix game (2) at $\\zeta = ( \\mu , \\nu )$ be $\\mathsf { O p t G a p } ( \\zeta ) = \\left| f _ { \\tau } ( \\mu , \\nu ) - f _ { \\tau } ( \\mu _ { \\tau } ^ { \\star } , \\nu _ { \\tau } ^ { \\star } ) \\right|$ . ",
|
| 475 |
+
"bbox": [
|
| 476 |
+
173,
|
| 477 |
+
299,
|
| 478 |
+
826,
|
| 479 |
+
406
|
| 480 |
+
],
|
| 481 |
+
"page_idx": 4
|
| 482 |
+
},
|
| 483 |
+
{
|
| 484 |
+
"type": "text",
|
| 485 |
+
"text": "2.2 Proposed extragradient methods: PU and OMWU ",
|
| 486 |
+
"text_level": 1,
|
| 487 |
+
"bbox": [
|
| 488 |
+
173,
|
| 489 |
+
415,
|
| 490 |
+
563,
|
| 491 |
+
430
|
| 492 |
+
],
|
| 493 |
+
"page_idx": 4
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"type": "text",
|
| 497 |
+
"text": "To begin, assume we are given a pair of policies $z _ { 1 } \\in \\Delta ( \\mathcal { A } )$ , $z _ { 2 } \\in \\Delta ( B )$ employed by each player respectively. If we proceed with fictitious play, i.e. player 1 (resp. player 2) aims to optimize its own policy by assuming the opponent’s policy is fixed as $z _ { 2 }$ (resp. $z _ { 1 }$ ), the saddle-point optimization problem (2) is then decoupled into two independent min/max optimization problems: ",
|
| 498 |
+
"bbox": [
|
| 499 |
+
173,
|
| 500 |
+
435,
|
| 501 |
+
826,
|
| 502 |
+
492
|
| 503 |
+
],
|
| 504 |
+
"page_idx": 4
|
| 505 |
+
},
|
| 506 |
+
{
|
| 507 |
+
"type": "equation",
|
| 508 |
+
"img_path": "images/3ae3b2a7eb130747659835faa8d01046b129d51408a7d9c08d0dd4aa00a0a848.jpg",
|
| 509 |
+
"text": "$$\n\\operatorname* { m a x } _ { \\mu \\in \\Delta ( A ) } \\mu ^ { \\top } A z _ { 2 } + \\tau \\mathcal { H } ( \\mu ) - \\tau \\mathcal { H } ( z _ { 2 } ) \\qquad \\mathrm { a n d } \\qquad \\operatorname* { m i n } _ { \\nu \\in \\Delta ( B ) } z _ { 1 } ^ { \\top } A \\nu + \\tau \\mathcal { H } ( z _ { 1 } ) - \\tau \\mathcal { H } ( \\nu ) ,\n$$",
|
| 510 |
+
"text_format": "latex",
|
| 511 |
+
"bbox": [
|
| 512 |
+
212,
|
| 513 |
+
496,
|
| 514 |
+
781,
|
| 515 |
+
523
|
| 516 |
+
],
|
| 517 |
+
"page_idx": 4
|
| 518 |
+
},
|
| 519 |
+
{
|
| 520 |
+
"type": "text",
|
| 521 |
+
"text": "which are naturally solved via mirror descent / ascent with KL divergence. Specifically, one step of mirror descent / ascent takes the form ",
|
| 522 |
+
"bbox": [
|
| 523 |
+
176,
|
| 524 |
+
527,
|
| 525 |
+
825,
|
| 526 |
+
556
|
| 527 |
+
],
|
| 528 |
+
"page_idx": 4
|
| 529 |
+
},
|
| 530 |
+
{
|
| 531 |
+
"type": "equation",
|
| 532 |
+
"img_path": "images/a1e0157530e8a4c8bc9d0803138f0eba224ba8f93f9a46dce23f9f9876f0d1d5.jpg",
|
| 533 |
+
"text": "$$\n\\begin{array} { r } { \\left\\{ \\begin{array} { l l } { \\mu ^ { ( t + 1 ) } ( a ) \\propto \\mu ^ { ( t ) } ( a ) ^ { 1 - \\eta \\tau } \\exp ( \\eta [ A z _ { 2 } ] _ { a } ) , } & { \\mathrm { f o r ~ a l l ~ } a \\in \\mathcal { A } , } \\\\ { \\nu ^ { ( t + 1 ) } ( b ) \\propto \\nu ^ { ( t ) } ( b ) ^ { 1 - \\eta \\tau } \\exp ( - \\eta [ A ^ { \\top } z _ { 1 } ] _ { b } ) , } & { \\mathrm { f o r ~ a l l ~ } b \\in \\mathcal { B } , } \\end{array} \\right. } \\end{array}\n$$",
|
| 534 |
+
"text_format": "latex",
|
| 535 |
+
"bbox": [
|
| 536 |
+
295,
|
| 537 |
+
558,
|
| 538 |
+
700,
|
| 539 |
+
609
|
| 540 |
+
],
|
| 541 |
+
"page_idx": 4
|
| 542 |
+
},
|
| 543 |
+
{
|
| 544 |
+
"type": "text",
|
| 545 |
+
"text": "where $\\eta$ is the learning rate. The above update rule forms the basis of our algorithm design. ",
|
| 546 |
+
"bbox": [
|
| 547 |
+
171,
|
| 548 |
+
612,
|
| 549 |
+
772,
|
| 550 |
+
627
|
| 551 |
+
],
|
| 552 |
+
"page_idx": 4
|
| 553 |
+
},
|
| 554 |
+
{
|
| 555 |
+
"type": "text",
|
| 556 |
+
"text": "Motivation: a form of implicit updates with linear convergence. It turns out, if we could select the policy pair $( z _ { 1 } , z _ { 2 } ) = \\hat { \\zeta } ^ { ( t + 1 ) } : \\stackrel { \\bullet } { = } ( \\mu ^ { ( t + 1 ) } , \\nu ^ { ( t + 1 ) } )$ as the ones to be taken in the future, and call the resulting update rule as the Implicit Update (IU) method: ",
|
| 557 |
+
"bbox": [
|
| 558 |
+
174,
|
| 559 |
+
632,
|
| 560 |
+
825,
|
| 561 |
+
675
|
| 562 |
+
],
|
| 563 |
+
"page_idx": 4
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"type": "equation",
|
| 567 |
+
"img_path": "images/57dd0f2231a644892d0d759e737298493ccd260916dd0c48818739b8fbeb7aef.jpg",
|
| 568 |
+
"text": "$$\n\\begin{array} { r } { \\left\\{ \\mu ^ { ( t + 1 ) } ( a ) \\propto \\mu ^ { ( t ) } ( a ) ^ { 1 - \\eta \\tau } \\exp ( \\eta [ A \\nu ^ { ( t + 1 ) } ] _ { a } ) , \\right. \\left. \\mathrm { f o r ~ a l l } \\ : a \\in \\mathcal { A } , \\right. } \\\\ { \\nu ^ { ( t + 1 ) } ( b ) \\propto \\nu ^ { ( t ) } ( b ) ^ { 1 - \\eta \\tau } \\exp ( - \\eta [ A ^ { \\top } \\mu ^ { ( t + 1 ) } ] _ { b } ) , \\left. \\mathrm { f o r ~ a l l } \\ : b \\in \\mathcal { B } . \\right. } \\end{array}\n$$",
|
| 569 |
+
"text_format": "latex",
|
| 570 |
+
"bbox": [
|
| 571 |
+
351,
|
| 572 |
+
679,
|
| 573 |
+
784,
|
| 574 |
+
729
|
| 575 |
+
],
|
| 576 |
+
"page_idx": 4
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"type": "text",
|
| 580 |
+
"text": "Though unrealistic — since it uses the future updates — it leads to a one-step convergence to the QRE when $\\eta = 1 / \\tau$ (see the optimality condition in (3)). Encouragingly, we have the following linear convergence guarantee of IU when adopting a general learning rate. ",
|
| 581 |
+
"bbox": [
|
| 582 |
+
176,
|
| 583 |
+
733,
|
| 584 |
+
823,
|
| 585 |
+
776
|
| 586 |
+
],
|
| 587 |
+
"page_idx": 4
|
| 588 |
+
},
|
| 589 |
+
{
|
| 590 |
+
"type": "text",
|
| 591 |
+
"text": "Proposition 1 (Linear convergence of IU). Assume $0 < \\eta \\leq 1 / \\tau$ , then for all $t \\geq 0$ , the iterates $\\zeta ^ { ( t ) } : = ( \\mu ^ { ( t ) } , \\nu ^ { ( t ) } )$ of the $I U$ method in (5) satisfy $\\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\parallel \\zeta ^ { ( t ) } \\big ) \\leq ( 1 - \\eta \\tau ) ^ { t } \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\parallel \\zeta ^ { ( 0 ) } \\big )$ . ",
|
| 592 |
+
"bbox": [
|
| 593 |
+
174,
|
| 594 |
+
779,
|
| 595 |
+
821,
|
| 596 |
+
811
|
| 597 |
+
],
|
| 598 |
+
"page_idx": 4
|
| 599 |
+
},
|
| 600 |
+
{
|
| 601 |
+
"type": "text",
|
| 602 |
+
"text": "In words, the IU method achieves an appealing linear rate of convergence that is independent of the problem dimension. Motivated by this observation, we seek to design algorithms where the policies $( z _ { 1 } , z _ { 2 } )$ employed in (4) serve as good predictions of $( \\mu ^ { ( t + 1 ) } , \\nu ^ { ( t + \\bar { 1 ) } } )$ , such that the resulting algorithms are both practical and retain the appealing convergence rate of IU. ",
|
| 603 |
+
"bbox": [
|
| 604 |
+
173,
|
| 605 |
+
819,
|
| 606 |
+
825,
|
| 607 |
+
877
|
| 608 |
+
],
|
| 609 |
+
"page_idx": 4
|
| 610 |
+
},
|
| 611 |
+
{
|
| 612 |
+
"type": "text",
|
| 613 |
+
"text": "Proposed algorithms. We propose two extragradient algorithms for solving the entropy-regularized matrix game, namely the Predictive Update $( P U )$ method and the Optimistic Multiplicative Weights ",
|
| 614 |
+
"bbox": [
|
| 615 |
+
173,
|
| 616 |
+
882,
|
| 617 |
+
823,
|
| 618 |
+
912
|
| 619 |
+
],
|
| 620 |
+
"page_idx": 4
|
| 621 |
+
},
|
| 622 |
+
{
|
| 623 |
+
"type": "text",
|
| 624 |
+
"text": "Algorithm 1: The PU method ",
|
| 625 |
+
"text_level": 1,
|
| 626 |
+
"bbox": [
|
| 627 |
+
174,
|
| 628 |
+
94,
|
| 629 |
+
374,
|
| 630 |
+
109
|
| 631 |
+
],
|
| 632 |
+
"page_idx": 5
|
| 633 |
+
},
|
| 634 |
+
{
|
| 635 |
+
"type": "text",
|
| 636 |
+
"text": "Algorithm 2: The OMWU method ",
|
| 637 |
+
"text_level": 1,
|
| 638 |
+
"bbox": [
|
| 639 |
+
500,
|
| 640 |
+
94,
|
| 641 |
+
732,
|
| 642 |
+
109
|
| 643 |
+
],
|
| 644 |
+
"page_idx": 5
|
| 645 |
+
},
|
| 646 |
+
{
|
| 647 |
+
"type": "text",
|
| 648 |
+
"text": "1 initialization: $\\mu ^ { ( 0 ) }$ , $\\nu ^ { ( 0 ) }$ ",
|
| 649 |
+
"bbox": [
|
| 650 |
+
163,
|
| 651 |
+
113,
|
| 652 |
+
339,
|
| 653 |
+
128
|
| 654 |
+
],
|
| 655 |
+
"page_idx": 5
|
| 656 |
+
},
|
| 657 |
+
{
|
| 658 |
+
"type": "text",
|
| 659 |
+
"text": "2 for $t = 0 , 1 , 2 , \\cdots$ do ",
|
| 660 |
+
"bbox": [
|
| 661 |
+
160,
|
| 662 |
+
128,
|
| 663 |
+
323,
|
| 664 |
+
142
|
| 665 |
+
],
|
| 666 |
+
"page_idx": 5
|
| 667 |
+
},
|
| 668 |
+
{
|
| 669 |
+
"type": "text",
|
| 670 |
+
"text": "2 for $t = 0 , 1 , 2 , \\cdots$ do ",
|
| 671 |
+
"bbox": [
|
| 672 |
+
485,
|
| 673 |
+
130,
|
| 674 |
+
645,
|
| 675 |
+
142
|
| 676 |
+
],
|
| 677 |
+
"page_idx": 5
|
| 678 |
+
},
|
| 679 |
+
{
|
| 680 |
+
"type": "text",
|
| 681 |
+
"text": "3 Update $\\bar { \\mu }$ and $\\bar { \\nu }$ according to ",
|
| 682 |
+
"bbox": [
|
| 683 |
+
160,
|
| 684 |
+
143,
|
| 685 |
+
388,
|
| 686 |
+
156
|
| 687 |
+
],
|
| 688 |
+
"page_idx": 5
|
| 689 |
+
},
|
| 690 |
+
{
|
| 691 |
+
"type": "text",
|
| 692 |
+
"text": "3 ",
|
| 693 |
+
"bbox": [
|
| 694 |
+
485,
|
| 695 |
+
145,
|
| 696 |
+
495,
|
| 697 |
+
155
|
| 698 |
+
],
|
| 699 |
+
"page_idx": 5
|
| 700 |
+
},
|
| 701 |
+
{
|
| 702 |
+
"type": "text",
|
| 703 |
+
"text": "Update $\\bar { \\mu }$ and $\\bar { \\nu }$ according to ",
|
| 704 |
+
"bbox": [
|
| 705 |
+
517,
|
| 706 |
+
143,
|
| 707 |
+
714,
|
| 708 |
+
156
|
| 709 |
+
],
|
| 710 |
+
"page_idx": 5
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"type": "equation",
|
| 714 |
+
"img_path": "images/2682d9ed9998363f783340595232da07def6ae1b62593530707c1b063a745524.jpg",
|
| 715 |
+
"text": "$$\n\\begin{array} { r } { \\left\\{ \\bar { \\mu } ^ { ( t + 1 ) } ( a ) \\propto { \\mu } ^ { ( t ) } ( a ) ^ { 1 - \\eta \\tau } \\exp ( \\eta [ A \\nu ^ { ( t ) } ] _ { a } ) , \\right. \\qquad } \\\\ { \\left. \\bar { \\nu } ^ { ( t + 1 ) } ( b ) \\propto \\nu ^ { ( t ) } ( b ) ^ { 1 - \\eta \\tau } \\exp ( - \\eta [ A ^ { \\top } { \\mu } ^ { ( t ) } ] _ { b } ) . \\right. } \\end{array}\n$$",
|
| 716 |
+
"text_format": "latex",
|
| 717 |
+
"bbox": [
|
| 718 |
+
205,
|
| 719 |
+
165,
|
| 720 |
+
482,
|
| 721 |
+
203
|
| 722 |
+
],
|
| 723 |
+
"page_idx": 5
|
| 724 |
+
},
|
| 725 |
+
{
|
| 726 |
+
"type": "equation",
|
| 727 |
+
"img_path": "images/07b2e5847f06f569d996bfff2cef68ba4e7e99ebccd9cf274448fbddd63318e4.jpg",
|
| 728 |
+
"text": "$$\n\\begin{array} { r } { \\left\\{ \\bar { \\mu } ^ { ( t + 1 ) } ( a ) \\propto { \\mu } ^ { ( t ) } ( a ) ^ { 1 - \\eta \\tau } \\exp ( \\eta [ A \\bar { \\nu } ^ { ( t ) } ] _ { a } ) , \\right. \\mathrm { ~ } } \\\\ { \\left. \\bar { \\nu } ^ { ( t + 1 ) } ( b ) \\propto \\nu ^ { ( t ) } ( b ) ^ { 1 - \\eta \\tau } \\exp ( - \\eta [ A ^ { \\top } \\bar { \\mu } ^ { ( t ) } ] _ { b } ) . \\right. } \\end{array}\n$$",
|
| 729 |
+
"text_format": "latex",
|
| 730 |
+
"bbox": [
|
| 731 |
+
531,
|
| 732 |
+
164,
|
| 733 |
+
808,
|
| 734 |
+
203
|
| 735 |
+
],
|
| 736 |
+
"page_idx": 5
|
| 737 |
+
},
|
| 738 |
+
{
|
| 739 |
+
"type": "text",
|
| 740 |
+
"text": "4 Update $\\mu$ and $\\nu$ according to ",
|
| 741 |
+
"bbox": [
|
| 742 |
+
160,
|
| 743 |
+
210,
|
| 744 |
+
390,
|
| 745 |
+
224
|
| 746 |
+
],
|
| 747 |
+
"page_idx": 5
|
| 748 |
+
},
|
| 749 |
+
{
|
| 750 |
+
"type": "text",
|
| 751 |
+
"text": "Update $\\mu$ and $\\nu$ according to ",
|
| 752 |
+
"bbox": [
|
| 753 |
+
524,
|
| 754 |
+
210,
|
| 755 |
+
714,
|
| 756 |
+
224
|
| 757 |
+
],
|
| 758 |
+
"page_idx": 5
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"type": "equation",
|
| 762 |
+
"img_path": "images/213e7eea9f5e11eb13db31a851e4ce7321f6ff73c1f9234656441c96d93ddf5f.jpg",
|
| 763 |
+
"text": "$$\n\\begin{array} { r } { \\left\\{ \\begin{array} { l l } { \\mu ^ { ( t + 1 ) } ( a ) \\propto \\mu ^ { ( t ) } ( a ) ^ { 1 - \\eta \\tau } \\exp ( \\eta [ A \\bar { \\nu } ^ { ( t + 1 ) } ] _ { a } ) , } \\\\ { \\nu ^ { ( t + 1 ) } ( b ) \\propto \\nu ^ { ( t ) } ( b ) ^ { 1 - \\eta \\tau } \\exp ( - \\eta [ A ^ { \\top } \\bar { \\mu } ^ { ( t + 1 ) } ] _ { b } ) . } \\end{array} \\right. } \\end{array}\n$$",
|
| 764 |
+
"text_format": "latex",
|
| 765 |
+
"bbox": [
|
| 766 |
+
205,
|
| 767 |
+
232,
|
| 768 |
+
496,
|
| 769 |
+
271
|
| 770 |
+
],
|
| 771 |
+
"page_idx": 5
|
| 772 |
+
},
|
| 773 |
+
{
|
| 774 |
+
"type": "equation",
|
| 775 |
+
"img_path": "images/46d9ad77b55d1cc670745026c47b3d350c5530aabb855c6c7b775b9a5c3422f4.jpg",
|
| 776 |
+
"text": "$$\n\\begin{array} { r } { \\left\\{ \\begin{array} { l l } { \\mu ^ { ( t + 1 ) } ( a ) \\propto \\mu ^ { ( t ) } ( a ) ^ { 1 - \\eta \\tau } \\exp ( \\eta [ A \\bar { \\nu } ^ { ( t + 1 ) } ] _ { a } ) , } \\\\ { \\nu ^ { ( t + 1 ) } ( b ) \\propto \\nu ^ { ( t ) } ( b ) ^ { 1 - \\eta \\tau } \\exp ( - \\eta [ A ^ { \\top } \\bar { \\mu } ^ { ( t + 1 ) } ] _ { b } ) . } \\end{array} \\right. } \\end{array}\n$$",
|
| 777 |
+
"text_format": "latex",
|
| 778 |
+
"bbox": [
|
| 779 |
+
531,
|
| 780 |
+
232,
|
| 781 |
+
823,
|
| 782 |
+
271
|
| 783 |
+
],
|
| 784 |
+
"page_idx": 5
|
| 785 |
+
},
|
| 786 |
+
{
|
| 787 |
+
"type": "text",
|
| 788 |
+
"text": "Update (OMWU) method, the latter adapted from Rakhlin and Sridharan (2013); Daskalakis et al. (2011). Detailed procedures can be found in Algorithm 1 and Algorithm 2, respectively. On a high level, both algorithms maintain two intertwined sequences $\\{ ( \\mu ^ { ( \\bar { t } ) } , \\nu ^ { ( t ) } ) \\} _ { t \\geq 0 }$ and $\\{ ( \\bar { \\mu } ^ { ( \\dot { t } ) } , \\bar { \\nu } ^ { ( t ) } ) \\} _ { t \\geq 0 }$ , and in each iteration $t = 0 , 1 , \\ldots$ , proceed in two steps: ",
|
| 789 |
+
"bbox": [
|
| 790 |
+
173,
|
| 791 |
+
318,
|
| 792 |
+
826,
|
| 793 |
+
377
|
| 794 |
+
],
|
| 795 |
+
"page_idx": 5
|
| 796 |
+
},
|
| 797 |
+
{
|
| 798 |
+
"type": "text",
|
| 799 |
+
"text": "• The midpoint $( \\bar { \\mu } ^ { ( t + 1 ) } , \\bar { \\nu } ^ { ( t + 1 ) } )$ serves as a prediction of $( \\mu ^ { ( t + 1 ) } , \\nu ^ { ( t + 1 ) } )$ by running one step of mirror descent / ascent (cf. (4)) from either $( z _ { 1 } , z _ { 2 } ) = ( \\mu ^ { ( t ) } , \\nu ^ { ( t ) } )$ (for PU) or $( z _ { 1 } , z _ { 2 } ) = ( \\bar { \\mu } ^ { ( t ) } , \\bar { \\nu } ^ { ( t ) } )$ (for OMWU). ",
|
| 800 |
+
"bbox": [
|
| 801 |
+
173,
|
| 802 |
+
382,
|
| 803 |
+
826,
|
| 804 |
+
428
|
| 805 |
+
],
|
| 806 |
+
"page_idx": 5
|
| 807 |
+
},
|
| 808 |
+
{
|
| 809 |
+
"type": "text",
|
| 810 |
+
"text": "• The update of $( \\boldsymbol { \\mu } ^ { ( t + 1 ) } , \\boldsymbol { \\nu } ^ { ( t + 1 ) } )$ then mimics the implicit update (5) using the prediction $( \\bar { \\mu } ^ { ( t + 1 ) } , \\bar { \\nu } ^ { ( t + 1 ) } )$ obtained above. ",
|
| 811 |
+
"bbox": [
|
| 812 |
+
173,
|
| 813 |
+
431,
|
| 814 |
+
821,
|
| 815 |
+
463
|
| 816 |
+
],
|
| 817 |
+
"page_idx": 5
|
| 818 |
+
},
|
| 819 |
+
{
|
| 820 |
+
"type": "text",
|
| 821 |
+
"text": "When the proposed algorithms converge, both $( \\mu ^ { ( t ) } , \\nu ^ { ( t ) } )$ and $( \\bar { \\mu } ^ { ( t ) } , \\bar { \\nu } ^ { ( t ) } )$ converge to the same point. The two players are completely symmetric and adopt the same learning rate, and require only first-order information provided by the sampling oracle. While the two algorithms resemble each other in many aspects, a key difference lies in the query and use of the sampling oracle: in each iteration, OMWU makes a single call to the sampling oracle for gradient evaluation, while PU calls the sampling oracle twice. It is worth noting that, when $\\tau = 0$ (i.e., no entropy regularization is enforced), the OMWU method in Algorithm 2 reduces to the method analyzed in Rakhlin and Sridharan (2013); Daskalakis and Panageas (2018a); Wei et al. (2021b) without entropy regularization. ",
|
| 822 |
+
"bbox": [
|
| 823 |
+
173,
|
| 824 |
+
469,
|
| 825 |
+
825,
|
| 826 |
+
583
|
| 827 |
+
],
|
| 828 |
+
"page_idx": 5
|
| 829 |
+
},
|
| 830 |
+
{
|
| 831 |
+
"type": "text",
|
| 832 |
+
"text": "Remark 1. It is worth highlighting that the proposed algorithms are different from Mertikopoulos et al. (2018a), as the extragradient is only applied to the bilinear term but not the entropy regularization term. This seemingly small, but important, difference leads to a more concise closed-form update rule and a cleaner analysis, as shall be seen momentarily. ",
|
| 833 |
+
"bbox": [
|
| 834 |
+
174,
|
| 835 |
+
588,
|
| 836 |
+
825,
|
| 837 |
+
643
|
| 838 |
+
],
|
| 839 |
+
"page_idx": 5
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"type": "text",
|
| 843 |
+
"text": "2.3 Performance guarantees ",
|
| 844 |
+
"text_level": 1,
|
| 845 |
+
"bbox": [
|
| 846 |
+
174,
|
| 847 |
+
664,
|
| 848 |
+
383,
|
| 849 |
+
679
|
| 850 |
+
],
|
| 851 |
+
"page_idx": 5
|
| 852 |
+
},
|
| 853 |
+
{
|
| 854 |
+
"type": "text",
|
| 855 |
+
"text": "We are now positioned to present our main theorem concerning the last-iterate convergence of PU and OMWU for solving (2). ",
|
| 856 |
+
"bbox": [
|
| 857 |
+
173,
|
| 858 |
+
685,
|
| 859 |
+
823,
|
| 860 |
+
714
|
| 861 |
+
],
|
| 862 |
+
"page_idx": 5
|
| 863 |
+
},
|
| 864 |
+
{
|
| 865 |
+
"type": "text",
|
| 866 |
+
"text": "Theorem 1 (Last-iterate convergence of PU and OMWU). Suppose that the learning rates $\\eta = \\eta _ { \\mathsf { P U } }$ of $P U$ in Algorithm $I$ and $\\eta = \\eta$ OMWU of OMWU in Algorithm 2 satisfy ",
|
| 867 |
+
"bbox": [
|
| 868 |
+
173,
|
| 869 |
+
718,
|
| 870 |
+
823,
|
| 871 |
+
748
|
| 872 |
+
],
|
| 873 |
+
"page_idx": 5
|
| 874 |
+
},
|
| 875 |
+
{
|
| 876 |
+
"type": "equation",
|
| 877 |
+
"img_path": "images/1da1f00199904a4b6c97f5e9c76f9a3df834798c6291155ebe93e90d726b2886.jpg",
|
| 878 |
+
"text": "$$\n0 < \\eta _ { \\mathsf { P U } } \\leq \\frac { 1 } { \\tau + 2 \\left\\| A \\right\\| _ { \\infty } } , a n d 0 < \\eta _ { \\mathsf { O M W U } } \\leq \\operatorname* { m i n } \\left\\{ \\frac { 1 } { 2 \\tau + 2 \\left\\| A \\right\\| _ { \\infty } } , \\frac { 1 } { 4 \\left\\| A \\right\\| _ { \\infty } } \\right\\} .\n$$",
|
| 879 |
+
"text_format": "latex",
|
| 880 |
+
"bbox": [
|
| 881 |
+
220,
|
| 882 |
+
756,
|
| 883 |
+
776,
|
| 884 |
+
791
|
| 885 |
+
],
|
| 886 |
+
"page_idx": 5
|
| 887 |
+
},
|
| 888 |
+
{
|
| 889 |
+
"type": "text",
|
| 890 |
+
"text": "Then for any $t \\geq 0$ , the iterates $\\zeta ^ { ( t ) } = ( \\mu ^ { ( t ) } , \\nu ^ { ( t ) } )$ and $\\bar { \\zeta } ^ { ( t ) } = ( \\bar { \\mu } ^ { ( t ) } , \\bar { \\nu } ^ { ( t ) } )$ of $P U$ and OMWU achieve ",
|
| 891 |
+
"bbox": [
|
| 892 |
+
173,
|
| 893 |
+
799,
|
| 894 |
+
825,
|
| 895 |
+
818
|
| 896 |
+
],
|
| 897 |
+
"page_idx": 5
|
| 898 |
+
},
|
| 899 |
+
{
|
| 900 |
+
"type": "text",
|
| 901 |
+
"text": "• Linear convergence of policies in KL divergence and entrywise log-ratios: ",
|
| 902 |
+
"text_level": 1,
|
| 903 |
+
"bbox": [
|
| 904 |
+
174,
|
| 905 |
+
821,
|
| 906 |
+
702,
|
| 907 |
+
838
|
| 908 |
+
],
|
| 909 |
+
"page_idx": 5
|
| 910 |
+
},
|
| 911 |
+
{
|
| 912 |
+
"type": "equation",
|
| 913 |
+
"img_path": "images/9c9ca3a937c7498ac455080e8128e2e20c04105f9efdeb52eea00b6054ed9608.jpg",
|
| 914 |
+
"text": "$$\n\\begin{array} { r l } & { \\operatorname* { m a x } \\left\\{ \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\| \\zeta ^ { ( t ) } \\big ) , \\frac { 1 } { 2 } \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\| \\bar { \\zeta } ^ { ( t + 1 ) } \\big ) \\right\\} \\leq ( 1 - \\eta \\tau ) ^ { t } \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\| \\zeta ^ { ( 0 ) } \\big ) , } \\\\ & { \\left\\| \\log \\frac { \\zeta ^ { ( t ) } } { \\zeta _ { \\tau } ^ { \\star } } \\right\\| _ { \\infty } \\leq 2 ( 1 - \\eta \\tau ) ^ { t } \\left\\| \\log \\frac { \\zeta ^ { ( 0 ) } } { \\zeta _ { \\tau } ^ { \\star } } \\right\\| _ { \\infty } + \\frac { 8 \\| A \\| _ { \\infty } } { \\tau } ( 1 - \\eta \\tau ) ^ { t / 2 } \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\| \\zeta ^ { ( 0 ) } \\big ) ^ { 1 / 2 } . } \\end{array}\n$$",
|
| 915 |
+
"text_format": "latex",
|
| 916 |
+
"bbox": [
|
| 917 |
+
212,
|
| 918 |
+
843,
|
| 919 |
+
758,
|
| 920 |
+
912
|
| 921 |
+
],
|
| 922 |
+
"page_idx": 5
|
| 923 |
+
},
|
| 924 |
+
{
|
| 925 |
+
"type": "text",
|
| 926 |
+
"text": "• Linear convergence of values in optimality and duality gaps: ",
|
| 927 |
+
"text_level": 1,
|
| 928 |
+
"bbox": [
|
| 929 |
+
173,
|
| 930 |
+
90,
|
| 931 |
+
607,
|
| 932 |
+
107
|
| 933 |
+
],
|
| 934 |
+
"page_idx": 6
|
| 935 |
+
},
|
| 936 |
+
{
|
| 937 |
+
"type": "equation",
|
| 938 |
+
"img_path": "images/3702b093fedebbe58ca32ee1f3510f7779f94f8298da5387a3f661e7ff847f92.jpg",
|
| 939 |
+
"text": "$$\n\\begin{array} { r l } & { \\mathsf { O p t G a p } _ { \\tau } ( \\bar { \\zeta } ^ { ( t ) } ) \\leq \\eta ^ { - 1 } \\cdot \\frac { 1 } { 1 - ( \\tau + \\| A \\| _ { \\infty } ) \\eta } \\cdot \\frac { ( 1 - \\eta \\tau ) ^ { t } } { 1 - ( 1 - \\eta \\tau ) ^ { t } } \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\| \\zeta ^ { ( 0 ) } \\big ) , } \\\\ & { \\mathsf { D u a l G a p } _ { \\tau } ( \\bar { \\zeta } ^ { ( t ) } ) \\leq \\left( \\eta ^ { - 1 } + 2 \\tau ^ { - 1 } \\| A \\| _ { \\infty } ^ { 2 } \\right) ( 1 - \\eta \\tau ) ^ { t - 1 } \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\| \\zeta ^ { ( 0 ) } \\big ) . } \\end{array}\n$$",
|
| 940 |
+
"text_format": "latex",
|
| 941 |
+
"bbox": [
|
| 942 |
+
251,
|
| 943 |
+
109,
|
| 944 |
+
745,
|
| 945 |
+
174
|
| 946 |
+
],
|
| 947 |
+
"page_idx": 6
|
| 948 |
+
},
|
| 949 |
+
{
|
| 950 |
+
"type": "text",
|
| 951 |
+
"text": "Remark 2. Setting $\\mu ^ { ( 0 ) }$ and $\\nu ^ { ( 0 ) }$ to be uniform policies leads to a universal bound ",
|
| 952 |
+
"bbox": [
|
| 953 |
+
168,
|
| 954 |
+
176,
|
| 955 |
+
717,
|
| 956 |
+
193
|
| 957 |
+
],
|
| 958 |
+
"page_idx": 6
|
| 959 |
+
},
|
| 960 |
+
{
|
| 961 |
+
"type": "equation",
|
| 962 |
+
"img_path": "images/9dc3554546cd0d7e8714d2a8742b8e7b54e42d716313e7ff2b236bd90a374b64.jpg",
|
| 963 |
+
"text": "$$\n\\begin{array} { r } { \\mathsf { K L } \\big ( \\zeta _ { \\tau } ^ { \\star } \\| \\zeta ^ { ( 0 ) } \\big ) = \\log | \\mathcal { A } | + \\log | \\mathcal { B } | - \\mathcal { H } ( \\mu _ { \\tau } ^ { \\star } ) - \\mathcal { H } ( \\nu _ { \\tau } ^ { \\star } ) \\leq \\log | \\mathcal { A } | + \\log | \\mathcal { B } | . } \\end{array}\n$$",
|
| 964 |
+
"text_format": "latex",
|
| 965 |
+
"bbox": [
|
| 966 |
+
253,
|
| 967 |
+
196,
|
| 968 |
+
741,
|
| 969 |
+
217
|
| 970 |
+
],
|
| 971 |
+
"page_idx": 6
|
| 972 |
+
},
|
| 973 |
+
{
|
| 974 |
+
"type": "text",
|
| 975 |
+
"text": "Remark 3. Similar results continue to hold even when the two players use different regularization parameters $\\tau _ { \\mu } , \\tau _ { \\nu } > 0$ in (2), as long as the regularization parameter $\\tau$ is replaced by max $\\{ \\tau _ { \\mu } , \\tau _ { \\nu } \\}$ in the upper bounds of the learning rate, and the contraction parameter is replaced by $1 - \\operatorname* { m i n } \\{ \\tau _ { \\mu } , \\tau _ { \\nu } \\} \\eta$ . ",
|
| 976 |
+
"bbox": [
|
| 977 |
+
173,
|
| 978 |
+
219,
|
| 979 |
+
825,
|
| 980 |
+
263
|
| 981 |
+
],
|
| 982 |
+
"page_idx": 6
|
| 983 |
+
},
|
| 984 |
+
{
|
| 985 |
+
"type": "text",
|
| 986 |
+
"text": "Theorem 1 characterizes the convergence of the last-iterates $\\zeta ^ { ( t ) }$ and $\\bar { \\zeta } ^ { ( t ) }$ of PU and OMWU as long as the learning rate lies within the specified ranges. While PU doubles the number of calls to the sampling oracle, it also allows roughly as large as twice the learning rate compared with OMWU (cf. (6)). Compared with the vast literature analyzing the average-iterate performance of variants of extragradient methods, our results contribute towards characterizing the last-iterate convergence of multiplicative update methods in the presence of entropy regularization and simplex constraints, which to the best of our knowledge, are the first of its kind. Several remarks are in order. ",
|
| 987 |
+
"bbox": [
|
| 988 |
+
173,
|
| 989 |
+
272,
|
| 990 |
+
825,
|
| 991 |
+
372
|
| 992 |
+
],
|
| 993 |
+
"page_idx": 6
|
| 994 |
+
},
|
| 995 |
+
{
|
| 996 |
+
"type": "text",
|
| 997 |
+
"text": "Linear convergence to QRE. To achieve an $\\epsilon$ -accurate estimate of the QRE in terms of the KL divergence, the bound (7a) tells that it is sufficient to take ",
|
| 998 |
+
"bbox": [
|
| 999 |
+
173,
|
| 1000 |
+
377,
|
| 1001 |
+
823,
|
| 1002 |
+
406
|
| 1003 |
+
],
|
| 1004 |
+
"page_idx": 6
|
| 1005 |
+
},
|
| 1006 |
+
{
|
| 1007 |
+
"type": "equation",
|
| 1008 |
+
"img_path": "images/be090bb26a96971777947c1ec6bf114494d258ce86d67ce09fbd7459d4b4d822.jpg",
|
| 1009 |
+
"text": "$$\n\\frac { 1 } { \\eta \\tau } \\log \\left( \\frac { \\log | \\cal { A } | + \\log | \\cal { B } | } { \\epsilon } \\right)\n$$",
|
| 1010 |
+
"text_format": "latex",
|
| 1011 |
+
"bbox": [
|
| 1012 |
+
403,
|
| 1013 |
+
410,
|
| 1014 |
+
593,
|
| 1015 |
+
445
|
| 1016 |
+
],
|
| 1017 |
+
"page_idx": 6
|
| 1018 |
+
},
|
| 1019 |
+
{
|
| 1020 |
+
"type": "text",
|
| 1021 |
+
"text": "iterations using either PU or OMWU. Notably, this iteration complexity does not depend on any hidden constants and only depends double logarithmically on the cardinality of action spaces, which is almost dimension-free. Maximizing the learning rate, the iteration complexity is bounded by $( 1 + \\| A \\| _ { \\infty } / \\tau ) \\log ( 1 / \\epsilon )$ (modulo log factors), which only depends on the ratio $\\| A \\| _ { \\infty } / \\tau$ . ",
|
| 1022 |
+
"bbox": [
|
| 1023 |
+
173,
|
| 1024 |
+
448,
|
| 1025 |
+
825,
|
| 1026 |
+
505
|
| 1027 |
+
],
|
| 1028 |
+
"page_idx": 6
|
| 1029 |
+
},
|
| 1030 |
+
{
|
| 1031 |
+
"type": "text",
|
| 1032 |
+
"text": "Entrywise error of the policy log-ratios. Both PU and OMWU enjoy strong entrywise guarantees in the sense we can guarantee the convergence of the $\\ell _ { \\infty }$ norm of the log-ratios between the learned policy pair and the QRE at the same dimension-free linear rate (cf. (7b)), which suggests the policy pair converges in a somewhat uniform manner across the entire action space. ",
|
| 1033 |
+
"bbox": [
|
| 1034 |
+
173,
|
| 1035 |
+
510,
|
| 1036 |
+
825,
|
| 1037 |
+
565
|
| 1038 |
+
],
|
| 1039 |
+
"page_idx": 6
|
| 1040 |
+
},
|
| 1041 |
+
{
|
| 1042 |
+
"type": "text",
|
| 1043 |
+
"text": "Linear convergence of optimality and duality gaps. Our theorem also establishes the last-iterate convergence of the game values in terms of the optimality gap (cf. (7c)) and the duality gap (cf. (7d)) for both PU and OMWU. In particular, as will be seen, bounding the optimality gap of matrix games turns out to be the key enabler for generalizing our algorithms to Markov games, and bounding the duality gap allows to directly translate our results to finding a NE of unregularized matrix games. ",
|
| 1044 |
+
"bbox": [
|
| 1045 |
+
173,
|
| 1046 |
+
571,
|
| 1047 |
+
825,
|
| 1048 |
+
642
|
| 1049 |
+
],
|
| 1050 |
+
"page_idx": 6
|
| 1051 |
+
},
|
| 1052 |
+
{
|
| 1053 |
+
"type": "text",
|
| 1054 |
+
"text": "Last-iterate convergence to approximate NE. The entropy-regularized matrix game can be thought as a smooth surrogate of the unregularized matrix game (1); in particular, it is possible to find an $\\epsilon$ -NE by setting $\\tau$ sufficiently small in (2). According to (Zhang et al., 2020, Definition 2.1), a policy pair $\\bar { \\zeta } = ( \\mu , \\bar { \\nu } )$ is an $\\epsilon$ -NE if it satisfies $\\begin{array} { r } { \\mathsf { D u a l G a p } ( \\zeta ) : = \\operatorname* { m a x } _ { \\mu ^ { \\prime } \\in \\Delta ( A ) } f ( \\mu ^ { \\prime } , \\nu ) - \\operatorname* { m i n } _ { \\nu ^ { \\prime } \\in \\Delta ( B ) } f ( \\mu , \\nu ^ { \\prime } ) \\leq \\epsilon } \\end{array}$ . ",
|
| 1055 |
+
"bbox": [
|
| 1056 |
+
173,
|
| 1057 |
+
647,
|
| 1058 |
+
826,
|
| 1059 |
+
705
|
| 1060 |
+
],
|
| 1061 |
+
"page_idx": 6
|
| 1062 |
+
},
|
| 1063 |
+
{
|
| 1064 |
+
"type": "text",
|
| 1065 |
+
"text": "Observe that setting $\\begin{array} { r } { \\tau = \\frac { \\epsilon / 4 } { \\log | \\mathcal { A } | + \\log | \\mathcal { B } | } } \\end{array}$ guarantees that $| f _ { \\tau } ( \\mu , \\nu ) - f ( \\mu , \\nu ) | < \\epsilon / 4$ uniformly over $( \\mu , \\nu ) \\in \\Delta ( \\mathcal { A } ) \\times \\Delta ( \\mathcal { B } )$ in view of the boundedness of the Shannon entropy $\\mathcal { H } ( \\cdot )$ . Theorem 7 (cf. (7d)) also ensures that our proposed algorithms find an approximate QRE $\\bar { \\zeta } ^ { ( T ) }$ such that $\\mathsf { D u a l G a p } _ { \\tau } \\bigl ( \\bar { \\zeta } ^ { ( T ) } \\bigr ) \\leq$ $\\epsilon / 2$ after taking $\\begin{array} { r } { T = \\widetilde { O } \\left( \\frac { 1 } { \\eta \\epsilon } \\right) } \\end{array}$ iterations, which is no more than $\\begin{array} { r } { \\widetilde { O } \\left( 1 + \\frac { \\| A \\| _ { \\infty } } { \\epsilon } \\right) } \\end{array}$ iterations with optimized learning rates. It follows immediately that ",
|
| 1066 |
+
"bbox": [
|
| 1067 |
+
173,
|
| 1068 |
+
712,
|
| 1069 |
+
826,
|
| 1070 |
+
799
|
| 1071 |
+
],
|
| 1072 |
+
"page_idx": 6
|
| 1073 |
+
},
|
| 1074 |
+
{
|
| 1075 |
+
"type": "text",
|
| 1076 |
+
"text": "[ $\\mathtt { \\mathtt { M a l G a p } } ( \\bar { \\zeta } ^ { ( T ) } ) \\le \\mathtt { D u a l G a p } _ { \\tau } ( \\bar { \\zeta } ^ { ( T ) } ) + \\operatorname* { m a x } _ { \\mu ^ { \\prime } , \\nu ^ { \\prime } } \\Big | f _ { \\tau } ( \\mu ^ { \\prime } , \\bar { \\nu } ^ { ( T ) } ) - f _ { \\tau } \\big ( \\bar { \\mu } ^ { ( T ) } , \\nu ^ { \\prime } \\big ) - \\big ( f ( \\mu ^ { \\prime } , \\bar { \\nu } ^ { ( T ) } ) - f ( \\bar { \\mu } ^ { ( T ) } , \\nu ^ { \\prime } ) \\big ) \\Big | \\le \\epsilon ,$ and therefore $\\bar { \\zeta } ^ { ( T ) }$ is an $\\mathrm { \\epsilon - N E }$ . Intriguingly, unlike prior work (Daskalakis and Panageas, 2018a; Wei et al., 2021b) that analyzed the last-iterate convergence of OMWU in the unregularized setting $\\mathit { \\Pi } _ { \\mathcal { T } } = 0 \\mathit { \\Pi } _ { \\mathcal { c } }$ ), our last-iterate convergence does not require the NE of (1) to be unique. ",
|
| 1077 |
+
"bbox": [
|
| 1078 |
+
173,
|
| 1079 |
+
801,
|
| 1080 |
+
831,
|
| 1081 |
+
877
|
| 1082 |
+
],
|
| 1083 |
+
"page_idx": 6
|
| 1084 |
+
},
|
| 1085 |
+
{
|
| 1086 |
+
"type": "text",
|
| 1087 |
+
"text": "Rationality. Another attractive feature of the algorithms developed above is being rational (as introduced in Bowling and Veloso (2001)) in the sense that the algorithm returns the best-response policy of one player when the opponent takes any fixed stationary policy. More specially, in terms of matrix games, when player 2 sticks to a stationary policy $\\nu$ , the update of player 1 reduces to $\\mu ^ { ( t + 1 ) } ( a ) \\bar { \\propto } \\mu ^ { ( t ) } ( a ) ^ { 1 - \\bar { \\eta tau } } \\exp ( \\eta [ A \\nu ] _ { a } )$ . In this case, Theorem 1 can be established in exactly the same fashion by restricting attention only to the updates of $\\mu ^ { ( t ) }$ . ",
|
| 1088 |
+
"bbox": [
|
| 1089 |
+
171,
|
| 1090 |
+
882,
|
| 1091 |
+
823,
|
| 1092 |
+
912
|
| 1093 |
+
],
|
| 1094 |
+
"page_idx": 6
|
| 1095 |
+
},
|
| 1096 |
+
{
|
| 1097 |
+
"type": "image",
|
| 1098 |
+
"img_path": "images/4e4a6167f1bf0473e8d7a714c8b874f38c7ea875b79a1295462ea5a711766939.jpg",
|
| 1099 |
+
"image_caption": [
|
| 1100 |
+
"Figure 1: Performance illustration of the PU and OMWU methods for solving entropy-regularized matrix games with $| \\mathcal { A } | = | \\mathcal { B } | = 1 0 0$ , where the entries of the payoff matrix $A$ is generated independently from the uniform distribution on $[ - 1 , 1 ]$ . The learning rates are fixed as $\\eta = 0 . 1$ . The left panel plots various error metrics of convergence w.r.t. the iteration count with $\\tau = 0 . 0 1$ , while the right panel plots these error metrics at 1000-th iteration with different choices of $\\tau$ . "
|
| 1101 |
+
],
|
| 1102 |
+
"image_footnote": [],
|
| 1103 |
+
"bbox": [
|
| 1104 |
+
205,
|
| 1105 |
+
95,
|
| 1106 |
+
792,
|
| 1107 |
+
272
|
| 1108 |
+
],
|
| 1109 |
+
"page_idx": 7
|
| 1110 |
+
},
|
| 1111 |
+
{
|
| 1112 |
+
"type": "text",
|
| 1113 |
+
"text": "",
|
| 1114 |
+
"bbox": [
|
| 1115 |
+
173,
|
| 1116 |
+
375,
|
| 1117 |
+
825,
|
| 1118 |
+
438
|
| 1119 |
+
],
|
| 1120 |
+
"page_idx": 7
|
| 1121 |
+
},
|
| 1122 |
+
{
|
| 1123 |
+
"type": "text",
|
| 1124 |
+
"text": "No-regret learning of OMWU. Besides convergence to equilibria, in game-theoretical settings, it is often desirable to design and implement no-regret algorithms, which are capable of providing black-box guarantees over arbitrary sequences played by the opponent (Cesa-Bianchi and Lugosi, 2006; Rakhlin and Sridharan, 2013). Fortunately, it turns out that entropy regularization not only accelerates the convergence, but also enables no-regret learning somewhat “for free”: it encourages exploration by putting a positive mass on every action, therefore guards against adversaries. By using a properly chosen learning rate schedule, the proposed OMWU (Algorithm 2) can be further established as a no-regret algorithm; the details can be found in (Cen et al., 2021). ",
|
| 1125 |
+
"bbox": [
|
| 1126 |
+
173,
|
| 1127 |
+
443,
|
| 1128 |
+
826,
|
| 1129 |
+
554
|
| 1130 |
+
],
|
| 1131 |
+
"page_idx": 7
|
| 1132 |
+
},
|
| 1133 |
+
{
|
| 1134 |
+
"type": "text",
|
| 1135 |
+
"text": "3 Zero-sum Markov games with entropy regularization ",
|
| 1136 |
+
"text_level": 1,
|
| 1137 |
+
"bbox": [
|
| 1138 |
+
176,
|
| 1139 |
+
569,
|
| 1140 |
+
650,
|
| 1141 |
+
585
|
| 1142 |
+
],
|
| 1143 |
+
"page_idx": 7
|
| 1144 |
+
},
|
| 1145 |
+
{
|
| 1146 |
+
"type": "text",
|
| 1147 |
+
"text": "Leveraging the success of PU and OMWU in solving the entropy-regularized matrix games, this section extends our current analysis to solve the zero-sum two-player Markov game with entropy regularization, which is again formulated as finding the equilibrium of a saddle-point optimization problem. We start by introducing its basic setup, which will be followed by the proposed policy extragradient method with its theoretical guarantees. ",
|
| 1148 |
+
"bbox": [
|
| 1149 |
+
174,
|
| 1150 |
+
594,
|
| 1151 |
+
825,
|
| 1152 |
+
665
|
| 1153 |
+
],
|
| 1154 |
+
"page_idx": 7
|
| 1155 |
+
},
|
| 1156 |
+
{
|
| 1157 |
+
"type": "text",
|
| 1158 |
+
"text": "3.1 Background and problem formulation ",
|
| 1159 |
+
"text_level": 1,
|
| 1160 |
+
"bbox": [
|
| 1161 |
+
176,
|
| 1162 |
+
676,
|
| 1163 |
+
478,
|
| 1164 |
+
691
|
| 1165 |
+
],
|
| 1166 |
+
"page_idx": 7
|
| 1167 |
+
},
|
| 1168 |
+
{
|
| 1169 |
+
"type": "text",
|
| 1170 |
+
"text": "We consider a discounted Markov Game (MG) which is defined as $\\mathcal { M } = \\{ { S , A , B , P , r , \\gamma } \\}$ , with discrete state space $s$ , action spaces of two players $\\mathcal { A }$ and $\\boldsymbol { B }$ , transition probability $P$ , reward function $r : \\mathcal { S } \\times \\mathcal { A } \\times \\mathcal { B } [ 0 , 1 ]$ and discount factor $\\gamma \\in [ 0 , 1 )$ . A policy $\\mu : { \\mathcal { S } } \\Delta ( { \\mathcal { A } } )$ (resp. $\\nu : S \\to \\Delta ( B ) )$ defines how player 1 (resp. player 2) reacts to a given state $s$ , where the probability of taking action $a \\in { \\mathcal { A } }$ (resp. $b \\in B ,$ ) is $\\mu ( a | s )$ (resp. $\\nu ( b | s ) )$ . The transition probability kernel $P : \\mathcal { S } \\times \\mathcal { A } \\times \\mathcal { B } \\Delta ( \\mathcal { S } )$ defines the dynamics of the Markov game, where $P ( s ^ { \\prime } | s , a , b )$ specifies the probability of transiting to state $s ^ { \\prime }$ from state $s$ when the players take actions $a$ and $b$ respectively. ",
|
| 1171 |
+
"bbox": [
|
| 1172 |
+
173,
|
| 1173 |
+
696,
|
| 1174 |
+
825,
|
| 1175 |
+
795
|
| 1176 |
+
],
|
| 1177 |
+
"page_idx": 7
|
| 1178 |
+
},
|
| 1179 |
+
{
|
| 1180 |
+
"type": "text",
|
| 1181 |
+
"text": "Motivated by entropy regularization in Markov decision processes (MDP) (Geist et al., 2019), we consider an entropy-regularized variant of MG, where the value function is defined as ",
|
| 1182 |
+
"bbox": [
|
| 1183 |
+
173,
|
| 1184 |
+
800,
|
| 1185 |
+
825,
|
| 1186 |
+
829
|
| 1187 |
+
],
|
| 1188 |
+
"page_idx": 7
|
| 1189 |
+
},
|
| 1190 |
+
{
|
| 1191 |
+
"type": "equation",
|
| 1192 |
+
"img_path": "images/cd4673555690bc9e7c2c134b940cbf6ce609f112d808752990bda7448b1a8d47.jpg",
|
| 1193 |
+
"text": "$$\nV _ { \\tau } ^ { \\mu , \\nu } ( s ) : = \\mathbb { E } \\left[ \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } \\left( r ( s _ { t } , a _ { t } , b _ { t } ) - \\tau \\log \\mu ( a _ { t } | s _ { t } ) + \\tau \\log \\nu ( b _ { t } | s _ { t } ) \\right) ~ \\middle | s _ { 0 } = s \\right] ,\n$$",
|
| 1194 |
+
"text_format": "latex",
|
| 1195 |
+
"bbox": [
|
| 1196 |
+
233,
|
| 1197 |
+
834,
|
| 1198 |
+
763,
|
| 1199 |
+
878
|
| 1200 |
+
],
|
| 1201 |
+
"page_idx": 7
|
| 1202 |
+
},
|
| 1203 |
+
{
|
| 1204 |
+
"type": "text",
|
| 1205 |
+
"text": "where the quantity $\\tau \\geq 0$ denotes the regularization parameter, and the expectation is evaluated over the randomness of the transition kernel as well as the policies. The regularized Q-function $Q _ { \\tau } ^ { \\mu , \\nu }$ of a ",
|
| 1206 |
+
"bbox": [
|
| 1207 |
+
173,
|
| 1208 |
+
882,
|
| 1209 |
+
823,
|
| 1210 |
+
911
|
| 1211 |
+
],
|
| 1212 |
+
"page_idx": 7
|
| 1213 |
+
},
|
| 1214 |
+
{
|
| 1215 |
+
"type": "text",
|
| 1216 |
+
"text": "policy pair $( \\mu , \\nu )$ is related to $V _ { \\tau } ^ { \\mu , \\nu }$ as ",
|
| 1217 |
+
"bbox": [
|
| 1218 |
+
174,
|
| 1219 |
+
90,
|
| 1220 |
+
424,
|
| 1221 |
+
106
|
| 1222 |
+
],
|
| 1223 |
+
"page_idx": 8
|
| 1224 |
+
},
|
| 1225 |
+
{
|
| 1226 |
+
"type": "equation",
|
| 1227 |
+
"img_path": "images/f3de3e826aa012c6e58901a8e3b22025d2112634453e74fd01df633f9fe9af64.jpg",
|
| 1228 |
+
"text": "$$\n\\begin{array} { r } { Q _ { \\tau } ^ { \\mu , \\nu } ( s , a , b ) = r ( s , a , b ) + \\gamma \\mathbb { E } _ { s ^ { \\prime } \\sim P ( \\cdot \\mid s , a , b ) } \\bigl [ V _ { \\tau } ^ { \\mu , \\nu } \\bigl ( s ^ { \\prime } \\bigr ) \\bigr ] . } \\end{array}\n$$",
|
| 1229 |
+
"text_format": "latex",
|
| 1230 |
+
"bbox": [
|
| 1231 |
+
320,
|
| 1232 |
+
113,
|
| 1233 |
+
678,
|
| 1234 |
+
132
|
| 1235 |
+
],
|
| 1236 |
+
"page_idx": 8
|
| 1237 |
+
},
|
| 1238 |
+
{
|
| 1239 |
+
"type": "text",
|
| 1240 |
+
"text": "We will call $V _ { \\tau } ^ { \\mu , \\nu }$ and $Q _ { \\tau } ^ { \\mu , \\nu }$ the soft value function and soft $Q$ -function, respectively. A policy pair $( \\mu _ { \\tau } ^ { \\star } , \\nu _ { \\tau } ^ { \\star } )$ is said to be the quantal response equilibrium (QRE) of the entropy-regularized MG, if its value attains the minimax value of the entropy-regularized MG over all states $s \\in S$ , i.e. ",
|
| 1241 |
+
"bbox": [
|
| 1242 |
+
173,
|
| 1243 |
+
137,
|
| 1244 |
+
825,
|
| 1245 |
+
180
|
| 1246 |
+
],
|
| 1247 |
+
"page_idx": 8
|
| 1248 |
+
},
|
| 1249 |
+
{
|
| 1250 |
+
"type": "equation",
|
| 1251 |
+
"img_path": "images/d2ba603658bc41dc17088723550a675658f813d3ce0cc45bf0c47f8124f2dac5.jpg",
|
| 1252 |
+
"text": "$$\nV _ { \\tau } ^ { \\star } ( s ) = \\operatorname* { m a x } _ { \\mu } \\operatorname* { m i n } _ { \\nu } V _ { \\tau } ^ { \\mu , \\nu } ( s ) = \\operatorname* { m i n } _ { \\nu } \\operatorname* { m a x } _ { \\mu } V _ { \\tau } ^ { \\mu , \\nu } ( s ) : = V _ { \\tau } ^ { \\mu _ { \\tau } ^ { \\star } , \\nu _ { \\tau } ^ { \\star } } ( s ) ,\n$$",
|
| 1253 |
+
"text_format": "latex",
|
| 1254 |
+
"bbox": [
|
| 1255 |
+
289,
|
| 1256 |
+
185,
|
| 1257 |
+
707,
|
| 1258 |
+
213
|
| 1259 |
+
],
|
| 1260 |
+
"page_idx": 8
|
| 1261 |
+
},
|
| 1262 |
+
{
|
| 1263 |
+
"type": "text",
|
| 1264 |
+
"text": "where ${ \\cal V } _ { \\tau } ^ { \\star }$ is called the optimal minimax soft value function, and similarly $Q _ { \\tau } ^ { \\star } : = Q _ { \\tau } ^ { \\mu _ { \\tau } ^ { \\star } , \\nu _ { \\tau } ^ { \\star } }$ is called the optimal minimax soft Q-function. ",
|
| 1265 |
+
"bbox": [
|
| 1266 |
+
173,
|
| 1267 |
+
224,
|
| 1268 |
+
823,
|
| 1269 |
+
253
|
| 1270 |
+
],
|
| 1271 |
+
"page_idx": 8
|
| 1272 |
+
},
|
| 1273 |
+
{
|
| 1274 |
+
"type": "text",
|
| 1275 |
+
"text": "Goal. Our goal is to find the QRE of the entropy-regularized MG in a decentralized manner where the players only observe its own reward without accessing the opponent’s actions. By setting the regularization parameter sufficiently small $\\tau$ , this also allows us to find an approximate NE of the unregularized MG. ",
|
| 1276 |
+
"bbox": [
|
| 1277 |
+
173,
|
| 1278 |
+
260,
|
| 1279 |
+
825,
|
| 1280 |
+
315
|
| 1281 |
+
],
|
| 1282 |
+
"page_idx": 8
|
| 1283 |
+
},
|
| 1284 |
+
{
|
| 1285 |
+
"type": "text",
|
| 1286 |
+
"text": "3.2 From value iteration to policy extragradient methods ",
|
| 1287 |
+
"text_level": 1,
|
| 1288 |
+
"bbox": [
|
| 1289 |
+
173,
|
| 1290 |
+
327,
|
| 1291 |
+
581,
|
| 1292 |
+
342
|
| 1293 |
+
],
|
| 1294 |
+
"page_idx": 8
|
| 1295 |
+
},
|
| 1296 |
+
{
|
| 1297 |
+
"type": "text",
|
| 1298 |
+
"text": "Entropy-regularized value iteration. It is known that classical dynamic programming approaches such as value iteration can be extended to solve MG (Perolat et al., 2015), where each iteration amounts to solving a series of matrix games for each state. Similar to the single-agent case (Cen et al., 2020), we can extend these approaches to solve the entropy-regularized MG. Setting the stage, let us introduce the per-state $\\mathbf { Q }$ -value matrix $Q ( s ) : = Q ( s , \\cdot , \\cdot ) \\bar { \\in \\mathbb { R } ^ { | \\mathcal { A } | \\times | \\mathcal { B } | } }$ for every $s \\in S$ , where the element indexed by the action pair $( a , b )$ is $Q ( s , a , b )$ . Similarly, we define the per-state policies $\\mu ( s ) : = \\mu ( \\cdot | s ) \\in \\Delta ( { \\dot { A } } )$ and $\\nu ( s ) : = \\nu ( \\cdot | s ) \\in \\Delta ( B )$ for both players. ",
|
| 1299 |
+
"bbox": [
|
| 1300 |
+
173,
|
| 1301 |
+
347,
|
| 1302 |
+
826,
|
| 1303 |
+
448
|
| 1304 |
+
],
|
| 1305 |
+
"page_idx": 8
|
| 1306 |
+
},
|
| 1307 |
+
{
|
| 1308 |
+
"type": "text",
|
| 1309 |
+
"text": "In parallel to the original Bellman operator, we denote the soft Bellman operator $\\mathcal { T } _ { \\tau }$ as ",
|
| 1310 |
+
"bbox": [
|
| 1311 |
+
176,
|
| 1312 |
+
452,
|
| 1313 |
+
740,
|
| 1314 |
+
468
|
| 1315 |
+
],
|
| 1316 |
+
"page_idx": 8
|
| 1317 |
+
},
|
| 1318 |
+
{
|
| 1319 |
+
"type": "equation",
|
| 1320 |
+
"img_path": "images/2059d579c93d1602d1e72439a83460122b2d3c51c9a3214e4c20249b2456a9c9.jpg",
|
| 1321 |
+
"text": "$$\n\\mathcal { T } _ { \\tau } ( Q ) ( s , a , b ) : = r ( s , a , b ) + \\gamma \\mathbb { E } _ { s ^ { \\prime } \\sim P ( \\cdot \\vert s , a , b ) } \\left[ \\operatorname* { m a x } _ { \\mu ( s ^ { \\prime } ) \\in \\Delta ( A ) } \\operatorname* { m i n } _ { \\nu ( s ^ { \\prime } ) \\in \\Delta ( B ) } f _ { \\tau } \\left( Q ( s ^ { \\prime } ) ; \\mu ( s ^ { \\prime } ) , \\nu ( s ^ { \\prime } ) \\right) \\right] ,\n$$",
|
| 1322 |
+
"text_format": "latex",
|
| 1323 |
+
"bbox": [
|
| 1324 |
+
189,
|
| 1325 |
+
473,
|
| 1326 |
+
805,
|
| 1327 |
+
508
|
| 1328 |
+
],
|
| 1329 |
+
"page_idx": 8
|
| 1330 |
+
},
|
| 1331 |
+
{
|
| 1332 |
+
"type": "text",
|
| 1333 |
+
"text": "where for each per-state Q-value matrix $Q ( s )$ , we introduce an entropy-regularized matrix game in the form of ",
|
| 1334 |
+
"bbox": [
|
| 1335 |
+
173,
|
| 1336 |
+
515,
|
| 1337 |
+
823,
|
| 1338 |
+
545
|
| 1339 |
+
],
|
| 1340 |
+
"page_idx": 8
|
| 1341 |
+
},
|
| 1342 |
+
{
|
| 1343 |
+
"type": "equation",
|
| 1344 |
+
"img_path": "images/be159e0ed6417306b53ac4f96ca81f6ad60044ec1cf69fdcd8c0e3698c23ba8a.jpg",
|
| 1345 |
+
"text": "$$\n\\displaystyle \\operatorname* { m a x } _ { \\mu \\in \\Delta ( A ) } \\operatorname* { m i n } _ { \\nu \\in \\Delta ( B ) } f _ { \\tau } \\big ( Q ( s ) ; \\mu ( s ) , \\nu ( s ) \\big ) : = \\mu ( s ) ^ { \\top } Q ( s ) \\nu ( s ) - \\tau \\mathcal { H } ( \\mu ( s ) ) + \\tau \\mathcal { H } ( \\nu ( s ) ) .\n$$",
|
| 1346 |
+
"text_format": "latex",
|
| 1347 |
+
"bbox": [
|
| 1348 |
+
222,
|
| 1349 |
+
547,
|
| 1350 |
+
774,
|
| 1351 |
+
577
|
| 1352 |
+
],
|
| 1353 |
+
"page_idx": 8
|
| 1354 |
+
},
|
| 1355 |
+
{
|
| 1356 |
+
"type": "text",
|
| 1357 |
+
"text": "The entropy-regularized value iteration then proceeds as ",
|
| 1358 |
+
"bbox": [
|
| 1359 |
+
173,
|
| 1360 |
+
582,
|
| 1361 |
+
544,
|
| 1362 |
+
597
|
| 1363 |
+
],
|
| 1364 |
+
"page_idx": 8
|
| 1365 |
+
},
|
| 1366 |
+
{
|
| 1367 |
+
"type": "equation",
|
| 1368 |
+
"img_path": "images/0dfa08fcfd9718653016b28f0cfebe97143b66b36a3bdfdbe1308aa1e3e265a4.jpg",
|
| 1369 |
+
"text": "$$\nQ ^ { ( t + 1 ) } = T _ { \\tau } ( Q ^ { ( t ) } ) ,\n$$",
|
| 1370 |
+
"text_format": "latex",
|
| 1371 |
+
"bbox": [
|
| 1372 |
+
431,
|
| 1373 |
+
603,
|
| 1374 |
+
565,
|
| 1375 |
+
623
|
| 1376 |
+
],
|
| 1377 |
+
"page_idx": 8
|
| 1378 |
+
},
|
| 1379 |
+
{
|
| 1380 |
+
"type": "text",
|
| 1381 |
+
"text": "where $Q ^ { ( 0 ) }$ is an initialization. By definition, the optimal minimax soft Q-function obeys ${ \\cal T } _ { \\tau } ( Q _ { \\tau } ^ { \\star } ) =$ $Q _ { \\tau } ^ { \\star }$ and therefore corresponds to the fix point of the soft Bellman operator. Given the above entropyregularized value iteration, the following lemma states its iterates contract linearly to the optimal minimax soft Q-function at a rate of the discount factor $\\gamma$ . ",
|
| 1382 |
+
"bbox": [
|
| 1383 |
+
173,
|
| 1384 |
+
631,
|
| 1385 |
+
825,
|
| 1386 |
+
688
|
| 1387 |
+
],
|
| 1388 |
+
"page_idx": 8
|
| 1389 |
+
},
|
| 1390 |
+
{
|
| 1391 |
+
"type": "text",
|
| 1392 |
+
"text": "Proposition 2. The entropy-regularized value iteration (10) converges at a linear rate, i.e. $\\parallel Q ^ { ( t ) } -$ \n$Q _ { \\tau } ^ { \\star } \\| _ { \\infty } \\leq \\gamma ^ { t } \\| Q ^ { ( 0 ) } - Q _ { \\tau } ^ { \\star } \\| _ { \\infty }$ . ",
|
| 1393 |
+
"bbox": [
|
| 1394 |
+
174,
|
| 1395 |
+
691,
|
| 1396 |
+
823,
|
| 1397 |
+
723
|
| 1398 |
+
],
|
| 1399 |
+
"page_idx": 8
|
| 1400 |
+
},
|
| 1401 |
+
{
|
| 1402 |
+
"type": "text",
|
| 1403 |
+
"text": "Approximate value iteration via policy extragradient methods. Proposition 2 suggests that the optimal minimax soft Q-function of the entropy-regularized MG can be found by solving a series of entropy-regularized matrix games induced by $\\bar { \\{ Q ^ { ( t ) } \\} } _ { t \\geq 0 }$ in (10), a task that can be accomplished by adopting the fast extragradient methods developed earlier. To proceed, we first define the following sampling oracle, which makes it rigorous that the proposed algorithm does not require access to the Q-function of the entire MG, but only its own single-agent Q-function when playing against the opponent’s policy. ",
|
| 1404 |
+
"bbox": [
|
| 1405 |
+
173,
|
| 1406 |
+
733,
|
| 1407 |
+
826,
|
| 1408 |
+
833
|
| 1409 |
+
],
|
| 1410 |
+
"page_idx": 8
|
| 1411 |
+
},
|
| 1412 |
+
{
|
| 1413 |
+
"type": "text",
|
| 1414 |
+
"text": "Definition 2 (Sampling oracle for Markov games). Given any policy pair $\\mu ( s ) , \\nu ( s )$ and $Q$ -value matrix $Q ( s )$ for any $s \\in S$ , the sampling oracle returns ",
|
| 1415 |
+
"bbox": [
|
| 1416 |
+
173,
|
| 1417 |
+
835,
|
| 1418 |
+
823,
|
| 1419 |
+
864
|
| 1420 |
+
],
|
| 1421 |
+
"page_idx": 8
|
| 1422 |
+
},
|
| 1423 |
+
{
|
| 1424 |
+
"type": "equation",
|
| 1425 |
+
"img_path": "images/82747ffa22810c00af6604f8c9488a585bb5fc2f6fc901ae3a034f3572b482f0.jpg",
|
| 1426 |
+
"text": "$$\n[ Q ( s ) \\nu ( s ) ] _ { a } = \\mathbb { E } _ { b \\sim \\nu ( s ) } \\left[ Q ( s , a , b ) \\right] , \\qquad a n d \\qquad [ Q ( s ) ^ { \\top } \\mu ( s ) ] _ { b } = \\mathbb { E } _ { a \\sim \\mu ( s ) } \\left[ Q ( s , a , b ) \\right]\n$$",
|
| 1427 |
+
"text_format": "latex",
|
| 1428 |
+
"bbox": [
|
| 1429 |
+
212,
|
| 1430 |
+
871,
|
| 1431 |
+
785,
|
| 1432 |
+
892
|
| 1433 |
+
],
|
| 1434 |
+
"page_idx": 8
|
| 1435 |
+
},
|
| 1436 |
+
{
|
| 1437 |
+
"type": "text",
|
| 1438 |
+
"text": "for any $a \\in { \\mathcal { A } }$ and $b \\in B$ . ",
|
| 1439 |
+
"bbox": [
|
| 1440 |
+
171,
|
| 1441 |
+
896,
|
| 1442 |
+
343,
|
| 1443 |
+
911
|
| 1444 |
+
],
|
| 1445 |
+
"page_idx": 8
|
| 1446 |
+
},
|
| 1447 |
+
{
|
| 1448 |
+
"type": "text",
|
| 1449 |
+
"text": "Algorithm 3: Policy Extragradient Method for Entropy-regularized Markov Game ",
|
| 1450 |
+
"text_level": 1,
|
| 1451 |
+
"bbox": [
|
| 1452 |
+
173,
|
| 1453 |
+
94,
|
| 1454 |
+
714,
|
| 1455 |
+
109
|
| 1456 |
+
],
|
| 1457 |
+
"page_idx": 9
|
| 1458 |
+
},
|
| 1459 |
+
{
|
| 1460 |
+
"type": "text",
|
| 1461 |
+
"text": "1 initialization: $Q ^ { ( 0 ) } = 0$ . \n2 for $t = 0 , 1 , 2 , \\cdots , T _ { \\mathrm { m a i n } } \\mathrm { { \\bf d o } }$ \n3 Let $Q ^ { ( t ) }$ denote $\\begin{array} { r } { Q ^ { ( t ) } ( s , a , b ) = r ( s , a , b ) + \\gamma \\mathbb { E } _ { s ^ { \\prime } \\sim P ( \\cdot | s , a , b ) } V ^ { ( t ) } ( s ^ { \\prime } ) . } \\end{array}$ (11) \n4 Invoke PU (Algorithm 1) or OMWU (Algorithm 2) for $T _ { \\mathrm { s u b } }$ iterations to solve the following entropy-regularized matrix game for every state $s$ , where the initialization is set as uniform distributions: $\\operatorname* { m a x } _ { \\mu ( s ) \\in \\Delta ( A ) } \\operatorname* { m i n } _ { \\nu ( s ) \\in \\Delta ( B ) } f _ { \\tau } \\big ( Q ^ { ( t ) } ( s ) ; \\mu ( s ) , \\nu ( s ) \\big ) .$ Return the last iterate $\\bar { \\mu } ^ { ( t , T _ { \\mathrm { s u b } } ) } ( s ) , \\bar { \\nu } ^ { ( t , T _ { \\mathrm { s u b } } ) } ( s )$ . \n5 Set $V ^ { ( t + 1 ) } ( s ) = f _ { \\tau } \\left( Q ^ { ( t ) } ( s ) ; \\bar { \\mu } ^ { ( t , \\bar { T } _ { \\mathrm { s u b } } ) } ( s ) , \\bar { \\nu } ^ { ( t , \\bar { T } _ { \\mathrm { s u b } } ) } ( s ) \\right)$ . ",
|
| 1462 |
+
"bbox": [
|
| 1463 |
+
158,
|
| 1464 |
+
113,
|
| 1465 |
+
803,
|
| 1466 |
+
299
|
| 1467 |
+
],
|
| 1468 |
+
"page_idx": 9
|
| 1469 |
+
},
|
| 1470 |
+
{
|
| 1471 |
+
"type": "text",
|
| 1472 |
+
"text": "Encouragingly, by judiciously setting the number of iterations in both the outer loop (for updating the Q-value matrices) and the inner loop (for updating the QRE of the corresponding Q-value matrix), we are guaranteed to find the QRE of the entropy-regularized MG in a small number of iterations without solving the iteration-varying matrix games exactly, as dictated by the following theorem. ",
|
| 1473 |
+
"bbox": [
|
| 1474 |
+
173,
|
| 1475 |
+
337,
|
| 1476 |
+
826,
|
| 1477 |
+
393
|
| 1478 |
+
],
|
| 1479 |
+
"page_idx": 9
|
| 1480 |
+
},
|
| 1481 |
+
{
|
| 1482 |
+
"type": "text",
|
| 1483 |
+
"text": "Theorem 2. Assume $| { \\mathcal { A } } | \\geq | { \\mathcal { B } } |$ and $\\tau \\leq 1$ . Setting $\\begin{array} { r } { \\eta = \\frac { 1 - \\gamma } { 2 ( 1 + \\tau ( \\log | \\mathcal { A } | + 1 - \\gamma ) ) } } \\end{array}$ , the total iterations (namely, the product $T _ { \\mathrm { m a i n } } \\cdot T _ { \\mathrm { s u b , } }$ ) required for Algorithm 3 to achieve $\\left\\| Q ^ { ( { \\vec { T } _ { \\operatorname* { m i n } } } ) } - Q _ { \\tau } ^ { \\star } \\right\\| _ { \\infty } \\leq \\epsilon$ is at most $\\begin{array} { r } { O \\left( \\frac { ( \\log | \\cal { A } | + 1 / \\tau ) } { ( 1 - \\gamma ) ^ { 2 } } \\left( \\log \\frac { \\log | \\cal { A } | } { ( 1 - \\gamma ) \\epsilon } \\right) ^ { 2 } \\right) } \\end{array}$ . ",
|
| 1484 |
+
"bbox": [
|
| 1485 |
+
173,
|
| 1486 |
+
398,
|
| 1487 |
+
825,
|
| 1488 |
+
468
|
| 1489 |
+
],
|
| 1490 |
+
"page_idx": 9
|
| 1491 |
+
},
|
| 1492 |
+
{
|
| 1493 |
+
"type": "text",
|
| 1494 |
+
"text": "Theorem 2 ensures that within $\\begin{array} { r } { \\widetilde O \\left( \\frac { 1 } { \\tau ( 1 - \\gamma ) ^ { 2 } } \\log ^ { 2 } \\left( \\frac { 1 } { \\epsilon } \\right) \\right) } \\end{array}$ iterations, Algorithm 3 finds a pair of policies whose value is close to the optimal minimax soft Q-function $Q _ { \\tau } ^ { \\star }$ in an entrywise manner to a prescribed accuracy $\\epsilon$ . Remarkably, the iteration complexity is independent of the dimensions of the state space and the action space (up to log factors). ",
|
| 1495 |
+
"bbox": [
|
| 1496 |
+
174,
|
| 1497 |
+
481,
|
| 1498 |
+
825,
|
| 1499 |
+
546
|
| 1500 |
+
],
|
| 1501 |
+
"page_idx": 9
|
| 1502 |
+
},
|
| 1503 |
+
{
|
| 1504 |
+
"type": "text",
|
| 1505 |
+
"text": "Remark 4 (Duality gap and solving the unregularized MG). Solving the entropy-regularized MG provides a viable strategy to find an $\\epsilon$ -approximate NE of the unregularized MG, where the optimality of a policy pair is typically gauged by the duality gap. Fortunately, this can be achieved similar to the case of matrix games, and we refer interested readers to Cen et al. (2021) for details. ",
|
| 1506 |
+
"bbox": [
|
| 1507 |
+
174,
|
| 1508 |
+
551,
|
| 1509 |
+
825,
|
| 1510 |
+
608
|
| 1511 |
+
],
|
| 1512 |
+
"page_idx": 9
|
| 1513 |
+
},
|
| 1514 |
+
{
|
| 1515 |
+
"type": "text",
|
| 1516 |
+
"text": "4 Conclusions ",
|
| 1517 |
+
"text_level": 1,
|
| 1518 |
+
"bbox": [
|
| 1519 |
+
174,
|
| 1520 |
+
632,
|
| 1521 |
+
305,
|
| 1522 |
+
648
|
| 1523 |
+
],
|
| 1524 |
+
"page_idx": 9
|
| 1525 |
+
},
|
| 1526 |
+
{
|
| 1527 |
+
"type": "text",
|
| 1528 |
+
"text": "This paper develops provably efficient policy extragradient methods (PU and OMWU) for entropyregularized matrix games and Markov games, whose last iterates are guaranteed to converge linearly to the quantal response equilibrium at a linear rate. Encouragingly, the rate of convergence is independent of the dimension of the problem, i.e. the sizes of the space space and the action space. In addition, the last iterates of the proposed algorithms can also be used to locate Nash equilibria for the unregularized competitive games without assuming the uniqueness of the Nash equilibria by judiciously tuning the amount of regularization. This work opens up interesting opportunities for further investigations of policy extragradient methods for solving competitive games. For example, can we develop a two-time-scale policy extragradient algorithms for Markov games where the Qfunction is updated simultaneously with the policy but potentially at a different time scale, using samples, such as in an actor-critic algorithm (Konda and Tsitsiklis, 2000)? ",
|
| 1529 |
+
"bbox": [
|
| 1530 |
+
174,
|
| 1531 |
+
660,
|
| 1532 |
+
826,
|
| 1533 |
+
813
|
| 1534 |
+
],
|
| 1535 |
+
"page_idx": 9
|
| 1536 |
+
},
|
| 1537 |
+
{
|
| 1538 |
+
"type": "text",
|
| 1539 |
+
"text": "Acknowledgments and Disclosure of Funding ",
|
| 1540 |
+
"text_level": 1,
|
| 1541 |
+
"bbox": [
|
| 1542 |
+
174,
|
| 1543 |
+
835,
|
| 1544 |
+
553,
|
| 1545 |
+
853
|
| 1546 |
+
],
|
| 1547 |
+
"page_idx": 9
|
| 1548 |
+
},
|
| 1549 |
+
{
|
| 1550 |
+
"type": "text",
|
| 1551 |
+
"text": "S. Cen and Y. Chi are supported in part by the grants ONR N00014-18-1-2142 and N00014-19-1- 2404, ARO W911NF-18-1-0303, NSF CCF-1901199, CCF-2007911 and CCF-2106778. Y. Wei is supported in part by the NSF grants CCF-2007911, DMS-2147546/2015447 and CCF-2106778. ",
|
| 1552 |
+
"bbox": [
|
| 1553 |
+
176,
|
| 1554 |
+
869,
|
| 1555 |
+
826,
|
| 1556 |
+
911
|
| 1557 |
+
],
|
| 1558 |
+
"page_idx": 9
|
| 1559 |
+
},
|
| 1560 |
+
{
|
| 1561 |
+
"type": "text",
|
| 1562 |
+
"text": "References \nArora, S., Hazan, E., and Kale, S. (2012). The multiplicative weights update method: a meta-algorithm and applications. Theory of Computing, 8(1):121–164. \nBai, Y. and Jin, C. (2020). Provable self-play algorithms for competitive reinforcement learning. In International Conference on Machine Learning, pages 551–560. PMLR. \nBowling, M. and Veloso, M. (2001). Rational and convergent learning in stochastic games. In Proceedings of the 17th international joint conference on Artificial intelligence-Volume 2, pages 1021–1026. \nCen, S., Cheng, C., Chen, Y., Wei, Y., and Chi, Y. (2020). Fast global convergence of natural policy gradient methods with entropy regularization. arXiv preprint arXiv:2007.06558. \nCen, S., Wei, Y., and Chi, Y. (2021). Fast policy extragradient methods for competitive games with entropy regularization. arXiv preprint arXiv:2105.15186. \nCesa-Bianchi, N. and Lugosi, G. (2006). Prediction, learning, and games. Cambridge university press. \nDaskalakis, C., Deckelbaum, A., and Kim, A. (2011). Near-optimal no-regret algorithms for zero-sum games. In Proceedings of the twenty-second annual ACM-SIAM symposium on Discrete Algorithms, pages 235–254. SIAM. \nDaskalakis, C., Foster, D. J., and Golowich, N. (2020). Independent policy gradient methods for competitive reinforcement learning. In Advances in Neural Information Processing Systems, volume 33, pages 5527–5540. \nDaskalakis, C., Ilyas, A., Syrgkanis, V., and Zeng, H. (2018). Training GANs with optimism. In International Conference on Learning Representations (ICLR 2018). \nDaskalakis, C. and Panageas, I. (2018a). Last-iterate convergence: Zero-sum games and constrained min-max optimization. arXiv preprint arXiv:1807.04252. \nDaskalakis, C. and Panageas, I. (2018b). The limit points of (optimistic) gradient descent in minmax optimization. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 9256–9266. \nFreund, Y. and Schapire, R. E. (1999). Adaptive game playing using multiplicative weights. Games and Economic Behavior, 29(1-2):79–103. \nGeist, M., Scherrer, B., and Pietquin, O. (2019). A theory of regularized Markov decision processes. In International Conference on Machine Learning, pages 2160–2169. \nGoodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial networks. arXiv preprint arXiv:1406.2661. \nHarker, P. T. and Pang, J.-S. (1990). Finite-dimensional variational inequality and nonlinear complementarity problems: a survey of theory, algorithms and applications. Mathematical programming, 48(1):161–220. \nHofbauer, J. and Sandholm, W. H. (2002). On the global convergence of stochastic fictitious play. Econometrica, 70(6):2265–2294. \nHsieh, Y.-G., Iutzeler, F., Malick, J., and Mertikopoulos, P. (2019). On the convergence of single-call stochastic extra-gradient methods. arXiv preprint arXiv:1908.08465. \nKonda, V. R. and Tsitsiklis, J. N. (2000). Actor-critic algorithms. In Advances in neural information processing systems, pages 1008–1014. Citeseer. \nKorpelevich, G. M. (1976). The extragradient method for finding saddle points and other problems. Matecon, 12:747–756. \nLan, G. (2021). Policy mirror descent for reinforcement learning: Linear convergence, new sampling complexity, and generalized problem classes. arXiv preprint arXiv:2102.00135. \nLei, Q., Nagarajan, S. G., Panageas, I., and Wang, X. (2021). Last iterate convergence in no-regret learning: constrained min-max optimization for convex-concave landscapes. In International Conference on Artificial Intelligence and Statistics, pages 1441–1449. PMLR. \nLiang, T. and Stokes, J. (2019). Interaction matters: A note on non-asymptotic local convergence of generative adversarial networks. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 907–915. PMLR. \nLittman, M. L. (1994). Markov games as a framework for multi-agent reinforcement learning. In Machine Learning Proceedings, pages 157–163. Elsevier. \nMcKelvey, R. D. and Palfrey, T. R. (1995). Quantal response equilibria for normal form games. Games and economic behavior, 10(1):6–38. \nMei, J., Xiao, C., Szepesvari, C., and Schuurmans, D. (2020). On the global convergence rates of softmax policy gradient methods. arXiv preprint arXiv:2005.06392. \nMertikopoulos, P., Lecouat, B., Zenati, H., Foo, C.-S., Chandrasekhar, V., and Piliouras, G. (2018a). Optimistic mirror descent in saddle-point problems: Going the extra (gradient) mile. In International Conference on Learning Representations. \nMertikopoulos, P., Papadimitriou, C., and Piliouras, G. (2018b). Cycles in adversarial regularized learning. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 2703–2717. SIAM. \nMertikopoulos, P. and Sandholm, W. H. (2016). Learning in games via reinforcement and regularization. Mathematics of Operations Research, 41(4):1297–1324. \nMokhtari, A., Ozdaglar, A., and Pattathil, S. (2020a). A unified analysis of extra-gradient and optimistic gradient methods for saddle point problems: Proximal point approach. In International Conference on Artificial Intelligence and Statistics, pages 1497–1507. PMLR. \nMokhtari, A., Ozdaglar, A. E., and Pattathil, S. (2020b). Convergence rate of $O ( 1 / k )$ for optimistic gradient and extragradient methods in smooth convex-concave saddle point problems. SIAM Journal on Optimization, 30(4):3230–3251. \nNemirovski, A. (2004). Prox-method with rate of convergence $O ( 1 / t )$ for variational inequalities with Lipschitz continuous monotone operators and smooth convex-concave saddle point problems. SIAM Journal on Optimization, 15(1):229–251. \nNeu, G., Jonsson, A., and Gómez, V. (2017). A unified view of entropy-regularized Markov decision processes. arXiv preprint arXiv:1705.07798. \nPerolat, J., Scherrer, B., Piot, B., and Pietquin, O. (2015). Approximate dynamic programming for two-player zero-sum Markov games. In International Conference on Machine Learning, pages 1321–1329. PMLR. \nRakhlin, A. and Sridharan, K. (2013). Optimization, learning, and games with predictable sequences. arXiv preprint arXiv:1311.1869. \nSavas, Y., Ahmadi, M., Tanaka, T., and Topcu, U. (2019). Entropy-regularized stochastic games. In 2019 IEEE 58th Conference on Decision and Control (CDC), pages 5955–5962. IEEE. \nShapley, L. S. (1953). Stochastic games. Proceedings of the National Academy of Sciences, 39(10):1095–1100. \nSyrgkanis, V., Agarwal, A., Luo, H., and Schapire, R. E. (2015). Fast convergence of regularized learning in games. In Proceedings of the 28th International Conference on Neural Information Processing Systems-Volume 2, pages 2989–2997. \nTseng, P. (1995). On linear convergence of iterative methods for the variational inequality problem. Journal of Computational and Applied Mathematics, 60(1-2):237–252. \nWei, C.-Y., Lee, C.-W., Zhang, M., and Luo, H. (2021a). Last-iterate convergence of decentralized optimistic gradient descent/ascent in infinite-horizon competitive Markov games. arXiv preprint arXiv:2102.04540. \nWei, C.-Y., Lee, C.-W., Zhang, M., and Luo, H. (2021b). Linear last-iterate convergence in constrained saddle-point optimization. In International Conference on Learning Representations (ICLR). \nXie, Q., Chen, Y., Wang, Z., and Yang, Z. (2020). Learning zero-sum simultaneous-move Markov games using function approximation and correlated equilibrium. In Conference on Learning Theory, pages 3674–3682. PMLR. \nYadav, A., Shah, S., Xu, Z., Jacobs, D., and Goldstein, T. (2017). Stabilizing adversarial nets with prediction methods. arXiv preprint arXiv:1705.07364. \nZhan, W., Cen, S., Huang, B., Chen, Y., Lee, J. D., and Chi, Y. (2021). Policy mirror descent for regularized reinforcement learning: A generalized framework with linear convergence. arXiv preprint arXiv:2105.11066. \nZhang, K., Kakade, S., Basar, T., and Yang, L. (2020). Model-based multi-agent RL in zero-sum Markov games with near-optimal sample complexity. Advances in Neural Information Processing Systems, 33. \nZhao, Y., Tian, Y., Lee, J. D., and Du, S. S. (2021). Provably efficient policy gradient methods for two-player zero-sum Markov games. arXiv preprint arXiv:2102.08903. ",
|
| 1563 |
+
"bbox": [
|
| 1564 |
+
169,
|
| 1565 |
+
78,
|
| 1566 |
+
828,
|
| 1567 |
+
916
|
| 1568 |
+
],
|
| 1569 |
+
"page_idx": 10
|
| 1570 |
+
},
|
| 1571 |
+
{
|
| 1572 |
+
"type": "text",
|
| 1573 |
+
"text": "",
|
| 1574 |
+
"bbox": [
|
| 1575 |
+
169,
|
| 1576 |
+
56,
|
| 1577 |
+
828,
|
| 1578 |
+
919
|
| 1579 |
+
],
|
| 1580 |
+
"page_idx": 11
|
| 1581 |
+
},
|
| 1582 |
+
{
|
| 1583 |
+
"type": "text",
|
| 1584 |
+
"text": "",
|
| 1585 |
+
"bbox": [
|
| 1586 |
+
171,
|
| 1587 |
+
89,
|
| 1588 |
+
828,
|
| 1589 |
+
402
|
| 1590 |
+
],
|
| 1591 |
+
"page_idx": 12
|
| 1592 |
+
}
|
| 1593 |
+
]
|
parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/ZIyj0E58vzlo/ZIyj0E58vzlo_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/k-0oq5eNjh/k-0oq5eNjh.md
ADDED
|
@@ -0,0 +1,234 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# A Shading-Guided Generative Implicit Model for Shape-Accurate 3D-Aware Image Synthesis
|
| 2 |
+
|
| 3 |
+
Xingang Pan1 Xudong Xu2 Chen Change Loy3 Christian Theobalt1 Bo Dai3
|
| 4 |
+
|
| 5 |
+
1Max Planck Institute for Informatics 2The Chinese University of Hong Kong {xpan,theobalt}@mpi-inf.mpg.de xx018@ie.cuhk.edu.hk 3S-Lab, Nanyang Technological University {ccloy, bo.dai}@ntu.edu.sg
|
| 6 |
+
|
| 7 |
+
# Abstract
|
| 8 |
+
|
| 9 |
+
The advancement of generative radiance fields has pushed the boundary of 3Daware image synthesis. Motivated by the observation that a 3D object should look realistic from multiple viewpoints, these methods introduce a multi-view constraint as regularization to learn valid 3D radiance fields from 2D images. Despite the progress, they often fall short of capturing accurate 3D shapes due to the shapecolor ambiguity, limiting their applicability in downstream tasks. In this work, we address this ambiguity by proposing a novel shading-guided generative implicit model that is able to learn a starkly improved shape representation. Our key insight is that an accurate 3D shape should also yield a realistic rendering under different lighting conditions. This multi-lighting constraint is realized by modeling illumination explicitly and performing shading with various lighting conditions. Gradients are derived by feeding the synthesized images to a discriminator. To compensate for the additional computational burden of calculating surface normals, we further devise an efficient volume rendering strategy via surface tracking, reducing the training and inference time by $24 \%$ and $48 \%$ , respectively. Our experiments on multiple datasets show that the proposed approach achieves photorealistic 3D-aware image synthesis while capturing accurate underlying 3D shapes. We demonstrate improved performance of our approach on 3D shape reconstruction against existing methods, and show its applicability on image relighting. Our code will be released at https://github.com/XingangPan/ShadeGAN.
|
| 10 |
+
|
| 11 |
+
# 1 Introduction
|
| 12 |
+
|
| 13 |
+
Advanced deep generative models, e.g., StyleGAN [1, 2] and BigGAN [3], have achieved great successes in natural image synthesis. While producing impressive results, these 2D representationbased models cannot synthesize novel views of an instance in a 3D-consistent manner. They also fall short of representing an explicit 3D object shape. To overcome such limitations, researchers have proposed new deep generative models that represent 3D scenes as neural radiance fields [4, 5]. Such 3D-aware generative models allow explicit control of viewpoint while preserving 3D consistency during image synthesis. Perhaps a more fascinating merit is that they have shown the great potential of learning 3D shapes in an unsupervised manner from just a collection of unconstrained 2D images. If we could train a 3D-aware generative model that learns accurate 3D object shapes, it would broaden various downstream applications such as 3D shape reconstruction and image relighting.
|
| 14 |
+
|
| 15 |
+
Existing attempts for 3D-aware image synthesis [4, 5] tend to learn coarse 3D shapes that are inaccurate and noisy, as shown in Fig.1 (a). We found that such inaccuracy arises from an inevitable ambiguity inherent in the training strategy adopted by these methods. In particular, a form of regularization, which we refer to as "multi-view constraint", is used to enforce the 3D representation to look realistic from different viewpoints. The constraint is commonly implemented by first projecting the generator’s outputs (e.g., radiance fields [6]) to randomly sampled viewpoints, and then feeding them to a discriminator as fake images for training. While such a constraint enables these models to synthesize images in a 3D-aware manner, it suffers from the shape-color ambiguity, i.e., small variations of shape could lead to similar RGB images that look equally plausible to the discriminator, as the color of many objects is locally smooth. Consequently, inaccurate shapes are concealed under this constraint.
|
| 16 |
+
|
| 17 |
+

|
| 18 |
+
Figure 1: Motivation. (a) Previous methods like pi-GAN [4] resort to the "multi-view constraint", where the 3D representation is projected to different viewpoints as fake images to the discriminator. The extracted 3D meshes are often inaccurate due to the shape-color ambiguity. (b) The proposed approach ShadeGAN further adopts a "multi-lighting constraint", which motivates the 3D representation to look realistic under different lighting conditions. This constraint effectively addresses the ambiguity, giving rise to more natural and precise 3D shapes.
|
| 19 |
+
|
| 20 |
+
In this work, we propose a novel shading-guided generative implicit model (ShadeGAN) to address the aforementioned ambiguity. In particular, ShadeGAN learns more accurate 3D shapes by explicitly modeling shading, i.e., the interaction of illumination and shape. We believe that an accurate 3D shape should look realistic not only from different viewpoints, but also under different lighting conditions, i.e., satisfying the "multi-lighting constraint". This idea shares similar intuition with photometric stereo [7], which shows that accurate surface normal could be recovered from images taken under different lighting conditions. Note that the multi-lighting constraint is feasible as real-world images used for training are often taken under various lighting conditions. To fulfill this constraint, ShadeGAN takes a relightable color field as the intermediate representation, which approximates the albedo but does not necessarily satisfy viewpoint independence. The color field is shaded under a randomly sampled lighting condition during rendering. Since image appearance via such a shading process is strongly dependent on surface normals, inaccurate 3D shape representations will be much more clearly revealed than in earlier shading-agnostic generative models. Hence, by satisfying the multi-lighting constraint, ShadeGAN is encouraged to infer more accurate 3D shapes as shown in Fig.1 (b).
|
| 21 |
+
|
| 22 |
+
The above shading process requires the calculation of the normal direction via back-propagation through the generator, and such calculation needs to be repeated dozens of times for a pixel in volume rendering [4, 5], introducing additional computational overhead. Existing efficient volume rendering techniques [8, 9, 10, 11, 12] mainly target static scenes, and could not be directly applied to generative models due to their dynamic nature. Therefore, to improve the rendering speed of ShadeGAN, we formulate an efficient surface tracking network to estimate the rendered object surface conditioned on the latent code. This enables us to save rendering computations by just querying points near the predicted surface, leading to $24 \%$ and $48 \%$ reduction of training and inference time without affecting the quality of rendered images.
|
| 23 |
+
|
| 24 |
+
Comprehensive experiments are conducted across multiple datasets to verify the effectiveness of ShadeGAN. The results show that our approach is capable of synthesizing photorealistic images while capturing more accurate underlying 3D shapes than previous generative methods. The learned distribution of 3D shapes enables various downstream tasks like 3D shape reconstruction, where our approach significantly outperforms other baselines on the BFM dataset [13]. Besides, modeling the shading process enables explicit control over lighting conditions, achieving image relighting effect. Our contributions can be summarized as follows: 1) We address the shape-color ambiguity in existing 3D-aware image synthesis methods with a shading-guided generative model that satisfies the proposed multi-lighting constraint. In this way, ShadeGAN is able to learn more accurate 3D shapes for better image synthesis. 2) We devise an efficient rendering technique via surface tracking, which significantly saves training and inference time for volume rendering-based generative models. 3) We show that ShadeGAN learns to disentangle shading and color that well approximates the albedo, achieving natural relighting effects in image synthesis.
|
| 25 |
+
|
| 26 |
+
# 2 Related Work
|
| 27 |
+
|
| 28 |
+
Neural volume rendering. Starting from the seminal work of neural radiance fields (NeRF) [6], neural volume rendering has gained much popularity in representing 3D scenes and synthesizing novel views. By integrating coordinate-based neural networks with volume rendering, NeRF performs high-fidelity view synthesis in a 3D consistent manner. Several attempts have been proposed to extend or improve NeRF. For instance, [14, 15, 16] further model illumination, and learn to disentangle reflectance with shading given well-aligned multi-view and multi-lighting images. Besides, many studies accelerate the rendering of static scenes from the perspective of spatial sparsity [8, 9], architectural design [10, 11], or efficient rendering [17, 12]. However, it is not trivial to apply these illumination and acceleration techniques to volume rendering-based generative models [5, 4], as they typically learn from unposed and unpaired images, and represent dynamic scenes that change with respect to the input latent codes.
|
| 29 |
+
|
| 30 |
+
In this work, we take the first attempt to model illumination in volume rendering-based generative models, which serves as a regularization for accurate 3D shape learning. We further devise an efficient rendering technique for our approach, which shares similar insight with [12], but does not rely on ground truth depth for training and it is not limited to a small viewpoint range.
|
| 31 |
+
|
| 32 |
+
Generative 3D-aware image synthesis. Generative adversarial networks (GANs) [18] are capable of generating photorealistic images of high-resolution, but lack explicit control over camera viewpoint. In order to enable them to synthesis images in a 3D-aware manner, many recent approaches investigate how 3D representations could be incorporated into GANs [19, 20, 21, 22, 23, 24, 25, 26, 27, 5, 4, 28, 29, 30]. While some works directly learn from 3D data [19, 20, 21, 22, 30], in this work we focus on approaches that only have access to unconstrained 2D images, which is a more practical setting. Several attempts [23, 24, 25] adopt 3D voxel features with learned neural rendering. These methods produce realistic 3D-aware synthesis, but the 3D voxels are not interpretable, i.e., they cannot be transferred to 3D shapes. By leveraging differentiable renderer, [26] and [27] learn interpretable 3D voxels and meshes respectively, but [26] suffers from limited visual quality due to low voxel resolution while the learned 3D shapes of [27] exhibit noticeable distortions. The success of NeRF has motivated researchers to use radiance fields as the intermediate 3D representation in GANs [5, 4, 28]. While achieving impressive 3D-aware image synthesis with multi-view consistency, the extracted 3D shapes of these approaches are often imprecise and noisy. Our main goal in this work is to address the inaccurate shape by explicitly modeling illumination in the rendering process. This innovation helps achieve better 3D-aware image synthesis with broader applications.
|
| 33 |
+
|
| 34 |
+
Unsupervised 3D shape learning from 2D images. Our work is also related to unsupervised approaches that learn 3D object shapes from unconstrained, monocular view 2D images. While several approaches use external 3D shape templates or 2D key-points as weak supervisions to facilitate learning [31, 32, 33, 34, 35, 36, 37], in this work we consider the harder setting where only 2D images are available. To tackle this problem, most approaches adopt an “analysis-by-synthesis” paradigm [38, 39, 40]. Specifically, they design photo-geometric autoencoders to infer the 3D shape and viewpoint of each image with a reconstruction loss. While succeed in learning the 3D shapes for some object categories, these approaches typically rely on certain regularization to prevent trivial solutions, like the commonly used symmetry assumption on object shapes [39, 40, 31, 32]. Such assumption tends to produce symmetric results that may overlook the asymmetric aspects of objects. Recently, GAN2Shape [41] shows that it is possible to recover 3D shapes for images generated by 2D GANs. This method, however, requires inefficient instance-specific training, and recovers depth maps instead of full 3D representations.
|
| 35 |
+
|
| 36 |
+

|
| 37 |
+
Figure 2: Method overview. Our generator $g _ { \theta }$ models a relightable color field conditioned on a latent code $z \sim p _ { z }$ . To synthesis an image, it performs volume rendering under a random camera pose $\xi \sim p _ { \xi }$ . The rendering process also performs shading with a randomly sampled lighting condition $\mu \sim p _ { \mu }$ . The discriminator learns to distinguish the synthesized images with real images from the training dataset, and the whole model is trained with a GAN loss. Although our model is trained from unconstrained 2D images, it allows explicit control over camera pose and lighting condition during inference.
|
| 38 |
+
|
| 39 |
+
The proposed 3D-aware generative model also serves as a powerful approach for unsupervised 3D shape learning. Compared with aforementioned autoencoder-based methods, our GAN-based approach avoids the need to infer the viewpoint of each image, and does not rely on strong regularizations. In experiments, we demonstrate superior performance over recent state-of-the-art approaches Unsup3d [39] and GAN2Shape [41].
|
| 40 |
+
|
| 41 |
+
# 3 Methodology
|
| 42 |
+
|
| 43 |
+
We consider the problem of 3D-aware image synthesis by learning from a collection of unconstrained and unlabeled 2D images. We argue that modeling shading, i.e., the interaction of illumination and shape, in a generative implicit model enables unsupervised learning of more accurate 3D object shapes. In the following, we first provide some preliminaries on neural radiance fields (NeRF) [6], and then introduce our shading-guided generative implicit model.
|
| 44 |
+
|
| 45 |
+
# 3.1 Preliminaries on Neural Radiance Fields
|
| 46 |
+
|
| 47 |
+
As a deep implicit model, NeRF [6] uses an MLP network to represent a 3D scene as a radiance field. The MLP $f _ { \theta } : ( x , d ) ( \sigma , c )$ takes a 3D coordinate $\pmb { x } \in \mathbb { R } ^ { 3 }$ and a viewing direction $\ b { d } \in \mathbb { S } ^ { 2 }$ as inputs, and outputs a volume density $\sigma \in \mathbb { R } ^ { + }$ and a color $c \in \mathbb { R } ^ { 3 }$ . To render an image under a given camera pose, each pixel color $C$ of the image is obtained via volume rendering along its corresponding camera ray $\bar { \boldsymbol { r } ( t ) } = \boldsymbol { o } + t \boldsymbol { d }$ with near and far bounds $t _ { n }$ and $t _ { f }$ as below:
|
| 48 |
+
|
| 49 |
+
$$
|
| 50 |
+
C ( r ) = \int _ { t _ { n } } ^ { t _ { f } } T ( t ) \sigma ( r ( t ) ) c ( r ( t ) , d ) d t , { \mathrm { ~ w h e r e ~ } } T ( t ) = \exp ( - \int _ { t _ { n } } ^ { t } \sigma ( r ( s ) ) d s ) .
|
| 51 |
+
$$
|
| 52 |
+
|
| 53 |
+
In practice, this volume rendering is implemented with a discretized form using stratified and hierarchical sampling. As this rendering process is differentiable, NeRF could be directly optimized via posed images of a static scene. After training, NeRF allows the rendering of images under new camera poses, achieving high-quality novel view synthesis.
|
| 54 |
+
|
| 55 |
+
# 3.2 Shading-Guided Generative Implicit Model
|
| 56 |
+
|
| 57 |
+
In this work, we are interested in developing a generative implicit model that explicitly models the shading process for 3D-aware image synthesis. To achieve this, we make two extensions to the MLP network in NeRF. First, similar to most deep generative models, it is further conditioned on a latent code $_ z$ sampled from a prior distribution $\bar { \mathcal { N } ( 0 , I ) } ^ { d }$ . Second, instead of directly outputting the color $^ c$ , it outputs a relightable pre-cosine color term $\pmb { a } \in \mathbb { R } ^ { 3 }$ , which is conceptually similar to albedo in the way that it could be shaded under a given lighting condition. While albedo is viewpoint-independent, in this work we do not strictly enforce such independence for $\textbf { \em a }$ in order to account for dataset bias. Thus, our generator $g _ { \theta } : ( x , \dot { d } , z ) ( \sigma , \pmb { a } )$ takes a coordinate $_ { \textbf { \em x } }$ , a viewing direction $^ d$ , and a latent code $_ z$ as inputs, and outputs a volume density $\sigma$ and a pre-cosine color $\textbf { \em a }$ . Note that here $\sigma$ is independent of $^ d$ , while the dependence of $\textbf { \em a }$ on $^ d$ is optional. To obtain the color $C$ of a camera ray $\mathbf { \boldsymbol { r } } ( t ) \mathbf { \dot { \alpha } } = \mathbf { \boldsymbol { o } } + t \mathbf { \dot { \alpha } }$ with near and far bounds $t _ { n }$ and $t _ { f }$ , we calculate the final pre-cosine color $\pmb { A }$ via:
|
| 58 |
+
|
| 59 |
+
$$
|
| 60 |
+
A ( r , z ) = \int _ { t _ { n } } ^ { t _ { f } } T ( t , z ) \sigma ( r ( t ) , z ) a ( r ( t ) , d , z ) d t , { \mathrm { ~ w h e r e ~ } } T ( t , z ) = \exp ( - \int _ { t _ { n } } ^ { t } \sigma ( r ( s ) , z ) d s ) .
|
| 61 |
+
$$
|
| 62 |
+
|
| 63 |
+
We also calculate the normal direction $\textbf { \em n }$ with:
|
| 64 |
+
|
| 65 |
+
$$
|
| 66 |
+
n ( r , z ) = \hat { n } ( r , z ) / \| \hat { n } ( r , z ) \| _ { 2 } , \mathrm { ~ w h e r e ~ } \hat { n } ( r , z ) = - \int _ { t _ { n } } ^ { t _ { f } } T ( t , z ) \sigma ( r ( t ) , z ) \nabla _ { r ( t ) } \sigma ( r ( t ) , z ) d t ,
|
| 67 |
+
$$
|
| 68 |
+
|
| 69 |
+
where $\nabla _ { \pmb { r } ( t ) } \sigma ( \pmb { r } ( t ) , \pmb { z } )$ is the derivative of volume density $\sigma$ with respect to its input coordinate, which naturally captures the local normal direction, and could be calculated via back-propagation. Then the final color $C$ is obtained via Lambertian shading as:
|
| 70 |
+
|
| 71 |
+
$$
|
| 72 |
+
C ( \boldsymbol { r } , z ) = A ( r , z ) ( k _ { a } + k _ { d } \mathrm { m a x } ( 0 , \boldsymbol { l } \cdot \boldsymbol { n } ( r , z ) ) ) ,
|
| 73 |
+
$$
|
| 74 |
+
|
| 75 |
+
where $\ b { l } \in \mathbb S ^ { 2 }$ is the lighting direction, $k _ { a }$ and $k _ { d }$ are the ambient and diffuse coefficients. We provide more discussions on this shading formulation at the end of this subsection.
|
| 76 |
+
|
| 77 |
+
Camera and Lighting Sampling. Eq.(2 - 4) describe the process of rendering a pixel color given a camera ray $\mathbf { } _ { \pmb { r } ( t ) }$ and a lighting condition $\mu = ( l , k _ { a } , k _ { d } )$ . Generating a full image $\pmb { I _ { g } } \in \mathbb { R } ^ { 3 \times H \times W }$ requires one to sample a camera pose $\boldsymbol { \xi }$ and a lighting condition $\pmb { \mu }$ in addition to the latent code $_ z$ i.e., $I _ { g } = G _ { \theta } ( z , \xi , \mu )$ . In our setting, the camera pose $\boldsymbol { \xi }$ could be described by pitch and yaw angles, and is sampled from a prior Gaussian or uniform distribution $p _ { \xi }$ , as also done in previous works [4, 5]. Sampling the camera pose randomly during training would motivate the learned 3D scene to look realistic from different viewpoints. While this multi-view constraint is beneficial for learning a valid 3D representation, it is often insufficient to infer the accurate 3D object shape. Thus, in our approach, we further introduce a multi-lighting constraint by also randomly sampling a lighting condition $\pmb { \mu }$ from a prior distribution $p _ { \mu }$ . In practice, $p _ { \mu }$ could be estimated from the dataset using existing approaches like [39]. We also show in our experiments that a simple and manually tuned prior distribution could also produce reasonable results. As the shading process is sensitive to the normal direction due to the diffuse term $k _ { d } \mathrm { m a x } ( 0 , l \cdot n ( r , z ) )$ in Eq.(4), this multi-lighting constraint would regularize the model to learn more accurate 3D shapes that produce natural shading, as shown in Fig.1 (b).
|
| 78 |
+
|
| 79 |
+
Training. Our generative model follows the paradigm of GANs [18], where the generator is trained together with a discriminator $D$ with parameters $\phi$ in an adversarial manner. During training, the generator generates fake images $I _ { g } = \bar { G } _ { \theta } ( z , \xi , \mu )$ by sampling the latent code $_ z$ , camera pose $\boldsymbol { \xi }$ and lighting condition $\pmb { \mu }$ from their corresponding prior distributions $p _ { z } , p _ { \xi }$ , and $p _ { \mu }$ . Let $\pmb { I }$ denotes real images sampled from the data distribution $p _ { I }$ . We train our model with a non-saturating GAN loss with $R _ { 1 }$ regularization [42]:
|
| 80 |
+
|
| 81 |
+
$$
|
| 82 |
+
\begin{array} { r } { \mathcal { L } ( \theta , \phi ) = \mathbf { E } _ { z \sim p _ { z } , \xi \sim p _ { \xi } , \mu \sim p _ { \mu } } \left[ f \Big ( D _ { \phi } ( G _ { \theta } ( z , \xi , \mu ) ) \Big ) \right] + \mathbf { E } _ { I \sim p _ { \mathcal { D } } } \left[ f ( - D _ { \phi } ( I ) ) + \lambda \| \nabla D _ { \phi } ( I ) \| ^ { 2 } \right] , } \end{array}
|
| 83 |
+
$$
|
| 84 |
+
|
| 85 |
+
where $f ( u ) = - \log ( 1 + \exp ( - u ) )$ , and $\lambda$ controls the strength of regularization. More implementation details are provided in the supplementary material.
|
| 86 |
+
|
| 87 |
+
Discussion. Note that in $\operatorname { E q . } ( 2 \ \cdot \ 4 )$ , we perform shading after $\pmb { A }$ and $\textbf { \em n }$ are obtained via volume rendering. An alternative way is to perform shading at each local spatial point as $\begin{array} { l l l } { c ( { \boldsymbol { r } } ( t ) , d , z ) } & { = } & { { \boldsymbol { a } } ( { \boldsymbol { r } } ( t ) , d , z ) ( k _ { a } + { k _ { d } } { \operatorname* { m a x } } ( 0 , { l } \cdot { \boldsymbol { n } } ( { \boldsymbol { r } } ( t ) , z ) ) ) } \end{array}$ , where $n ( r ( t ) , { \bar { z } } ) =$ $- \nabla _ { \pmb { r } ( t ) } \sigma ( \pmb { r } ( t ) , z ) / \| \nabla _ { \pmb { r } ( t ) } \sigma ( \pmb { r } ( t ) , z ) \| _ { 2 }$ is the local normal. Then we could perform volume rendering using $\mathbf { } c ( \pmb { r } ( t ) , z )$ to get the final pixel color. In practice, we observe that this formulation obtains suboptimal results. An intuitive reason is that in this formulation, the normal direction is normalized at each local point, neglecting the magnitude of $\nabla _ { \pmb { r } ( t ) } \sigma ( \pmb { r } ( t ) , \pmb { z } )$ , which tends to be larger near the object surfaces. We provide more analysis in experiments and the supplementary material.
|
| 88 |
+
|
| 89 |
+
The Lambertian shading we used is an approximation to the real illumination scenario. While serving as a good regularization for improving the learned 3D shape, it could possibly introduce an additional gap between the distribution of generated images and that of real images. To compensate for such risk, we could optionally let the predicted $\textbf { \em a }$ be conditioned on the lighting condition, i.e., $\pmb { a } = \pmb { a } ( \pmb { r } ( t ) , \pmb { d } , \pmb { \mu } , z )$ . Thus, in cases where the lighting condition deviates from the real data distribution, the generator could learn to adjust the value of $\textbf { \em a }$ and reduce the aforementioned gap. We show the benefit of this design in the experiments.
|
| 90 |
+
|
| 91 |
+

|
| 92 |
+
Figure 3: (a) Conventional volume rendering samples dozens of points within a predefined near and far bounds $t _ { n }$ and $t _ { f }$ . (b) We propose an efficient volume rendering technique via surface tracking. Before rendering, our surface tracking network $S _ { \psi }$ predicts an initial guess of the surface position $s$ conditioned on the latent code $_ z$ and camera pose $\xi$ . Then we sample points near $s$ , which requires fewer samples. Finally, we use the volume rendered depth $d$ as the ground truth to train $S _ { \psi }$ . During training, $S _ { \psi }$ is able to predict depth $s$ that well approximates the real surface depth $d$ .
|
| 93 |
+
|
| 94 |
+
# 3.3 Efficient Volume Rendering via Surface Tracking
|
| 95 |
+
|
| 96 |
+
Similar to NeRF, we implement volume rendering with a discretized integral, which typically requires to sample dozens of points along a camera ray, as shown in Fig. 3 (a). In our approach, we also need to perform back-propagation across the generator in Eq.(3) to get the normal direction for each point, which introduces additional computational cost. To achieve more efficient volume rendering, a natural idea is to exploit spatial sparsity. Usually, the weight $T ( t , z ) \sigma ( \pmb { r } ( t ) , z )$ in volume rendering would concentrate on the object surface position during training. Thus, if we know the rough surface position before rendering, we could sample points near the surface to save computation. While for a static scene it is possible to store such spatial sparsity in a sparse voxel grid [8, 9], this technique cannot be directly applied to our generative model, as the 3D scene keeps changing with respect to the input latent code.
|
| 97 |
+
|
| 98 |
+
To achieve more efficient volume rendering in our generative implicit model, we further propose a surface tracking network $S$ that learns to mimic the surface position conditioned on the latent code. In particular, the volume rendering naturally allows the depth estimation of the object surface via:
|
| 99 |
+
|
| 100 |
+
$$
|
| 101 |
+
t _ { s } ( \pmb { r } , z ) = \int _ { t _ { n } } ^ { t _ { f } } T ( t , z ) \sigma ( \pmb { r } ( t ) , z ) t d t ,
|
| 102 |
+
$$
|
| 103 |
+
|
| 104 |
+
where $T ( t , z )$ is defined the same way as in Eq.(2). Thus, given a camera pose $\boldsymbol { \xi }$ and a latent code $_ z$ , we could render the full depth map $\pmb { t } _ { s } ( z , \pmb { \xi } )$ . As shown in Fig. 3 (b), we mimic $\pmb { t } _ { s } ( z , \pmb { \xi } )$ with the surface tracking network $S _ { \psi }$ , which is a light-weighted convolutional neural network that takes $z , \xi$ as inputs and outputs a depth map. The depth mimic loss is:
|
| 105 |
+
|
| 106 |
+
$$
|
| 107 |
+
\mathcal { L } ( \psi ) = \mathbf { E } _ { z \sim p _ { z } , \xi \sim p _ { \xi } } \left[ \| S _ { \psi } ( z , \xi ) - t _ { s } ( z , \xi ) \| _ { 1 } + \operatorname* { P r e c } ( S _ { \psi } ( z , \xi ) , t _ { s } ( d ( z , \xi ) ) \right] ,
|
| 108 |
+
$$
|
| 109 |
+
|
| 110 |
+
where Prec is the perceptual loss that motivates $S _ { \psi }$ to better capture edges of the surface.
|
| 111 |
+
|
| 112 |
+
During training, $S _ { \psi }$ is optimized jointly with the generator and the discriminator. Thus, each time after we sample a latent code $_ z$ and a camera pose $\boldsymbol { \xi }$ , we can get an initial guess of the depth map as $S _ { \psi } ( z , \pmb { \xi } )$ . Then for a pixel with predicted depth $s$ , we could perform volume rendering in Eq.(2,3,6) with near bound $t _ { n } = s - \Delta _ { i } / 2$ and far bound $t _ { f } = s + \Delta _ { i } / 2$ , where $\Delta _ { i }$ is the interval for volume rendering that decreases as the training iteration $i$ grows. Specifically, we start with a large interval $\Delta _ { m a x }$ and decrease to $\Delta _ { m i n }$ with an exponential schedule. As $\Delta _ { i }$ decreases, the number of points used for rendering $m$ also decreases accordingly. Note that the computational cost of our efficient surface tracking network is marginal compared to the generator, as the former only needs a single forward pass to render an image while the latter will be queried for $H \times W \times m$ times. Thus, the reduction of $m$ would significantly accelerate the training and inference speed for ShadeGAN.
|
| 113 |
+
|
| 114 |
+

|
| 115 |
+
Figure 4: Qualitative comparison on BFM (top), CelebA (middle), and Cats (bottom) datasets. "Albedo" refers to the pre-cosine color that approximates albedo. Our approach synthesizes more accurate 3D shapes than pi-GAN and GRAF, and also learns to disentangle shading with albedo.
|
| 116 |
+
|
| 117 |
+
# 4 Experiments
|
| 118 |
+
|
| 119 |
+
In this section, we evaluate the proposed ShadeGAN on 3D-aware image synthesis. We also show that ShadeGAN learns much more accurate 3D shapes than previous methods, and in the meantime allows explicit control over lighting conditions. The datasets used include CelebA [43], BFM [13], and Cats [44], all of which contain only unconstrained 2D RGB images.
|
| 120 |
+
|
| 121 |
+
Implementation. In terms of model architectures, we adopt a SIREN-based MLP [45] as the generator and a convolutional neural network as the discriminator following [4]. For the prior distribution of lighting conditions, we use Unsup3d [39] to estimate the lighting conditions of real data and subsequently fit a multivariate Gaussian distribution of $\pmb { \mu } = ( l , k _ { a } , k _ { d } )$ as the prior. A hand-crafted prior distribution is also included in the ablation study. In quantitative study, we let the pre-cosine color $\textbf { \em a }$ be conditioned on the lighting condition $\pmb { \mu }$ as well as the viewing direction $^ d$ unless otherwise stated. In qualitative study, we observe that removing view conditioning achieves slightly better 3D shapes for CelebA and BFM datasets. Thus, we show results without view conditioning for these two datasets in the main paper, and put those with view conditioning in Fig. 4 of the supplementary material. Other implementation details are also provided in the supplementary.
|
| 122 |
+
|
| 123 |
+

|
| 124 |
+
Figure 5: Generated face images and their 3D meshes.
|
| 125 |
+
|
| 126 |
+

|
| 127 |
+
Figure 6: Qualitative ablation. See the main text for discussions.
|
| 128 |
+
|
| 129 |
+
Table 1: Comparisons on the BFM dataset. We report FID $\mathsf { \bar { \rho } _ { 1 2 8 } } 2 \mathsf { \bar { \rho } _ { 1 } }$ for image synthesis, and SIDE $\bar { ( } \times 1 0 ^ { - 2 } )$ and MAD (deg.) for the accuracy of 3D shapes. ‘-’ indicates not available. Results of pi-GAN and Ours are averaged over 5 runs.
|
| 130 |
+
|
| 131 |
+
<table><tr><td>Method</td><td>FID↓</td><td>SIDE↓</td><td>MAD↓</td></tr><tr><td>Supervised</td><td>-</td><td>0.410</td><td>10.78</td></tr><tr><td>Unsup3d [39]</td><td>-</td><td>0.793</td><td>16.51</td></tr><tr><td>GAN2Shape [41]</td><td>-</td><td>0.756</td><td>14.81</td></tr><tr><td>GRAF[5]</td><td>53.4</td><td>1.857</td><td>26.60</td></tr><tr><td>pi-GAN [4]</td><td>16.7±0.2</td><td>0.727±0.012</td><td>20.09±0.23</td></tr><tr><td>Ours</td><td>17.7±0.2</td><td>0.607±0.007</td><td>14.52±0.11</td></tr></table>
|
| 132 |
+
|
| 133 |
+
Table 2: Comparisons on the CelebA and Cats datasets. The image resolution is $1 2 8 ^ { 2 }$ .
|
| 134 |
+
|
| 135 |
+
<table><tr><td>Dataset</td><td>Method FID↓</td><td>MAD↓</td></tr><tr><td rowspan="3">CelebA</td><td>GRAF</td><td>43.0 30.48</td></tr><tr><td>pi-GAN 15.7</td><td>27.22</td></tr><tr><td>Ours 16.2</td><td>20.49</td></tr><tr><td rowspan="3">Cats</td><td>GRAF</td><td>30.3 65.47</td></tr><tr><td>pi-GAN</td><td>10.7 33.48</td></tr><tr><td>Ours 10.3</td><td>25.47</td></tr></table>
|
| 136 |
+
|
| 137 |
+
Comparison with baselines. We compare ShadeGAN with two state-of-the-art generative implicit models, namely GRAF [5] and pi-GAN [4]. Specifically, Fig. 4 includes both synthesized images as well as their corresponding 3D meshes, which are obtained by performing marching cubes on the volume density $\sigma$ . While GRAF and pi-GAN could synthesize images with controllable poses, their learned 3D shapes are inaccurate and noisy. In contrast, our approach not only synthesizes photorealistic 3D-consistent images, but also learns much more accurate 3D shapes and surface normals, indicating the effectiveness of the proposed multi-lighting constraint as a regularization. More synthesized images and their corresponding shapes are included in Fig.5. Besides more accurate 3D shapes, ShadeGAN can also learn the albedo and diffuse shading components inherently. As shown in Fig. 4, although not perfect, ShadeGAN has managed to disentangle shading and albedo with satisfying quality, as such disentanglement is a natural solution to the multi-lighting constraint.
|
| 138 |
+
|
| 139 |
+
The quality of learned 3D shapes is quantitatively evaluated on the BFM dataset. Specifically, we use each of the generative implicit models to generate 50k images and their corresponding depth maps. Image-depth pairs from each model are used as training data to train an additional convolutional neural network (CNN) that learns to predict the depth map of an input image. We then test each trained CNN on the BFM test set and compare its predictions to the ground-truth depth maps as a measurement of the quality of learned 3D shapes. Following [39], we report the scale-invariant depth error (SIDE) and mean angle deviation (MAD) metrics. The results are included in Tab. 1, where ShadeGAN significantly outperforms GRAF and pi-GAN. Besides, ShadeGAN also outperforms other advanced unsupervised 3D shape learning approaches including Unsup3d [39] and GAN2Shape [41], demonstrating its large potential in unsupervised 3D shapes learning. In terms of image quality, Tab. 1 includes the FID [46] scores of images synthesized by different models, where the FID score of ShadeGAN is slightly inferior to pi-GAN in BFM and CelebA. Intuitively, this is caused by the gap between our approximated shading (i.e. Lambertian shading) and the real illumination, which can be potentially avoided by adopting more realistic shading models and improving the lighting prior.
|
| 140 |
+
|
| 141 |
+
In Tab. 2, we also show the quantitative results of different models on CelebA and Cats. To evaluate the learned shape, we use each generative implicit model to generate 2k front-view images and their corresponding depth maps. While these datasets do not have ground truth depth, we report MAD obtained by testing pretrained Unsup3d models [39] on these generated image-depth pairs as a reference. As we can observe, results on CelebA and Cat are consistent with those on the BFM dataset.
|
| 142 |
+
|
| 143 |
+
Table 3: Ablation study on the BFM dataset.
|
| 144 |
+
|
| 145 |
+
<table><tr><td>No.</td><td>Method</td><td>FID↓</td><td>SIDE↓</td><td>MAD</td></tr><tr><td>(1)</td><td>ShadeGAN</td><td>17.7</td><td>0.607</td><td>14.52</td></tr><tr><td>(2)</td><td>local shading</td><td>30.1</td><td>0.754</td><td>18.18</td></tr><tr><td>(3)</td><td>w/o light</td><td>19.2</td><td>0.618</td><td>14.53</td></tr><tr><td>(4)</td><td>w/o view</td><td>18.6</td><td>0.622</td><td>14.88</td></tr><tr><td>(5)</td><td>manual prior</td><td>20.2</td><td>0.643</td><td>15.38</td></tr><tr><td>(6)</td><td>+efficient</td><td>18.2</td><td>0.673</td><td>14.72</td></tr></table>
|
| 146 |
+
|
| 147 |
+
Table 4: Training and inference time cost on CelebA. The efficient volume rendering significantly improves training and inference speed.
|
| 148 |
+
|
| 149 |
+
<table><tr><td>Method</td><td>Train (h) Inference (s) FID</td></tr><tr><td>ShadeGAN</td><td>92.3 0.343</td></tr><tr><td>70.2</td><td>16.4 0.179 16.2</td></tr><tr><td>+efficient pi-GAN</td><td></td></tr><tr><td>56.8</td><td>0.204 15.7</td></tr><tr><td>+efficient 46.9</td><td>0.114 15.9</td></tr></table>
|
| 150 |
+
|
| 151 |
+

|
| 152 |
+
Figure 7: Visualization of depths predicted by our depth tracking network and those calculated via volume rendering.
|
| 153 |
+
|
| 154 |
+

|
| 155 |
+
Figure 8: Illumination-aware image synthesis. ShadeGAN allows explicit control over the lighting. The pre-cosine color (albedo) is independent of lighting in (a) and is conditioned on lighting in (b). We show results of adding a specular term in (c).
|
| 156 |
+
|
| 157 |
+
Ablation studies. We further study the effects of several design choices in ShadeGAN. First, we perform local points-specific shading as mentioned in the discussion of Sec. 3.2. As Tab. 3 No.(2) and Fig. 6 (b) show, the results of such a local shading strategy are notably worse than the original one, which indicates that taking the magnitude of $\nabla _ { \pmb { x } } \sigma$ into account is beneficial. Besides, the results of Tab. 3 No.(3) and No.(4) imply that removing $\textbf { \em a }$ ’s dependence on the lighting $\pmb { \mu }$ or the viewpoint $^ d$ could lead to a slight performance drop. The results of using a simple manually tuned lighting prior are provided in Tab. 3 No.(5) and Fig. 6 (c), which are only moderately worse than the results of using a data-driven prior, and the generated shapes are still significantly better than the ones produced by existing approaches.
|
| 158 |
+
|
| 159 |
+
To verify the effectiveness of the proposed efficient volume rendering technique, we include its effects on image quality and training/inference time in Tab. 3 No.(6) and Tab. 4. It is observed that the efficient volume rendering has marginal effects on the performance, but significantly reduces the training and inference time by $24 \%$ and $48 \%$ for ShadeGAN. Moreover, in Fig. 7 we visualize the depth maps predicted by our surface tracking network and those obtained via volume rendering. It is shown that under varying identities and camera poses, the surface tracking network could consistently predict depth values that are quite close to the real surface positions, so that we can sample points near the predicted surface for rendering without sacrificing image quality.
|
| 160 |
+
|
| 161 |
+
Illumination-aware image synthesis. As ShadeGAN models the shading process, it by design allows explicit control over the lighting condition. We provide such illumination-aware image synthesis results in Fig.8, where ShadeGAN generates promising images under different lighting directions. We also show that in cases where the predicted $\textbf { \em a }$ is conditioned on the lighting condition $\pmb { \mu }$ , $\textbf { \em a }$ would slightly change $w . r . t .$ the lighting condition, e.g., it would be brighter in areas having a overly dim shading in order to make the final image more natural. Besides, we could optionally add a specular term $k _ { s } \operatorname* { m a x } ( 0 , \boldsymbol { h } \cdot \boldsymbol { n } ) ^ { p }$ in Eq. 4 (i.e., Blinn-Phong shading [47], where $^ { h }$ is the bisector of the angle between the viewpoint and the lighting direction) to create specular highlight effects, as shown in Fig.8 (c).
|
| 162 |
+
|
| 163 |
+
GAN inversion. ShadeGAN could also be used to reconstruct a given target image by performing GAN inversion. As shown in Fig. 9 such inversion allows us to obtain several factors of the image, including the 3D shape, surface normal, approximated albedo, and shading. Besides, we can further perform view synthesis and relighting by changing the viewpoint and lighting condition. The implementation of GAN inversion is provided in the supplementary material.
|
| 164 |
+
|
| 165 |
+

|
| 166 |
+
Figure 9: GAN inversion for real image editing.
|
| 167 |
+
|
| 168 |
+
Discussions. As the Lambertian shading we used is an approximation to the real illumination, the albedo learned by ShadeGAN is not perfectly disentangled. Our approach does not consider the spatially-varying material properties of objects as well. In the future, we intend to incorporate more sophisticated shading models to learn better disentangled generative reflectance fields.
|
| 169 |
+
|
| 170 |
+
# 5 Conclusion
|
| 171 |
+
|
| 172 |
+
In this work, we present ShadeGAN, a new generative implicit model for shape-accurate 3D-aware image synthesis. We have shown that the multi-lighting constraint, achieved in ShadeGAN by explicit illumination modeling, significantly helps learning accurate 3D shapes from 2D images. ShadeGAN also allows us to control the lighting condition during image synthesis, achieving natural image relighting effects. To reduce the computational cost, we have further devised a light-weighted surface tracking network, which enables an efficient volume rendering technique for generative implicit models, achieving significant acceleration on both training and inference speed. A generative model with shape-accurate 3D representation could broaden its applications in vision and graphics, and our work has taken a solid step towards this goal.
|
| 173 |
+
|
| 174 |
+
Acknowledgment. We would like to thank Eric R. Chan for sharing the codebase of pi-GAN. This study is supported under the ERC Consolidator Grant 4DRepLy (770784). This study is also supported under the RIE2020 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).
|
| 175 |
+
|
| 176 |
+
# References
|
| 177 |
+
|
| 178 |
+
[1] T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” in CVPR, 2019.
|
| 179 |
+
[2] T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, “Analyzing and improving the image quality of stylegan,” in CVPR, 2020.
|
| 180 |
+
[3] A. Brock, J. Donahue, and K. Simonyan, “Large scale gan training for high fidelity natural image synthesis,” in ICLR, 2019.
|
| 181 |
+
[4] E. R. Chan, M. Monteiro, P. Kellnhofer, J. Wu, and G. Wetzstein, “pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis,” in CVPR, 2021.
|
| 182 |
+
[5] K. Schwarz, Y. Liao, M. Niemeyer, and A. Geiger, “Graf: Generative radiance fields for 3d-aware image synthesis,” in NeurIPS, 2020.
|
| 183 |
+
[6] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” in ECCV, 2020.
|
| 184 |
+
[7] R. J. Woodham, “Photometric method for determining surface orientation from multiple images,” Optical engineering, vol. 19, no. 1, p. 191139, 1980.
|
| 185 |
+
[8] L. Liu, J. Gu, K. Zaw Lin, T.-S. Chua, and C. Theobalt, “Neural sparse voxel fields,” in NeurIPS, 2020.
|
| 186 |
+
[9] A. Yu, R. Li, M. Tancik, H. Li, R. Ng, and A. Kanazawa, “PlenOctrees for real-time rendering of neural radiance fields,” in arXiv, 2021.
|
| 187 |
+
[10] C. Reiser, S. Peng, Y. Liao, and A. Geiger, “Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,” arXiv preprint arXiv:2103.13744, 2021.
|
| 188 |
+
[11] S. J. Garbin, M. Kowalski, M. Johnson, J. Shotton, and J. Valentin, “Fastnerf: High-fidelity neural rendering at 200fps,” arXiv preprint arXiv:2103.10380, 2021.
|
| 189 |
+
[12] T. Neff, P. Stadlbauer, M. Parger, A. Kurz, C. R. A. Chaitanya, A. Kaplanyan, and M. Steinberger, “Donerf: Towards real-time rendering of neural radiance fields using depth oracle networks,” arXiv preprint arXiv:2103.03231, 2021.
|
| 190 |
+
[13] P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter, “A 3d face model for pose and illumination invariant face recognition,” in 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, Ieee, 2009.
|
| 191 |
+
[14] M. Boss, R. Braun, V. Jampani, J. T. Barron, C. Liu, and H. Lensch, “Nerd: Neural reflectance decomposition from image collections,” arXiv preprint arXiv:2012.03918, 2020.
|
| 192 |
+
[15] S. Bi, Z. Xu, P. Srinivasan, B. Mildenhall, K. Sunkavalli, M. Hašan, Y. Hold-Geoffroy, D. Kriegman, and R. Ramamoorthi, “Neural reflectance fields for appearance acquisition,” arXiv preprint arXiv:2008.03824, 2020.
|
| 193 |
+
[16] P. P. Srinivasan, B. Deng, X. Zhang, M. Tancik, B. Mildenhall, and J. T. Barron, “Nerv: Neural reflectance and visibility fields for relighting and view synthesis,” in CVPR, 2021.
|
| 194 |
+
[17] D. B. Lindell, J. N. Martel, and G. Wetzstein, “Autoint: Automatic integration for fast neural volume rendering,” arXiv preprint arXiv:2012.01714, 2020.
|
| 195 |
+
[18] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in NIPS, 2014.
|
| 196 |
+
[19] J. Wu, C. Zhang, T. Xue, B. Freeman, and J. Tenenbaum, “Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling,” in NIPS, 2016.
|
| 197 |
+
[20] H. A. Alhaija, S. K. Mustikovela, A. Geiger, and C. Rother, “Geometric image synthesis,” in ACCV, Springer, 2018.
|
| 198 |
+
[21] X. Chen, D. Cohen-Or, B. Chen, and N. J. Mitra, “Towards a neural graphics pipeline for controllable image generation,” Computer Graphics Forum, vol. 40, no. 2, 2021.
|
| 199 |
+
[22] J.-Y. Zhu, Z. Zhang, C. Zhang, J. Wu, A. Torralba, J. Tenenbaum, and B. Freeman, “Visual object networks: Image generation with disentangled 3d representations,” in NeurIPS, 2018.
|
| 200 |
+
[23] T. Nguyen-Phuoc, C. Li, L. Theis, C. Richardt, and Y.-L. Yang, “Hologan: Unsupervised learning of 3d representations from natural images,” in ICCV, 2019.
|
| 201 |
+
[24] T. Nguyen-Phuoc, C. Richardt, L. Mai, Y.-L. Yang, and N. Mitra, “Blockgan: Learning 3d object-aware scene representations from unlabelled images,” in NeurIPS, Nov 2020.
|
| 202 |
+
[25] Y. Liao, K. Schwarz, L. Mescheder, and A. Geiger, “Towards unsupervised learning of generative models for 3d controllable image synthesis,” in CVPR, 2020.
|
| 203 |
+
[26] P. Henzler, N. J. Mitra, and T. Ritschel, “Escaping plato’s cave: 3d shape from adversarial rendering,” in ICCV, 2019.
|
| 204 |
+
[27] A. Szabó, G. Meishvili, and P. Favaro, “Unsupervised generative 3d shape learning from natural images,” arXiv preprint arXiv:1910.00287, 2019.
|
| 205 |
+
[28] M. Niemeyer and A. Geiger, “Giraffe: Representing scenes as compositional generative neural feature fields,” in CVPR, 2021.
|
| 206 |
+
[29] A. Tewari, M. Elgharib, G. Bharaj, F. Bernard, H.-P. Seidel, P. Pérez, M. Zollhofer, and C. Theobalt, “Stylerig: Rigging stylegan for 3d control over portrait images,” in CVPR, 2020.
|
| 207 |
+
[30] T. Yenamandra, A. Tewari, F. Bernard, H.-P. Seidel, M. Elgharib, D. Cremers, and C. Theobalt, “i3dmm: Deep implicit 3d morphable model of human heads,” in CVPR, 2021.
|
| 208 |
+
[31] S. Tulsiani, N. Kulkarni, and A. Gupta, “Implicit mesh reconstruction from unannotated image collections,” arXiv preprint arXiv:2007.08504, 2020.
|
| 209 |
+
[32] S. Goel, A. Kanazawa, and J. Malik, “Shape and viewpoint without keypoints,” ECCV, 2020.
|
| 210 |
+
[33] L. Tran and X. Liu, “Nonlinear 3d face morphable model,” in CVPR, 2018.
|
| 211 |
+
[34] A. Kanazawa, S. Tulsiani, A. A. Efros, and J. Malik, “Learning category-specific mesh reconstruction from image collections,” in ECCV, 2018.
|
| 212 |
+
[35] A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik, “End-to-end recovery of human shape and pose,” in CVPR, 2018.
|
| 213 |
+
[36] S. Sanyal, T. Bolkart, H. Feng, and M. J. Black, “Learning to regress 3d face shape and expression from an image without 3d supervision,” in CVPR, 2019.
|
| 214 |
+
[37] J. Shang, T. Shen, S. Li, L. Zhou, M. Zhen, T. Fang, and L. Quan, “Self-supervised monocular 3d face reconstruction by occlusion-aware multi-view geometry consistency,” ECCV, 2020.
|
| 215 |
+
[38] M. Sahasrabudhe, Z. Shu, E. Bartrum, R. Alp Guler, D. Samaras, and I. Kokkinos, “Lifting autoencoders: Unsupervised learning of a fully-disentangled 3d morphable model using deep non-rigid structure from motion,” in ICCV Workshops, 2019.
|
| 216 |
+
[39] S. Wu, C. Rupprecht, and A. Vedaldi, “Unsupervised learning of probably symmetric deformable 3d objects from images in the wild,” in CVPR, 2020.
|
| 217 |
+
[40] X. Li, S. Liu, K. Kim, S. De Mello, V. Jampani, M.-H. Yang, and J. Kautz, “Self-supervised single-view 3d reconstruction via semantic consistency,” ECCV, 2020.
|
| 218 |
+
[41] X. Pan, B. Dai, Z. Liu, C. C. Loy, and P. Luo, “Do 2d gans know 3d shape? unsupervised 3d shape reconstruction from 2d image gans,” in ICLR, 2021.
|
| 219 |
+
[42] L. Mescheder, A. Geiger, and S. Nowozin, “Which training methods for gans do actually converge?,” in ICML, 2018.
|
| 220 |
+
[43] Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in ICCV, 2015.
|
| 221 |
+
[44] W. Zhang, J. Sun, and X. Tang, “Cat head detection-how to effectively exploit shape and texture features,” in ECCV, 2008.
|
| 222 |
+
[45] V. Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wetzstein, “Implicit neural representations with periodic activation functions,” in NeurIPS, 2020.
|
| 223 |
+
[46] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” in NeurIPS, 2017.
|
| 224 |
+
[47] B. T. Phong, “Illumination for computer generated pictures,” Communications of the ACM, vol. 18, no. 6, pp. 311–317, 1975.
|
| 225 |
+
[48] E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” in AAAI, 2018.
|
| 226 |
+
[49] R. Liu, J. Lehman, P. Molino, F. P. Such, E. Frank, A. Sergeev, and J. Yosinski, “An intriguing failing of convolutional neural networks and the coordconv solution,” in NeurIPS, 2018.
|
| 227 |
+
[50] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR, 2016.
|
| 228 |
+
[51] Y. Wu and K. He, “Group normalization,” in ECCV, 2018.
|
| 229 |
+
[52] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in ICML, 2015.
|
| 230 |
+
[53] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.
|
| 231 |
+
[54] “Unsup3d.” https://github.com/elliottwu/unsup3d.
|
| 232 |
+
[55] “Celeba.” http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html.
|
| 233 |
+
[56] “Cats.” https://web.archive.org/web/20150520175645/http://137.189.35.203/ WebUI/CatDatabase/catData.html.
|
| 234 |
+
[57] Y. Choi, Y. Uh, J. Yoo, and J.-W. Ha, “Stargan v2: Diverse image synthesis for multiple domains,” in CVPR, 2020.
|
parse/train/k-0oq5eNjh/k-0oq5eNjh_content_list.json
ADDED
|
@@ -0,0 +1,974 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "A Shading-Guided Generative Implicit Model for Shape-Accurate 3D-Aware Image Synthesis ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
200,
|
| 8 |
+
122,
|
| 9 |
+
797,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Xingang Pan1 Xudong Xu2 Chen Change Loy3 Christian Theobalt1 Bo Dai3 ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
186,
|
| 19 |
+
223,
|
| 20 |
+
805,
|
| 21 |
+
241
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "1Max Planck Institute for Informatics 2The Chinese University of Hong Kong {xpan,theobalt}@mpi-inf.mpg.de xx018@ie.cuhk.edu.hk 3S-Lab, Nanyang Technological University {ccloy, bo.dai}@ntu.edu.sg ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
235,
|
| 30 |
+
261,
|
| 31 |
+
764,
|
| 32 |
+
318
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Abstract ",
|
| 39 |
+
"text_level": 1,
|
| 40 |
+
"bbox": [
|
| 41 |
+
462,
|
| 42 |
+
354,
|
| 43 |
+
535,
|
| 44 |
+
371
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "The advancement of generative radiance fields has pushed the boundary of 3Daware image synthesis. Motivated by the observation that a 3D object should look realistic from multiple viewpoints, these methods introduce a multi-view constraint as regularization to learn valid 3D radiance fields from 2D images. Despite the progress, they often fall short of capturing accurate 3D shapes due to the shapecolor ambiguity, limiting their applicability in downstream tasks. In this work, we address this ambiguity by proposing a novel shading-guided generative implicit model that is able to learn a starkly improved shape representation. Our key insight is that an accurate 3D shape should also yield a realistic rendering under different lighting conditions. This multi-lighting constraint is realized by modeling illumination explicitly and performing shading with various lighting conditions. Gradients are derived by feeding the synthesized images to a discriminator. To compensate for the additional computational burden of calculating surface normals, we further devise an efficient volume rendering strategy via surface tracking, reducing the training and inference time by $24 \\%$ and $48 \\%$ , respectively. Our experiments on multiple datasets show that the proposed approach achieves photorealistic 3D-aware image synthesis while capturing accurate underlying 3D shapes. We demonstrate improved performance of our approach on 3D shape reconstruction against existing methods, and show its applicability on image relighting. Our code will be released at https://github.com/XingangPan/ShadeGAN. ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
233,
|
| 53 |
+
383,
|
| 54 |
+
766,
|
| 55 |
+
660
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "1 Introduction ",
|
| 62 |
+
"text_level": 1,
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
684,
|
| 66 |
+
310,
|
| 67 |
+
702
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Advanced deep generative models, e.g., StyleGAN [1, 2] and BigGAN [3], have achieved great successes in natural image synthesis. While producing impressive results, these 2D representationbased models cannot synthesize novel views of an instance in a 3D-consistent manner. They also fall short of representing an explicit 3D object shape. To overcome such limitations, researchers have proposed new deep generative models that represent 3D scenes as neural radiance fields [4, 5]. Such 3D-aware generative models allow explicit control of viewpoint while preserving 3D consistency during image synthesis. Perhaps a more fascinating merit is that they have shown the great potential of learning 3D shapes in an unsupervised manner from just a collection of unconstrained 2D images. If we could train a 3D-aware generative model that learns accurate 3D object shapes, it would broaden various downstream applications such as 3D shape reconstruction and image relighting. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
715,
|
| 77 |
+
825,
|
| 78 |
+
854
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Existing attempts for 3D-aware image synthesis [4, 5] tend to learn coarse 3D shapes that are inaccurate and noisy, as shown in Fig.1 (a). We found that such inaccuracy arises from an inevitable ambiguity inherent in the training strategy adopted by these methods. In particular, a form of regularization, which we refer to as \"multi-view constraint\", is used to enforce the 3D representation to look realistic from different viewpoints. The constraint is commonly implemented by first projecting the generator’s outputs (e.g., radiance fields [6]) to randomly sampled viewpoints, and then feeding them to a discriminator as fake images for training. While such a constraint enables these models to synthesize images in a 3D-aware manner, it suffers from the shape-color ambiguity, i.e., small variations of shape could lead to similar RGB images that look equally plausible to the discriminator, as the color of many objects is locally smooth. Consequently, inaccurate shapes are concealed under this constraint. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
176,
|
| 87 |
+
861,
|
| 88 |
+
825,
|
| 89 |
+
902
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/e70236e34de27c607f69ef44fbb3dc035034e964eca5ef93cf793f863a16cdda.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: Motivation. (a) Previous methods like pi-GAN [4] resort to the \"multi-view constraint\", where the 3D representation is projected to different viewpoints as fake images to the discriminator. The extracted 3D meshes are often inaccurate due to the shape-color ambiguity. (b) The proposed approach ShadeGAN further adopts a \"multi-lighting constraint\", which motivates the 3D representation to look realistic under different lighting conditions. This constraint effectively addresses the ambiguity, giving rise to more natural and precise 3D shapes. "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
199,
|
| 102 |
+
88,
|
| 103 |
+
790,
|
| 104 |
+
304
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "",
|
| 111 |
+
"bbox": [
|
| 112 |
+
174,
|
| 113 |
+
407,
|
| 114 |
+
825,
|
| 115 |
+
518
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "In this work, we propose a novel shading-guided generative implicit model (ShadeGAN) to address the aforementioned ambiguity. In particular, ShadeGAN learns more accurate 3D shapes by explicitly modeling shading, i.e., the interaction of illumination and shape. We believe that an accurate 3D shape should look realistic not only from different viewpoints, but also under different lighting conditions, i.e., satisfying the \"multi-lighting constraint\". This idea shares similar intuition with photometric stereo [7], which shows that accurate surface normal could be recovered from images taken under different lighting conditions. Note that the multi-lighting constraint is feasible as real-world images used for training are often taken under various lighting conditions. To fulfill this constraint, ShadeGAN takes a relightable color field as the intermediate representation, which approximates the albedo but does not necessarily satisfy viewpoint independence. The color field is shaded under a randomly sampled lighting condition during rendering. Since image appearance via such a shading process is strongly dependent on surface normals, inaccurate 3D shape representations will be much more clearly revealed than in earlier shading-agnostic generative models. Hence, by satisfying the multi-lighting constraint, ShadeGAN is encouraged to infer more accurate 3D shapes as shown in Fig.1 (b). ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
525,
|
| 125 |
+
825,
|
| 126 |
+
732
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "The above shading process requires the calculation of the normal direction via back-propagation through the generator, and such calculation needs to be repeated dozens of times for a pixel in volume rendering [4, 5], introducing additional computational overhead. Existing efficient volume rendering techniques [8, 9, 10, 11, 12] mainly target static scenes, and could not be directly applied to generative models due to their dynamic nature. Therefore, to improve the rendering speed of ShadeGAN, we formulate an efficient surface tracking network to estimate the rendered object surface conditioned on the latent code. This enables us to save rendering computations by just querying points near the predicted surface, leading to $24 \\%$ and $48 \\%$ reduction of training and inference time without affecting the quality of rendered images. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
738,
|
| 136 |
+
825,
|
| 137 |
+
863
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "Comprehensive experiments are conducted across multiple datasets to verify the effectiveness of ShadeGAN. The results show that our approach is capable of synthesizing photorealistic images while capturing more accurate underlying 3D shapes than previous generative methods. The learned distribution of 3D shapes enables various downstream tasks like 3D shape reconstruction, where our approach significantly outperforms other baselines on the BFM dataset [13]. Besides, modeling the shading process enables explicit control over lighting conditions, achieving image relighting effect. Our contributions can be summarized as follows: 1) We address the shape-color ambiguity in existing 3D-aware image synthesis methods with a shading-guided generative model that satisfies the proposed multi-lighting constraint. In this way, ShadeGAN is able to learn more accurate 3D shapes for better image synthesis. 2) We devise an efficient rendering technique via surface tracking, which significantly saves training and inference time for volume rendering-based generative models. 3) We show that ShadeGAN learns to disentangle shading and color that well approximates the albedo, achieving natural relighting effects in image synthesis. ",
|
| 144 |
+
"bbox": [
|
| 145 |
+
176,
|
| 146 |
+
869,
|
| 147 |
+
825,
|
| 148 |
+
911
|
| 149 |
+
],
|
| 150 |
+
"page_idx": 1
|
| 151 |
+
},
|
| 152 |
+
{
|
| 153 |
+
"type": "text",
|
| 154 |
+
"text": "",
|
| 155 |
+
"bbox": [
|
| 156 |
+
174,
|
| 157 |
+
90,
|
| 158 |
+
825,
|
| 159 |
+
229
|
| 160 |
+
],
|
| 161 |
+
"page_idx": 2
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"type": "text",
|
| 165 |
+
"text": "2 Related Work ",
|
| 166 |
+
"text_level": 1,
|
| 167 |
+
"bbox": [
|
| 168 |
+
174,
|
| 169 |
+
251,
|
| 170 |
+
321,
|
| 171 |
+
268
|
| 172 |
+
],
|
| 173 |
+
"page_idx": 2
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"type": "text",
|
| 177 |
+
"text": "Neural volume rendering. Starting from the seminal work of neural radiance fields (NeRF) [6], neural volume rendering has gained much popularity in representing 3D scenes and synthesizing novel views. By integrating coordinate-based neural networks with volume rendering, NeRF performs high-fidelity view synthesis in a 3D consistent manner. Several attempts have been proposed to extend or improve NeRF. For instance, [14, 15, 16] further model illumination, and learn to disentangle reflectance with shading given well-aligned multi-view and multi-lighting images. Besides, many studies accelerate the rendering of static scenes from the perspective of spatial sparsity [8, 9], architectural design [10, 11], or efficient rendering [17, 12]. However, it is not trivial to apply these illumination and acceleration techniques to volume rendering-based generative models [5, 4], as they typically learn from unposed and unpaired images, and represent dynamic scenes that change with respect to the input latent codes. ",
|
| 178 |
+
"bbox": [
|
| 179 |
+
174,
|
| 180 |
+
284,
|
| 181 |
+
825,
|
| 182 |
+
435
|
| 183 |
+
],
|
| 184 |
+
"page_idx": 2
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"type": "text",
|
| 188 |
+
"text": "In this work, we take the first attempt to model illumination in volume rendering-based generative models, which serves as a regularization for accurate 3D shape learning. We further devise an efficient rendering technique for our approach, which shares similar insight with [12], but does not rely on ground truth depth for training and it is not limited to a small viewpoint range. ",
|
| 189 |
+
"bbox": [
|
| 190 |
+
174,
|
| 191 |
+
443,
|
| 192 |
+
825,
|
| 193 |
+
498
|
| 194 |
+
],
|
| 195 |
+
"page_idx": 2
|
| 196 |
+
},
|
| 197 |
+
{
|
| 198 |
+
"type": "text",
|
| 199 |
+
"text": "Generative 3D-aware image synthesis. Generative adversarial networks (GANs) [18] are capable of generating photorealistic images of high-resolution, but lack explicit control over camera viewpoint. In order to enable them to synthesis images in a 3D-aware manner, many recent approaches investigate how 3D representations could be incorporated into GANs [19, 20, 21, 22, 23, 24, 25, 26, 27, 5, 4, 28, 29, 30]. While some works directly learn from 3D data [19, 20, 21, 22, 30], in this work we focus on approaches that only have access to unconstrained 2D images, which is a more practical setting. Several attempts [23, 24, 25] adopt 3D voxel features with learned neural rendering. These methods produce realistic 3D-aware synthesis, but the 3D voxels are not interpretable, i.e., they cannot be transferred to 3D shapes. By leveraging differentiable renderer, [26] and [27] learn interpretable 3D voxels and meshes respectively, but [26] suffers from limited visual quality due to low voxel resolution while the learned 3D shapes of [27] exhibit noticeable distortions. The success of NeRF has motivated researchers to use radiance fields as the intermediate 3D representation in GANs [5, 4, 28]. While achieving impressive 3D-aware image synthesis with multi-view consistency, the extracted 3D shapes of these approaches are often imprecise and noisy. Our main goal in this work is to address the inaccurate shape by explicitly modeling illumination in the rendering process. This innovation helps achieve better 3D-aware image synthesis with broader applications. ",
|
| 200 |
+
"bbox": [
|
| 201 |
+
174,
|
| 202 |
+
505,
|
| 203 |
+
825,
|
| 204 |
+
724
|
| 205 |
+
],
|
| 206 |
+
"page_idx": 2
|
| 207 |
+
},
|
| 208 |
+
{
|
| 209 |
+
"type": "text",
|
| 210 |
+
"text": "Unsupervised 3D shape learning from 2D images. Our work is also related to unsupervised approaches that learn 3D object shapes from unconstrained, monocular view 2D images. While several approaches use external 3D shape templates or 2D key-points as weak supervisions to facilitate learning [31, 32, 33, 34, 35, 36, 37], in this work we consider the harder setting where only 2D images are available. To tackle this problem, most approaches adopt an “analysis-by-synthesis” paradigm [38, 39, 40]. Specifically, they design photo-geometric autoencoders to infer the 3D shape and viewpoint of each image with a reconstruction loss. While succeed in learning the 3D shapes for some object categories, these approaches typically rely on certain regularization to prevent trivial solutions, like the commonly used symmetry assumption on object shapes [39, 40, 31, 32]. Such assumption tends to produce symmetric results that may overlook the asymmetric aspects of objects. Recently, GAN2Shape [41] shows that it is possible to recover 3D shapes for images generated by 2D GANs. This method, however, requires inefficient instance-specific training, and recovers depth maps instead of full 3D representations. ",
|
| 211 |
+
"bbox": [
|
| 212 |
+
174,
|
| 213 |
+
731,
|
| 214 |
+
826,
|
| 215 |
+
911
|
| 216 |
+
],
|
| 217 |
+
"page_idx": 2
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"type": "image",
|
| 221 |
+
"img_path": "images/3df78f76629cf3932216f630e608875ef6f29d9065d720b47914c9ffadf01cbe.jpg",
|
| 222 |
+
"image_caption": [
|
| 223 |
+
"Figure 2: Method overview. Our generator $g _ { \\theta }$ models a relightable color field conditioned on a latent code $z \\sim p _ { z }$ . To synthesis an image, it performs volume rendering under a random camera pose $\\xi \\sim p _ { \\xi }$ . The rendering process also performs shading with a randomly sampled lighting condition $\\mu \\sim p _ { \\mu }$ . The discriminator learns to distinguish the synthesized images with real images from the training dataset, and the whole model is trained with a GAN loss. Although our model is trained from unconstrained 2D images, it allows explicit control over camera pose and lighting condition during inference. "
|
| 224 |
+
],
|
| 225 |
+
"image_footnote": [],
|
| 226 |
+
"bbox": [
|
| 227 |
+
197,
|
| 228 |
+
89,
|
| 229 |
+
795,
|
| 230 |
+
238
|
| 231 |
+
],
|
| 232 |
+
"page_idx": 3
|
| 233 |
+
},
|
| 234 |
+
{
|
| 235 |
+
"type": "text",
|
| 236 |
+
"text": "The proposed 3D-aware generative model also serves as a powerful approach for unsupervised 3D shape learning. Compared with aforementioned autoencoder-based methods, our GAN-based approach avoids the need to infer the viewpoint of each image, and does not rely on strong regularizations. In experiments, we demonstrate superior performance over recent state-of-the-art approaches Unsup3d [39] and GAN2Shape [41]. ",
|
| 237 |
+
"bbox": [
|
| 238 |
+
174,
|
| 239 |
+
356,
|
| 240 |
+
825,
|
| 241 |
+
426
|
| 242 |
+
],
|
| 243 |
+
"page_idx": 3
|
| 244 |
+
},
|
| 245 |
+
{
|
| 246 |
+
"type": "text",
|
| 247 |
+
"text": "3 Methodology ",
|
| 248 |
+
"text_level": 1,
|
| 249 |
+
"bbox": [
|
| 250 |
+
174,
|
| 251 |
+
445,
|
| 252 |
+
313,
|
| 253 |
+
463
|
| 254 |
+
],
|
| 255 |
+
"page_idx": 3
|
| 256 |
+
},
|
| 257 |
+
{
|
| 258 |
+
"type": "text",
|
| 259 |
+
"text": "We consider the problem of 3D-aware image synthesis by learning from a collection of unconstrained and unlabeled 2D images. We argue that modeling shading, i.e., the interaction of illumination and shape, in a generative implicit model enables unsupervised learning of more accurate 3D object shapes. In the following, we first provide some preliminaries on neural radiance fields (NeRF) [6], and then introduce our shading-guided generative implicit model. ",
|
| 260 |
+
"bbox": [
|
| 261 |
+
173,
|
| 262 |
+
476,
|
| 263 |
+
826,
|
| 264 |
+
546
|
| 265 |
+
],
|
| 266 |
+
"page_idx": 3
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"type": "text",
|
| 270 |
+
"text": "3.1 Preliminaries on Neural Radiance Fields ",
|
| 271 |
+
"text_level": 1,
|
| 272 |
+
"bbox": [
|
| 273 |
+
174,
|
| 274 |
+
561,
|
| 275 |
+
496,
|
| 276 |
+
577
|
| 277 |
+
],
|
| 278 |
+
"page_idx": 3
|
| 279 |
+
},
|
| 280 |
+
{
|
| 281 |
+
"type": "text",
|
| 282 |
+
"text": "As a deep implicit model, NeRF [6] uses an MLP network to represent a 3D scene as a radiance field. The MLP $f _ { \\theta } : ( x , d ) ( \\sigma , c )$ takes a 3D coordinate $\\pmb { x } \\in \\mathbb { R } ^ { 3 }$ and a viewing direction $\\ b { d } \\in \\mathbb { S } ^ { 2 }$ as inputs, and outputs a volume density $\\sigma \\in \\mathbb { R } ^ { + }$ and a color $c \\in \\mathbb { R } ^ { 3 }$ . To render an image under a given camera pose, each pixel color $C$ of the image is obtained via volume rendering along its corresponding camera ray $\\bar { \\boldsymbol { r } ( t ) } = \\boldsymbol { o } + t \\boldsymbol { d }$ with near and far bounds $t _ { n }$ and $t _ { f }$ as below: ",
|
| 283 |
+
"bbox": [
|
| 284 |
+
173,
|
| 285 |
+
587,
|
| 286 |
+
825,
|
| 287 |
+
657
|
| 288 |
+
],
|
| 289 |
+
"page_idx": 3
|
| 290 |
+
},
|
| 291 |
+
{
|
| 292 |
+
"type": "equation",
|
| 293 |
+
"img_path": "images/148bd61ff1ac563e5aa6995b31c07940bebff2bff2fe719d8d05ea30547993a6.jpg",
|
| 294 |
+
"text": "$$\nC ( r ) = \\int _ { t _ { n } } ^ { t _ { f } } T ( t ) \\sigma ( r ( t ) ) c ( r ( t ) , d ) d t , { \\mathrm { ~ w h e r e ~ } } T ( t ) = \\exp ( - \\int _ { t _ { n } } ^ { t } \\sigma ( r ( s ) ) d s ) .\n$$",
|
| 295 |
+
"text_format": "latex",
|
| 296 |
+
"bbox": [
|
| 297 |
+
240,
|
| 298 |
+
662,
|
| 299 |
+
758,
|
| 300 |
+
699
|
| 301 |
+
],
|
| 302 |
+
"page_idx": 3
|
| 303 |
+
},
|
| 304 |
+
{
|
| 305 |
+
"type": "text",
|
| 306 |
+
"text": "In practice, this volume rendering is implemented with a discretized form using stratified and hierarchical sampling. As this rendering process is differentiable, NeRF could be directly optimized via posed images of a static scene. After training, NeRF allows the rendering of images under new camera poses, achieving high-quality novel view synthesis. ",
|
| 307 |
+
"bbox": [
|
| 308 |
+
173,
|
| 309 |
+
703,
|
| 310 |
+
823,
|
| 311 |
+
760
|
| 312 |
+
],
|
| 313 |
+
"page_idx": 3
|
| 314 |
+
},
|
| 315 |
+
{
|
| 316 |
+
"type": "text",
|
| 317 |
+
"text": "3.2 Shading-Guided Generative Implicit Model ",
|
| 318 |
+
"text_level": 1,
|
| 319 |
+
"bbox": [
|
| 320 |
+
174,
|
| 321 |
+
773,
|
| 322 |
+
516,
|
| 323 |
+
790
|
| 324 |
+
],
|
| 325 |
+
"page_idx": 3
|
| 326 |
+
},
|
| 327 |
+
{
|
| 328 |
+
"type": "text",
|
| 329 |
+
"text": "In this work, we are interested in developing a generative implicit model that explicitly models the shading process for 3D-aware image synthesis. To achieve this, we make two extensions to the MLP network in NeRF. First, similar to most deep generative models, it is further conditioned on a latent code $_ z$ sampled from a prior distribution $\\bar { \\mathcal { N } ( 0 , I ) } ^ { d }$ . Second, instead of directly outputting the color $^ c$ , it outputs a relightable pre-cosine color term $\\pmb { a } \\in \\mathbb { R } ^ { 3 }$ , which is conceptually similar to albedo in the way that it could be shaded under a given lighting condition. While albedo is viewpoint-independent, in this work we do not strictly enforce such independence for $\\textbf { \\em a }$ in order to account for dataset bias. Thus, our generator $g _ { \\theta } : ( x , \\dot { d } , z ) ( \\sigma , \\pmb { a } )$ takes a coordinate $_ { \\textbf { \\em x } }$ , a viewing direction $^ d$ , and a latent code $_ z$ as inputs, and outputs a volume density $\\sigma$ and a pre-cosine color $\\textbf { \\em a }$ . Note that here $\\sigma$ is independent of $^ d$ , while the dependence of $\\textbf { \\em a }$ on $^ d$ is optional. To obtain the color $C$ of a camera ray $\\mathbf { \\boldsymbol { r } } ( t ) \\mathbf { \\dot { \\alpha } } = \\mathbf { \\boldsymbol { o } } + t \\mathbf { \\dot { \\alpha } }$ with near and far bounds $t _ { n }$ and $t _ { f }$ , we calculate the final pre-cosine color $\\pmb { A }$ via: ",
|
| 330 |
+
"bbox": [
|
| 331 |
+
173,
|
| 332 |
+
799,
|
| 333 |
+
826,
|
| 334 |
+
912
|
| 335 |
+
],
|
| 336 |
+
"page_idx": 3
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"type": "text",
|
| 340 |
+
"text": "",
|
| 341 |
+
"bbox": [
|
| 342 |
+
173,
|
| 343 |
+
90,
|
| 344 |
+
825,
|
| 345 |
+
133
|
| 346 |
+
],
|
| 347 |
+
"page_idx": 4
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"type": "equation",
|
| 351 |
+
"img_path": "images/de6bff1d9c03789957ea84d393a5b61da5c3ee0ba5b0147ae13d705a518556d9.jpg",
|
| 352 |
+
"text": "$$\nA ( r , z ) = \\int _ { t _ { n } } ^ { t _ { f } } T ( t , z ) \\sigma ( r ( t ) , z ) a ( r ( t ) , d , z ) d t , { \\mathrm { ~ w h e r e ~ } } T ( t , z ) = \\exp ( - \\int _ { t _ { n } } ^ { t } \\sigma ( r ( s ) , z ) d s ) .\n$$",
|
| 353 |
+
"text_format": "latex",
|
| 354 |
+
"bbox": [
|
| 355 |
+
187,
|
| 356 |
+
138,
|
| 357 |
+
805,
|
| 358 |
+
176
|
| 359 |
+
],
|
| 360 |
+
"page_idx": 4
|
| 361 |
+
},
|
| 362 |
+
{
|
| 363 |
+
"type": "text",
|
| 364 |
+
"text": "We also calculate the normal direction $\\textbf { \\em n }$ with: ",
|
| 365 |
+
"bbox": [
|
| 366 |
+
173,
|
| 367 |
+
195,
|
| 368 |
+
478,
|
| 369 |
+
210
|
| 370 |
+
],
|
| 371 |
+
"page_idx": 4
|
| 372 |
+
},
|
| 373 |
+
{
|
| 374 |
+
"type": "equation",
|
| 375 |
+
"img_path": "images/615d9d46e8362f868a5d8626fdcf2438e9103186244a39b3a6222e00a91c97a0.jpg",
|
| 376 |
+
"text": "$$\nn ( r , z ) = \\hat { n } ( r , z ) / \\| \\hat { n } ( r , z ) \\| _ { 2 } , \\mathrm { ~ w h e r e ~ } \\hat { n } ( r , z ) = - \\int _ { t _ { n } } ^ { t _ { f } } T ( t , z ) \\sigma ( r ( t ) , z ) \\nabla _ { r ( t ) } \\sigma ( r ( t ) , z ) d t ,\n$$",
|
| 377 |
+
"text_format": "latex",
|
| 378 |
+
"bbox": [
|
| 379 |
+
187,
|
| 380 |
+
213,
|
| 381 |
+
807,
|
| 382 |
+
250
|
| 383 |
+
],
|
| 384 |
+
"page_idx": 4
|
| 385 |
+
},
|
| 386 |
+
{
|
| 387 |
+
"type": "text",
|
| 388 |
+
"text": "where $\\nabla _ { \\pmb { r } ( t ) } \\sigma ( \\pmb { r } ( t ) , \\pmb { z } )$ is the derivative of volume density $\\sigma$ with respect to its input coordinate, which naturally captures the local normal direction, and could be calculated via back-propagation. Then the final color $C$ is obtained via Lambertian shading as: ",
|
| 389 |
+
"bbox": [
|
| 390 |
+
174,
|
| 391 |
+
270,
|
| 392 |
+
825,
|
| 393 |
+
311
|
| 394 |
+
],
|
| 395 |
+
"page_idx": 4
|
| 396 |
+
},
|
| 397 |
+
{
|
| 398 |
+
"type": "equation",
|
| 399 |
+
"img_path": "images/c0d401ea13e053667cafb357089e021e1fdc89a73111abf70af0525756ef2c7d.jpg",
|
| 400 |
+
"text": "$$\nC ( \\boldsymbol { r } , z ) = A ( r , z ) ( k _ { a } + k _ { d } \\mathrm { m a x } ( 0 , \\boldsymbol { l } \\cdot \\boldsymbol { n } ( r , z ) ) ) ,\n$$",
|
| 401 |
+
"text_format": "latex",
|
| 402 |
+
"bbox": [
|
| 403 |
+
336,
|
| 404 |
+
318,
|
| 405 |
+
661,
|
| 406 |
+
335
|
| 407 |
+
],
|
| 408 |
+
"page_idx": 4
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"type": "text",
|
| 412 |
+
"text": "where $\\ b { l } \\in \\mathbb S ^ { 2 }$ is the lighting direction, $k _ { a }$ and $k _ { d }$ are the ambient and diffuse coefficients. We provide more discussions on this shading formulation at the end of this subsection. ",
|
| 413 |
+
"bbox": [
|
| 414 |
+
173,
|
| 415 |
+
342,
|
| 416 |
+
821,
|
| 417 |
+
371
|
| 418 |
+
],
|
| 419 |
+
"page_idx": 4
|
| 420 |
+
},
|
| 421 |
+
{
|
| 422 |
+
"type": "text",
|
| 423 |
+
"text": "Camera and Lighting Sampling. Eq.(2 - 4) describe the process of rendering a pixel color given a camera ray $\\mathbf { } _ { \\pmb { r } ( t ) }$ and a lighting condition $\\mu = ( l , k _ { a } , k _ { d } )$ . Generating a full image $\\pmb { I _ { g } } \\in \\mathbb { R } ^ { 3 \\times H \\times W }$ requires one to sample a camera pose $\\boldsymbol { \\xi }$ and a lighting condition $\\pmb { \\mu }$ in addition to the latent code $_ z$ i.e., $I _ { g } = G _ { \\theta } ( z , \\xi , \\mu )$ . In our setting, the camera pose $\\boldsymbol { \\xi }$ could be described by pitch and yaw angles, and is sampled from a prior Gaussian or uniform distribution $p _ { \\xi }$ , as also done in previous works [4, 5]. Sampling the camera pose randomly during training would motivate the learned 3D scene to look realistic from different viewpoints. While this multi-view constraint is beneficial for learning a valid 3D representation, it is often insufficient to infer the accurate 3D object shape. Thus, in our approach, we further introduce a multi-lighting constraint by also randomly sampling a lighting condition $\\pmb { \\mu }$ from a prior distribution $p _ { \\mu }$ . In practice, $p _ { \\mu }$ could be estimated from the dataset using existing approaches like [39]. We also show in our experiments that a simple and manually tuned prior distribution could also produce reasonable results. As the shading process is sensitive to the normal direction due to the diffuse term $k _ { d } \\mathrm { m a x } ( 0 , l \\cdot n ( r , z ) )$ in Eq.(4), this multi-lighting constraint would regularize the model to learn more accurate 3D shapes that produce natural shading, as shown in Fig.1 (b). ",
|
| 424 |
+
"bbox": [
|
| 425 |
+
173,
|
| 426 |
+
376,
|
| 427 |
+
825,
|
| 428 |
+
584
|
| 429 |
+
],
|
| 430 |
+
"page_idx": 4
|
| 431 |
+
},
|
| 432 |
+
{
|
| 433 |
+
"type": "text",
|
| 434 |
+
"text": "Training. Our generative model follows the paradigm of GANs [18], where the generator is trained together with a discriminator $D$ with parameters $\\phi$ in an adversarial manner. During training, the generator generates fake images $I _ { g } = \\bar { G } _ { \\theta } ( z , \\xi , \\mu )$ by sampling the latent code $_ z$ , camera pose $\\boldsymbol { \\xi }$ and lighting condition $\\pmb { \\mu }$ from their corresponding prior distributions $p _ { z } , p _ { \\xi }$ , and $p _ { \\mu }$ . Let $\\pmb { I }$ denotes real images sampled from the data distribution $p _ { I }$ . We train our model with a non-saturating GAN loss with $R _ { 1 }$ regularization [42]: ",
|
| 435 |
+
"bbox": [
|
| 436 |
+
173,
|
| 437 |
+
588,
|
| 438 |
+
825,
|
| 439 |
+
674
|
| 440 |
+
],
|
| 441 |
+
"page_idx": 4
|
| 442 |
+
},
|
| 443 |
+
{
|
| 444 |
+
"type": "equation",
|
| 445 |
+
"img_path": "images/27664b7a031302ab4dcf83d6fe716a378f028c0dd84969c55934320bf1287cc3.jpg",
|
| 446 |
+
"text": "$$\n\\begin{array} { r } { \\mathcal { L } ( \\theta , \\phi ) = \\mathbf { E } _ { z \\sim p _ { z } , \\xi \\sim p _ { \\xi } , \\mu \\sim p _ { \\mu } } \\left[ f \\Big ( D _ { \\phi } ( G _ { \\theta } ( z , \\xi , \\mu ) ) \\Big ) \\right] + \\mathbf { E } _ { I \\sim p _ { \\mathcal { D } } } \\left[ f ( - D _ { \\phi } ( I ) ) + \\lambda \\| \\nabla D _ { \\phi } ( I ) \\| ^ { 2 } \\right] , } \\end{array}\n$$",
|
| 447 |
+
"text_format": "latex",
|
| 448 |
+
"bbox": [
|
| 449 |
+
181,
|
| 450 |
+
678,
|
| 451 |
+
812,
|
| 452 |
+
705
|
| 453 |
+
],
|
| 454 |
+
"page_idx": 4
|
| 455 |
+
},
|
| 456 |
+
{
|
| 457 |
+
"type": "text",
|
| 458 |
+
"text": "where $f ( u ) = - \\log ( 1 + \\exp ( - u ) )$ , and $\\lambda$ controls the strength of regularization. More implementation details are provided in the supplementary material. ",
|
| 459 |
+
"bbox": [
|
| 460 |
+
173,
|
| 461 |
+
717,
|
| 462 |
+
825,
|
| 463 |
+
744
|
| 464 |
+
],
|
| 465 |
+
"page_idx": 4
|
| 466 |
+
},
|
| 467 |
+
{
|
| 468 |
+
"type": "text",
|
| 469 |
+
"text": "Discussion. Note that in $\\operatorname { E q . } ( 2 \\ \\cdot \\ 4 )$ , we perform shading after $\\pmb { A }$ and $\\textbf { \\em n }$ are obtained via volume rendering. An alternative way is to perform shading at each local spatial point as $\\begin{array} { l l l } { c ( { \\boldsymbol { r } } ( t ) , d , z ) } & { = } & { { \\boldsymbol { a } } ( { \\boldsymbol { r } } ( t ) , d , z ) ( k _ { a } + { k _ { d } } { \\operatorname* { m a x } } ( 0 , { l } \\cdot { \\boldsymbol { n } } ( { \\boldsymbol { r } } ( t ) , z ) ) ) } \\end{array}$ , where $n ( r ( t ) , { \\bar { z } } ) =$ $- \\nabla _ { \\pmb { r } ( t ) } \\sigma ( \\pmb { r } ( t ) , z ) / \\| \\nabla _ { \\pmb { r } ( t ) } \\sigma ( \\pmb { r } ( t ) , z ) \\| _ { 2 }$ is the local normal. Then we could perform volume rendering using $\\mathbf { } c ( \\pmb { r } ( t ) , z )$ to get the final pixel color. In practice, we observe that this formulation obtains suboptimal results. An intuitive reason is that in this formulation, the normal direction is normalized at each local point, neglecting the magnitude of $\\nabla _ { \\pmb { r } ( t ) } \\sigma ( \\pmb { r } ( t ) , \\pmb { z } )$ , which tends to be larger near the object surfaces. We provide more analysis in experiments and the supplementary material. ",
|
| 470 |
+
"bbox": [
|
| 471 |
+
173,
|
| 472 |
+
750,
|
| 473 |
+
825,
|
| 474 |
+
864
|
| 475 |
+
],
|
| 476 |
+
"page_idx": 4
|
| 477 |
+
},
|
| 478 |
+
{
|
| 479 |
+
"type": "text",
|
| 480 |
+
"text": "The Lambertian shading we used is an approximation to the real illumination scenario. While serving as a good regularization for improving the learned 3D shape, it could possibly introduce an additional gap between the distribution of generated images and that of real images. To compensate for such risk, we could optionally let the predicted $\\textbf { \\em a }$ be conditioned on the lighting condition, i.e., $\\pmb { a } = \\pmb { a } ( \\pmb { r } ( t ) , \\pmb { d } , \\pmb { \\mu } , z )$ . Thus, in cases where the lighting condition deviates from the real data distribution, the generator could learn to adjust the value of $\\textbf { \\em a }$ and reduce the aforementioned gap. We show the benefit of this design in the experiments. ",
|
| 481 |
+
"bbox": [
|
| 482 |
+
174,
|
| 483 |
+
869,
|
| 484 |
+
825,
|
| 485 |
+
911
|
| 486 |
+
],
|
| 487 |
+
"page_idx": 4
|
| 488 |
+
},
|
| 489 |
+
{
|
| 490 |
+
"type": "image",
|
| 491 |
+
"img_path": "images/13cea0618dccbea116e983843c0cd9f62bd1c06784ebcbfc985b927fbfddc5a6.jpg",
|
| 492 |
+
"image_caption": [
|
| 493 |
+
"Figure 3: (a) Conventional volume rendering samples dozens of points within a predefined near and far bounds $t _ { n }$ and $t _ { f }$ . (b) We propose an efficient volume rendering technique via surface tracking. Before rendering, our surface tracking network $S _ { \\psi }$ predicts an initial guess of the surface position $s$ conditioned on the latent code $_ z$ and camera pose $\\xi$ . Then we sample points near $s$ , which requires fewer samples. Finally, we use the volume rendered depth $d$ as the ground truth to train $S _ { \\psi }$ . During training, $S _ { \\psi }$ is able to predict depth $s$ that well approximates the real surface depth $d$ . "
|
| 494 |
+
],
|
| 495 |
+
"image_footnote": [],
|
| 496 |
+
"bbox": [
|
| 497 |
+
212,
|
| 498 |
+
88,
|
| 499 |
+
782,
|
| 500 |
+
238
|
| 501 |
+
],
|
| 502 |
+
"page_idx": 5
|
| 503 |
+
},
|
| 504 |
+
{
|
| 505 |
+
"type": "text",
|
| 506 |
+
"text": "",
|
| 507 |
+
"bbox": [
|
| 508 |
+
173,
|
| 509 |
+
340,
|
| 510 |
+
826,
|
| 511 |
+
397
|
| 512 |
+
],
|
| 513 |
+
"page_idx": 5
|
| 514 |
+
},
|
| 515 |
+
{
|
| 516 |
+
"type": "text",
|
| 517 |
+
"text": "3.3 Efficient Volume Rendering via Surface Tracking ",
|
| 518 |
+
"text_level": 1,
|
| 519 |
+
"bbox": [
|
| 520 |
+
173,
|
| 521 |
+
412,
|
| 522 |
+
557,
|
| 523 |
+
428
|
| 524 |
+
],
|
| 525 |
+
"page_idx": 5
|
| 526 |
+
},
|
| 527 |
+
{
|
| 528 |
+
"type": "text",
|
| 529 |
+
"text": "Similar to NeRF, we implement volume rendering with a discretized integral, which typically requires to sample dozens of points along a camera ray, as shown in Fig. 3 (a). In our approach, we also need to perform back-propagation across the generator in Eq.(3) to get the normal direction for each point, which introduces additional computational cost. To achieve more efficient volume rendering, a natural idea is to exploit spatial sparsity. Usually, the weight $T ( t , z ) \\sigma ( \\pmb { r } ( t ) , z )$ in volume rendering would concentrate on the object surface position during training. Thus, if we know the rough surface position before rendering, we could sample points near the surface to save computation. While for a static scene it is possible to store such spatial sparsity in a sparse voxel grid [8, 9], this technique cannot be directly applied to our generative model, as the 3D scene keeps changing with respect to the input latent code. ",
|
| 530 |
+
"bbox": [
|
| 531 |
+
173,
|
| 532 |
+
436,
|
| 533 |
+
825,
|
| 534 |
+
577
|
| 535 |
+
],
|
| 536 |
+
"page_idx": 5
|
| 537 |
+
},
|
| 538 |
+
{
|
| 539 |
+
"type": "text",
|
| 540 |
+
"text": "To achieve more efficient volume rendering in our generative implicit model, we further propose a surface tracking network $S$ that learns to mimic the surface position conditioned on the latent code. In particular, the volume rendering naturally allows the depth estimation of the object surface via: ",
|
| 541 |
+
"bbox": [
|
| 542 |
+
174,
|
| 543 |
+
582,
|
| 544 |
+
825,
|
| 545 |
+
625
|
| 546 |
+
],
|
| 547 |
+
"page_idx": 5
|
| 548 |
+
},
|
| 549 |
+
{
|
| 550 |
+
"type": "equation",
|
| 551 |
+
"img_path": "images/cb470f938725f63638e16b6d1bdae1ace599b2e6bb81268017227a1323023ef6.jpg",
|
| 552 |
+
"text": "$$\nt _ { s } ( \\pmb { r } , z ) = \\int _ { t _ { n } } ^ { t _ { f } } T ( t , z ) \\sigma ( \\pmb { r } ( t ) , z ) t d t ,\n$$",
|
| 553 |
+
"text_format": "latex",
|
| 554 |
+
"bbox": [
|
| 555 |
+
370,
|
| 556 |
+
628,
|
| 557 |
+
625,
|
| 558 |
+
666
|
| 559 |
+
],
|
| 560 |
+
"page_idx": 5
|
| 561 |
+
},
|
| 562 |
+
{
|
| 563 |
+
"type": "text",
|
| 564 |
+
"text": "where $T ( t , z )$ is defined the same way as in Eq.(2). Thus, given a camera pose $\\boldsymbol { \\xi }$ and a latent code $_ z$ , we could render the full depth map $\\pmb { t } _ { s } ( z , \\pmb { \\xi } )$ . As shown in Fig. 3 (b), we mimic $\\pmb { t } _ { s } ( z , \\pmb { \\xi } )$ with the surface tracking network $S _ { \\psi }$ , which is a light-weighted convolutional neural network that takes $z , \\xi$ as inputs and outputs a depth map. The depth mimic loss is: ",
|
| 565 |
+
"bbox": [
|
| 566 |
+
173,
|
| 567 |
+
670,
|
| 568 |
+
823,
|
| 569 |
+
726
|
| 570 |
+
],
|
| 571 |
+
"page_idx": 5
|
| 572 |
+
},
|
| 573 |
+
{
|
| 574 |
+
"type": "equation",
|
| 575 |
+
"img_path": "images/460066a996920a66f0ff732ae92c88c7c38df63162666935a9ec66ced95bbe8e.jpg",
|
| 576 |
+
"text": "$$\n\\mathcal { L } ( \\psi ) = \\mathbf { E } _ { z \\sim p _ { z } , \\xi \\sim p _ { \\xi } } \\left[ \\| S _ { \\psi } ( z , \\xi ) - t _ { s } ( z , \\xi ) \\| _ { 1 } + \\operatorname* { P r e c } ( S _ { \\psi } ( z , \\xi ) , t _ { s } ( d ( z , \\xi ) ) \\right] ,\n$$",
|
| 577 |
+
"text_format": "latex",
|
| 578 |
+
"bbox": [
|
| 579 |
+
245,
|
| 580 |
+
731,
|
| 581 |
+
751,
|
| 582 |
+
748
|
| 583 |
+
],
|
| 584 |
+
"page_idx": 5
|
| 585 |
+
},
|
| 586 |
+
{
|
| 587 |
+
"type": "text",
|
| 588 |
+
"text": "where Prec is the perceptual loss that motivates $S _ { \\psi }$ to better capture edges of the surface. ",
|
| 589 |
+
"bbox": [
|
| 590 |
+
181,
|
| 591 |
+
752,
|
| 592 |
+
754,
|
| 593 |
+
767
|
| 594 |
+
],
|
| 595 |
+
"page_idx": 5
|
| 596 |
+
},
|
| 597 |
+
{
|
| 598 |
+
"type": "text",
|
| 599 |
+
"text": "During training, $S _ { \\psi }$ is optimized jointly with the generator and the discriminator. Thus, each time after we sample a latent code $_ z$ and a camera pose $\\boldsymbol { \\xi }$ , we can get an initial guess of the depth map as $S _ { \\psi } ( z , \\pmb { \\xi } )$ . Then for a pixel with predicted depth $s$ , we could perform volume rendering in Eq.(2,3,6) with near bound $t _ { n } = s - \\Delta _ { i } / 2$ and far bound $t _ { f } = s + \\Delta _ { i } / 2$ , where $\\Delta _ { i }$ is the interval for volume rendering that decreases as the training iteration $i$ grows. Specifically, we start with a large interval $\\Delta _ { m a x }$ and decrease to $\\Delta _ { m i n }$ with an exponential schedule. As $\\Delta _ { i }$ decreases, the number of points used for rendering $m$ also decreases accordingly. Note that the computational cost of our efficient surface tracking network is marginal compared to the generator, as the former only needs a single forward pass to render an image while the latter will be queried for $H \\times W \\times m$ times. Thus, the reduction of $m$ would significantly accelerate the training and inference speed for ShadeGAN. ",
|
| 600 |
+
"bbox": [
|
| 601 |
+
173,
|
| 602 |
+
772,
|
| 603 |
+
825,
|
| 604 |
+
911
|
| 605 |
+
],
|
| 606 |
+
"page_idx": 5
|
| 607 |
+
},
|
| 608 |
+
{
|
| 609 |
+
"type": "image",
|
| 610 |
+
"img_path": "images/9decfdb73ca7cefeb28379532c0c155364b0e9f7ae537e8e5bafbf3a02fc1307.jpg",
|
| 611 |
+
"image_caption": [
|
| 612 |
+
"Figure 4: Qualitative comparison on BFM (top), CelebA (middle), and Cats (bottom) datasets. \"Albedo\" refers to the pre-cosine color that approximates albedo. Our approach synthesizes more accurate 3D shapes than pi-GAN and GRAF, and also learns to disentangle shading with albedo. "
|
| 613 |
+
],
|
| 614 |
+
"image_footnote": [],
|
| 615 |
+
"bbox": [
|
| 616 |
+
173,
|
| 617 |
+
83,
|
| 618 |
+
823,
|
| 619 |
+
592
|
| 620 |
+
],
|
| 621 |
+
"page_idx": 6
|
| 622 |
+
},
|
| 623 |
+
{
|
| 624 |
+
"type": "text",
|
| 625 |
+
"text": "4 Experiments ",
|
| 626 |
+
"text_level": 1,
|
| 627 |
+
"bbox": [
|
| 628 |
+
174,
|
| 629 |
+
676,
|
| 630 |
+
312,
|
| 631 |
+
694
|
| 632 |
+
],
|
| 633 |
+
"page_idx": 6
|
| 634 |
+
},
|
| 635 |
+
{
|
| 636 |
+
"type": "text",
|
| 637 |
+
"text": "In this section, we evaluate the proposed ShadeGAN on 3D-aware image synthesis. We also show that ShadeGAN learns much more accurate 3D shapes than previous methods, and in the meantime allows explicit control over lighting conditions. The datasets used include CelebA [43], BFM [13], and Cats [44], all of which contain only unconstrained 2D RGB images. ",
|
| 638 |
+
"bbox": [
|
| 639 |
+
174,
|
| 640 |
+
710,
|
| 641 |
+
825,
|
| 642 |
+
766
|
| 643 |
+
],
|
| 644 |
+
"page_idx": 6
|
| 645 |
+
},
|
| 646 |
+
{
|
| 647 |
+
"type": "text",
|
| 648 |
+
"text": "Implementation. In terms of model architectures, we adopt a SIREN-based MLP [45] as the generator and a convolutional neural network as the discriminator following [4]. For the prior distribution of lighting conditions, we use Unsup3d [39] to estimate the lighting conditions of real data and subsequently fit a multivariate Gaussian distribution of $\\pmb { \\mu } = ( l , k _ { a } , k _ { d } )$ as the prior. A hand-crafted prior distribution is also included in the ablation study. In quantitative study, we let the pre-cosine color $\\textbf { \\em a }$ be conditioned on the lighting condition $\\pmb { \\mu }$ as well as the viewing direction $^ d$ unless otherwise stated. In qualitative study, we observe that removing view conditioning achieves slightly better 3D shapes for CelebA and BFM datasets. Thus, we show results without view conditioning for these two datasets in the main paper, and put those with view conditioning in Fig. 4 of the supplementary material. Other implementation details are also provided in the supplementary. ",
|
| 649 |
+
"bbox": [
|
| 650 |
+
174,
|
| 651 |
+
772,
|
| 652 |
+
825,
|
| 653 |
+
911
|
| 654 |
+
],
|
| 655 |
+
"page_idx": 6
|
| 656 |
+
},
|
| 657 |
+
{
|
| 658 |
+
"type": "image",
|
| 659 |
+
"img_path": "images/ff1406392fc76f588c855654ee978051ca4f72b859038a7281c035f357932a49.jpg",
|
| 660 |
+
"image_caption": [
|
| 661 |
+
"Figure 5: Generated face images and their 3D meshes. "
|
| 662 |
+
],
|
| 663 |
+
"image_footnote": [],
|
| 664 |
+
"bbox": [
|
| 665 |
+
176,
|
| 666 |
+
94,
|
| 667 |
+
588,
|
| 668 |
+
256
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 7
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "image",
|
| 674 |
+
"img_path": "images/fe3b5050f89ebed1134c6077e6c683709d44f588ffd089dcfa626190fe2e58b0.jpg",
|
| 675 |
+
"image_caption": [
|
| 676 |
+
"Figure 6: Qualitative ablation. See the main text for discussions. "
|
| 677 |
+
],
|
| 678 |
+
"image_footnote": [],
|
| 679 |
+
"bbox": [
|
| 680 |
+
598,
|
| 681 |
+
89,
|
| 682 |
+
818,
|
| 683 |
+
253
|
| 684 |
+
],
|
| 685 |
+
"page_idx": 7
|
| 686 |
+
},
|
| 687 |
+
{
|
| 688 |
+
"type": "table",
|
| 689 |
+
"img_path": "images/37c5b438f30e7a5db08d4781c2696b4cf7631abc8f204bc065f66b8226b6d9c0.jpg",
|
| 690 |
+
"table_caption": [
|
| 691 |
+
"Table 1: Comparisons on the BFM dataset. We report FID $\\mathsf { \\bar { \\rho } _ { 1 2 8 } } 2 \\mathsf { \\bar { \\rho } _ { 1 } }$ for image synthesis, and SIDE $\\bar { ( } \\times 1 0 ^ { - 2 } )$ and MAD (deg.) for the accuracy of 3D shapes. ‘-’ indicates not available. Results of pi-GAN and Ours are averaged over 5 runs. "
|
| 692 |
+
],
|
| 693 |
+
"table_footnote": [],
|
| 694 |
+
"table_body": "<table><tr><td>Method</td><td>FID↓</td><td>SIDE↓</td><td>MAD↓</td></tr><tr><td>Supervised</td><td>-</td><td>0.410</td><td>10.78</td></tr><tr><td>Unsup3d [39]</td><td>-</td><td>0.793</td><td>16.51</td></tr><tr><td>GAN2Shape [41]</td><td>-</td><td>0.756</td><td>14.81</td></tr><tr><td>GRAF[5]</td><td>53.4</td><td>1.857</td><td>26.60</td></tr><tr><td>pi-GAN [4]</td><td>16.7±0.2</td><td>0.727±0.012</td><td>20.09±0.23</td></tr><tr><td>Ours</td><td>17.7±0.2</td><td>0.607±0.007</td><td>14.52±0.11</td></tr></table>",
|
| 695 |
+
"bbox": [
|
| 696 |
+
176,
|
| 697 |
+
361,
|
| 698 |
+
485,
|
| 699 |
+
458
|
| 700 |
+
],
|
| 701 |
+
"page_idx": 7
|
| 702 |
+
},
|
| 703 |
+
{
|
| 704 |
+
"type": "table",
|
| 705 |
+
"img_path": "images/62a8e2501c5b1d3ca0121a7e5620daecb3b63f685adf2cb601b106f8a17ef698.jpg",
|
| 706 |
+
"table_caption": [
|
| 707 |
+
"Table 2: Comparisons on the CelebA and Cats datasets. The image resolution is $1 2 8 ^ { 2 }$ . "
|
| 708 |
+
],
|
| 709 |
+
"table_footnote": [],
|
| 710 |
+
"table_body": "<table><tr><td>Dataset</td><td>Method FID↓</td><td>MAD↓</td></tr><tr><td rowspan=\"3\">CelebA</td><td>GRAF</td><td>43.0 30.48</td></tr><tr><td>pi-GAN 15.7</td><td>27.22</td></tr><tr><td>Ours 16.2</td><td>20.49</td></tr><tr><td rowspan=\"3\">Cats</td><td>GRAF</td><td>30.3 65.47</td></tr><tr><td>pi-GAN</td><td>10.7 33.48</td></tr><tr><td>Ours 10.3</td><td>25.47</td></tr></table>",
|
| 711 |
+
"bbox": [
|
| 712 |
+
539,
|
| 713 |
+
340,
|
| 714 |
+
785,
|
| 715 |
+
438
|
| 716 |
+
],
|
| 717 |
+
"page_idx": 7
|
| 718 |
+
},
|
| 719 |
+
{
|
| 720 |
+
"type": "text",
|
| 721 |
+
"text": "Comparison with baselines. We compare ShadeGAN with two state-of-the-art generative implicit models, namely GRAF [5] and pi-GAN [4]. Specifically, Fig. 4 includes both synthesized images as well as their corresponding 3D meshes, which are obtained by performing marching cubes on the volume density $\\sigma$ . While GRAF and pi-GAN could synthesize images with controllable poses, their learned 3D shapes are inaccurate and noisy. In contrast, our approach not only synthesizes photorealistic 3D-consistent images, but also learns much more accurate 3D shapes and surface normals, indicating the effectiveness of the proposed multi-lighting constraint as a regularization. More synthesized images and their corresponding shapes are included in Fig.5. Besides more accurate 3D shapes, ShadeGAN can also learn the albedo and diffuse shading components inherently. As shown in Fig. 4, although not perfect, ShadeGAN has managed to disentangle shading and albedo with satisfying quality, as such disentanglement is a natural solution to the multi-lighting constraint. ",
|
| 722 |
+
"bbox": [
|
| 723 |
+
174,
|
| 724 |
+
483,
|
| 725 |
+
825,
|
| 726 |
+
636
|
| 727 |
+
],
|
| 728 |
+
"page_idx": 7
|
| 729 |
+
},
|
| 730 |
+
{
|
| 731 |
+
"type": "text",
|
| 732 |
+
"text": "The quality of learned 3D shapes is quantitatively evaluated on the BFM dataset. Specifically, we use each of the generative implicit models to generate 50k images and their corresponding depth maps. Image-depth pairs from each model are used as training data to train an additional convolutional neural network (CNN) that learns to predict the depth map of an input image. We then test each trained CNN on the BFM test set and compare its predictions to the ground-truth depth maps as a measurement of the quality of learned 3D shapes. Following [39], we report the scale-invariant depth error (SIDE) and mean angle deviation (MAD) metrics. The results are included in Tab. 1, where ShadeGAN significantly outperforms GRAF and pi-GAN. Besides, ShadeGAN also outperforms other advanced unsupervised 3D shape learning approaches including Unsup3d [39] and GAN2Shape [41], demonstrating its large potential in unsupervised 3D shapes learning. In terms of image quality, Tab. 1 includes the FID [46] scores of images synthesized by different models, where the FID score of ShadeGAN is slightly inferior to pi-GAN in BFM and CelebA. Intuitively, this is caused by the gap between our approximated shading (i.e. Lambertian shading) and the real illumination, which can be potentially avoided by adopting more realistic shading models and improving the lighting prior. ",
|
| 733 |
+
"bbox": [
|
| 734 |
+
174,
|
| 735 |
+
642,
|
| 736 |
+
825,
|
| 737 |
+
849
|
| 738 |
+
],
|
| 739 |
+
"page_idx": 7
|
| 740 |
+
},
|
| 741 |
+
{
|
| 742 |
+
"type": "text",
|
| 743 |
+
"text": "In Tab. 2, we also show the quantitative results of different models on CelebA and Cats. To evaluate the learned shape, we use each generative implicit model to generate 2k front-view images and their corresponding depth maps. While these datasets do not have ground truth depth, we report MAD obtained by testing pretrained Unsup3d models [39] on these generated image-depth pairs as a reference. As we can observe, results on CelebA and Cat are consistent with those on the BFM dataset. ",
|
| 744 |
+
"bbox": [
|
| 745 |
+
174,
|
| 746 |
+
856,
|
| 747 |
+
823,
|
| 748 |
+
911
|
| 749 |
+
],
|
| 750 |
+
"page_idx": 7
|
| 751 |
+
},
|
| 752 |
+
{
|
| 753 |
+
"type": "table",
|
| 754 |
+
"img_path": "images/ba899c687321409fe866c3a87146b4f6964363b9371761ec3866b32c54db6765.jpg",
|
| 755 |
+
"table_caption": [
|
| 756 |
+
"Table 3: Ablation study on the BFM dataset. "
|
| 757 |
+
],
|
| 758 |
+
"table_footnote": [],
|
| 759 |
+
"table_body": "<table><tr><td>No.</td><td>Method</td><td>FID↓</td><td>SIDE↓</td><td>MAD</td></tr><tr><td>(1)</td><td>ShadeGAN</td><td>17.7</td><td>0.607</td><td>14.52</td></tr><tr><td>(2)</td><td>local shading</td><td>30.1</td><td>0.754</td><td>18.18</td></tr><tr><td>(3)</td><td>w/o light</td><td>19.2</td><td>0.618</td><td>14.53</td></tr><tr><td>(4)</td><td>w/o view</td><td>18.6</td><td>0.622</td><td>14.88</td></tr><tr><td>(5)</td><td>manual prior</td><td>20.2</td><td>0.643</td><td>15.38</td></tr><tr><td>(6)</td><td>+efficient</td><td>18.2</td><td>0.673</td><td>14.72</td></tr></table>",
|
| 760 |
+
"bbox": [
|
| 761 |
+
181,
|
| 762 |
+
104,
|
| 763 |
+
480,
|
| 764 |
+
193
|
| 765 |
+
],
|
| 766 |
+
"page_idx": 8
|
| 767 |
+
},
|
| 768 |
+
{
|
| 769 |
+
"type": "table",
|
| 770 |
+
"img_path": "images/e97def077d4f7816a730e4c59ce0bee44747549c6c4529baf7b9887d53f4f6cd.jpg",
|
| 771 |
+
"table_caption": [
|
| 772 |
+
"Table 4: Training and inference time cost on CelebA. The efficient volume rendering significantly improves training and inference speed. "
|
| 773 |
+
],
|
| 774 |
+
"table_footnote": [],
|
| 775 |
+
"table_body": "<table><tr><td>Method</td><td>Train (h) Inference (s) FID</td></tr><tr><td>ShadeGAN</td><td>92.3 0.343</td></tr><tr><td>70.2</td><td>16.4 0.179 16.2</td></tr><tr><td>+efficient pi-GAN</td><td></td></tr><tr><td>56.8</td><td>0.204 15.7</td></tr><tr><td>+efficient 46.9</td><td>0.114 15.9</td></tr></table>",
|
| 776 |
+
"bbox": [
|
| 777 |
+
539,
|
| 778 |
+
131,
|
| 779 |
+
785,
|
| 780 |
+
194
|
| 781 |
+
],
|
| 782 |
+
"page_idx": 8
|
| 783 |
+
},
|
| 784 |
+
{
|
| 785 |
+
"type": "image",
|
| 786 |
+
"img_path": "images/605119fad261eeda2f3f6171d8b0a69a2bba7d38bc9125de9daaf68b5c4a72a6.jpg",
|
| 787 |
+
"image_caption": [
|
| 788 |
+
"Figure 7: Visualization of depths predicted by our depth tracking network and those calculated via volume rendering. "
|
| 789 |
+
],
|
| 790 |
+
"image_footnote": [],
|
| 791 |
+
"bbox": [
|
| 792 |
+
178,
|
| 793 |
+
210,
|
| 794 |
+
375,
|
| 795 |
+
344
|
| 796 |
+
],
|
| 797 |
+
"page_idx": 8
|
| 798 |
+
},
|
| 799 |
+
{
|
| 800 |
+
"type": "image",
|
| 801 |
+
"img_path": "images/e5eaac7f139b81812227a4fdbdb6b7b552f02368d49dbe54d2a17a3ae796f554.jpg",
|
| 802 |
+
"image_caption": [
|
| 803 |
+
"Figure 8: Illumination-aware image synthesis. ShadeGAN allows explicit control over the lighting. The pre-cosine color (albedo) is independent of lighting in (a) and is conditioned on lighting in (b). We show results of adding a specular term in (c). "
|
| 804 |
+
],
|
| 805 |
+
"image_footnote": [],
|
| 806 |
+
"bbox": [
|
| 807 |
+
397,
|
| 808 |
+
209,
|
| 809 |
+
810,
|
| 810 |
+
349
|
| 811 |
+
],
|
| 812 |
+
"page_idx": 8
|
| 813 |
+
},
|
| 814 |
+
{
|
| 815 |
+
"type": "text",
|
| 816 |
+
"text": "",
|
| 817 |
+
"bbox": [
|
| 818 |
+
174,
|
| 819 |
+
428,
|
| 820 |
+
823,
|
| 821 |
+
457
|
| 822 |
+
],
|
| 823 |
+
"page_idx": 8
|
| 824 |
+
},
|
| 825 |
+
{
|
| 826 |
+
"type": "text",
|
| 827 |
+
"text": "Ablation studies. We further study the effects of several design choices in ShadeGAN. First, we perform local points-specific shading as mentioned in the discussion of Sec. 3.2. As Tab. 3 No.(2) and Fig. 6 (b) show, the results of such a local shading strategy are notably worse than the original one, which indicates that taking the magnitude of $\\nabla _ { \\pmb { x } } \\sigma$ into account is beneficial. Besides, the results of Tab. 3 No.(3) and No.(4) imply that removing $\\textbf { \\em a }$ ’s dependence on the lighting $\\pmb { \\mu }$ or the viewpoint $^ d$ could lead to a slight performance drop. The results of using a simple manually tuned lighting prior are provided in Tab. 3 No.(5) and Fig. 6 (c), which are only moderately worse than the results of using a data-driven prior, and the generated shapes are still significantly better than the ones produced by existing approaches. ",
|
| 828 |
+
"bbox": [
|
| 829 |
+
174,
|
| 830 |
+
462,
|
| 831 |
+
825,
|
| 832 |
+
588
|
| 833 |
+
],
|
| 834 |
+
"page_idx": 8
|
| 835 |
+
},
|
| 836 |
+
{
|
| 837 |
+
"type": "text",
|
| 838 |
+
"text": "To verify the effectiveness of the proposed efficient volume rendering technique, we include its effects on image quality and training/inference time in Tab. 3 No.(6) and Tab. 4. It is observed that the efficient volume rendering has marginal effects on the performance, but significantly reduces the training and inference time by $24 \\%$ and $48 \\%$ for ShadeGAN. Moreover, in Fig. 7 we visualize the depth maps predicted by our surface tracking network and those obtained via volume rendering. It is shown that under varying identities and camera poses, the surface tracking network could consistently predict depth values that are quite close to the real surface positions, so that we can sample points near the predicted surface for rendering without sacrificing image quality. ",
|
| 839 |
+
"bbox": [
|
| 840 |
+
173,
|
| 841 |
+
593,
|
| 842 |
+
825,
|
| 843 |
+
705
|
| 844 |
+
],
|
| 845 |
+
"page_idx": 8
|
| 846 |
+
},
|
| 847 |
+
{
|
| 848 |
+
"type": "text",
|
| 849 |
+
"text": "Illumination-aware image synthesis. As ShadeGAN models the shading process, it by design allows explicit control over the lighting condition. We provide such illumination-aware image synthesis results in Fig.8, where ShadeGAN generates promising images under different lighting directions. We also show that in cases where the predicted $\\textbf { \\em a }$ is conditioned on the lighting condition $\\pmb { \\mu }$ , $\\textbf { \\em a }$ would slightly change $w . r . t .$ the lighting condition, e.g., it would be brighter in areas having a overly dim shading in order to make the final image more natural. Besides, we could optionally add a specular term $k _ { s } \\operatorname* { m a x } ( 0 , \\boldsymbol { h } \\cdot \\boldsymbol { n } ) ^ { p }$ in Eq. 4 (i.e., Blinn-Phong shading [47], where $^ { h }$ is the bisector of the angle between the viewpoint and the lighting direction) to create specular highlight effects, as shown in Fig.8 (c). ",
|
| 850 |
+
"bbox": [
|
| 851 |
+
173,
|
| 852 |
+
710,
|
| 853 |
+
825,
|
| 854 |
+
835
|
| 855 |
+
],
|
| 856 |
+
"page_idx": 8
|
| 857 |
+
},
|
| 858 |
+
{
|
| 859 |
+
"type": "text",
|
| 860 |
+
"text": "GAN inversion. ShadeGAN could also be used to reconstruct a given target image by performing GAN inversion. As shown in Fig. 9 such inversion allows us to obtain several factors of the image, including the 3D shape, surface normal, approximated albedo, and shading. Besides, we can further perform view synthesis and relighting by changing the viewpoint and lighting condition. The implementation of GAN inversion is provided in the supplementary material. ",
|
| 861 |
+
"bbox": [
|
| 862 |
+
174,
|
| 863 |
+
842,
|
| 864 |
+
825,
|
| 865 |
+
911
|
| 866 |
+
],
|
| 867 |
+
"page_idx": 8
|
| 868 |
+
},
|
| 869 |
+
{
|
| 870 |
+
"type": "image",
|
| 871 |
+
"img_path": "images/eaffb2b5570a81485756fcf7580d30cecd842145f036815f3eddac6ce95a8091.jpg",
|
| 872 |
+
"image_caption": [
|
| 873 |
+
"Figure 9: GAN inversion for real image editing. "
|
| 874 |
+
],
|
| 875 |
+
"image_footnote": [],
|
| 876 |
+
"bbox": [
|
| 877 |
+
176,
|
| 878 |
+
88,
|
| 879 |
+
823,
|
| 880 |
+
150
|
| 881 |
+
],
|
| 882 |
+
"page_idx": 9
|
| 883 |
+
},
|
| 884 |
+
{
|
| 885 |
+
"type": "text",
|
| 886 |
+
"text": "Discussions. As the Lambertian shading we used is an approximation to the real illumination, the albedo learned by ShadeGAN is not perfectly disentangled. Our approach does not consider the spatially-varying material properties of objects as well. In the future, we intend to incorporate more sophisticated shading models to learn better disentangled generative reflectance fields. ",
|
| 887 |
+
"bbox": [
|
| 888 |
+
173,
|
| 889 |
+
167,
|
| 890 |
+
825,
|
| 891 |
+
223
|
| 892 |
+
],
|
| 893 |
+
"page_idx": 9
|
| 894 |
+
},
|
| 895 |
+
{
|
| 896 |
+
"type": "text",
|
| 897 |
+
"text": "5 Conclusion ",
|
| 898 |
+
"text_level": 1,
|
| 899 |
+
"bbox": [
|
| 900 |
+
174,
|
| 901 |
+
242,
|
| 902 |
+
299,
|
| 903 |
+
260
|
| 904 |
+
],
|
| 905 |
+
"page_idx": 9
|
| 906 |
+
},
|
| 907 |
+
{
|
| 908 |
+
"type": "text",
|
| 909 |
+
"text": "In this work, we present ShadeGAN, a new generative implicit model for shape-accurate 3D-aware image synthesis. We have shown that the multi-lighting constraint, achieved in ShadeGAN by explicit illumination modeling, significantly helps learning accurate 3D shapes from 2D images. ShadeGAN also allows us to control the lighting condition during image synthesis, achieving natural image relighting effects. To reduce the computational cost, we have further devised a light-weighted surface tracking network, which enables an efficient volume rendering technique for generative implicit models, achieving significant acceleration on both training and inference speed. A generative model with shape-accurate 3D representation could broaden its applications in vision and graphics, and our work has taken a solid step towards this goal. ",
|
| 910 |
+
"bbox": [
|
| 911 |
+
173,
|
| 912 |
+
273,
|
| 913 |
+
825,
|
| 914 |
+
398
|
| 915 |
+
],
|
| 916 |
+
"page_idx": 9
|
| 917 |
+
},
|
| 918 |
+
{
|
| 919 |
+
"type": "text",
|
| 920 |
+
"text": "Acknowledgment. We would like to thank Eric R. Chan for sharing the codebase of pi-GAN. This study is supported under the ERC Consolidator Grant 4DRepLy (770784). This study is also supported under the RIE2020 Industry Alignment Fund – Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). ",
|
| 921 |
+
"bbox": [
|
| 922 |
+
174,
|
| 923 |
+
412,
|
| 924 |
+
825,
|
| 925 |
+
469
|
| 926 |
+
],
|
| 927 |
+
"page_idx": 9
|
| 928 |
+
},
|
| 929 |
+
{
|
| 930 |
+
"type": "text",
|
| 931 |
+
"text": "References ",
|
| 932 |
+
"text_level": 1,
|
| 933 |
+
"bbox": [
|
| 934 |
+
174,
|
| 935 |
+
488,
|
| 936 |
+
266,
|
| 937 |
+
505
|
| 938 |
+
],
|
| 939 |
+
"page_idx": 9
|
| 940 |
+
},
|
| 941 |
+
{
|
| 942 |
+
"type": "text",
|
| 943 |
+
"text": "[1] T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” in CVPR, 2019. \n[2] T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, “Analyzing and improving the image quality of stylegan,” in CVPR, 2020. \n[3] A. Brock, J. Donahue, and K. Simonyan, “Large scale gan training for high fidelity natural image synthesis,” in ICLR, 2019. \n[4] E. R. Chan, M. Monteiro, P. Kellnhofer, J. Wu, and G. Wetzstein, “pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis,” in CVPR, 2021. \n[5] K. Schwarz, Y. Liao, M. Niemeyer, and A. Geiger, “Graf: Generative radiance fields for 3d-aware image synthesis,” in NeurIPS, 2020. \n[6] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” in ECCV, 2020. \n[7] R. J. Woodham, “Photometric method for determining surface orientation from multiple images,” Optical engineering, vol. 19, no. 1, p. 191139, 1980. \n[8] L. Liu, J. Gu, K. Zaw Lin, T.-S. Chua, and C. Theobalt, “Neural sparse voxel fields,” in NeurIPS, 2020. \n[9] A. Yu, R. Li, M. Tancik, H. Li, R. Ng, and A. Kanazawa, “PlenOctrees for real-time rendering of neural radiance fields,” in arXiv, 2021. \n[10] C. Reiser, S. Peng, Y. Liao, and A. Geiger, “Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps,” arXiv preprint arXiv:2103.13744, 2021. \n[11] S. J. Garbin, M. Kowalski, M. Johnson, J. Shotton, and J. Valentin, “Fastnerf: High-fidelity neural rendering at 200fps,” arXiv preprint arXiv:2103.10380, 2021. \n[12] T. Neff, P. Stadlbauer, M. Parger, A. Kurz, C. R. A. Chaitanya, A. Kaplanyan, and M. Steinberger, “Donerf: Towards real-time rendering of neural radiance fields using depth oracle networks,” arXiv preprint arXiv:2103.03231, 2021. \n[13] P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter, “A 3d face model for pose and illumination invariant face recognition,” in 2009 Sixth IEEE International Conference on Advanced Video and Signal Based Surveillance, Ieee, 2009. \n[14] M. Boss, R. Braun, V. Jampani, J. T. Barron, C. Liu, and H. Lensch, “Nerd: Neural reflectance decomposition from image collections,” arXiv preprint arXiv:2012.03918, 2020. \n[15] S. Bi, Z. Xu, P. Srinivasan, B. Mildenhall, K. Sunkavalli, M. Hašan, Y. Hold-Geoffroy, D. Kriegman, and R. Ramamoorthi, “Neural reflectance fields for appearance acquisition,” arXiv preprint arXiv:2008.03824, 2020. \n[16] P. P. Srinivasan, B. Deng, X. Zhang, M. Tancik, B. Mildenhall, and J. T. Barron, “Nerv: Neural reflectance and visibility fields for relighting and view synthesis,” in CVPR, 2021. \n[17] D. B. Lindell, J. N. Martel, and G. Wetzstein, “Autoint: Automatic integration for fast neural volume rendering,” arXiv preprint arXiv:2012.01714, 2020. \n[18] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in NIPS, 2014. \n[19] J. Wu, C. Zhang, T. Xue, B. Freeman, and J. Tenenbaum, “Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling,” in NIPS, 2016. \n[20] H. A. Alhaija, S. K. Mustikovela, A. Geiger, and C. Rother, “Geometric image synthesis,” in ACCV, Springer, 2018. \n[21] X. Chen, D. Cohen-Or, B. Chen, and N. J. Mitra, “Towards a neural graphics pipeline for controllable image generation,” Computer Graphics Forum, vol. 40, no. 2, 2021. \n[22] J.-Y. Zhu, Z. Zhang, C. Zhang, J. Wu, A. Torralba, J. Tenenbaum, and B. Freeman, “Visual object networks: Image generation with disentangled 3d representations,” in NeurIPS, 2018. \n[23] T. Nguyen-Phuoc, C. Li, L. Theis, C. Richardt, and Y.-L. Yang, “Hologan: Unsupervised learning of 3d representations from natural images,” in ICCV, 2019. \n[24] T. Nguyen-Phuoc, C. Richardt, L. Mai, Y.-L. Yang, and N. Mitra, “Blockgan: Learning 3d object-aware scene representations from unlabelled images,” in NeurIPS, Nov 2020. \n[25] Y. Liao, K. Schwarz, L. Mescheder, and A. Geiger, “Towards unsupervised learning of generative models for 3d controllable image synthesis,” in CVPR, 2020. \n[26] P. Henzler, N. J. Mitra, and T. Ritschel, “Escaping plato’s cave: 3d shape from adversarial rendering,” in ICCV, 2019. \n[27] A. Szabó, G. Meishvili, and P. Favaro, “Unsupervised generative 3d shape learning from natural images,” arXiv preprint arXiv:1910.00287, 2019. \n[28] M. Niemeyer and A. Geiger, “Giraffe: Representing scenes as compositional generative neural feature fields,” in CVPR, 2021. \n[29] A. Tewari, M. Elgharib, G. Bharaj, F. Bernard, H.-P. Seidel, P. Pérez, M. Zollhofer, and C. Theobalt, “Stylerig: Rigging stylegan for 3d control over portrait images,” in CVPR, 2020. \n[30] T. Yenamandra, A. Tewari, F. Bernard, H.-P. Seidel, M. Elgharib, D. Cremers, and C. Theobalt, “i3dmm: Deep implicit 3d morphable model of human heads,” in CVPR, 2021. \n[31] S. Tulsiani, N. Kulkarni, and A. Gupta, “Implicit mesh reconstruction from unannotated image collections,” arXiv preprint arXiv:2007.08504, 2020. \n[32] S. Goel, A. Kanazawa, and J. Malik, “Shape and viewpoint without keypoints,” ECCV, 2020. \n[33] L. Tran and X. Liu, “Nonlinear 3d face morphable model,” in CVPR, 2018. \n[34] A. Kanazawa, S. Tulsiani, A. A. Efros, and J. Malik, “Learning category-specific mesh reconstruction from image collections,” in ECCV, 2018. \n[35] A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik, “End-to-end recovery of human shape and pose,” in CVPR, 2018. \n[36] S. Sanyal, T. Bolkart, H. Feng, and M. J. Black, “Learning to regress 3d face shape and expression from an image without 3d supervision,” in CVPR, 2019. \n[37] J. Shang, T. Shen, S. Li, L. Zhou, M. Zhen, T. Fang, and L. Quan, “Self-supervised monocular 3d face reconstruction by occlusion-aware multi-view geometry consistency,” ECCV, 2020. \n[38] M. Sahasrabudhe, Z. Shu, E. Bartrum, R. Alp Guler, D. Samaras, and I. Kokkinos, “Lifting autoencoders: Unsupervised learning of a fully-disentangled 3d morphable model using deep non-rigid structure from motion,” in ICCV Workshops, 2019. \n[39] S. Wu, C. Rupprecht, and A. Vedaldi, “Unsupervised learning of probably symmetric deformable 3d objects from images in the wild,” in CVPR, 2020. \n[40] X. Li, S. Liu, K. Kim, S. De Mello, V. Jampani, M.-H. Yang, and J. Kautz, “Self-supervised single-view 3d reconstruction via semantic consistency,” ECCV, 2020. \n[41] X. Pan, B. Dai, Z. Liu, C. C. Loy, and P. Luo, “Do 2d gans know 3d shape? unsupervised 3d shape reconstruction from 2d image gans,” in ICLR, 2021. \n[42] L. Mescheder, A. Geiger, and S. Nowozin, “Which training methods for gans do actually converge?,” in ICML, 2018. \n[43] Z. Liu, P. Luo, X. Wang, and X. Tang, “Deep learning face attributes in the wild,” in ICCV, 2015. \n[44] W. Zhang, J. Sun, and X. Tang, “Cat head detection-how to effectively exploit shape and texture features,” in ECCV, 2008. \n[45] V. Sitzmann, J. Martel, A. Bergman, D. Lindell, and G. Wetzstein, “Implicit neural representations with periodic activation functions,” in NeurIPS, 2020. \n[46] M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” in NeurIPS, 2017. \n[47] B. T. Phong, “Illumination for computer generated pictures,” Communications of the ACM, vol. 18, no. 6, pp. 311–317, 1975. \n[48] E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville, “Film: Visual reasoning with a general conditioning layer,” in AAAI, 2018. \n[49] R. Liu, J. Lehman, P. Molino, F. P. Such, E. Frank, A. Sergeev, and J. Yosinski, “An intriguing failing of convolutional neural networks and the coordconv solution,” in NeurIPS, 2018. \n[50] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR, 2016. \n[51] Y. Wu and K. He, “Group normalization,” in ECCV, 2018. \n[52] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in ICML, 2015. \n[53] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. \n[54] “Unsup3d.” https://github.com/elliottwu/unsup3d. \n[55] “Celeba.” http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html. \n[56] “Cats.” https://web.archive.org/web/20150520175645/http://137.189.35.203/ WebUI/CatDatabase/catData.html. \n[57] Y. Choi, Y. Uh, J. Yoo, and J.-W. Ha, “Stargan v2: Diverse image synthesis for multiple domains,” in CVPR, 2020. ",
|
| 944 |
+
"bbox": [
|
| 945 |
+
178,
|
| 946 |
+
515,
|
| 947 |
+
826,
|
| 948 |
+
911
|
| 949 |
+
],
|
| 950 |
+
"page_idx": 9
|
| 951 |
+
},
|
| 952 |
+
{
|
| 953 |
+
"type": "text",
|
| 954 |
+
"text": "",
|
| 955 |
+
"bbox": [
|
| 956 |
+
171,
|
| 957 |
+
99,
|
| 958 |
+
828,
|
| 959 |
+
907
|
| 960 |
+
],
|
| 961 |
+
"page_idx": 10
|
| 962 |
+
},
|
| 963 |
+
{
|
| 964 |
+
"type": "text",
|
| 965 |
+
"text": "",
|
| 966 |
+
"bbox": [
|
| 967 |
+
171,
|
| 968 |
+
92,
|
| 969 |
+
828,
|
| 970 |
+
715
|
| 971 |
+
],
|
| 972 |
+
"page_idx": 11
|
| 973 |
+
}
|
| 974 |
+
]
|
parse/train/k-0oq5eNjh/k-0oq5eNjh_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/k-0oq5eNjh/k-0oq5eNjh_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/rJgzzJHtDB/rJgzzJHtDB.md
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# TRIPLE WINS: BOOSTING ACCURACY, ROBUSTNESS AND EFFICIENCY TOGETHER BY ENABLING INPUTADAPTIVE INFERENCE
|
| 2 |
+
|
| 3 |
+
Ting-Kuei Hu∗, Tianlong Chen∗, Haotao Wang Zhangyang Wang
|
| 4 |
+
|
| 5 |
+
Department of Computer Science and Engineering Texas A&M University, USA {tkhu,wiwjp619,htwang,atlaswang}@tamu.edu
|
| 6 |
+
|
| 7 |
+
# ABSTRACT
|
| 8 |
+
|
| 9 |
+
Deep networks were recently suggested to face the odds between accuracy (on clean natural images) and robustness (on adversarially perturbed images) (Tsipras et al., 2019). Such a dilemma is shown to be rooted in the inherently higher sample complexity (Schmidt et al., 2018) and/or model capacity (Nakkiran, 2019), for learning a high-accuracy and robust classifier. In view of that, give a classification task, growing the model capacity appears to help draw a win-win between accuracy and robustness, yet at the expense of model size and latency, therefore posing challenges for resource-constrained applications. Is it possible to co-design model accuracy, robustness and efficiency to achieve their triple wins?
|
| 10 |
+
|
| 11 |
+
This paper studies multi-exit networks associated with input-adaptive efficient inference, showing their strong promise in achieving a “sweet point” in cooptimizing model accuracy, robustness and efficiency. Our proposed solution, dubbed Robust Dynamic Inference Networks (RDI-Nets), allows for each input (either clean or adversarial) to adaptively choose one of the multiple output layers (early branches or the final one) to output its prediction. That multi-loss adaptivity adds new variations and flexibility to adversarial attacks and defenses, on which we present a systematical investigation. We show experimentally that by equipping existing backbones with such robust adaptive inference, the resulting RDI-Nets can achieve better accuracy and robustness, yet with over $30 \%$ computational savings, compared to the defended original models.
|
| 12 |
+
|
| 13 |
+
# 1 INTRODUCTION
|
| 14 |
+
|
| 15 |
+
Deep networks, despite their high predictive accuracy, are notoriously vulnerable to adversarial attacks (Goodfellow et al., 2015; Biggio et al., 2013; Szegedy et al., 2014; Papernot et al., 2016). While many defense methods have been proposed to increase a model’s robustness to adversarial examples, they were typically observed to hamper its accuracy on original clean images. Tsipras et al. (2019) first pointed out the inherent tension between the goals of adversarial robustness and standard accuracy in deep networks, whose provable existence was shown in a simplified setting. Zhang et al. (2019) theoretically quantified the accuracy-robustness trade-off, in terms of the gap between the risk for adversarial examples versus the risk for non-adversarial examples.
|
| 16 |
+
|
| 17 |
+
It is intriguing to consider whether and why the model accuracy and robustness have to be at odds. Schmidt et al. (2018) demonstrated that the number of samples needed to achieve adversarially robust generalization is polynomially larger than that needed for standard generalization, under the adversarial training setting. A similar conclusion was concurred by Sun et al. (2019) in the standard training setting. Tsipras et al. (2019) considered the accuracy-robustness trade-off as an inherent trait of the data distribution itself, indicating that this phenomenon persists even in the limit of infinite data. Nakkiran (2019) argued from a different perspective, that the complexity (e.g. capacity) of a robust classifier must be higher than that of a standard classifier. Therefore, replacing a largercapacity classifier might effectively alleviate the trade-off. Overall, those existing works appear to suggest that, while accuracy and robustness are likely to trade off for a fixed classification model and on a given dataset, such trade-off might be effectively alleviated (“win-win”), if supplying more training data and/or replacing a larger-capacity classifier.
|
| 18 |
+
|
| 19 |
+
On a separate note, deep networks also face the pressing challenge to be deployed on resourceconstrained platforms due to the prosperity of smart Internet-of-Things (IoT) devices. Many IoT applications naturally demand security and trustworthiness, e.g., , biometrics and identity verification, but can only afford limited latency, memory and energy budget. Hereby we extend the question: can we achieve a triple-win, i.e., , an accurate and robust classfier while keeping it efficient?
|
| 20 |
+
|
| 21 |
+
This paper makes an attempt in providing a positive answer to the above question. Rather than proposing a specific design of robust light-weight models, we reduce the average computation loads by input-adaptive routing to achieve triple-win. To this end, we introduce the input-adaptive dynamic inference (Teerapittayanon et al., 2017; Wang et al., 2018a), an emerging efficient inference scheme in contrast to the (non-adaptive) model compression, to the adversarial defense field for the first time. Given any deep network backbone (e.g., , ResNet, MobileNet), we first follow (Teerapittayanon et al., 2017) to augment it with multiple early-branch output layers in addition to the original final output. Each input, regardless of clean or adversarial samples, adaptively chooses which output layer to take for its own prediction. Therefore, a large portion of input inferences can be terminated early when the samples can already be inferred with high confidence.
|
| 22 |
+
|
| 23 |
+
Up to our best knowledge, no existing work studied adversarial attacks and defenses for an adaptive multi-output model, as the multiple sources of losses provide much larger flexibility to compose attacks (and therefore defenses), compared to the typical single-loss backbone. We present a systematical exploration on how to (white-box) attack and defense our proposed multi-output network with adaptive inference, demonstrating that the composition of multiple-loss information is critical in making the attack/defense strong. Fig. 1 illustrates our proposed Robust Dynamic Inference Networks (RDI-Nets). We show experimentally that the input-adaptive inference and multiloss flexibility can be our friend in achieving the desired “triple wins”. With our best defended RDI-Nets, we achieve better accuracy and robustness, yet with over $30 \%$ inference computational savings, compared to the defended original models as well as existing solutions co-designing robustness and efficiency (Gui et al., 2019; Guo et al., 2018). The codes can be referenced from https://github.com/TAMU-VITA/triple-wins.
|
| 24 |
+
|
| 25 |
+

|
| 26 |
+
Figure 1: Our proposed RDI-Net framework, a defended multi-output network enabling dynamic inference. Each image, being it clean or adversarially perturbed, adaptively picks one branch to exit.
|
| 27 |
+
|
| 28 |
+
# 2 RELATED WORK
|
| 29 |
+
|
| 30 |
+
# 2.1 ADVERSARIAL DEFENSE
|
| 31 |
+
|
| 32 |
+
A magnitude of defend approaches have been proposed (Kurakin et al., 2017; Xu et al., 2018; Song et al., 2018; Liao et al., 2018), although many were quickly evaded by new attacks (Carlini & Wagner, 2017; Baluja & Fischer, 2018). One strong defense algorithm that has so far not been fully compromised is adversarial training (Madry et al., 2018). It searches for adversarial images to augment the training procedure, although at the price of higher training costs (but not affecting inference efficiency). However, almost all existing attacks and defenses focus on a single-output classification (or other task) model. We are unaware of prior studies directly addressing attacks/defenses to more complicated networks with multiple possible outputs.
|
| 33 |
+
|
| 34 |
+
One related row of works are to exploit model ensemble (Tramer et al. \` , 2018; Strauss et al., 2017) in adversarial training. The gains of the defended ensemble compared to a single model could be viewed as the benefits of either the benefits of diversity (generating stronger and more transferable perturbations), or the increasing model capacity (consider the ensembled multiple models as a compound one). Unfortunately, ensemble methods could amplify the inference complexity and be detrimental for efficiency. Besides, it is also known that injecting randomization at inference time helps mitigate adversarial effects (Xie et al., 2018; Cohen et al., 2019). Yet up to our best knowledge, no work has studied non-random, but rather input-dependent inference for defense.
|
| 35 |
+
|
| 36 |
+
# 2.2 EFFICIENT INFERENCE
|
| 37 |
+
|
| 38 |
+
Research in improving deep network efficiency could be categorized into two streams: the static way that designs compact models or compresses heavy models, while the compact/compressed models remain fixed for all inputs at inference; and the dynamic way, that at inference the inputs can choose different computational paths adaptively, and the simpler inputs usually take less computation to make predictions. We briefly review the literature below.
|
| 39 |
+
|
| 40 |
+
Static: Compact Network Design and Model Compression. Many compact architectures have been specifically designed for resource-constrained applications, by adopting lightweight depthwise convolutions (Sandler et al., 2018), and group-wise convolutions with channel-shuffling (Zhang et al., 2018), to just name a few. For model compression, Han et al. (2015) first proposed to sparsify deep models by removing non-significant synapses and then re-training to restore performance. Structured pruning was later on introduced for more hardware friendliness (Wen et al., 2016). Layer factorization (Tai et al., 2016; Yu et al., 2017), quantization (Wu et al., 2016), model distillation (Wang et al., 2018c) and weight sharing (Wu et al., 2018) have also been respectively found effective.
|
| 41 |
+
|
| 42 |
+
Dynamic: Input-Adaptive Inference. Higher inference efficiency could be also accomplished by enabling input-conditional execution. Teerapittayanon et al. (2017); Huang et al. (2018); Kaya et al. (2019) leveraged intermediate features to augment multiple side branch classifiers to enable early predictions. Their methodology sets up the foundation for our work. Other efforts (Figurnov et al., 2017; Wang et al., 2018a;b; 2019) allow for an input to choose between passing through or skipping each layer. The approach could be integrated with RDI-Nets too, which we leave as future work.
|
| 43 |
+
|
| 44 |
+
# 2.3 BRIDGING ROBUSTNESS WITH EFFICIENCY
|
| 45 |
+
|
| 46 |
+
A few studies recently try to link deep learning robustness and efficiency. Guo et al. (2018) observed that in a sparse deep network, appropriately sparsified weights improve robustness, whereas over-sparsification (e.g., less than $5 \%$ nonzero weights) in turn makes the model more fragile. Two latest works (Ye et al., 2019; Gui et al., 2019) examined the robustness of compressed models, and concluded similar observations that the relationship between mode size and robustness depends on compression methods and are often non-monotonic. Lin et al. (2019) found that activation quantization may hurt robustness, but can be turned into effective defense if enforcing continuity constraints.
|
| 47 |
+
|
| 48 |
+
Different from above methods that tackle robustness from static compact/compressed models, the proposed RDI-Nets are the first to address robustness from the dynamic input-adaptive inference. Our experiment results demonstrate the consistent superiority of RDI-Nets over those static methods (Section 4.3). Moreover, applying dynamic inference top of those static methods may further boost the robustness and efficiency, which we leave as future work.
|
| 49 |
+
|
| 50 |
+
# 3 APPROACH
|
| 51 |
+
|
| 52 |
+
With the goal of achieving inference efficiency, we first look at the setting of multi-output networks and the specific design of RDI-Net in Section 3.1. Then we define three forms of adversarial attacks for multi-output networks in Section 3.2 and their corresponding defense methods in Section 3.3.
|
| 53 |
+
|
| 54 |
+
Note that RDI-Nets achieve “triple wins”via reducing the average computation loads through inputadaptive routing. It is not to be confused with any specifically-designed robust light-weight model.
|
| 55 |
+
|
| 56 |
+
# 3.1 DESIGNING RDI-NETS FOR HIGHER INFERENCE EFFICIENCY
|
| 57 |
+
|
| 58 |
+
Given an input image $x$ , an $N$ -output network can produce a set of predictions $[ \hat { y _ { 1 } } , . . . , \hat { y _ { N } } ]$ by a set of transformations $[ f _ { \theta _ { 1 } } ( \cdot ) , . . . , f _ { \theta _ { N } } ( \cdot ) ]$ . $\theta _ { i }$ denote the model parameter of $f _ { \theta _ { i } }$ , $i = 1 , . . . , N$ , and $f _ { \theta _ { i } } \mathbf { s }$ will typically share some weights. With an input $x$ , one can express $\hat { y } _ { i } = f _ { \boldsymbol { \theta } _ { i } } ( \boldsymbol { x } )$ . We assume that the final prediction will be one chosen (NOT fused) from $[ \hat { y _ { 1 } } , . . . , \hat { y _ { N } } ]$ via some deterministic strategy.
|
| 59 |
+
|
| 60 |
+
We now look at RDI-Nets as a specific instance of multi-output networks, specifically designed for the goal of more efficient, input-adaptive inference. As shown in Fig. 1, for any deep network (e.g., , ResNet, MobileNet), we could append $K$ side branches (with negligible overhead) to allow for early-exit predictions. In other words, it becomes a $( K + 1 )$ -output network, and the subneworks with the $K + 1$ exits, from the lowest to the highest (the original final output), correspond to $[ f _ { \theta _ { 1 } } ( \cdot ) , . . . , f _ { \theta _ { K + 1 } } ( \cdot ) ]$ . They share their weights in a nested fashion: ${ \theta } _ { 1 } \subseteq { \theta } _ { 2 } . . . \subseteq { \theta } _ { K + 1 }$ , with $\theta _ { K + 1 }$ including the entire network’s parameters.
|
| 61 |
+
|
| 62 |
+
Our deterministic strategy in selecting one final output follows (Teerapittayanon et al., 2017). We set a confidence threshold $t _ { k }$ for each $k$ -th exit, $k = 1 , . . . , K + 1$ , and each input $x$ will terminate inference and output its prediction in the earliest exit (smallest $k$ ), whose softmax entropy (as a confidence measure) falls below $t _ { k }$ . All computations after the $k$ -th exit will not be activated for this $x$ . Such a progressive and early-halting mechanism effectively saves unnecessary computation for most easier-to-classify samples, and applies in both training and inference. Note that, if efficiency is not the concern, instead of choosing (the earliest one), we could have designed an adaptive or randomized fusion of all $f _ { \theta _ { i } }$ predictions: but that falls beyond the goal of this work.
|
| 63 |
+
|
| 64 |
+
The training objective for RDI-Nets could be written as
|
| 65 |
+
|
| 66 |
+
$$
|
| 67 |
+
L _ { R D I } = \sum _ { i = 1 } ^ { K + 1 } w _ { i } [ ( \phi ( f _ { i } ( \theta _ { i } | x ) , y ) + \phi ( f _ { i } ( \theta _ { i } | x ^ { a d v } ) , y ) ] ,
|
| 68 |
+
$$
|
| 69 |
+
|
| 70 |
+
For each exit loss, we minimize a hybrid loss of accuracy (on clean $x$ ) and robustness (on $x ^ { a d v }$ ). The K + 1 exits are balanced with a group of weights {wi}K+1i=1 . More details about RDI-Net structures, hyperparameters, and inference branch selection can be founded in Appendix A, B, and C.
|
| 71 |
+
|
| 72 |
+
In what follows, we discuss three ways to generate $x ^ { a d v }$ in RDI-Nets, and then their defenses.
|
| 73 |
+
|
| 74 |
+
# 3.2 THREE ATTACK FORMS ON MULTI-OUTPUT NETWORKS
|
| 75 |
+
|
| 76 |
+
We consider white box attacks in this paper. Attackers have access to the model’s parameters, and aim to generate an adversarial image $x ^ { a d v }$ to fool the model by perturbing an input $x$ within a given magnitude bound.
|
| 77 |
+
|
| 78 |
+
We next discuss three attack forms for an $N$ -output network. Note that they are independent of, and to be distinguished from attacker algorithms (e.g., , PGD, C&W, FGSM): the former depicts the optimization formulation, that can be solved any of the attacker algorithms.
|
| 79 |
+
|
| 80 |
+
Single Attack Naively extending from attacking single-output networks, a single attack is defined to maximally fool one $f _ { \theta _ { i } } ( \cdot )$ only, expressed as:
|
| 81 |
+
|
| 82 |
+
$$
|
| 83 |
+
x _ { i } ^ { a d v } = \underset { x ^ { \prime } \in | x ^ { \prime } - x | _ { \infty } \leq \epsilon } { \arg \operatorname* { m a x } } | \phi ( f _ { \theta _ { i } } ( x ^ { \prime } ) , y ) | ,
|
| 84 |
+
$$
|
| 85 |
+
|
| 86 |
+
where $y$ is the ground truth label, and $\phi$ is the loss for $f _ { \theta _ { i } }$ (we assume softmax for all). $\epsilon$ is the perturbation radius and we adopt $\ell _ { \infty }$ ball for an empirically strong attacker. Naturally, an $N$ -output network can have $N$ different single attacks. However, each single attack is derived without being aware of other parallel outputs. The found $x _ { i } ^ { a d v }$ is not necessarily transferable to other $f _ { \boldsymbol { \theta } _ { j } } \mathbf { s } \left( j \neq i \right)$ , and therefore can be easily bypassed if $x$ is re-routed through other outputs to make its prediction.
|
| 87 |
+
|
| 88 |
+
Average Attack Our second attack maximizes the average of all $f _ { \theta _ { i } }$ losses, so that the found $x ^ { a d v }$ remains in effect no matter which one $f _ { \theta _ { i } }$ is chosen to output the prediction for $x$ :
|
| 89 |
+
|
| 90 |
+
$$
|
| 91 |
+
x _ { a v g } ^ { a d v } = \underset { x ^ { \prime } \in | x ^ { \prime } - x | _ { \infty } \leq \epsilon } { \arg \operatorname* { m a x } } | \frac { 1 } { N } \sum _ { j = 1 } ^ { N } \phi ( f _ { \theta _ { j } } ( x ^ { \prime } ) , y ) | ,
|
| 92 |
+
$$
|
| 93 |
+
|
| 94 |
+
The average attack addresses takes into account the attack transferablity and involves all $\theta _ { j } \mathbf { s }$ into optimization. However, while only one output will be selected for each sample at inference, the average strategy might weaken the individual defense strength of each $f _ { \theta _ { i } }$ .
|
| 95 |
+
|
| 96 |
+
Max-Average Attack Our third attack aims to emphasize individual output defense strength, more than simply maximizing an all-averaged loss. We first solve the $N$ single attacks $x _ { i } ^ { a d v }$ as described in Eqn. 2, and denote their collection as $\Omega$ . We then solve the max-average attack via the following:
|
| 97 |
+
|
| 98 |
+
$$
|
| 99 |
+
i ^ { * } = \arg \operatorname* { m a x } _ { i } | \frac { 1 } { N } \sum _ { j = 1 } ^ { N } \phi ( f _ { \theta _ { j } } ( x _ { i } ^ { a d v } ) , y ) | .
|
| 100 |
+
$$
|
| 101 |
+
|
| 102 |
+
Note Eqn. 4 differs from Eqn. 3 by adding an $\Omega$ constraint to balance between “commodity” and “specificity”. The found $\hat { x } _ { m a x } ^ { a d v }$ both strongly increases the averaged loss values from all $f _ { i } \mathbf { s }$ (therefore possessing transferablity), and maximally fools one individual $f _ { \boldsymbol { \theta } _ { i } } \mathbf { s }$ as it is selected from the collection $\Omega$ of single attacks.
|
| 103 |
+
|
| 104 |
+
# 3.3 DEFENCE ON MULTI-OUTPUT NETWORKS
|
| 105 |
+
|
| 106 |
+
For simplicity and fair comparison, we focus on adversarial training (Madry et al., 2018) as our defense framework, where the three above defined attack forms can be plugged-in to generate adversarial images to augment training, as follows ( $\Theta$ is the union of learnable parameters):
|
| 107 |
+
|
| 108 |
+
$$
|
| 109 |
+
\Xi \Theta , \mathrm { w h e r e } \theta _ { i } = \underset { \theta ^ { \prime } } { \arg \operatorname* { m i n } } | \phi ( f _ { \theta _ { i } } ( x ) , y ) + \phi ( f _ { \theta _ { i } } ( x ^ { a d v } ) , y ) | . ,
|
| 110 |
+
$$
|
| 111 |
+
|
| 112 |
+
where $x ^ { a d v } \in \{ x _ { i } ^ { a d v } , x _ { a v g } ^ { a d v } , x _ { m a x } ^ { a d v } \} { } _ { }$ . As $f _ { i } \mathbf { s }$ partially share their weights $\theta _ { i }$ in a multi-output network, the updates from different $f _ { i } \mathbf { s }$ will be averaged on the shared parameters.
|
| 113 |
+
|
| 114 |
+
# 4 EXPERIMENTAL RESULTS
|
| 115 |
+
|
| 116 |
+
4.1 EXPERIMENTAL SETUP
|
| 117 |
+
|
| 118 |
+
Evaluation Metrics We evaluate accuracy, robustness, and efficiency, using the metrics below:
|
| 119 |
+
|
| 120 |
+
• Testing Accuracy (TA): the classification accuracy on the original clean test set. • Adversarial Testing Accuracy (ATA): Given an attacker, ATA stands for the classification accuracy on the attacked test set. It is the same as the “robust accuracy” in (Zhang et al., 2019). • Mega Flops (MFlops): The number of million floating-point multiplication operations consumed on the inference, averaged over the entire testing set.
|
| 121 |
+
|
| 122 |
+
Datasets and Benchmark Models We evaluate three representative CNN models on two popular datasets: SmallCNN on MNIST (Chen et al., 2018); ResNet-38 (He et al., 2016) and MobileNet-V2 (Sandler et al., 2018) on CIFAR-10. The three networks span from simplest to more complicated, and covers a compact backbone. All three models are defended by adversarial training, constituting strong baselines. Table 1 reports the models, datasets, the attacker algorithm used in attack & defense, and thee TA/ATA/MFlops performance of three defended models.
|
| 123 |
+
|
| 124 |
+
Attack and Defense on RDI-Nets We build RDI-Nets by appending side branch outputs for each backbone. For SmallCNN, we add two side branches $K = 2$ ). For ResNet-38 and MobileNet-V2, we have $K = 6$ and $K = 2$ , respectively. The branches are designed to cause negligible overheads: more details of their structure and positions can be referenced in Appendix B. We call those result models RDI-SmallCNN, RDI-ResNet38 and RDI-MobileNetV2 hereinafter.
|
| 125 |
+
|
| 126 |
+
We then generate attacks using our three defined forms. Each attack form could be solved with various attacker algorithms (e.g., PGD, C&W, FGSM), and by default we solve it with the same attacker used for each backbone in Table 1. If we fix one attacker algorithm (e.g., PGD), then TA/ATA for a single-output network can be measured without ambiguity. Yet for $( K { + } 1 )$ -output RDI-Nets, there could be at least $K { + 3 }$ different ATA numbers for one defended model, depending on what attack form in Section 3.1 to apply $\scriptstyle { K + 1 }$ single attacks, 1 average attack, and 1 maxaverage attack). For example, we denote by ATA (Branch1) the ATA number when applying the single attack generated from the first side output branch (e.g., $x _ { 1 } ^ { a d v } .$ ); similarly elsewhere.
|
| 127 |
+
|
| 128 |
+
We also defend RDI-Nets using adversarial training, using the forms of adversarial images to augment training. By default, we adopt three adversarial training defense schemes: Main Branch (single attack using $x _ { K + 1 } ^ { a d v } ) ^ { 1 }$ 1,Average (using $x _ { a v g . } ^ { a d v . }$ ),and Max-Average (using $x _ { a v g } ^ { m a x }$ ), in addition to the undefended RDI-Nets (using standard training) denoted as Standard.
|
| 129 |
+
|
| 130 |
+
We cross evaluate ATAs of different defenses and attacks, since an ideal defense shall protect against all possible attack forms. To faithfully indicate the actual robustness, we choose the lowest number among all $K + 3$ ATAs, denoted as ATA (Worst-Case), as the robustness measure for an RDI-Net.
|
| 131 |
+
|
| 132 |
+
Table 1: Benchmarking results of adverserial training of three networks. PGD-40 denotes running the projected gradient descent attacker (Madry et al., 2018) for 40 iterations. We set the perturbation size as 0.3 for MNIST and 8/255 for CIFAR-10 in $\ell _ { \infty }$ norm (adopted by all following experiments).
|
| 133 |
+
|
| 134 |
+
<table><tr><td>Model</td><td>Dataset</td><td>Defend</td><td>Attack</td><td>TA</td><td>ATA</td><td>MFlops</td></tr><tr><td>SmalICNN</td><td>MNIST</td><td>PGD-40</td><td>PGD-40</td><td>99.49%</td><td>96.31%</td><td>9.25</td></tr><tr><td>ResNet-38</td><td>CIFAR-10</td><td>PGD-10</td><td>PGD-20</td><td>83.62%</td><td>42.29%</td><td>79.42</td></tr><tr><td>MobileNetV2</td><td>CIFAR-10</td><td>PGD-10</td><td>PGD-20</td><td>84.42%</td><td>46.92%</td><td>86.91</td></tr></table>
|
| 135 |
+
|
| 136 |
+
# 4.2 EVALUATION AND ANALYSIS
|
| 137 |
+
|
| 138 |
+
MNIST Experiments The MNIST experimental results on RDI-SmallCNN are summarized in table 2, with several meaningful observations to be drawn. First, the undefended models (Standard) are easily compromised by all attack forms. Second, The single attack-defended model (Main Branch) achieves the best ATA against the same type of attack, i.e., ATA (Main Branch), and also seems to boost the closest output branch’s robustness, i.e., ATA (Branch 2). However, its defense effect on the further-away Branch 1 is degraded, and also shows to be fragile under two stronger attacks (Average, and Max-Average). Third, both Average and Max-Average defenses achieve good TAs, as well as ATAs against all attack forms (and therefore Worst-Case), with Max-Average slightly better at both (the margins are small due to the data/task simplicity; see next two).
|
| 139 |
+
|
| 140 |
+
Moreover, compared to the strong baseline of SmallCNN defended by PGD (40 iterations)-based adversarial training, RDI-SmallCNN with Max-Average defense wins in terms of both TA and ATA. Impressively, that comes together with $3 4 . 3 0 \%$ computational savings compared to the baseline. Here the different defense forms do not appear to alter the inference efficiency much: they all save around $34 \% - 3 6 \%$ MFlops compared to the backbone.
|
| 141 |
+
|
| 142 |
+
Table 2: The performance of RDI-SmallCNN. The ”Average MFlops” is calculated by averaging the total flop costs consumed over the inference of the entire set (different samples take different FLOPs due to input-adaptive inference). The perturbation size and step size are 0.3 and 0.01, respectively.
|
| 143 |
+
|
| 144 |
+
<table><tr><td rowspan=1 colspan=1>Defense Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>99.48%</td><td rowspan=1 colspan=1>99.50%</td><td rowspan=1 colspan=1>99.51%</td><td rowspan=1 colspan=1>99.52%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch 1)</td><td rowspan=1 colspan=1>6.60%</td><td rowspan=1 colspan=1>60.50%</td><td rowspan=1 colspan=1>98.69%</td><td rowspan=1 colspan=1>98.52%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch 2)</td><td rowspan=1 colspan=1>3.16%</td><td rowspan=1 colspan=1>98.14%</td><td rowspan=1 colspan=1>97.64%</td><td rowspan=1 colspan=1>97.62%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>1.32%</td><td rowspan=1 colspan=1>96.70%</td><td rowspan=1 colspan=1>96.30%</td><td rowspan=1 colspan=1>96.43%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>2.61%</td><td rowspan=1 colspan=1>61.35%</td><td rowspan=1 colspan=1>97.37%</td><td rowspan=1 colspan=1>97.42%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>2.10%</td><td rowspan=1 colspan=1>61.83%</td><td rowspan=1 colspan=1>96.82%</td><td rowspan=1 colspan=1>96.89%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>1.32%</td><td rowspan=1 colspan=1>60.50%</td><td rowspan=1 colspan=1>96.30%</td><td rowspan=1 colspan=1>96.43%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>5.89</td><td rowspan=1 colspan=1>5.89</td><td rowspan=1 colspan=1>5.95</td><td rowspan=1 colspan=1>6.08</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>36.40%</td><td rowspan=1 colspan=1>36.40%</td><td rowspan=1 colspan=1>35.70%</td><td rowspan=1 colspan=1>34.30%</td></tr></table>
|
| 145 |
+
|
| 146 |
+
CIFAR-10 Experiments The results on RDI-ResNet38 and RDI-MobileNetV2 are presented in Tables 3 and 4, respectively. Most findings seem to concur with MNIST experiments. Specifically, on the more complicated CIFAR-10 classification task, Max-Average defense achieves much more obvious margins over Average defense, in terms of ATA (Worst-Case): $2 . 7 9 \%$ for RDI-ResNet38, and $1 . 0 6 \%$ for RDI-MobileNetV2. Interestingly, the Average defense is not even the strongest in defending average attacks, as Max-Average defense can achieve higher ATA (Average) in both cases. We conjecture that averaging all branch losses might “over-smooth” and diminish useful gradients.
|
| 147 |
+
|
| 148 |
+
Compared to the defended ResNet-38 and MobileNet-V2 backbones, RDI-Nets with Max-Average defense achieve higher TAs and ATAs for both. Especially, the ATA (Worst-Case) of RDI-ResNet38 surpasses the ATA of ResNet-38 defended by PGD-adversarial training by $1 . 0 3 \%$ , while saving around $30 \%$ inference budget. We find that different defenses on CIFAR-10 have more notable impacts on computational saving. Seemingly, a stronger defense (Max-Average) requires inputs to go through the scrutiny of more layers on average, before outputting confident enough predictions: a sensible observation as we expect.
|
| 149 |
+
|
| 150 |
+
Visualization of Adaptive Inference Behaviors We visualize the exiting behaviors of RDIResNet38 in Fig 4.2. We plot each branch exiting percentage on clean set and adversarial sets (worst-case) of examples. A few interesting observations can be found. First, we observe that the single-attack defended model can be easily fooled as adversarial examples can be routed through other less-defended outputs (due to the limited transferability of attacks between different outputs). Second, the two stronger defenses (Average and Max-Average) show much more uniform usage of multiple outputs. Their routing behaviors for clean examples are almost identical. For adversarial examples, Max-Average tends to call upon the full inference more often (i.e., more “conservative”).
|
| 151 |
+
|
| 152 |
+
Table 3: The performance evaluation on RDI-ResNet38. The perturbation size and step size are 8/255 and 2/255, respectively.
|
| 153 |
+
|
| 154 |
+
<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>0.12%</td><td rowspan=1 colspan=1>12.02%</td><td rowspan=1 colspan=1>71.56%</td><td rowspan=1 colspan=1>69.71%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>5.58%</td><td rowspan=1 colspan=1>66.67%</td><td rowspan=1 colspan=1>63.11%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch3)</td><td rowspan=1 colspan=1>0.04%</td><td rowspan=1 colspan=1>42.73%</td><td rowspan=1 colspan=1>60.65%</td><td rowspan=1 colspan=1>60.72%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch4)</td><td rowspan=1 colspan=1>0.06%</td><td rowspan=1 colspan=1>34.95%</td><td rowspan=1 colspan=1>50.17%</td><td rowspan=1 colspan=1>47.82%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch5)</td><td rowspan=1 colspan=1>0.06%</td><td rowspan=1 colspan=1>41.77%</td><td rowspan=1 colspan=1>44.83%</td><td rowspan=1 colspan=1>45.53%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch6)</td><td rowspan=1 colspan=1>0.11%</td><td rowspan=1 colspan=1>41.68%</td><td rowspan=1 colspan=1>45.83%</td><td rowspan=1 colspan=1>44.12%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>0.13%</td><td rowspan=1 colspan=1>42.74%</td><td rowspan=1 colspan=1>47.52%</td><td rowspan=1 colspan=1>49.82%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>9.14%</td><td rowspan=1 colspan=1>42.09%</td><td rowspan=1 colspan=1>43.32%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>7.15%</td><td rowspan=1 colspan=1>40.53%</td><td rowspan=1 colspan=1>43.43%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>5.58%</td><td rowspan=1 colspan=1>40.53%</td><td rowspan=1 colspan=1>43.32%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>29.41</td><td rowspan=1 colspan=1>48.27</td><td rowspan=1 colspan=1>56.90</td><td rowspan=1 colspan=1>57.81</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>62.96%</td><td rowspan=1 colspan=1>39.20%</td><td rowspan=1 colspan=1>28.35%</td><td rowspan=1 colspan=1>27.20%</td></tr></table>
|
| 155 |
+
|
| 156 |
+
Table 4: The performance evaluation on RDI-MobilenetV2. The perturbation size and step size are 8/255 and 2/255, respectively.
|
| 157 |
+
|
| 158 |
+
<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>93.22%</td><td rowspan=1 colspan=1>85.28%</td><td rowspan=1 colspan=1>82.14%</td><td rowspan=1 colspan=1>84.91%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>0.35%</td><td rowspan=1 colspan=1>37.40%</td><td rowspan=1 colspan=1>67.65%</td><td rowspan=1 colspan=1>71.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>47.35%</td><td rowspan=1 colspan=1>50.38%</td><td rowspan=1 colspan=1>50.15%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>46.69%</td><td rowspan=1 colspan=1>49.33%</td><td rowspan=1 colspan=1>46.99%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>35.20%</td><td rowspan=1 colspan=1>45.93%</td><td rowspan=1 colspan=1>47.00%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>36.66%</td><td rowspan=1 colspan=1>49.33%</td><td rowspan=1 colspan=1>50.18%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>35.20%</td><td rowspan=1 colspan=1>45.93%</td><td rowspan=1 colspan=1>46.99%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>49.78</td><td rowspan=1 colspan=1>52.81</td><td rowspan=1 colspan=1>58.23</td><td rowspan=1 colspan=1>60.84</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>42.72%</td><td rowspan=1 colspan=1>39.23%</td><td rowspan=1 colspan=1>33.00%</td><td rowspan=1 colspan=1>29.99%</td></tr></table>
|
| 159 |
+
|
| 160 |
+

|
| 161 |
+
Figure 2: The exiting behaviours of RDI-ResNet38 defended by (a) Single attack defense (Main Branch); (b) Average defense; and (c) Max-Average defense.
|
| 162 |
+
|
| 163 |
+
# 4.3 COMPARISON WITH DEFENDED SPARSE NETWORKS
|
| 164 |
+
|
| 165 |
+
An alternative to achieve accuracy-robust-efficiency trade-off is by defending a sparse or compressed model. Inspired by (Guo et al., 2018; Gui et al., 2019), we compare RDI-Net with Max-Average defense to the following baseline: first compressing the network with a state-of-the-art model compression method (Huang & Wang, 2018), and then defend the compressed network using the PGD-10 adversarial training. We sample different sparsity ratios in (Huang & Wang, 2018) to obtain models of different complexities. Fig. 6 in Appendix visualizes the comparison on ResNet-38: for either method, we sample a few models of different MFLOPs. At similar inference costs (e.g., 49.38M for pruning $^ +$ defense, and 48.35M for RDI-Nets), our proposed approach consistently achieves higher ATAs $( > 2 \% )$ ) than the strong pruning $^ +$ defense baseline, with higher TAs.
|
| 166 |
+
|
| 167 |
+
We also compare with the latest ATMC algorithm (Gui et al., 2019) that jointly optimizes robustness and efficiency, applied the same ResNet-38 backbone. As shown in Table 5, at comparable MFlops, RDI-ResNet-38 surpasses ATMC by $0 . 3 \%$ in terms of ATA, with a similar TA.
|
| 168 |
+
|
| 169 |
+
Table 5: Performance comparison between RDIResNet38 and ATMC.
|
| 170 |
+
|
| 171 |
+
<table><tr><td>Methods</td><td>TA</td><td>ATA</td><td>MFlops</td></tr><tr><td>ATMC (Gui et al. (2019))</td><td>83.81</td><td>43.02</td><td>56.82</td></tr><tr><td>RDI-ResNet-38(Worst-Case)</td><td>83.79</td><td>43.32</td><td>57.81</td></tr></table>
|
| 172 |
+
|
| 173 |
+
# 4.4 GENERALIZED ROBUSTNESS AGAINST OTHER ATTACKERS
|
| 174 |
+
|
| 175 |
+
In the aforementioned experiments, we have only evaluated on RDI-Nets against “deterministic” PGD-based adversarial images. We show that RDI-Nets also achieve better generalized robustness against other “randomized” or unseen attackers. We create the new “random attack”: that attack will randomly combine the multi-exit losses, and summarize the results in Table 6. We also follow the similar setting in Gui et al. (2019) and report the results against FGSM (Goodfellow et al., 2015) and WRM (Sinha et al., 2018) attacker, in Tables 7 and 8 respectively (more complete results can be found in Appendix D).
|
| 176 |
+
|
| 177 |
+
Table 6: Performance on RDI-ResNet38 against random attack. The perturbation size and step size are 8/255 and 2/255, respectively. More details of random attack can be referenced in Appendix D.
|
| 178 |
+
|
| 179 |
+
<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Random)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>10.33%</td><td rowspan=1 colspan=1>43.11%</td><td rowspan=1 colspan=1>44.86%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>27.33</td><td rowspan=1 colspan=1>52.36</td><td rowspan=1 colspan=1>55.21</td><td rowspan=1 colspan=1>56.54</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>65.58%</td><td rowspan=1 colspan=1>34.07%</td><td rowspan=1 colspan=1>30.48%</td><td rowspan=1 colspan=1>28.80%</td></tr></table>
|
| 180 |
+
|
| 181 |
+
Table 7: Performance on RDI-ResNet38 (defended with PGD) against FGSM attack (perturbation size is 8/255). The original defended ResNet38 by PGD under the same attack has ATA $5 1 . 1 1 \%$ .
|
| 182 |
+
|
| 183 |
+
<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>11.51%</td><td rowspan=1 colspan=1>51.45%</td><td rowspan=1 colspan=1>53.64%</td><td rowspan=1 colspan=1>54.72%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>11.41%</td><td rowspan=1 colspan=1>50.21%</td><td rowspan=1 colspan=1>51.81%</td><td rowspan=1 colspan=1>53.20%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>52.40%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>51.05%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>65.74</td><td rowspan=1 colspan=1>55.27</td><td rowspan=1 colspan=1>58.27</td><td rowspan=1 colspan=1>59.67</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>17.21%</td><td rowspan=1 colspan=1>30.40%</td><td rowspan=1 colspan=1>26.40%</td><td rowspan=1 colspan=1>24.86%</td></tr></table>
|
| 184 |
+
|
| 185 |
+
Table 8: Performance on RDI-ResNet38 (defended with PGD) against WRM attack (perturbation size is 0.3). The original defended ResNet38 by PGD under the same attack has ATA $8 3 . 3 5 \%$ .
|
| 186 |
+
|
| 187 |
+
<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>34.42%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>26.48%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.40%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>50.05</td><td rowspan=1 colspan=1>50.46</td><td rowspan=1 colspan=1>52.89</td><td rowspan=1 colspan=1>52.38</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>36.98%</td><td rowspan=1 colspan=1>36.46%</td><td rowspan=1 colspan=1>33.40%</td><td rowspan=1 colspan=1>34.04%</td></tr></table>
|
| 188 |
+
|
| 189 |
+
# 5 DISCUSSION AND ANALYSIS
|
| 190 |
+
|
| 191 |
+
Intuition: Multi-Output Networks as Special Ensembles Our intuition on defending multioutput networks arises from the success of ensemble defense in improving both accuracy and robustness (Tramer et al. \` , 2018; Strauss et al., 2017), which also aligns with the model capacity hypothesis (Nakkiran, 2019). A general multi-output network (Xu et al., 2019) could be decomposed by an ensemble of single-output models, with weight re-using enforced among them. It is thus more compact than an ensemble of independent models, and the extent of sharing weight calibrates ensemble diversity versus efficiency. Therefore, we expect a defended multi-output network to (mostly) inherit the strong accuracy/robustness of ensemble defense, while keeping the inference cost lower.
|
| 192 |
+
|
| 193 |
+
Do ”Triple Wins” Go Against the Model Capacity Needs? We point out that our seemingly “free” efficiency gains (e.g., not sacrificing TA/ATA) do not go against the current belief that a more accurate and robust classifier relies on a larger model capacity (Nakkiran, 2019). From the visualization, there remains to be a portion of clean/adversarial examples that have to utilize the full inference to predict well. In other words, the full model capacity is still necessary to achieve our current TAs/ATAs. Meanwhile, just like in standard classification (Wang et al., 2018a), not all adversarial examples are born equally. Many of them can be predicted using fewer inference costs (taking earlier exits). Therefore, RDI-Nets reduces the “effective model capacity” averaged on all testing samples for overall higher inference efficiency, while not altering the full model capacity.
|
| 194 |
+
|
| 195 |
+
# 6 CONCLUSION
|
| 196 |
+
|
| 197 |
+
This paper targets to simultaneously achieve high accuracy and robustness and meanwhile keeping inference costs lower. We introduce the multi-output network and input-adaptive dynamic inference, as a strong tool to the adversarial defense field for the first time. Our RDI-Nets achieve the “triple wins” of better accuracy, stronger robustness, and around $30 \%$ inference computational savings. Our future work will extend RDI-Nets to more dynamic inference mechanisms.
|
| 198 |
+
|
| 199 |
+
# 7 ACKNOWLEDGEMENT
|
| 200 |
+
|
| 201 |
+
We would like to thank Dr. Yang Yang from Walmart Technology for highly helpful discussions throughout this project.
|
| 202 |
+
|
| 203 |
+
# REFERENCES
|
| 204 |
+
|
| 205 |
+
Shumeet Baluja and Ian Fischer. Adversarial transformation networks: Learning to generate adversarial examples. In AAAI, 2018.
|
| 206 |
+
|
| 207 |
+
Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndi ˇ c, Pavel Laskov, Gior- ´ gio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In ECML, 2013.
|
| 208 |
+
|
| 209 |
+
Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In $S P$ , 2017.
|
| 210 |
+
|
| 211 |
+
Tian Qi Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. In NeurIPS, 2018.
|
| 212 |
+
|
| 213 |
+
Jeremy M Cohen, Elan Rosenfeld, and J Zico Kolter. Certified adversarial robustness via randomized smoothing. In ICML, 2019.
|
| 214 |
+
|
| 215 |
+
Michael Figurnov, Maxwell D Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov. Spatially adaptive computation time for residual networks. In CVPR, 2017.
|
| 216 |
+
|
| 217 |
+
Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015.
|
| 218 |
+
|
| 219 |
+
Shupeng Gui, Haotao Wang, Haichuan Yang, Chen Yu, Zhangyang Wang, and Ji Liu. Model compression with adversarial robustness: A unified optimization framework. In NeurIPS, pp. 1283– 1294, 2019.
|
| 220 |
+
|
| 221 |
+
Yiwen Guo, Chao Zhang, Changshui Zhang, and Yurong Chen. Sparse dnns with improved adversarial robustness. In NeurIPS, 2018.
|
| 222 |
+
|
| 223 |
+
Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In NeurIPS, 2015.
|
| 224 |
+
|
| 225 |
+
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016.
|
| 226 |
+
|
| 227 |
+
Hanzhang Hu, Debadeepta Dey, J. Andrew Bagnell, and Martial Hebert. Anytime neural networks via joint optimization of auxiliary losses. In AAAI, 2019.
|
| 228 |
+
|
| 229 |
+
Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Weinberger. Multi-scale dense networks for resource efficient image classification. In ICLR, 2018.
|
| 230 |
+
|
| 231 |
+
Zehao Huang and Naiyan Wang. Data-driven sparse structure selection for deep neural networks. In ECCV, 2018.
|
| 232 |
+
|
| 233 |
+
Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitras. Shallow-deep networks: Understanding and mitigating network overthinking. In ICML, 2019.
|
| 234 |
+
|
| 235 |
+
Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. In ICLR, 2017.
|
| 236 |
+
|
| 237 |
+
Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In CVPR, 2018.
|
| 238 |
+
|
| 239 |
+
Ji Lin, Chuang Gan, and Song Han. Defensive quantization: When efficiency meets robustness. In ICLR, 2019.
|
| 240 |
+
|
| 241 |
+
Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018.
|
| 242 |
+
|
| 243 |
+
Preetum Nakkiran. Adversarial robustness may be at odds with simplicity. arXiv, 2019.
|
| 244 |
+
|
| 245 |
+
Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In EuroS&P, 2016.
|
| 246 |
+
|
| 247 |
+
Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018.
|
| 248 |
+
|
| 249 |
+
Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarially robust generalization requires more data. In NeurIPS, 2018.
|
| 250 |
+
|
| 251 |
+
Aman Sinha, Hongseok Namkoong, and John Duchi. Certifying Some Distributional Robustness with Principled Adversarial Training. In ICLR, 2018.
|
| 252 |
+
|
| 253 |
+
Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In ICLR, 2018.
|
| 254 |
+
|
| 255 |
+
Thilo Strauss, Markus Hanselmann, Andrej Junginger, and Holger Ulmer. Ensemble methods as a defense to adversarial perturbations against deep neural networks. arXiv, 2017.
|
| 256 |
+
|
| 257 |
+
Ke Sun, Zhanxing Zhu, and Zhouchen Lin. Towards understanding adversarial examples systematically: Exploring data size, task and model factors. arXiv, 2019.
|
| 258 |
+
|
| 259 |
+
Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014.
|
| 260 |
+
|
| 261 |
+
Cheng Tai, Tong Xiao, Yi Zhang, Xiaogang Wang, et al. Convolutional neural networks with lowrank regularization. In ICLR, 2016.
|
| 262 |
+
|
| 263 |
+
Surat Teerapittayanon, Bradley McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural networks. In ICPR, 2017.
|
| 264 |
+
|
| 265 |
+
Florian Tramer, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick Mc-\` Daniel. Ensemble adversarial training: Attacks and defenses. In ICLR, 2018.
|
| 266 |
+
|
| 267 |
+
Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019.
|
| 268 |
+
|
| 269 |
+
Xin Wang, Fisher Yu, Zi-Yi Dou, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in convolutional networks. In ECCV, 2018a.
|
| 270 |
+
|
| 271 |
+
Yue Wang, Tan Nguyen, Yang Zhao, Zhangyang Wang, Yingyan Lin, and Richard Baraniuk. Energynet: Energy-efficient dynamic inference. 2018b.
|
| 272 |
+
|
| 273 |
+
Yue Wang, Jianghao Shen, Ting-Kuei Hu, Pengfei Xu, Tan Nguyen, Richard Baraniuk, Zhangyang Wang, and Yingyan Lin. Dual dynamic inference: Enabling more efficient, adaptive and controllable deep inference. arXiv preprint arXiv:1907.04523, 2019.
|
| 274 |
+
|
| 275 |
+
Yunhe Wang, Chang Xu, Chao Xu, and Dacheng Tao. Adversarial learning of portable student networks. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018c.
|
| 276 |
+
|
| 277 |
+
Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In NeurIPS, 2016.
|
| 278 |
+
|
| 279 |
+
Jiaxiang Wu, Cong Leng, Yuhang Wang, Qinghao Hu, and Jian Cheng. Quantized convolutional neural networks for mobile devices. In CVPR, 2016.
|
| 280 |
+
|
| 281 |
+
Junru Wu, Yue Wang, Zhenyu Wu, Zhangyang Wang, Ashok Veeraraghavan, and Yingyan Lin. Deep $k$ -means: Re-training and parameter sharing with harder cluster assignments for compressing deep convolutions. In ICML, 2018.
|
| 282 |
+
|
| 283 |
+
Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In ICLR, 2018.
|
| 284 |
+
|
| 285 |
+
Donna Xu, Yaxin Shi, Ivor W Tsang, Yew-Soon Ong, Chen Gong, and Xiaobo Shen. A survey on multi-output learning. arXiv, 2019.
|
| 286 |
+
|
| 287 |
+
Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018.
|
| 288 |
+
|
| 289 |
+
Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, and Xue Lin. Adversarial robustness vs model compression, or both? In ICCV, 2019.
|
| 290 |
+
|
| 291 |
+
Xiyu Yu, Tongliang Liu, Xinchao Wang, and Dacheng Tao. On compressing deep models by low rank and sparse decomposition. In CVPR, 2017.
|
| 292 |
+
|
| 293 |
+
Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically principled trade-off between robustness and accuracy. arXiv, 2019.
|
| 294 |
+
|
| 295 |
+
Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018.
|
| 296 |
+
|
| 297 |
+
# A LEARNING DETAILS OF RDI-NETS
|
| 298 |
+
|
| 299 |
+
MNIST We adopt the network architecture from (Chen et al., 2018) with four convolutions and three full-connected layers. We train for 13100 iterations with a batch size of 256. The learning rate is initialized as 0.033 and is lowered by 10 at 12000th and 12900th iteration. For hybrid loss, the weights $\{ w _ { i } \} _ { i = 1 } ^ { N + 1 }$ are set as $\{ 1 , 1 , 1 \}$ for simplicity. For adversarial defense/attack, we perform 40-steps PGD for both defense and evaluation. The perturbation size and step size are set as 0.3 and 0.01.
|
| 300 |
+
|
| 301 |
+
CIFAR-10 We take ResNet-38 and MobileNetV2 as the backbone architectures. For RDIResNet38, we initialize learning rate as 0.1 and decay it by a factor of 10 at 32000th and 48000th iteration. The learning procedure stops at 55000 iteration. For RDI-MobileNetV2, the learning rate is set to 0.05 and is lowered by 10 times at 62000th and 70000th iteration. We stop the learn$\{ w _ { i } \} _ { i = 1 } ^ { N + 1 }$ 76000 iteration. For hybrid loss, we follow tof RDI-ResNet38 and RDI-MobileNetV2 as ectively. For adversarial defense/attack, the $\{ 0 . 5 , 0 . 5 , 0 . 7 , 0 . 7 , 0 . 9 , 0 . 9 , 2 \}$ 019)and are $\{ 0 . 5 , 0 . 5 , 1 \}$
|
| 302 |
+
set as $8 / 2 5 5$ and 2/255. 10-steps PGD is performed for defense and 20-steps PGD is utilized for evaluation.
|
| 303 |
+
|
| 304 |
+
# B NETWORK STRUCTURE OF RDI-NETS
|
| 305 |
+
|
| 306 |
+
To build RDI-Nets, we follow the similar setting in Teerapittayanon et al. (2017) by appending additional branch classifiers at equidistant points throughout a given network, as illustrated in Fig 3, Fig 4 and Fig 5. A few pooling operations, light-weight convolutions and fully-connected layers are appended to each branch classifiers. Note that the extra flops introduced by side branch classifiers are less than $2 \%$ than the original ResNet-38 or MobileNetV2.
|
| 307 |
+
|
| 308 |
+

|
| 309 |
+
Figure 3: Network architecture of RDI-SmallCNN. Two branch classifiers are inserted after 1st convolutional layer and 3rd convolutional layer in the original SmallCNN.
|
| 310 |
+
|
| 311 |
+

|
| 312 |
+
Figure 4: Network architecture of RDI-ResNet38. In each residual block group, two branch classifiers are inserted after $1 { s t }$ residual block and 4th residual block.
|
| 313 |
+
|
| 314 |
+
# C INPUT-ADAPTIVE INFERENCE FOR RDI-NETS
|
| 315 |
+
|
| 316 |
+
Similar to the deterministic strategy in Teerapittayanon et al. (2017), we adopt the entropy as the measure of the prediction confidence. Given a prediction vector $y \in \mathbb { R } ^ { C }$ , where $C$ is the number of
|
| 317 |
+
|
| 318 |
+

|
| 319 |
+
Figure 5: Network architecture of RDI-MobilenetV2. Two branch classfiers are inserted after 3rd inverted residual block and 11th inverted residual block in the orignal MobilenetV2.
|
| 320 |
+
|
| 321 |
+
class, the entropy of $y$ is defined as follow,
|
| 322 |
+
|
| 323 |
+
$$
|
| 324 |
+
- \sum _ { c = 1 } ^ { C } ( y _ { c } + \epsilon ) l o g ( y _ { c } + \epsilon ) ,
|
| 325 |
+
$$
|
| 326 |
+
|
| 327 |
+
where $\epsilon$ is a small positive constant used for robust entropy computation. To perform fast inference on a $( K { + } 1 )$ -output RDI-Net, we need to determine $K$ threshold numbers, i.e., $\{ t _ { i } \} _ { i = 1 } ^ { K }$ , so that the input $x$ will exit at ith branch if the entropy of $y _ { i }$ is larger than $t _ { i }$ . To choose $\{ t _ { i } \} _ { i = 1 } ^ { K }$ , Huang et al. (2018) provides a good starting point by fixing exiting probability of each branch classifiers equally on validation set so that each sample can equally contribute to inference. We follow this strategy but adjust the thresholds to make the contribution of middle branches slightly larger than the early branches. The threshold numbers for RDI-SmallCNN, RDI-ResNet38, and RDI-MobilenetV2 are set to be $\{ 0 . 0 2 3 , 0 . 0 1 4 \}$ , $\{ 0 . 3 2 , 0 . 3 6 , 0 . 3 9 , 0 . 8 3 , 1 . 1 2 , 1 . 3 5 \}$ , and $\{ 0 . 2 6 7 , 0 . 7 6 5 \}$ , respectively.
|
| 328 |
+
|
| 329 |
+

|
| 330 |
+
Figure 6: Performance comparison between RDI-Net and the pruning $^ +$ defense baseline. Each marker represents a model, whose size is proportional to its MFlops. $\gamma$ is the sparsity trade-off parameter: the larger the sparser (smaller model).
|
| 331 |
+
|
| 332 |
+
# D GENERALIZED ROBUSTNESS
|
| 333 |
+
|
| 334 |
+
Here, we introduce the attack form of random attack and report the complete results against FGSM (Goodfellow et al., 2015) and WRM (Sinha et al., 2018) attacker under various attack forms, in Tables 9 and 10, respectively.
|
| 335 |
+
|
| 336 |
+
Random Attack the attack exploits multi-loss flexibility by randomly fusing all $f _ { \theta _ { i } }$ losses. Given a $N$ -output network, we have a fusion vector $C \in \mathbb { R } ^ { N } \sim \mathbf { \bar { \mathbb { D } } }$ , where $\mathbb { D }$ is some distribution (uniform
|
| 337 |
+
|
| 338 |
+
by default). We denote $c _ { j }$ as the $j$ th element of $C$ and $x _ { r d m } ^ { a d v }$ can be found by:
|
| 339 |
+
|
| 340 |
+
$$
|
| 341 |
+
x _ { r d m } ^ { a d v } = \underset { x ^ { \prime } \in | x ^ { \prime } - x | _ { \infty } \leq \epsilon } { \arg \operatorname* { m a x } } | \frac { 1 } { N } \sum _ { j = 1 } ^ { N } c _ { j } \phi ( f _ { \theta _ { j } } ( x ^ { \prime } ) , y ) | .
|
| 342 |
+
$$
|
| 343 |
+
|
| 344 |
+
It is expected to challenge our defense, due to the infinitely many ways of randomly fusing outputs.
|
| 345 |
+
|
| 346 |
+
Table 9: The performance evaluation on RDI-ResNet38 (defended with PGD) against FGSM attack. The perturbation size is 8/255. The ATA of the original defended ResNet38 by PGD under the same attacker is $5 1 . 1 1 \%$ .
|
| 347 |
+
|
| 348 |
+
<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>20.69%</td><td rowspan=1 colspan=1>66.06%</td><td rowspan=1 colspan=1>72.77%</td><td rowspan=1 colspan=1>72.76%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>16.15%</td><td rowspan=1 colspan=1>53.87%</td><td rowspan=1 colspan=1>70.40%</td><td rowspan=1 colspan=1>69.71%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch3)</td><td rowspan=1 colspan=1>8.13%</td><td rowspan=1 colspan=1>63.70%</td><td rowspan=1 colspan=1>64.19%</td><td rowspan=1 colspan=1>65.14%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch4)</td><td rowspan=1 colspan=1>10.09%</td><td rowspan=1 colspan=1>56.67%</td><td rowspan=1 colspan=1>58.45%</td><td rowspan=1 colspan=1>58.20%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch5)</td><td rowspan=1 colspan=1>9.45%</td><td rowspan=1 colspan=1>50.81%</td><td rowspan=1 colspan=1>52.76%</td><td rowspan=1 colspan=1>52.96%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch6)</td><td rowspan=1 colspan=1>10.22%</td><td rowspan=1 colspan=1>50.34%</td><td rowspan=1 colspan=1>53.17%</td><td rowspan=1 colspan=1>51.05%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>11.51%</td><td rowspan=1 colspan=1>51.45%</td><td rowspan=1 colspan=1>53.64%</td><td rowspan=1 colspan=1>54.72%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>11.41%</td><td rowspan=1 colspan=1>50.21%</td><td rowspan=1 colspan=1>51.81%</td><td rowspan=1 colspan=1>53.20%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>52.40%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>51.05%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>65.74</td><td rowspan=1 colspan=1>55.27</td><td rowspan=1 colspan=1>58.27</td><td rowspan=1 colspan=1>59.67</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>17.21%</td><td rowspan=1 colspan=1>30.40%</td><td rowspan=1 colspan=1>26.40%</td><td rowspan=1 colspan=1>24.86%</td></tr></table>
|
| 349 |
+
|
| 350 |
+
Table 10: The performance evaluation on RDI-ResNet38 (defended with PGD) against WRM attack. The perturbation size is 0.3. The ATA of the original defended ResNet38 by PGD under the same attacker is $8 3 . 3 5 \%$ .
|
| 351 |
+
|
| 352 |
+
<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>46.60%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>71.33%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch3)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.41%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch4)</td><td rowspan=1 colspan=1>33.41%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch5)</td><td rowspan=1 colspan=1>42.35%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.41%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch6)</td><td rowspan=1 colspan=1>47.77%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.40%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>34.42%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>26.48%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.40%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>50.05</td><td rowspan=1 colspan=1>50.46</td><td rowspan=1 colspan=1>52.89</td><td rowspan=1 colspan=1>52.38</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>36.98%</td><td rowspan=1 colspan=1>36.46%</td><td rowspan=1 colspan=1>33.40%</td><td rowspan=1 colspan=1>34.04%</td></tr></table>
|
parse/train/rJgzzJHtDB/rJgzzJHtDB_content_list.json
ADDED
|
@@ -0,0 +1,1873 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "TRIPLE WINS: BOOSTING ACCURACY, ROBUSTNESS AND EFFICIENCY TOGETHER BY ENABLING INPUTADAPTIVE INFERENCE ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
176,
|
| 8 |
+
98,
|
| 9 |
+
823,
|
| 10 |
+
171
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Ting-Kuei Hu∗, Tianlong Chen∗, Haotao Wang Zhangyang Wang ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
184,
|
| 19 |
+
194,
|
| 20 |
+
645,
|
| 21 |
+
210
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Department of Computer Science and Engineering Texas A&M University, USA {tkhu,wiwjp619,htwang,atlaswang}@tamu.edu ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
184,
|
| 30 |
+
210,
|
| 31 |
+
584,
|
| 32 |
+
251
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "ABSTRACT ",
|
| 39 |
+
"text_level": 1,
|
| 40 |
+
"bbox": [
|
| 41 |
+
454,
|
| 42 |
+
287,
|
| 43 |
+
544,
|
| 44 |
+
303
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "Deep networks were recently suggested to face the odds between accuracy (on clean natural images) and robustness (on adversarially perturbed images) (Tsipras et al., 2019). Such a dilemma is shown to be rooted in the inherently higher sample complexity (Schmidt et al., 2018) and/or model capacity (Nakkiran, 2019), for learning a high-accuracy and robust classifier. In view of that, give a classification task, growing the model capacity appears to help draw a win-win between accuracy and robustness, yet at the expense of model size and latency, therefore posing challenges for resource-constrained applications. Is it possible to co-design model accuracy, robustness and efficiency to achieve their triple wins? ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
233,
|
| 53 |
+
320,
|
| 54 |
+
764,
|
| 55 |
+
445
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "This paper studies multi-exit networks associated with input-adaptive efficient inference, showing their strong promise in achieving a “sweet point” in cooptimizing model accuracy, robustness and efficiency. Our proposed solution, dubbed Robust Dynamic Inference Networks (RDI-Nets), allows for each input (either clean or adversarial) to adaptively choose one of the multiple output layers (early branches or the final one) to output its prediction. That multi-loss adaptivity adds new variations and flexibility to adversarial attacks and defenses, on which we present a systematical investigation. We show experimentally that by equipping existing backbones with such robust adaptive inference, the resulting RDI-Nets can achieve better accuracy and robustness, yet with over $30 \\%$ computational savings, compared to the defended original models. ",
|
| 62 |
+
"bbox": [
|
| 63 |
+
233,
|
| 64 |
+
449,
|
| 65 |
+
764,
|
| 66 |
+
602
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "1 INTRODUCTION ",
|
| 73 |
+
"text_level": 1,
|
| 74 |
+
"bbox": [
|
| 75 |
+
176,
|
| 76 |
+
619,
|
| 77 |
+
336,
|
| 78 |
+
635
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Deep networks, despite their high predictive accuracy, are notoriously vulnerable to adversarial attacks (Goodfellow et al., 2015; Biggio et al., 2013; Szegedy et al., 2014; Papernot et al., 2016). While many defense methods have been proposed to increase a model’s robustness to adversarial examples, they were typically observed to hamper its accuracy on original clean images. Tsipras et al. (2019) first pointed out the inherent tension between the goals of adversarial robustness and standard accuracy in deep networks, whose provable existence was shown in a simplified setting. Zhang et al. (2019) theoretically quantified the accuracy-robustness trade-off, in terms of the gap between the risk for adversarial examples versus the risk for non-adversarial examples. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
640,
|
| 88 |
+
825,
|
| 89 |
+
751
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "text",
|
| 95 |
+
"text": "It is intriguing to consider whether and why the model accuracy and robustness have to be at odds. Schmidt et al. (2018) demonstrated that the number of samples needed to achieve adversarially robust generalization is polynomially larger than that needed for standard generalization, under the adversarial training setting. A similar conclusion was concurred by Sun et al. (2019) in the standard training setting. Tsipras et al. (2019) considered the accuracy-robustness trade-off as an inherent trait of the data distribution itself, indicating that this phenomenon persists even in the limit of infinite data. Nakkiran (2019) argued from a different perspective, that the complexity (e.g. capacity) of a robust classifier must be higher than that of a standard classifier. Therefore, replacing a largercapacity classifier might effectively alleviate the trade-off. Overall, those existing works appear to suggest that, while accuracy and robustness are likely to trade off for a fixed classification model and on a given dataset, such trade-off might be effectively alleviated (“win-win”), if supplying more training data and/or replacing a larger-capacity classifier. ",
|
| 96 |
+
"bbox": [
|
| 97 |
+
174,
|
| 98 |
+
758,
|
| 99 |
+
825,
|
| 100 |
+
897
|
| 101 |
+
],
|
| 102 |
+
"page_idx": 0
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"type": "text",
|
| 106 |
+
"text": "",
|
| 107 |
+
"bbox": [
|
| 108 |
+
173,
|
| 109 |
+
103,
|
| 110 |
+
823,
|
| 111 |
+
132
|
| 112 |
+
],
|
| 113 |
+
"page_idx": 1
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"type": "text",
|
| 117 |
+
"text": "On a separate note, deep networks also face the pressing challenge to be deployed on resourceconstrained platforms due to the prosperity of smart Internet-of-Things (IoT) devices. Many IoT applications naturally demand security and trustworthiness, e.g., , biometrics and identity verification, but can only afford limited latency, memory and energy budget. Hereby we extend the question: can we achieve a triple-win, i.e., , an accurate and robust classfier while keeping it efficient? ",
|
| 118 |
+
"bbox": [
|
| 119 |
+
174,
|
| 120 |
+
138,
|
| 121 |
+
823,
|
| 122 |
+
208
|
| 123 |
+
],
|
| 124 |
+
"page_idx": 1
|
| 125 |
+
},
|
| 126 |
+
{
|
| 127 |
+
"type": "text",
|
| 128 |
+
"text": "This paper makes an attempt in providing a positive answer to the above question. Rather than proposing a specific design of robust light-weight models, we reduce the average computation loads by input-adaptive routing to achieve triple-win. To this end, we introduce the input-adaptive dynamic inference (Teerapittayanon et al., 2017; Wang et al., 2018a), an emerging efficient inference scheme in contrast to the (non-adaptive) model compression, to the adversarial defense field for the first time. Given any deep network backbone (e.g., , ResNet, MobileNet), we first follow (Teerapittayanon et al., 2017) to augment it with multiple early-branch output layers in addition to the original final output. Each input, regardless of clean or adversarial samples, adaptively chooses which output layer to take for its own prediction. Therefore, a large portion of input inferences can be terminated early when the samples can already be inferred with high confidence. ",
|
| 129 |
+
"bbox": [
|
| 130 |
+
173,
|
| 131 |
+
215,
|
| 132 |
+
825,
|
| 133 |
+
354
|
| 134 |
+
],
|
| 135 |
+
"page_idx": 1
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"type": "text",
|
| 139 |
+
"text": "Up to our best knowledge, no existing work studied adversarial attacks and defenses for an adaptive multi-output model, as the multiple sources of losses provide much larger flexibility to compose attacks (and therefore defenses), compared to the typical single-loss backbone. We present a systematical exploration on how to (white-box) attack and defense our proposed multi-output network with adaptive inference, demonstrating that the composition of multiple-loss information is critical in making the attack/defense strong. Fig. 1 illustrates our proposed Robust Dynamic Inference Networks (RDI-Nets). We show experimentally that the input-adaptive inference and multiloss flexibility can be our friend in achieving the desired “triple wins”. With our best defended RDI-Nets, we achieve better accuracy and robustness, yet with over $30 \\%$ inference computational savings, compared to the defended original models as well as existing solutions co-designing robustness and efficiency (Gui et al., 2019; Guo et al., 2018). The codes can be referenced from https://github.com/TAMU-VITA/triple-wins. ",
|
| 140 |
+
"bbox": [
|
| 141 |
+
174,
|
| 142 |
+
361,
|
| 143 |
+
825,
|
| 144 |
+
527
|
| 145 |
+
],
|
| 146 |
+
"page_idx": 1
|
| 147 |
+
},
|
| 148 |
+
{
|
| 149 |
+
"type": "image",
|
| 150 |
+
"img_path": "images/4b918d7859876af3bc28f2b330882c0236a965a6486f21219ecdc2f03aba9340.jpg",
|
| 151 |
+
"image_caption": [
|
| 152 |
+
"Figure 1: Our proposed RDI-Net framework, a defended multi-output network enabling dynamic inference. Each image, being it clean or adversarially perturbed, adaptively picks one branch to exit. "
|
| 153 |
+
],
|
| 154 |
+
"image_footnote": [],
|
| 155 |
+
"bbox": [
|
| 156 |
+
236,
|
| 157 |
+
549,
|
| 158 |
+
746,
|
| 159 |
+
698
|
| 160 |
+
],
|
| 161 |
+
"page_idx": 1
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"type": "text",
|
| 165 |
+
"text": "2 RELATED WORK ",
|
| 166 |
+
"text_level": 1,
|
| 167 |
+
"bbox": [
|
| 168 |
+
176,
|
| 169 |
+
761,
|
| 170 |
+
344,
|
| 171 |
+
776
|
| 172 |
+
],
|
| 173 |
+
"page_idx": 1
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"type": "text",
|
| 177 |
+
"text": "2.1 ADVERSARIAL DEFENSE ",
|
| 178 |
+
"text_level": 1,
|
| 179 |
+
"bbox": [
|
| 180 |
+
176,
|
| 181 |
+
790,
|
| 182 |
+
387,
|
| 183 |
+
804
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 1
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "text",
|
| 189 |
+
"text": "A magnitude of defend approaches have been proposed (Kurakin et al., 2017; Xu et al., 2018; Song et al., 2018; Liao et al., 2018), although many were quickly evaded by new attacks (Carlini & Wagner, 2017; Baluja & Fischer, 2018). One strong defense algorithm that has so far not been fully compromised is adversarial training (Madry et al., 2018). It searches for adversarial images to augment the training procedure, although at the price of higher training costs (but not affecting inference efficiency). However, almost all existing attacks and defenses focus on a single-output classification (or other task) model. We are unaware of prior studies directly addressing attacks/defenses to more complicated networks with multiple possible outputs. ",
|
| 190 |
+
"bbox": [
|
| 191 |
+
174,
|
| 192 |
+
811,
|
| 193 |
+
825,
|
| 194 |
+
924
|
| 195 |
+
],
|
| 196 |
+
"page_idx": 1
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"type": "text",
|
| 200 |
+
"text": "One related row of works are to exploit model ensemble (Tramer et al. \\` , 2018; Strauss et al., 2017) in adversarial training. The gains of the defended ensemble compared to a single model could be viewed as the benefits of either the benefits of diversity (generating stronger and more transferable perturbations), or the increasing model capacity (consider the ensembled multiple models as a compound one). Unfortunately, ensemble methods could amplify the inference complexity and be detrimental for efficiency. Besides, it is also known that injecting randomization at inference time helps mitigate adversarial effects (Xie et al., 2018; Cohen et al., 2019). Yet up to our best knowledge, no work has studied non-random, but rather input-dependent inference for defense. ",
|
| 201 |
+
"bbox": [
|
| 202 |
+
173,
|
| 203 |
+
103,
|
| 204 |
+
825,
|
| 205 |
+
215
|
| 206 |
+
],
|
| 207 |
+
"page_idx": 2
|
| 208 |
+
},
|
| 209 |
+
{
|
| 210 |
+
"type": "text",
|
| 211 |
+
"text": "2.2 EFFICIENT INFERENCE ",
|
| 212 |
+
"text_level": 1,
|
| 213 |
+
"bbox": [
|
| 214 |
+
176,
|
| 215 |
+
228,
|
| 216 |
+
372,
|
| 217 |
+
243
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "Research in improving deep network efficiency could be categorized into two streams: the static way that designs compact models or compresses heavy models, while the compact/compressed models remain fixed for all inputs at inference; and the dynamic way, that at inference the inputs can choose different computational paths adaptively, and the simpler inputs usually take less computation to make predictions. We briefly review the literature below. ",
|
| 224 |
+
"bbox": [
|
| 225 |
+
174,
|
| 226 |
+
250,
|
| 227 |
+
825,
|
| 228 |
+
319
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "Static: Compact Network Design and Model Compression. Many compact architectures have been specifically designed for resource-constrained applications, by adopting lightweight depthwise convolutions (Sandler et al., 2018), and group-wise convolutions with channel-shuffling (Zhang et al., 2018), to just name a few. For model compression, Han et al. (2015) first proposed to sparsify deep models by removing non-significant synapses and then re-training to restore performance. Structured pruning was later on introduced for more hardware friendliness (Wen et al., 2016). Layer factorization (Tai et al., 2016; Yu et al., 2017), quantization (Wu et al., 2016), model distillation (Wang et al., 2018c) and weight sharing (Wu et al., 2018) have also been respectively found effective. ",
|
| 235 |
+
"bbox": [
|
| 236 |
+
174,
|
| 237 |
+
327,
|
| 238 |
+
825,
|
| 239 |
+
438
|
| 240 |
+
],
|
| 241 |
+
"page_idx": 2
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"type": "text",
|
| 245 |
+
"text": "Dynamic: Input-Adaptive Inference. Higher inference efficiency could be also accomplished by enabling input-conditional execution. Teerapittayanon et al. (2017); Huang et al. (2018); Kaya et al. (2019) leveraged intermediate features to augment multiple side branch classifiers to enable early predictions. Their methodology sets up the foundation for our work. Other efforts (Figurnov et al., 2017; Wang et al., 2018a;b; 2019) allow for an input to choose between passing through or skipping each layer. The approach could be integrated with RDI-Nets too, which we leave as future work. ",
|
| 246 |
+
"bbox": [
|
| 247 |
+
174,
|
| 248 |
+
445,
|
| 249 |
+
825,
|
| 250 |
+
529
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 2
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "text",
|
| 256 |
+
"text": "2.3 BRIDGING ROBUSTNESS WITH EFFICIENCY ",
|
| 257 |
+
"text_level": 1,
|
| 258 |
+
"bbox": [
|
| 259 |
+
174,
|
| 260 |
+
542,
|
| 261 |
+
514,
|
| 262 |
+
558
|
| 263 |
+
],
|
| 264 |
+
"page_idx": 2
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"type": "text",
|
| 268 |
+
"text": "A few studies recently try to link deep learning robustness and efficiency. Guo et al. (2018) observed that in a sparse deep network, appropriately sparsified weights improve robustness, whereas over-sparsification (e.g., less than $5 \\%$ nonzero weights) in turn makes the model more fragile. Two latest works (Ye et al., 2019; Gui et al., 2019) examined the robustness of compressed models, and concluded similar observations that the relationship between mode size and robustness depends on compression methods and are often non-monotonic. Lin et al. (2019) found that activation quantization may hurt robustness, but can be turned into effective defense if enforcing continuity constraints. ",
|
| 269 |
+
"bbox": [
|
| 270 |
+
174,
|
| 271 |
+
564,
|
| 272 |
+
825,
|
| 273 |
+
661
|
| 274 |
+
],
|
| 275 |
+
"page_idx": 2
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"type": "text",
|
| 279 |
+
"text": "Different from above methods that tackle robustness from static compact/compressed models, the proposed RDI-Nets are the first to address robustness from the dynamic input-adaptive inference. Our experiment results demonstrate the consistent superiority of RDI-Nets over those static methods (Section 4.3). Moreover, applying dynamic inference top of those static methods may further boost the robustness and efficiency, which we leave as future work. ",
|
| 280 |
+
"bbox": [
|
| 281 |
+
174,
|
| 282 |
+
667,
|
| 283 |
+
825,
|
| 284 |
+
738
|
| 285 |
+
],
|
| 286 |
+
"page_idx": 2
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"type": "text",
|
| 290 |
+
"text": "3 APPROACH",
|
| 291 |
+
"text_level": 1,
|
| 292 |
+
"bbox": [
|
| 293 |
+
174,
|
| 294 |
+
756,
|
| 295 |
+
299,
|
| 296 |
+
771
|
| 297 |
+
],
|
| 298 |
+
"page_idx": 2
|
| 299 |
+
},
|
| 300 |
+
{
|
| 301 |
+
"type": "text",
|
| 302 |
+
"text": "With the goal of achieving inference efficiency, we first look at the setting of multi-output networks and the specific design of RDI-Net in Section 3.1. Then we define three forms of adversarial attacks for multi-output networks in Section 3.2 and their corresponding defense methods in Section 3.3. ",
|
| 303 |
+
"bbox": [
|
| 304 |
+
176,
|
| 305 |
+
782,
|
| 306 |
+
825,
|
| 307 |
+
825
|
| 308 |
+
],
|
| 309 |
+
"page_idx": 2
|
| 310 |
+
},
|
| 311 |
+
{
|
| 312 |
+
"type": "text",
|
| 313 |
+
"text": "Note that RDI-Nets achieve “triple wins”via reducing the average computation loads through inputadaptive routing. It is not to be confused with any specifically-designed robust light-weight model. ",
|
| 314 |
+
"bbox": [
|
| 315 |
+
176,
|
| 316 |
+
832,
|
| 317 |
+
820,
|
| 318 |
+
861
|
| 319 |
+
],
|
| 320 |
+
"page_idx": 2
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"type": "text",
|
| 324 |
+
"text": "3.1 DESIGNING RDI-NETS FOR HIGHER INFERENCE EFFICIENCY",
|
| 325 |
+
"text_level": 1,
|
| 326 |
+
"bbox": [
|
| 327 |
+
174,
|
| 328 |
+
873,
|
| 329 |
+
638,
|
| 330 |
+
888
|
| 331 |
+
],
|
| 332 |
+
"page_idx": 2
|
| 333 |
+
},
|
| 334 |
+
{
|
| 335 |
+
"type": "text",
|
| 336 |
+
"text": "Given an input image $x$ , an $N$ -output network can produce a set of predictions $[ \\hat { y _ { 1 } } , . . . , \\hat { y _ { N } } ]$ by a set of transformations $[ f _ { \\theta _ { 1 } } ( \\cdot ) , . . . , f _ { \\theta _ { N } } ( \\cdot ) ]$ . $\\theta _ { i }$ denote the model parameter of $f _ { \\theta _ { i } }$ , $i = 1 , . . . , N$ , and $f _ { \\theta _ { i } } \\mathbf { s }$ will typically share some weights. With an input $x$ , one can express $\\hat { y } _ { i } = f _ { \\boldsymbol { \\theta } _ { i } } ( \\boldsymbol { x } )$ . We assume that the final prediction will be one chosen (NOT fused) from $[ \\hat { y _ { 1 } } , . . . , \\hat { y _ { N } } ]$ via some deterministic strategy. ",
|
| 337 |
+
"bbox": [
|
| 338 |
+
173,
|
| 339 |
+
895,
|
| 340 |
+
826,
|
| 341 |
+
924
|
| 342 |
+
],
|
| 343 |
+
"page_idx": 2
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"type": "text",
|
| 347 |
+
"text": "",
|
| 348 |
+
"bbox": [
|
| 349 |
+
171,
|
| 350 |
+
103,
|
| 351 |
+
823,
|
| 352 |
+
133
|
| 353 |
+
],
|
| 354 |
+
"page_idx": 3
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"type": "text",
|
| 358 |
+
"text": "We now look at RDI-Nets as a specific instance of multi-output networks, specifically designed for the goal of more efficient, input-adaptive inference. As shown in Fig. 1, for any deep network (e.g., , ResNet, MobileNet), we could append $K$ side branches (with negligible overhead) to allow for early-exit predictions. In other words, it becomes a $( K + 1 )$ -output network, and the subneworks with the $K + 1$ exits, from the lowest to the highest (the original final output), correspond to $[ f _ { \\theta _ { 1 } } ( \\cdot ) , . . . , f _ { \\theta _ { K + 1 } } ( \\cdot ) ]$ . They share their weights in a nested fashion: ${ \\theta } _ { 1 } \\subseteq { \\theta } _ { 2 } . . . \\subseteq { \\theta } _ { K + 1 }$ , with $\\theta _ { K + 1 }$ including the entire network’s parameters. ",
|
| 359 |
+
"bbox": [
|
| 360 |
+
174,
|
| 361 |
+
138,
|
| 362 |
+
825,
|
| 363 |
+
237
|
| 364 |
+
],
|
| 365 |
+
"page_idx": 3
|
| 366 |
+
},
|
| 367 |
+
{
|
| 368 |
+
"type": "text",
|
| 369 |
+
"text": "Our deterministic strategy in selecting one final output follows (Teerapittayanon et al., 2017). We set a confidence threshold $t _ { k }$ for each $k$ -th exit, $k = 1 , . . . , K + 1$ , and each input $x$ will terminate inference and output its prediction in the earliest exit (smallest $k$ ), whose softmax entropy (as a confidence measure) falls below $t _ { k }$ . All computations after the $k$ -th exit will not be activated for this $x$ . Such a progressive and early-halting mechanism effectively saves unnecessary computation for most easier-to-classify samples, and applies in both training and inference. Note that, if efficiency is not the concern, instead of choosing (the earliest one), we could have designed an adaptive or randomized fusion of all $f _ { \\theta _ { i } }$ predictions: but that falls beyond the goal of this work. ",
|
| 370 |
+
"bbox": [
|
| 371 |
+
173,
|
| 372 |
+
242,
|
| 373 |
+
826,
|
| 374 |
+
356
|
| 375 |
+
],
|
| 376 |
+
"page_idx": 3
|
| 377 |
+
},
|
| 378 |
+
{
|
| 379 |
+
"type": "text",
|
| 380 |
+
"text": "The training objective for RDI-Nets could be written as ",
|
| 381 |
+
"bbox": [
|
| 382 |
+
174,
|
| 383 |
+
361,
|
| 384 |
+
539,
|
| 385 |
+
376
|
| 386 |
+
],
|
| 387 |
+
"page_idx": 3
|
| 388 |
+
},
|
| 389 |
+
{
|
| 390 |
+
"type": "equation",
|
| 391 |
+
"img_path": "images/1534c8d3e104be71dfc694d6f7968441bd16bd4ce354a5ff180fc2038c7cac9c.jpg",
|
| 392 |
+
"text": "$$\nL _ { R D I } = \\sum _ { i = 1 } ^ { K + 1 } w _ { i } [ ( \\phi ( f _ { i } ( \\theta _ { i } | x ) , y ) + \\phi ( f _ { i } ( \\theta _ { i } | x ^ { a d v } ) , y ) ] ,\n$$",
|
| 393 |
+
"text_format": "latex",
|
| 394 |
+
"bbox": [
|
| 395 |
+
320,
|
| 396 |
+
383,
|
| 397 |
+
676,
|
| 398 |
+
428
|
| 399 |
+
],
|
| 400 |
+
"page_idx": 3
|
| 401 |
+
},
|
| 402 |
+
{
|
| 403 |
+
"type": "text",
|
| 404 |
+
"text": "For each exit loss, we minimize a hybrid loss of accuracy (on clean $x$ ) and robustness (on $x ^ { a d v }$ ). The K + 1 exits are balanced with a group of weights {wi}K+1i=1 . More details about RDI-Net structures, hyperparameters, and inference branch selection can be founded in Appendix A, B, and C. ",
|
| 405 |
+
"bbox": [
|
| 406 |
+
173,
|
| 407 |
+
438,
|
| 408 |
+
825,
|
| 409 |
+
482
|
| 410 |
+
],
|
| 411 |
+
"page_idx": 3
|
| 412 |
+
},
|
| 413 |
+
{
|
| 414 |
+
"type": "text",
|
| 415 |
+
"text": "In what follows, we discuss three ways to generate $x ^ { a d v }$ in RDI-Nets, and then their defenses. ",
|
| 416 |
+
"bbox": [
|
| 417 |
+
173,
|
| 418 |
+
487,
|
| 419 |
+
785,
|
| 420 |
+
503
|
| 421 |
+
],
|
| 422 |
+
"page_idx": 3
|
| 423 |
+
},
|
| 424 |
+
{
|
| 425 |
+
"type": "text",
|
| 426 |
+
"text": "3.2 THREE ATTACK FORMS ON MULTI-OUTPUT NETWORKS ",
|
| 427 |
+
"text_level": 1,
|
| 428 |
+
"bbox": [
|
| 429 |
+
173,
|
| 430 |
+
515,
|
| 431 |
+
604,
|
| 432 |
+
530
|
| 433 |
+
],
|
| 434 |
+
"page_idx": 3
|
| 435 |
+
},
|
| 436 |
+
{
|
| 437 |
+
"type": "text",
|
| 438 |
+
"text": "We consider white box attacks in this paper. Attackers have access to the model’s parameters, and aim to generate an adversarial image $x ^ { a d v }$ to fool the model by perturbing an input $x$ within a given magnitude bound. ",
|
| 439 |
+
"bbox": [
|
| 440 |
+
174,
|
| 441 |
+
535,
|
| 442 |
+
825,
|
| 443 |
+
578
|
| 444 |
+
],
|
| 445 |
+
"page_idx": 3
|
| 446 |
+
},
|
| 447 |
+
{
|
| 448 |
+
"type": "text",
|
| 449 |
+
"text": "We next discuss three attack forms for an $N$ -output network. Note that they are independent of, and to be distinguished from attacker algorithms (e.g., , PGD, C&W, FGSM): the former depicts the optimization formulation, that can be solved any of the attacker algorithms. ",
|
| 450 |
+
"bbox": [
|
| 451 |
+
174,
|
| 452 |
+
584,
|
| 453 |
+
825,
|
| 454 |
+
627
|
| 455 |
+
],
|
| 456 |
+
"page_idx": 3
|
| 457 |
+
},
|
| 458 |
+
{
|
| 459 |
+
"type": "text",
|
| 460 |
+
"text": "Single Attack Naively extending from attacking single-output networks, a single attack is defined to maximally fool one $f _ { \\theta _ { i } } ( \\cdot )$ only, expressed as: ",
|
| 461 |
+
"bbox": [
|
| 462 |
+
173,
|
| 463 |
+
637,
|
| 464 |
+
823,
|
| 465 |
+
666
|
| 466 |
+
],
|
| 467 |
+
"page_idx": 3
|
| 468 |
+
},
|
| 469 |
+
{
|
| 470 |
+
"type": "equation",
|
| 471 |
+
"img_path": "images/0f12a75d30067fa9a202d696ac6091633593c59ab930acb6e67621402566d256.jpg",
|
| 472 |
+
"text": "$$\nx _ { i } ^ { a d v } = \\underset { x ^ { \\prime } \\in | x ^ { \\prime } - x | _ { \\infty } \\leq \\epsilon } { \\arg \\operatorname* { m a x } } | \\phi ( f _ { \\theta _ { i } } ( x ^ { \\prime } ) , y ) | ,\n$$",
|
| 473 |
+
"text_format": "latex",
|
| 474 |
+
"bbox": [
|
| 475 |
+
377,
|
| 476 |
+
675,
|
| 477 |
+
619,
|
| 478 |
+
705
|
| 479 |
+
],
|
| 480 |
+
"page_idx": 3
|
| 481 |
+
},
|
| 482 |
+
{
|
| 483 |
+
"type": "text",
|
| 484 |
+
"text": "where $y$ is the ground truth label, and $\\phi$ is the loss for $f _ { \\theta _ { i } }$ (we assume softmax for all). $\\epsilon$ is the perturbation radius and we adopt $\\ell _ { \\infty }$ ball for an empirically strong attacker. Naturally, an $N$ -output network can have $N$ different single attacks. However, each single attack is derived without being aware of other parallel outputs. The found $x _ { i } ^ { a d v }$ is not necessarily transferable to other $f _ { \\boldsymbol { \\theta } _ { j } } \\mathbf { s } \\left( j \\neq i \\right)$ , and therefore can be easily bypassed if $x$ is re-routed through other outputs to make its prediction. ",
|
| 485 |
+
"bbox": [
|
| 486 |
+
174,
|
| 487 |
+
713,
|
| 488 |
+
825,
|
| 489 |
+
785
|
| 490 |
+
],
|
| 491 |
+
"page_idx": 3
|
| 492 |
+
},
|
| 493 |
+
{
|
| 494 |
+
"type": "text",
|
| 495 |
+
"text": "Average Attack Our second attack maximizes the average of all $f _ { \\theta _ { i } }$ losses, so that the found $x ^ { a d v }$ remains in effect no matter which one $f _ { \\theta _ { i } }$ is chosen to output the prediction for $x$ : ",
|
| 496 |
+
"bbox": [
|
| 497 |
+
174,
|
| 498 |
+
794,
|
| 499 |
+
820,
|
| 500 |
+
823
|
| 501 |
+
],
|
| 502 |
+
"page_idx": 3
|
| 503 |
+
},
|
| 504 |
+
{
|
| 505 |
+
"type": "equation",
|
| 506 |
+
"img_path": "images/26003983610aaf33834be5c287c8bb41a61b05533d04d50bf1e2dc3efc28d944.jpg",
|
| 507 |
+
"text": "$$\nx _ { a v g } ^ { a d v } = \\underset { x ^ { \\prime } \\in | x ^ { \\prime } - x | _ { \\infty } \\leq \\epsilon } { \\arg \\operatorname* { m a x } } | \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\phi ( f _ { \\theta _ { j } } ( x ^ { \\prime } ) , y ) | ,\n$$",
|
| 508 |
+
"text_format": "latex",
|
| 509 |
+
"bbox": [
|
| 510 |
+
352,
|
| 511 |
+
832,
|
| 512 |
+
643,
|
| 513 |
+
877
|
| 514 |
+
],
|
| 515 |
+
"page_idx": 3
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"type": "text",
|
| 519 |
+
"text": "The average attack addresses takes into account the attack transferablity and involves all $\\theta _ { j } \\mathbf { s }$ into optimization. However, while only one output will be selected for each sample at inference, the average strategy might weaken the individual defense strength of each $f _ { \\theta _ { i } }$ . ",
|
| 520 |
+
"bbox": [
|
| 521 |
+
174,
|
| 522 |
+
883,
|
| 523 |
+
825,
|
| 524 |
+
928
|
| 525 |
+
],
|
| 526 |
+
"page_idx": 3
|
| 527 |
+
},
|
| 528 |
+
{
|
| 529 |
+
"type": "text",
|
| 530 |
+
"text": "Max-Average Attack Our third attack aims to emphasize individual output defense strength, more than simply maximizing an all-averaged loss. We first solve the $N$ single attacks $x _ { i } ^ { a d v }$ as described in Eqn. 2, and denote their collection as $\\Omega$ . We then solve the max-average attack via the following: ",
|
| 531 |
+
"bbox": [
|
| 532 |
+
173,
|
| 533 |
+
103,
|
| 534 |
+
826,
|
| 535 |
+
146
|
| 536 |
+
],
|
| 537 |
+
"page_idx": 4
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"type": "equation",
|
| 541 |
+
"img_path": "images/e97970b19e3ceb70e93975fd631cc746e864d0cc194361cce628a4dfa6044609.jpg",
|
| 542 |
+
"text": "$$\ni ^ { * } = \\arg \\operatorname* { m a x } _ { i } | \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } \\phi ( f _ { \\theta _ { j } } ( x _ { i } ^ { a d v } ) , y ) | .\n$$",
|
| 543 |
+
"text_format": "latex",
|
| 544 |
+
"bbox": [
|
| 545 |
+
486,
|
| 546 |
+
150,
|
| 547 |
+
748,
|
| 548 |
+
195
|
| 549 |
+
],
|
| 550 |
+
"page_idx": 4
|
| 551 |
+
},
|
| 552 |
+
{
|
| 553 |
+
"type": "text",
|
| 554 |
+
"text": "Note Eqn. 4 differs from Eqn. 3 by adding an $\\Omega$ constraint to balance between “commodity” and “specificity”. The found $\\hat { x } _ { m a x } ^ { a d v }$ both strongly increases the averaged loss values from all $f _ { i } \\mathbf { s }$ (therefore possessing transferablity), and maximally fools one individual $f _ { \\boldsymbol { \\theta } _ { i } } \\mathbf { s }$ as it is selected from the collection $\\Omega$ of single attacks. ",
|
| 555 |
+
"bbox": [
|
| 556 |
+
173,
|
| 557 |
+
204,
|
| 558 |
+
825,
|
| 559 |
+
261
|
| 560 |
+
],
|
| 561 |
+
"page_idx": 4
|
| 562 |
+
},
|
| 563 |
+
{
|
| 564 |
+
"type": "text",
|
| 565 |
+
"text": "3.3 DEFENCE ON MULTI-OUTPUT NETWORKS ",
|
| 566 |
+
"text_level": 1,
|
| 567 |
+
"bbox": [
|
| 568 |
+
173,
|
| 569 |
+
271,
|
| 570 |
+
508,
|
| 571 |
+
285
|
| 572 |
+
],
|
| 573 |
+
"page_idx": 4
|
| 574 |
+
},
|
| 575 |
+
{
|
| 576 |
+
"type": "text",
|
| 577 |
+
"text": "For simplicity and fair comparison, we focus on adversarial training (Madry et al., 2018) as our defense framework, where the three above defined attack forms can be plugged-in to generate adversarial images to augment training, as follows ( $\\Theta$ is the union of learnable parameters): ",
|
| 578 |
+
"bbox": [
|
| 579 |
+
174,
|
| 580 |
+
291,
|
| 581 |
+
825,
|
| 582 |
+
333
|
| 583 |
+
],
|
| 584 |
+
"page_idx": 4
|
| 585 |
+
},
|
| 586 |
+
{
|
| 587 |
+
"type": "equation",
|
| 588 |
+
"img_path": "images/d31b20c698ce1b0dc4fac3e593a1e0ec703d349a996301de75ef45554fd059cf.jpg",
|
| 589 |
+
"text": "$$\n\\Xi \\Theta , \\mathrm { w h e r e } \\theta _ { i } = \\underset { \\theta ^ { \\prime } } { \\arg \\operatorname* { m i n } } | \\phi ( f _ { \\theta _ { i } } ( x ) , y ) + \\phi ( f _ { \\theta _ { i } } ( x ^ { a d v } ) , y ) | . ,\n$$",
|
| 590 |
+
"text_format": "latex",
|
| 591 |
+
"bbox": [
|
| 592 |
+
316,
|
| 593 |
+
337,
|
| 594 |
+
700,
|
| 595 |
+
363
|
| 596 |
+
],
|
| 597 |
+
"page_idx": 4
|
| 598 |
+
},
|
| 599 |
+
{
|
| 600 |
+
"type": "text",
|
| 601 |
+
"text": "where $x ^ { a d v } \\in \\{ x _ { i } ^ { a d v } , x _ { a v g } ^ { a d v } , x _ { m a x } ^ { a d v } \\} { } _ { }$ . As $f _ { i } \\mathbf { s }$ partially share their weights $\\theta _ { i }$ in a multi-output network, the updates from different $f _ { i } \\mathbf { s }$ will be averaged on the shared parameters. ",
|
| 602 |
+
"bbox": [
|
| 603 |
+
174,
|
| 604 |
+
367,
|
| 605 |
+
823,
|
| 606 |
+
398
|
| 607 |
+
],
|
| 608 |
+
"page_idx": 4
|
| 609 |
+
},
|
| 610 |
+
{
|
| 611 |
+
"type": "text",
|
| 612 |
+
"text": "4 EXPERIMENTAL RESULTS ",
|
| 613 |
+
"text_level": 1,
|
| 614 |
+
"bbox": [
|
| 615 |
+
176,
|
| 616 |
+
405,
|
| 617 |
+
419,
|
| 618 |
+
421
|
| 619 |
+
],
|
| 620 |
+
"page_idx": 4
|
| 621 |
+
},
|
| 622 |
+
{
|
| 623 |
+
"type": "text",
|
| 624 |
+
"text": "4.1 EXPERIMENTAL SETUP ",
|
| 625 |
+
"bbox": [
|
| 626 |
+
174,
|
| 627 |
+
429,
|
| 628 |
+
377,
|
| 629 |
+
444
|
| 630 |
+
],
|
| 631 |
+
"page_idx": 4
|
| 632 |
+
},
|
| 633 |
+
{
|
| 634 |
+
"type": "text",
|
| 635 |
+
"text": "Evaluation Metrics We evaluate accuracy, robustness, and efficiency, using the metrics below: ",
|
| 636 |
+
"bbox": [
|
| 637 |
+
171,
|
| 638 |
+
449,
|
| 639 |
+
808,
|
| 640 |
+
464
|
| 641 |
+
],
|
| 642 |
+
"page_idx": 4
|
| 643 |
+
},
|
| 644 |
+
{
|
| 645 |
+
"type": "text",
|
| 646 |
+
"text": "• Testing Accuracy (TA): the classification accuracy on the original clean test set. • Adversarial Testing Accuracy (ATA): Given an attacker, ATA stands for the classification accuracy on the attacked test set. It is the same as the “robust accuracy” in (Zhang et al., 2019). • Mega Flops (MFlops): The number of million floating-point multiplication operations consumed on the inference, averaged over the entire testing set. ",
|
| 647 |
+
"bbox": [
|
| 648 |
+
181,
|
| 649 |
+
465,
|
| 650 |
+
825,
|
| 651 |
+
541
|
| 652 |
+
],
|
| 653 |
+
"page_idx": 4
|
| 654 |
+
},
|
| 655 |
+
{
|
| 656 |
+
"type": "text",
|
| 657 |
+
"text": "Datasets and Benchmark Models We evaluate three representative CNN models on two popular datasets: SmallCNN on MNIST (Chen et al., 2018); ResNet-38 (He et al., 2016) and MobileNet-V2 (Sandler et al., 2018) on CIFAR-10. The three networks span from simplest to more complicated, and covers a compact backbone. All three models are defended by adversarial training, constituting strong baselines. Table 1 reports the models, datasets, the attacker algorithm used in attack & defense, and thee TA/ATA/MFlops performance of three defended models. ",
|
| 658 |
+
"bbox": [
|
| 659 |
+
173,
|
| 660 |
+
545,
|
| 661 |
+
825,
|
| 662 |
+
628
|
| 663 |
+
],
|
| 664 |
+
"page_idx": 4
|
| 665 |
+
},
|
| 666 |
+
{
|
| 667 |
+
"type": "text",
|
| 668 |
+
"text": "Attack and Defense on RDI-Nets We build RDI-Nets by appending side branch outputs for each backbone. For SmallCNN, we add two side branches $K = 2$ ). For ResNet-38 and MobileNet-V2, we have $K = 6$ and $K = 2$ , respectively. The branches are designed to cause negligible overheads: more details of their structure and positions can be referenced in Appendix B. We call those result models RDI-SmallCNN, RDI-ResNet38 and RDI-MobileNetV2 hereinafter. ",
|
| 669 |
+
"bbox": [
|
| 670 |
+
174,
|
| 671 |
+
636,
|
| 672 |
+
825,
|
| 673 |
+
707
|
| 674 |
+
],
|
| 675 |
+
"page_idx": 4
|
| 676 |
+
},
|
| 677 |
+
{
|
| 678 |
+
"type": "text",
|
| 679 |
+
"text": "We then generate attacks using our three defined forms. Each attack form could be solved with various attacker algorithms (e.g., PGD, C&W, FGSM), and by default we solve it with the same attacker used for each backbone in Table 1. If we fix one attacker algorithm (e.g., PGD), then TA/ATA for a single-output network can be measured without ambiguity. Yet for $( K { + } 1 )$ -output RDI-Nets, there could be at least $K { + 3 }$ different ATA numbers for one defended model, depending on what attack form in Section 3.1 to apply $\\scriptstyle { K + 1 }$ single attacks, 1 average attack, and 1 maxaverage attack). For example, we denote by ATA (Branch1) the ATA number when applying the single attack generated from the first side output branch (e.g., $x _ { 1 } ^ { a d v } .$ ); similarly elsewhere. ",
|
| 680 |
+
"bbox": [
|
| 681 |
+
173,
|
| 682 |
+
713,
|
| 683 |
+
825,
|
| 684 |
+
827
|
| 685 |
+
],
|
| 686 |
+
"page_idx": 4
|
| 687 |
+
},
|
| 688 |
+
{
|
| 689 |
+
"type": "text",
|
| 690 |
+
"text": "We also defend RDI-Nets using adversarial training, using the forms of adversarial images to augment training. By default, we adopt three adversarial training defense schemes: Main Branch (single attack using $x _ { K + 1 } ^ { a d v } ) ^ { 1 }$ 1,Average (using $x _ { a v g . } ^ { a d v . }$ ),and Max-Average (using $x _ { a v g } ^ { m a x }$ ), in addition to the undefended RDI-Nets (using standard training) denoted as Standard. ",
|
| 691 |
+
"bbox": [
|
| 692 |
+
174,
|
| 693 |
+
832,
|
| 694 |
+
825,
|
| 695 |
+
888
|
| 696 |
+
],
|
| 697 |
+
"page_idx": 4
|
| 698 |
+
},
|
| 699 |
+
{
|
| 700 |
+
"type": "text",
|
| 701 |
+
"text": "We cross evaluate ATAs of different defenses and attacks, since an ideal defense shall protect against all possible attack forms. To faithfully indicate the actual robustness, we choose the lowest number among all $K + 3$ ATAs, denoted as ATA (Worst-Case), as the robustness measure for an RDI-Net. ",
|
| 702 |
+
"bbox": [
|
| 703 |
+
174,
|
| 704 |
+
103,
|
| 705 |
+
823,
|
| 706 |
+
146
|
| 707 |
+
],
|
| 708 |
+
"page_idx": 5
|
| 709 |
+
},
|
| 710 |
+
{
|
| 711 |
+
"type": "table",
|
| 712 |
+
"img_path": "images/85bcd4c0df192db7f050451a9e7a7443020f43dfd8f3a0715eb8aeca89ba742e.jpg",
|
| 713 |
+
"table_caption": [
|
| 714 |
+
"Table 1: Benchmarking results of adverserial training of three networks. PGD-40 denotes running the projected gradient descent attacker (Madry et al., 2018) for 40 iterations. We set the perturbation size as 0.3 for MNIST and 8/255 for CIFAR-10 in $\\ell _ { \\infty }$ norm (adopted by all following experiments). "
|
| 715 |
+
],
|
| 716 |
+
"table_footnote": [],
|
| 717 |
+
"table_body": "<table><tr><td>Model</td><td>Dataset</td><td>Defend</td><td>Attack</td><td>TA</td><td>ATA</td><td>MFlops</td></tr><tr><td>SmalICNN</td><td>MNIST</td><td>PGD-40</td><td>PGD-40</td><td>99.49%</td><td>96.31%</td><td>9.25</td></tr><tr><td>ResNet-38</td><td>CIFAR-10</td><td>PGD-10</td><td>PGD-20</td><td>83.62%</td><td>42.29%</td><td>79.42</td></tr><tr><td>MobileNetV2</td><td>CIFAR-10</td><td>PGD-10</td><td>PGD-20</td><td>84.42%</td><td>46.92%</td><td>86.91</td></tr></table>",
|
| 718 |
+
"bbox": [
|
| 719 |
+
269,
|
| 720 |
+
205,
|
| 721 |
+
725,
|
| 722 |
+
255
|
| 723 |
+
],
|
| 724 |
+
"page_idx": 5
|
| 725 |
+
},
|
| 726 |
+
{
|
| 727 |
+
"type": "text",
|
| 728 |
+
"text": "4.2 EVALUATION AND ANALYSIS ",
|
| 729 |
+
"text_level": 1,
|
| 730 |
+
"bbox": [
|
| 731 |
+
174,
|
| 732 |
+
268,
|
| 733 |
+
415,
|
| 734 |
+
282
|
| 735 |
+
],
|
| 736 |
+
"page_idx": 5
|
| 737 |
+
},
|
| 738 |
+
{
|
| 739 |
+
"type": "text",
|
| 740 |
+
"text": "MNIST Experiments The MNIST experimental results on RDI-SmallCNN are summarized in table 2, with several meaningful observations to be drawn. First, the undefended models (Standard) are easily compromised by all attack forms. Second, The single attack-defended model (Main Branch) achieves the best ATA against the same type of attack, i.e., ATA (Main Branch), and also seems to boost the closest output branch’s robustness, i.e., ATA (Branch 2). However, its defense effect on the further-away Branch 1 is degraded, and also shows to be fragile under two stronger attacks (Average, and Max-Average). Third, both Average and Max-Average defenses achieve good TAs, as well as ATAs against all attack forms (and therefore Worst-Case), with Max-Average slightly better at both (the margins are small due to the data/task simplicity; see next two). ",
|
| 741 |
+
"bbox": [
|
| 742 |
+
173,
|
| 743 |
+
290,
|
| 744 |
+
825,
|
| 745 |
+
416
|
| 746 |
+
],
|
| 747 |
+
"page_idx": 5
|
| 748 |
+
},
|
| 749 |
+
{
|
| 750 |
+
"type": "text",
|
| 751 |
+
"text": "Moreover, compared to the strong baseline of SmallCNN defended by PGD (40 iterations)-based adversarial training, RDI-SmallCNN with Max-Average defense wins in terms of both TA and ATA. Impressively, that comes together with $3 4 . 3 0 \\%$ computational savings compared to the baseline. Here the different defense forms do not appear to alter the inference efficiency much: they all save around $34 \\% - 3 6 \\%$ MFlops compared to the backbone. ",
|
| 752 |
+
"bbox": [
|
| 753 |
+
174,
|
| 754 |
+
421,
|
| 755 |
+
825,
|
| 756 |
+
492
|
| 757 |
+
],
|
| 758 |
+
"page_idx": 5
|
| 759 |
+
},
|
| 760 |
+
{
|
| 761 |
+
"type": "table",
|
| 762 |
+
"img_path": "images/63449c166c45ceeb14145af351563b34ca83fe6c02e2bd841fdc347e7c1eaf9a.jpg",
|
| 763 |
+
"table_caption": [
|
| 764 |
+
"Table 2: The performance of RDI-SmallCNN. The ”Average MFlops” is calculated by averaging the total flop costs consumed over the inference of the entire set (different samples take different FLOPs due to input-adaptive inference). The perturbation size and step size are 0.3 and 0.01, respectively. "
|
| 765 |
+
],
|
| 766 |
+
"table_footnote": [],
|
| 767 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defense Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>99.48%</td><td rowspan=1 colspan=1>99.50%</td><td rowspan=1 colspan=1>99.51%</td><td rowspan=1 colspan=1>99.52%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch 1)</td><td rowspan=1 colspan=1>6.60%</td><td rowspan=1 colspan=1>60.50%</td><td rowspan=1 colspan=1>98.69%</td><td rowspan=1 colspan=1>98.52%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch 2)</td><td rowspan=1 colspan=1>3.16%</td><td rowspan=1 colspan=1>98.14%</td><td rowspan=1 colspan=1>97.64%</td><td rowspan=1 colspan=1>97.62%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>1.32%</td><td rowspan=1 colspan=1>96.70%</td><td rowspan=1 colspan=1>96.30%</td><td rowspan=1 colspan=1>96.43%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>2.61%</td><td rowspan=1 colspan=1>61.35%</td><td rowspan=1 colspan=1>97.37%</td><td rowspan=1 colspan=1>97.42%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>2.10%</td><td rowspan=1 colspan=1>61.83%</td><td rowspan=1 colspan=1>96.82%</td><td rowspan=1 colspan=1>96.89%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>1.32%</td><td rowspan=1 colspan=1>60.50%</td><td rowspan=1 colspan=1>96.30%</td><td rowspan=1 colspan=1>96.43%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>5.89</td><td rowspan=1 colspan=1>5.89</td><td rowspan=1 colspan=1>5.95</td><td rowspan=1 colspan=1>6.08</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>36.40%</td><td rowspan=1 colspan=1>36.40%</td><td rowspan=1 colspan=1>35.70%</td><td rowspan=1 colspan=1>34.30%</td></tr></table>",
|
| 768 |
+
"bbox": [
|
| 769 |
+
238,
|
| 770 |
+
558,
|
| 771 |
+
753,
|
| 772 |
+
707
|
| 773 |
+
],
|
| 774 |
+
"page_idx": 5
|
| 775 |
+
},
|
| 776 |
+
{
|
| 777 |
+
"type": "text",
|
| 778 |
+
"text": "CIFAR-10 Experiments The results on RDI-ResNet38 and RDI-MobileNetV2 are presented in Tables 3 and 4, respectively. Most findings seem to concur with MNIST experiments. Specifically, on the more complicated CIFAR-10 classification task, Max-Average defense achieves much more obvious margins over Average defense, in terms of ATA (Worst-Case): $2 . 7 9 \\%$ for RDI-ResNet38, and $1 . 0 6 \\%$ for RDI-MobileNetV2. Interestingly, the Average defense is not even the strongest in defending average attacks, as Max-Average defense can achieve higher ATA (Average) in both cases. We conjecture that averaging all branch losses might “over-smooth” and diminish useful gradients. ",
|
| 779 |
+
"bbox": [
|
| 780 |
+
173,
|
| 781 |
+
724,
|
| 782 |
+
825,
|
| 783 |
+
824
|
| 784 |
+
],
|
| 785 |
+
"page_idx": 5
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"type": "text",
|
| 789 |
+
"text": "Compared to the defended ResNet-38 and MobileNet-V2 backbones, RDI-Nets with Max-Average defense achieve higher TAs and ATAs for both. Especially, the ATA (Worst-Case) of RDI-ResNet38 surpasses the ATA of ResNet-38 defended by PGD-adversarial training by $1 . 0 3 \\%$ , while saving around $30 \\%$ inference budget. We find that different defenses on CIFAR-10 have more notable impacts on computational saving. Seemingly, a stronger defense (Max-Average) requires inputs to go through the scrutiny of more layers on average, before outputting confident enough predictions: a sensible observation as we expect. ",
|
| 790 |
+
"bbox": [
|
| 791 |
+
174,
|
| 792 |
+
829,
|
| 793 |
+
825,
|
| 794 |
+
928
|
| 795 |
+
],
|
| 796 |
+
"page_idx": 5
|
| 797 |
+
},
|
| 798 |
+
{
|
| 799 |
+
"type": "text",
|
| 800 |
+
"text": "Visualization of Adaptive Inference Behaviors We visualize the exiting behaviors of RDIResNet38 in Fig 4.2. We plot each branch exiting percentage on clean set and adversarial sets (worst-case) of examples. A few interesting observations can be found. First, we observe that the single-attack defended model can be easily fooled as adversarial examples can be routed through other less-defended outputs (due to the limited transferability of attacks between different outputs). Second, the two stronger defenses (Average and Max-Average) show much more uniform usage of multiple outputs. Their routing behaviors for clean examples are almost identical. For adversarial examples, Max-Average tends to call upon the full inference more often (i.e., more “conservative”). ",
|
| 801 |
+
"bbox": [
|
| 802 |
+
173,
|
| 803 |
+
103,
|
| 804 |
+
825,
|
| 805 |
+
215
|
| 806 |
+
],
|
| 807 |
+
"page_idx": 6
|
| 808 |
+
},
|
| 809 |
+
{
|
| 810 |
+
"type": "table",
|
| 811 |
+
"img_path": "images/7e8fdd045fc49a62794ceea58897bf0854402ef923d46092bd0bdc46c5542bae.jpg",
|
| 812 |
+
"table_caption": [
|
| 813 |
+
"Table 3: The performance evaluation on RDI-ResNet38. The perturbation size and step size are 8/255 and 2/255, respectively. "
|
| 814 |
+
],
|
| 815 |
+
"table_footnote": [],
|
| 816 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>0.12%</td><td rowspan=1 colspan=1>12.02%</td><td rowspan=1 colspan=1>71.56%</td><td rowspan=1 colspan=1>69.71%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>5.58%</td><td rowspan=1 colspan=1>66.67%</td><td rowspan=1 colspan=1>63.11%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch3)</td><td rowspan=1 colspan=1>0.04%</td><td rowspan=1 colspan=1>42.73%</td><td rowspan=1 colspan=1>60.65%</td><td rowspan=1 colspan=1>60.72%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch4)</td><td rowspan=1 colspan=1>0.06%</td><td rowspan=1 colspan=1>34.95%</td><td rowspan=1 colspan=1>50.17%</td><td rowspan=1 colspan=1>47.82%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch5)</td><td rowspan=1 colspan=1>0.06%</td><td rowspan=1 colspan=1>41.77%</td><td rowspan=1 colspan=1>44.83%</td><td rowspan=1 colspan=1>45.53%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch6)</td><td rowspan=1 colspan=1>0.11%</td><td rowspan=1 colspan=1>41.68%</td><td rowspan=1 colspan=1>45.83%</td><td rowspan=1 colspan=1>44.12%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>0.13%</td><td rowspan=1 colspan=1>42.74%</td><td rowspan=1 colspan=1>47.52%</td><td rowspan=1 colspan=1>49.82%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>9.14%</td><td rowspan=1 colspan=1>42.09%</td><td rowspan=1 colspan=1>43.32%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>7.15%</td><td rowspan=1 colspan=1>40.53%</td><td rowspan=1 colspan=1>43.43%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>5.58%</td><td rowspan=1 colspan=1>40.53%</td><td rowspan=1 colspan=1>43.32%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>29.41</td><td rowspan=1 colspan=1>48.27</td><td rowspan=1 colspan=1>56.90</td><td rowspan=1 colspan=1>57.81</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>62.96%</td><td rowspan=1 colspan=1>39.20%</td><td rowspan=1 colspan=1>28.35%</td><td rowspan=1 colspan=1>27.20%</td></tr></table>",
|
| 817 |
+
"bbox": [
|
| 818 |
+
238,
|
| 819 |
+
265,
|
| 820 |
+
758,
|
| 821 |
+
472
|
| 822 |
+
],
|
| 823 |
+
"page_idx": 6
|
| 824 |
+
},
|
| 825 |
+
{
|
| 826 |
+
"type": "table",
|
| 827 |
+
"img_path": "images/a96aa2ca65f67fe8e416245d1c2e465a1b39c149ac432c180949e628899cf307.jpg",
|
| 828 |
+
"table_caption": [
|
| 829 |
+
"Table 4: The performance evaluation on RDI-MobilenetV2. The perturbation size and step size are 8/255 and 2/255, respectively. "
|
| 830 |
+
],
|
| 831 |
+
"table_footnote": [],
|
| 832 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>93.22%</td><td rowspan=1 colspan=1>85.28%</td><td rowspan=1 colspan=1>82.14%</td><td rowspan=1 colspan=1>84.91%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>0.35%</td><td rowspan=1 colspan=1>37.40%</td><td rowspan=1 colspan=1>67.65%</td><td rowspan=1 colspan=1>71.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>47.35%</td><td rowspan=1 colspan=1>50.38%</td><td rowspan=1 colspan=1>50.15%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>46.69%</td><td rowspan=1 colspan=1>49.33%</td><td rowspan=1 colspan=1>46.99%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>35.20%</td><td rowspan=1 colspan=1>45.93%</td><td rowspan=1 colspan=1>47.00%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>36.66%</td><td rowspan=1 colspan=1>49.33%</td><td rowspan=1 colspan=1>50.18%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>0%</td><td rowspan=1 colspan=1>35.20%</td><td rowspan=1 colspan=1>45.93%</td><td rowspan=1 colspan=1>46.99%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>49.78</td><td rowspan=1 colspan=1>52.81</td><td rowspan=1 colspan=1>58.23</td><td rowspan=1 colspan=1>60.84</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>42.72%</td><td rowspan=1 colspan=1>39.23%</td><td rowspan=1 colspan=1>33.00%</td><td rowspan=1 colspan=1>29.99%</td></tr></table>",
|
| 833 |
+
"bbox": [
|
| 834 |
+
240,
|
| 835 |
+
531,
|
| 836 |
+
758,
|
| 837 |
+
683
|
| 838 |
+
],
|
| 839 |
+
"page_idx": 6
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"type": "image",
|
| 843 |
+
"img_path": "images/5dbda2b36e7c91cbf75d9f8b9e046d38f02ad67e8977986d0348cc86493362e1.jpg",
|
| 844 |
+
"image_caption": [
|
| 845 |
+
"Figure 2: The exiting behaviours of RDI-ResNet38 defended by (a) Single attack defense (Main Branch); (b) Average defense; and (c) Max-Average defense. "
|
| 846 |
+
],
|
| 847 |
+
"image_footnote": [],
|
| 848 |
+
"bbox": [
|
| 849 |
+
186,
|
| 850 |
+
705,
|
| 851 |
+
813,
|
| 852 |
+
833
|
| 853 |
+
],
|
| 854 |
+
"page_idx": 6
|
| 855 |
+
},
|
| 856 |
+
{
|
| 857 |
+
"type": "text",
|
| 858 |
+
"text": "4.3 COMPARISON WITH DEFENDED SPARSE NETWORKS ",
|
| 859 |
+
"text_level": 1,
|
| 860 |
+
"bbox": [
|
| 861 |
+
173,
|
| 862 |
+
875,
|
| 863 |
+
575,
|
| 864 |
+
890
|
| 865 |
+
],
|
| 866 |
+
"page_idx": 6
|
| 867 |
+
},
|
| 868 |
+
{
|
| 869 |
+
"type": "text",
|
| 870 |
+
"text": "An alternative to achieve accuracy-robust-efficiency trade-off is by defending a sparse or compressed model. Inspired by (Guo et al., 2018; Gui et al., 2019), we compare RDI-Net with Max-Average defense to the following baseline: first compressing the network with a state-of-the-art model compression method (Huang & Wang, 2018), and then defend the compressed network using the PGD-10 adversarial training. We sample different sparsity ratios in (Huang & Wang, 2018) to obtain models of different complexities. Fig. 6 in Appendix visualizes the comparison on ResNet-38: for either method, we sample a few models of different MFLOPs. At similar inference costs (e.g., 49.38M for pruning $^ +$ defense, and 48.35M for RDI-Nets), our proposed approach consistently achieves higher ATAs $( > 2 \\% )$ ) than the strong pruning $^ +$ defense baseline, with higher TAs. ",
|
| 871 |
+
"bbox": [
|
| 872 |
+
174,
|
| 873 |
+
895,
|
| 874 |
+
825,
|
| 875 |
+
924
|
| 876 |
+
],
|
| 877 |
+
"page_idx": 6
|
| 878 |
+
},
|
| 879 |
+
{
|
| 880 |
+
"type": "text",
|
| 881 |
+
"text": "",
|
| 882 |
+
"bbox": [
|
| 883 |
+
174,
|
| 884 |
+
103,
|
| 885 |
+
825,
|
| 886 |
+
202
|
| 887 |
+
],
|
| 888 |
+
"page_idx": 7
|
| 889 |
+
},
|
| 890 |
+
{
|
| 891 |
+
"type": "text",
|
| 892 |
+
"text": "We also compare with the latest ATMC algorithm (Gui et al., 2019) that jointly optimizes robustness and efficiency, applied the same ResNet-38 backbone. As shown in Table 5, at comparable MFlops, RDI-ResNet-38 surpasses ATMC by $0 . 3 \\%$ in terms of ATA, with a similar TA. ",
|
| 893 |
+
"bbox": [
|
| 894 |
+
174,
|
| 895 |
+
208,
|
| 896 |
+
482,
|
| 897 |
+
277
|
| 898 |
+
],
|
| 899 |
+
"page_idx": 7
|
| 900 |
+
},
|
| 901 |
+
{
|
| 902 |
+
"type": "table",
|
| 903 |
+
"img_path": "images/84213bfba8a5c35a4a2cb2056d6c629ba2e1f1835603a59f87f36f73d7bd11f1.jpg",
|
| 904 |
+
"table_caption": [
|
| 905 |
+
"Table 5: Performance comparison between RDIResNet38 and ATMC. "
|
| 906 |
+
],
|
| 907 |
+
"table_footnote": [],
|
| 908 |
+
"table_body": "<table><tr><td>Methods</td><td>TA</td><td>ATA</td><td>MFlops</td></tr><tr><td>ATMC (Gui et al. (2019))</td><td>83.81</td><td>43.02</td><td>56.82</td></tr><tr><td>RDI-ResNet-38(Worst-Case)</td><td>83.79</td><td>43.32</td><td>57.81</td></tr></table>",
|
| 909 |
+
"bbox": [
|
| 910 |
+
500,
|
| 911 |
+
236,
|
| 912 |
+
823,
|
| 913 |
+
272
|
| 914 |
+
],
|
| 915 |
+
"page_idx": 7
|
| 916 |
+
},
|
| 917 |
+
{
|
| 918 |
+
"type": "text",
|
| 919 |
+
"text": "",
|
| 920 |
+
"bbox": [
|
| 921 |
+
181,
|
| 922 |
+
279,
|
| 923 |
+
578,
|
| 924 |
+
291
|
| 925 |
+
],
|
| 926 |
+
"page_idx": 7
|
| 927 |
+
},
|
| 928 |
+
{
|
| 929 |
+
"type": "text",
|
| 930 |
+
"text": "4.4 GENERALIZED ROBUSTNESS AGAINST OTHER ATTACKERS ",
|
| 931 |
+
"text_level": 1,
|
| 932 |
+
"bbox": [
|
| 933 |
+
174,
|
| 934 |
+
308,
|
| 935 |
+
622,
|
| 936 |
+
321
|
| 937 |
+
],
|
| 938 |
+
"page_idx": 7
|
| 939 |
+
},
|
| 940 |
+
{
|
| 941 |
+
"type": "text",
|
| 942 |
+
"text": "In the aforementioned experiments, we have only evaluated on RDI-Nets against “deterministic” PGD-based adversarial images. We show that RDI-Nets also achieve better generalized robustness against other “randomized” or unseen attackers. We create the new “random attack”: that attack will randomly combine the multi-exit losses, and summarize the results in Table 6. We also follow the similar setting in Gui et al. (2019) and report the results against FGSM (Goodfellow et al., 2015) and WRM (Sinha et al., 2018) attacker, in Tables 7 and 8 respectively (more complete results can be found in Appendix D). ",
|
| 943 |
+
"bbox": [
|
| 944 |
+
173,
|
| 945 |
+
329,
|
| 946 |
+
825,
|
| 947 |
+
426
|
| 948 |
+
],
|
| 949 |
+
"page_idx": 7
|
| 950 |
+
},
|
| 951 |
+
{
|
| 952 |
+
"type": "table",
|
| 953 |
+
"img_path": "images/73790f5169e44eac88846478f6450ec8034ddda70df06e89e7e81cc574c95a9f.jpg",
|
| 954 |
+
"table_caption": [
|
| 955 |
+
"Table 6: Performance on RDI-ResNet38 against random attack. The perturbation size and step size are 8/255 and 2/255, respectively. More details of random attack can be referenced in Appendix D. "
|
| 956 |
+
],
|
| 957 |
+
"table_footnote": [],
|
| 958 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Random)</td><td rowspan=1 colspan=1>0.01%</td><td rowspan=1 colspan=1>10.33%</td><td rowspan=1 colspan=1>43.11%</td><td rowspan=1 colspan=1>44.86%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>27.33</td><td rowspan=1 colspan=1>52.36</td><td rowspan=1 colspan=1>55.21</td><td rowspan=1 colspan=1>56.54</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>65.58%</td><td rowspan=1 colspan=1>34.07%</td><td rowspan=1 colspan=1>30.48%</td><td rowspan=1 colspan=1>28.80%</td></tr></table>",
|
| 959 |
+
"bbox": [
|
| 960 |
+
238,
|
| 961 |
+
486,
|
| 962 |
+
758,
|
| 963 |
+
566
|
| 964 |
+
],
|
| 965 |
+
"page_idx": 7
|
| 966 |
+
},
|
| 967 |
+
{
|
| 968 |
+
"type": "table",
|
| 969 |
+
"img_path": "images/b90a770e9590e37f9b612c5d679bec049c91f50be7d933c37434b30c299769bd.jpg",
|
| 970 |
+
"table_caption": [
|
| 971 |
+
"Table 7: Performance on RDI-ResNet38 (defended with PGD) against FGSM attack (perturbation size is 8/255). The original defended ResNet38 by PGD under the same attack has ATA $5 1 . 1 1 \\%$ . "
|
| 972 |
+
],
|
| 973 |
+
"table_footnote": [],
|
| 974 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>11.51%</td><td rowspan=1 colspan=1>51.45%</td><td rowspan=1 colspan=1>53.64%</td><td rowspan=1 colspan=1>54.72%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>11.41%</td><td rowspan=1 colspan=1>50.21%</td><td rowspan=1 colspan=1>51.81%</td><td rowspan=1 colspan=1>53.20%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>52.40%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>51.05%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>65.74</td><td rowspan=1 colspan=1>55.27</td><td rowspan=1 colspan=1>58.27</td><td rowspan=1 colspan=1>59.67</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>17.21%</td><td rowspan=1 colspan=1>30.40%</td><td rowspan=1 colspan=1>26.40%</td><td rowspan=1 colspan=1>24.86%</td></tr></table>",
|
| 975 |
+
"bbox": [
|
| 976 |
+
238,
|
| 977 |
+
627,
|
| 978 |
+
758,
|
| 979 |
+
750
|
| 980 |
+
],
|
| 981 |
+
"page_idx": 7
|
| 982 |
+
},
|
| 983 |
+
{
|
| 984 |
+
"type": "table",
|
| 985 |
+
"img_path": "images/486c035ca8ccc4458a9d567ea0437fba314c9337e22867458d453ce4d464b718.jpg",
|
| 986 |
+
"table_caption": [
|
| 987 |
+
"Table 8: Performance on RDI-ResNet38 (defended with PGD) against WRM attack (perturbation size is 0.3). The original defended ResNet38 by PGD under the same attack has ATA $8 3 . 3 5 \\%$ . "
|
| 988 |
+
],
|
| 989 |
+
"table_footnote": [],
|
| 990 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>34.42%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>26.48%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.40%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>50.05</td><td rowspan=1 colspan=1>50.46</td><td rowspan=1 colspan=1>52.89</td><td rowspan=1 colspan=1>52.38</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>36.98%</td><td rowspan=1 colspan=1>36.46%</td><td rowspan=1 colspan=1>33.40%</td><td rowspan=1 colspan=1>34.04%</td></tr></table>",
|
| 991 |
+
"bbox": [
|
| 992 |
+
240,
|
| 993 |
+
809,
|
| 994 |
+
758,
|
| 995 |
+
933
|
| 996 |
+
],
|
| 997 |
+
"page_idx": 7
|
| 998 |
+
},
|
| 999 |
+
{
|
| 1000 |
+
"type": "text",
|
| 1001 |
+
"text": "5 DISCUSSION AND ANALYSIS ",
|
| 1002 |
+
"text_level": 1,
|
| 1003 |
+
"bbox": [
|
| 1004 |
+
176,
|
| 1005 |
+
102,
|
| 1006 |
+
442,
|
| 1007 |
+
117
|
| 1008 |
+
],
|
| 1009 |
+
"page_idx": 8
|
| 1010 |
+
},
|
| 1011 |
+
{
|
| 1012 |
+
"type": "text",
|
| 1013 |
+
"text": "Intuition: Multi-Output Networks as Special Ensembles Our intuition on defending multioutput networks arises from the success of ensemble defense in improving both accuracy and robustness (Tramer et al. \\` , 2018; Strauss et al., 2017), which also aligns with the model capacity hypothesis (Nakkiran, 2019). A general multi-output network (Xu et al., 2019) could be decomposed by an ensemble of single-output models, with weight re-using enforced among them. It is thus more compact than an ensemble of independent models, and the extent of sharing weight calibrates ensemble diversity versus efficiency. Therefore, we expect a defended multi-output network to (mostly) inherit the strong accuracy/robustness of ensemble defense, while keeping the inference cost lower. ",
|
| 1014 |
+
"bbox": [
|
| 1015 |
+
173,
|
| 1016 |
+
127,
|
| 1017 |
+
823,
|
| 1018 |
+
239
|
| 1019 |
+
],
|
| 1020 |
+
"page_idx": 8
|
| 1021 |
+
},
|
| 1022 |
+
{
|
| 1023 |
+
"type": "text",
|
| 1024 |
+
"text": "Do ”Triple Wins” Go Against the Model Capacity Needs? We point out that our seemingly “free” efficiency gains (e.g., not sacrificing TA/ATA) do not go against the current belief that a more accurate and robust classifier relies on a larger model capacity (Nakkiran, 2019). From the visualization, there remains to be a portion of clean/adversarial examples that have to utilize the full inference to predict well. In other words, the full model capacity is still necessary to achieve our current TAs/ATAs. Meanwhile, just like in standard classification (Wang et al., 2018a), not all adversarial examples are born equally. Many of them can be predicted using fewer inference costs (taking earlier exits). Therefore, RDI-Nets reduces the “effective model capacity” averaged on all testing samples for overall higher inference efficiency, while not altering the full model capacity. ",
|
| 1025 |
+
"bbox": [
|
| 1026 |
+
174,
|
| 1027 |
+
247,
|
| 1028 |
+
825,
|
| 1029 |
+
373
|
| 1030 |
+
],
|
| 1031 |
+
"page_idx": 8
|
| 1032 |
+
},
|
| 1033 |
+
{
|
| 1034 |
+
"type": "text",
|
| 1035 |
+
"text": "6 CONCLUSION ",
|
| 1036 |
+
"text_level": 1,
|
| 1037 |
+
"bbox": [
|
| 1038 |
+
174,
|
| 1039 |
+
382,
|
| 1040 |
+
318,
|
| 1041 |
+
397
|
| 1042 |
+
],
|
| 1043 |
+
"page_idx": 8
|
| 1044 |
+
},
|
| 1045 |
+
{
|
| 1046 |
+
"type": "text",
|
| 1047 |
+
"text": "This paper targets to simultaneously achieve high accuracy and robustness and meanwhile keeping inference costs lower. We introduce the multi-output network and input-adaptive dynamic inference, as a strong tool to the adversarial defense field for the first time. Our RDI-Nets achieve the “triple wins” of better accuracy, stronger robustness, and around $30 \\%$ inference computational savings. Our future work will extend RDI-Nets to more dynamic inference mechanisms. ",
|
| 1048 |
+
"bbox": [
|
| 1049 |
+
174,
|
| 1050 |
+
407,
|
| 1051 |
+
825,
|
| 1052 |
+
477
|
| 1053 |
+
],
|
| 1054 |
+
"page_idx": 8
|
| 1055 |
+
},
|
| 1056 |
+
{
|
| 1057 |
+
"type": "text",
|
| 1058 |
+
"text": "7 ACKNOWLEDGEMENT ",
|
| 1059 |
+
"text_level": 1,
|
| 1060 |
+
"bbox": [
|
| 1061 |
+
176,
|
| 1062 |
+
486,
|
| 1063 |
+
388,
|
| 1064 |
+
501
|
| 1065 |
+
],
|
| 1066 |
+
"page_idx": 8
|
| 1067 |
+
},
|
| 1068 |
+
{
|
| 1069 |
+
"type": "text",
|
| 1070 |
+
"text": "We would like to thank Dr. Yang Yang from Walmart Technology for highly helpful discussions throughout this project. ",
|
| 1071 |
+
"bbox": [
|
| 1072 |
+
176,
|
| 1073 |
+
511,
|
| 1074 |
+
823,
|
| 1075 |
+
539
|
| 1076 |
+
],
|
| 1077 |
+
"page_idx": 8
|
| 1078 |
+
},
|
| 1079 |
+
{
|
| 1080 |
+
"type": "text",
|
| 1081 |
+
"text": "REFERENCES ",
|
| 1082 |
+
"text_level": 1,
|
| 1083 |
+
"bbox": [
|
| 1084 |
+
174,
|
| 1085 |
+
560,
|
| 1086 |
+
287,
|
| 1087 |
+
575
|
| 1088 |
+
],
|
| 1089 |
+
"page_idx": 8
|
| 1090 |
+
},
|
| 1091 |
+
{
|
| 1092 |
+
"type": "text",
|
| 1093 |
+
"text": "Shumeet Baluja and Ian Fischer. Adversarial transformation networks: Learning to generate adversarial examples. In AAAI, 2018. ",
|
| 1094 |
+
"bbox": [
|
| 1095 |
+
174,
|
| 1096 |
+
583,
|
| 1097 |
+
821,
|
| 1098 |
+
612
|
| 1099 |
+
],
|
| 1100 |
+
"page_idx": 8
|
| 1101 |
+
},
|
| 1102 |
+
{
|
| 1103 |
+
"type": "text",
|
| 1104 |
+
"text": "Battista Biggio, Igino Corona, Davide Maiorca, Blaine Nelson, Nedim Srndi ˇ c, Pavel Laskov, Gior- ´ gio Giacinto, and Fabio Roli. Evasion attacks against machine learning at test time. In ECML, 2013. ",
|
| 1105 |
+
"bbox": [
|
| 1106 |
+
173,
|
| 1107 |
+
622,
|
| 1108 |
+
825,
|
| 1109 |
+
665
|
| 1110 |
+
],
|
| 1111 |
+
"page_idx": 8
|
| 1112 |
+
},
|
| 1113 |
+
{
|
| 1114 |
+
"type": "text",
|
| 1115 |
+
"text": "Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In $S P$ , 2017. ",
|
| 1116 |
+
"bbox": [
|
| 1117 |
+
171,
|
| 1118 |
+
675,
|
| 1119 |
+
823,
|
| 1120 |
+
704
|
| 1121 |
+
],
|
| 1122 |
+
"page_idx": 8
|
| 1123 |
+
},
|
| 1124 |
+
{
|
| 1125 |
+
"type": "text",
|
| 1126 |
+
"text": "Tian Qi Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. In NeurIPS, 2018. ",
|
| 1127 |
+
"bbox": [
|
| 1128 |
+
171,
|
| 1129 |
+
713,
|
| 1130 |
+
825,
|
| 1131 |
+
742
|
| 1132 |
+
],
|
| 1133 |
+
"page_idx": 8
|
| 1134 |
+
},
|
| 1135 |
+
{
|
| 1136 |
+
"type": "text",
|
| 1137 |
+
"text": "Jeremy M Cohen, Elan Rosenfeld, and J Zico Kolter. Certified adversarial robustness via randomized smoothing. In ICML, 2019. ",
|
| 1138 |
+
"bbox": [
|
| 1139 |
+
171,
|
| 1140 |
+
752,
|
| 1141 |
+
825,
|
| 1142 |
+
781
|
| 1143 |
+
],
|
| 1144 |
+
"page_idx": 8
|
| 1145 |
+
},
|
| 1146 |
+
{
|
| 1147 |
+
"type": "text",
|
| 1148 |
+
"text": "Michael Figurnov, Maxwell D Collins, Yukun Zhu, Li Zhang, Jonathan Huang, Dmitry Vetrov, and Ruslan Salakhutdinov. Spatially adaptive computation time for residual networks. In CVPR, 2017. ",
|
| 1149 |
+
"bbox": [
|
| 1150 |
+
173,
|
| 1151 |
+
790,
|
| 1152 |
+
825,
|
| 1153 |
+
833
|
| 1154 |
+
],
|
| 1155 |
+
"page_idx": 8
|
| 1156 |
+
},
|
| 1157 |
+
{
|
| 1158 |
+
"type": "text",
|
| 1159 |
+
"text": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. ",
|
| 1160 |
+
"bbox": [
|
| 1161 |
+
169,
|
| 1162 |
+
843,
|
| 1163 |
+
823,
|
| 1164 |
+
871
|
| 1165 |
+
],
|
| 1166 |
+
"page_idx": 8
|
| 1167 |
+
},
|
| 1168 |
+
{
|
| 1169 |
+
"type": "text",
|
| 1170 |
+
"text": "Shupeng Gui, Haotao Wang, Haichuan Yang, Chen Yu, Zhangyang Wang, and Ji Liu. Model compression with adversarial robustness: A unified optimization framework. In NeurIPS, pp. 1283– 1294, 2019. ",
|
| 1171 |
+
"bbox": [
|
| 1172 |
+
176,
|
| 1173 |
+
881,
|
| 1174 |
+
823,
|
| 1175 |
+
922
|
| 1176 |
+
],
|
| 1177 |
+
"page_idx": 8
|
| 1178 |
+
},
|
| 1179 |
+
{
|
| 1180 |
+
"type": "text",
|
| 1181 |
+
"text": "Yiwen Guo, Chao Zhang, Changshui Zhang, and Yurong Chen. Sparse dnns with improved adversarial robustness. In NeurIPS, 2018. ",
|
| 1182 |
+
"bbox": [
|
| 1183 |
+
173,
|
| 1184 |
+
103,
|
| 1185 |
+
823,
|
| 1186 |
+
132
|
| 1187 |
+
],
|
| 1188 |
+
"page_idx": 9
|
| 1189 |
+
},
|
| 1190 |
+
{
|
| 1191 |
+
"type": "text",
|
| 1192 |
+
"text": "Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In NeurIPS, 2015. ",
|
| 1193 |
+
"bbox": [
|
| 1194 |
+
173,
|
| 1195 |
+
140,
|
| 1196 |
+
823,
|
| 1197 |
+
170
|
| 1198 |
+
],
|
| 1199 |
+
"page_idx": 9
|
| 1200 |
+
},
|
| 1201 |
+
{
|
| 1202 |
+
"type": "text",
|
| 1203 |
+
"text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. ",
|
| 1204 |
+
"bbox": [
|
| 1205 |
+
174,
|
| 1206 |
+
178,
|
| 1207 |
+
821,
|
| 1208 |
+
208
|
| 1209 |
+
],
|
| 1210 |
+
"page_idx": 9
|
| 1211 |
+
},
|
| 1212 |
+
{
|
| 1213 |
+
"type": "text",
|
| 1214 |
+
"text": "Hanzhang Hu, Debadeepta Dey, J. Andrew Bagnell, and Martial Hebert. Anytime neural networks via joint optimization of auxiliary losses. In AAAI, 2019. ",
|
| 1215 |
+
"bbox": [
|
| 1216 |
+
173,
|
| 1217 |
+
215,
|
| 1218 |
+
823,
|
| 1219 |
+
246
|
| 1220 |
+
],
|
| 1221 |
+
"page_idx": 9
|
| 1222 |
+
},
|
| 1223 |
+
{
|
| 1224 |
+
"type": "text",
|
| 1225 |
+
"text": "Gao Huang, Danlu Chen, Tianhong Li, Felix Wu, Laurens van der Maaten, and Kilian Weinberger. Multi-scale dense networks for resource efficient image classification. In ICLR, 2018. ",
|
| 1226 |
+
"bbox": [
|
| 1227 |
+
173,
|
| 1228 |
+
253,
|
| 1229 |
+
823,
|
| 1230 |
+
284
|
| 1231 |
+
],
|
| 1232 |
+
"page_idx": 9
|
| 1233 |
+
},
|
| 1234 |
+
{
|
| 1235 |
+
"type": "text",
|
| 1236 |
+
"text": "Zehao Huang and Naiyan Wang. Data-driven sparse structure selection for deep neural networks. In ECCV, 2018. ",
|
| 1237 |
+
"bbox": [
|
| 1238 |
+
173,
|
| 1239 |
+
291,
|
| 1240 |
+
825,
|
| 1241 |
+
320
|
| 1242 |
+
],
|
| 1243 |
+
"page_idx": 9
|
| 1244 |
+
},
|
| 1245 |
+
{
|
| 1246 |
+
"type": "text",
|
| 1247 |
+
"text": "Yigitcan Kaya, Sanghyun Hong, and Tudor Dumitras. Shallow-deep networks: Understanding and mitigating network overthinking. In ICML, 2019. ",
|
| 1248 |
+
"bbox": [
|
| 1249 |
+
173,
|
| 1250 |
+
329,
|
| 1251 |
+
823,
|
| 1252 |
+
359
|
| 1253 |
+
],
|
| 1254 |
+
"page_idx": 9
|
| 1255 |
+
},
|
| 1256 |
+
{
|
| 1257 |
+
"type": "text",
|
| 1258 |
+
"text": "Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial machine learning at scale. In ICLR, 2017. ",
|
| 1259 |
+
"bbox": [
|
| 1260 |
+
173,
|
| 1261 |
+
367,
|
| 1262 |
+
823,
|
| 1263 |
+
396
|
| 1264 |
+
],
|
| 1265 |
+
"page_idx": 9
|
| 1266 |
+
},
|
| 1267 |
+
{
|
| 1268 |
+
"type": "text",
|
| 1269 |
+
"text": "Fangzhou Liao, Ming Liang, Yinpeng Dong, Tianyu Pang, Xiaolin Hu, and Jun Zhu. Defense against adversarial attacks using high-level representation guided denoiser. In CVPR, 2018. ",
|
| 1270 |
+
"bbox": [
|
| 1271 |
+
173,
|
| 1272 |
+
405,
|
| 1273 |
+
823,
|
| 1274 |
+
434
|
| 1275 |
+
],
|
| 1276 |
+
"page_idx": 9
|
| 1277 |
+
},
|
| 1278 |
+
{
|
| 1279 |
+
"type": "text",
|
| 1280 |
+
"text": "Ji Lin, Chuang Gan, and Song Han. Defensive quantization: When efficiency meets robustness. In ICLR, 2019. ",
|
| 1281 |
+
"bbox": [
|
| 1282 |
+
173,
|
| 1283 |
+
443,
|
| 1284 |
+
823,
|
| 1285 |
+
472
|
| 1286 |
+
],
|
| 1287 |
+
"page_idx": 9
|
| 1288 |
+
},
|
| 1289 |
+
{
|
| 1290 |
+
"type": "text",
|
| 1291 |
+
"text": "Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. ",
|
| 1292 |
+
"bbox": [
|
| 1293 |
+
173,
|
| 1294 |
+
479,
|
| 1295 |
+
823,
|
| 1296 |
+
510
|
| 1297 |
+
],
|
| 1298 |
+
"page_idx": 9
|
| 1299 |
+
},
|
| 1300 |
+
{
|
| 1301 |
+
"type": "text",
|
| 1302 |
+
"text": "Preetum Nakkiran. Adversarial robustness may be at odds with simplicity. arXiv, 2019. ",
|
| 1303 |
+
"bbox": [
|
| 1304 |
+
173,
|
| 1305 |
+
517,
|
| 1306 |
+
750,
|
| 1307 |
+
534
|
| 1308 |
+
],
|
| 1309 |
+
"page_idx": 9
|
| 1310 |
+
},
|
| 1311 |
+
{
|
| 1312 |
+
"type": "text",
|
| 1313 |
+
"text": "Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In EuroS&P, 2016. ",
|
| 1314 |
+
"bbox": [
|
| 1315 |
+
173,
|
| 1316 |
+
541,
|
| 1317 |
+
823,
|
| 1318 |
+
571
|
| 1319 |
+
],
|
| 1320 |
+
"page_idx": 9
|
| 1321 |
+
},
|
| 1322 |
+
{
|
| 1323 |
+
"type": "text",
|
| 1324 |
+
"text": "Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. ",
|
| 1325 |
+
"bbox": [
|
| 1326 |
+
171,
|
| 1327 |
+
579,
|
| 1328 |
+
823,
|
| 1329 |
+
608
|
| 1330 |
+
],
|
| 1331 |
+
"page_idx": 9
|
| 1332 |
+
},
|
| 1333 |
+
{
|
| 1334 |
+
"type": "text",
|
| 1335 |
+
"text": "Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Madry. Adversarially robust generalization requires more data. In NeurIPS, 2018. ",
|
| 1336 |
+
"bbox": [
|
| 1337 |
+
171,
|
| 1338 |
+
617,
|
| 1339 |
+
823,
|
| 1340 |
+
647
|
| 1341 |
+
],
|
| 1342 |
+
"page_idx": 9
|
| 1343 |
+
},
|
| 1344 |
+
{
|
| 1345 |
+
"type": "text",
|
| 1346 |
+
"text": "Aman Sinha, Hongseok Namkoong, and John Duchi. Certifying Some Distributional Robustness with Principled Adversarial Training. In ICLR, 2018. ",
|
| 1347 |
+
"bbox": [
|
| 1348 |
+
173,
|
| 1349 |
+
655,
|
| 1350 |
+
823,
|
| 1351 |
+
684
|
| 1352 |
+
],
|
| 1353 |
+
"page_idx": 9
|
| 1354 |
+
},
|
| 1355 |
+
{
|
| 1356 |
+
"type": "text",
|
| 1357 |
+
"text": "Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In ICLR, 2018. ",
|
| 1358 |
+
"bbox": [
|
| 1359 |
+
176,
|
| 1360 |
+
693,
|
| 1361 |
+
823,
|
| 1362 |
+
736
|
| 1363 |
+
],
|
| 1364 |
+
"page_idx": 9
|
| 1365 |
+
},
|
| 1366 |
+
{
|
| 1367 |
+
"type": "text",
|
| 1368 |
+
"text": "Thilo Strauss, Markus Hanselmann, Andrej Junginger, and Holger Ulmer. Ensemble methods as a defense to adversarial perturbations against deep neural networks. arXiv, 2017. ",
|
| 1369 |
+
"bbox": [
|
| 1370 |
+
173,
|
| 1371 |
+
743,
|
| 1372 |
+
820,
|
| 1373 |
+
773
|
| 1374 |
+
],
|
| 1375 |
+
"page_idx": 9
|
| 1376 |
+
},
|
| 1377 |
+
{
|
| 1378 |
+
"type": "text",
|
| 1379 |
+
"text": "Ke Sun, Zhanxing Zhu, and Zhouchen Lin. Towards understanding adversarial examples systematically: Exploring data size, task and model factors. arXiv, 2019. ",
|
| 1380 |
+
"bbox": [
|
| 1381 |
+
174,
|
| 1382 |
+
781,
|
| 1383 |
+
820,
|
| 1384 |
+
811
|
| 1385 |
+
],
|
| 1386 |
+
"page_idx": 9
|
| 1387 |
+
},
|
| 1388 |
+
{
|
| 1389 |
+
"type": "text",
|
| 1390 |
+
"text": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. ",
|
| 1391 |
+
"bbox": [
|
| 1392 |
+
174,
|
| 1393 |
+
819,
|
| 1394 |
+
821,
|
| 1395 |
+
849
|
| 1396 |
+
],
|
| 1397 |
+
"page_idx": 9
|
| 1398 |
+
},
|
| 1399 |
+
{
|
| 1400 |
+
"type": "text",
|
| 1401 |
+
"text": "Cheng Tai, Tong Xiao, Yi Zhang, Xiaogang Wang, et al. Convolutional neural networks with lowrank regularization. In ICLR, 2016. ",
|
| 1402 |
+
"bbox": [
|
| 1403 |
+
174,
|
| 1404 |
+
857,
|
| 1405 |
+
821,
|
| 1406 |
+
887
|
| 1407 |
+
],
|
| 1408 |
+
"page_idx": 9
|
| 1409 |
+
},
|
| 1410 |
+
{
|
| 1411 |
+
"type": "text",
|
| 1412 |
+
"text": "Surat Teerapittayanon, Bradley McDanel, and H. T. Kung. Branchynet: Fast inference via early exiting from deep neural networks. In ICPR, 2017. ",
|
| 1413 |
+
"bbox": [
|
| 1414 |
+
176,
|
| 1415 |
+
895,
|
| 1416 |
+
820,
|
| 1417 |
+
924
|
| 1418 |
+
],
|
| 1419 |
+
"page_idx": 9
|
| 1420 |
+
},
|
| 1421 |
+
{
|
| 1422 |
+
"type": "text",
|
| 1423 |
+
"text": "Florian Tramer, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick Mc-\\` Daniel. Ensemble adversarial training: Attacks and defenses. In ICLR, 2018. ",
|
| 1424 |
+
"bbox": [
|
| 1425 |
+
173,
|
| 1426 |
+
103,
|
| 1427 |
+
823,
|
| 1428 |
+
132
|
| 1429 |
+
],
|
| 1430 |
+
"page_idx": 10
|
| 1431 |
+
},
|
| 1432 |
+
{
|
| 1433 |
+
"type": "text",
|
| 1434 |
+
"text": "Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In ICLR, 2019. ",
|
| 1435 |
+
"bbox": [
|
| 1436 |
+
171,
|
| 1437 |
+
143,
|
| 1438 |
+
821,
|
| 1439 |
+
172
|
| 1440 |
+
],
|
| 1441 |
+
"page_idx": 10
|
| 1442 |
+
},
|
| 1443 |
+
{
|
| 1444 |
+
"type": "text",
|
| 1445 |
+
"text": "Xin Wang, Fisher Yu, Zi-Yi Dou, and Joseph E Gonzalez. Skipnet: Learning dynamic routing in convolutional networks. In ECCV, 2018a. ",
|
| 1446 |
+
"bbox": [
|
| 1447 |
+
173,
|
| 1448 |
+
183,
|
| 1449 |
+
821,
|
| 1450 |
+
213
|
| 1451 |
+
],
|
| 1452 |
+
"page_idx": 10
|
| 1453 |
+
},
|
| 1454 |
+
{
|
| 1455 |
+
"type": "text",
|
| 1456 |
+
"text": "Yue Wang, Tan Nguyen, Yang Zhao, Zhangyang Wang, Yingyan Lin, and Richard Baraniuk. Energynet: Energy-efficient dynamic inference. 2018b. ",
|
| 1457 |
+
"bbox": [
|
| 1458 |
+
173,
|
| 1459 |
+
223,
|
| 1460 |
+
823,
|
| 1461 |
+
253
|
| 1462 |
+
],
|
| 1463 |
+
"page_idx": 10
|
| 1464 |
+
},
|
| 1465 |
+
{
|
| 1466 |
+
"type": "text",
|
| 1467 |
+
"text": "Yue Wang, Jianghao Shen, Ting-Kuei Hu, Pengfei Xu, Tan Nguyen, Richard Baraniuk, Zhangyang Wang, and Yingyan Lin. Dual dynamic inference: Enabling more efficient, adaptive and controllable deep inference. arXiv preprint arXiv:1907.04523, 2019. ",
|
| 1468 |
+
"bbox": [
|
| 1469 |
+
176,
|
| 1470 |
+
263,
|
| 1471 |
+
823,
|
| 1472 |
+
308
|
| 1473 |
+
],
|
| 1474 |
+
"page_idx": 10
|
| 1475 |
+
},
|
| 1476 |
+
{
|
| 1477 |
+
"type": "text",
|
| 1478 |
+
"text": "Yunhe Wang, Chang Xu, Chao Xu, and Dacheng Tao. Adversarial learning of portable student networks. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018c. ",
|
| 1479 |
+
"bbox": [
|
| 1480 |
+
171,
|
| 1481 |
+
318,
|
| 1482 |
+
825,
|
| 1483 |
+
348
|
| 1484 |
+
],
|
| 1485 |
+
"page_idx": 10
|
| 1486 |
+
},
|
| 1487 |
+
{
|
| 1488 |
+
"type": "text",
|
| 1489 |
+
"text": "Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In NeurIPS, 2016. ",
|
| 1490 |
+
"bbox": [
|
| 1491 |
+
174,
|
| 1492 |
+
358,
|
| 1493 |
+
823,
|
| 1494 |
+
387
|
| 1495 |
+
],
|
| 1496 |
+
"page_idx": 10
|
| 1497 |
+
},
|
| 1498 |
+
{
|
| 1499 |
+
"type": "text",
|
| 1500 |
+
"text": "Jiaxiang Wu, Cong Leng, Yuhang Wang, Qinghao Hu, and Jian Cheng. Quantized convolutional neural networks for mobile devices. In CVPR, 2016. ",
|
| 1501 |
+
"bbox": [
|
| 1502 |
+
173,
|
| 1503 |
+
398,
|
| 1504 |
+
823,
|
| 1505 |
+
428
|
| 1506 |
+
],
|
| 1507 |
+
"page_idx": 10
|
| 1508 |
+
},
|
| 1509 |
+
{
|
| 1510 |
+
"type": "text",
|
| 1511 |
+
"text": "Junru Wu, Yue Wang, Zhenyu Wu, Zhangyang Wang, Ashok Veeraraghavan, and Yingyan Lin. Deep $k$ -means: Re-training and parameter sharing with harder cluster assignments for compressing deep convolutions. In ICML, 2018. ",
|
| 1512 |
+
"bbox": [
|
| 1513 |
+
174,
|
| 1514 |
+
439,
|
| 1515 |
+
823,
|
| 1516 |
+
482
|
| 1517 |
+
],
|
| 1518 |
+
"page_idx": 10
|
| 1519 |
+
},
|
| 1520 |
+
{
|
| 1521 |
+
"type": "text",
|
| 1522 |
+
"text": "Cihang Xie, Jianyu Wang, Zhishuai Zhang, Zhou Ren, and Alan Yuille. Mitigating adversarial effects through randomization. In ICLR, 2018. ",
|
| 1523 |
+
"bbox": [
|
| 1524 |
+
168,
|
| 1525 |
+
492,
|
| 1526 |
+
823,
|
| 1527 |
+
522
|
| 1528 |
+
],
|
| 1529 |
+
"page_idx": 10
|
| 1530 |
+
},
|
| 1531 |
+
{
|
| 1532 |
+
"type": "text",
|
| 1533 |
+
"text": "Donna Xu, Yaxin Shi, Ivor W Tsang, Yew-Soon Ong, Chen Gong, and Xiaobo Shen. A survey on multi-output learning. arXiv, 2019. ",
|
| 1534 |
+
"bbox": [
|
| 1535 |
+
173,
|
| 1536 |
+
532,
|
| 1537 |
+
823,
|
| 1538 |
+
563
|
| 1539 |
+
],
|
| 1540 |
+
"page_idx": 10
|
| 1541 |
+
},
|
| 1542 |
+
{
|
| 1543 |
+
"type": "text",
|
| 1544 |
+
"text": "Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018. ",
|
| 1545 |
+
"bbox": [
|
| 1546 |
+
173,
|
| 1547 |
+
573,
|
| 1548 |
+
823,
|
| 1549 |
+
602
|
| 1550 |
+
],
|
| 1551 |
+
"page_idx": 10
|
| 1552 |
+
},
|
| 1553 |
+
{
|
| 1554 |
+
"type": "text",
|
| 1555 |
+
"text": "Shaokai Ye, Kaidi Xu, Sijia Liu, Hao Cheng, Jan-Henrik Lambrechts, Huan Zhang, Aojun Zhou, Kaisheng Ma, Yanzhi Wang, and Xue Lin. Adversarial robustness vs model compression, or both? In ICCV, 2019. ",
|
| 1556 |
+
"bbox": [
|
| 1557 |
+
174,
|
| 1558 |
+
613,
|
| 1559 |
+
823,
|
| 1560 |
+
656
|
| 1561 |
+
],
|
| 1562 |
+
"page_idx": 10
|
| 1563 |
+
},
|
| 1564 |
+
{
|
| 1565 |
+
"type": "text",
|
| 1566 |
+
"text": "Xiyu Yu, Tongliang Liu, Xinchao Wang, and Dacheng Tao. On compressing deep models by low rank and sparse decomposition. In CVPR, 2017. ",
|
| 1567 |
+
"bbox": [
|
| 1568 |
+
169,
|
| 1569 |
+
667,
|
| 1570 |
+
823,
|
| 1571 |
+
696
|
| 1572 |
+
],
|
| 1573 |
+
"page_idx": 10
|
| 1574 |
+
},
|
| 1575 |
+
{
|
| 1576 |
+
"type": "text",
|
| 1577 |
+
"text": "Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P Xing, Laurent El Ghaoui, and Michael I Jordan. Theoretically principled trade-off between robustness and accuracy. arXiv, 2019. ",
|
| 1578 |
+
"bbox": [
|
| 1579 |
+
174,
|
| 1580 |
+
708,
|
| 1581 |
+
820,
|
| 1582 |
+
737
|
| 1583 |
+
],
|
| 1584 |
+
"page_idx": 10
|
| 1585 |
+
},
|
| 1586 |
+
{
|
| 1587 |
+
"type": "text",
|
| 1588 |
+
"text": "Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, 2018. ",
|
| 1589 |
+
"bbox": [
|
| 1590 |
+
173,
|
| 1591 |
+
748,
|
| 1592 |
+
821,
|
| 1593 |
+
776
|
| 1594 |
+
],
|
| 1595 |
+
"page_idx": 10
|
| 1596 |
+
},
|
| 1597 |
+
{
|
| 1598 |
+
"type": "text",
|
| 1599 |
+
"text": "A LEARNING DETAILS OF RDI-NETS ",
|
| 1600 |
+
"text_level": 1,
|
| 1601 |
+
"bbox": [
|
| 1602 |
+
174,
|
| 1603 |
+
806,
|
| 1604 |
+
501,
|
| 1605 |
+
821
|
| 1606 |
+
],
|
| 1607 |
+
"page_idx": 10
|
| 1608 |
+
},
|
| 1609 |
+
{
|
| 1610 |
+
"type": "text",
|
| 1611 |
+
"text": "MNIST We adopt the network architecture from (Chen et al., 2018) with four convolutions and three full-connected layers. We train for 13100 iterations with a batch size of 256. The learning rate is initialized as 0.033 and is lowered by 10 at 12000th and 12900th iteration. For hybrid loss, the weights $\\{ w _ { i } \\} _ { i = 1 } ^ { N + 1 }$ are set as $\\{ 1 , 1 , 1 \\}$ for simplicity. For adversarial defense/attack, we perform 40-steps PGD for both defense and evaluation. The perturbation size and step size are set as 0.3 and 0.01. ",
|
| 1612 |
+
"bbox": [
|
| 1613 |
+
173,
|
| 1614 |
+
838,
|
| 1615 |
+
825,
|
| 1616 |
+
922
|
| 1617 |
+
],
|
| 1618 |
+
"page_idx": 10
|
| 1619 |
+
},
|
| 1620 |
+
{
|
| 1621 |
+
"type": "text",
|
| 1622 |
+
"text": "CIFAR-10 We take ResNet-38 and MobileNetV2 as the backbone architectures. For RDIResNet38, we initialize learning rate as 0.1 and decay it by a factor of 10 at 32000th and 48000th iteration. The learning procedure stops at 55000 iteration. For RDI-MobileNetV2, the learning rate is set to 0.05 and is lowered by 10 times at 62000th and 70000th iteration. We stop the learn$\\{ w _ { i } \\} _ { i = 1 } ^ { N + 1 }$ 76000 iteration. For hybrid loss, we follow tof RDI-ResNet38 and RDI-MobileNetV2 as ectively. For adversarial defense/attack, the $\\{ 0 . 5 , 0 . 5 , 0 . 7 , 0 . 7 , 0 . 9 , 0 . 9 , 2 \\}$ 019)and are $\\{ 0 . 5 , 0 . 5 , 1 \\}$ \nset as $8 / 2 5 5$ and 2/255. 10-steps PGD is performed for defense and 20-steps PGD is utilized for evaluation. ",
|
| 1623 |
+
"bbox": [
|
| 1624 |
+
173,
|
| 1625 |
+
103,
|
| 1626 |
+
825,
|
| 1627 |
+
231
|
| 1628 |
+
],
|
| 1629 |
+
"page_idx": 11
|
| 1630 |
+
},
|
| 1631 |
+
{
|
| 1632 |
+
"type": "text",
|
| 1633 |
+
"text": "B NETWORK STRUCTURE OF RDI-NETS ",
|
| 1634 |
+
"text_level": 1,
|
| 1635 |
+
"bbox": [
|
| 1636 |
+
176,
|
| 1637 |
+
255,
|
| 1638 |
+
526,
|
| 1639 |
+
272
|
| 1640 |
+
],
|
| 1641 |
+
"page_idx": 11
|
| 1642 |
+
},
|
| 1643 |
+
{
|
| 1644 |
+
"type": "text",
|
| 1645 |
+
"text": "To build RDI-Nets, we follow the similar setting in Teerapittayanon et al. (2017) by appending additional branch classifiers at equidistant points throughout a given network, as illustrated in Fig 3, Fig 4 and Fig 5. A few pooling operations, light-weight convolutions and fully-connected layers are appended to each branch classifiers. Note that the extra flops introduced by side branch classifiers are less than $2 \\%$ than the original ResNet-38 or MobileNetV2. ",
|
| 1646 |
+
"bbox": [
|
| 1647 |
+
173,
|
| 1648 |
+
290,
|
| 1649 |
+
825,
|
| 1650 |
+
359
|
| 1651 |
+
],
|
| 1652 |
+
"page_idx": 11
|
| 1653 |
+
},
|
| 1654 |
+
{
|
| 1655 |
+
"type": "image",
|
| 1656 |
+
"img_path": "images/82af5566478d1d34d525841fa230e6fcb3015f2578147dfb37d255ed3e14f362.jpg",
|
| 1657 |
+
"image_caption": [
|
| 1658 |
+
"Figure 3: Network architecture of RDI-SmallCNN. Two branch classifiers are inserted after 1st convolutional layer and 3rd convolutional layer in the original SmallCNN. "
|
| 1659 |
+
],
|
| 1660 |
+
"image_footnote": [],
|
| 1661 |
+
"bbox": [
|
| 1662 |
+
281,
|
| 1663 |
+
380,
|
| 1664 |
+
714,
|
| 1665 |
+
560
|
| 1666 |
+
],
|
| 1667 |
+
"page_idx": 11
|
| 1668 |
+
},
|
| 1669 |
+
{
|
| 1670 |
+
"type": "image",
|
| 1671 |
+
"img_path": "images/72d0a0e8f76b621498484d67b9cdfa2a8cb6c771f8e3a2e58b704923d89eaaee.jpg",
|
| 1672 |
+
"image_caption": [
|
| 1673 |
+
"Figure 4: Network architecture of RDI-ResNet38. In each residual block group, two branch classifiers are inserted after $1 { s t }$ residual block and 4th residual block. "
|
| 1674 |
+
],
|
| 1675 |
+
"image_footnote": [],
|
| 1676 |
+
"bbox": [
|
| 1677 |
+
176,
|
| 1678 |
+
632,
|
| 1679 |
+
823,
|
| 1680 |
+
795
|
| 1681 |
+
],
|
| 1682 |
+
"page_idx": 11
|
| 1683 |
+
},
|
| 1684 |
+
{
|
| 1685 |
+
"type": "text",
|
| 1686 |
+
"text": "C INPUT-ADAPTIVE INFERENCE FOR RDI-NETS ",
|
| 1687 |
+
"text_level": 1,
|
| 1688 |
+
"bbox": [
|
| 1689 |
+
174,
|
| 1690 |
+
861,
|
| 1691 |
+
589,
|
| 1692 |
+
877
|
| 1693 |
+
],
|
| 1694 |
+
"page_idx": 11
|
| 1695 |
+
},
|
| 1696 |
+
{
|
| 1697 |
+
"type": "text",
|
| 1698 |
+
"text": "Similar to the deterministic strategy in Teerapittayanon et al. (2017), we adopt the entropy as the measure of the prediction confidence. Given a prediction vector $y \\in \\mathbb { R } ^ { C }$ , where $C$ is the number of ",
|
| 1699 |
+
"bbox": [
|
| 1700 |
+
174,
|
| 1701 |
+
895,
|
| 1702 |
+
825,
|
| 1703 |
+
924
|
| 1704 |
+
],
|
| 1705 |
+
"page_idx": 11
|
| 1706 |
+
},
|
| 1707 |
+
{
|
| 1708 |
+
"type": "image",
|
| 1709 |
+
"img_path": "images/10c09fd684a097f8c76b530e4513ea7c0bcae954dd0be1138ad38ca37f8c9ced.jpg",
|
| 1710 |
+
"image_caption": [
|
| 1711 |
+
"Figure 5: Network architecture of RDI-MobilenetV2. Two branch classfiers are inserted after 3rd inverted residual block and 11th inverted residual block in the orignal MobilenetV2. "
|
| 1712 |
+
],
|
| 1713 |
+
"image_footnote": [],
|
| 1714 |
+
"bbox": [
|
| 1715 |
+
176,
|
| 1716 |
+
102,
|
| 1717 |
+
815,
|
| 1718 |
+
257
|
| 1719 |
+
],
|
| 1720 |
+
"page_idx": 12
|
| 1721 |
+
},
|
| 1722 |
+
{
|
| 1723 |
+
"type": "text",
|
| 1724 |
+
"text": "class, the entropy of $y$ is defined as follow, ",
|
| 1725 |
+
"bbox": [
|
| 1726 |
+
176,
|
| 1727 |
+
310,
|
| 1728 |
+
454,
|
| 1729 |
+
324
|
| 1730 |
+
],
|
| 1731 |
+
"page_idx": 12
|
| 1732 |
+
},
|
| 1733 |
+
{
|
| 1734 |
+
"type": "equation",
|
| 1735 |
+
"img_path": "images/406c57f0c03adb16082db14f1ffac0ef5b9f6f32e69bfa68056c39c3c90161a5.jpg",
|
| 1736 |
+
"text": "$$\n- \\sum _ { c = 1 } ^ { C } ( y _ { c } + \\epsilon ) l o g ( y _ { c } + \\epsilon ) ,\n$$",
|
| 1737 |
+
"text_format": "latex",
|
| 1738 |
+
"bbox": [
|
| 1739 |
+
408,
|
| 1740 |
+
327,
|
| 1741 |
+
588,
|
| 1742 |
+
369
|
| 1743 |
+
],
|
| 1744 |
+
"page_idx": 12
|
| 1745 |
+
},
|
| 1746 |
+
{
|
| 1747 |
+
"type": "text",
|
| 1748 |
+
"text": "where $\\epsilon$ is a small positive constant used for robust entropy computation. To perform fast inference on a $( K { + } 1 )$ -output RDI-Net, we need to determine $K$ threshold numbers, i.e., $\\{ t _ { i } \\} _ { i = 1 } ^ { K }$ , so that the input $x$ will exit at ith branch if the entropy of $y _ { i }$ is larger than $t _ { i }$ . To choose $\\{ t _ { i } \\} _ { i = 1 } ^ { K }$ , Huang et al. (2018) provides a good starting point by fixing exiting probability of each branch classifiers equally on validation set so that each sample can equally contribute to inference. We follow this strategy but adjust the thresholds to make the contribution of middle branches slightly larger than the early branches. The threshold numbers for RDI-SmallCNN, RDI-ResNet38, and RDI-MobilenetV2 are set to be $\\{ 0 . 0 2 3 , 0 . 0 1 4 \\}$ , $\\{ 0 . 3 2 , 0 . 3 6 , 0 . 3 9 , 0 . 8 3 , 1 . 1 2 , 1 . 3 5 \\}$ , and $\\{ 0 . 2 6 7 , 0 . 7 6 5 \\}$ , respectively. ",
|
| 1749 |
+
"bbox": [
|
| 1750 |
+
173,
|
| 1751 |
+
371,
|
| 1752 |
+
825,
|
| 1753 |
+
484
|
| 1754 |
+
],
|
| 1755 |
+
"page_idx": 12
|
| 1756 |
+
},
|
| 1757 |
+
{
|
| 1758 |
+
"type": "image",
|
| 1759 |
+
"img_path": "images/000aeef68d976dcd4bcb2683319e472f60fc2b5c2ea5e35a61d7cd81310a2651.jpg",
|
| 1760 |
+
"image_caption": [
|
| 1761 |
+
"Figure 6: Performance comparison between RDI-Net and the pruning $^ +$ defense baseline. Each marker represents a model, whose size is proportional to its MFlops. $\\gamma$ is the sparsity trade-off parameter: the larger the sparser (smaller model). "
|
| 1762 |
+
],
|
| 1763 |
+
"image_footnote": [],
|
| 1764 |
+
"bbox": [
|
| 1765 |
+
235,
|
| 1766 |
+
493,
|
| 1767 |
+
761,
|
| 1768 |
+
726
|
| 1769 |
+
],
|
| 1770 |
+
"page_idx": 12
|
| 1771 |
+
},
|
| 1772 |
+
{
|
| 1773 |
+
"type": "text",
|
| 1774 |
+
"text": "D GENERALIZED ROBUSTNESS ",
|
| 1775 |
+
"text_level": 1,
|
| 1776 |
+
"bbox": [
|
| 1777 |
+
176,
|
| 1778 |
+
808,
|
| 1779 |
+
447,
|
| 1780 |
+
824
|
| 1781 |
+
],
|
| 1782 |
+
"page_idx": 12
|
| 1783 |
+
},
|
| 1784 |
+
{
|
| 1785 |
+
"type": "text",
|
| 1786 |
+
"text": "Here, we introduce the attack form of random attack and report the complete results against FGSM (Goodfellow et al., 2015) and WRM (Sinha et al., 2018) attacker under various attack forms, in Tables 9 and 10, respectively. ",
|
| 1787 |
+
"bbox": [
|
| 1788 |
+
174,
|
| 1789 |
+
838,
|
| 1790 |
+
825,
|
| 1791 |
+
881
|
| 1792 |
+
],
|
| 1793 |
+
"page_idx": 12
|
| 1794 |
+
},
|
| 1795 |
+
{
|
| 1796 |
+
"type": "text",
|
| 1797 |
+
"text": "Random Attack the attack exploits multi-loss flexibility by randomly fusing all $f _ { \\theta _ { i } }$ losses. Given a $N$ -output network, we have a fusion vector $C \\in \\mathbb { R } ^ { N } \\sim \\mathbf { \\bar { \\mathbb { D } } }$ , where $\\mathbb { D }$ is some distribution (uniform ",
|
| 1798 |
+
"bbox": [
|
| 1799 |
+
173,
|
| 1800 |
+
895,
|
| 1801 |
+
823,
|
| 1802 |
+
924
|
| 1803 |
+
],
|
| 1804 |
+
"page_idx": 12
|
| 1805 |
+
},
|
| 1806 |
+
{
|
| 1807 |
+
"type": "text",
|
| 1808 |
+
"text": "by default). We denote $c _ { j }$ as the $j$ th element of $C$ and $x _ { r d m } ^ { a d v }$ can be found by: ",
|
| 1809 |
+
"bbox": [
|
| 1810 |
+
171,
|
| 1811 |
+
102,
|
| 1812 |
+
679,
|
| 1813 |
+
119
|
| 1814 |
+
],
|
| 1815 |
+
"page_idx": 13
|
| 1816 |
+
},
|
| 1817 |
+
{
|
| 1818 |
+
"type": "equation",
|
| 1819 |
+
"img_path": "images/28e92d0716ec25a8d1f030ffb8bf7032e43b53ce05dd126816a6e711930f0bec.jpg",
|
| 1820 |
+
"text": "$$\nx _ { r d m } ^ { a d v } = \\underset { x ^ { \\prime } \\in | x ^ { \\prime } - x | _ { \\infty } \\leq \\epsilon } { \\arg \\operatorname* { m a x } } | \\frac { 1 } { N } \\sum _ { j = 1 } ^ { N } c _ { j } \\phi ( f _ { \\theta _ { j } } ( x ^ { \\prime } ) , y ) | .\n$$",
|
| 1821 |
+
"text_format": "latex",
|
| 1822 |
+
"bbox": [
|
| 1823 |
+
343,
|
| 1824 |
+
126,
|
| 1825 |
+
651,
|
| 1826 |
+
170
|
| 1827 |
+
],
|
| 1828 |
+
"page_idx": 13
|
| 1829 |
+
},
|
| 1830 |
+
{
|
| 1831 |
+
"type": "text",
|
| 1832 |
+
"text": "It is expected to challenge our defense, due to the infinitely many ways of randomly fusing outputs. ",
|
| 1833 |
+
"bbox": [
|
| 1834 |
+
169,
|
| 1835 |
+
176,
|
| 1836 |
+
823,
|
| 1837 |
+
193
|
| 1838 |
+
],
|
| 1839 |
+
"page_idx": 13
|
| 1840 |
+
},
|
| 1841 |
+
{
|
| 1842 |
+
"type": "table",
|
| 1843 |
+
"img_path": "images/73dd03f838a33ed3a986050754a5aba50dcc648cce43a746d04b85a28340b50e.jpg",
|
| 1844 |
+
"table_caption": [
|
| 1845 |
+
"Table 9: The performance evaluation on RDI-ResNet38 (defended with PGD) against FGSM attack. The perturbation size is 8/255. The ATA of the original defended ResNet38 by PGD under the same attacker is $5 1 . 1 1 \\%$ . "
|
| 1846 |
+
],
|
| 1847 |
+
"table_footnote": [],
|
| 1848 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>20.69%</td><td rowspan=1 colspan=1>66.06%</td><td rowspan=1 colspan=1>72.77%</td><td rowspan=1 colspan=1>72.76%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>16.15%</td><td rowspan=1 colspan=1>53.87%</td><td rowspan=1 colspan=1>70.40%</td><td rowspan=1 colspan=1>69.71%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch3)</td><td rowspan=1 colspan=1>8.13%</td><td rowspan=1 colspan=1>63.70%</td><td rowspan=1 colspan=1>64.19%</td><td rowspan=1 colspan=1>65.14%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch4)</td><td rowspan=1 colspan=1>10.09%</td><td rowspan=1 colspan=1>56.67%</td><td rowspan=1 colspan=1>58.45%</td><td rowspan=1 colspan=1>58.20%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch5)</td><td rowspan=1 colspan=1>9.45%</td><td rowspan=1 colspan=1>50.81%</td><td rowspan=1 colspan=1>52.76%</td><td rowspan=1 colspan=1>52.96%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch6)</td><td rowspan=1 colspan=1>10.22%</td><td rowspan=1 colspan=1>50.34%</td><td rowspan=1 colspan=1>53.17%</td><td rowspan=1 colspan=1>51.05%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>11.51%</td><td rowspan=1 colspan=1>51.45%</td><td rowspan=1 colspan=1>53.64%</td><td rowspan=1 colspan=1>54.72%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>11.41%</td><td rowspan=1 colspan=1>50.21%</td><td rowspan=1 colspan=1>51.81%</td><td rowspan=1 colspan=1>53.20%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>52.40%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>2.09%</td><td rowspan=1 colspan=1>47.53%</td><td rowspan=1 colspan=1>50.63%</td><td rowspan=1 colspan=1>51.05%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>65.74</td><td rowspan=1 colspan=1>55.27</td><td rowspan=1 colspan=1>58.27</td><td rowspan=1 colspan=1>59.67</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>17.21%</td><td rowspan=1 colspan=1>30.40%</td><td rowspan=1 colspan=1>26.40%</td><td rowspan=1 colspan=1>24.86%</td></tr></table>",
|
| 1849 |
+
"bbox": [
|
| 1850 |
+
232,
|
| 1851 |
+
257,
|
| 1852 |
+
761,
|
| 1853 |
+
469
|
| 1854 |
+
],
|
| 1855 |
+
"page_idx": 13
|
| 1856 |
+
},
|
| 1857 |
+
{
|
| 1858 |
+
"type": "table",
|
| 1859 |
+
"img_path": "images/5ae3a3d8d9957674fc0c542c4157bdf82ab39e324ece4f6e93c4ef53137242d6.jpg",
|
| 1860 |
+
"table_caption": [
|
| 1861 |
+
"Table 10: The performance evaluation on RDI-ResNet38 (defended with PGD) against WRM attack. The perturbation size is 0.3. The ATA of the original defended ResNet38 by PGD under the same attacker is $8 3 . 3 5 \\%$ . "
|
| 1862 |
+
],
|
| 1863 |
+
"table_footnote": [],
|
| 1864 |
+
"table_body": "<table><tr><td rowspan=1 colspan=1>Defence Method</td><td rowspan=1 colspan=1>Standard</td><td rowspan=1 colspan=1>Main Branch</td><td rowspan=1 colspan=1>Average</td><td rowspan=1 colspan=1>Max-Average</td></tr><tr><td rowspan=1 colspan=1>TA</td><td rowspan=1 colspan=1>92.43%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch1)</td><td rowspan=1 colspan=1>46.60%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch2)</td><td rowspan=1 colspan=1>71.33%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.79%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch3)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.41%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch4)</td><td rowspan=1 colspan=1>33.41%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch5)</td><td rowspan=1 colspan=1>42.35%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.41%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Branch6)</td><td rowspan=1 colspan=1>47.77%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.40%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Main Branch)</td><td rowspan=1 colspan=1>34.42%</td><td rowspan=1 colspan=1>83.74%</td><td rowspan=1 colspan=1>82.42%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Average)</td><td rowspan=1 colspan=1>26.48%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>ATA (Max-Average)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.73%</td><td rowspan=1 colspan=1>82.40%</td><td rowspan=1 colspan=1>83.78%</td></tr><tr><td rowspan=1 colspan=1>ATA (Worst-Case)</td><td rowspan=1 colspan=1>23.51%</td><td rowspan=1 colspan=1>83.69%</td><td rowspan=1 colspan=1>82.36%</td><td rowspan=1 colspan=1>83.77%</td></tr><tr><td rowspan=1 colspan=1>Average MFlops</td><td rowspan=1 colspan=1>50.05</td><td rowspan=1 colspan=1>50.46</td><td rowspan=1 colspan=1>52.89</td><td rowspan=1 colspan=1>52.38</td></tr><tr><td rowspan=1 colspan=1>Computation Saving</td><td rowspan=1 colspan=1>36.98%</td><td rowspan=1 colspan=1>36.46%</td><td rowspan=1 colspan=1>33.40%</td><td rowspan=1 colspan=1>34.04%</td></tr></table>",
|
| 1865 |
+
"bbox": [
|
| 1866 |
+
232,
|
| 1867 |
+
536,
|
| 1868 |
+
761,
|
| 1869 |
+
747
|
| 1870 |
+
],
|
| 1871 |
+
"page_idx": 13
|
| 1872 |
+
}
|
| 1873 |
+
]
|
parse/train/rJgzzJHtDB/rJgzzJHtDB_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/rJgzzJHtDB/rJgzzJHtDB_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/rJljdh4KDH/rJljdh4KDH.md
ADDED
|
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MULTI-SCALE REPRESENTATION LEARNING FOR SPATIAL FEATURE DISTRIBUTIONS USING GRID CELLS
|
| 2 |
+
|
| 3 |
+
Gengchen Mai1, Krzysztof Janowicz1, Bo Yan2, Rui $\mathbf { Z } \mathbf { h } \mathbf { u } ^ { 1 }$ , Ling Cai1 & Ni Lao3
|
| 4 |
+
1STKO Lab, University of California, Santa Barbara, CA, USA, 93106
|
| 5 |
+
{gengchen_mai,janowicz,ruizhu,lingcai}@ucsb.edu
|
| 6 |
+
2LinkedIn Corporation, Mountain View, CA, USA, 94043
|
| 7 |
+
boyan1@linkedin.com
|
| 8 |
+
3SayMosaic Inc., Palo Alto, CA, USA, 94303
|
| 9 |
+
ni.lao@mosaix.ai
|
| 10 |
+
|
| 11 |
+
# ABSTRACT
|
| 12 |
+
|
| 13 |
+
Unsupervised text encoding models have recently fueled substantial progress in Natural Language Processing (NLP). The key idea is to use neural networks to convert words in texts to vector space representations (embeddings) based on word positions in a sentence and their contexts, which are suitable for end-to-end training of downstream tasks. We see a strikingly similar situation in spatial analysis, which focuses on incorporating both absolute positions and spatial contexts of geographic objects such as Points of Interest (POIs) into models. A general-purpose representation model for space is valuable for a multitude of tasks. However, no such general model exists to date beyond simply applying discretization or feedforward nets to coordinates, and little effort has been put into jointly modeling distributions with vastly different characteristics, which commonly emerges from GIS data. Meanwhile, Nobel Prize-winning Neuroscience research shows that grid cells in mammals provide a multi-scale periodic representation that functions as a metric for location encoding and is critical for recognizing places and for path-integration. Therefore, we propose a representation learning model called Space2Vec to encode the absolute positions and spatial relationships of places. We conduct experiments on two real-world geographic data for two different tasks: 1) predicting types of POIs given their positions and context, 2) image classification leveraging their geo-locations. Results show that because of its multiscale representations, Space2Vec outperforms well-established ML approaches such as RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for location modeling and image classification tasks. Detailed analysis shows that all baselines can at most well handle distribution at one scale but show poor performances in other scales. In contrast, Space2Vec ’s multi-scale representation can handle distributions at different scales.
|
| 14 |
+
|
| 15 |
+
# 1 INTRODUCTION
|
| 16 |
+
|
| 17 |
+
Unsupervised text encoding models such as Word2Vec (Mikolov et al., 2013), Glove (Pennington et al., 2014), ELMo (Peters et al., 2018), and BERT (Devlin et al., 2018) have been effectively utilized in many Natural Language Processing (NLP) tasks. At their core they train models which encode words into vector space representations based on their positions in the text and their context. A similar situation can be encountered in the field of Geographic Information Science (GIScience). For example, spatial interpolation aims at predicting an attribute value, e.g., elevation, at an unsampled location based on the known attribute values of nearby samples. Geographic information has become an important component to many tasks such as fine-grained image classification (Mac Aodha et al., 2019), point cloud classification and semantic segmentation (Qi et al., 2017), reasoning about Point of Interest (POI) type similarity (Yan et al., 2017), land cover classification (Kussul et al., 2017), and geographic question answering (Mai et al., 2019b). Developing a general model for vector space representation of any point in space would pave the way for many future applications.
|
| 18 |
+
|
| 19 |
+

|
| 20 |
+
Figure 1: The challenge of joint modeling distributions with very different characteristics. (a)(b) The POI locations (red dots) in Las Vegas and Space2Vec predicted conditional likelihood of Women’s Clothing (with a clustered distribution) and Education (with an even distribution). The dark area in (b) indicates that the downtown area has more POIs of other types than education. (c) Ripley’s K curves of POI types for which Space2Vec has the largest and smallest improvement over wrap (Mac Aodha et al., 2019). Each curve represents the number of POIs of a certain type inside certain radios centered at every POI of that type; (d) Ripley’s K curves renormalized by POI densities and shown in log-scale. To efficiently achieve multi-scale representation Space2Vec concatenates the grid cell encoding of 64 scales (with wave lengths ranging from 50 meters to $4 0 k$ meters) as the first layer of a deep model, and trains with POI data in an unsupervised fashion.
|
| 21 |
+
|
| 22 |
+
However, existing models often utilize specific methods to deal with geographic information and often disregards geographic coordinates. For example, Place2Vec (Yan et al., 2017) converts the coordinates of POIs into spatially collocated POI pairs within certain distance bins, and does not preserve information about the (cardinal) direction between points. Li et al. (2017) propose DCRNN for traffic forecasting in which the traffic sensor network is converted to a distance weighted graph which necessarily forfeits information about the spatial layout of sensors. There is, however, no general representation model beyond simply applying discretization (Berg et al., 2014; Tang et al., 2015) or feed-forward nets (Chu et al., 2019; Mac Aodha et al., 2019) to coordinates.
|
| 23 |
+
|
| 24 |
+
A key challenge in developing a general-purpose representation model for space is how to deal with mixtures of distributions with very different characteristics (see an example in Figure 1), which often emerges in spatial datasets (McKenzie et al., 2015). For example, there are POI types with clustered distributions such as women’s clothing, while there are other POI types with regular distributions such as education. These feature distributions co-exist in the same space, and yet we want a single representation to accommodate all of them in a task such as location-aware image classification (Mac Aodha et al., 2019). Ripley’s K is a spatial analysis method used to describe point patterns over a given area of interest. Figure 1c shows the K plot of several POI types in Las Vegas. One can see that as the radius grows the numbers of POIs increase at different rates for different POI types. In order to see the relative change of density at different scales, we renormalize the curves by each POI type’s density and show it in log scale in Figure 1d. One can see two distinct POI type groups with different distribution patterns with clustered and even distributions. If we want to model the distribution of these POIs by discretizing the study area into tiles, we have to use small grid sizes for women’s clothing while using larger grid sizes for educations because smaller grid sizes lead to over- parameterization of the model and overfitting. In order to jointly describe these distributions and their patterns, we need an encoding method which supports multi-scale representations.
|
| 25 |
+
|
| 26 |
+
Nobel Prize winning Neuroscience research (Abbott & Callaway, 2014) has demonstrated that grid cells in mammals provide a multi-scale periodic representation that functions as a metric for location encoding, which is critical for integrating self-motion. Moreover, Blair et al. (2007) show that the multi-scale periodic representation of grid cells can be simulated by summing three cosine grating functions oriented $6 0 ^ { \circ }$ apart, which may be regarded as a simple Fourier model of the hexagonal lattice. This research inspired us to encode locations with multi-scale periodic representations. Our assumption is that decomposed geographic coordinates helps machine learning models, such as deep neural nets, and multi-scale representations deal with the inefficiency of intrinsically single-scale methods such as RFB kernels or discretization (tile embeddings). To validate this intuition, we propose an encoder-decoder framework to encode the distribution of point-features2 in space and train such a model in an unsupervised manner. This idea of using sinusoid functions with different frequencies to encode positions is similar to the position encoding proposed in the Transformer model (Vaswani et al., 2017). However, the position encoding model of Transformer deals with a discrete 1D space – the positions of words in a sentence – while our model works on higher dimensional continuous spaces such as the surface of earth.
|
| 27 |
+
|
| 28 |
+
# In summary, the contributions of our work are as follows:
|
| 29 |
+
|
| 30 |
+
1. We propose an encoder-decoder encoding framework called Space2Vec using sinusoid functions with different frequencies to model absolute positions and spatial contexts. We also propose a multi-head attention mechanism based on context points. To the best of our knowledge, this is the first attention model that explicitly considers the spatial relationships between the query point and context points.
|
| 31 |
+
2. We conduct experiments on two real world geographic data for two different tasks: 1) predicting types of POIs given their positions and context, 2) image classification leveraging their geo-locations. Space2Vec outperforms well-established encoding methods such as RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for location modeling and image classification.
|
| 32 |
+
3. To understand the advantages of Space2Vec we visualize the firing patterns (response maps) of location models’ encoding layer neurons and show how they handle spatial structures at different scales by integrating multi-scale representations. Furthermore the firing patterns for the spatial context models neurons give insight into how the grid-like cells capture the decreasing distance effect with multi-scale representations.
|
| 33 |
+
|
| 34 |
+
# 2 PROBLEM FORMULATION
|
| 35 |
+
|
| 36 |
+
Distributed representation of point-features in space can be formulated as follows. Given a set of points $\mathcal { P } _ { - } = \left\{ p _ { i } \right\}$ , i.e., Points of Interests (POIs), in $L { \mathrm { - } } \mathbf { D }$ space $( L = 2 , 3$ ) define a function $f _ { \mathcal { P } , \theta } ( \mathbf { x } ) : \mathbb { R } ^ { L } \mathbf { \bar { \mathbb { R } } } ^ { \dot { d } }$ $( L \ll d )$ , which is parameterized by $\theta$ and maps any coordinate $\mathbf { x }$ in space to a vector representation of $d$ dimension. Each point (e.g., a restaurant) $p _ { i } = ( \mathbf { x } _ { i } , \mathbf { v } _ { i } )$ is associated with a location $\mathbf { x } _ { i }$ and attributes $\mathbf { v } _ { i }$ (i.e., POI features such as type, name, capacity, etc.). The function $f _ { \mathcal { P } , \theta } ( \mathbf { x } )$ encodes the probability distribution of point features over space and can give a representation of any point in the space. Attributes (e.g. place types such as Museum) and coordinate of point can be seen as analogies to words and word positions in commonly used word embedding models.
|
| 37 |
+
|
| 38 |
+
# 3 RELATED WORK
|
| 39 |
+
|
| 40 |
+
There has been theoretical research on neural network based path integration/spatial localization models and their relationships with grid cells. Both Cueva & Wei (2018) and Banino et al. (2018) showed that grid-like spatial response patterns emerge in trained networks for navigation tasks which demonstrate that grid cells are critical for vector-based navigation. Moreover, Gao et al. (2019) propose a representational model for grid cells in navigation tasks which has good quality such as magnified local isometry. All these research is focusing on understanding the relationship between the grid-like spatial response patterns and navigation tasks from a theoretical perspective. In contrast, our goal focuses on utilizing these theoretical results on real world data in geoinformatics.
|
| 41 |
+
|
| 42 |
+
Radial Basis Function (RBF) kernel is a well-established approach to generating learning friendly representation from points in space for machine learning algorithms such as SVM classification (Baudat & Anouar, 2001) and regression (Bierens, 1994). However, the representation is example based – i.e., the resultant model uses the positions of training examples as the centers of Gaussian kernel functions (Maz’ya & Schmidt, 1996). In comparison, the grid cell based location encoding relies on sine and cosine functions, and the resultant model is inductive and does not store training examples.
|
| 43 |
+
|
| 44 |
+
Recently the computer vision community shows increasing interests in incorporating geographic information (e.g. coordinate encoding) into neural network architectures for multiple tasks such as image classification (Tang et al., 2015) and fine grained recognition (Berg et al., 2014; Chu et al., 2019; Mac Aodha et al., 2019). Both Berg et al. (2014) and Tang et al. (2015) proposed to discretize the study area into regular grids. To model the geographical prior distribution of the image categories, the grid id is used for GPS encoding instead of the raw coordinates. However, choosing the correct discretization is challenging (Openshaw, 1984; Fotheringham & Wong, 1991), and incorrect choices can significantly affect the final performance (Moat et al., 2018; Lechner et al., 2012). In addition, discretization does not scale well in terms of memory use. To overcome these difficulties, both Chu et al. (2019) and Mac Aodha et al. (2019) advocated the idea of inductive location encoders which directly encode coordinates into a location embedding. However, both of them directly feed the coordinates into a feed-forward neural network (Chu et al., 2019) or residual blocks (Mac Aodha et al., 2019) without any feature decomposition strategy. Our experiments show that this direct encoding approach is insufficient to capture the spatial feature distribution and Space2Vec significantly outperforms them by integrating spatial representations of different scales.
|
| 45 |
+
|
| 46 |
+
# 4 METHOD
|
| 47 |
+
|
| 48 |
+
We solve distributed representation of point-features in space (defined in Section 2) with an encoderdecoder architecture:
|
| 49 |
+
|
| 50 |
+
1. Given a point $p _ { i } = ( \mathbf { x } _ { i } , \mathbf { v } _ { i } )$ a point space encoder $E n c ^ { ( x ) } ( \ v r )$ encodes location $\mathbf { x } _ { i }$ into a location embedding $\mathbf { e } [ \mathbf { x } _ { i } ] \in \mathbb { R } ^ { d ^ { ( x ) } }$ and a point feature encoder $E n c ^ { ( v ) } ( )$ encodes its feature into a feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ] \in \mathbb { R } ^ { d ^ { ( v ) } }$ . $\mathbf { e } = [ \mathbf { e } [ \mathbf { x } _ { i } ] ; \mathbf { e } [ \mathbf { v } _ { i } ] ] \in \mathbb { R } ^ { d }$ is the full representation of point $p _ { i } \in \mathcal { P }$ , where $d = d ^ { ( x ) } + d ^ { ( v ) }$ . $[ ; ]$ represents vector concatenation. In contrast, geographic entities not in $\mathcal { P }$ within the studied space can be represented by their location embedding $\mathbf { e } [ \mathbf { x } _ { j } ]$ since its $\mathbf { v } _ { i }$ is unknown. 2. We developed two types of decoders which can be used independently or jointly. A location decoder $D e c _ { s } ( )$ reconstructs point feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ given location embedding $\mathbf { e } [ \mathbf { x } _ { i } ]$ , and a spatial context decoder $D e c _ { c } ( )$ reconstructs the feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ of point $p _ { i }$ based on the space and feature embeddings $\left\{ \mathbf { e } _ { i 1 } , . . . , \mathbf { e } _ { i j } , . . . , \mathbf { e } _ { i n } \right\}$ of nearest neighboring points $\{ p _ { i 1 } , . . . , p _ { i j } , . . . , p _ { i n } \}$ , where $n$ is a hyper-parameter.
|
| 51 |
+
|
| 52 |
+
# 4.1 ENCODER
|
| 53 |
+
|
| 54 |
+
Point Feature Encoder Each point $p _ { i } = ( \mathbf { x } _ { i } , \mathbf { v } _ { i } )$ in a point set $\mathcal { P }$ is often associated with features such as the air pollution station data associate with some air quality measures, a set of POIs with POI types and names, a set of points from survey and mapping with elevation values, a set of points from geological survey with mineral content measure, and so on. The point feature encoder $E n c ^ { ( v ) } ( )$ encodes such features $\mathbf { v } _ { i }$ into a feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ] \in \mathbb { R } ^ { d ^ { ( v ) } }$ . The implementation of $E n c ^ { ( v ) } ( )$ depends on the nature of these features. For example, if each point represents a POI with multiple POI types (as in this study), the feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ can simply be the mean of each POI types’ embeddings $\mathbf { e } [ \mathbf { v } _ { i } ] = \frac { 1 } { H } \sum _ { h = 1 } ^ { H } \mathbf { t } _ { h } ^ { ( \gamma ) }$ , where $\mathbf { t } _ { h } ^ { \left( \gamma \right) }$ indicates the $h$ th POI type embedding of a POI $p _ { i }$ with $H$ POI types. We apply $L _ { 2 }$ normalization to the POI type embedding matrix.
|
| 55 |
+
|
| 56 |
+
Point Space Encoder A part of the novelty of this paper is from the point space encoder $E n c ^ { ( x ) } ( \ v r )$ . We first introduce Theorem 1 which provide an analytical solution $\phi ( \mathbf { x } )$ as the base of encoding any location $\mathbf { x } \in \mathbb { R } ^ { 2 }$ in 2D space to a distributed representation:
|
| 57 |
+
|
| 58 |
+
Theorem 1. Let $\Psi ( \mathbf { x } ) = ( e ^ { i \langle \mathbf { a } _ { j } , \mathbf { x } \rangle } , j = 1 , 2 , 3 ) ^ { T } \in \mathbb { C } ^ { 3 }$ where $e ^ { i \theta } = \cos \theta + i \sin \theta$ is the Euler notation of complex values; $\langle \mathbf { a } _ { j } , \mathbf { x } \rangle$ is the inner product of ${ \bf a } _ { j }$ and $\mathbf { x } .$ $\mathfrak { c } . \ \mathbf { a } _ { 1 } , \mathbf { a } _ { 2 } , \mathbf { a } _ { 3 } \in \mathbb { R } ^ { 2 }$ are $2 D$ vectors such that the angle between ${ \bf a } _ { k }$ and ${ \bf a } _ { l }$ is $2 \pi / 3 , \forall j$ , $\| \mathbf { a } _ { j } \| = 2 { \sqrt { \alpha } }$ . Let $\mathbf { C } \in \mathbb { C } ^ { 3 \times 3 }$ be a random complex matrix such as $\mathbf { C } ^ { * } \mathbf { C } = \mathbf { I } .$ . Then $\phi ( \mathbf { x } ) = \mathbf { C } \Psi ( \mathbf { x } )$ , $M ( \Delta \mathbf { x } ) = \mathbf { C } d i a g ( \pmb { \Psi } ( \Delta \mathbf { x } ) ) \mathbf { C } ^ { * }$ satisfies
|
| 59 |
+
|
| 60 |
+
$$
|
| 61 |
+
\phi ( \mathbf { x } + \Delta \mathbf { x } ) = M ( \Delta \mathbf { x } ) \phi ( \mathbf { x } )
|
| 62 |
+
$$
|
| 63 |
+
|
| 64 |
+
and
|
| 65 |
+
|
| 66 |
+
$$
|
| 67 |
+
\langle \phi ( \mathbf { x } + \Delta \mathbf { x } ) , \phi ( \mathbf { x } ) \rangle = d ( 1 - \alpha \| \Delta \mathbf { x } \| ^ { 2 } )
|
| 68 |
+
$$
|
| 69 |
+
|
| 70 |
+
where $d = 3$ is the dimension of $\phi ( \mathbf { x } )$ and $\Delta \mathbf { x }$ is a small displacement from x.
|
| 71 |
+
|
| 72 |
+
The proof of Theorem 1 can be seen in Gao et al. (2019). $\phi ( \mathbf { x } ) = \mathbf { C } \pmb { \Psi } ( \mathbf { x } ) \in \mathbb { C } ^ { 3 }$ amounts to a 6-dimension real value vector and each dimension shows a hexagon firing pattern which models the grid cell behavior. Because of the periodicity of $s i n ( )$ and $c o s ( )$ , this single scale representation $\phi ( \mathbf { x } )$ does not form a global codebook of 2D positions, i.e. there can be $\mathbf x \neq \mathbf y$ , but $\phi ( \mathbf { x } ) = \phi ( \mathbf { y } )$ .
|
| 73 |
+
|
| 74 |
+
Inspired by Theorem 1 and the multi-scale periodic representation of grid cells in mammals (Abbott & Callaway, 2014) we set up our point space encoder $\mathbf { e } [ \mathbf { x } ] \stackrel { - } { = } E n c _ { t h e o r y } ^ { ( x ) } ( \mathbf { x } )$ to use sine and cosine functions of different frequencies to encode positions in space. Given any point $\mathbf { x }$ in the studied 2D space, the space encoder $E n c _ { t h e o r y } ^ { ( x ) } ( \mathbf { x } ) \ = \ \mathbf { N N } ( P E ^ { ( t ) } ( \mathbf { x } ) )$ where $P E ^ { ( t ) } ( \mathbf { x } ) \ =$ $[ P E _ { 0 } ^ { ( t ) } ( { \bf x } ) ; . . . ; P E _ { s } ^ { ( t ) } ( { \bf x } ) ; . . . ; P E _ { S - 1 } ^ { ( t ) } ( { \bf x } ) ]$ is a concatenation of multi-scale representations of $d ^ { ( x ) } =$ $6 S$ dimensions. Here $S$ is the total number of grid scales and $s = 0 , 1 , 2 , . . . , S - 1 . \ \mathbf { N N } ( )$ represents fully connected ReLU layers. Let $\mathbf { a } _ { 1 } = [ 1 , 0 ] ^ { T } , \mathbf { a } _ { 2 } = [ - 1 / 2 , \sqrt { 3 } / 2 ] ^ { T } , \mathbf { a } _ { 3 } = [ - 1 / 2 , - \sqrt { 3 } / 2 ] ^ { T } \in \mathbb { R } ^ { 2 }$ be three unit vectors and the angle between any of them is $2 \pi / 3$ . $\lambda _ { m i n } , \lambda _ { m a x }$ are the minimum and maximum grid scale and $\begin{array} { r } { g = \frac { \lambda _ { m a x } } { \lambda _ { m i n } } } \end{array}$ . At each scale $s , P E _ { s } ^ { ( t ) } ( \mathbf { x } ) = [ P E _ { s , 1 } ^ { ( t ) } ( \mathbf { x } ) ; P E _ { s , 2 } ^ { ( t ) } ( \mathbf { x } ) ; P E _ { s , 3 } ^ { ( t ) } ( \mathbf { x } ) ]$ is a concatenation of three components, where
|
| 75 |
+
|
| 76 |
+
$$
|
| 77 |
+
P E _ { s , j } ^ { ( t ) } ( \mathbf { x } ) = [ \cos ( \frac { \langle \mathbf { x } , \mathbf { a } _ { j } \rangle } { \lambda _ { m i n } \cdot g ^ { s / ( S - 1 ) } } ) ; \sin ( \frac { \langle \mathbf { x } , \mathbf { a } _ { j } \rangle } { \lambda _ { m i n } \cdot g ^ { s / ( S - 1 ) } } ) ] \forall j = 1 , 2 , 3 ;
|
| 78 |
+
$$
|
| 79 |
+
|
| 80 |
+
$\mathbf { N N } ( )$ and $P E ^ { ( t ) } ( \mathbf { x } )$ are analogies of $\mathbf { C }$ and $\Psi ( \mathbf { x } )$ in Theorem 1.
|
| 81 |
+
|
| 82 |
+
Similarly we can define another space encoder $E n c _ { g r i d } ^ { ( x ) } ( { \bf x } ) ~ = ~ { \bf N N } ( P E ^ { ( g ) } ( { \bf x } ) )$ inspired by the position encoding model of Transformer (Vaswani et al., 2017), where $\begin{array} { r l } { P E ^ { ( g ) } ( \mathbf { x } ) } & { { } = } \end{array}$ $[ { \bar { P E _ { 0 } ^ { ( g ) } } } ( \mathbf { x } ) ; . . . ; { \bar { P E _ { s } ^ { ( g ) } ( \mathbf { x } ) } } ; . . . ; { \bar { P E _ { S - 1 } ^ { ( g ) } ( \mathbf { x } ) } } ]$ is still a concatenation of its multi-scale representations, while $P E _ { s } ^ { ( g ) } ( \mathbf { x } ) = [ P E _ { s , 1 } ^ { ( g ) } ( \mathbf { x } ) ; P E _ { s , 2 } ^ { ( g ) } ( \mathbf { x } ) ]$ handles each component $l$ of $\mathbf { x }$ separately:
|
| 83 |
+
|
| 84 |
+
$$
|
| 85 |
+
P E _ { s , l } ^ { ( g ) } ( \mathbf { x } ) = [ \cos ( \frac { \mathbf { x } ^ { [ l ] } } { \lambda _ { m i n } \cdot g ^ { s / ( S - 1 ) } } ) ; \sin ( \frac { \mathbf { x } ^ { [ l ] } } { \lambda _ { m i n } \cdot g ^ { s / ( S - 1 ) } } ) ] \forall l = 1 , 2
|
| 86 |
+
$$
|
| 87 |
+
|
| 88 |
+
# 4.2 DECODER
|
| 89 |
+
|
| 90 |
+
Two types of decoders are designed for two major types of GIS problems: location modeling and spatial context modeling (See Section 5.1).
|
| 91 |
+
|
| 92 |
+
Location Decoder $D e c _ { s } ( )$ directly reconstructs point feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ given its space embedding $\mathbf { e } [ \mathbf { x } _ { i } ]$ . We use one layer feed-forward neural network $\mathbf { N N } _ { \mathrm { d e c } } ($ q
|
| 93 |
+
|
| 94 |
+
$$
|
| 95 |
+
\mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime } = D e c _ { s } ( \mathbf { x } _ { i } ; \theta _ { \mathrm { d e c } _ { s } } ) = \mathbf { N } \mathbf { N } _ { \mathrm { d e c } } ( \mathbf { e } [ \mathbf { x } _ { i } ] )
|
| 96 |
+
$$
|
| 97 |
+
|
| 98 |
+
For training we use inner product to compare the reconstructed feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime }$ against the real feature embeddings of $\mathbf { e } [ \mathbf { v } _ { i } ]$ and other negative points (see training detail in Sec 4.3).
|
| 99 |
+
|
| 100 |
+
Spatial Context Decoder $D e c _ { c } ( )$ reconstructs the feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ of the center point $p _ { i }$ based on the space and feature embeddings $\left\{ \mathbf { e } _ { i 1 } , . . . , \mathbf { e } _ { i j } , . . . , \mathbf { e } _ { i n } \right\}$ of $n$ nearby points $\{ p _ { i 1 } , . . . , p _ { i j } , . . . , p _ { i n } \}$ . Note that the feed-in order of context points should not affect the prediction results, which can be achieved by permutation invariant neural network architectures (Zaheer et al., 2017) like PointNet (Qi et al., 2017).
|
| 101 |
+
|
| 102 |
+
$$
|
| 103 |
+
\mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime } = D e c _ { c } ( \mathbf { x } _ { i } , \{ \mathbf { e } _ { i 1 } , . . . , \mathbf { e } _ { i j } , . . . , \mathbf { e } _ { i n } \} ; \theta _ { \mathrm { d e c } _ { c } } ) = g ( \frac { 1 } { K } \sum _ { k = 1 } ^ { K } \sum _ { j = 1 } ^ { n } \alpha _ { i j k } \mathbf { e } [ \mathbf { v } _ { i j } ] )
|
| 104 |
+
$$
|
| 105 |
+
|
| 106 |
+
Here g is an activation function such as sigmoid. αijk “ $\begin{array} { r } { \alpha _ { i j k } = \frac { e x p ( \sigma _ { i j k } ) } { \sum _ { o = 1 } ^ { n } e x p ( \sigma _ { i o k } ) } } \end{array}$ is the attention of $p _ { i }$ with its $j$ th neighbor through the $k$ th attention head, and
|
| 107 |
+
|
| 108 |
+
$$
|
| 109 |
+
\sigma _ { i j k } = L e a k y R e L U ( \mathbf { a } _ { k } ^ { T } [ \mathbf { e } [ \mathbf { v } _ { i } ] _ { i n i t } ; \mathbf { e } [ \mathbf { v } _ { i j } ] ; \mathbf { e } [ \mathbf { x } _ { i } - \mathbf { x } _ { i j } ] ] )
|
| 110 |
+
$$
|
| 111 |
+
|
| 112 |
+
where $\mathbf { a } _ { k } \in \mathbb { R } ^ { 2 d ^ { ( v ) } + d ^ { ( x ) } }$ is the attention parameter in the $k$ th attention head. The multi-head attention mechanism is inspired by Graph Attention Network (Velickovi ˇ c et al., 2018) and Mai et al. (2019a). ´
|
| 113 |
+
|
| 114 |
+
To represent the spatial relationship (distance and direction) between each context point $\begin{array} { r l } { p _ { i j } } & { { } = } \end{array}$ $( \mathbf { x } _ { i j } , \mathbf { v } _ { i j } )$ and the center point $p _ { i } = ( \mathbf { x } _ { i } , \mathbf { v } _ { i } )$ , we use the space encoder $E n c ^ { ( x ) } ( \ v r )$ to encode the displacement between them $\Delta { \bf x } _ { i j } = { \bf x } _ { i } - { \bf x } _ { i j }$ . Note that we are modeling the spatial interactions between the center point and $n$ context points simultaneously.
|
| 115 |
+
|
| 116 |
+
In Eq. $7 , \mathbf { e } [ \mathbf { v } _ { i } ] _ { i n i t }$ indicates the initial guess of the feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ of point $p _ { i }$ which is computed by using another multi-head attention layer as Eq. 6 where the weight $\begin{array} { r } { \alpha _ { i j k } ^ { \prime } = \frac { e x p ( \sigma _ { i j k } ^ { \prime } ) } { \sum _ { o = 1 } ^ { n } e x p ( \sigma _ { i o k } ^ { \prime } ) } } \end{array}$ . Here, $\sigma _ { i j k } ^ { \prime }$ is computed as Eq. 8 where the query embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ is excluded.
|
| 117 |
+
|
| 118 |
+
$$
|
| 119 |
+
\sigma _ { i j k } ^ { \prime } = L e a k y R e L U ( \mathbf { a } _ { k } ^ { \prime T } [ \mathbf { e } [ \mathbf { v } _ { i j } ] ; \mathbf { e } [ \mathbf { x } _ { i } - \mathbf { x } _ { i j } ] ] )
|
| 120 |
+
$$
|
| 121 |
+
|
| 122 |
+
# 4.3 UNSUPERVISED TRAINING
|
| 123 |
+
|
| 124 |
+
The unsupervised learning task can simply be maximizing the log likelihood of observing the true point $p _ { i }$ at position $\mathbf { x } _ { i }$ among all the points in $\mathcal { P }$
|
| 125 |
+
|
| 126 |
+
$$
|
| 127 |
+
\mathcal { L } _ { \mathcal { P } } ( \boldsymbol { \theta } ) = - \sum _ { p _ { i } \in \mathcal { P } } \log P ( p _ { i } | p _ { i 1 } , . . . , p _ { i j } , . . . , p _ { i n } ) = - \sum _ { p _ { i } \in \mathcal { P } } \log \frac { \exp ( \mathbf { e } [ \mathbf { v } _ { i } ] ^ { T } \mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime } ) } { \sum _ { p _ { o } \in \mathcal { P } } \exp ( \mathbf { e } [ \mathbf { v } _ { o } ] ^ { T } \mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime } ) }
|
| 128 |
+
$$
|
| 129 |
+
|
| 130 |
+
Here only the feature embedding of $p _ { i }$ is used (without location embedding) to prevent revealing the identities of the point candidates, and $\theta = [ \theta _ { \mathrm { { e n c } } } ; \theta _ { \mathrm { { d e c } } } ]$
|
| 131 |
+
|
| 132 |
+
Negative sampling by Mikolov et al. (2013) can be used to improve the efficiency of training
|
| 133 |
+
|
| 134 |
+
$$
|
| 135 |
+
\mathcal { L } _ { \mathcal { P } } ^ { \prime } ( \boldsymbol { \theta } ) = - \sum _ { p _ { i } \in \mathcal { P } } \Big ( \log \sigma ( \mathbf { e } [ \mathbf { v } _ { i } ] ^ { T } \mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime } ) + \frac { 1 } { | \mathcal { N } _ { i } | } \sum _ { p _ { o } \in \mathcal { N } _ { i } } \log \sigma ( - \mathbf { e } [ \mathbf { v } _ { o } ] ^ { T } \mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime } ) \Big )
|
| 136 |
+
$$
|
| 137 |
+
|
| 138 |
+
Here ${ \mathcal { N } } _ { i } \subseteq { \mathcal { P } }$ is a set of sampled negative points for $p _ { i } \ ( p _ { i } \notin { \mathcal { N } } _ { i } )$ and $\sigma ( x ) = 1 / ( 1 + e ^ { - x } )$ .
|
| 139 |
+
|
| 140 |
+
# 5 EXPERIMENT
|
| 141 |
+
|
| 142 |
+
In this section we compare Space2Vec with commonly used position encoding methods, and analyze them both quantitatively and qualitatively.
|
| 143 |
+
|
| 144 |
+
Baselines Our baselines include 1) direct directly applying feed-forward nets (Chu et al., 2019); 2) tile discretization (Berg et al., 2014; Adams et al., 2015; Tang et al., 2015); 3) wrap feed-forward nets with coordinate wrapping (Mac Aodha et al., 2019); and 4) rbf Radial Basis Function (RBF) kernels (Baudat & Anouar, 2001; Bierens, 1994). See Appendix A.1 for details of the baselines.
|
| 145 |
+
|
| 146 |
+
# 5.1 POI TYPE CLASSIFICATION TASKS
|
| 147 |
+
|
| 148 |
+
Dataset and Tasks To test the proposed model, we conduct experiments on geographic datasets with POI position and type information. We utilize the open-source dataset published by Yelp Data Challenge and select all POIs within the Las Vegas downtown area3. There are 21,830 POIs with 1,191 different POI types in this dataset. Note that each POI may be associated with one or more types, and we do not use any other meta-data such as business names, reviews for this study. We project geographic coordinates into projection coordinates using the NAD83/Conus Albers projection coordinate system4. The POIs are split into training, validation, and test dataset with ratios $8 0 \% { : } 1 0 \% { : } 1 0 \%$ . We create two tasks setups which represent different types of modeling need in Geographic Information Science:
|
| 149 |
+
|
| 150 |
+
Location Modeling predicts the feature information associated with a POI based on its location $\mathbf { x } _ { i }$ represented by the location decoder $D e c _ { s } ( )$ . This represents a large number of location prediction problems such as image fine grained recognition with geographic prior (Chu et al., 2019), and species potential distribution prediction (Zuo et al., 2008). Spatial Context Modeling predicts the feature information associated with a POI based on its context $\left\{ \mathbf { e } _ { i 1 } , . . . , \mathbf { e } _ { i j } , . . . , \mathbf { e } _ { i n } \right\}$ represented by the spatial context decoder $D e c _ { c } ( )$ . This represents a collections of spatial context prediction problem such as spatial context based facade image classification (Yan et al., 2018), and all spatial interpolation problems.
|
| 151 |
+
|
| 152 |
+
We use POI prediction metrics to evaluate these models. Given the real point feature embedding $\mathbf { e } [ \mathbf { v } _ { i } ]$ and $N$ negative feature embeddings ${ \mathcal { N } } _ { i } = \{ \mathbf { e } [ \mathbf { v } _ { i } ] ^ { - } \}$ , we compare the predicted $\mathbf { e } [ \mathbf { v } _ { i } ] ^ { \prime }$ with them by cosine distance. The cosine scores are used to rank $\mathbf { e } [ \mathbf { v } _ { i } ]$ and $N$ negative samples. The negative feature embeddings are the feature embeddings of points $p _ { j }$ randomly sampled from $\mathcal { P }$ and $p _ { i } \neq p _ { j }$ . We evaluate each model using Negative Log-Likelihood (NLL), Mean Reciprocal Rank (MRR) and $\mathrm { H I T } @ 5$ (the chance of the true POI being ranked to top 5. We train and test each model 10 times to estimate standard deviations. See Appendix A.2 for hyper-parameter selection details.
|
| 153 |
+
|
| 154 |
+
# 5.1.1 LOCATION MODELING EVALUATION
|
| 155 |
+
|
| 156 |
+
We first study location modeling with the location decoder $D e c _ { s } ( )$ in Section 4.2. We use a negative sample size of $N = 1 0 0$ . Table 1 shows the average metrics of different models with their best hyperparameter setting on the validation set. We can see that direct and theorydiag are less competitive, only beating the random selection baseline. Other methods with single scale representations – including tile, wrap, and $r b f$ – perform better. The best results come from various version of the grid cell models, which are capable of dealing with multi-scale representations.
|
| 157 |
+
|
| 158 |
+

|
| 159 |
+
Figure 2: Embedding clustering of (a) direct; (b) tile with the best cell size $c = 5 0 0$ ; (c) wrap $( h = 3 , o =$ 512); (d) $r b f$ with the best $\sigma$ (1k) and 200 anchor points (red) and (e)(f)(h) theory models with different $\lambda _ { m i n }$ but fixed $\lambda _ { m a x } = 4 0 k$ and $S = 6 4$ . All models use 1 hidden ReLU layers of 512 neurons except wrap.
|
| 160 |
+
|
| 161 |
+
Table 1: The evaluation results of different location models on the validation and test dataset.
|
| 162 |
+
|
| 163 |
+
<table><tr><td rowspan="2"></td><td rowspan="2">Train NLL</td><td colspan="3">Validation</td><td colspan="2">Testing</td></tr><tr><td>NLL</td><td>MRR</td><td>HIT@5</td><td>MRR</td><td>HIT@5</td></tr><tr><td>random</td><td></td><td>-</td><td>0.052 (0.002)</td><td>4.8 (0.5)</td><td>0.051 (0.002)</td><td>5.0 (0.5)</td></tr><tr><td>direct</td><td>1.285</td><td>1.332</td><td>0.089 (0.001)</td><td>10.6 (0.2)</td><td>0.090 (0.001)</td><td>11.3 (0.2)</td></tr><tr><td>tile (c=500)</td><td>1.118</td><td>1.261</td><td>0.123 (0.001)</td><td>16.8 (0.2)</td><td>0.120 (0.001)</td><td>17.1 (0.3)</td></tr><tr><td>wrap(h=3,0=512)</td><td>1.222</td><td>1.288</td><td>0.112 (0.001)</td><td>14.6 (0.1)</td><td>0.119 (0.001)</td><td>15.8 (0.2)</td></tr><tr><td>rbf (σ=1k)</td><td>1.209</td><td>1.279</td><td>0.115 (0.001)</td><td>15.2 (0.2)</td><td>0.123 (0.001)</td><td>16.8 (0.3)</td></tr><tr><td>grid (Xmin=50)</td><td>1.156</td><td>1.258</td><td>0.128 (0.001)</td><td>18.1 (0.3)</td><td>0.139 (0.001)</td><td>20.0 (0.2)</td></tr><tr><td>hexa (Xmin=50)</td><td>1.230</td><td>1.297</td><td>0.107 (0.001)</td><td>14.0 (0.2)</td><td>0.105 (0.001)</td><td>14.5 (0.2)</td></tr><tr><td>theorydiag (Xmin=50)</td><td>1.277</td><td>1.324</td><td>0.094 (0.001)</td><td>12.3 (0.3)</td><td>0.094 (0.002)</td><td>11.2 (0.3)</td></tr><tr><td>theory (Xmin=1k)</td><td>1.207</td><td>1.281</td><td>0.123 (0.002)</td><td>16.3 (0.5)</td><td>0.121 (0.001)</td><td>16.2 (0.1)</td></tr><tr><td>theory (Xmin=500)</td><td>1.188</td><td>1.269</td><td>0.132 (0.001)</td><td>17.6 (0.3)</td><td>0.129 (0.001)</td><td>17.7 (0.2)</td></tr><tr><td>theory (Xmin=50)</td><td>1.098</td><td>1.249</td><td>0.137 (0.002)</td><td>19.4 (0.1)</td><td>0.144 (0.001)</td><td>20.0 (0.2)</td></tr></table>
|
| 164 |
+
|
| 165 |
+
In order to understand the reason for the superiority of grid cell models we provide qualitative analysis of their representations. We apply hierarchical clustering to the location embeddings produced by studied models using cosine distance as the distance metric (See Fig. 2). we can see that when restricted to large grid sizes $( \lambda _ { m i n } = 1 k )$ ), theory has similar representation (Fig. 2d, 2e, and Fig. 4d, 4e) and performance compared to $r b f$ $\mathit { \Pi } _ { \overline { { \sigma } } } = 1 k \mathit { \Pi } _ { \overline { { \theta } } }$ ). However it is able to significantly outperform $r b f$ $\sigma = 1 k$ ) (and tile and wrap) when small grid sizes $\langle \lambda _ { m i n } = 5 0 0 , 5 0 )$ are available. The relative improvements over $r b f$ $( \sigma = 1 k )$ ) are $- 0 . 2 \%$ , $+ 0 . 6 \%$ , $+ 2 . 1 \%$ MRR for $\lambda _ { m i n } { = } 1 \mathrm { k }$ , 500, 50 respectively.
|
| 166 |
+
|
| 167 |
+
# 5.1.2 MULTI-SCALE ANALYSIS OF LOCATION MODELING
|
| 168 |
+
|
| 169 |
+
In order to show how our multi-scale location representation model will affect the prediction of POI types with different distribution patterns, we classify all 1,191 POI types into three groups based on radius $r$ , which is derived from each POI types’ renormalized Ripley’s K curve (See Figure 1d for examples). It indicates the $\mathbf { X }$ axis value of the intersection between the curve and the line of $y = 3 . 0$ A lower $r$ indicates a more clustered distribution patterns. These three groups are listed below:
|
| 170 |
+
|
| 171 |
+
1. Clustered $( r \leqslant 1 0 0 m )$ ): POI types with clustered distribution patterns;
|
| 172 |
+
2. Middle ( $1 0 0 m < r < 2 0 0 m$ ): POI types with less extreme scales;
|
| 173 |
+
3. Even $r \geqslant 2 0 0 m$ ): POI types with even distribution patterns.
|
| 174 |
+
|
| 175 |
+
Table 2 shows the performance $( M R R )$ of direct, tile, wrap, rbf , and our theory model on the test dataset of the location modeling task with respect to these three different POI distribution groups. The numbers in pq indicate the MRR difference betweeb a baseline and theory. # POI refers to total number of POI belong to each group5. We can see that 1) The two neural net approaches (direct and wrap) have no scale related parameter and are not performing ideally across all scales, with direct
|
| 176 |
+
|
| 177 |
+
Table 2: Comparing performances in different POI groups. We classify all 1,191 POI types into three groups based on the radius $r$ of their root types, where their renormalized Ripley’s K curve (See Figure 1d) reach 3.0: 1) Clustered $( r \leqslant 1 0 0 m )$ ): POI types with clustered distribution patterns; 2) Middle $1 0 0 m < r < 2 0 0 m$ ): POI types with unclear distribution patterns; 3) Even $( r \geqslant 2 0 0 m )$ : POI types with even distribution patterns. The MRR of wrap and theory on those three groups are shown. The numbers in pq indicate the difference between the MRR of a baseline model and the MRR of theory with respect to a specific group. $\# P O I$ refers to the total number of POIs belonging to each group. Root Types indicates the root categories of those POI types belong to each group.
|
| 178 |
+
|
| 179 |
+
<table><tr><td>POI Groups</td><td>Clustered (r≤100m)</td><td>Middle (100m<r<200m)</td><td>Even (r≥ 200m)</td></tr><tr><td>direct</td><td>0.080 (-0.047)</td><td>0.108 (-0.030)</td><td>0.084 (-0.047)</td></tr><tr><td>wrap</td><td>0.106 (-0.021)</td><td>0.126 (-0.012)</td><td>0.122 (-0.009)</td></tr><tr><td>tile</td><td>0.108 (-0.019)</td><td>0.135 (-0.003)</td><td>0.111 (-0.020)</td></tr><tr><td>rbf</td><td>0.112 (-0.015)</td><td>0.136 (-0.002)</td><td>0.119 (-0.012)</td></tr><tr><td>theory</td><td>0.127 (-)</td><td>0.138(-)</td><td>0.131(-)</td></tr><tr><td>#POI</td><td>16,016</td><td>7,443</td><td>3,915</td></tr><tr><td>Root Types</td><td>Restaurants; Shopping; Food; Nightlife; Automotive; Active Life; Arts & Entertainment; Financial Services</td><td>Beauty & Spas; Health & Medical; Local Services;Hotels & Travel; Professional Services; Public Services & Government</td><td>Home Services; Event Planning & Services; Pets; Education</td></tr></table>
|
| 180 |
+
|
| 181 |
+
performs worse because of its simple single layer network. 2) The two approaches with built-in scale parameter (tile and $r b f$ ) have to trade off the performance of different scales. Their best parameter settings lead to close performances to that of Space2Vec at the middle scale, while performing poorly in both clustered and regular groups. These observation clearly shows that all baselines can at most well handle distribution at one scale but show poor performances in other scales. In contrast, Space2Vec’s multi-scale representation can handle distributions at different scales.
|
| 182 |
+
|
| 183 |
+
# 5.1.3 SPATIAL CONTEXT MODELING EVALUATION
|
| 184 |
+
|
| 185 |
+
Next, we evaluate the spatial context decoder $D e c _ { c } ( )$ in Sec. 4.2. We use the same evaluation set up as location modeling. The context points are obtained by querying the $n$ -th nearest points using PostGIS $( n = 1 0$ ). As for validation and test datasets, we make sure the center points are all unknown during the training phase. Table 3 shows the evaluation results of different models for spatial context modeling. The baseline approaches (direct, tile, wrap, $r b f$ ) generally perform poorly in context modeling. We designed specialized version of these approaches (polar, polar_tile, scaled_rbf) with polar coordinates, which lead to significantly improvements. Note that these are models proposed by us specialized for context modeling and therefore are less general than the grid cell approaches.
|
| 186 |
+
|
| 187 |
+
Table 3: The evaluation results of different spatial context models on the validation and test dataset. All encoders contains a 1 hidden layer FFN. All grid cell encoders set $\lambda _ { m i n } { = } 1 0$ , $\lambda _ { m a x } { = } 1 0 \mathbf { k }$ .
|
| 188 |
+
|
| 189 |
+
<table><tr><td rowspan="2">Space2Vec</td><td rowspan="2">Train NLL</td><td colspan="3"></td><td colspan="2">Testing</td></tr><tr><td>NLL</td><td>Validation MRR</td><td>HIT@5</td><td>MRR</td><td>HIT@5</td></tr><tr><td>none</td><td>1.163</td><td>1.297</td><td>0.159 (0.002)</td><td>22.4 (0.5)</td><td>0.167 (0.006)</td><td>23.4 (0.7)</td></tr><tr><td>direct</td><td>1.151</td><td>1.282</td><td>0.170 (0.002)</td><td>24.6 (0.4)</td><td>0.175 (0.003)</td><td>24.7 (0.5)</td></tr><tr><td>polar</td><td>1.157</td><td>1.283</td><td>0.176 (0.004)</td><td>25.4 (0.4)</td><td>0.178 (0.006)</td><td>24.9 (0.1)</td></tr><tr><td>tile(c = 50)</td><td>1.163</td><td>1.298</td><td>0.173 (0.004)</td><td>24.0 (0.6)</td><td>0.173 (0.001)</td><td>23.4 (0.1)</td></tr><tr><td>polar_tile(S =64)</td><td>1.161</td><td>1.282</td><td>0.173 (0.003)</td><td>25.0 (0.1)</td><td>0.177 (0.001)</td><td>24.5 (0.3)</td></tr><tr><td>wrap (h=2,0=512)</td><td>1.167</td><td>1.291</td><td>0.159 (0.001)</td><td>23.0 (0.1)</td><td>0.170 (0.001)</td><td>23.9 (0.2)</td></tr><tr><td>rbf (σ= 50)</td><td>1.160</td><td>1.281</td><td>0.179 (0.002)</td><td>25.2 (0.6)</td><td>0.172 (0.001)</td><td>25.0 (0.1)</td></tr><tr><td>scaled_rbf (o=40,β=0.1)</td><td>1.150</td><td>1.272</td><td>0.177 (0.002)</td><td>25.7 (0.1)</td><td>0.181 (0.001)</td><td>25.3 (0.1)</td></tr><tr><td>grid(Xmin=10)</td><td>1.172</td><td>1.285</td><td>0.178 (0.004)</td><td>24.9 (0.5)</td><td>0.181 (0.001)</td><td>25.1 (0.3)</td></tr><tr><td>hexa (Xmin=10)</td><td>1.156</td><td>1.289</td><td>0.173 (0.002)</td><td>24.0 (0.2)</td><td>0.183 (0.002)</td><td>25.3 (0.2)</td></tr><tr><td>theorydiag (Xmin = 10)</td><td>1.156</td><td>1.287</td><td>0.168 (0.001)</td><td>24.1 (0.4)</td><td>0.174 (0.005)</td><td>24.9 (0.1)</td></tr><tr><td>theory(Xmin=200)</td><td>1.168</td><td>1.295</td><td>0.159 (0.001)</td><td>23.1 (0.2)</td><td>0.170 (0.001)</td><td>23.2 (0.2)</td></tr><tr><td>theory(Xmin=50)</td><td>1.157</td><td>1.275</td><td>0.171 (0.001)</td><td>24.2 (0.3)</td><td>0.173 (0.001)</td><td>24.8 (0.4)</td></tr><tr><td>theory(Xmin=10)</td><td>1.158</td><td>1.280</td><td>0.177 (0.003)</td><td>25.2 (0.3)</td><td>0.185 (0.002)</td><td>25.7 (0.3)</td></tr></table>
|
| 190 |
+
|
| 191 |
+

|
| 192 |
+
Figure 3: Embedding clustering in the original space of (a) direct; (b) polar; (c) wrap, $h { = } 2 , o { = } 5 1 2$ ; (d) polar_tile, $S = 6 4$ , (e) scaled_ $. r b f$ , $\sigma = 4 0$ , $\beta { = } 0 . 1$ ; and (f) theory, $\lambda _ { m i n } = 1 0$ , $\lambda _ { m a x } = 1 0 k$ , $S = 6 4$ . $( \mathrm { g } ) ( \mathrm { h } ) ( \mathrm { i } ) ( \mathrm { j } ) ( \mathrm { k } ) ( \mathrm { l } )$ are the clustering results of the same models in the polar-distance space using $\log ( \Vert \ \Delta \mathbf { x } _ { i j } \ \Vert + 1 )$ . All models use 1 hidden ReLU (except wrap) layers of 512 neurons. Most models except wrap can capture a shift when distance is around $e ^ { 5 } - 1 \approx 1 5 0$ meters.
|
| 193 |
+
|
| 194 |
+
Nevertheless the grid cell approaches are able to perform better than the specialized approaches on the test dataset while have competitive performance on validation dataset. See Appendix ?? for the visualization of context models. Actually the gains are small for all baseline approaches also. The reason is that we expect location encoding to be less important when context information is accessible. Similarly as discussed in (Gao et al., 2019), it is when there is a lack of visual clues that the grid cells of animals are the most helpful for their navigation.
|
| 195 |
+
|
| 196 |
+
Figure 6 shows the location embedding clustering results in both Cartesian and polar coordinate systems. We can see that direct (Fig. 3a, 3g) only captures the distance information when the context POI is very close $( l o g ( \parallel \Delta \mathbf { x } _ { i j } \parallel + 1 ) \leqslant 5 $ ) while in the farther spatial context it purely models the direction information. polar (Fig. 3b, 3h) has the similar behaviors but captures the distance information in a more fine-grained manner. wrap (Fig. 3c, 3i) mainly focuses on differentiating relative positions in farther spatial context cont which might explain its lower performance6. polar_tile (Fig. 3d) mostly responds to distance information. Interestingly, scaled_rbf and theory have similar representations in the polar coordinate system (Fig. 3k, 3l) and similar performance (Table 3). While scaled_rbf captures the gradually decreased distance effect with a scaled kernel size which becomes larger in farther distance, theory achieves this by integrating representations of different scales.
|
| 197 |
+
|
| 198 |
+
# 5.2 FINE-GRAINED IMAGE CLASSIFICATION TASKS
|
| 199 |
+
|
| 200 |
+
To demonstrate the generalizability of Space2Vec for space representation we utilized the proposed point space encoder $E n c ^ { ( x ) } ( \ v r )$ model in a well-known computer vision task: fine-grained image classification. As we discussed in Section 3, many studies (Berg et al., 2014; Chu et al., 2019; Mac Aodha et al., 2019) have shown that geographic prior information - where (and when) the image is taken - is very important additional information for the fine-grained image classification task and can substantially improve the model performance. For example, the appearance information is usually not sufficient to differentiate two visually similar species. In this case, the geographic prior becomes much more important because these two species may have very different spatial prior distributions such as the example of European Toads and Spiny Toads in Figure 1 of Mac Aodha et al. (2019).
|
| 201 |
+
|
| 202 |
+
We adopt the task setup of Mac Aodha et al. (2019). During training we have a set of tuples $D =$ $\{ ( I _ { i } , \mathbf { x } _ { i } , y _ { i } , p _ { i } ) \mid i = 1 , . . . , N \}$ where $I _ { i }$ indicates an image, $y _ { i } \in \{ 1 , 2 , . . . , C \}$ is the corresponding class label (species category), $\mathbf { x } _ { i } = \left[ l o n g i t u d e _ { i } , l a t i t u d e _ { i } \right]$ is the geographic coordinates where the image was taken, and $p _ { i }$ is the id of the photographer who took this image. At training time, a location encoder is trained to capture the spatial prior information $P ( y \mid \mathbf { x } )$ . At inference time, $p _ { i }$ information is not available and the final image classification prediction is calculated based on the combination of two models: 1) the trained location encoder which captures the spatial priors $P ( y \mid \mathbf { x } )$ and 2) the pretrained image classification model, InceptionV3 network (Szegedy et al., 2016), which captures ${ \bar { P } } ( y \mid I )$ . Bayesian theory has been used to derive the joint distribution $P ( y \mid I , \mathbf { x } )$ . See Mac Aodha et al. (2019) for detail explanation as well as the loss function. Note that while Space2Vec outperforms specialized density estimation methods such as Adaptive Kernel (Berg et al., 2014), it would be interesting to explore early fusion Space2Vec ’s representations with the image module.
|
| 203 |
+
|
| 204 |
+
Table 4: Fine-grained image classification results on two datasets: BirdSnap: and NABirds:. The classification accuracy is calculated by combining image classification predictions $P ( y \mid I )$ with different spatial priors $P ( y \mid \mathbf { x } )$ . The grid and theory model use 1 hidden ReLU layers of 512 neurons. The evaluation results of the baseline models are from Table 1 of Mac Aodha et al. (2019).
|
| 205 |
+
|
| 206 |
+
<table><tr><td></td><td>BirdSnapt</td><td>NABirdst</td></tr><tr><td>No Prior (i.e.uniform)</td><td>70.07</td><td>76.08</td></tr><tr><td>Nearest Neighbor (num)</td><td>77.76</td><td>79.99</td></tr><tr><td>Nearest Neighbor (spatial)</td><td>77.98</td><td>80.79</td></tr><tr><td>Adaptive Kernel (Berg et al., 2014)</td><td>78.65</td><td>81.11</td></tr><tr><td>tile (Tang et al.,2015) (location only)</td><td>77.19</td><td>79.58</td></tr><tr><td>wrap (Mac Aodha et al.,2019) (location only)</td><td>78.65</td><td>81.15</td></tr><tr><td>rbf (σ=1k)</td><td>78.56</td><td>81.13</td></tr><tr><td>grid (入min=0.0001,入max=360,S= 64)</td><td>79.44</td><td>81.28</td></tr><tr><td>theory (入min=0.0001, λmax=360,S=64)</td><td>79.35</td><td>81.59</td></tr></table>
|
| 207 |
+
|
| 208 |
+
We use two versions of our point space encoder $E n c ^ { ( x ) } ( \ v r )$ model (grid, theory) as the location encoder to capture the spatial prior information $P ( y \mid \mathbf { x } )$ . The evaluation results of our models as well as multiple baselines are shown in Table 4. We can see that both grid, theory outperform previous models as well as that of Mac Aodha et al. (2019) on two fine-grained image classification datasets with significant sizes: BirdSnap:, NABirds:. theory shows superiority over grid on NABirds: while fail to outperform grid on BirdSnap:. Note that we only pick baseline models which capture spatial-only prior and drop models which additionally consider time information. Both grid and theory use 1 hidden ReLU layers of 512 neurons for $\mathbf { N N } ( )$ and they have the same hyperparameters: $\lambda _ { m i n } { = } 0 . 0 0 0 1$ , $\lambda _ { m a x } { = } 3 6 0$ , $S = 6 4$ . Like Mac Aodha et al. (2019), the location embedding size $d ^ { ( x ) }$ is 1024 and we train the location encoder for 30 epochs. Our implementation is based on the original code7 of Mac Aodha et al. (2019) for both model training and evaluation phase.
|
| 209 |
+
|
| 210 |
+
# 6 CONCLUSION
|
| 211 |
+
|
| 212 |
+
We introduced an encoder-decoder framework as a general-purpose representation model for space inspired by biological grid cells’ multi-scale periodic representations. The model is an inductive learning model and can be trained in an unsupervised manner. We conduct two experiments on POI type prediction based on 1) POI locations and 2) nearby POIs. The evaluation results demonstrate the effectiveness of our model. Our analysis reveals that it is the ability to integrate representations of different scales that makes the grid cell models outperform other baselines on these two tasks. In the future, we hope to incorporate the presented framework to more complex GIS tasks such as social network analysis, and sea surface temperature prediction.
|
| 213 |
+
|
| 214 |
+
# ACKNOWLEDGMENTS
|
| 215 |
+
|
| 216 |
+
The presented work is partially funded by the NSF award 1936677 C-Accel Pilot - Track A1 (Open Knowledge Network): Spatially-Explicit Models, Methods, And Services For Open Knowledge Networks, Esri Inc., and Microsoft AI for Earth Grant: Deep Species Spatio-temporal Distribution Modeling for Biodiversity Hotspot Prediction. We thank Dr. Ruiqi Gao for discussions about grid cells, Dr. Wenyun Zuo for discussion about species potential distribution prediction and Dr. Yingjie Hu for his suggestions about the introduction section.
|
| 217 |
+
|
| 218 |
+
# REFERENCES
|
| 219 |
+
|
| 220 |
+
Alison Abbott and Ewen Callaway. Nobel prize for decoding brain’s sense of place. Nature News, 514(7521):153, 2014.
|
| 221 |
+
|
| 222 |
+
Benjamin Adams, Grant McKenzie, and Mark Gahegan. Frankenplace: interactive thematic mapping for ad hoc exploratory search. In Proceedings of the 24th international conference on world wide web, pp. 12–22. International World Wide Web Conferences Steering Committee, 2015.
|
| 223 |
+
|
| 224 |
+
Andrea Banino, Caswell Barry, Benigno Uria, Charles Blundell, Timothy Lillicrap, Piotr Mirowski, Alexander Pritzel, Martin J Chadwick, Thomas Degris, Joseph Modayil, et al. Vector-based navigation using grid-like representations in artificial agents. Nature, 557(7705):429, 2018.
|
| 225 |
+
|
| 226 |
+
G Baudat and F Anouar. Kernel-based methods and function approximation. volume 2, pp. 1244 – 1249 vol.2, 02 2001. ISBN 0-7803-7044-9. doi: 10.1109/IJCNN.2001.939539.
|
| 227 |
+
|
| 228 |
+
Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Birdsnap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2011–2018, 2014.
|
| 229 |
+
|
| 230 |
+
Herman J. Bierens. The nadaraya–watson kernel regression function estimator. Topics in Advanced Econometrics, 16:212–247, 1994.
|
| 231 |
+
|
| 232 |
+
Hugh T Blair, Adam C Welday, and Kechen Zhang. Scale-invariant memory representations emerge from moire interference between grid fields that produce theta oscillations: a computational model. Journal of Neuroscience, 27(12):3211–3229, 2007.
|
| 233 |
+
|
| 234 |
+
Grace Chu, Brian Potetz, Weijun Wang, Andrew Howard, Yang Song, Fernando Brucher, Thomas Leung, and Hartwig Adam. Geo-aware networks for fine grained recognition. arXiv preprint arXiv:1906.01737, 2019.
|
| 235 |
+
|
| 236 |
+
Christopher J Cueva and Xue-Xin Wei. Emergence of grid-like representations by training recurrent neural networks to perform spatial localization. arXiv preprint arXiv:1803.07770, 2018.
|
| 237 |
+
|
| 238 |
+
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.
|
| 239 |
+
|
| 240 |
+
A Stewart Fotheringham and David WS Wong. The modifiable areal unit problem in multivariate statistical analysis. Environment and planning A, 23(7):1025–1044, 1991.
|
| 241 |
+
|
| 242 |
+
Ruiqi Gao, Jianwen Xie, Song-Chun Zhu, and Ying Nian Wu. Learning grid cells as vector representation of self-position coupled with matrix representation of self-motion. In Proceedings of ICLR 2019, 2019.
|
| 243 |
+
|
| 244 |
+
Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii Shelestov. Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote Sensing Letters, 14(5):778–782, 2017.
|
| 245 |
+
|
| 246 |
+
Alex M Lechner, William T Langford, Simon D Jones, Sarah A Bekessy, and Ascelin Gordon. Investigating species–environment relationships at multiple scales: Differentiating between intrinsic scale and the modifiable areal unit problem. Ecological Complexity, 11:91–102, 2012.
|
| 247 |
+
|
| 248 |
+
Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926, 2017.
|
| 249 |
+
|
| 250 |
+
Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained image classification. arXiv preprint arXiv:1906.05272, 2019.
|
| 251 |
+
|
| 252 |
+
Gengchen Mai, Krzysztof Janowicz, Bo Yan, Rui Zhu, Ling Cai, and Ni Lao. Contextual graph attention for answering logical queries over incomplete knowledge graphs. In Proceedings of the 10th International Conference on Knowledge Capture, pp. 171–178, 2019a.
|
| 253 |
+
|
| 254 |
+
Gengchen Mai, Bo Yan, Krzysztof Janowicz, and Rui Zhu. Relaxing unanswerable geographic questions using a spatially explicit knowledge graph embedding model. In AGILE: The 22nd Annual International Conference on Geographic Information Science, pp. 21–39. Springer, 2019b.
|
| 255 |
+
|
| 256 |
+
V Maz’ya and G Schmidt. On approximate approximations using gaussian kernels. IMA Journal of Numerical Analysis, 16:13–29, 01 1996.
|
| 257 |
+
|
| 258 |
+
Grant McKenzie, Krzysztof Janowicz, Song Gao, Jiue-An Yang, and Yingjie Hu. Poi pulse: A multi-granular, semantic signature–based information observatory for the interactive visualization of big geosocial data. Cartographica: The International Journal for Geographic Information and Geovisualization, 50(2):71–85, 2015.
|
| 259 |
+
|
| 260 |
+
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pp. 3111–3119, 2013.
|
| 261 |
+
|
| 262 |
+
Justin Moat, Steven P Bachman, Richard Field, and Doreen S Boyd. Refining area of occupancy to address the modifiable areal unit problem in ecology and conservation. Conservation biology, 32 (6):1278–1289, 2018.
|
| 263 |
+
|
| 264 |
+
Stan Openshaw. The modifiable areal unit problem. Concepts and techniques in modern geography, 1984.
|
| 265 |
+
|
| 266 |
+
Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532–1543, 2014.
|
| 267 |
+
|
| 268 |
+
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.
|
| 269 |
+
|
| 270 |
+
Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 652–660, 2017.
|
| 271 |
+
|
| 272 |
+
Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2818–2826, 2016.
|
| 273 |
+
|
| 274 |
+
Kevin Tang, Manohar Paluri, Li Fei-Fei, Rob Fergus, and Lubomir Bourdev. Improving image classification with location context. In Proceedings of the IEEE international conference on computer vision, pp. 1008–1016, 2015.
|
| 275 |
+
|
| 276 |
+
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.
|
| 277 |
+
|
| 278 |
+
Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´ Bengio. Graph attention networks. In ICLR 2018, 2018.
|
| 279 |
+
|
| 280 |
+
Bo Yan, Krzysztof Janowicz, Gengchen Mai, and Song Gao. From itdl to place2vec: Reasoning about place type similarity and relatedness by learning embeddings from augmented spatial contexts. In Proceedings of the 25th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, pp. 35. ACM, 2017.
|
| 281 |
+
|
| 282 |
+
Bo Yan, Krzysztof Janowicz, Gengchen Mai, and Rui Zhu. xnet+ sc: Classifying places based on images by incorporating spatial contexts. In 10th International Conference on Geographic Information Science (GIScience 2018). Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, 2018.
|
| 283 |
+
|
| 284 |
+
Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan R Salakhutdinov, and Alexander J Smola. Deep sets. In Advances in neural information processing systems, pp. 3391–3401, 2017.
|
| 285 |
+
|
| 286 |
+
Wenyun Zuo, Ni Lao, Yuying Geng, and Keping Ma. Geosvm: an efficient and effective tool to predict species’ potential distributions. Journal of Plant Ecology, 1(2):143–145, 2008.
|
| 287 |
+
|
| 288 |
+
# A APPENDIX
|
| 289 |
+
|
| 290 |
+
# A.1 BASELINES
|
| 291 |
+
|
| 292 |
+
To help understand the mechanism of distributed space representation we compare multiple ways of encoding spatial information. Different models use different point space encoder $E n c ^ { ( x ) } ( \ v r )$ to encode either location $\mathbf { x } _ { i }$ (for location modeling loc) or the displacement between the center point and one context point $\Delta { \bf x } _ { i j } = { \bf x } _ { i } - { \bf x } _ { i j }$ (for spatial context modeling cont)8.
|
| 293 |
+
|
| 294 |
+
• random shuffles the order of the correct POI and $N$ negative samples randomly as the predicted ranking. This shows the lower bound of each metrics.
|
| 295 |
+
• direct directly encode location $\mathbf { x } _ { i }$ (or $\Delta { \bf x } _ { i j }$ for cont) into a location embedding $\mathbf { e } [ \mathbf { x } _ { i } ]$ (or $\mathbf { e } [ \Delta \mathbf { x } _ { i j } ] )$ using a feed-forward neural networks $\left( \mathrm { F F N s } \right) ^ { 9 }$ , denoted as $E n c _ { d i r e c t } ^ { ( x ) } ( { \bf x } )$ without decomposing coordinates into a multi-scale periodic representation. This is essentially the GPS encoding method used by Chu et al. (2019). Note that Chu et al. (2019) is not open sourced and we end up implementing the model architecture ourselves.
|
| 296 |
+
• tile divides the study area $A _ { l o c }$ (for loc) or the range of spatial context defined by $\lambda _ { m a x }$ , $A _ { c o n t }$ , (for cont) into grids with equal grid sizes $c$ . Each grid has an embedding to be used as the encoding for every location $\mathbf { x } _ { i }$ or displacement $\Delta { \bf x } _ { i j }$ fall into this grid. This is a common practice by many previous work when dealing with coordinate data (Berg et al., 2014; Adams et al., 2015; Tang et al., 2015). wrap is a location encoder model recently introduced by Mac Aodha et al. (2019). It first normalizes $\mathbf { x }$ (or $\Delta \mathbf { x } )$ ) into the range $[ - 1 , 1 ]$ and uses a coordinate wrap mechanism $[ \sin ( \pi \mathbf { x } ^ { [ l ] } ) ; \cos ( \pi \mathbf { x } ^ { [ l ] } ) ]$ to convert each dimension of $\mathbf { x }$ into 2 numbers. This is then passed through an initial fully connected layer, followed by a series of $h$ residual blocks, each consisting of two fully connected layers $\scriptstyle { \dot { } } _ { o }$ hidden neurons) with a dropout layer in between. We adopt the official code of Mac Aodha et al. $( 2 0 1 9 ) ^ { 1 0 }$ for this implementation.
|
| 297 |
+
• $r b f$ randomly samples $M$ points from the training dataset as RBF anchor points $\{ \mathbf { x } _ { m } ^ { a n c h o r } , m \stackrel { \cdot } { = } 1 . . . \bar { M } \}$ (or samples $M \Delta \mathbf { x } _ { m } ^ { a n c h o r }$ from $A _ { c o n t }$ for cont) 11, and use gaussian kernels $\exp \big ( - \frac { \parallel \mathbf { x } _ { i } - \mathbf { x } _ { m } ^ { a n c h o r } \parallel ^ { 2 } } { 2 \sigma ^ { 2 } } \big )$ (or $\exp \big ( - \frac { \parallel \Delta \mathbf { x } _ { i j } - \Delta \mathbf { x } _ { m } ^ { a n c h o r } \parallel ^ { 2 } } { 2 \sigma ^ { 2 } } \big )$ for cont) on each anchor points, where $\sigma$ is the kernel size. Each point $p _ { i }$ has a $M$ -dimension RBF feature vector which is fed into a FNN to obtain the spatial embedding. This is a strong baseline for representing floating number features in machine learning models.
|
| 298 |
+
• grid as described in Section 4.1 inspired by the position encoding in Transformer (Vaswani et al., 2017).
|
| 299 |
+
• hexa Same as grid but use $s i n ( \theta ) , s i n ( \theta + 2 \pi / 3 )$ , and $s i n ( \theta + 4 \pi / 3 )$ in $P E _ { s , l } ^ { ( g ) } ( \mathbf { x } )$ .
|
| 300 |
+
• theory as described in Section 4.1, uses the theoretical models (Gao et al., 2019) as the first layer of $E n c _ { t h e o r y } ^ { ( x ) } ( \mathbf { x } )$ or $E n c _ { t h e o r y } ^ { ( x ) } ( \Delta \mathbf { x } _ { i j } )$ .
|
| 301 |
+
• theorydiag further constrains $\mathbf { N N } ( )$ as a block diagonal matrix, with each scale as a block.
|
| 302 |
+
|
| 303 |
+
We also have the following baselines which are specific to the spatial context modeling task.
|
| 304 |
+
|
| 305 |
+
• none the decoder $D e c _ { c } ( )$ does not consider the spatial relationship between the center point and context points but only the co-locate patterns such as Place2Vec (Yan et al., 2017). That means we drop the $\mathbf { e } [ \Delta \mathbf { x } _ { i j } ]$ from the attention mechanism in Equ. 7 and 8. polar first converts the displacement $\Delta \mathbf { x } _ { i j }$ into polar coordinates $( r , \theta )$ centered at the center point where $r = l o g ( \parallel \Delta \mathbf { x } _ { i j } \parallel + 1 ) ^ { - }$ . Then it uses $[ r , \theta ]$ as the input for a FFN to obtain the spatial relationship embedding in Equ. 7. We find out that it has a significant performance improvement over the variation with $r = \parallel \Delta \mathbf { x } _ { i j } \parallel$ .
|
| 306 |
+
|
| 307 |
+
• polar_tile is a modified version of tile but the grids are extracted from polar coordinates $( r , \theta )$ centered at the center point where $r = l o g ( \parallel \Delta \mathbf { x } _ { i j } \parallel + 1 )$ . Instead of using grid size $c$ , we use the number of grids along $\theta$ (or $r$ ) axis, $F$ , as the only hyperparameter. Similarly, We find that $r = l o g ( \parallel \Delta \mathbf { x } _ { i j } \parallel + \mathrm { 1 } \bar { ) }$ outperform $r = \parallel \Delta \mathbf { x } _ { i j } \parallel$ significantly. • scaled_rbf is a modified version of $r b f$ for cont whose kernel size is proportional to the distance between the current anchor point and the origin, $\parallel \Delta \mathbf { x } _ { m } ^ { a n c \hat { h } o r ^ { * } } \parallel$ . That is ´ 2σ2scbasic kernel size and $\exp \big ( - \frac { \parallel \Delta \mathbf { x } _ { i j } - \Delta \mathbf { x } _ { m } ^ { a n c h o r } \parallel ^ { 2 } } { 2 \sigma _ { s c a l e d } ^ { 2 } } \big )$ $\beta$ d scaled “ \` m is kernel rescale factor, a constant. We developed this me $\sigma$ anism to help RFB to deal with relations at different scale, and we observe that it produces significantly better result than vanilla RBFs.
|
| 308 |
+
|
| 309 |
+
# A.2 HYPER-PARAMETER SELECTION
|
| 310 |
+
|
| 311 |
+
We perform grid search for all methods based on their performance on the validation sets.
|
| 312 |
+
|
| 313 |
+
Location Modeling The hyper-parameters of theory models are based on grid search with $d ^ { ( v ) } \ = \ ( 3 2 , 6 4 , 1 2 \bar { 8 } , 2 5 6 )$ , $d ^ { ( x ) } \ = \ ( 3 2 , 6 4 , 1 2 8 , 2 5 6 )$ , $S \ = \ ( 4 , 8 , 1 6 , 3 2 , 6 4 , 1 2 8 )$ , and $\lambda _ { m i n } ~ =$ $( 1 , 5 , 1 0 , 5 0 , 1 0 0 , 2 0 0 , 5 0 0 , 1 k )$ while $\lambda _ { m a x } = 4 0 k$ is decided based on the total size of the study area. We find out the best performances of different grid cell based models are obtained when $d ^ { ( v ) } = 6 4$ , $d ^ { ( x ) } = 6 4$ , $S \ : = \ : 6 4$ , and $\lambda _ { m i n } = 5 0$ . In terms of tile, the hyper-parameters are selected from $c = ( 1 0 , 5 0 , 1 0 0 , 2 0 0 , 5 0 0 , 1 0 0 0 )$ while $c = 5 0 0$ gives us the best performance.As for $r b f$ , we do grid search on the hyper-parameters: $M \ = \ ( 1 0 , 5 0 , 1 0 0 , 2 0 0 , 4 0 0 , 8 0 0 )$ and $\sigma = ( \mathrm { 1 0 ^ { 2 } } , \mathrm { 1 0 ^ { 3 } } , \mathrm { 1 0 ^ { 4 } } , \mathrm { 1 0 ^ { 5 } } , \mathrm { 1 0 ^ { 6 } } , \mathrm { 1 0 ^ { 7 } } )$ . The best performance of $r b f$ is obtain when $M = 2 0 0$ and $\sigma = 1 0 ^ { 3 }$ . As for wrap, grid search is performed on: $h = ( 1 , 2 , 3 , 4 )$ and $o = ( 6 4 , 1 2 8 , 2 5 6 , 5 1 2 )$ while $h = 3$ and $o = 5 1 2$ gives us the best result. All models use FFNs in their $E n c ^ { ( x ) } ( \ v r )$ except wrap. The number of layers $f$ and the number of hidden state neurons $u$ of the FFN are selected from $f = ( 1 , 2 , 3 )$ and $u = ( 1 2 8 , 2 5 6 , 5 1 2 )$ . We find out $f = 1$ and $u = 5 1 2$ give the best performance for direct, tile, $r b f$ , and theory. So we use them for every model for a fair comparison.
|
| 314 |
+
|
| 315 |
+
Spatial Context Modeling Grid search is used for hyperparameter tuning and the best performance of different grid cell models is obtain when $d ^ { ( v ) } = 6 \dot { 4 }$ , $d ^ { ( \bar { x } ) } = 6 4$ , $S = 6 4$ , and $\lambda _ { m i n } = 1 0$ . We set $\lambda _ { m a x } = 1 0 k$ based on the maximum displacement between context points and center points to make the location encoding unique. As for multiple baseline models, grid search is used again to obtain the best model. The best model hyperparameters are shown in () besides the model names in Table 3. Note that both $r b f$ and scaled_rbf achieve the best performance with $M = 1 0 0$ .
|
| 316 |
+
|
| 317 |
+

|
| 318 |
+
A.3 FIRING PATTERN FOR THE NEURONS
|
| 319 |
+
Figure 4: The firing pattern for the first 8 neurons (out of 64) given different encoders in location modeling.
|
| 320 |
+
|
| 321 |
+

|
| 322 |
+
|
| 323 |
+

|
| 324 |
+
Figure 5: Embedding clustering in the original space of (a)(b)(c)(d) theory with different $\lambda _ { m i n }$ , but the same $\lambda _ { m a x } = 1 0 k$ and $S = 6 4$ . (e)(f)(g)(h) are the embedding clustering results of the same models in the polar-distance space. All models use 1 hidden ReLU layers of 512 neurons.
|
| 325 |
+
Figure 6: Embedding clustering of RBF models with different kernel rescalar factor $\beta$ (a)(b)(c)(d) in the original space; (e)(f)(g)(h) in the polar-distance space. Here $\beta { = } 0 . 0$ indicates the original RBF model. All models use $\sigma { = } 1 0 \mathrm { m }$ as the basic kernel size and 1 hidden ReLU layers of 512 neurons.
|
parse/train/rJljdh4KDH/rJljdh4KDH_content_list.json
ADDED
|
@@ -0,0 +1,1659 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "MULTI-SCALE REPRESENTATION LEARNING FOR SPATIAL FEATURE DISTRIBUTIONS USING GRID CELLS ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
174,
|
| 8 |
+
99,
|
| 9 |
+
826,
|
| 10 |
+
146
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Gengchen Mai1, Krzysztof Janowicz1, Bo Yan2, Rui $\\mathbf { Z } \\mathbf { h } \\mathbf { u } ^ { 1 }$ , Ling Cai1 & Ni Lao3 \n1STKO Lab, University of California, Santa Barbara, CA, USA, 93106 \n{gengchen_mai,janowicz,ruizhu,lingcai}@ucsb.edu \n2LinkedIn Corporation, Mountain View, CA, USA, 94043 \nboyan1@linkedin.com \n3SayMosaic Inc., Palo Alto, CA, USA, 94303 \nni.lao@mosaix.ai ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
184,
|
| 19 |
+
167,
|
| 20 |
+
740,
|
| 21 |
+
268
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "ABSTRACT ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
454,
|
| 31 |
+
305,
|
| 32 |
+
544,
|
| 33 |
+
320
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "Unsupervised text encoding models have recently fueled substantial progress in Natural Language Processing (NLP). The key idea is to use neural networks to convert words in texts to vector space representations (embeddings) based on word positions in a sentence and their contexts, which are suitable for end-to-end training of downstream tasks. We see a strikingly similar situation in spatial analysis, which focuses on incorporating both absolute positions and spatial contexts of geographic objects such as Points of Interest (POIs) into models. A general-purpose representation model for space is valuable for a multitude of tasks. However, no such general model exists to date beyond simply applying discretization or feedforward nets to coordinates, and little effort has been put into jointly modeling distributions with vastly different characteristics, which commonly emerges from GIS data. Meanwhile, Nobel Prize-winning Neuroscience research shows that grid cells in mammals provide a multi-scale periodic representation that functions as a metric for location encoding and is critical for recognizing places and for path-integration. Therefore, we propose a representation learning model called Space2Vec to encode the absolute positions and spatial relationships of places. We conduct experiments on two real-world geographic data for two different tasks: 1) predicting types of POIs given their positions and context, 2) image classification leveraging their geo-locations. Results show that because of its multiscale representations, Space2Vec outperforms well-established ML approaches such as RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for location modeling and image classification tasks. Detailed analysis shows that all baselines can at most well handle distribution at one scale but show poor performances in other scales. In contrast, Space2Vec ’s multi-scale representation can handle distributions at different scales. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
232,
|
| 42 |
+
338,
|
| 43 |
+
766,
|
| 44 |
+
683
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 INTRODUCTION ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
176,
|
| 54 |
+
708,
|
| 55 |
+
336,
|
| 56 |
+
724
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "Unsupervised text encoding models such as Word2Vec (Mikolov et al., 2013), Glove (Pennington et al., 2014), ELMo (Peters et al., 2018), and BERT (Devlin et al., 2018) have been effectively utilized in many Natural Language Processing (NLP) tasks. At their core they train models which encode words into vector space representations based on their positions in the text and their context. A similar situation can be encountered in the field of Geographic Information Science (GIScience). For example, spatial interpolation aims at predicting an attribute value, e.g., elevation, at an unsampled location based on the known attribute values of nearby samples. Geographic information has become an important component to many tasks such as fine-grained image classification (Mac Aodha et al., 2019), point cloud classification and semantic segmentation (Qi et al., 2017), reasoning about Point of Interest (POI) type similarity (Yan et al., 2017), land cover classification (Kussul et al., 2017), and geographic question answering (Mai et al., 2019b). Developing a general model for vector space representation of any point in space would pave the way for many future applications. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
733,
|
| 66 |
+
825,
|
| 67 |
+
900
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "image",
|
| 73 |
+
"img_path": "images/d51a01c0d293200f92299629bbfb4a0ec7530397835bdb9650ef5b4daa40806d.jpg",
|
| 74 |
+
"image_caption": [
|
| 75 |
+
"Figure 1: The challenge of joint modeling distributions with very different characteristics. (a)(b) The POI locations (red dots) in Las Vegas and Space2Vec predicted conditional likelihood of Women’s Clothing (with a clustered distribution) and Education (with an even distribution). The dark area in (b) indicates that the downtown area has more POIs of other types than education. (c) Ripley’s K curves of POI types for which Space2Vec has the largest and smallest improvement over wrap (Mac Aodha et al., 2019). Each curve represents the number of POIs of a certain type inside certain radios centered at every POI of that type; (d) Ripley’s K curves renormalized by POI densities and shown in log-scale. To efficiently achieve multi-scale representation Space2Vec concatenates the grid cell encoding of 64 scales (with wave lengths ranging from 50 meters to $4 0 k$ meters) as the first layer of a deep model, and trains with POI data in an unsupervised fashion. "
|
| 76 |
+
],
|
| 77 |
+
"image_footnote": [],
|
| 78 |
+
"bbox": [
|
| 79 |
+
181,
|
| 80 |
+
104,
|
| 81 |
+
816,
|
| 82 |
+
265
|
| 83 |
+
],
|
| 84 |
+
"page_idx": 1
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"type": "text",
|
| 88 |
+
"text": "However, existing models often utilize specific methods to deal with geographic information and often disregards geographic coordinates. For example, Place2Vec (Yan et al., 2017) converts the coordinates of POIs into spatially collocated POI pairs within certain distance bins, and does not preserve information about the (cardinal) direction between points. Li et al. (2017) propose DCRNN for traffic forecasting in which the traffic sensor network is converted to a distance weighted graph which necessarily forfeits information about the spatial layout of sensors. There is, however, no general representation model beyond simply applying discretization (Berg et al., 2014; Tang et al., 2015) or feed-forward nets (Chu et al., 2019; Mac Aodha et al., 2019) to coordinates. ",
|
| 89 |
+
"bbox": [
|
| 90 |
+
174,
|
| 91 |
+
398,
|
| 92 |
+
825,
|
| 93 |
+
510
|
| 94 |
+
],
|
| 95 |
+
"page_idx": 1
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"type": "text",
|
| 99 |
+
"text": "A key challenge in developing a general-purpose representation model for space is how to deal with mixtures of distributions with very different characteristics (see an example in Figure 1), which often emerges in spatial datasets (McKenzie et al., 2015). For example, there are POI types with clustered distributions such as women’s clothing, while there are other POI types with regular distributions such as education. These feature distributions co-exist in the same space, and yet we want a single representation to accommodate all of them in a task such as location-aware image classification (Mac Aodha et al., 2019). Ripley’s K is a spatial analysis method used to describe point patterns over a given area of interest. Figure 1c shows the K plot of several POI types in Las Vegas. One can see that as the radius grows the numbers of POIs increase at different rates for different POI types. In order to see the relative change of density at different scales, we renormalize the curves by each POI type’s density and show it in log scale in Figure 1d. One can see two distinct POI type groups with different distribution patterns with clustered and even distributions. If we want to model the distribution of these POIs by discretizing the study area into tiles, we have to use small grid sizes for women’s clothing while using larger grid sizes for educations because smaller grid sizes lead to over- parameterization of the model and overfitting. In order to jointly describe these distributions and their patterns, we need an encoding method which supports multi-scale representations. ",
|
| 100 |
+
"bbox": [
|
| 101 |
+
174,
|
| 102 |
+
517,
|
| 103 |
+
825,
|
| 104 |
+
739
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "Nobel Prize winning Neuroscience research (Abbott & Callaway, 2014) has demonstrated that grid cells in mammals provide a multi-scale periodic representation that functions as a metric for location encoding, which is critical for integrating self-motion. Moreover, Blair et al. (2007) show that the multi-scale periodic representation of grid cells can be simulated by summing three cosine grating functions oriented $6 0 ^ { \\circ }$ apart, which may be regarded as a simple Fourier model of the hexagonal lattice. This research inspired us to encode locations with multi-scale periodic representations. Our assumption is that decomposed geographic coordinates helps machine learning models, such as deep neural nets, and multi-scale representations deal with the inefficiency of intrinsically single-scale methods such as RFB kernels or discretization (tile embeddings). To validate this intuition, we propose an encoder-decoder framework to encode the distribution of point-features2 in space and train such a model in an unsupervised manner. This idea of using sinusoid functions with different frequencies to encode positions is similar to the position encoding proposed in the Transformer model (Vaswani et al., 2017). However, the position encoding model of Transformer deals with a discrete 1D space – the positions of words in a sentence – while our model works on higher dimensional continuous spaces such as the surface of earth. ",
|
| 111 |
+
"bbox": [
|
| 112 |
+
174,
|
| 113 |
+
746,
|
| 114 |
+
825,
|
| 115 |
+
885
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
103,
|
| 125 |
+
825,
|
| 126 |
+
174
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 2
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "In summary, the contributions of our work are as follows: ",
|
| 133 |
+
"text_level": 1,
|
| 134 |
+
"bbox": [
|
| 135 |
+
174,
|
| 136 |
+
180,
|
| 137 |
+
576,
|
| 138 |
+
195
|
| 139 |
+
],
|
| 140 |
+
"page_idx": 2
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"type": "text",
|
| 144 |
+
"text": "1. We propose an encoder-decoder encoding framework called Space2Vec using sinusoid functions with different frequencies to model absolute positions and spatial contexts. We also propose a multi-head attention mechanism based on context points. To the best of our knowledge, this is the first attention model that explicitly considers the spatial relationships between the query point and context points. \n2. We conduct experiments on two real world geographic data for two different tasks: 1) predicting types of POIs given their positions and context, 2) image classification leveraging their geo-locations. Space2Vec outperforms well-established encoding methods such as RBF kernels, multi-layer feed-forward nets, and tile embedding approaches for location modeling and image classification. \n3. To understand the advantages of Space2Vec we visualize the firing patterns (response maps) of location models’ encoding layer neurons and show how they handle spatial structures at different scales by integrating multi-scale representations. Furthermore the firing patterns for the spatial context models neurons give insight into how the grid-like cells capture the decreasing distance effect with multi-scale representations. ",
|
| 145 |
+
"bbox": [
|
| 146 |
+
212,
|
| 147 |
+
207,
|
| 148 |
+
825,
|
| 149 |
+
425
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 2
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "2 PROBLEM FORMULATION ",
|
| 156 |
+
"text_level": 1,
|
| 157 |
+
"bbox": [
|
| 158 |
+
176,
|
| 159 |
+
444,
|
| 160 |
+
416,
|
| 161 |
+
459
|
| 162 |
+
],
|
| 163 |
+
"page_idx": 2
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"type": "text",
|
| 167 |
+
"text": "Distributed representation of point-features in space can be formulated as follows. Given a set of points $\\mathcal { P } _ { - } = \\left\\{ p _ { i } \\right\\}$ , i.e., Points of Interests (POIs), in $L { \\mathrm { - } } \\mathbf { D }$ space $( L = 2 , 3$ ) define a function $f _ { \\mathcal { P } , \\theta } ( \\mathbf { x } ) : \\mathbb { R } ^ { L } \\mathbf { \\bar { \\mathbb { R } } } ^ { \\dot { d } }$ $( L \\ll d )$ , which is parameterized by $\\theta$ and maps any coordinate $\\mathbf { x }$ in space to a vector representation of $d$ dimension. Each point (e.g., a restaurant) $p _ { i } = ( \\mathbf { x } _ { i } , \\mathbf { v } _ { i } )$ is associated with a location $\\mathbf { x } _ { i }$ and attributes $\\mathbf { v } _ { i }$ (i.e., POI features such as type, name, capacity, etc.). The function $f _ { \\mathcal { P } , \\theta } ( \\mathbf { x } )$ encodes the probability distribution of point features over space and can give a representation of any point in the space. Attributes (e.g. place types such as Museum) and coordinate of point can be seen as analogies to words and word positions in commonly used word embedding models. ",
|
| 168 |
+
"bbox": [
|
| 169 |
+
173,
|
| 170 |
+
468,
|
| 171 |
+
825,
|
| 172 |
+
579
|
| 173 |
+
],
|
| 174 |
+
"page_idx": 2
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"type": "text",
|
| 178 |
+
"text": "3 RELATED WORK ",
|
| 179 |
+
"text_level": 1,
|
| 180 |
+
"bbox": [
|
| 181 |
+
176,
|
| 182 |
+
593,
|
| 183 |
+
344,
|
| 184 |
+
608
|
| 185 |
+
],
|
| 186 |
+
"page_idx": 2
|
| 187 |
+
},
|
| 188 |
+
{
|
| 189 |
+
"type": "text",
|
| 190 |
+
"text": "There has been theoretical research on neural network based path integration/spatial localization models and their relationships with grid cells. Both Cueva & Wei (2018) and Banino et al. (2018) showed that grid-like spatial response patterns emerge in trained networks for navigation tasks which demonstrate that grid cells are critical for vector-based navigation. Moreover, Gao et al. (2019) propose a representational model for grid cells in navigation tasks which has good quality such as magnified local isometry. All these research is focusing on understanding the relationship between the grid-like spatial response patterns and navigation tasks from a theoretical perspective. In contrast, our goal focuses on utilizing these theoretical results on real world data in geoinformatics. ",
|
| 191 |
+
"bbox": [
|
| 192 |
+
174,
|
| 193 |
+
617,
|
| 194 |
+
825,
|
| 195 |
+
728
|
| 196 |
+
],
|
| 197 |
+
"page_idx": 2
|
| 198 |
+
},
|
| 199 |
+
{
|
| 200 |
+
"type": "text",
|
| 201 |
+
"text": "Radial Basis Function (RBF) kernel is a well-established approach to generating learning friendly representation from points in space for machine learning algorithms such as SVM classification (Baudat & Anouar, 2001) and regression (Bierens, 1994). However, the representation is example based – i.e., the resultant model uses the positions of training examples as the centers of Gaussian kernel functions (Maz’ya & Schmidt, 1996). In comparison, the grid cell based location encoding relies on sine and cosine functions, and the resultant model is inductive and does not store training examples. ",
|
| 202 |
+
"bbox": [
|
| 203 |
+
173,
|
| 204 |
+
734,
|
| 205 |
+
825,
|
| 206 |
+
819
|
| 207 |
+
],
|
| 208 |
+
"page_idx": 2
|
| 209 |
+
},
|
| 210 |
+
{
|
| 211 |
+
"type": "text",
|
| 212 |
+
"text": "Recently the computer vision community shows increasing interests in incorporating geographic information (e.g. coordinate encoding) into neural network architectures for multiple tasks such as image classification (Tang et al., 2015) and fine grained recognition (Berg et al., 2014; Chu et al., 2019; Mac Aodha et al., 2019). Both Berg et al. (2014) and Tang et al. (2015) proposed to discretize the study area into regular grids. To model the geographical prior distribution of the image categories, the grid id is used for GPS encoding instead of the raw coordinates. However, choosing the correct discretization is challenging (Openshaw, 1984; Fotheringham & Wong, 1991), and incorrect choices can significantly affect the final performance (Moat et al., 2018; Lechner et al., 2012). In addition, discretization does not scale well in terms of memory use. To overcome these difficulties, both Chu et al. (2019) and Mac Aodha et al. (2019) advocated the idea of inductive location encoders which directly encode coordinates into a location embedding. However, both of them directly feed the coordinates into a feed-forward neural network (Chu et al., 2019) or residual blocks (Mac Aodha et al., 2019) without any feature decomposition strategy. Our experiments show that this direct encoding approach is insufficient to capture the spatial feature distribution and Space2Vec significantly outperforms them by integrating spatial representations of different scales. ",
|
| 213 |
+
"bbox": [
|
| 214 |
+
174,
|
| 215 |
+
825,
|
| 216 |
+
825,
|
| 217 |
+
924
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "",
|
| 224 |
+
"bbox": [
|
| 225 |
+
174,
|
| 226 |
+
103,
|
| 227 |
+
825,
|
| 228 |
+
215
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 3
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "4 METHOD ",
|
| 235 |
+
"text_level": 1,
|
| 236 |
+
"bbox": [
|
| 237 |
+
174,
|
| 238 |
+
227,
|
| 239 |
+
282,
|
| 240 |
+
243
|
| 241 |
+
],
|
| 242 |
+
"page_idx": 3
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"type": "text",
|
| 246 |
+
"text": "We solve distributed representation of point-features in space (defined in Section 2) with an encoderdecoder architecture: ",
|
| 247 |
+
"bbox": [
|
| 248 |
+
176,
|
| 249 |
+
251,
|
| 250 |
+
825,
|
| 251 |
+
280
|
| 252 |
+
],
|
| 253 |
+
"page_idx": 3
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"type": "text",
|
| 257 |
+
"text": "1. Given a point $p _ { i } = ( \\mathbf { x } _ { i } , \\mathbf { v } _ { i } )$ a point space encoder $E n c ^ { ( x ) } ( \\ v r )$ encodes location $\\mathbf { x } _ { i }$ into a location embedding $\\mathbf { e } [ \\mathbf { x } _ { i } ] \\in \\mathbb { R } ^ { d ^ { ( x ) } }$ and a point feature encoder $E n c ^ { ( v ) } ( )$ encodes its feature into a feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ] \\in \\mathbb { R } ^ { d ^ { ( v ) } }$ . $\\mathbf { e } = [ \\mathbf { e } [ \\mathbf { x } _ { i } ] ; \\mathbf { e } [ \\mathbf { v } _ { i } ] ] \\in \\mathbb { R } ^ { d }$ is the full representation of point $p _ { i } \\in \\mathcal { P }$ , where $d = d ^ { ( x ) } + d ^ { ( v ) }$ . $[ ; ]$ represents vector concatenation. In contrast, geographic entities not in $\\mathcal { P }$ within the studied space can be represented by their location embedding $\\mathbf { e } [ \\mathbf { x } _ { j } ]$ since its $\\mathbf { v } _ { i }$ is unknown. 2. We developed two types of decoders which can be used independently or jointly. A location decoder $D e c _ { s } ( )$ reconstructs point feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ given location embedding $\\mathbf { e } [ \\mathbf { x } _ { i } ]$ , and a spatial context decoder $D e c _ { c } ( )$ reconstructs the feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ of point $p _ { i }$ based on the space and feature embeddings $\\left\\{ \\mathbf { e } _ { i 1 } , . . . , \\mathbf { e } _ { i j } , . . . , \\mathbf { e } _ { i n } \\right\\}$ of nearest neighboring points $\\{ p _ { i 1 } , . . . , p _ { i j } , . . . , p _ { i n } \\}$ , where $n$ is a hyper-parameter. ",
|
| 258 |
+
"bbox": [
|
| 259 |
+
212,
|
| 260 |
+
287,
|
| 261 |
+
826,
|
| 262 |
+
458
|
| 263 |
+
],
|
| 264 |
+
"page_idx": 3
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"type": "text",
|
| 268 |
+
"text": "4.1 ENCODER ",
|
| 269 |
+
"text_level": 1,
|
| 270 |
+
"bbox": [
|
| 271 |
+
174,
|
| 272 |
+
469,
|
| 273 |
+
284,
|
| 274 |
+
483
|
| 275 |
+
],
|
| 276 |
+
"page_idx": 3
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"type": "text",
|
| 280 |
+
"text": "Point Feature Encoder Each point $p _ { i } = ( \\mathbf { x } _ { i } , \\mathbf { v } _ { i } )$ in a point set $\\mathcal { P }$ is often associated with features such as the air pollution station data associate with some air quality measures, a set of POIs with POI types and names, a set of points from survey and mapping with elevation values, a set of points from geological survey with mineral content measure, and so on. The point feature encoder $E n c ^ { ( v ) } ( )$ encodes such features $\\mathbf { v } _ { i }$ into a feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ] \\in \\mathbb { R } ^ { d ^ { ( v ) } }$ . The implementation of $E n c ^ { ( v ) } ( )$ depends on the nature of these features. For example, if each point represents a POI with multiple POI types (as in this study), the feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ can simply be the mean of each POI types’ embeddings $\\mathbf { e } [ \\mathbf { v } _ { i } ] = \\frac { 1 } { H } \\sum _ { h = 1 } ^ { H } \\mathbf { t } _ { h } ^ { ( \\gamma ) }$ , where $\\mathbf { t } _ { h } ^ { \\left( \\gamma \\right) }$ indicates the $h$ th POI type embedding of a POI $p _ { i }$ with $H$ POI types. We apply $L _ { 2 }$ normalization to the POI type embedding matrix. ",
|
| 281 |
+
"bbox": [
|
| 282 |
+
173,
|
| 283 |
+
488,
|
| 284 |
+
825,
|
| 285 |
+
631
|
| 286 |
+
],
|
| 287 |
+
"page_idx": 3
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"type": "text",
|
| 291 |
+
"text": "Point Space Encoder A part of the novelty of this paper is from the point space encoder $E n c ^ { ( x ) } ( \\ v r )$ . We first introduce Theorem 1 which provide an analytical solution $\\phi ( \\mathbf { x } )$ as the base of encoding any location $\\mathbf { x } \\in \\mathbb { R } ^ { 2 }$ in 2D space to a distributed representation: ",
|
| 292 |
+
"bbox": [
|
| 293 |
+
174,
|
| 294 |
+
645,
|
| 295 |
+
825,
|
| 296 |
+
688
|
| 297 |
+
],
|
| 298 |
+
"page_idx": 3
|
| 299 |
+
},
|
| 300 |
+
{
|
| 301 |
+
"type": "text",
|
| 302 |
+
"text": "Theorem 1. Let $\\Psi ( \\mathbf { x } ) = ( e ^ { i \\langle \\mathbf { a } _ { j } , \\mathbf { x } \\rangle } , j = 1 , 2 , 3 ) ^ { T } \\in \\mathbb { C } ^ { 3 }$ where $e ^ { i \\theta } = \\cos \\theta + i \\sin \\theta$ is the Euler notation of complex values; $\\langle \\mathbf { a } _ { j } , \\mathbf { x } \\rangle$ is the inner product of ${ \\bf a } _ { j }$ and $\\mathbf { x } .$ $\\mathfrak { c } . \\ \\mathbf { a } _ { 1 } , \\mathbf { a } _ { 2 } , \\mathbf { a } _ { 3 } \\in \\mathbb { R } ^ { 2 }$ are $2 D$ vectors such that the angle between ${ \\bf a } _ { k }$ and ${ \\bf a } _ { l }$ is $2 \\pi / 3 , \\forall j$ , $\\| \\mathbf { a } _ { j } \\| = 2 { \\sqrt { \\alpha } }$ . Let $\\mathbf { C } \\in \\mathbb { C } ^ { 3 \\times 3 }$ be a random complex matrix such as $\\mathbf { C } ^ { * } \\mathbf { C } = \\mathbf { I } .$ . Then $\\phi ( \\mathbf { x } ) = \\mathbf { C } \\Psi ( \\mathbf { x } )$ , $M ( \\Delta \\mathbf { x } ) = \\mathbf { C } d i a g ( \\pmb { \\Psi } ( \\Delta \\mathbf { x } ) ) \\mathbf { C } ^ { * }$ satisfies ",
|
| 303 |
+
"bbox": [
|
| 304 |
+
173,
|
| 305 |
+
690,
|
| 306 |
+
825,
|
| 307 |
+
751
|
| 308 |
+
],
|
| 309 |
+
"page_idx": 3
|
| 310 |
+
},
|
| 311 |
+
{
|
| 312 |
+
"type": "equation",
|
| 313 |
+
"img_path": "images/919b88dbed41cf09915ef7383dacf5779d85e5e73cc2b3f8cb1459da7d988c3d.jpg",
|
| 314 |
+
"text": "$$\n\\phi ( \\mathbf { x } + \\Delta \\mathbf { x } ) = M ( \\Delta \\mathbf { x } ) \\phi ( \\mathbf { x } )\n$$",
|
| 315 |
+
"text_format": "latex",
|
| 316 |
+
"bbox": [
|
| 317 |
+
405,
|
| 318 |
+
753,
|
| 319 |
+
593,
|
| 320 |
+
770
|
| 321 |
+
],
|
| 322 |
+
"page_idx": 3
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"type": "text",
|
| 326 |
+
"text": "and ",
|
| 327 |
+
"bbox": [
|
| 328 |
+
173,
|
| 329 |
+
775,
|
| 330 |
+
202,
|
| 331 |
+
787
|
| 332 |
+
],
|
| 333 |
+
"page_idx": 3
|
| 334 |
+
},
|
| 335 |
+
{
|
| 336 |
+
"type": "equation",
|
| 337 |
+
"img_path": "images/1edb46244251d5f1801b3ecaba3f5cb6a3e869194165762498b7a334776fe611.jpg",
|
| 338 |
+
"text": "$$\n\\langle \\phi ( \\mathbf { x } + \\Delta \\mathbf { x } ) , \\phi ( \\mathbf { x } ) \\rangle = d ( 1 - \\alpha \\| \\Delta \\mathbf { x } \\| ^ { 2 } )\n$$",
|
| 339 |
+
"text_format": "latex",
|
| 340 |
+
"bbox": [
|
| 341 |
+
369,
|
| 342 |
+
786,
|
| 343 |
+
629,
|
| 344 |
+
803
|
| 345 |
+
],
|
| 346 |
+
"page_idx": 3
|
| 347 |
+
},
|
| 348 |
+
{
|
| 349 |
+
"type": "text",
|
| 350 |
+
"text": "where $d = 3$ is the dimension of $\\phi ( \\mathbf { x } )$ and $\\Delta \\mathbf { x }$ is a small displacement from x. ",
|
| 351 |
+
"bbox": [
|
| 352 |
+
174,
|
| 353 |
+
804,
|
| 354 |
+
684,
|
| 355 |
+
819
|
| 356 |
+
],
|
| 357 |
+
"page_idx": 3
|
| 358 |
+
},
|
| 359 |
+
{
|
| 360 |
+
"type": "text",
|
| 361 |
+
"text": "The proof of Theorem 1 can be seen in Gao et al. (2019). $\\phi ( \\mathbf { x } ) = \\mathbf { C } \\pmb { \\Psi } ( \\mathbf { x } ) \\in \\mathbb { C } ^ { 3 }$ amounts to a 6-dimension real value vector and each dimension shows a hexagon firing pattern which models the grid cell behavior. Because of the periodicity of $s i n ( )$ and $c o s ( )$ , this single scale representation $\\phi ( \\mathbf { x } )$ does not form a global codebook of 2D positions, i.e. there can be $\\mathbf x \\neq \\mathbf y$ , but $\\phi ( \\mathbf { x } ) = \\phi ( \\mathbf { y } )$ . ",
|
| 362 |
+
"bbox": [
|
| 363 |
+
174,
|
| 364 |
+
828,
|
| 365 |
+
823,
|
| 366 |
+
886
|
| 367 |
+
],
|
| 368 |
+
"page_idx": 3
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"type": "text",
|
| 372 |
+
"text": "Inspired by Theorem 1 and the multi-scale periodic representation of grid cells in mammals (Abbott & Callaway, 2014) we set up our point space encoder $\\mathbf { e } [ \\mathbf { x } ] \\stackrel { - } { = } E n c _ { t h e o r y } ^ { ( x ) } ( \\mathbf { x } )$ to use sine and cosine functions of different frequencies to encode positions in space. Given any point $\\mathbf { x }$ in the studied 2D space, the space encoder $E n c _ { t h e o r y } ^ { ( x ) } ( \\mathbf { x } ) \\ = \\ \\mathbf { N N } ( P E ^ { ( t ) } ( \\mathbf { x } ) )$ where $P E ^ { ( t ) } ( \\mathbf { x } ) \\ =$ $[ P E _ { 0 } ^ { ( t ) } ( { \\bf x } ) ; . . . ; P E _ { s } ^ { ( t ) } ( { \\bf x } ) ; . . . ; P E _ { S - 1 } ^ { ( t ) } ( { \\bf x } ) ]$ is a concatenation of multi-scale representations of $d ^ { ( x ) } =$ $6 S$ dimensions. Here $S$ is the total number of grid scales and $s = 0 , 1 , 2 , . . . , S - 1 . \\ \\mathbf { N N } ( )$ represents fully connected ReLU layers. Let $\\mathbf { a } _ { 1 } = [ 1 , 0 ] ^ { T } , \\mathbf { a } _ { 2 } = [ - 1 / 2 , \\sqrt { 3 } / 2 ] ^ { T } , \\mathbf { a } _ { 3 } = [ - 1 / 2 , - \\sqrt { 3 } / 2 ] ^ { T } \\in \\mathbb { R } ^ { 2 }$ be three unit vectors and the angle between any of them is $2 \\pi / 3$ . $\\lambda _ { m i n } , \\lambda _ { m a x }$ are the minimum and maximum grid scale and $\\begin{array} { r } { g = \\frac { \\lambda _ { m a x } } { \\lambda _ { m i n } } } \\end{array}$ . At each scale $s , P E _ { s } ^ { ( t ) } ( \\mathbf { x } ) = [ P E _ { s , 1 } ^ { ( t ) } ( \\mathbf { x } ) ; P E _ { s , 2 } ^ { ( t ) } ( \\mathbf { x } ) ; P E _ { s , 3 } ^ { ( t ) } ( \\mathbf { x } ) ]$ is a concatenation of three components, where ",
|
| 373 |
+
"bbox": [
|
| 374 |
+
173,
|
| 375 |
+
892,
|
| 376 |
+
825,
|
| 377 |
+
925
|
| 378 |
+
],
|
| 379 |
+
"page_idx": 3
|
| 380 |
+
},
|
| 381 |
+
{
|
| 382 |
+
"type": "text",
|
| 383 |
+
"text": "",
|
| 384 |
+
"bbox": [
|
| 385 |
+
173,
|
| 386 |
+
102,
|
| 387 |
+
826,
|
| 388 |
+
234
|
| 389 |
+
],
|
| 390 |
+
"page_idx": 4
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"type": "equation",
|
| 394 |
+
"img_path": "images/6439bbc100d2d88f5b620ba119ef46dfd81abcf8fc141857ce56878dcb309ccf.jpg",
|
| 395 |
+
"text": "$$\nP E _ { s , j } ^ { ( t ) } ( \\mathbf { x } ) = [ \\cos ( \\frac { \\langle \\mathbf { x } , \\mathbf { a } _ { j } \\rangle } { \\lambda _ { m i n } \\cdot g ^ { s / ( S - 1 ) } } ) ; \\sin ( \\frac { \\langle \\mathbf { x } , \\mathbf { a } _ { j } \\rangle } { \\lambda _ { m i n } \\cdot g ^ { s / ( S - 1 ) } } ) ] \\forall j = 1 , 2 , 3 ;\n$$",
|
| 396 |
+
"text_format": "latex",
|
| 397 |
+
"bbox": [
|
| 398 |
+
264,
|
| 399 |
+
236,
|
| 400 |
+
733,
|
| 401 |
+
270
|
| 402 |
+
],
|
| 403 |
+
"page_idx": 4
|
| 404 |
+
},
|
| 405 |
+
{
|
| 406 |
+
"type": "text",
|
| 407 |
+
"text": "$\\mathbf { N N } ( )$ and $P E ^ { ( t ) } ( \\mathbf { x } )$ are analogies of $\\mathbf { C }$ and $\\Psi ( \\mathbf { x } )$ in Theorem 1. ",
|
| 408 |
+
"bbox": [
|
| 409 |
+
174,
|
| 410 |
+
281,
|
| 411 |
+
604,
|
| 412 |
+
297
|
| 413 |
+
],
|
| 414 |
+
"page_idx": 4
|
| 415 |
+
},
|
| 416 |
+
{
|
| 417 |
+
"type": "text",
|
| 418 |
+
"text": "Similarly we can define another space encoder $E n c _ { g r i d } ^ { ( x ) } ( { \\bf x } ) ~ = ~ { \\bf N N } ( P E ^ { ( g ) } ( { \\bf x } ) )$ inspired by the position encoding model of Transformer (Vaswani et al., 2017), where $\\begin{array} { r l } { P E ^ { ( g ) } ( \\mathbf { x } ) } & { { } = } \\end{array}$ $[ { \\bar { P E _ { 0 } ^ { ( g ) } } } ( \\mathbf { x } ) ; . . . ; { \\bar { P E _ { s } ^ { ( g ) } ( \\mathbf { x } ) } } ; . . . ; { \\bar { P E _ { S - 1 } ^ { ( g ) } ( \\mathbf { x } ) } } ]$ is still a concatenation of its multi-scale representations, while $P E _ { s } ^ { ( g ) } ( \\mathbf { x } ) = [ P E _ { s , 1 } ^ { ( g ) } ( \\mathbf { x } ) ; P E _ { s , 2 } ^ { ( g ) } ( \\mathbf { x } ) ]$ handles each component $l$ of $\\mathbf { x }$ separately: ",
|
| 419 |
+
"bbox": [
|
| 420 |
+
173,
|
| 421 |
+
305,
|
| 422 |
+
826,
|
| 423 |
+
380
|
| 424 |
+
],
|
| 425 |
+
"page_idx": 4
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"type": "equation",
|
| 429 |
+
"img_path": "images/1d2c4e3477ee6deade4d260042297ca5a6a75818ced60c8d99d64c850a2fb7d5.jpg",
|
| 430 |
+
"text": "$$\nP E _ { s , l } ^ { ( g ) } ( \\mathbf { x } ) = [ \\cos ( \\frac { \\mathbf { x } ^ { [ l ] } } { \\lambda _ { m i n } \\cdot g ^ { s / ( S - 1 ) } } ) ; \\sin ( \\frac { \\mathbf { x } ^ { [ l ] } } { \\lambda _ { m i n } \\cdot g ^ { s / ( S - 1 ) } } ) ] \\forall l = 1 , 2\n$$",
|
| 431 |
+
"text_format": "latex",
|
| 432 |
+
"bbox": [
|
| 433 |
+
274,
|
| 434 |
+
382,
|
| 435 |
+
723,
|
| 436 |
+
419
|
| 437 |
+
],
|
| 438 |
+
"page_idx": 4
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"type": "text",
|
| 442 |
+
"text": "4.2 DECODER ",
|
| 443 |
+
"text_level": 1,
|
| 444 |
+
"bbox": [
|
| 445 |
+
174,
|
| 446 |
+
412,
|
| 447 |
+
285,
|
| 448 |
+
428
|
| 449 |
+
],
|
| 450 |
+
"page_idx": 4
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"type": "text",
|
| 454 |
+
"text": "Two types of decoders are designed for two major types of GIS problems: location modeling and spatial context modeling (See Section 5.1). ",
|
| 455 |
+
"bbox": [
|
| 456 |
+
174,
|
| 457 |
+
431,
|
| 458 |
+
823,
|
| 459 |
+
460
|
| 460 |
+
],
|
| 461 |
+
"page_idx": 4
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"type": "text",
|
| 465 |
+
"text": "Location Decoder $D e c _ { s } ( )$ directly reconstructs point feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ given its space embedding $\\mathbf { e } [ \\mathbf { x } _ { i } ]$ . We use one layer feed-forward neural network $\\mathbf { N N } _ { \\mathrm { d e c } } ($ q ",
|
| 466 |
+
"bbox": [
|
| 467 |
+
173,
|
| 468 |
+
474,
|
| 469 |
+
823,
|
| 470 |
+
503
|
| 471 |
+
],
|
| 472 |
+
"page_idx": 4
|
| 473 |
+
},
|
| 474 |
+
{
|
| 475 |
+
"type": "equation",
|
| 476 |
+
"img_path": "images/c48a8c7ecbe0def9b2c82672484632ea4947daa4f433a0f9e8e9edda2b2eb1e7.jpg",
|
| 477 |
+
"text": "$$\n\\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime } = D e c _ { s } ( \\mathbf { x } _ { i } ; \\theta _ { \\mathrm { d e c } _ { s } } ) = \\mathbf { N } \\mathbf { N } _ { \\mathrm { d e c } } ( \\mathbf { e } [ \\mathbf { x } _ { i } ] )\n$$",
|
| 478 |
+
"text_format": "latex",
|
| 479 |
+
"bbox": [
|
| 480 |
+
356,
|
| 481 |
+
507,
|
| 482 |
+
640,
|
| 483 |
+
523
|
| 484 |
+
],
|
| 485 |
+
"page_idx": 4
|
| 486 |
+
},
|
| 487 |
+
{
|
| 488 |
+
"type": "text",
|
| 489 |
+
"text": "For training we use inner product to compare the reconstructed feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime }$ against the real feature embeddings of $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ and other negative points (see training detail in Sec 4.3). ",
|
| 490 |
+
"bbox": [
|
| 491 |
+
169,
|
| 492 |
+
527,
|
| 493 |
+
823,
|
| 494 |
+
556
|
| 495 |
+
],
|
| 496 |
+
"page_idx": 4
|
| 497 |
+
},
|
| 498 |
+
{
|
| 499 |
+
"type": "text",
|
| 500 |
+
"text": "Spatial Context Decoder $D e c _ { c } ( )$ reconstructs the feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ of the center point $p _ { i }$ based on the space and feature embeddings $\\left\\{ \\mathbf { e } _ { i 1 } , . . . , \\mathbf { e } _ { i j } , . . . , \\mathbf { e } _ { i n } \\right\\}$ of $n$ nearby points $\\{ p _ { i 1 } , . . . , p _ { i j } , . . . , p _ { i n } \\}$ . Note that the feed-in order of context points should not affect the prediction results, which can be achieved by permutation invariant neural network architectures (Zaheer et al., 2017) like PointNet (Qi et al., 2017). ",
|
| 501 |
+
"bbox": [
|
| 502 |
+
174,
|
| 503 |
+
569,
|
| 504 |
+
826,
|
| 505 |
+
641
|
| 506 |
+
],
|
| 507 |
+
"page_idx": 4
|
| 508 |
+
},
|
| 509 |
+
{
|
| 510 |
+
"type": "equation",
|
| 511 |
+
"img_path": "images/cc9f21c2f4471345e2b9726dd97f71df84569cbfc3e2d848290a87008af3ac4d.jpg",
|
| 512 |
+
"text": "$$\n\\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime } = D e c _ { c } ( \\mathbf { x } _ { i } , \\{ \\mathbf { e } _ { i 1 } , . . . , \\mathbf { e } _ { i j } , . . . , \\mathbf { e } _ { i n } \\} ; \\theta _ { \\mathrm { d e c } _ { c } } ) = g ( \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\sum _ { j = 1 } ^ { n } \\alpha _ { i j k } \\mathbf { e } [ \\mathbf { v } _ { i j } ] )\n$$",
|
| 513 |
+
"text_format": "latex",
|
| 514 |
+
"bbox": [
|
| 515 |
+
251,
|
| 516 |
+
643,
|
| 517 |
+
745,
|
| 518 |
+
686
|
| 519 |
+
],
|
| 520 |
+
"page_idx": 4
|
| 521 |
+
},
|
| 522 |
+
{
|
| 523 |
+
"type": "text",
|
| 524 |
+
"text": "Here g is an activation function such as sigmoid. αijk “ $\\begin{array} { r } { \\alpha _ { i j k } = \\frac { e x p ( \\sigma _ { i j k } ) } { \\sum _ { o = 1 } ^ { n } e x p ( \\sigma _ { i o k } ) } } \\end{array}$ is the attention of $p _ { i }$ with its $j$ th neighbor through the $k$ th attention head, and ",
|
| 525 |
+
"bbox": [
|
| 526 |
+
174,
|
| 527 |
+
691,
|
| 528 |
+
823,
|
| 529 |
+
726
|
| 530 |
+
],
|
| 531 |
+
"page_idx": 4
|
| 532 |
+
},
|
| 533 |
+
{
|
| 534 |
+
"type": "equation",
|
| 535 |
+
"img_path": "images/eea7c5aaad49682e0174fd7a4a34688391a773471e5a7b35138a73a3f202ef55.jpg",
|
| 536 |
+
"text": "$$\n\\sigma _ { i j k } = L e a k y R e L U ( \\mathbf { a } _ { k } ^ { T } [ \\mathbf { e } [ \\mathbf { v } _ { i } ] _ { i n i t } ; \\mathbf { e } [ \\mathbf { v } _ { i j } ] ; \\mathbf { e } [ \\mathbf { x } _ { i } - \\mathbf { x } _ { i j } ] ] )\n$$",
|
| 537 |
+
"text_format": "latex",
|
| 538 |
+
"bbox": [
|
| 539 |
+
310,
|
| 540 |
+
729,
|
| 541 |
+
687,
|
| 542 |
+
748
|
| 543 |
+
],
|
| 544 |
+
"page_idx": 4
|
| 545 |
+
},
|
| 546 |
+
{
|
| 547 |
+
"type": "text",
|
| 548 |
+
"text": "where $\\mathbf { a } _ { k } \\in \\mathbb { R } ^ { 2 d ^ { ( v ) } + d ^ { ( x ) } }$ is the attention parameter in the $k$ th attention head. The multi-head attention mechanism is inspired by Graph Attention Network (Velickovi ˇ c et al., 2018) and Mai et al. (2019a). ´ ",
|
| 549 |
+
"bbox": [
|
| 550 |
+
171,
|
| 551 |
+
755,
|
| 552 |
+
823,
|
| 553 |
+
785
|
| 554 |
+
],
|
| 555 |
+
"page_idx": 4
|
| 556 |
+
},
|
| 557 |
+
{
|
| 558 |
+
"type": "text",
|
| 559 |
+
"text": "To represent the spatial relationship (distance and direction) between each context point $\\begin{array} { r l } { p _ { i j } } & { { } = } \\end{array}$ $( \\mathbf { x } _ { i j } , \\mathbf { v } _ { i j } )$ and the center point $p _ { i } = ( \\mathbf { x } _ { i } , \\mathbf { v } _ { i } )$ , we use the space encoder $E n c ^ { ( x ) } ( \\ v r )$ to encode the displacement between them $\\Delta { \\bf x } _ { i j } = { \\bf x } _ { i } - { \\bf x } _ { i j }$ . Note that we are modeling the spatial interactions between the center point and $n$ context points simultaneously. ",
|
| 560 |
+
"bbox": [
|
| 561 |
+
173,
|
| 562 |
+
790,
|
| 563 |
+
825,
|
| 564 |
+
851
|
| 565 |
+
],
|
| 566 |
+
"page_idx": 4
|
| 567 |
+
},
|
| 568 |
+
{
|
| 569 |
+
"type": "text",
|
| 570 |
+
"text": "In Eq. $7 , \\mathbf { e } [ \\mathbf { v } _ { i } ] _ { i n i t }$ indicates the initial guess of the feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ of point $p _ { i }$ which is computed by using another multi-head attention layer as Eq. 6 where the weight $\\begin{array} { r } { \\alpha _ { i j k } ^ { \\prime } = \\frac { e x p ( \\sigma _ { i j k } ^ { \\prime } ) } { \\sum _ { o = 1 } ^ { n } e x p ( \\sigma _ { i o k } ^ { \\prime } ) } } \\end{array}$ . Here, $\\sigma _ { i j k } ^ { \\prime }$ is computed as Eq. 8 where the query embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ is excluded. ",
|
| 571 |
+
"bbox": [
|
| 572 |
+
173,
|
| 573 |
+
856,
|
| 574 |
+
825,
|
| 575 |
+
910
|
| 576 |
+
],
|
| 577 |
+
"page_idx": 4
|
| 578 |
+
},
|
| 579 |
+
{
|
| 580 |
+
"type": "equation",
|
| 581 |
+
"img_path": "images/d666837d39bcbc7a24c7b161a0062bc55a8b9831ec4741cbe96acbcd48820048.jpg",
|
| 582 |
+
"text": "$$\n\\sigma _ { i j k } ^ { \\prime } = L e a k y R e L U ( \\mathbf { a } _ { k } ^ { \\prime T } [ \\mathbf { e } [ \\mathbf { v } _ { i j } ] ; \\mathbf { e } [ \\mathbf { x } _ { i } - \\mathbf { x } _ { i j } ] ] )\n$$",
|
| 583 |
+
"text_format": "latex",
|
| 584 |
+
"bbox": [
|
| 585 |
+
341,
|
| 586 |
+
917,
|
| 587 |
+
656,
|
| 588 |
+
938
|
| 589 |
+
],
|
| 590 |
+
"page_idx": 4
|
| 591 |
+
},
|
| 592 |
+
{
|
| 593 |
+
"type": "text",
|
| 594 |
+
"text": "4.3 UNSUPERVISED TRAINING ",
|
| 595 |
+
"text_level": 1,
|
| 596 |
+
"bbox": [
|
| 597 |
+
174,
|
| 598 |
+
103,
|
| 599 |
+
400,
|
| 600 |
+
117
|
| 601 |
+
],
|
| 602 |
+
"page_idx": 5
|
| 603 |
+
},
|
| 604 |
+
{
|
| 605 |
+
"type": "text",
|
| 606 |
+
"text": "The unsupervised learning task can simply be maximizing the log likelihood of observing the true point $p _ { i }$ at position $\\mathbf { x } _ { i }$ among all the points in $\\mathcal { P }$ ",
|
| 607 |
+
"bbox": [
|
| 608 |
+
174,
|
| 609 |
+
122,
|
| 610 |
+
826,
|
| 611 |
+
151
|
| 612 |
+
],
|
| 613 |
+
"page_idx": 5
|
| 614 |
+
},
|
| 615 |
+
{
|
| 616 |
+
"type": "equation",
|
| 617 |
+
"img_path": "images/3dc824b7500ca5b23a62820474f3af1dde0dc3d94c376e95052cb9301dfd0bed.jpg",
|
| 618 |
+
"text": "$$\n\\mathcal { L } _ { \\mathcal { P } } ( \\boldsymbol { \\theta } ) = - \\sum _ { p _ { i } \\in \\mathcal { P } } \\log P ( p _ { i } | p _ { i 1 } , . . . , p _ { i j } , . . . , p _ { i n } ) = - \\sum _ { p _ { i } \\in \\mathcal { P } } \\log \\frac { \\exp ( \\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { T } \\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime } ) } { \\sum _ { p _ { o } \\in \\mathcal { P } } \\exp ( \\mathbf { e } [ \\mathbf { v } _ { o } ] ^ { T } \\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime } ) }\n$$",
|
| 619 |
+
"text_format": "latex",
|
| 620 |
+
"bbox": [
|
| 621 |
+
214,
|
| 622 |
+
154,
|
| 623 |
+
782,
|
| 624 |
+
194
|
| 625 |
+
],
|
| 626 |
+
"page_idx": 5
|
| 627 |
+
},
|
| 628 |
+
{
|
| 629 |
+
"type": "text",
|
| 630 |
+
"text": "Here only the feature embedding of $p _ { i }$ is used (without location embedding) to prevent revealing the identities of the point candidates, and $\\theta = [ \\theta _ { \\mathrm { { e n c } } } ; \\theta _ { \\mathrm { { d e c } } } ]$ ",
|
| 631 |
+
"bbox": [
|
| 632 |
+
174,
|
| 633 |
+
195,
|
| 634 |
+
828,
|
| 635 |
+
226
|
| 636 |
+
],
|
| 637 |
+
"page_idx": 5
|
| 638 |
+
},
|
| 639 |
+
{
|
| 640 |
+
"type": "text",
|
| 641 |
+
"text": "Negative sampling by Mikolov et al. (2013) can be used to improve the efficiency of training ",
|
| 642 |
+
"bbox": [
|
| 643 |
+
176,
|
| 644 |
+
231,
|
| 645 |
+
782,
|
| 646 |
+
246
|
| 647 |
+
],
|
| 648 |
+
"page_idx": 5
|
| 649 |
+
},
|
| 650 |
+
{
|
| 651 |
+
"type": "equation",
|
| 652 |
+
"img_path": "images/6753cf1bc3ffeb52dba95fbe5d509f8651bb39370b0c8e73baa138ff88e67298.jpg",
|
| 653 |
+
"text": "$$\n\\mathcal { L } _ { \\mathcal { P } } ^ { \\prime } ( \\boldsymbol { \\theta } ) = - \\sum _ { p _ { i } \\in \\mathcal { P } } \\Big ( \\log \\sigma ( \\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { T } \\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime } ) + \\frac { 1 } { | \\mathcal { N } _ { i } | } \\sum _ { p _ { o } \\in \\mathcal { N } _ { i } } \\log \\sigma ( - \\mathbf { e } [ \\mathbf { v } _ { o } ] ^ { T } \\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime } ) \\Big )\n$$",
|
| 654 |
+
"text_format": "latex",
|
| 655 |
+
"bbox": [
|
| 656 |
+
243,
|
| 657 |
+
247,
|
| 658 |
+
753,
|
| 659 |
+
286
|
| 660 |
+
],
|
| 661 |
+
"page_idx": 5
|
| 662 |
+
},
|
| 663 |
+
{
|
| 664 |
+
"type": "text",
|
| 665 |
+
"text": "Here ${ \\mathcal { N } } _ { i } \\subseteq { \\mathcal { P } }$ is a set of sampled negative points for $p _ { i } \\ ( p _ { i } \\notin { \\mathcal { N } } _ { i } )$ and $\\sigma ( x ) = 1 / ( 1 + e ^ { - x } )$ . ",
|
| 666 |
+
"bbox": [
|
| 667 |
+
171,
|
| 668 |
+
289,
|
| 669 |
+
767,
|
| 670 |
+
305
|
| 671 |
+
],
|
| 672 |
+
"page_idx": 5
|
| 673 |
+
},
|
| 674 |
+
{
|
| 675 |
+
"type": "text",
|
| 676 |
+
"text": "5 EXPERIMENT ",
|
| 677 |
+
"text_level": 1,
|
| 678 |
+
"bbox": [
|
| 679 |
+
174,
|
| 680 |
+
315,
|
| 681 |
+
318,
|
| 682 |
+
333
|
| 683 |
+
],
|
| 684 |
+
"page_idx": 5
|
| 685 |
+
},
|
| 686 |
+
{
|
| 687 |
+
"type": "text",
|
| 688 |
+
"text": "In this section we compare Space2Vec with commonly used position encoding methods, and analyze them both quantitatively and qualitatively. ",
|
| 689 |
+
"bbox": [
|
| 690 |
+
173,
|
| 691 |
+
339,
|
| 692 |
+
823,
|
| 693 |
+
369
|
| 694 |
+
],
|
| 695 |
+
"page_idx": 5
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"type": "text",
|
| 699 |
+
"text": "Baselines Our baselines include 1) direct directly applying feed-forward nets (Chu et al., 2019); 2) tile discretization (Berg et al., 2014; Adams et al., 2015; Tang et al., 2015); 3) wrap feed-forward nets with coordinate wrapping (Mac Aodha et al., 2019); and 4) rbf Radial Basis Function (RBF) kernels (Baudat & Anouar, 2001; Bierens, 1994). See Appendix A.1 for details of the baselines. ",
|
| 700 |
+
"bbox": [
|
| 701 |
+
174,
|
| 702 |
+
382,
|
| 703 |
+
825,
|
| 704 |
+
439
|
| 705 |
+
],
|
| 706 |
+
"page_idx": 5
|
| 707 |
+
},
|
| 708 |
+
{
|
| 709 |
+
"type": "text",
|
| 710 |
+
"text": "5.1 POI TYPE CLASSIFICATION TASKS ",
|
| 711 |
+
"text_level": 1,
|
| 712 |
+
"bbox": [
|
| 713 |
+
176,
|
| 714 |
+
448,
|
| 715 |
+
455,
|
| 716 |
+
462
|
| 717 |
+
],
|
| 718 |
+
"page_idx": 5
|
| 719 |
+
},
|
| 720 |
+
{
|
| 721 |
+
"type": "text",
|
| 722 |
+
"text": "Dataset and Tasks To test the proposed model, we conduct experiments on geographic datasets with POI position and type information. We utilize the open-source dataset published by Yelp Data Challenge and select all POIs within the Las Vegas downtown area3. There are 21,830 POIs with 1,191 different POI types in this dataset. Note that each POI may be associated with one or more types, and we do not use any other meta-data such as business names, reviews for this study. We project geographic coordinates into projection coordinates using the NAD83/Conus Albers projection coordinate system4. The POIs are split into training, validation, and test dataset with ratios $8 0 \\% { : } 1 0 \\% { : } 1 0 \\%$ . We create two tasks setups which represent different types of modeling need in Geographic Information Science: ",
|
| 723 |
+
"bbox": [
|
| 724 |
+
173,
|
| 725 |
+
467,
|
| 726 |
+
825,
|
| 727 |
+
592
|
| 728 |
+
],
|
| 729 |
+
"page_idx": 5
|
| 730 |
+
},
|
| 731 |
+
{
|
| 732 |
+
"type": "text",
|
| 733 |
+
"text": "Location Modeling predicts the feature information associated with a POI based on its location $\\mathbf { x } _ { i }$ represented by the location decoder $D e c _ { s } ( )$ . This represents a large number of location prediction problems such as image fine grained recognition with geographic prior (Chu et al., 2019), and species potential distribution prediction (Zuo et al., 2008). Spatial Context Modeling predicts the feature information associated with a POI based on its context $\\left\\{ \\mathbf { e } _ { i 1 } , . . . , \\mathbf { e } _ { i j } , . . . , \\mathbf { e } _ { i n } \\right\\}$ represented by the spatial context decoder $D e c _ { c } ( )$ . This represents a collections of spatial context prediction problem such as spatial context based facade image classification (Yan et al., 2018), and all spatial interpolation problems. ",
|
| 734 |
+
"bbox": [
|
| 735 |
+
215,
|
| 736 |
+
602,
|
| 737 |
+
825,
|
| 738 |
+
718
|
| 739 |
+
],
|
| 740 |
+
"page_idx": 5
|
| 741 |
+
},
|
| 742 |
+
{
|
| 743 |
+
"type": "text",
|
| 744 |
+
"text": "We use POI prediction metrics to evaluate these models. Given the real point feature embedding $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ and $N$ negative feature embeddings ${ \\mathcal { N } } _ { i } = \\{ \\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { - } \\}$ , we compare the predicted $\\mathbf { e } [ \\mathbf { v } _ { i } ] ^ { \\prime }$ with them by cosine distance. The cosine scores are used to rank $\\mathbf { e } [ \\mathbf { v } _ { i } ]$ and $N$ negative samples. The negative feature embeddings are the feature embeddings of points $p _ { j }$ randomly sampled from $\\mathcal { P }$ and $p _ { i } \\neq p _ { j }$ . We evaluate each model using Negative Log-Likelihood (NLL), Mean Reciprocal Rank (MRR) and $\\mathrm { H I T } @ 5$ (the chance of the true POI being ranked to top 5. We train and test each model 10 times to estimate standard deviations. See Appendix A.2 for hyper-parameter selection details. ",
|
| 745 |
+
"bbox": [
|
| 746 |
+
173,
|
| 747 |
+
728,
|
| 748 |
+
825,
|
| 749 |
+
827
|
| 750 |
+
],
|
| 751 |
+
"page_idx": 5
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"type": "text",
|
| 755 |
+
"text": "5.1.1 LOCATION MODELING EVALUATION ",
|
| 756 |
+
"text_level": 1,
|
| 757 |
+
"bbox": [
|
| 758 |
+
176,
|
| 759 |
+
838,
|
| 760 |
+
480,
|
| 761 |
+
853
|
| 762 |
+
],
|
| 763 |
+
"page_idx": 5
|
| 764 |
+
},
|
| 765 |
+
{
|
| 766 |
+
"type": "text",
|
| 767 |
+
"text": "We first study location modeling with the location decoder $D e c _ { s } ( )$ in Section 4.2. We use a negative sample size of $N = 1 0 0$ . Table 1 shows the average metrics of different models with their best hyperparameter setting on the validation set. We can see that direct and theorydiag are less competitive, only beating the random selection baseline. Other methods with single scale representations – including tile, wrap, and $r b f$ – perform better. The best results come from various version of the grid cell models, which are capable of dealing with multi-scale representations. ",
|
| 768 |
+
"bbox": [
|
| 769 |
+
178,
|
| 770 |
+
859,
|
| 771 |
+
821,
|
| 772 |
+
887
|
| 773 |
+
],
|
| 774 |
+
"page_idx": 5
|
| 775 |
+
},
|
| 776 |
+
{
|
| 777 |
+
"type": "image",
|
| 778 |
+
"img_path": "images/9d23055766f68f88b9affd265a715f83240e297e834f2e503eef7c623966f7b8.jpg",
|
| 779 |
+
"image_caption": [
|
| 780 |
+
"Figure 2: Embedding clustering of (a) direct; (b) tile with the best cell size $c = 5 0 0$ ; (c) wrap $( h = 3 , o =$ 512); (d) $r b f$ with the best $\\sigma$ (1k) and 200 anchor points (red) and (e)(f)(h) theory models with different $\\lambda _ { m i n }$ but fixed $\\lambda _ { m a x } = 4 0 k$ and $S = 6 4$ . All models use 1 hidden ReLU layers of 512 neurons except wrap. "
|
| 781 |
+
],
|
| 782 |
+
"image_footnote": [],
|
| 783 |
+
"bbox": [
|
| 784 |
+
174,
|
| 785 |
+
98,
|
| 786 |
+
818,
|
| 787 |
+
202
|
| 788 |
+
],
|
| 789 |
+
"page_idx": 6
|
| 790 |
+
},
|
| 791 |
+
{
|
| 792 |
+
"type": "table",
|
| 793 |
+
"img_path": "images/1d6f6b314e1e6a4c32bd2c2f1deca868f0e009968e3dbf0f4cffb640a34691dd.jpg",
|
| 794 |
+
"table_caption": [
|
| 795 |
+
"Table 1: The evaluation results of different location models on the validation and test dataset. "
|
| 796 |
+
],
|
| 797 |
+
"table_footnote": [],
|
| 798 |
+
"table_body": "<table><tr><td rowspan=\"2\"></td><td rowspan=\"2\">Train NLL</td><td colspan=\"3\">Validation</td><td colspan=\"2\">Testing</td></tr><tr><td>NLL</td><td>MRR</td><td>HIT@5</td><td>MRR</td><td>HIT@5</td></tr><tr><td>random</td><td></td><td>-</td><td>0.052 (0.002)</td><td>4.8 (0.5)</td><td>0.051 (0.002)</td><td>5.0 (0.5)</td></tr><tr><td>direct</td><td>1.285</td><td>1.332</td><td>0.089 (0.001)</td><td>10.6 (0.2)</td><td>0.090 (0.001)</td><td>11.3 (0.2)</td></tr><tr><td>tile (c=500)</td><td>1.118</td><td>1.261</td><td>0.123 (0.001)</td><td>16.8 (0.2)</td><td>0.120 (0.001)</td><td>17.1 (0.3)</td></tr><tr><td>wrap(h=3,0=512)</td><td>1.222</td><td>1.288</td><td>0.112 (0.001)</td><td>14.6 (0.1)</td><td>0.119 (0.001)</td><td>15.8 (0.2)</td></tr><tr><td>rbf (σ=1k)</td><td>1.209</td><td>1.279</td><td>0.115 (0.001)</td><td>15.2 (0.2)</td><td>0.123 (0.001)</td><td>16.8 (0.3)</td></tr><tr><td>grid (Xmin=50)</td><td>1.156</td><td>1.258</td><td>0.128 (0.001)</td><td>18.1 (0.3)</td><td>0.139 (0.001)</td><td>20.0 (0.2)</td></tr><tr><td>hexa (Xmin=50)</td><td>1.230</td><td>1.297</td><td>0.107 (0.001)</td><td>14.0 (0.2)</td><td>0.105 (0.001)</td><td>14.5 (0.2)</td></tr><tr><td>theorydiag (Xmin=50)</td><td>1.277</td><td>1.324</td><td>0.094 (0.001)</td><td>12.3 (0.3)</td><td>0.094 (0.002)</td><td>11.2 (0.3)</td></tr><tr><td>theory (Xmin=1k)</td><td>1.207</td><td>1.281</td><td>0.123 (0.002)</td><td>16.3 (0.5)</td><td>0.121 (0.001)</td><td>16.2 (0.1)</td></tr><tr><td>theory (Xmin=500)</td><td>1.188</td><td>1.269</td><td>0.132 (0.001)</td><td>17.6 (0.3)</td><td>0.129 (0.001)</td><td>17.7 (0.2)</td></tr><tr><td>theory (Xmin=50)</td><td>1.098</td><td>1.249</td><td>0.137 (0.002)</td><td>19.4 (0.1)</td><td>0.144 (0.001)</td><td>20.0 (0.2)</td></tr></table>",
|
| 799 |
+
"bbox": [
|
| 800 |
+
186,
|
| 801 |
+
289,
|
| 802 |
+
810,
|
| 803 |
+
463
|
| 804 |
+
],
|
| 805 |
+
"page_idx": 6
|
| 806 |
+
},
|
| 807 |
+
{
|
| 808 |
+
"type": "text",
|
| 809 |
+
"text": "",
|
| 810 |
+
"bbox": [
|
| 811 |
+
174,
|
| 812 |
+
479,
|
| 813 |
+
826,
|
| 814 |
+
535
|
| 815 |
+
],
|
| 816 |
+
"page_idx": 6
|
| 817 |
+
},
|
| 818 |
+
{
|
| 819 |
+
"type": "text",
|
| 820 |
+
"text": "In order to understand the reason for the superiority of grid cell models we provide qualitative analysis of their representations. We apply hierarchical clustering to the location embeddings produced by studied models using cosine distance as the distance metric (See Fig. 2). we can see that when restricted to large grid sizes $( \\lambda _ { m i n } = 1 k )$ ), theory has similar representation (Fig. 2d, 2e, and Fig. 4d, 4e) and performance compared to $r b f$ $\\mathit { \\Pi } _ { \\overline { { \\sigma } } } = 1 k \\mathit { \\Pi } _ { \\overline { { \\theta } } }$ ). However it is able to significantly outperform $r b f$ $\\sigma = 1 k$ ) (and tile and wrap) when small grid sizes $\\langle \\lambda _ { m i n } = 5 0 0 , 5 0 )$ are available. The relative improvements over $r b f$ $( \\sigma = 1 k )$ ) are $- 0 . 2 \\%$ , $+ 0 . 6 \\%$ , $+ 2 . 1 \\%$ MRR for $\\lambda _ { m i n } { = } 1 \\mathrm { k }$ , 500, 50 respectively. ",
|
| 821 |
+
"bbox": [
|
| 822 |
+
173,
|
| 823 |
+
542,
|
| 824 |
+
826,
|
| 825 |
+
640
|
| 826 |
+
],
|
| 827 |
+
"page_idx": 6
|
| 828 |
+
},
|
| 829 |
+
{
|
| 830 |
+
"type": "text",
|
| 831 |
+
"text": "5.1.2 MULTI-SCALE ANALYSIS OF LOCATION MODELING ",
|
| 832 |
+
"text_level": 1,
|
| 833 |
+
"bbox": [
|
| 834 |
+
173,
|
| 835 |
+
652,
|
| 836 |
+
588,
|
| 837 |
+
667
|
| 838 |
+
],
|
| 839 |
+
"page_idx": 6
|
| 840 |
+
},
|
| 841 |
+
{
|
| 842 |
+
"type": "text",
|
| 843 |
+
"text": "In order to show how our multi-scale location representation model will affect the prediction of POI types with different distribution patterns, we classify all 1,191 POI types into three groups based on radius $r$ , which is derived from each POI types’ renormalized Ripley’s K curve (See Figure 1d for examples). It indicates the $\\mathbf { X }$ axis value of the intersection between the curve and the line of $y = 3 . 0$ A lower $r$ indicates a more clustered distribution patterns. These three groups are listed below: ",
|
| 844 |
+
"bbox": [
|
| 845 |
+
174,
|
| 846 |
+
674,
|
| 847 |
+
825,
|
| 848 |
+
743
|
| 849 |
+
],
|
| 850 |
+
"page_idx": 6
|
| 851 |
+
},
|
| 852 |
+
{
|
| 853 |
+
"type": "text",
|
| 854 |
+
"text": "1. Clustered $( r \\leqslant 1 0 0 m )$ ): POI types with clustered distribution patterns; \n2. Middle ( $1 0 0 m < r < 2 0 0 m$ ): POI types with less extreme scales; \n3. Even $r \\geqslant 2 0 0 m$ ): POI types with even distribution patterns. ",
|
| 855 |
+
"bbox": [
|
| 856 |
+
210,
|
| 857 |
+
753,
|
| 858 |
+
689,
|
| 859 |
+
808
|
| 860 |
+
],
|
| 861 |
+
"page_idx": 6
|
| 862 |
+
},
|
| 863 |
+
{
|
| 864 |
+
"type": "text",
|
| 865 |
+
"text": "Table 2 shows the performance $( M R R )$ of direct, tile, wrap, rbf , and our theory model on the test dataset of the location modeling task with respect to these three different POI distribution groups. The numbers in pq indicate the MRR difference betweeb a baseline and theory. # POI refers to total number of POI belong to each group5. We can see that 1) The two neural net approaches (direct and wrap) have no scale related parameter and are not performing ideally across all scales, with direct ",
|
| 866 |
+
"bbox": [
|
| 867 |
+
173,
|
| 868 |
+
818,
|
| 869 |
+
826,
|
| 870 |
+
888
|
| 871 |
+
],
|
| 872 |
+
"page_idx": 6
|
| 873 |
+
},
|
| 874 |
+
{
|
| 875 |
+
"type": "text",
|
| 876 |
+
"text": "Table 2: Comparing performances in different POI groups. We classify all 1,191 POI types into three groups based on the radius $r$ of their root types, where their renormalized Ripley’s K curve (See Figure 1d) reach 3.0: 1) Clustered $( r \\leqslant 1 0 0 m )$ ): POI types with clustered distribution patterns; 2) Middle $1 0 0 m < r < 2 0 0 m$ ): POI types with unclear distribution patterns; 3) Even $( r \\geqslant 2 0 0 m )$ : POI types with even distribution patterns. The MRR of wrap and theory on those three groups are shown. The numbers in pq indicate the difference between the MRR of a baseline model and the MRR of theory with respect to a specific group. $\\# P O I$ refers to the total number of POIs belonging to each group. Root Types indicates the root categories of those POI types belong to each group. ",
|
| 877 |
+
"bbox": [
|
| 878 |
+
173,
|
| 879 |
+
99,
|
| 880 |
+
825,
|
| 881 |
+
213
|
| 882 |
+
],
|
| 883 |
+
"page_idx": 7
|
| 884 |
+
},
|
| 885 |
+
{
|
| 886 |
+
"type": "table",
|
| 887 |
+
"img_path": "images/b4a2453a061c559c938ddf5d040a4ae30124f9dc6dbf80091e300e1dca531b21.jpg",
|
| 888 |
+
"table_caption": [],
|
| 889 |
+
"table_footnote": [],
|
| 890 |
+
"table_body": "<table><tr><td>POI Groups</td><td>Clustered (r≤100m)</td><td>Middle (100m<r<200m)</td><td>Even (r≥ 200m)</td></tr><tr><td>direct</td><td>0.080 (-0.047)</td><td>0.108 (-0.030)</td><td>0.084 (-0.047)</td></tr><tr><td>wrap</td><td>0.106 (-0.021)</td><td>0.126 (-0.012)</td><td>0.122 (-0.009)</td></tr><tr><td>tile</td><td>0.108 (-0.019)</td><td>0.135 (-0.003)</td><td>0.111 (-0.020)</td></tr><tr><td>rbf</td><td>0.112 (-0.015)</td><td>0.136 (-0.002)</td><td>0.119 (-0.012)</td></tr><tr><td>theory</td><td>0.127 (-)</td><td>0.138(-)</td><td>0.131(-)</td></tr><tr><td>#POI</td><td>16,016</td><td>7,443</td><td>3,915</td></tr><tr><td>Root Types</td><td>Restaurants; Shopping; Food; Nightlife; Automotive; Active Life; Arts & Entertainment; Financial Services</td><td>Beauty & Spas; Health & Medical; Local Services;Hotels & Travel; Professional Services; Public Services & Government</td><td>Home Services; Event Planning & Services; Pets; Education</td></tr></table>",
|
| 891 |
+
"bbox": [
|
| 892 |
+
176,
|
| 893 |
+
224,
|
| 894 |
+
813,
|
| 895 |
+
387
|
| 896 |
+
],
|
| 897 |
+
"page_idx": 7
|
| 898 |
+
},
|
| 899 |
+
{
|
| 900 |
+
"type": "text",
|
| 901 |
+
"text": "performs worse because of its simple single layer network. 2) The two approaches with built-in scale parameter (tile and $r b f$ ) have to trade off the performance of different scales. Their best parameter settings lead to close performances to that of Space2Vec at the middle scale, while performing poorly in both clustered and regular groups. These observation clearly shows that all baselines can at most well handle distribution at one scale but show poor performances in other scales. In contrast, Space2Vec’s multi-scale representation can handle distributions at different scales. ",
|
| 902 |
+
"bbox": [
|
| 903 |
+
174,
|
| 904 |
+
407,
|
| 905 |
+
825,
|
| 906 |
+
492
|
| 907 |
+
],
|
| 908 |
+
"page_idx": 7
|
| 909 |
+
},
|
| 910 |
+
{
|
| 911 |
+
"type": "text",
|
| 912 |
+
"text": "5.1.3 SPATIAL CONTEXT MODELING EVALUATION ",
|
| 913 |
+
"text_level": 1,
|
| 914 |
+
"bbox": [
|
| 915 |
+
174,
|
| 916 |
+
508,
|
| 917 |
+
539,
|
| 918 |
+
523
|
| 919 |
+
],
|
| 920 |
+
"page_idx": 7
|
| 921 |
+
},
|
| 922 |
+
{
|
| 923 |
+
"type": "text",
|
| 924 |
+
"text": "Next, we evaluate the spatial context decoder $D e c _ { c } ( )$ in Sec. 4.2. We use the same evaluation set up as location modeling. The context points are obtained by querying the $n$ -th nearest points using PostGIS $( n = 1 0$ ). As for validation and test datasets, we make sure the center points are all unknown during the training phase. Table 3 shows the evaluation results of different models for spatial context modeling. The baseline approaches (direct, tile, wrap, $r b f$ ) generally perform poorly in context modeling. We designed specialized version of these approaches (polar, polar_tile, scaled_rbf) with polar coordinates, which lead to significantly improvements. Note that these are models proposed by us specialized for context modeling and therefore are less general than the grid cell approaches. ",
|
| 925 |
+
"bbox": [
|
| 926 |
+
173,
|
| 927 |
+
530,
|
| 928 |
+
825,
|
| 929 |
+
642
|
| 930 |
+
],
|
| 931 |
+
"page_idx": 7
|
| 932 |
+
},
|
| 933 |
+
{
|
| 934 |
+
"type": "table",
|
| 935 |
+
"img_path": "images/948af1ee6726653363251fd159a4e39bd29f1db72c110f638aa1ec569de6cdd6.jpg",
|
| 936 |
+
"table_caption": [
|
| 937 |
+
"Table 3: The evaluation results of different spatial context models on the validation and test dataset. All encoders contains a 1 hidden layer FFN. All grid cell encoders set $\\lambda _ { m i n } { = } 1 0$ , $\\lambda _ { m a x } { = } 1 0 \\mathbf { k }$ . "
|
| 938 |
+
],
|
| 939 |
+
"table_footnote": [],
|
| 940 |
+
"table_body": "<table><tr><td rowspan=\"2\">Space2Vec</td><td rowspan=\"2\">Train NLL</td><td colspan=\"3\"></td><td colspan=\"2\">Testing</td></tr><tr><td>NLL</td><td>Validation MRR</td><td>HIT@5</td><td>MRR</td><td>HIT@5</td></tr><tr><td>none</td><td>1.163</td><td>1.297</td><td>0.159 (0.002)</td><td>22.4 (0.5)</td><td>0.167 (0.006)</td><td>23.4 (0.7)</td></tr><tr><td>direct</td><td>1.151</td><td>1.282</td><td>0.170 (0.002)</td><td>24.6 (0.4)</td><td>0.175 (0.003)</td><td>24.7 (0.5)</td></tr><tr><td>polar</td><td>1.157</td><td>1.283</td><td>0.176 (0.004)</td><td>25.4 (0.4)</td><td>0.178 (0.006)</td><td>24.9 (0.1)</td></tr><tr><td>tile(c = 50)</td><td>1.163</td><td>1.298</td><td>0.173 (0.004)</td><td>24.0 (0.6)</td><td>0.173 (0.001)</td><td>23.4 (0.1)</td></tr><tr><td>polar_tile(S =64)</td><td>1.161</td><td>1.282</td><td>0.173 (0.003)</td><td>25.0 (0.1)</td><td>0.177 (0.001)</td><td>24.5 (0.3)</td></tr><tr><td>wrap (h=2,0=512)</td><td>1.167</td><td>1.291</td><td>0.159 (0.001)</td><td>23.0 (0.1)</td><td>0.170 (0.001)</td><td>23.9 (0.2)</td></tr><tr><td>rbf (σ= 50)</td><td>1.160</td><td>1.281</td><td>0.179 (0.002)</td><td>25.2 (0.6)</td><td>0.172 (0.001)</td><td>25.0 (0.1)</td></tr><tr><td>scaled_rbf (o=40,β=0.1)</td><td>1.150</td><td>1.272</td><td>0.177 (0.002)</td><td>25.7 (0.1)</td><td>0.181 (0.001)</td><td>25.3 (0.1)</td></tr><tr><td>grid(Xmin=10)</td><td>1.172</td><td>1.285</td><td>0.178 (0.004)</td><td>24.9 (0.5)</td><td>0.181 (0.001)</td><td>25.1 (0.3)</td></tr><tr><td>hexa (Xmin=10)</td><td>1.156</td><td>1.289</td><td>0.173 (0.002)</td><td>24.0 (0.2)</td><td>0.183 (0.002)</td><td>25.3 (0.2)</td></tr><tr><td>theorydiag (Xmin = 10)</td><td>1.156</td><td>1.287</td><td>0.168 (0.001)</td><td>24.1 (0.4)</td><td>0.174 (0.005)</td><td>24.9 (0.1)</td></tr><tr><td>theory(Xmin=200)</td><td>1.168</td><td>1.295</td><td>0.159 (0.001)</td><td>23.1 (0.2)</td><td>0.170 (0.001)</td><td>23.2 (0.2)</td></tr><tr><td>theory(Xmin=50)</td><td>1.157</td><td>1.275</td><td>0.171 (0.001)</td><td>24.2 (0.3)</td><td>0.173 (0.001)</td><td>24.8 (0.4)</td></tr><tr><td>theory(Xmin=10)</td><td>1.158</td><td>1.280</td><td>0.177 (0.003)</td><td>25.2 (0.3)</td><td>0.185 (0.002)</td><td>25.7 (0.3)</td></tr></table>",
|
| 941 |
+
"bbox": [
|
| 942 |
+
178,
|
| 943 |
+
709,
|
| 944 |
+
818,
|
| 945 |
+
920
|
| 946 |
+
],
|
| 947 |
+
"page_idx": 7
|
| 948 |
+
},
|
| 949 |
+
{
|
| 950 |
+
"type": "image",
|
| 951 |
+
"img_path": "images/5fa30412df780e2bde0ae609aeaa14c71381c22820f850498f2fd03a727bedd7.jpg",
|
| 952 |
+
"image_caption": [
|
| 953 |
+
"Figure 3: Embedding clustering in the original space of (a) direct; (b) polar; (c) wrap, $h { = } 2 , o { = } 5 1 2$ ; (d) polar_tile, $S = 6 4$ , (e) scaled_ $. r b f$ , $\\sigma = 4 0$ , $\\beta { = } 0 . 1$ ; and (f) theory, $\\lambda _ { m i n } = 1 0$ , $\\lambda _ { m a x } = 1 0 k$ , $S = 6 4$ . $( \\mathrm { g } ) ( \\mathrm { h } ) ( \\mathrm { i } ) ( \\mathrm { j } ) ( \\mathrm { k } ) ( \\mathrm { l } )$ are the clustering results of the same models in the polar-distance space using $\\log ( \\Vert \\ \\Delta \\mathbf { x } _ { i j } \\ \\Vert + 1 )$ . All models use 1 hidden ReLU (except wrap) layers of 512 neurons. Most models except wrap can capture a shift when distance is around $e ^ { 5 } - 1 \\approx 1 5 0$ meters. "
|
| 954 |
+
],
|
| 955 |
+
"image_footnote": [],
|
| 956 |
+
"bbox": [
|
| 957 |
+
173,
|
| 958 |
+
98,
|
| 959 |
+
820,
|
| 960 |
+
286
|
| 961 |
+
],
|
| 962 |
+
"page_idx": 8
|
| 963 |
+
},
|
| 964 |
+
{
|
| 965 |
+
"type": "text",
|
| 966 |
+
"text": "Nevertheless the grid cell approaches are able to perform better than the specialized approaches on the test dataset while have competitive performance on validation dataset. See Appendix ?? for the visualization of context models. Actually the gains are small for all baseline approaches also. The reason is that we expect location encoding to be less important when context information is accessible. Similarly as discussed in (Gao et al., 2019), it is when there is a lack of visual clues that the grid cells of animals are the most helpful for their navigation. ",
|
| 967 |
+
"bbox": [
|
| 968 |
+
173,
|
| 969 |
+
391,
|
| 970 |
+
825,
|
| 971 |
+
476
|
| 972 |
+
],
|
| 973 |
+
"page_idx": 8
|
| 974 |
+
},
|
| 975 |
+
{
|
| 976 |
+
"type": "text",
|
| 977 |
+
"text": "Figure 6 shows the location embedding clustering results in both Cartesian and polar coordinate systems. We can see that direct (Fig. 3a, 3g) only captures the distance information when the context POI is very close $( l o g ( \\parallel \\Delta \\mathbf { x } _ { i j } \\parallel + 1 ) \\leqslant 5 $ ) while in the farther spatial context it purely models the direction information. polar (Fig. 3b, 3h) has the similar behaviors but captures the distance information in a more fine-grained manner. wrap (Fig. 3c, 3i) mainly focuses on differentiating relative positions in farther spatial context cont which might explain its lower performance6. polar_tile (Fig. 3d) mostly responds to distance information. Interestingly, scaled_rbf and theory have similar representations in the polar coordinate system (Fig. 3k, 3l) and similar performance (Table 3). While scaled_rbf captures the gradually decreased distance effect with a scaled kernel size which becomes larger in farther distance, theory achieves this by integrating representations of different scales. ",
|
| 978 |
+
"bbox": [
|
| 979 |
+
174,
|
| 980 |
+
482,
|
| 981 |
+
825,
|
| 982 |
+
621
|
| 983 |
+
],
|
| 984 |
+
"page_idx": 8
|
| 985 |
+
},
|
| 986 |
+
{
|
| 987 |
+
"type": "text",
|
| 988 |
+
"text": "5.2 FINE-GRAINED IMAGE CLASSIFICATION TASKS ",
|
| 989 |
+
"text_level": 1,
|
| 990 |
+
"bbox": [
|
| 991 |
+
174,
|
| 992 |
+
640,
|
| 993 |
+
544,
|
| 994 |
+
654
|
| 995 |
+
],
|
| 996 |
+
"page_idx": 8
|
| 997 |
+
},
|
| 998 |
+
{
|
| 999 |
+
"type": "text",
|
| 1000 |
+
"text": "To demonstrate the generalizability of Space2Vec for space representation we utilized the proposed point space encoder $E n c ^ { ( x ) } ( \\ v r )$ model in a well-known computer vision task: fine-grained image classification. As we discussed in Section 3, many studies (Berg et al., 2014; Chu et al., 2019; Mac Aodha et al., 2019) have shown that geographic prior information - where (and when) the image is taken - is very important additional information for the fine-grained image classification task and can substantially improve the model performance. For example, the appearance information is usually not sufficient to differentiate two visually similar species. In this case, the geographic prior becomes much more important because these two species may have very different spatial prior distributions such as the example of European Toads and Spiny Toads in Figure 1 of Mac Aodha et al. (2019). ",
|
| 1001 |
+
"bbox": [
|
| 1002 |
+
173,
|
| 1003 |
+
666,
|
| 1004 |
+
825,
|
| 1005 |
+
792
|
| 1006 |
+
],
|
| 1007 |
+
"page_idx": 8
|
| 1008 |
+
},
|
| 1009 |
+
{
|
| 1010 |
+
"type": "text",
|
| 1011 |
+
"text": "We adopt the task setup of Mac Aodha et al. (2019). During training we have a set of tuples $D =$ $\\{ ( I _ { i } , \\mathbf { x } _ { i } , y _ { i } , p _ { i } ) \\mid i = 1 , . . . , N \\}$ where $I _ { i }$ indicates an image, $y _ { i } \\in \\{ 1 , 2 , . . . , C \\}$ is the corresponding class label (species category), $\\mathbf { x } _ { i } = \\left[ l o n g i t u d e _ { i } , l a t i t u d e _ { i } \\right]$ is the geographic coordinates where the image was taken, and $p _ { i }$ is the id of the photographer who took this image. At training time, a location encoder is trained to capture the spatial prior information $P ( y \\mid \\mathbf { x } )$ . At inference time, $p _ { i }$ information is not available and the final image classification prediction is calculated based on the combination of two models: 1) the trained location encoder which captures the spatial priors $P ( y \\mid \\mathbf { x } )$ and 2) the pretrained image classification model, InceptionV3 network (Szegedy et al., 2016), which captures ${ \\bar { P } } ( y \\mid I )$ . Bayesian theory has been used to derive the joint distribution $P ( y \\mid I , \\mathbf { x } )$ . See Mac Aodha et al. (2019) for detail explanation as well as the loss function. Note that while Space2Vec outperforms specialized density estimation methods such as Adaptive Kernel (Berg et al., 2014), it would be interesting to explore early fusion Space2Vec ’s representations with the image module. ",
|
| 1012 |
+
"bbox": [
|
| 1013 |
+
174,
|
| 1014 |
+
799,
|
| 1015 |
+
825,
|
| 1016 |
+
885
|
| 1017 |
+
],
|
| 1018 |
+
"page_idx": 8
|
| 1019 |
+
},
|
| 1020 |
+
{
|
| 1021 |
+
"type": "table",
|
| 1022 |
+
"img_path": "images/1a57346574c5190339ff62cd988dc647115dbfb4eddf8a8daefc93c167e3c65a.jpg",
|
| 1023 |
+
"table_caption": [
|
| 1024 |
+
"Table 4: Fine-grained image classification results on two datasets: BirdSnap: and NABirds:. The classification accuracy is calculated by combining image classification predictions $P ( y \\mid I )$ with different spatial priors $P ( y \\mid \\mathbf { x } )$ . The grid and theory model use 1 hidden ReLU layers of 512 neurons. The evaluation results of the baseline models are from Table 1 of Mac Aodha et al. (2019). "
|
| 1025 |
+
],
|
| 1026 |
+
"table_footnote": [],
|
| 1027 |
+
"table_body": "<table><tr><td></td><td>BirdSnapt</td><td>NABirdst</td></tr><tr><td>No Prior (i.e.uniform)</td><td>70.07</td><td>76.08</td></tr><tr><td>Nearest Neighbor (num)</td><td>77.76</td><td>79.99</td></tr><tr><td>Nearest Neighbor (spatial)</td><td>77.98</td><td>80.79</td></tr><tr><td>Adaptive Kernel (Berg et al., 2014)</td><td>78.65</td><td>81.11</td></tr><tr><td>tile (Tang et al.,2015) (location only)</td><td>77.19</td><td>79.58</td></tr><tr><td>wrap (Mac Aodha et al.,2019) (location only)</td><td>78.65</td><td>81.15</td></tr><tr><td>rbf (σ=1k)</td><td>78.56</td><td>81.13</td></tr><tr><td>grid (入min=0.0001,入max=360,S= 64)</td><td>79.44</td><td>81.28</td></tr><tr><td>theory (入min=0.0001, λmax=360,S=64)</td><td>79.35</td><td>81.59</td></tr></table>",
|
| 1028 |
+
"bbox": [
|
| 1029 |
+
264,
|
| 1030 |
+
169,
|
| 1031 |
+
725,
|
| 1032 |
+
306
|
| 1033 |
+
],
|
| 1034 |
+
"page_idx": 9
|
| 1035 |
+
},
|
| 1036 |
+
{
|
| 1037 |
+
"type": "text",
|
| 1038 |
+
"text": "",
|
| 1039 |
+
"bbox": [
|
| 1040 |
+
174,
|
| 1041 |
+
332,
|
| 1042 |
+
825,
|
| 1043 |
+
416
|
| 1044 |
+
],
|
| 1045 |
+
"page_idx": 9
|
| 1046 |
+
},
|
| 1047 |
+
{
|
| 1048 |
+
"type": "text",
|
| 1049 |
+
"text": "We use two versions of our point space encoder $E n c ^ { ( x ) } ( \\ v r )$ model (grid, theory) as the location encoder to capture the spatial prior information $P ( y \\mid \\mathbf { x } )$ . The evaluation results of our models as well as multiple baselines are shown in Table 4. We can see that both grid, theory outperform previous models as well as that of Mac Aodha et al. (2019) on two fine-grained image classification datasets with significant sizes: BirdSnap:, NABirds:. theory shows superiority over grid on NABirds: while fail to outperform grid on BirdSnap:. Note that we only pick baseline models which capture spatial-only prior and drop models which additionally consider time information. Both grid and theory use 1 hidden ReLU layers of 512 neurons for $\\mathbf { N N } ( )$ and they have the same hyperparameters: $\\lambda _ { m i n } { = } 0 . 0 0 0 1$ , $\\lambda _ { m a x } { = } 3 6 0$ , $S = 6 4$ . Like Mac Aodha et al. (2019), the location embedding size $d ^ { ( x ) }$ is 1024 and we train the location encoder for 30 epochs. Our implementation is based on the original code7 of Mac Aodha et al. (2019) for both model training and evaluation phase. ",
|
| 1050 |
+
"bbox": [
|
| 1051 |
+
173,
|
| 1052 |
+
422,
|
| 1053 |
+
825,
|
| 1054 |
+
578
|
| 1055 |
+
],
|
| 1056 |
+
"page_idx": 9
|
| 1057 |
+
},
|
| 1058 |
+
{
|
| 1059 |
+
"type": "text",
|
| 1060 |
+
"text": "6 CONCLUSION ",
|
| 1061 |
+
"text_level": 1,
|
| 1062 |
+
"bbox": [
|
| 1063 |
+
176,
|
| 1064 |
+
602,
|
| 1065 |
+
318,
|
| 1066 |
+
617
|
| 1067 |
+
],
|
| 1068 |
+
"page_idx": 9
|
| 1069 |
+
},
|
| 1070 |
+
{
|
| 1071 |
+
"type": "text",
|
| 1072 |
+
"text": "We introduced an encoder-decoder framework as a general-purpose representation model for space inspired by biological grid cells’ multi-scale periodic representations. The model is an inductive learning model and can be trained in an unsupervised manner. We conduct two experiments on POI type prediction based on 1) POI locations and 2) nearby POIs. The evaluation results demonstrate the effectiveness of our model. Our analysis reveals that it is the ability to integrate representations of different scales that makes the grid cell models outperform other baselines on these two tasks. In the future, we hope to incorporate the presented framework to more complex GIS tasks such as social network analysis, and sea surface temperature prediction. ",
|
| 1073 |
+
"bbox": [
|
| 1074 |
+
174,
|
| 1075 |
+
632,
|
| 1076 |
+
825,
|
| 1077 |
+
743
|
| 1078 |
+
],
|
| 1079 |
+
"page_idx": 9
|
| 1080 |
+
},
|
| 1081 |
+
{
|
| 1082 |
+
"type": "text",
|
| 1083 |
+
"text": "ACKNOWLEDGMENTS ",
|
| 1084 |
+
"text_level": 1,
|
| 1085 |
+
"bbox": [
|
| 1086 |
+
176,
|
| 1087 |
+
771,
|
| 1088 |
+
326,
|
| 1089 |
+
784
|
| 1090 |
+
],
|
| 1091 |
+
"page_idx": 9
|
| 1092 |
+
},
|
| 1093 |
+
{
|
| 1094 |
+
"type": "text",
|
| 1095 |
+
"text": "The presented work is partially funded by the NSF award 1936677 C-Accel Pilot - Track A1 (Open Knowledge Network): Spatially-Explicit Models, Methods, And Services For Open Knowledge Networks, Esri Inc., and Microsoft AI for Earth Grant: Deep Species Spatio-temporal Distribution Modeling for Biodiversity Hotspot Prediction. We thank Dr. Ruiqi Gao for discussions about grid cells, Dr. Wenyun Zuo for discussion about species potential distribution prediction and Dr. Yingjie Hu for his suggestions about the introduction section. ",
|
| 1096 |
+
"bbox": [
|
| 1097 |
+
174,
|
| 1098 |
+
797,
|
| 1099 |
+
825,
|
| 1100 |
+
881
|
| 1101 |
+
],
|
| 1102 |
+
"page_idx": 9
|
| 1103 |
+
},
|
| 1104 |
+
{
|
| 1105 |
+
"type": "text",
|
| 1106 |
+
"text": "REFERENCES ",
|
| 1107 |
+
"text_level": 1,
|
| 1108 |
+
"bbox": [
|
| 1109 |
+
176,
|
| 1110 |
+
102,
|
| 1111 |
+
287,
|
| 1112 |
+
118
|
| 1113 |
+
],
|
| 1114 |
+
"page_idx": 10
|
| 1115 |
+
},
|
| 1116 |
+
{
|
| 1117 |
+
"type": "text",
|
| 1118 |
+
"text": "Alison Abbott and Ewen Callaway. Nobel prize for decoding brain’s sense of place. Nature News, 514(7521):153, 2014. ",
|
| 1119 |
+
"bbox": [
|
| 1120 |
+
173,
|
| 1121 |
+
126,
|
| 1122 |
+
823,
|
| 1123 |
+
154
|
| 1124 |
+
],
|
| 1125 |
+
"page_idx": 10
|
| 1126 |
+
},
|
| 1127 |
+
{
|
| 1128 |
+
"type": "text",
|
| 1129 |
+
"text": "Benjamin Adams, Grant McKenzie, and Mark Gahegan. Frankenplace: interactive thematic mapping for ad hoc exploratory search. In Proceedings of the 24th international conference on world wide web, pp. 12–22. International World Wide Web Conferences Steering Committee, 2015. ",
|
| 1130 |
+
"bbox": [
|
| 1131 |
+
173,
|
| 1132 |
+
161,
|
| 1133 |
+
825,
|
| 1134 |
+
205
|
| 1135 |
+
],
|
| 1136 |
+
"page_idx": 10
|
| 1137 |
+
},
|
| 1138 |
+
{
|
| 1139 |
+
"type": "text",
|
| 1140 |
+
"text": "Andrea Banino, Caswell Barry, Benigno Uria, Charles Blundell, Timothy Lillicrap, Piotr Mirowski, Alexander Pritzel, Martin J Chadwick, Thomas Degris, Joseph Modayil, et al. Vector-based navigation using grid-like representations in artificial agents. Nature, 557(7705):429, 2018. ",
|
| 1141 |
+
"bbox": [
|
| 1142 |
+
174,
|
| 1143 |
+
213,
|
| 1144 |
+
825,
|
| 1145 |
+
256
|
| 1146 |
+
],
|
| 1147 |
+
"page_idx": 10
|
| 1148 |
+
},
|
| 1149 |
+
{
|
| 1150 |
+
"type": "text",
|
| 1151 |
+
"text": "G Baudat and F Anouar. Kernel-based methods and function approximation. volume 2, pp. 1244 – 1249 vol.2, 02 2001. ISBN 0-7803-7044-9. doi: 10.1109/IJCNN.2001.939539. ",
|
| 1152 |
+
"bbox": [
|
| 1153 |
+
173,
|
| 1154 |
+
265,
|
| 1155 |
+
825,
|
| 1156 |
+
294
|
| 1157 |
+
],
|
| 1158 |
+
"page_idx": 10
|
| 1159 |
+
},
|
| 1160 |
+
{
|
| 1161 |
+
"type": "text",
|
| 1162 |
+
"text": "Thomas Berg, Jiongxin Liu, Seung Woo Lee, Michelle L Alexander, David W Jacobs, and Peter N Belhumeur. Birdsnap: Large-scale fine-grained visual categorization of birds. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2011–2018, 2014. ",
|
| 1163 |
+
"bbox": [
|
| 1164 |
+
176,
|
| 1165 |
+
301,
|
| 1166 |
+
825,
|
| 1167 |
+
344
|
| 1168 |
+
],
|
| 1169 |
+
"page_idx": 10
|
| 1170 |
+
},
|
| 1171 |
+
{
|
| 1172 |
+
"type": "text",
|
| 1173 |
+
"text": "Herman J. Bierens. The nadaraya–watson kernel regression function estimator. Topics in Advanced Econometrics, 16:212–247, 1994. ",
|
| 1174 |
+
"bbox": [
|
| 1175 |
+
169,
|
| 1176 |
+
352,
|
| 1177 |
+
823,
|
| 1178 |
+
381
|
| 1179 |
+
],
|
| 1180 |
+
"page_idx": 10
|
| 1181 |
+
},
|
| 1182 |
+
{
|
| 1183 |
+
"type": "text",
|
| 1184 |
+
"text": "Hugh T Blair, Adam C Welday, and Kechen Zhang. Scale-invariant memory representations emerge from moire interference between grid fields that produce theta oscillations: a computational model. Journal of Neuroscience, 27(12):3211–3229, 2007. ",
|
| 1185 |
+
"bbox": [
|
| 1186 |
+
173,
|
| 1187 |
+
390,
|
| 1188 |
+
823,
|
| 1189 |
+
433
|
| 1190 |
+
],
|
| 1191 |
+
"page_idx": 10
|
| 1192 |
+
},
|
| 1193 |
+
{
|
| 1194 |
+
"type": "text",
|
| 1195 |
+
"text": "Grace Chu, Brian Potetz, Weijun Wang, Andrew Howard, Yang Song, Fernando Brucher, Thomas Leung, and Hartwig Adam. Geo-aware networks for fine grained recognition. arXiv preprint arXiv:1906.01737, 2019. ",
|
| 1196 |
+
"bbox": [
|
| 1197 |
+
176,
|
| 1198 |
+
440,
|
| 1199 |
+
825,
|
| 1200 |
+
483
|
| 1201 |
+
],
|
| 1202 |
+
"page_idx": 10
|
| 1203 |
+
},
|
| 1204 |
+
{
|
| 1205 |
+
"type": "text",
|
| 1206 |
+
"text": "Christopher J Cueva and Xue-Xin Wei. Emergence of grid-like representations by training recurrent neural networks to perform spatial localization. arXiv preprint arXiv:1803.07770, 2018. ",
|
| 1207 |
+
"bbox": [
|
| 1208 |
+
173,
|
| 1209 |
+
491,
|
| 1210 |
+
823,
|
| 1211 |
+
521
|
| 1212 |
+
],
|
| 1213 |
+
"page_idx": 10
|
| 1214 |
+
},
|
| 1215 |
+
{
|
| 1216 |
+
"type": "text",
|
| 1217 |
+
"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. ",
|
| 1218 |
+
"bbox": [
|
| 1219 |
+
171,
|
| 1220 |
+
529,
|
| 1221 |
+
821,
|
| 1222 |
+
559
|
| 1223 |
+
],
|
| 1224 |
+
"page_idx": 10
|
| 1225 |
+
},
|
| 1226 |
+
{
|
| 1227 |
+
"type": "text",
|
| 1228 |
+
"text": "A Stewart Fotheringham and David WS Wong. The modifiable areal unit problem in multivariate statistical analysis. Environment and planning A, 23(7):1025–1044, 1991. ",
|
| 1229 |
+
"bbox": [
|
| 1230 |
+
171,
|
| 1231 |
+
565,
|
| 1232 |
+
823,
|
| 1233 |
+
595
|
| 1234 |
+
],
|
| 1235 |
+
"page_idx": 10
|
| 1236 |
+
},
|
| 1237 |
+
{
|
| 1238 |
+
"type": "text",
|
| 1239 |
+
"text": "Ruiqi Gao, Jianwen Xie, Song-Chun Zhu, and Ying Nian Wu. Learning grid cells as vector representation of self-position coupled with matrix representation of self-motion. In Proceedings of ICLR 2019, 2019. ",
|
| 1240 |
+
"bbox": [
|
| 1241 |
+
173,
|
| 1242 |
+
603,
|
| 1243 |
+
823,
|
| 1244 |
+
645
|
| 1245 |
+
],
|
| 1246 |
+
"page_idx": 10
|
| 1247 |
+
},
|
| 1248 |
+
{
|
| 1249 |
+
"type": "text",
|
| 1250 |
+
"text": "Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii Shelestov. Deep learning classification of land cover and crop types using remote sensing data. IEEE Geoscience and Remote Sensing Letters, 14(5):778–782, 2017. ",
|
| 1251 |
+
"bbox": [
|
| 1252 |
+
173,
|
| 1253 |
+
654,
|
| 1254 |
+
823,
|
| 1255 |
+
696
|
| 1256 |
+
],
|
| 1257 |
+
"page_idx": 10
|
| 1258 |
+
},
|
| 1259 |
+
{
|
| 1260 |
+
"type": "text",
|
| 1261 |
+
"text": "Alex M Lechner, William T Langford, Simon D Jones, Sarah A Bekessy, and Ascelin Gordon. Investigating species–environment relationships at multiple scales: Differentiating between intrinsic scale and the modifiable areal unit problem. Ecological Complexity, 11:91–102, 2012. ",
|
| 1262 |
+
"bbox": [
|
| 1263 |
+
174,
|
| 1264 |
+
705,
|
| 1265 |
+
825,
|
| 1266 |
+
748
|
| 1267 |
+
],
|
| 1268 |
+
"page_idx": 10
|
| 1269 |
+
},
|
| 1270 |
+
{
|
| 1271 |
+
"type": "text",
|
| 1272 |
+
"text": "Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926, 2017. ",
|
| 1273 |
+
"bbox": [
|
| 1274 |
+
171,
|
| 1275 |
+
756,
|
| 1276 |
+
823,
|
| 1277 |
+
785
|
| 1278 |
+
],
|
| 1279 |
+
"page_idx": 10
|
| 1280 |
+
},
|
| 1281 |
+
{
|
| 1282 |
+
"type": "text",
|
| 1283 |
+
"text": "Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained image classification. arXiv preprint arXiv:1906.05272, 2019. ",
|
| 1284 |
+
"bbox": [
|
| 1285 |
+
171,
|
| 1286 |
+
792,
|
| 1287 |
+
825,
|
| 1288 |
+
821
|
| 1289 |
+
],
|
| 1290 |
+
"page_idx": 10
|
| 1291 |
+
},
|
| 1292 |
+
{
|
| 1293 |
+
"type": "text",
|
| 1294 |
+
"text": "Gengchen Mai, Krzysztof Janowicz, Bo Yan, Rui Zhu, Ling Cai, and Ni Lao. Contextual graph attention for answering logical queries over incomplete knowledge graphs. In Proceedings of the 10th International Conference on Knowledge Capture, pp. 171–178, 2019a. ",
|
| 1295 |
+
"bbox": [
|
| 1296 |
+
176,
|
| 1297 |
+
829,
|
| 1298 |
+
823,
|
| 1299 |
+
873
|
| 1300 |
+
],
|
| 1301 |
+
"page_idx": 10
|
| 1302 |
+
},
|
| 1303 |
+
{
|
| 1304 |
+
"type": "text",
|
| 1305 |
+
"text": "Gengchen Mai, Bo Yan, Krzysztof Janowicz, and Rui Zhu. Relaxing unanswerable geographic questions using a spatially explicit knowledge graph embedding model. In AGILE: The 22nd Annual International Conference on Geographic Information Science, pp. 21–39. Springer, 2019b. ",
|
| 1306 |
+
"bbox": [
|
| 1307 |
+
174,
|
| 1308 |
+
881,
|
| 1309 |
+
825,
|
| 1310 |
+
924
|
| 1311 |
+
],
|
| 1312 |
+
"page_idx": 10
|
| 1313 |
+
},
|
| 1314 |
+
{
|
| 1315 |
+
"type": "text",
|
| 1316 |
+
"text": "V Maz’ya and G Schmidt. On approximate approximations using gaussian kernels. IMA Journal of Numerical Analysis, 16:13–29, 01 1996. ",
|
| 1317 |
+
"bbox": [
|
| 1318 |
+
173,
|
| 1319 |
+
103,
|
| 1320 |
+
825,
|
| 1321 |
+
132
|
| 1322 |
+
],
|
| 1323 |
+
"page_idx": 11
|
| 1324 |
+
},
|
| 1325 |
+
{
|
| 1326 |
+
"type": "text",
|
| 1327 |
+
"text": "Grant McKenzie, Krzysztof Janowicz, Song Gao, Jiue-An Yang, and Yingjie Hu. Poi pulse: A multi-granular, semantic signature–based information observatory for the interactive visualization of big geosocial data. Cartographica: The International Journal for Geographic Information and Geovisualization, 50(2):71–85, 2015. ",
|
| 1328 |
+
"bbox": [
|
| 1329 |
+
174,
|
| 1330 |
+
140,
|
| 1331 |
+
826,
|
| 1332 |
+
196
|
| 1333 |
+
],
|
| 1334 |
+
"page_idx": 11
|
| 1335 |
+
},
|
| 1336 |
+
{
|
| 1337 |
+
"type": "text",
|
| 1338 |
+
"text": "Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pp. 3111–3119, 2013. ",
|
| 1339 |
+
"bbox": [
|
| 1340 |
+
174,
|
| 1341 |
+
207,
|
| 1342 |
+
825,
|
| 1343 |
+
250
|
| 1344 |
+
],
|
| 1345 |
+
"page_idx": 11
|
| 1346 |
+
},
|
| 1347 |
+
{
|
| 1348 |
+
"type": "text",
|
| 1349 |
+
"text": "Justin Moat, Steven P Bachman, Richard Field, and Doreen S Boyd. Refining area of occupancy to address the modifiable areal unit problem in ecology and conservation. Conservation biology, 32 (6):1278–1289, 2018. ",
|
| 1350 |
+
"bbox": [
|
| 1351 |
+
173,
|
| 1352 |
+
257,
|
| 1353 |
+
826,
|
| 1354 |
+
301
|
| 1355 |
+
],
|
| 1356 |
+
"page_idx": 11
|
| 1357 |
+
},
|
| 1358 |
+
{
|
| 1359 |
+
"type": "text",
|
| 1360 |
+
"text": "Stan Openshaw. The modifiable areal unit problem. Concepts and techniques in modern geography, 1984. ",
|
| 1361 |
+
"bbox": [
|
| 1362 |
+
173,
|
| 1363 |
+
309,
|
| 1364 |
+
825,
|
| 1365 |
+
338
|
| 1366 |
+
],
|
| 1367 |
+
"page_idx": 11
|
| 1368 |
+
},
|
| 1369 |
+
{
|
| 1370 |
+
"type": "text",
|
| 1371 |
+
"text": "Jeffrey Pennington, Richard Socher, and Christopher Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pp. 1532–1543, 2014. ",
|
| 1372 |
+
"bbox": [
|
| 1373 |
+
174,
|
| 1374 |
+
347,
|
| 1375 |
+
826,
|
| 1376 |
+
390
|
| 1377 |
+
],
|
| 1378 |
+
"page_idx": 11
|
| 1379 |
+
},
|
| 1380 |
+
{
|
| 1381 |
+
"type": "text",
|
| 1382 |
+
"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. ",
|
| 1383 |
+
"bbox": [
|
| 1384 |
+
173,
|
| 1385 |
+
398,
|
| 1386 |
+
826,
|
| 1387 |
+
440
|
| 1388 |
+
],
|
| 1389 |
+
"page_idx": 11
|
| 1390 |
+
},
|
| 1391 |
+
{
|
| 1392 |
+
"type": "text",
|
| 1393 |
+
"text": "Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 652–660, 2017. ",
|
| 1394 |
+
"bbox": [
|
| 1395 |
+
173,
|
| 1396 |
+
450,
|
| 1397 |
+
823,
|
| 1398 |
+
493
|
| 1399 |
+
],
|
| 1400 |
+
"page_idx": 11
|
| 1401 |
+
},
|
| 1402 |
+
{
|
| 1403 |
+
"type": "text",
|
| 1404 |
+
"text": "Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2818–2826, 2016. ",
|
| 1405 |
+
"bbox": [
|
| 1406 |
+
173,
|
| 1407 |
+
501,
|
| 1408 |
+
823,
|
| 1409 |
+
545
|
| 1410 |
+
],
|
| 1411 |
+
"page_idx": 11
|
| 1412 |
+
},
|
| 1413 |
+
{
|
| 1414 |
+
"type": "text",
|
| 1415 |
+
"text": "Kevin Tang, Manohar Paluri, Li Fei-Fei, Rob Fergus, and Lubomir Bourdev. Improving image classification with location context. In Proceedings of the IEEE international conference on computer vision, pp. 1008–1016, 2015. ",
|
| 1416 |
+
"bbox": [
|
| 1417 |
+
173,
|
| 1418 |
+
553,
|
| 1419 |
+
823,
|
| 1420 |
+
597
|
| 1421 |
+
],
|
| 1422 |
+
"page_idx": 11
|
| 1423 |
+
},
|
| 1424 |
+
{
|
| 1425 |
+
"type": "text",
|
| 1426 |
+
"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. ",
|
| 1427 |
+
"bbox": [
|
| 1428 |
+
174,
|
| 1429 |
+
604,
|
| 1430 |
+
825,
|
| 1431 |
+
648
|
| 1432 |
+
],
|
| 1433 |
+
"page_idx": 11
|
| 1434 |
+
},
|
| 1435 |
+
{
|
| 1436 |
+
"type": "text",
|
| 1437 |
+
"text": "Petar Velickovi ˇ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua ´ Bengio. Graph attention networks. In ICLR 2018, 2018. ",
|
| 1438 |
+
"bbox": [
|
| 1439 |
+
171,
|
| 1440 |
+
656,
|
| 1441 |
+
823,
|
| 1442 |
+
685
|
| 1443 |
+
],
|
| 1444 |
+
"page_idx": 11
|
| 1445 |
+
},
|
| 1446 |
+
{
|
| 1447 |
+
"type": "text",
|
| 1448 |
+
"text": "Bo Yan, Krzysztof Janowicz, Gengchen Mai, and Song Gao. From itdl to place2vec: Reasoning about place type similarity and relatedness by learning embeddings from augmented spatial contexts. In Proceedings of the 25th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems, pp. 35. ACM, 2017. ",
|
| 1449 |
+
"bbox": [
|
| 1450 |
+
173,
|
| 1451 |
+
694,
|
| 1452 |
+
826,
|
| 1453 |
+
751
|
| 1454 |
+
],
|
| 1455 |
+
"page_idx": 11
|
| 1456 |
+
},
|
| 1457 |
+
{
|
| 1458 |
+
"type": "text",
|
| 1459 |
+
"text": "Bo Yan, Krzysztof Janowicz, Gengchen Mai, and Rui Zhu. xnet+ sc: Classifying places based on images by incorporating spatial contexts. In 10th International Conference on Geographic Information Science (GIScience 2018). Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, 2018. ",
|
| 1460 |
+
"bbox": [
|
| 1461 |
+
174,
|
| 1462 |
+
760,
|
| 1463 |
+
826,
|
| 1464 |
+
803
|
| 1465 |
+
],
|
| 1466 |
+
"page_idx": 11
|
| 1467 |
+
},
|
| 1468 |
+
{
|
| 1469 |
+
"type": "text",
|
| 1470 |
+
"text": "Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan R Salakhutdinov, and Alexander J Smola. Deep sets. In Advances in neural information processing systems, pp. 3391–3401, 2017. ",
|
| 1471 |
+
"bbox": [
|
| 1472 |
+
174,
|
| 1473 |
+
811,
|
| 1474 |
+
826,
|
| 1475 |
+
854
|
| 1476 |
+
],
|
| 1477 |
+
"page_idx": 11
|
| 1478 |
+
},
|
| 1479 |
+
{
|
| 1480 |
+
"type": "text",
|
| 1481 |
+
"text": "Wenyun Zuo, Ni Lao, Yuying Geng, and Keping Ma. Geosvm: an efficient and effective tool to predict species’ potential distributions. Journal of Plant Ecology, 1(2):143–145, 2008. ",
|
| 1482 |
+
"bbox": [
|
| 1483 |
+
176,
|
| 1484 |
+
863,
|
| 1485 |
+
826,
|
| 1486 |
+
892
|
| 1487 |
+
],
|
| 1488 |
+
"page_idx": 11
|
| 1489 |
+
},
|
| 1490 |
+
{
|
| 1491 |
+
"type": "text",
|
| 1492 |
+
"text": "A APPENDIX ",
|
| 1493 |
+
"text_level": 1,
|
| 1494 |
+
"bbox": [
|
| 1495 |
+
176,
|
| 1496 |
+
102,
|
| 1497 |
+
299,
|
| 1498 |
+
118
|
| 1499 |
+
],
|
| 1500 |
+
"page_idx": 12
|
| 1501 |
+
},
|
| 1502 |
+
{
|
| 1503 |
+
"type": "text",
|
| 1504 |
+
"text": "A.1 BASELINES",
|
| 1505 |
+
"text_level": 1,
|
| 1506 |
+
"bbox": [
|
| 1507 |
+
176,
|
| 1508 |
+
133,
|
| 1509 |
+
302,
|
| 1510 |
+
148
|
| 1511 |
+
],
|
| 1512 |
+
"page_idx": 12
|
| 1513 |
+
},
|
| 1514 |
+
{
|
| 1515 |
+
"type": "text",
|
| 1516 |
+
"text": "To help understand the mechanism of distributed space representation we compare multiple ways of encoding spatial information. Different models use different point space encoder $E n c ^ { ( x ) } ( \\ v r )$ to encode either location $\\mathbf { x } _ { i }$ (for location modeling loc) or the displacement between the center point and one context point $\\Delta { \\bf x } _ { i j } = { \\bf x } _ { i } - { \\bf x } _ { i j }$ (for spatial context modeling cont)8. ",
|
| 1517 |
+
"bbox": [
|
| 1518 |
+
174,
|
| 1519 |
+
160,
|
| 1520 |
+
825,
|
| 1521 |
+
217
|
| 1522 |
+
],
|
| 1523 |
+
"page_idx": 12
|
| 1524 |
+
},
|
| 1525 |
+
{
|
| 1526 |
+
"type": "text",
|
| 1527 |
+
"text": "• random shuffles the order of the correct POI and $N$ negative samples randomly as the predicted ranking. This shows the lower bound of each metrics. \n• direct directly encode location $\\mathbf { x } _ { i }$ (or $\\Delta { \\bf x } _ { i j }$ for cont) into a location embedding $\\mathbf { e } [ \\mathbf { x } _ { i } ]$ (or $\\mathbf { e } [ \\Delta \\mathbf { x } _ { i j } ] )$ using a feed-forward neural networks $\\left( \\mathrm { F F N s } \\right) ^ { 9 }$ , denoted as $E n c _ { d i r e c t } ^ { ( x ) } ( { \\bf x } )$ without decomposing coordinates into a multi-scale periodic representation. This is essentially the GPS encoding method used by Chu et al. (2019). Note that Chu et al. (2019) is not open sourced and we end up implementing the model architecture ourselves. \n• tile divides the study area $A _ { l o c }$ (for loc) or the range of spatial context defined by $\\lambda _ { m a x }$ , $A _ { c o n t }$ , (for cont) into grids with equal grid sizes $c$ . Each grid has an embedding to be used as the encoding for every location $\\mathbf { x } _ { i }$ or displacement $\\Delta { \\bf x } _ { i j }$ fall into this grid. This is a common practice by many previous work when dealing with coordinate data (Berg et al., 2014; Adams et al., 2015; Tang et al., 2015). wrap is a location encoder model recently introduced by Mac Aodha et al. (2019). It first normalizes $\\mathbf { x }$ (or $\\Delta \\mathbf { x } )$ ) into the range $[ - 1 , 1 ]$ and uses a coordinate wrap mechanism $[ \\sin ( \\pi \\mathbf { x } ^ { [ l ] } ) ; \\cos ( \\pi \\mathbf { x } ^ { [ l ] } ) ]$ to convert each dimension of $\\mathbf { x }$ into 2 numbers. This is then passed through an initial fully connected layer, followed by a series of $h$ residual blocks, each consisting of two fully connected layers $\\scriptstyle { \\dot { } } _ { o }$ hidden neurons) with a dropout layer in between. We adopt the official code of Mac Aodha et al. $( 2 0 1 9 ) ^ { 1 0 }$ for this implementation. \n• $r b f$ randomly samples $M$ points from the training dataset as RBF anchor points $\\{ \\mathbf { x } _ { m } ^ { a n c h o r } , m \\stackrel { \\cdot } { = } 1 . . . \\bar { M } \\}$ (or samples $M \\Delta \\mathbf { x } _ { m } ^ { a n c h o r }$ from $A _ { c o n t }$ for cont) 11, and use gaussian kernels $\\exp \\big ( - \\frac { \\parallel \\mathbf { x } _ { i } - \\mathbf { x } _ { m } ^ { a n c h o r } \\parallel ^ { 2 } } { 2 \\sigma ^ { 2 } } \\big )$ (or $\\exp \\big ( - \\frac { \\parallel \\Delta \\mathbf { x } _ { i j } - \\Delta \\mathbf { x } _ { m } ^ { a n c h o r } \\parallel ^ { 2 } } { 2 \\sigma ^ { 2 } } \\big )$ for cont) on each anchor points, where $\\sigma$ is the kernel size. Each point $p _ { i }$ has a $M$ -dimension RBF feature vector which is fed into a FNN to obtain the spatial embedding. This is a strong baseline for representing floating number features in machine learning models. \n• grid as described in Section 4.1 inspired by the position encoding in Transformer (Vaswani et al., 2017). \n• hexa Same as grid but use $s i n ( \\theta ) , s i n ( \\theta + 2 \\pi / 3 )$ , and $s i n ( \\theta + 4 \\pi / 3 )$ in $P E _ { s , l } ^ { ( g ) } ( \\mathbf { x } )$ . \n• theory as described in Section 4.1, uses the theoretical models (Gao et al., 2019) as the first layer of $E n c _ { t h e o r y } ^ { ( x ) } ( \\mathbf { x } )$ or $E n c _ { t h e o r y } ^ { ( x ) } ( \\Delta \\mathbf { x } _ { i j } )$ . \n• theorydiag further constrains $\\mathbf { N N } ( )$ as a block diagonal matrix, with each scale as a block. ",
|
| 1528 |
+
"bbox": [
|
| 1529 |
+
214,
|
| 1530 |
+
224,
|
| 1531 |
+
826,
|
| 1532 |
+
724
|
| 1533 |
+
],
|
| 1534 |
+
"page_idx": 12
|
| 1535 |
+
},
|
| 1536 |
+
{
|
| 1537 |
+
"type": "text",
|
| 1538 |
+
"text": "We also have the following baselines which are specific to the spatial context modeling task. ",
|
| 1539 |
+
"bbox": [
|
| 1540 |
+
178,
|
| 1541 |
+
728,
|
| 1542 |
+
776,
|
| 1543 |
+
743
|
| 1544 |
+
],
|
| 1545 |
+
"page_idx": 12
|
| 1546 |
+
},
|
| 1547 |
+
{
|
| 1548 |
+
"type": "text",
|
| 1549 |
+
"text": "• none the decoder $D e c _ { c } ( )$ does not consider the spatial relationship between the center point and context points but only the co-locate patterns such as Place2Vec (Yan et al., 2017). That means we drop the $\\mathbf { e } [ \\Delta \\mathbf { x } _ { i j } ]$ from the attention mechanism in Equ. 7 and 8. polar first converts the displacement $\\Delta \\mathbf { x } _ { i j }$ into polar coordinates $( r , \\theta )$ centered at the center point where $r = l o g ( \\parallel \\Delta \\mathbf { x } _ { i j } \\parallel + 1 ) ^ { - }$ . Then it uses $[ r , \\theta ]$ as the input for a FFN to obtain the spatial relationship embedding in Equ. 7. We find out that it has a significant performance improvement over the variation with $r = \\parallel \\Delta \\mathbf { x } _ { i j } \\parallel$ . ",
|
| 1550 |
+
"bbox": [
|
| 1551 |
+
214,
|
| 1552 |
+
755,
|
| 1553 |
+
826,
|
| 1554 |
+
858
|
| 1555 |
+
],
|
| 1556 |
+
"page_idx": 12
|
| 1557 |
+
},
|
| 1558 |
+
{
|
| 1559 |
+
"type": "text",
|
| 1560 |
+
"text": "• polar_tile is a modified version of tile but the grids are extracted from polar coordinates $( r , \\theta )$ centered at the center point where $r = l o g ( \\parallel \\Delta \\mathbf { x } _ { i j } \\parallel + 1 )$ . Instead of using grid size $c$ , we use the number of grids along $\\theta$ (or $r$ ) axis, $F$ , as the only hyperparameter. Similarly, We find that $r = l o g ( \\parallel \\Delta \\mathbf { x } _ { i j } \\parallel + \\mathrm { 1 } \\bar { ) }$ outperform $r = \\parallel \\Delta \\mathbf { x } _ { i j } \\parallel$ significantly. • scaled_rbf is a modified version of $r b f$ for cont whose kernel size is proportional to the distance between the current anchor point and the origin, $\\parallel \\Delta \\mathbf { x } _ { m } ^ { a n c \\hat { h } o r ^ { * } } \\parallel$ . That is ´ 2σ2scbasic kernel size and $\\exp \\big ( - \\frac { \\parallel \\Delta \\mathbf { x } _ { i j } - \\Delta \\mathbf { x } _ { m } ^ { a n c h o r } \\parallel ^ { 2 } } { 2 \\sigma _ { s c a l e d } ^ { 2 } } \\big )$ $\\beta$ d scaled “ \\` m is kernel rescale factor, a constant. We developed this me $\\sigma$ anism to help RFB to deal with relations at different scale, and we observe that it produces significantly better result than vanilla RBFs. ",
|
| 1561 |
+
"bbox": [
|
| 1562 |
+
214,
|
| 1563 |
+
103,
|
| 1564 |
+
826,
|
| 1565 |
+
266
|
| 1566 |
+
],
|
| 1567 |
+
"page_idx": 13
|
| 1568 |
+
},
|
| 1569 |
+
{
|
| 1570 |
+
"type": "text",
|
| 1571 |
+
"text": "A.2 HYPER-PARAMETER SELECTION ",
|
| 1572 |
+
"text_level": 1,
|
| 1573 |
+
"bbox": [
|
| 1574 |
+
176,
|
| 1575 |
+
284,
|
| 1576 |
+
444,
|
| 1577 |
+
297
|
| 1578 |
+
],
|
| 1579 |
+
"page_idx": 13
|
| 1580 |
+
},
|
| 1581 |
+
{
|
| 1582 |
+
"type": "text",
|
| 1583 |
+
"text": "We perform grid search for all methods based on their performance on the validation sets. ",
|
| 1584 |
+
"bbox": [
|
| 1585 |
+
173,
|
| 1586 |
+
309,
|
| 1587 |
+
758,
|
| 1588 |
+
324
|
| 1589 |
+
],
|
| 1590 |
+
"page_idx": 13
|
| 1591 |
+
},
|
| 1592 |
+
{
|
| 1593 |
+
"type": "text",
|
| 1594 |
+
"text": "Location Modeling The hyper-parameters of theory models are based on grid search with $d ^ { ( v ) } \\ = \\ ( 3 2 , 6 4 , 1 2 \\bar { 8 } , 2 5 6 )$ , $d ^ { ( x ) } \\ = \\ ( 3 2 , 6 4 , 1 2 8 , 2 5 6 )$ , $S \\ = \\ ( 4 , 8 , 1 6 , 3 2 , 6 4 , 1 2 8 )$ , and $\\lambda _ { m i n } ~ =$ $( 1 , 5 , 1 0 , 5 0 , 1 0 0 , 2 0 0 , 5 0 0 , 1 k )$ while $\\lambda _ { m a x } = 4 0 k$ is decided based on the total size of the study area. We find out the best performances of different grid cell based models are obtained when $d ^ { ( v ) } = 6 4$ , $d ^ { ( x ) } = 6 4$ , $S \\ : = \\ : 6 4$ , and $\\lambda _ { m i n } = 5 0$ . In terms of tile, the hyper-parameters are selected from $c = ( 1 0 , 5 0 , 1 0 0 , 2 0 0 , 5 0 0 , 1 0 0 0 )$ while $c = 5 0 0$ gives us the best performance.As for $r b f$ , we do grid search on the hyper-parameters: $M \\ = \\ ( 1 0 , 5 0 , 1 0 0 , 2 0 0 , 4 0 0 , 8 0 0 )$ and $\\sigma = ( \\mathrm { 1 0 ^ { 2 } } , \\mathrm { 1 0 ^ { 3 } } , \\mathrm { 1 0 ^ { 4 } } , \\mathrm { 1 0 ^ { 5 } } , \\mathrm { 1 0 ^ { 6 } } , \\mathrm { 1 0 ^ { 7 } } )$ . The best performance of $r b f$ is obtain when $M = 2 0 0$ and $\\sigma = 1 0 ^ { 3 }$ . As for wrap, grid search is performed on: $h = ( 1 , 2 , 3 , 4 )$ and $o = ( 6 4 , 1 2 8 , 2 5 6 , 5 1 2 )$ while $h = 3$ and $o = 5 1 2$ gives us the best result. All models use FFNs in their $E n c ^ { ( x ) } ( \\ v r )$ except wrap. The number of layers $f$ and the number of hidden state neurons $u$ of the FFN are selected from $f = ( 1 , 2 , 3 )$ and $u = ( 1 2 8 , 2 5 6 , 5 1 2 )$ . We find out $f = 1$ and $u = 5 1 2$ give the best performance for direct, tile, $r b f$ , and theory. So we use them for every model for a fair comparison. ",
|
| 1595 |
+
"bbox": [
|
| 1596 |
+
173,
|
| 1597 |
+
338,
|
| 1598 |
+
825,
|
| 1599 |
+
523
|
| 1600 |
+
],
|
| 1601 |
+
"page_idx": 13
|
| 1602 |
+
},
|
| 1603 |
+
{
|
| 1604 |
+
"type": "text",
|
| 1605 |
+
"text": "Spatial Context Modeling Grid search is used for hyperparameter tuning and the best performance of different grid cell models is obtain when $d ^ { ( v ) } = 6 \\dot { 4 }$ , $d ^ { ( \\bar { x } ) } = 6 4$ , $S = 6 4$ , and $\\lambda _ { m i n } = 1 0$ . We set $\\lambda _ { m a x } = 1 0 k$ based on the maximum displacement between context points and center points to make the location encoding unique. As for multiple baseline models, grid search is used again to obtain the best model. The best model hyperparameters are shown in () besides the model names in Table 3. Note that both $r b f$ and scaled_rbf achieve the best performance with $M = 1 0 0$ . ",
|
| 1606 |
+
"bbox": [
|
| 1607 |
+
173,
|
| 1608 |
+
539,
|
| 1609 |
+
825,
|
| 1610 |
+
625
|
| 1611 |
+
],
|
| 1612 |
+
"page_idx": 13
|
| 1613 |
+
},
|
| 1614 |
+
{
|
| 1615 |
+
"type": "image",
|
| 1616 |
+
"img_path": "images/b0f57259acc4ea4381424987490656af81d512522bc2e28f31e31c5ecb333e03.jpg",
|
| 1617 |
+
"image_caption": [
|
| 1618 |
+
"A.3 FIRING PATTERN FOR THE NEURONS ",
|
| 1619 |
+
"Figure 4: The firing pattern for the first 8 neurons (out of 64) given different encoders in location modeling. "
|
| 1620 |
+
],
|
| 1621 |
+
"image_footnote": [],
|
| 1622 |
+
"bbox": [
|
| 1623 |
+
178,
|
| 1624 |
+
669,
|
| 1625 |
+
820,
|
| 1626 |
+
873
|
| 1627 |
+
],
|
| 1628 |
+
"page_idx": 13
|
| 1629 |
+
},
|
| 1630 |
+
{
|
| 1631 |
+
"type": "image",
|
| 1632 |
+
"img_path": "images/c58840159e48f35df9f7269cc34a4fcb95454959e7e134131513dbfe0cfe2691.jpg",
|
| 1633 |
+
"image_caption": [],
|
| 1634 |
+
"image_footnote": [],
|
| 1635 |
+
"bbox": [
|
| 1636 |
+
173,
|
| 1637 |
+
131,
|
| 1638 |
+
818,
|
| 1639 |
+
387
|
| 1640 |
+
],
|
| 1641 |
+
"page_idx": 14
|
| 1642 |
+
},
|
| 1643 |
+
{
|
| 1644 |
+
"type": "image",
|
| 1645 |
+
"img_path": "images/42c26ab4f3f54a2d8d4f51cc6de55ca1f930d86c5b7b3ca778d5560cf27f55a4.jpg",
|
| 1646 |
+
"image_caption": [
|
| 1647 |
+
"Figure 5: Embedding clustering in the original space of (a)(b)(c)(d) theory with different $\\lambda _ { m i n }$ , but the same $\\lambda _ { m a x } = 1 0 k$ and $S = 6 4$ . (e)(f)(g)(h) are the embedding clustering results of the same models in the polar-distance space. All models use 1 hidden ReLU layers of 512 neurons. ",
|
| 1648 |
+
"Figure 6: Embedding clustering of RBF models with different kernel rescalar factor $\\beta$ (a)(b)(c)(d) in the original space; (e)(f)(g)(h) in the polar-distance space. Here $\\beta { = } 0 . 0$ indicates the original RBF model. All models use $\\sigma { = } 1 0 \\mathrm { m }$ as the basic kernel size and 1 hidden ReLU layers of 512 neurons. "
|
| 1649 |
+
],
|
| 1650 |
+
"image_footnote": [],
|
| 1651 |
+
"bbox": [
|
| 1652 |
+
173,
|
| 1653 |
+
458,
|
| 1654 |
+
818,
|
| 1655 |
+
713
|
| 1656 |
+
],
|
| 1657 |
+
"page_idx": 14
|
| 1658 |
+
}
|
| 1659 |
+
]
|
parse/train/rJljdh4KDH/rJljdh4KDH_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/rJljdh4KDH/rJljdh4KDH_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/rJxGLlBtwH/rJxGLlBtwH.md
ADDED
|
@@ -0,0 +1,273 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ON THE INTERACTION BETWEEN SUPERVISION AND SELF-PLAY IN EMERGENT COMMUNICATION
|
| 2 |
+
|
| 3 |
+
Ryan Lowe∗, Abhinav Gupta∗ MILA
|
| 4 |
+
|
| 5 |
+
Jakob Foerster, Douwe Kiela Facebook AI Research
|
| 6 |
+
|
| 7 |
+
Joelle Pineau Facebook AI Research MILA
|
| 8 |
+
|
| 9 |
+
# ABSTRACT
|
| 10 |
+
|
| 11 |
+
A promising approach for teaching artificial agents to use natural language involves using human-in-the-loop training. However, recent work suggests that current machine learning methods are too data inefficient to be trained in this way from scratch. In this paper, we investigate the relationship between two categories of learning signals with the ultimate goal of improving sample efficiency: imitating human language data via supervised learning, and maximizing reward in a simulated multi-agent environment via self-play (as done in emergent communication), and introduce the term supervised self-play $( S 2 P )$ for algorithms using both of these signals. We find that first training agents via supervised learning on human data followed by self-play outperforms the converse, suggesting that it is not beneficial to emerge languages from scratch. We then empirically investigate various S2P schedules that begin with supervised learning in two environments: a Lewis signaling game with symbolic inputs, and an image-based referential game with natural language descriptions. Lastly, we introduce population based approaches to S2P, which further improves the performance over single-agent methods.1
|
| 12 |
+
|
| 13 |
+
# 1 INTRODUCTION
|
| 14 |
+
|
| 15 |
+
Language is one of the most important aspects of human intelligence; it allows humans to coordinate and share knowledge with each other. It is also crucial for human-machine interaction, as human language is a natural means by which to exchange information, give feedback, and specify goals. A promising approach for training agents to solve problems with natural language is to have a “human in the loop”, meaning we collect problem-specific data from humans interacting directly with our agents for learning. However, human-in-the-loop data is expensive and time-consuming to obtain as it requires continuously collecting human data as the agent’s policy improves, and recent work suggests that current machine learning methods (e.g. from deep reinforcement learning) are too data-inefficient to be trained in this way from scratch (Chevalier-Boisvert et al., 2019). Thus, an important open problem is: how can we make human-in-the-loop training as data efficient as possible?
|
| 16 |
+
|
| 17 |
+
To maximize data efficiency, it is important to fully leverage all available training signals. In this paper, we study two categories of such training methods: imitating human data via supervised learning, and self-play to maximize reward in a multi-agent environment, both of which provide rich signals for endowing agents with language-using capabilities. However, these are potentially competing objectives, as maximizing environmental reward can lead to the resulting communication protocol drifting from natural language (Lewis et al., 2017; Lee et al., 2019). The crucial question, then, is how do we best combine self-play and supervised updates? This question has received surprisingly little attention from the emergent communication literature, where the question of how to bridge the gap from emergent protocols to natural language is generally left for future work (Mordatch & Abbeel, 2018; Lazaridou et al., 2018; Cao et al., 2018).
|
| 18 |
+
|
| 19 |
+

|
| 20 |
+
Figure 1: (a) Diagram of the supervised self-play (S2P) procedure (phases 1-3) and the testing procedure considered in this work (phase 4). (b) The environments considered in this paper (Sec. 4).
|
| 21 |
+
|
| 22 |
+
Our goal in this paper is to investigate algorithms for combining supervised learning with self-play — which we call supervised self-play (S2P) algorithms — using two classic emergent communication tasks: a Lewis signaling game with symbolic inputs, and a more complicated image-based referential game with natural language descriptions. Our first finding is that supervised learning followed by self-play outperforms emergent communication with supervised fine-tuning in these environments, and we provide three reasons for why this is the case. We then empirically investigate several supervised-first S2P methods in our environments. Existing approaches in this area have used various ad-hoc schedules for alternating between the two kinds of updates (Lazaridou et al., 2017), but to our knowledge there has been no systematic study that has compared these approaches. Lastly, we propose the use of population-based methods for S2P, and find that it leads to improved performance in the more challenging image-based referential game. Our findings highlight the need for further work in combining supervised learning and self-play to develop more sample-efficient language learners.
|
| 23 |
+
|
| 24 |
+
# 2 RELATED WORK
|
| 25 |
+
|
| 26 |
+
In the past few years, there has been a renewed interest in the field of emergent communication (Sukhbaatar et al., 2016; Foerster et al., 2016; Lazaridou et al., 2017; Havrylov & Titov, 2017) culminating in 3 NeurIPS workshops. Empirical studies have showed that agents can autonomously evolve a communication protocol using discrete symbols when deployed in a multi-agent environment which helps them to play a cooperative or competitive game (Singh et al., 2019; Cao et al., 2018; Choi et al., 2018; Resnick\* et al., 2019; Evtimova et al., 2018).
|
| 27 |
+
|
| 28 |
+
While the idea of promoting coordination among agents through communication sounds promising, recent experiments (Lowe et al., 2019; Chaabouni et al., 2019; Kottur et al., 2017; Jaques et al., 2019) have emphasized the difficulty in learning meaningful emergent communication protocols even with centralized training.
|
| 29 |
+
|
| 30 |
+
Apart from the above advances in emergent communication, there has been a long outstanding goal of learning intelligent conversational agents to be able to interact with humans. This involves training the artificial agents in a way so that they achieve high scores while solving the task and their language is interpretable by humans or close to natural language. Recent works also add another axis orthogonal to communication where the agent also takes a discrete action in an interactive environment (de Vries et al., 2018; Mul et al., 2019). Lewis et al. (2017) introduced a negotiation task which involves learning linguistic and reasoning skills. They train models imitating human utterances using supervised learning and found that the model generated human-like captions but were poor negotiators. So they perform self-play with these pretrained agents in an interleaved manner and found that the performance improved drastically while avoiding language drift. Lee et al. (2019) also propose using an auxiliary task for grounding the communication to counter language drift. They use visual grounding to learn the semantics of the language while still generating messages that are close to English.
|
| 31 |
+
|
| 32 |
+
A recent trend on using population based training for multi-agent communication is a promising avenue for research using inspirations from language evolution literature (Smith et al., 2003; Kirby,
|
| 33 |
+
|
| 34 |
+
2014; Raviv & Arnon, 2018). Cultural transmission is one such technique which focuses on the structure and compression of languages, since a language must be used and learned by all individuals of the culture in which it resides and at the same time be suitable for a variety of tasks. Harding Graesser et al. (2019) shows the emergence of linguistic phenomena when a pool of agents contact each other giving rise to novel creole languages. Li & Bowling (2019); Cogswell et al. (2019); Tieleman et al. (2018) have also tried different ways of imposing cultural pressures on agents, by simulating a large population of them and pairing agents to solve a cooperative game with communication. They train the agent against a sampled generation of agents where the generation corresponds to the different languages of the different agent at different times in the history.
|
| 35 |
+
|
| 36 |
+
Our work is inspired from these works where we aim to formalize the recent advancements in using self-play in dialog modeling, through the lens of emergent communication.
|
| 37 |
+
|
| 38 |
+
# 3 METHODS
|
| 39 |
+
|
| 40 |
+
# 3.1 PROBLEM DEFINITION
|
| 41 |
+
|
| 42 |
+
Our agents are embedded in a multi-agent environment with $N$ agents where they receive observations $o \in O$ (which are functions of a hidden state $S$ ) and perform actions $a \in A$ . Some actions $A _ { L } \subset A$ involve sending a message $m \in A _ { L }$ over a discrete, costless communication channel (i.e. a cheap talk channel (Farrell & Rabin, 1996)). The agents are rewarded with a reward $r \in R$ for their performance in the environment. We assume throughout that the environment is cooperative and thus the agents are trained to maximize the sum of rewards $\begin{array} { r } { R = \sum _ { t = 1 : T } \sum _ { i = 1 : N } r _ { i , t } } \end{array}$ across both agents. This can be thought of as a cooperative partially-observable Markov game (Littman (1994)).
|
| 43 |
+
|
| 44 |
+
We define a target language $L ^ { * } \in { \mathcal { L } }$ , usually corresponding to natural language, that we want our agents to learn (we further assume $L ^ { * }$ can be used to achieve high task reward). In this paper, we consider a language $L \in { \mathcal { L } }$ to be simply a set of valid messages $A _ { L }$ and a mapping between observations and messages in the environment, $L : O \times A _ { L } \mapsto [ 0 , 1 ]$ . For example, in an English image-based referential game (Section 4) this corresponds to the mapping between images and image descriptions in English. We are given a dataset $\mathcal { D }$ consisting of $| \mathcal D |$ (observation, action) pairs, corresponding to $N _ { e }$ ‘experts’ (for us, $N _ { e } = 2$ ) playing the game using the target language $L ^ { * }$ . Our goal is to train agents to achieve a high reward in the game while speaking language $L ^ { * }$ with an ‘expert’. Specifically, we want our agents to generalize and to perform well on examples that are not contained in $\mathcal { D }$ .
|
| 45 |
+
|
| 46 |
+
To summarize, we want agents that can perform well on a collaborative task with English-speaking humans, and we can train them using a supervised dataset $\mathcal { D }$ and via self-play.
|
| 47 |
+
|
| 48 |
+
# 3.2 SUPERVISED SELF-PLAY (S2P)
|
| 49 |
+
|
| 50 |
+
In recent years, there have been several approaches to language learning that have combined supervised or imitation learning with self-play. In this paper, we propose an umbrella term for these algorithms called supervised self-play (S2P). S2P requires two things: (1) a multi-agent environment where at least one agent can send messages over a dedicated communication channel, along with a reward function that measures how well the agents are doing at some task; and (2) a supervised dataset $\mathcal { D }$ of experts acting and speaking language $L ^ { * }$ in the environment (such that they perform well on the task). Given these ingredients, we define S2P below (see Figure 2).
|
| 51 |
+
|
| 52 |
+
Definition 3.1. Supervised self-play (S2P). Supervised self-play is a class of language learning algorithms that combines: (1) self-play updates in a multi-agent language environment, and (2) supervised updates on an expert dataset $\mathcal { D }$ .
|
| 53 |
+
|
| 54 |
+
S2P algorithms can differ in how they combine self-play and supervised learning updates on $\mathcal { D }$ . When supervised learning is performed before self-play, we refer to the dataset $\mathcal { D }$ as the seed data. Why might we want to train our agents via self-play? Won’t their language diverge from $L ^ { * } ?$ One way to intuitively understand why S2P is beneficial is to think in terms of constraints. In our set-up, there are two known constraints on the target language $L ^ { * }$ : (1) it is consistent with the samples from the supervised dataset $\mathcal { D }$ , and (2) $L ^ { * }$ can be used to obtain a high reward in the environment. Thus, finding $L ^ { * }$ can be loosely viewed as a constrained optimization problem, and enforcing both constraints should clearly lead to better performance.
|
| 55 |
+
|
| 56 |
+
# 3.3 ALGORITHMS FOR S2P
|
| 57 |
+
|
| 58 |
+
Here we describe several methods for S2P training. Our goal is not to exhaustively enumerate all possible optimization strategies, but rather provide a categorization of some well-known ways to combine self-play and supervised learning. To help describe these methods, we further split the seed dataset $\mathcal { D }$ into $\mathcal { D } _ { t r a i n }$ , which is used for training, and $\mathcal { D } _ { v a l }$ which is used for early-stopping. We also visualize the schedules in Figure 2.
|
| 59 |
+
|
| 60 |
+
Emergent communication with supervised fine-tuning (sp2sup): We first perform self-play updates until the learning converges on the task performance. It is then followed by supervised updates on $\mathcal { D } _ { t r a i n }$ until the listener performance converges on the dataset $\mathcal { D } _ { v a l }$ .
|
| 61 |
+
|
| 62 |
+
Supervised learning with self-play (sup2sp): This is the complement of the above method which involves doing supervised updates until convergence on $\mathcal { D } _ { v a l }$ followed by self-play updates until convergence on the task performance.
|
| 63 |
+
|
| 64 |
+

|
| 65 |
+
Figure 2: A visual representation of the different S2P methods.
|
| 66 |
+
|
| 67 |
+
Random updates $( \tt r a n d )$ : This is the method used in (Lazaridou et al., 2017). At each time step, we sample a bernoulli random variable $z \sim B e r n o u l l i ( q )$ where $q$ is fixed. If $z = 1$ , we perform one supervised update, else we do one self-play update, and repeat until both losses convergence on $\mathcal { D } _ { v a l }$ .
|
| 68 |
+
|
| 69 |
+
Scheduled updates (sched): We first pretrain the listener and the speaker until convergence on $\mathcal { D } _ { v a l }$ . Then we create a schedule, where we perform $l$ self-play updates followed by $m$ supervised updates, and repeat until convergence on the dataset.
|
| 70 |
+
|
| 71 |
+
Scheduled updates with speaker freezing (sched_ $\pmb { \mathcal { \tt { f r } } } \mathbf { z }$ ): This method is based on the findings of Lewis et al. (2017), who do sched S2P while freezing the parameters of the speaker during self-play to reduce the amount of language drift. In our case, we freeze the parameters of the speaker after the initial supervised learning.
|
| 72 |
+
|
| 73 |
+
Scheduled updates with random speaker freezing (sched_rand_frz): Experimentally, we noticed that sched_frz didn’t perform well in self-play. Thus, we introduce a variation, we sample a bernoulli random variable $z \sim B e r n o u l l i ( r )$ where $r$ is fixed. If $z = 1$ , we freeze the parameters of the speaker during both self-play and supervised learning, else we allow updates to the speaker as well.
|
| 74 |
+
|
| 75 |
+
# 3.4 POPULATION-BASED S2P (POP-S2P)
|
| 76 |
+
|
| 77 |
+
As explained above, the goal of S2P is to produce agents that follow dataset $\mathcal { D }$ while maximizing reward in the environment. However, there are many such policies satisfying these criteria. This results in a large space of possible solutions, that increases as the environment grows more complex (but decreases with increasing $| \mathcal { D } | )$ . Experimentally, we find that this can result in diverse agent policies. We show this in Figure 3 by training 50 randomly initialized agents on the image-based referential game (defined in Sec. 4) the agents can often make diverse predictions for a given image (Figure 3a) and achieve variable performance when playing with other populations with a slight preference towards their own partner (the diagonal in Figure 3b).
|
| 78 |
+
|
| 79 |
+

|
| 80 |
+
Figure 3: Results from training $5 0 \ : \mathrm { S 2 P }$ agents on the IBR game with $| \mathcal { D } | = 1 0 0 0 0$ . (a) The agents have a range of predictions on many images. (b) When playing with each other, the agents exhibit uneven performance (color is mean reward, yellow is higher), indicating policy variability.
|
| 81 |
+
|
| 82 |
+
ment S2P by training a population of $N$ agents, and subsequently aggregating them back into a single agent (the ‘student’). We call this population-based S2P (Pop-S2P). While there are many feasible ways of doing this, in this paper we train the populations by simply randomizing the initial seed, and we aggregate the populations using a simple form of policy distillation (Rusu et al., 2016). Another simple technique to boost performance is via ensembling where we simply take the majority prediction at each time step.
|
| 83 |
+
|
| 84 |
+
# 4 ENVIRONMENTS & IMPLEMENTATION DETAILS
|
| 85 |
+
|
| 86 |
+
We consider environments based on classical problems in emergent communication. These environments are cooperative and involve the interaction between a speaker, who makes an observation and sends a message, and a listener, who observes the message and makes a prediction (see Figure 1b). Our goal is to train a listener such that it achieves high reward when playing with an expert speaking the target language $L ^ { * }$ on inputs unseen during training.2
|
| 87 |
+
|
| 88 |
+
Environment 1: Object Reconstruction (OR) Our first game is a Lewis signaling game (Lewis, 1969) and a simpler version of the Task & Talk game from Kottur et al. (2017), with a single turn and a much larger input space. The speaker agent observes an object with a certain set of properties, and must describe the object to the listener using a sequence of words. The listener then attempts to reconstruct the object. More specifically, the input space consists of $p$ properties (e.g. shape, color) of $t$ types each (e.g. triangle, square). The speaker observes a symbolic representation of the input, consisting of the concatenation of $p = 6$ one-hot vectors, each of length $t = 1 0$ . The number of possible inputs scales as $t ^ { p }$ . We define the vocabulary size (length of each one-hot vector sent from the speaker) as $| V | = 6 0$ , and the number of words (fixed length message) sent to be $T = 6$ .
|
| 89 |
+
|
| 90 |
+
For our target language $L ^ { * }$ for this task, we programatically generate a perfectly compositional language, by assigning each object a unique word. In other words, to describe a ‘blue shaded triangle’, we create a language where the output description would be “blue, triangle, shaded”, in some arbitrary order. By ‘unique symbol’, we mean that no two properties are assigned the same word. The speaker and listener policies are parameterized using a 2-layer linear network (results were similar with added non-linearity and significantly worse with 1-layer linear networks) with 200 hidden units. During both supervised learning and self-play, the listener is trained to minimize the cross-entropy loss over property predictions.
|
| 91 |
+
|
| 92 |
+
Environment 2: Image-Based Referential game with natural language (IBR) Our second game is the communication task introduced in Lee et al. (2018). The speaker observes a target image $d ^ { * }$ , and must describe the image using a set of words. The listener observes the target image along with $D$ distractor images sampled uniformly at random from the training set (for us, $D = 9$ ), and the message $y _ { d ^ { * } }$ from the speaker, and is rewarded for correctly selecting the target image. For this game, the target language $L ^ { * }$ is English — we obtain English image descriptions using caption data from MS COCO and Flickr30k. We set the vocabulary size $| V | = 1 0 0$ , and filter out any descriptions that contain more than $30 \%$ unknown tokens while keeping the maximum message length $T$ to 15.
|
| 93 |
+
|
| 94 |
+
Similar to (Mordatch & Abbeel, 2018; Sukhbaatar et al., 2016), we train our agents end-to-end with backpropagation. Since the speaker sends discrete messages, we use the Straight-Through version of Gumbel-Softmax (Jang et al., 2017; Maddison et al., 2017) to allow gradient flow to the speaker during self-play $\left( \mathcal { I } _ { \mathrm { s e l f - p l a y } } \right)$ . The speaker’s predictions are trained on the ground truth English captions $m ^ { * }$ using the cross entropy loss $\mathcal { I } _ { \mathrm { s p k } }$ -supervised. The listener is trained using the cross-entropy loss $\mathcal { I } _ { \mathrm { l s n } }$ -supervised where the logits are the reciprocal of the mean squared error which was found to perform better than directly minimizing MSE loss in Lee et al. (2018). The mean squared error is taken over the listener’s image representation $b _ { l s n }$ of the distractor (or target) image and the message representation given as input. The loss functions are defined as:
|
| 95 |
+
|
| 96 |
+
$$
|
| 97 |
+
\begin{array} { c } { { \mathcal { I } _ { \mathrm { s p k \mathrm { s u p e r v i s e d } } } ( d ^ { * } ) = \displaystyle - \sum _ { t = 1 } ^ { T } \log p _ { s p k } ( m _ { t } | m _ { < t } , d ^ { * } ) } } \\ { { \mathcal { I } _ { \mathrm { l s n \mathrm { - } s u p e r v i s e d } } ( m ^ { * } , d ^ { * } , D ) = \displaystyle - \sum _ { d = 1 } ^ { D + 1 } \log ( \mathsf { s o f t r m a x } ( 1 / p _ { l s n } ( m ^ { * } ) - b _ { \mathrm { l s n } } ( d ) ) ^ { 2 } ) } } \end{array}
|
| 98 |
+
$$
|
| 99 |
+
|
| 100 |
+

|
| 101 |
+
Figure 4: (a) Left: In the OR game, best performance (number of total samples required to achieve $9 5 \%$ test accuracy, lower is better) for S2P is achieved when all of the samples are in the seed. 0 on the $\mathbf { X }$ -axis corresponds to sp2sup and Optimal is the actual (minimum) number of samples required to solve this optimization problem (see Appenix B). Right: This is also the case in the IBR game, where performance is measured by the generalization accuracy using $1 0 \mathrm { k }$ total training samples (higher is better). (b) Adding more samples to initial supervised learning in the IBR game improves agents’ generalization to $L ^ { * }$ . (c) Even when we learn the perfect distribution with emergent communication in the OR game, it still performs worse than Pop-S2P (using sup2sp S2P).
|
| 102 |
+
|
| 103 |
+
$$
|
| 104 |
+
\mathcal { I } _ { \mathrm { s e l f - p l a y } } ( d ^ { * } , D ) = - \sum _ { d = 1 } ^ { D + 1 } \log ( \mathsf { s o f t m a x } ( 1 / p _ { l s n } ( y _ { d ^ { * } } ) - b _ { \mathrm { l s n } } ( d ) ) ^ { 2 } )
|
| 105 |
+
$$
|
| 106 |
+
|
| 107 |
+
where $y _ { d ^ { * } }$ is the concatenation of $T$ one-hot vectors $y _ { d ^ { * } } ^ { t } = \mathtt { S T - G u m b e l S o f t m a x } ( p _ { s p k } ^ { t } ) .$
|
| 108 |
+
|
| 109 |
+
We use the same architecture as described in Lee et al. (2018). The speaker and listener are parameterized by recurrent policies, both using an embedding layer of size 256 followed by a GRU (Cho et al., 2014) of size 512. We provide further hyperparameter details in Table 1 in the Appendix.
|
| 110 |
+
|
| 111 |
+
# 5 DO SUPERVISED LEARNING BEFORE SELF-PLAY
|
| 112 |
+
|
| 113 |
+
A central question in our work is how to combine supervised and self-play updates for effective pre-training of conversational agents. In this section, we study this question by conducting experiments with two schedules: training with emergent communication followed by supervised learning $( { \tt s p 2 s u p } )$ , and training with supervised learning followed by self-play (sup2sp). We also interpolate between these two regimes by performing the rand and sched on $0 < n < | \mathcal { D } |$ samples, followed by supervised fine-tuning on the remaining $| \mathcal { D } | - n$ samples.
|
| 114 |
+
|
| 115 |
+
Our first finding is that it is best to use all of your samples for supervised learning before doing self-play. This can be seen in Figure 4: when all of the samples are used first for supervised learning, the number of total samples required to solve the OR game drastically, and in the IBR game the accuracy for a fixed number of samples is maximized (Figure 4a). While this may seem to be common sense, it in fact runs counter to the prevailing wisdom in some emergent communication literature, where languages are emerged from scratch with the ultimate goal of translating them to natural language.
|
| 116 |
+
|
| 117 |
+
To better understand why it is best to do supervised learning first, we now conduct a set of targeted experiments using the environments from Section 4. Results of our experiments suggest three main explanations:
|
| 118 |
+
|
| 119 |
+
(1) Emerging a language is hard. For many environments, with emergent communication it’s often hard to find an equilibrium where the agents meaningfully communicate. The difficulty of ‘emergent language discovery’ has been well-known in emergent communication (Lowe et al., 2017), so we will only briefly discuss it here. In short, to discover a useful communication protocol agents have to coordinate repeatedly over time, which is difficult when agents are randomly initialized, particularly in environments with sparse reward. Compounding the difficulty is that, if neither agent communicates and both agents act optimally given their lack of knowledge, they converge to a Nash equilibrium called babbling equilibrium (Farrell & Rabin, 1996). This equilibrium must be overcome to learn a useful communication protocol. In S2P, the initial language supervision can help overcome the discovery problem, as it provides an initial policy for how agents could usefully communicate (Lewis et al., 2017).
|
| 120 |
+
|
| 121 |
+

|
| 122 |
+
Figure 5: Results from the OR game with 1 property and 10 types. When the supervised updates are performed first (supervised data available for words $0 - 3 )$ , then the self-play updates make sensible predictions for the unknown words $4 - 7$ . When the self-play updates are performed first, the subsequent supervised updates merely correct the predictions for words $1 - 4$ , without enforcing the constraint that each word should result in a separate type to solve the task.
|
| 123 |
+
|
| 124 |
+
(2) Emergent languages are different than natural language. Even if one does find an equilibrium where agents communicate and perform well on the task, the distribution of languages they find will usually be very different from natural language. This is a problem because, if the languages obtained through self-play are sufficiently different from $L ^ { * }$ , they will not be helpful for learning. This is seen for the OR game in Figure 4a, where 17 samples are required in the seed before S2P outperforms the supervised learning baseline. We speculate that this is due to the different pressures exerted during the emergence of artificial languages and human languages.
|
| 125 |
+
|
| 126 |
+
Thankfully, we can learn languages closer to $L ^ { * }$ by simply adding more samples to our initial supervised learning phase. We show this in Figure 4b, where we train populations of 50 agents on the IBR game and use Pop-S2P to produce a single distilled agent. With both 1K and 10K initial supervised samples, the distill agent generalizes to agents in the validation set of their population. However, the distilled agent trained with 10000 samples performs significantly better when playing with an expert agent speaking $L ^ { * }$ , indicating that the training agents from that population speak languages closer to $L ^ { * }$ .
|
| 127 |
+
|
| 128 |
+
(3) Starting with self-play violates constraints. Even if you have ‘perfect emergent communication’ that learns a distribution over languages under which $L ^ { * }$ has high probability, current methods of supervised fine-tuning do not properly learn from this distribution. What if we had all the correct learning pressures, such that we emerged a distribution over languages $\mathcal { L }$ with structure identical to $L ^ { * }$ , and then trained a Pop-S2P agent using this distribution? Surprisingly, we find that S2P with all of the samples in the seed performs better than even this optimistic case, in terms of providing useful information for training a Pop-S2P agent. We conduct an experiment in the OR game where we programmatically define a distribution over compositional languages $\mathcal { L } _ { c }$ , of which our target language $L ^ { * }$ is a sample. Each language $L \in \mathcal { L } _ { c }$ has the same structure and are obtained by randomly permuting the mapping between the word IDs and the corresponding type IDs, along with the order of properties in an utterance. Next, we compare two distilled policies using 50 populations: one is distilled from S2P populations (trained with $X$ samples), and the other is distilled from ‘perfect emergent communication’ and fine-tuned on $X$ samples. As can be seen in Figure 4c, we show that when we train a Pop-S2P agent on 50 of these compositional populations, we still need $3 X$ more samples than regular Pop-S2P (trained on ${ 5 0 } \thinspace \mathrm { S 2 P }$ agents with all of the samples in the seed) to reach $9 5 \%$ test accuracy3.
|
| 129 |
+
|
| 130 |
+
To understand why this happens, we conduct a case study in an even simpler setting: single-agent S2P in the OR game with $p = 1 , t = 1 0 , | V | = 1 0$ . We find that agents trained via emergent communication consistently learn to solve this task. However, as shown in Figure 5, when subsequently trained via supervised learning on $\mathcal { D }$ to learn $L ^ { * }$ , the learned language is no longer coherent (it maps different words to the same type) and doesn’t solve the task. On the other hand, agents trained first with supervised learning are able to learn a language that both solves the task and is consistent with $\mathcal { D }$ .
|
| 131 |
+
|
| 132 |
+
Intuitively, what’s happening is that the samples in $\mathcal { D }$ are also valid for solving the task, since we assume agents speaking $L ^ { * }$ can solve the task. Thus, self-play after supervised learning simply ‘fills in the gaps’ for examples not in $\mathcal { D }$ .4 Emergent languages that start with self-play, on the other hand, contain input-output mappings that are inconsistent with $L ^ { * }$ , which must be un-learned during subsequent supervised learning.
|
| 133 |
+
|
| 134 |
+
In theory, the above issue could be resolved using Pop-S2P; if the distilled agent could use the population of emergent languages to discover structural rules (e.g. discovering that the languages in the OR game in Figure 4c are compositional), it could use the samples from $\mathcal { D }$ to refine a posterior distribution over target languages that is consistent with these rules (e.g. learning the distribution of compositional languages consistent with $\mathcal { D }$ ). Current approaches to supervised fine-tuning in language, though, do not do this (Lazaridou et al., 2017; Lewis et al., 2017). An interesting direction for future work is examining how to apply Bayesian techniques to S2P.
|
| 135 |
+
|
| 136 |
+
# 6 EXPLORING VARIANTS OF S2P
|
| 137 |
+
|
| 138 |
+
# 6.1 POPULATION-BASED S2P
|
| 139 |
+
|
| 140 |
+
In this section, we aim to show that (1) S2P outperforms the supervised learning baseline, and (2) Pop-S2P outperforms S2P. We conduct our experiments in the more complex IBR game, since the agents must communicate in English, and measure performance by calculating the accuracy at different (fixed) numbers of samples. Our baseline is then the performance of a supervised learner on a fixed number of samples.
|
| 141 |
+
|
| 142 |
+
We show the results in Figure 6. We first note that, when both 1k and 10k samples are used for supervised learning, S2P (sched) outperforms the supervised learning baseline. We can also see that the population-based approach outperforms single agent S2P (sched) by a significant margin. We also compare our distillation method to an ensembling method that keeps all 50 populations at test time, and find that ensembling performs significantly better, although it is much less efficient. This suggests that there is room to push distilled Pop-S2P to even better performance.
|
| 143 |
+
|
| 144 |
+

|
| 145 |
+
Figure 6: S2P (sched) outperforms the supervised baseline in the IBR game, and is in turn outperformed by PopS2P.
|
| 146 |
+
|
| 147 |
+
# 6.2 EXAMINING S2P SCHEDULES
|
| 148 |
+
|
| 149 |
+
In this section, we aim to: (1) evaluate several S2P schedules empirically on the IBR game; and (2) attain a better understanding of S2P through quantitative experiments.
|
| 150 |
+
|
| 151 |
+
Parameter freezing improves S2P We show the results comparing different S2P schedules in Figure 7a. We find that in this more complex game, the sup2sp S2P performs much worse than the other options. We also see that adding freezing slightly improves the performance on the target language (Figure 8 in the Appendix also shows that it converges more quickly). We hypothesize that this is because it reduces the language drift that is experienced during each round of self-play updates (Lee et al., 2019). Overall, however, the difference between different S2P schedules is relatively small, and it’s unclear if the same ordering will hold in a different domain.
|
| 152 |
+
|
| 153 |
+

|
| 154 |
+
Figure 7: (a) Comparing test performances of different S2P methods on the IBR game. For each method, we picked the model that gave the best performance on $\mathcal { D } _ { v a l }$ . (b) 2D visualization of S2P (sched) performance over the course of training, in terms of performance on $L ^ { * }$ (vertical axis) and performance in self-play (horizontal axis). The zig-zag patterns indicates that most self-play updates result in a short-term decrease in target language performance. (c) Visualization of the role of the supervised and self-play updates in sched S2P.
|
| 155 |
+
|
| 156 |
+
Self-play acts as a regularizer What is the role of self-play in S2P? We can start to decipher this by taking a closer look at the sched S2P. We plot the training performance of this method in Figure 7b. Interestingly, we notice from the zig-zag pattern that the validation performance usually goes down after every set of self-play updates. However, the overall validation performance goes up after the next round of supervised updates. This is also reflected in the poor performance of the sup2sp S2P in Figure 6.
|
| 157 |
+
|
| 158 |
+
This phenomenon can be explained by framing self-play as a form of regularization: alternating between supervised and self-play updates is a way to satisfy the parallel constraints of ‘is consistent with the dataset $\mathcal { D } ^ { \ast }$ and ‘performs well on the task’. We visualize this pictorially in Figure 7b: while a set of self-play updates results in poor performance on $\mathcal { D }$ , eventually the learned language moves closer to satisfying both constraints.
|
| 159 |
+
|
| 160 |
+
# 7 DISCUSSION
|
| 161 |
+
|
| 162 |
+
In this work, we investigated the research question of how to combine supervised and self-play updates, with a focus on training agents to learn a language. However, this research question is not only important for language learning; it is also a important in equilibrium selection and learning social conventions (Lerer & Peysakhovich, 2019) in general games. For example, in robotics there may be a trade-off between performing a task well (moving an object to a certain place) and having your policy be interpretable by humans (so that they will not stumble over you). Examining how to combine supervised and self-play updates in these settings is an exciting direction for future work.
|
| 163 |
+
|
| 164 |
+
There are several axes of complexity not addressed in our environments and problem set-up. First, we consider only single-state environments, and agents don’t have to make temporally extended decisions. Second, we do not consider pre-training on large text corpora that are separate from the desired task (Radford et al., 2019; Devlin et al., 2018). Third, we limit our exploration of self-play to the multi-agent setting, which is not the case in works such as instruction following (Andreas & Klein, 2015). Introducing these elements may result in additional practical considerations for S2P learning, which we leave for future work. Our goal in this paper is not to determine the best method of S2P in all of these settings, but rather to inspire others to use the framing of ‘supervised self-play algorithms’ to make progress on sample efficient human-in-the-loop language learning.
|
| 165 |
+
|
| 166 |
+
# ACKNOWLEDGEMENTS
|
| 167 |
+
|
| 168 |
+
We are very grateful to Angeliki Lazaridou, with whom discussions at ICML 2019 and her simultaneous work (Lazaridou et al., 2020) shifted the direction of this work considerably. We also thank Jean Harb, Liam Fedus, Amy Zhang, Evgeny Naumov, Cinjon Resnick, Igor Mordatch, and others at MILA and Facebook AI Research for discussions related to the ideas in this paper. Special thanks to Arthur Szlam and Kavya Srinet for discussing their ongoing work with us. RL is supported in part by a Vanier Scholarship.
|
| 169 |
+
|
| 170 |
+
# REFERENCES
|
| 171 |
+
|
| 172 |
+
Jacob Andreas and Dan Klein. Alignment-based compositional semantics for instruction following. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 1165–1174, Lisbon, Portugal, September 2015. Association for Computational Linguistics. doi: 10.18653/v1/D15-1138. URL https://www.aclweb.org/anthology/D15-1138.
|
| 173 |
+
|
| 174 |
+
Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent communication through negotiation. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id ${ . } =$ Hk6WhagRW.
|
| 175 |
+
|
| 176 |
+
Rahma Chaabouni, Eugene Kharitonov, Emmanuel Dupoux, and Marco Baroni. Anti-efficient encoding in emergent communication. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d\textquotesingle Alché-Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 6290–6300. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8859-anti-efficient-encoding-in-emergent-communication.pdf.
|
| 177 |
+
|
| 178 |
+
Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Salem Lahlou, Lucas Willems, Chitwan Saharia, Thien Huu Nguyen, and Yoshua Bengio. BabyAI: First steps towards grounded language learning with a human in the loop. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ rJeXCo0cYX.
|
| 179 |
+
|
| 180 |
+
Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1724–1734, Doha, Qatar, October 2014. Association for Computational Linguistics. doi: 10.3115/v1/D14-1179. URL https://www.aclweb.org/ anthology/D14-1179.
|
| 181 |
+
|
| 182 |
+
Edward Choi, Angeliki Lazaridou, and Nando de Freitas. Multi-agent compositional communication learning from raw visual input. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rknt2Be0-.
|
| 183 |
+
|
| 184 |
+
Michael Cogswell, Jiasen Lu, Stefan Lee, Devi Parikh, and Dhruv Batra. Emergence of Compositional Language with Deep Generational Transmission. arXiv:1904.09067 [cs, stat], April 2019. arXiv: 1904.09067.
|
| 185 |
+
|
| 186 |
+
Harm de Vries, Kurt Shuster, Dhruv Batra, Devi Parikh, Jason Weston, and Douwe Kiela. Talk the walk: Navigating new york city through grounded dialogue. arXiv preprint arXiv:1807.03367, 2018.
|
| 187 |
+
|
| 188 |
+
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.
|
| 189 |
+
|
| 190 |
+
Katrina Evtimova, Andrew Drozdov, Douwe Kiela, and Kyunghyun Cho. Emergent communication in a multi-modal, multi-step referential game. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } }$ rJGZq6g0-.
|
| 191 |
+
|
| 192 |
+
Joseph Farrell and Matthew Rabin. Cheap talk. Journal of Economic Perspectives, 10(3): 103–118, September 1996. doi: 10.1257/jep.10.3.103. URL http://www.aeaweb.org/ articles?id=10.1257/jep.10.3.103.
|
| 193 |
+
|
| 194 |
+
Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to Communicate with Deep Multi-Agent Reinforcement Learning. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp. 2137–2145. Curran Associates, Inc., 2016.
|
| 195 |
+
|
| 196 |
+
Laura Harding Graesser, Kyunghyun Cho, and Douwe Kiela. Emergent linguistic phenomena in multi-agent communication games. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 3691–3701, Hong Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1384. URL https:// www.aclweb.org/anthology/D19-1384.
|
| 197 |
+
|
| 198 |
+
Serhii Havrylov and Ivan Titov. Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems 30, pp. 2149–2159. Curran Associates, Inc., 2017.
|
| 199 |
+
|
| 200 |
+
Eric Jang, Shixiang Gu, and Ben Poole. Categorical Reparameterization with GumbelSoftmax. In International Conference on Learning Representations, 2017. URL https: //openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ rkE3y85ee.
|
| 201 |
+
|
| 202 |
+
Natasha Jaques, Angeliki Lazaridou, Edward Hughes, Caglar Gulcehre, Pedro Ortega, Dj Strouse, Joel Z. Leibo, and Nando De Freitas. Social Influence as Intrinsic Motivation for Multi-Agent Deep Reinforcement Learning. In International Conference on Machine Learning, pp. 3040–3049, May 2019. URL http://proceedings.mlr.press/v97/jaques19a.html.
|
| 203 |
+
|
| 204 |
+
Simon Kirby. Iterated learning and the evolution of language. Current Opinion in Neurobiology, pp. 7, 2014.
|
| 205 |
+
|
| 206 |
+
Satwik Kottur, José Moura, Stefan Lee, and Dhruv Batra. Natural language does not emerge ‘naturally’ in multi-agent dialog. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 2962–2967, Copenhagen, Denmark, September 2017. Association for Computational Linguistics. doi: 10.18653/v1/D17-1321. URL https://www.aclweb.org/ anthology/D17-1321.
|
| 207 |
+
|
| 208 |
+
Angeliki Lazaridou, Alexander Peysakhovich, and Marco Baroni. Multi-Agent Cooperation and the Emergence of (Natural) Language. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id $\underline { { \underline { { \mathbf { \Pi } } } } }$ Hk8N3Sclg.
|
| 209 |
+
|
| 210 |
+
Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of linguistic communication from referential games with symbolic and pixel input. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id= HJGv1Z-AW.
|
| 211 |
+
|
| 212 |
+
Angeliki Lazaridou, Anna Potapenko, and Olivier Tieleman. Multi-agent communication meets natural language: Synergies between functional and structural language learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 7663–7674, Online, July 2020. Association for Computational Linguistics. URL https://www.aclweb.org/ anthology/2020.acl-main.685.
|
| 213 |
+
|
| 214 |
+
Jason Lee, Kyunghyun Cho, Jason Weston, and Douwe Kiela. Emergent translation in multiagent communication. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $_ { \cdot } =$ H1vEXaxA-.
|
| 215 |
+
|
| 216 |
+
Jason Lee, Kyunghyun Cho, and Douwe Kiela. Countering language drift via visual grounding. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 4376–4386, Hong Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1447. URL https://www.aclweb.org/anthology/D19-1447.
|
| 217 |
+
|
| 218 |
+
Adam Lerer and Alexander Peysakhovich. Learning existing social conventions via observationally augmented self-play. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pp. 107–114. ACM, 2019.
|
| 219 |
+
|
| 220 |
+
David Lewis. Convention: A philosophical study. Harvard University Press, 1969.
|
| 221 |
+
|
| 222 |
+
Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, and Dhruv Batra. Deal or no deal? endto-end learning of negotiation dialogues. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 2443–2453, Copenhagen, Denmark, September 2017. Association for Computational Linguistics. doi: 10.18653/v1/D17-1259. URL https: //www.aclweb.org/anthology/D17-1259.
|
| 223 |
+
|
| 224 |
+
Fushan Li and Michael Bowling. Ease-of-Teaching and Language Structure from Emergent Communication. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d\textquotesingle Alché-Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 15825–15835. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/9714-ease-ofteaching-and-language-structure-from-emergent-communication.pdf.
|
| 225 |
+
|
| 226 |
+
Michael L Littman. Markov games as a framework for multi-agent reinforcement learning. In International Conference on Machine Learning, volume 157, pp. 157–163, 1994.
|
| 227 |
+
|
| 228 |
+
Ryan Lowe, YI WU, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. MultiAgent Actor-Critic for Mixed Cooperative-Competitive Environments. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems 30, pp. 6379–6390. Curran Associates, Inc., 2017. URL http://papers.nips.cc/paper/7217-multi-agent-actor-critic-formixed-cooperative-competitive-environments.pdf.
|
| 229 |
+
|
| 230 |
+
Ryan Lowe, Jakob Foerster, Y-Lan Boureau, Joelle Pineau, and Yann Dauphin. On the pitfalls of measuring emergent communication. In Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems, AAMAS ’19, pp. 693–701. International Foundation for Autonomous Agents and Multiagent Systems, 2019. ISBN 978-1-4503-6309-9. URL http: //www.ifaamas.org/Proceedings/aamas2019/pdfs/p693.pdf.
|
| 231 |
+
|
| 232 |
+
Chris J. Maddison, Andriy Mnih, and Yee Whye Teh. The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id $\scriptstyle 1 = \ S 1$ jE5L5gl.
|
| 233 |
+
|
| 234 |
+
Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent populations. In AAAI Conference on Artificial Intelligence, 2018. URL https://aaai.org/ ocs/index.php/AAAI/AAAI18/paper/view/17007.
|
| 235 |
+
|
| 236 |
+
Mathijs Mul, Diane Bouchacourt, and Elia Bruni. Mastering emergent language: learning to guide in simulated navigation. arXiv:1908.05135 [cs], August 2019. arXiv: 1908.05135.
|
| 237 |
+
|
| 238 |
+
Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. URL https://d4mucfpksywv.cloudfront.net/better-language-models/ language_models_are_unsupervised_multitask_learners.pdf.
|
| 239 |
+
|
| 240 |
+
Limor Raviv and Inbal Arnon. Systematicity, but not compositionality: Examining the emergence of linguistic structure in children and adults using iterated learning. Cognition, 181:160–173, December 2018. ISSN 0010-0277.
|
| 241 |
+
|
| 242 |
+
Cinjon Resnick\*, Abhinav Gupta\*, Jakob N. Foerster, Andrew M. Dai, and Kyunghyun Cho. Capacity, bandwidth, and compositionality in emergent language learning. ArXiv, abs/1910.11424, 2019.
|
| 243 |
+
|
| 244 |
+
Andrei A Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy distillation. In International Conference on Learning Representations, 2016. URL https: //arxiv.org/pdf/1511.06295.pdf.
|
| 245 |
+
|
| 246 |
+
Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Individualized controlled continuous communication model for multiagent cooperative and competitive tasks. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id= rye7knCqK7.
|
| 247 |
+
|
| 248 |
+
Kenny Smith, Henry Brighton, and Simon Kirby. Complex Systems In Language Evolution: The Cultural Emergence Of Compositional Structure. Advances in Complex Systems (ACS), 6(04): 537–558, 2003. doi: 10.1142/S0219525903001055. URL https://ideas.repec.org/a/ wsi/acsxxx/v06y2003i04ns0219525903001055.html.
|
| 249 |
+
|
| 250 |
+
Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backpropagation. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp. 2244–2252. Curran Associates, Inc., 2016. URL http://papers.nips.cc/paper/6398-learning-multiagentcommunication-with-backpropagation.pdf.
|
| 251 |
+
|
| 252 |
+
Olivier Tieleman, Angeliki Lazaridou, Shibl Mourad, Charles Blundell, and Doina Precup. Shaping representations through communication. 2018. URL https://openreview.net/pdf?id $\underline { { \underline { { \mathbf { \Pi } } } } } =$ HkzL4hR9Ym.
|
| 253 |
+
|
| 254 |
+
# A HYPERPARAMETERS
|
| 255 |
+
|
| 256 |
+
We provide hyperparameter details in Table 1.
|
| 257 |
+
|
| 258 |
+
Table 1: Hyperparameters considered in S2P training.
|
| 259 |
+
|
| 260 |
+
<table><tr><td>Hyperparameter</td><td>Values</td></tr><tr><td>Learning rate</td><td>1e-2,1e-3,2e-3, 6e-3,1e-4, 5e-4,6e-4</td></tr><tr><td>Model architecture</td><td>Linear, Bilinear, Non-Linear</td></tr><tr><td>Number of encoders (perfect emcomm)</td><td>1,2,5,10,20,50,100,200,500,1000</td></tr><tr><td>Hidden layer size (Linear)</td><td>200,500,1000</td></tr><tr><td>Number of encoders (Pop-S2P)</td><td>20,40,50,60,80,100</td></tr><tr><td>Number of distractors</td><td>1,4,9</td></tr><tr><td>GRU hidden size Word embedding size</td><td>256</td></tr><tr><td>Image embedding size (from pretrained Resnet50)</td><td>512 2048</td></tr><tr><td>Batch size</td><td>1,512,1000</td></tr><tr><td>Random seeds</td><td></td></tr><tr><td>Optimizer</td><td>0,1, 2,3,4</td></tr><tr><td>Dropout</td><td>Adam, SGD</td></tr><tr><td>Gumbel relaxation temperature</td><td>0, 0.3</td></tr><tr><td>Vocabulary size</td><td>1</td></tr><tr><td>Max sentence length</td><td>100,200,500,1000,5000</td></tr><tr><td>m in sched</td><td>12,15,20,30, 50</td></tr><tr><td>lin sched</td><td>0,1, 30,40, 50, 70</td></tr><tr><td></td><td>0,30,40,50</td></tr><tr><td>qin rand</td><td>0.75</td></tr><tr><td>rinsched_rand_frz</td><td>0.5</td></tr><tr><td>Number of initial supervised steps (pretraining)</td><td>0,1000,2000,3000,5000</td></tr></table>
|
| 261 |
+
|
| 262 |
+
# B CALCULATION OF OPTIMAL SAMPLE COMPLEXITY IN OR GAME
|
| 263 |
+
|
| 264 |
+
Here we provide a quick calculation for how quickly a human might learn a new compositional language $L$ in the OR game in as few examples as possible, which we use as a baseline in Figure 4a. We assume a OR game with $p = 6$ properties, $t = 1 0$ types, $T = 6$ words sent per message (concatenated together), and $| V | = 6 0$ vocabulary size. If this language $L$ is compositional, then each word in the vocabulary is assigned to 1 type. Thus, we need to learn 60 total assignments. In this analysis we assume we can construct (i.e. hand-design) the samples seen by the human, and thus the final number should be considered something like a lower bound.
|
| 265 |
+
|
| 266 |
+
Since $T = 6$ , we get information about 6 word type assignments for every sample. However, this information is entangled as we don’t know which word corresponded to which type. Thus, we (1) divide the problem up by first constructing 9 (word sequence, object) sample pairs where none of the object types overlap between each sample. With this information, we are able to narrow down the word $\gets$ type assignments into 10 groups of 6 (that is, in each group we have 6 words corresponding to 6 types, but we don’t know which type belongs to which word). Note we don’t need 10 samples as the last one can be inferred by exclusion. (2) We then construct 5 more samples where each type belongs to a separate group. We can do this because $t > p$ . Because each type belongs to a separate group, cross-referencing the words observed from samples in (1) and (2) uniquely defines each word $\gets$ type assignment. Note again we don’t need 6 samples as the last one can be inferred by exclusion. This gives us a total of $9 + 5 = 1 4$ samples.
|
| 267 |
+
|
| 268 |
+
# C ADDITIONAL PLOTS
|
| 269 |
+
|
| 270 |
+
We show training curves for various S2P schedules.
|
| 271 |
+
|
| 272 |
+

|
| 273 |
+
Figure 8: Training curves for various S2P methods in the IBR game described in $\ S 4$
|
parse/train/rJxGLlBtwH/rJxGLlBtwH_content_list.json
ADDED
|
@@ -0,0 +1,1515 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "ON THE INTERACTION BETWEEN SUPERVISION AND SELF-PLAY IN EMERGENT COMMUNICATION ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
176,
|
| 8 |
+
101,
|
| 9 |
+
823,
|
| 10 |
+
146
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Ryan Lowe∗, Abhinav Gupta∗ MILA ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
184,
|
| 19 |
+
170,
|
| 20 |
+
385,
|
| 21 |
+
196
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Jakob Foerster, Douwe Kiela Facebook AI Research ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
184,
|
| 30 |
+
219,
|
| 31 |
+
387,
|
| 32 |
+
247
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Joelle Pineau Facebook AI Research MILA ",
|
| 39 |
+
"bbox": [
|
| 40 |
+
532,
|
| 41 |
+
219,
|
| 42 |
+
684,
|
| 43 |
+
261
|
| 44 |
+
],
|
| 45 |
+
"page_idx": 0
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"type": "text",
|
| 49 |
+
"text": "ABSTRACT ",
|
| 50 |
+
"text_level": 1,
|
| 51 |
+
"bbox": [
|
| 52 |
+
454,
|
| 53 |
+
297,
|
| 54 |
+
544,
|
| 55 |
+
313
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "A promising approach for teaching artificial agents to use natural language involves using human-in-the-loop training. However, recent work suggests that current machine learning methods are too data inefficient to be trained in this way from scratch. In this paper, we investigate the relationship between two categories of learning signals with the ultimate goal of improving sample efficiency: imitating human language data via supervised learning, and maximizing reward in a simulated multi-agent environment via self-play (as done in emergent communication), and introduce the term supervised self-play $( S 2 P )$ for algorithms using both of these signals. We find that first training agents via supervised learning on human data followed by self-play outperforms the converse, suggesting that it is not beneficial to emerge languages from scratch. We then empirically investigate various S2P schedules that begin with supervised learning in two environments: a Lewis signaling game with symbolic inputs, and an image-based referential game with natural language descriptions. Lastly, we introduce population based approaches to S2P, which further improves the performance over single-agent methods.1 ",
|
| 62 |
+
"bbox": [
|
| 63 |
+
233,
|
| 64 |
+
329,
|
| 65 |
+
766,
|
| 66 |
+
536
|
| 67 |
+
],
|
| 68 |
+
"page_idx": 0
|
| 69 |
+
},
|
| 70 |
+
{
|
| 71 |
+
"type": "text",
|
| 72 |
+
"text": "1 INTRODUCTION ",
|
| 73 |
+
"text_level": 1,
|
| 74 |
+
"bbox": [
|
| 75 |
+
176,
|
| 76 |
+
560,
|
| 77 |
+
336,
|
| 78 |
+
577
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Language is one of the most important aspects of human intelligence; it allows humans to coordinate and share knowledge with each other. It is also crucial for human-machine interaction, as human language is a natural means by which to exchange information, give feedback, and specify goals. A promising approach for training agents to solve problems with natural language is to have a “human in the loop”, meaning we collect problem-specific data from humans interacting directly with our agents for learning. However, human-in-the-loop data is expensive and time-consuming to obtain as it requires continuously collecting human data as the agent’s policy improves, and recent work suggests that current machine learning methods (e.g. from deep reinforcement learning) are too data-inefficient to be trained in this way from scratch (Chevalier-Boisvert et al., 2019). Thus, an important open problem is: how can we make human-in-the-loop training as data efficient as possible? ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
592,
|
| 88 |
+
825,
|
| 89 |
+
731
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "text",
|
| 95 |
+
"text": "To maximize data efficiency, it is important to fully leverage all available training signals. In this paper, we study two categories of such training methods: imitating human data via supervised learning, and self-play to maximize reward in a multi-agent environment, both of which provide rich signals for endowing agents with language-using capabilities. However, these are potentially competing objectives, as maximizing environmental reward can lead to the resulting communication protocol drifting from natural language (Lewis et al., 2017; Lee et al., 2019). The crucial question, then, is how do we best combine self-play and supervised updates? This question has received surprisingly little attention from the emergent communication literature, where the question of how to bridge the gap from emergent protocols to natural language is generally left for future work (Mordatch & Abbeel, 2018; Lazaridou et al., 2018; Cao et al., 2018). ",
|
| 96 |
+
"bbox": [
|
| 97 |
+
174,
|
| 98 |
+
737,
|
| 99 |
+
825,
|
| 100 |
+
876
|
| 101 |
+
],
|
| 102 |
+
"page_idx": 0
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"type": "image",
|
| 106 |
+
"img_path": "images/703108bb01f55e571503ecb2e42a48453f06610298cc29d52ebbedd2e3d38490.jpg",
|
| 107 |
+
"image_caption": [
|
| 108 |
+
"Figure 1: (a) Diagram of the supervised self-play (S2P) procedure (phases 1-3) and the testing procedure considered in this work (phase 4). (b) The environments considered in this paper (Sec. 4). "
|
| 109 |
+
],
|
| 110 |
+
"image_footnote": [],
|
| 111 |
+
"bbox": [
|
| 112 |
+
181,
|
| 113 |
+
101,
|
| 114 |
+
823,
|
| 115 |
+
257
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "Our goal in this paper is to investigate algorithms for combining supervised learning with self-play — which we call supervised self-play (S2P) algorithms — using two classic emergent communication tasks: a Lewis signaling game with symbolic inputs, and a more complicated image-based referential game with natural language descriptions. Our first finding is that supervised learning followed by self-play outperforms emergent communication with supervised fine-tuning in these environments, and we provide three reasons for why this is the case. We then empirically investigate several supervised-first S2P methods in our environments. Existing approaches in this area have used various ad-hoc schedules for alternating between the two kinds of updates (Lazaridou et al., 2017), but to our knowledge there has been no systematic study that has compared these approaches. Lastly, we propose the use of population-based methods for S2P, and find that it leads to improved performance in the more challenging image-based referential game. Our findings highlight the need for further work in combining supervised learning and self-play to develop more sample-efficient language learners. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
323,
|
| 125 |
+
825,
|
| 126 |
+
502
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "2 RELATED WORK ",
|
| 133 |
+
"text_level": 1,
|
| 134 |
+
"bbox": [
|
| 135 |
+
176,
|
| 136 |
+
523,
|
| 137 |
+
341,
|
| 138 |
+
539
|
| 139 |
+
],
|
| 140 |
+
"page_idx": 1
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"type": "text",
|
| 144 |
+
"text": "In the past few years, there has been a renewed interest in the field of emergent communication (Sukhbaatar et al., 2016; Foerster et al., 2016; Lazaridou et al., 2017; Havrylov & Titov, 2017) culminating in 3 NeurIPS workshops. Empirical studies have showed that agents can autonomously evolve a communication protocol using discrete symbols when deployed in a multi-agent environment which helps them to play a cooperative or competitive game (Singh et al., 2019; Cao et al., 2018; Choi et al., 2018; Resnick\\* et al., 2019; Evtimova et al., 2018). ",
|
| 145 |
+
"bbox": [
|
| 146 |
+
174,
|
| 147 |
+
554,
|
| 148 |
+
825,
|
| 149 |
+
638
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 1
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "While the idea of promoting coordination among agents through communication sounds promising, recent experiments (Lowe et al., 2019; Chaabouni et al., 2019; Kottur et al., 2017; Jaques et al., 2019) have emphasized the difficulty in learning meaningful emergent communication protocols even with centralized training. ",
|
| 156 |
+
"bbox": [
|
| 157 |
+
174,
|
| 158 |
+
645,
|
| 159 |
+
825,
|
| 160 |
+
702
|
| 161 |
+
],
|
| 162 |
+
"page_idx": 1
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"type": "text",
|
| 166 |
+
"text": "Apart from the above advances in emergent communication, there has been a long outstanding goal of learning intelligent conversational agents to be able to interact with humans. This involves training the artificial agents in a way so that they achieve high scores while solving the task and their language is interpretable by humans or close to natural language. Recent works also add another axis orthogonal to communication where the agent also takes a discrete action in an interactive environment (de Vries et al., 2018; Mul et al., 2019). Lewis et al. (2017) introduced a negotiation task which involves learning linguistic and reasoning skills. They train models imitating human utterances using supervised learning and found that the model generated human-like captions but were poor negotiators. So they perform self-play with these pretrained agents in an interleaved manner and found that the performance improved drastically while avoiding language drift. Lee et al. (2019) also propose using an auxiliary task for grounding the communication to counter language drift. They use visual grounding to learn the semantics of the language while still generating messages that are close to English. ",
|
| 167 |
+
"bbox": [
|
| 168 |
+
174,
|
| 169 |
+
708,
|
| 170 |
+
825,
|
| 171 |
+
888
|
| 172 |
+
],
|
| 173 |
+
"page_idx": 1
|
| 174 |
+
},
|
| 175 |
+
{
|
| 176 |
+
"type": "text",
|
| 177 |
+
"text": "A recent trend on using population based training for multi-agent communication is a promising avenue for research using inspirations from language evolution literature (Smith et al., 2003; Kirby, ",
|
| 178 |
+
"bbox": [
|
| 179 |
+
174,
|
| 180 |
+
895,
|
| 181 |
+
823,
|
| 182 |
+
924
|
| 183 |
+
],
|
| 184 |
+
"page_idx": 1
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"type": "text",
|
| 188 |
+
"text": "2014; Raviv & Arnon, 2018). Cultural transmission is one such technique which focuses on the structure and compression of languages, since a language must be used and learned by all individuals of the culture in which it resides and at the same time be suitable for a variety of tasks. Harding Graesser et al. (2019) shows the emergence of linguistic phenomena when a pool of agents contact each other giving rise to novel creole languages. Li & Bowling (2019); Cogswell et al. (2019); Tieleman et al. (2018) have also tried different ways of imposing cultural pressures on agents, by simulating a large population of them and pairing agents to solve a cooperative game with communication. They train the agent against a sampled generation of agents where the generation corresponds to the different languages of the different agent at different times in the history. ",
|
| 189 |
+
"bbox": [
|
| 190 |
+
174,
|
| 191 |
+
103,
|
| 192 |
+
825,
|
| 193 |
+
229
|
| 194 |
+
],
|
| 195 |
+
"page_idx": 2
|
| 196 |
+
},
|
| 197 |
+
{
|
| 198 |
+
"type": "text",
|
| 199 |
+
"text": "Our work is inspired from these works where we aim to formalize the recent advancements in using self-play in dialog modeling, through the lens of emergent communication. ",
|
| 200 |
+
"bbox": [
|
| 201 |
+
176,
|
| 202 |
+
236,
|
| 203 |
+
823,
|
| 204 |
+
263
|
| 205 |
+
],
|
| 206 |
+
"page_idx": 2
|
| 207 |
+
},
|
| 208 |
+
{
|
| 209 |
+
"type": "text",
|
| 210 |
+
"text": "3 METHODS ",
|
| 211 |
+
"text_level": 1,
|
| 212 |
+
"bbox": [
|
| 213 |
+
174,
|
| 214 |
+
284,
|
| 215 |
+
290,
|
| 216 |
+
299
|
| 217 |
+
],
|
| 218 |
+
"page_idx": 2
|
| 219 |
+
},
|
| 220 |
+
{
|
| 221 |
+
"type": "text",
|
| 222 |
+
"text": "3.1 PROBLEM DEFINITION ",
|
| 223 |
+
"text_level": 1,
|
| 224 |
+
"bbox": [
|
| 225 |
+
176,
|
| 226 |
+
314,
|
| 227 |
+
369,
|
| 228 |
+
328
|
| 229 |
+
],
|
| 230 |
+
"page_idx": 2
|
| 231 |
+
},
|
| 232 |
+
{
|
| 233 |
+
"type": "text",
|
| 234 |
+
"text": "Our agents are embedded in a multi-agent environment with $N$ agents where they receive observations $o \\in O$ (which are functions of a hidden state $S$ ) and perform actions $a \\in A$ . Some actions $A _ { L } \\subset A$ involve sending a message $m \\in A _ { L }$ over a discrete, costless communication channel (i.e. a cheap talk channel (Farrell & Rabin, 1996)). The agents are rewarded with a reward $r \\in R$ for their performance in the environment. We assume throughout that the environment is cooperative and thus the agents are trained to maximize the sum of rewards $\\begin{array} { r } { R = \\sum _ { t = 1 : T } \\sum _ { i = 1 : N } r _ { i , t } } \\end{array}$ across both agents. This can be thought of as a cooperative partially-observable Markov game (Littman (1994)). ",
|
| 235 |
+
"bbox": [
|
| 236 |
+
174,
|
| 237 |
+
340,
|
| 238 |
+
825,
|
| 239 |
+
438
|
| 240 |
+
],
|
| 241 |
+
"page_idx": 2
|
| 242 |
+
},
|
| 243 |
+
{
|
| 244 |
+
"type": "text",
|
| 245 |
+
"text": "We define a target language $L ^ { * } \\in { \\mathcal { L } }$ , usually corresponding to natural language, that we want our agents to learn (we further assume $L ^ { * }$ can be used to achieve high task reward). In this paper, we consider a language $L \\in { \\mathcal { L } }$ to be simply a set of valid messages $A _ { L }$ and a mapping between observations and messages in the environment, $L : O \\times A _ { L } \\mapsto [ 0 , 1 ]$ . For example, in an English image-based referential game (Section 4) this corresponds to the mapping between images and image descriptions in English. We are given a dataset $\\mathcal { D }$ consisting of $| \\mathcal D |$ (observation, action) pairs, corresponding to $N _ { e }$ ‘experts’ (for us, $N _ { e } = 2$ ) playing the game using the target language $L ^ { * }$ . Our goal is to train agents to achieve a high reward in the game while speaking language $L ^ { * }$ with an ‘expert’. Specifically, we want our agents to generalize and to perform well on examples that are not contained in $\\mathcal { D }$ . ",
|
| 246 |
+
"bbox": [
|
| 247 |
+
174,
|
| 248 |
+
444,
|
| 249 |
+
825,
|
| 250 |
+
583
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 2
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "text",
|
| 256 |
+
"text": "To summarize, we want agents that can perform well on a collaborative task with English-speaking humans, and we can train them using a supervised dataset $\\mathcal { D }$ and via self-play. ",
|
| 257 |
+
"bbox": [
|
| 258 |
+
173,
|
| 259 |
+
590,
|
| 260 |
+
821,
|
| 261 |
+
618
|
| 262 |
+
],
|
| 263 |
+
"page_idx": 2
|
| 264 |
+
},
|
| 265 |
+
{
|
| 266 |
+
"type": "text",
|
| 267 |
+
"text": "3.2 SUPERVISED SELF-PLAY (S2P) ",
|
| 268 |
+
"text_level": 1,
|
| 269 |
+
"bbox": [
|
| 270 |
+
176,
|
| 271 |
+
635,
|
| 272 |
+
431,
|
| 273 |
+
648
|
| 274 |
+
],
|
| 275 |
+
"page_idx": 2
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"type": "text",
|
| 279 |
+
"text": "In recent years, there have been several approaches to language learning that have combined supervised or imitation learning with self-play. In this paper, we propose an umbrella term for these algorithms called supervised self-play (S2P). S2P requires two things: (1) a multi-agent environment where at least one agent can send messages over a dedicated communication channel, along with a reward function that measures how well the agents are doing at some task; and (2) a supervised dataset $\\mathcal { D }$ of experts acting and speaking language $L ^ { * }$ in the environment (such that they perform well on the task). Given these ingredients, we define S2P below (see Figure 2). ",
|
| 280 |
+
"bbox": [
|
| 281 |
+
174,
|
| 282 |
+
660,
|
| 283 |
+
825,
|
| 284 |
+
758
|
| 285 |
+
],
|
| 286 |
+
"page_idx": 2
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"type": "text",
|
| 290 |
+
"text": "Definition 3.1. Supervised self-play (S2P). Supervised self-play is a class of language learning algorithms that combines: (1) self-play updates in a multi-agent language environment, and (2) supervised updates on an expert dataset $\\mathcal { D }$ . ",
|
| 291 |
+
"bbox": [
|
| 292 |
+
174,
|
| 293 |
+
761,
|
| 294 |
+
825,
|
| 295 |
+
803
|
| 296 |
+
],
|
| 297 |
+
"page_idx": 2
|
| 298 |
+
},
|
| 299 |
+
{
|
| 300 |
+
"type": "text",
|
| 301 |
+
"text": "S2P algorithms can differ in how they combine self-play and supervised learning updates on $\\mathcal { D }$ . When supervised learning is performed before self-play, we refer to the dataset $\\mathcal { D }$ as the seed data. Why might we want to train our agents via self-play? Won’t their language diverge from $L ^ { * } ?$ One way to intuitively understand why S2P is beneficial is to think in terms of constraints. In our set-up, there are two known constraints on the target language $L ^ { * }$ : (1) it is consistent with the samples from the supervised dataset $\\mathcal { D }$ , and (2) $L ^ { * }$ can be used to obtain a high reward in the environment. Thus, finding $L ^ { * }$ can be loosely viewed as a constrained optimization problem, and enforcing both constraints should clearly lead to better performance. ",
|
| 302 |
+
"bbox": [
|
| 303 |
+
174,
|
| 304 |
+
811,
|
| 305 |
+
825,
|
| 306 |
+
924
|
| 307 |
+
],
|
| 308 |
+
"page_idx": 2
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"type": "text",
|
| 312 |
+
"text": "3.3 ALGORITHMS FOR S2P ",
|
| 313 |
+
"text_level": 1,
|
| 314 |
+
"bbox": [
|
| 315 |
+
176,
|
| 316 |
+
103,
|
| 317 |
+
375,
|
| 318 |
+
117
|
| 319 |
+
],
|
| 320 |
+
"page_idx": 3
|
| 321 |
+
},
|
| 322 |
+
{
|
| 323 |
+
"type": "text",
|
| 324 |
+
"text": "Here we describe several methods for S2P training. Our goal is not to exhaustively enumerate all possible optimization strategies, but rather provide a categorization of some well-known ways to combine self-play and supervised learning. To help describe these methods, we further split the seed dataset $\\mathcal { D }$ into $\\mathcal { D } _ { t r a i n }$ , which is used for training, and $\\mathcal { D } _ { v a l }$ which is used for early-stopping. We also visualize the schedules in Figure 2. ",
|
| 325 |
+
"bbox": [
|
| 326 |
+
173,
|
| 327 |
+
131,
|
| 328 |
+
823,
|
| 329 |
+
200
|
| 330 |
+
],
|
| 331 |
+
"page_idx": 3
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"type": "text",
|
| 335 |
+
"text": "Emergent communication with supervised fine-tuning (sp2sup): We first perform self-play updates until the learning converges on the task performance. It is then followed by supervised updates on $\\mathcal { D } _ { t r a i n }$ until the listener performance converges on the dataset $\\mathcal { D } _ { v a l }$ . ",
|
| 336 |
+
"bbox": [
|
| 337 |
+
174,
|
| 338 |
+
208,
|
| 339 |
+
614,
|
| 340 |
+
277
|
| 341 |
+
],
|
| 342 |
+
"page_idx": 3
|
| 343 |
+
},
|
| 344 |
+
{
|
| 345 |
+
"type": "text",
|
| 346 |
+
"text": "Supervised learning with self-play (sup2sp): This is the complement of the above method which involves doing supervised updates until convergence on $\\mathcal { D } _ { v a l }$ followed by self-play updates until convergence on the task performance. ",
|
| 347 |
+
"bbox": [
|
| 348 |
+
173,
|
| 349 |
+
284,
|
| 350 |
+
614,
|
| 351 |
+
340
|
| 352 |
+
],
|
| 353 |
+
"page_idx": 3
|
| 354 |
+
},
|
| 355 |
+
{
|
| 356 |
+
"type": "image",
|
| 357 |
+
"img_path": "images/ef8b2045c855de037929134bc3d5a2d2107d04c2c4940ae917df37b304f53b77.jpg",
|
| 358 |
+
"image_caption": [
|
| 359 |
+
"Figure 2: A visual representation of the different S2P methods. "
|
| 360 |
+
],
|
| 361 |
+
"image_footnote": [],
|
| 362 |
+
"bbox": [
|
| 363 |
+
625,
|
| 364 |
+
224,
|
| 365 |
+
818,
|
| 366 |
+
366
|
| 367 |
+
],
|
| 368 |
+
"page_idx": 3
|
| 369 |
+
},
|
| 370 |
+
{
|
| 371 |
+
"type": "text",
|
| 372 |
+
"text": "Random updates $( \\tt r a n d )$ : This is the method used in (Lazaridou et al., 2017). At each time step, we sample a bernoulli random variable $z \\sim B e r n o u l l i ( q )$ where $q$ is fixed. If $z = 1$ , we perform one supervised update, else we do one self-play update, and repeat until both losses convergence on $\\mathcal { D } _ { v a l }$ . ",
|
| 373 |
+
"bbox": [
|
| 374 |
+
173,
|
| 375 |
+
347,
|
| 376 |
+
612,
|
| 377 |
+
417
|
| 378 |
+
],
|
| 379 |
+
"page_idx": 3
|
| 380 |
+
},
|
| 381 |
+
{
|
| 382 |
+
"type": "text",
|
| 383 |
+
"text": "Scheduled updates (sched): We first pretrain the listener and the speaker until convergence on $\\mathcal { D } _ { v a l }$ . Then we create a schedule, where we perform $l$ self-play updates followed by $m$ supervised updates, and repeat until convergence on the dataset. ",
|
| 384 |
+
"bbox": [
|
| 385 |
+
173,
|
| 386 |
+
424,
|
| 387 |
+
821,
|
| 388 |
+
467
|
| 389 |
+
],
|
| 390 |
+
"page_idx": 3
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"type": "text",
|
| 394 |
+
"text": "Scheduled updates with speaker freezing (sched_ $\\pmb { \\mathcal { \\tt { f r } } } \\mathbf { z }$ ): This method is based on the findings of Lewis et al. (2017), who do sched S2P while freezing the parameters of the speaker during self-play to reduce the amount of language drift. In our case, we freeze the parameters of the speaker after the initial supervised learning. ",
|
| 395 |
+
"bbox": [
|
| 396 |
+
173,
|
| 397 |
+
473,
|
| 398 |
+
825,
|
| 399 |
+
530
|
| 400 |
+
],
|
| 401 |
+
"page_idx": 3
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"type": "text",
|
| 405 |
+
"text": "Scheduled updates with random speaker freezing (sched_rand_frz): Experimentally, we noticed that sched_frz didn’t perform well in self-play. Thus, we introduce a variation, we sample a bernoulli random variable $z \\sim B e r n o u l l i ( r )$ where $r$ is fixed. If $z = 1$ , we freeze the parameters of the speaker during both self-play and supervised learning, else we allow updates to the speaker as well. ",
|
| 406 |
+
"bbox": [
|
| 407 |
+
173,
|
| 408 |
+
536,
|
| 409 |
+
825,
|
| 410 |
+
604
|
| 411 |
+
],
|
| 412 |
+
"page_idx": 3
|
| 413 |
+
},
|
| 414 |
+
{
|
| 415 |
+
"type": "text",
|
| 416 |
+
"text": "3.4 POPULATION-BASED S2P (POP-S2P) ",
|
| 417 |
+
"text_level": 1,
|
| 418 |
+
"bbox": [
|
| 419 |
+
178,
|
| 420 |
+
625,
|
| 421 |
+
468,
|
| 422 |
+
640
|
| 423 |
+
],
|
| 424 |
+
"page_idx": 3
|
| 425 |
+
},
|
| 426 |
+
{
|
| 427 |
+
"type": "text",
|
| 428 |
+
"text": "As explained above, the goal of S2P is to produce agents that follow dataset $\\mathcal { D }$ while maximizing reward in the environment. However, there are many such policies satisfying these criteria. This results in a large space of possible solutions, that increases as the environment grows more complex (but decreases with increasing $| \\mathcal { D } | )$ . Experimentally, we find that this can result in diverse agent policies. We show this in Figure 3 by training 50 randomly initialized agents on the image-based referential game (defined in Sec. 4) the agents can often make diverse predictions for a given image (Figure 3a) and achieve variable performance when playing with other populations with a slight preference towards their own partner (the diagonal in Figure 3b). ",
|
| 429 |
+
"bbox": [
|
| 430 |
+
174,
|
| 431 |
+
654,
|
| 432 |
+
485,
|
| 433 |
+
888
|
| 434 |
+
],
|
| 435 |
+
"page_idx": 3
|
| 436 |
+
},
|
| 437 |
+
{
|
| 438 |
+
"type": "image",
|
| 439 |
+
"img_path": "images/ee255ee673b36ee44a5471ef255777e187d1b7d72ab110c096a5503db0b32c5f.jpg",
|
| 440 |
+
"image_caption": [
|
| 441 |
+
"Figure 3: Results from training $5 0 \\ : \\mathrm { S 2 P }$ agents on the IBR game with $| \\mathcal { D } | = 1 0 0 0 0$ . (a) The agents have a range of predictions on many images. (b) When playing with each other, the agents exhibit uneven performance (color is mean reward, yellow is higher), indicating policy variability. "
|
| 442 |
+
],
|
| 443 |
+
"image_footnote": [],
|
| 444 |
+
"bbox": [
|
| 445 |
+
498,
|
| 446 |
+
618,
|
| 447 |
+
818,
|
| 448 |
+
805
|
| 449 |
+
],
|
| 450 |
+
"page_idx": 3
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"type": "text",
|
| 454 |
+
"text": "ment S2P by training a population of $N$ agents, and subsequently aggregating them back into a single agent (the ‘student’). We call this population-based S2P (Pop-S2P). While there are many feasible ways of doing this, in this paper we train the populations by simply randomizing the initial seed, and we aggregate the populations using a simple form of policy distillation (Rusu et al., 2016). Another simple technique to boost performance is via ensembling where we simply take the majority prediction at each time step. ",
|
| 455 |
+
"bbox": [
|
| 456 |
+
174,
|
| 457 |
+
910,
|
| 458 |
+
823,
|
| 459 |
+
924
|
| 460 |
+
],
|
| 461 |
+
"page_idx": 3
|
| 462 |
+
},
|
| 463 |
+
{
|
| 464 |
+
"type": "text",
|
| 465 |
+
"text": "",
|
| 466 |
+
"bbox": [
|
| 467 |
+
174,
|
| 468 |
+
103,
|
| 469 |
+
825,
|
| 470 |
+
174
|
| 471 |
+
],
|
| 472 |
+
"page_idx": 4
|
| 473 |
+
},
|
| 474 |
+
{
|
| 475 |
+
"type": "text",
|
| 476 |
+
"text": "4 ENVIRONMENTS & IMPLEMENTATION DETAILS ",
|
| 477 |
+
"text_level": 1,
|
| 478 |
+
"bbox": [
|
| 479 |
+
174,
|
| 480 |
+
194,
|
| 481 |
+
594,
|
| 482 |
+
209
|
| 483 |
+
],
|
| 484 |
+
"page_idx": 4
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"type": "text",
|
| 488 |
+
"text": "We consider environments based on classical problems in emergent communication. These environments are cooperative and involve the interaction between a speaker, who makes an observation and sends a message, and a listener, who observes the message and makes a prediction (see Figure 1b). Our goal is to train a listener such that it achieves high reward when playing with an expert speaking the target language $L ^ { * }$ on inputs unseen during training.2 ",
|
| 489 |
+
"bbox": [
|
| 490 |
+
174,
|
| 491 |
+
223,
|
| 492 |
+
826,
|
| 493 |
+
294
|
| 494 |
+
],
|
| 495 |
+
"page_idx": 4
|
| 496 |
+
},
|
| 497 |
+
{
|
| 498 |
+
"type": "text",
|
| 499 |
+
"text": "Environment 1: Object Reconstruction (OR) Our first game is a Lewis signaling game (Lewis, 1969) and a simpler version of the Task & Talk game from Kottur et al. (2017), with a single turn and a much larger input space. The speaker agent observes an object with a certain set of properties, and must describe the object to the listener using a sequence of words. The listener then attempts to reconstruct the object. More specifically, the input space consists of $p$ properties (e.g. shape, color) of $t$ types each (e.g. triangle, square). The speaker observes a symbolic representation of the input, consisting of the concatenation of $p = 6$ one-hot vectors, each of length $t = 1 0$ . The number of possible inputs scales as $t ^ { p }$ . We define the vocabulary size (length of each one-hot vector sent from the speaker) as $| V | = 6 0$ , and the number of words (fixed length message) sent to be $T = 6$ . ",
|
| 500 |
+
"bbox": [
|
| 501 |
+
174,
|
| 502 |
+
308,
|
| 503 |
+
825,
|
| 504 |
+
434
|
| 505 |
+
],
|
| 506 |
+
"page_idx": 4
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"type": "text",
|
| 510 |
+
"text": "For our target language $L ^ { * }$ for this task, we programatically generate a perfectly compositional language, by assigning each object a unique word. In other words, to describe a ‘blue shaded triangle’, we create a language where the output description would be “blue, triangle, shaded”, in some arbitrary order. By ‘unique symbol’, we mean that no two properties are assigned the same word. The speaker and listener policies are parameterized using a 2-layer linear network (results were similar with added non-linearity and significantly worse with 1-layer linear networks) with 200 hidden units. During both supervised learning and self-play, the listener is trained to minimize the cross-entropy loss over property predictions. ",
|
| 511 |
+
"bbox": [
|
| 512 |
+
173,
|
| 513 |
+
440,
|
| 514 |
+
825,
|
| 515 |
+
553
|
| 516 |
+
],
|
| 517 |
+
"page_idx": 4
|
| 518 |
+
},
|
| 519 |
+
{
|
| 520 |
+
"type": "text",
|
| 521 |
+
"text": "Environment 2: Image-Based Referential game with natural language (IBR) Our second game is the communication task introduced in Lee et al. (2018). The speaker observes a target image $d ^ { * }$ , and must describe the image using a set of words. The listener observes the target image along with $D$ distractor images sampled uniformly at random from the training set (for us, $D = 9$ ), and the message $y _ { d ^ { * } }$ from the speaker, and is rewarded for correctly selecting the target image. For this game, the target language $L ^ { * }$ is English — we obtain English image descriptions using caption data from MS COCO and Flickr30k. We set the vocabulary size $| V | = 1 0 0$ , and filter out any descriptions that contain more than $30 \\%$ unknown tokens while keeping the maximum message length $T$ to 15. ",
|
| 522 |
+
"bbox": [
|
| 523 |
+
173,
|
| 524 |
+
566,
|
| 525 |
+
825,
|
| 526 |
+
678
|
| 527 |
+
],
|
| 528 |
+
"page_idx": 4
|
| 529 |
+
},
|
| 530 |
+
{
|
| 531 |
+
"type": "text",
|
| 532 |
+
"text": "Similar to (Mordatch & Abbeel, 2018; Sukhbaatar et al., 2016), we train our agents end-to-end with backpropagation. Since the speaker sends discrete messages, we use the Straight-Through version of Gumbel-Softmax (Jang et al., 2017; Maddison et al., 2017) to allow gradient flow to the speaker during self-play $\\left( \\mathcal { I } _ { \\mathrm { s e l f - p l a y } } \\right)$ . The speaker’s predictions are trained on the ground truth English captions $m ^ { * }$ using the cross entropy loss $\\mathcal { I } _ { \\mathrm { s p k } }$ -supervised. The listener is trained using the cross-entropy loss $\\mathcal { I } _ { \\mathrm { l s n } }$ -supervised where the logits are the reciprocal of the mean squared error which was found to perform better than directly minimizing MSE loss in Lee et al. (2018). The mean squared error is taken over the listener’s image representation $b _ { l s n }$ of the distractor (or target) image and the message representation given as input. The loss functions are defined as: ",
|
| 533 |
+
"bbox": [
|
| 534 |
+
173,
|
| 535 |
+
684,
|
| 536 |
+
825,
|
| 537 |
+
810
|
| 538 |
+
],
|
| 539 |
+
"page_idx": 4
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"type": "equation",
|
| 543 |
+
"img_path": "images/2dd8470b3fc164713c47d720322785b8ff27603b47f38cb4a9e246d1bb590330.jpg",
|
| 544 |
+
"text": "$$\n\\begin{array} { c } { { \\mathcal { I } _ { \\mathrm { s p k \\mathrm { s u p e r v i s e d } } } ( d ^ { * } ) = \\displaystyle - \\sum _ { t = 1 } ^ { T } \\log p _ { s p k } ( m _ { t } | m _ { < t } , d ^ { * } ) } } \\\\ { { \\mathcal { I } _ { \\mathrm { l s n \\mathrm { - } s u p e r v i s e d } } ( m ^ { * } , d ^ { * } , D ) = \\displaystyle - \\sum _ { d = 1 } ^ { D + 1 } \\log ( \\mathsf { s o f t r m a x } ( 1 / p _ { l s n } ( m ^ { * } ) - b _ { \\mathrm { l s n } } ( d ) ) ^ { 2 } ) } } \\end{array}\n$$",
|
| 545 |
+
"text_format": "latex",
|
| 546 |
+
"bbox": [
|
| 547 |
+
254,
|
| 548 |
+
813,
|
| 549 |
+
745,
|
| 550 |
+
904
|
| 551 |
+
],
|
| 552 |
+
"page_idx": 4
|
| 553 |
+
},
|
| 554 |
+
{
|
| 555 |
+
"type": "image",
|
| 556 |
+
"img_path": "images/36a2a15820b0a29e43bca933b7bae40c31f6e543e408306e2109c996d8d44e9b.jpg",
|
| 557 |
+
"image_caption": [
|
| 558 |
+
"Figure 4: (a) Left: In the OR game, best performance (number of total samples required to achieve $9 5 \\%$ test accuracy, lower is better) for S2P is achieved when all of the samples are in the seed. 0 on the $\\mathbf { X }$ -axis corresponds to sp2sup and Optimal is the actual (minimum) number of samples required to solve this optimization problem (see Appenix B). Right: This is also the case in the IBR game, where performance is measured by the generalization accuracy using $1 0 \\mathrm { k }$ total training samples (higher is better). (b) Adding more samples to initial supervised learning in the IBR game improves agents’ generalization to $L ^ { * }$ . (c) Even when we learn the perfect distribution with emergent communication in the OR game, it still performs worse than Pop-S2P (using sup2sp S2P). "
|
| 559 |
+
],
|
| 560 |
+
"image_footnote": [],
|
| 561 |
+
"bbox": [
|
| 562 |
+
174,
|
| 563 |
+
99,
|
| 564 |
+
818,
|
| 565 |
+
300
|
| 566 |
+
],
|
| 567 |
+
"page_idx": 5
|
| 568 |
+
},
|
| 569 |
+
{
|
| 570 |
+
"type": "equation",
|
| 571 |
+
"img_path": "images/e733ad66108ac11e89f6e75134c4af351a42882c2b2213d02dfc3a411edd6f8b.jpg",
|
| 572 |
+
"text": "$$\n\\mathcal { I } _ { \\mathrm { s e l f - p l a y } } ( d ^ { * } , D ) = - \\sum _ { d = 1 } ^ { D + 1 } \\log ( \\mathsf { s o f t m a x } ( 1 / p _ { l s n } ( y _ { d ^ { * } } ) - b _ { \\mathrm { l s n } } ( d ) ) ^ { 2 } )\n$$",
|
| 573 |
+
"text_format": "latex",
|
| 574 |
+
"bbox": [
|
| 575 |
+
281,
|
| 576 |
+
445,
|
| 577 |
+
717,
|
| 578 |
+
489
|
| 579 |
+
],
|
| 580 |
+
"page_idx": 5
|
| 581 |
+
},
|
| 582 |
+
{
|
| 583 |
+
"type": "text",
|
| 584 |
+
"text": "where $y _ { d ^ { * } }$ is the concatenation of $T$ one-hot vectors $y _ { d ^ { * } } ^ { t } = \\mathtt { S T - G u m b e l S o f t m a x } ( p _ { s p k } ^ { t } ) .$ ",
|
| 585 |
+
"bbox": [
|
| 586 |
+
171,
|
| 587 |
+
500,
|
| 588 |
+
763,
|
| 589 |
+
516
|
| 590 |
+
],
|
| 591 |
+
"page_idx": 5
|
| 592 |
+
},
|
| 593 |
+
{
|
| 594 |
+
"type": "text",
|
| 595 |
+
"text": "We use the same architecture as described in Lee et al. (2018). The speaker and listener are parameterized by recurrent policies, both using an embedding layer of size 256 followed by a GRU (Cho et al., 2014) of size 512. We provide further hyperparameter details in Table 1 in the Appendix. ",
|
| 596 |
+
"bbox": [
|
| 597 |
+
174,
|
| 598 |
+
521,
|
| 599 |
+
825,
|
| 600 |
+
564
|
| 601 |
+
],
|
| 602 |
+
"page_idx": 5
|
| 603 |
+
},
|
| 604 |
+
{
|
| 605 |
+
"type": "text",
|
| 606 |
+
"text": "5 DO SUPERVISED LEARNING BEFORE SELF-PLAY ",
|
| 607 |
+
"text_level": 1,
|
| 608 |
+
"bbox": [
|
| 609 |
+
174,
|
| 610 |
+
584,
|
| 611 |
+
599,
|
| 612 |
+
599
|
| 613 |
+
],
|
| 614 |
+
"page_idx": 5
|
| 615 |
+
},
|
| 616 |
+
{
|
| 617 |
+
"type": "text",
|
| 618 |
+
"text": "A central question in our work is how to combine supervised and self-play updates for effective pre-training of conversational agents. In this section, we study this question by conducting experiments with two schedules: training with emergent communication followed by supervised learning $( { \\tt s p 2 s u p } )$ , and training with supervised learning followed by self-play (sup2sp). We also interpolate between these two regimes by performing the rand and sched on $0 < n < | \\mathcal { D } |$ samples, followed by supervised fine-tuning on the remaining $| \\mathcal { D } | - n$ samples. ",
|
| 619 |
+
"bbox": [
|
| 620 |
+
173,
|
| 621 |
+
614,
|
| 622 |
+
826,
|
| 623 |
+
699
|
| 624 |
+
],
|
| 625 |
+
"page_idx": 5
|
| 626 |
+
},
|
| 627 |
+
{
|
| 628 |
+
"type": "text",
|
| 629 |
+
"text": "Our first finding is that it is best to use all of your samples for supervised learning before doing self-play. This can be seen in Figure 4: when all of the samples are used first for supervised learning, the number of total samples required to solve the OR game drastically, and in the IBR game the accuracy for a fixed number of samples is maximized (Figure 4a). While this may seem to be common sense, it in fact runs counter to the prevailing wisdom in some emergent communication literature, where languages are emerged from scratch with the ultimate goal of translating them to natural language. ",
|
| 630 |
+
"bbox": [
|
| 631 |
+
174,
|
| 632 |
+
705,
|
| 633 |
+
825,
|
| 634 |
+
804
|
| 635 |
+
],
|
| 636 |
+
"page_idx": 5
|
| 637 |
+
},
|
| 638 |
+
{
|
| 639 |
+
"type": "text",
|
| 640 |
+
"text": "To better understand why it is best to do supervised learning first, we now conduct a set of targeted experiments using the environments from Section 4. Results of our experiments suggest three main explanations: ",
|
| 641 |
+
"bbox": [
|
| 642 |
+
174,
|
| 643 |
+
810,
|
| 644 |
+
825,
|
| 645 |
+
852
|
| 646 |
+
],
|
| 647 |
+
"page_idx": 5
|
| 648 |
+
},
|
| 649 |
+
{
|
| 650 |
+
"type": "text",
|
| 651 |
+
"text": "(1) Emerging a language is hard. For many environments, with emergent communication it’s often hard to find an equilibrium where the agents meaningfully communicate. The difficulty of ‘emergent language discovery’ has been well-known in emergent communication (Lowe et al., 2017), so we will only briefly discuss it here. In short, to discover a useful communication protocol agents have to coordinate repeatedly over time, which is difficult when agents are randomly initialized, particularly in environments with sparse reward. Compounding the difficulty is that, if neither agent communicates and both agents act optimally given their lack of knowledge, they converge to a Nash equilibrium called babbling equilibrium (Farrell & Rabin, 1996). This equilibrium must be overcome to learn a useful communication protocol. In S2P, the initial language supervision can help overcome the discovery problem, as it provides an initial policy for how agents could usefully communicate (Lewis et al., 2017). ",
|
| 652 |
+
"bbox": [
|
| 653 |
+
174,
|
| 654 |
+
867,
|
| 655 |
+
825,
|
| 656 |
+
924
|
| 657 |
+
],
|
| 658 |
+
"page_idx": 5
|
| 659 |
+
},
|
| 660 |
+
{
|
| 661 |
+
"type": "image",
|
| 662 |
+
"img_path": "images/f17032b986b46b95915336ddc8a41b75aa579c7407092aa3ab838bd0ac63cc6c.jpg",
|
| 663 |
+
"image_caption": [
|
| 664 |
+
"Figure 5: Results from the OR game with 1 property and 10 types. When the supervised updates are performed first (supervised data available for words $0 - 3 )$ , then the self-play updates make sensible predictions for the unknown words $4 - 7$ . When the self-play updates are performed first, the subsequent supervised updates merely correct the predictions for words $1 - 4$ , without enforcing the constraint that each word should result in a separate type to solve the task. "
|
| 665 |
+
],
|
| 666 |
+
"image_footnote": [],
|
| 667 |
+
"bbox": [
|
| 668 |
+
176,
|
| 669 |
+
103,
|
| 670 |
+
820,
|
| 671 |
+
239
|
| 672 |
+
],
|
| 673 |
+
"page_idx": 6
|
| 674 |
+
},
|
| 675 |
+
{
|
| 676 |
+
"type": "text",
|
| 677 |
+
"text": "",
|
| 678 |
+
"bbox": [
|
| 679 |
+
173,
|
| 680 |
+
348,
|
| 681 |
+
825,
|
| 682 |
+
446
|
| 683 |
+
],
|
| 684 |
+
"page_idx": 6
|
| 685 |
+
},
|
| 686 |
+
{
|
| 687 |
+
"type": "text",
|
| 688 |
+
"text": "(2) Emergent languages are different than natural language. Even if one does find an equilibrium where agents communicate and perform well on the task, the distribution of languages they find will usually be very different from natural language. This is a problem because, if the languages obtained through self-play are sufficiently different from $L ^ { * }$ , they will not be helpful for learning. This is seen for the OR game in Figure 4a, where 17 samples are required in the seed before S2P outperforms the supervised learning baseline. We speculate that this is due to the different pressures exerted during the emergence of artificial languages and human languages. ",
|
| 689 |
+
"bbox": [
|
| 690 |
+
173,
|
| 691 |
+
462,
|
| 692 |
+
825,
|
| 693 |
+
559
|
| 694 |
+
],
|
| 695 |
+
"page_idx": 6
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"type": "text",
|
| 699 |
+
"text": "Thankfully, we can learn languages closer to $L ^ { * }$ by simply adding more samples to our initial supervised learning phase. We show this in Figure 4b, where we train populations of 50 agents on the IBR game and use Pop-S2P to produce a single distilled agent. With both 1K and 10K initial supervised samples, the distill agent generalizes to agents in the validation set of their population. However, the distilled agent trained with 10000 samples performs significantly better when playing with an expert agent speaking $L ^ { * }$ , indicating that the training agents from that population speak languages closer to $L ^ { * }$ . ",
|
| 700 |
+
"bbox": [
|
| 701 |
+
174,
|
| 702 |
+
566,
|
| 703 |
+
825,
|
| 704 |
+
664
|
| 705 |
+
],
|
| 706 |
+
"page_idx": 6
|
| 707 |
+
},
|
| 708 |
+
{
|
| 709 |
+
"type": "text",
|
| 710 |
+
"text": "(3) Starting with self-play violates constraints. Even if you have ‘perfect emergent communication’ that learns a distribution over languages under which $L ^ { * }$ has high probability, current methods of supervised fine-tuning do not properly learn from this distribution. What if we had all the correct learning pressures, such that we emerged a distribution over languages $\\mathcal { L }$ with structure identical to $L ^ { * }$ , and then trained a Pop-S2P agent using this distribution? Surprisingly, we find that S2P with all of the samples in the seed performs better than even this optimistic case, in terms of providing useful information for training a Pop-S2P agent. We conduct an experiment in the OR game where we programmatically define a distribution over compositional languages $\\mathcal { L } _ { c }$ , of which our target language $L ^ { * }$ is a sample. Each language $L \\in \\mathcal { L } _ { c }$ has the same structure and are obtained by randomly permuting the mapping between the word IDs and the corresponding type IDs, along with the order of properties in an utterance. Next, we compare two distilled policies using 50 populations: one is distilled from S2P populations (trained with $X$ samples), and the other is distilled from ‘perfect emergent communication’ and fine-tuned on $X$ samples. As can be seen in Figure 4c, we show that when we train a Pop-S2P agent on 50 of these compositional populations, we still need $3 X$ more samples than regular Pop-S2P (trained on ${ 5 0 } \\thinspace \\mathrm { S 2 P }$ agents with all of the samples in the seed) to reach $9 5 \\%$ test accuracy3. ",
|
| 711 |
+
"bbox": [
|
| 712 |
+
174,
|
| 713 |
+
679,
|
| 714 |
+
825,
|
| 715 |
+
900
|
| 716 |
+
],
|
| 717 |
+
"page_idx": 6
|
| 718 |
+
},
|
| 719 |
+
{
|
| 720 |
+
"type": "text",
|
| 721 |
+
"text": "To understand why this happens, we conduct a case study in an even simpler setting: single-agent S2P in the OR game with $p = 1 , t = 1 0 , | V | = 1 0$ . We find that agents trained via emergent communication consistently learn to solve this task. However, as shown in Figure 5, when subsequently trained via supervised learning on $\\mathcal { D }$ to learn $L ^ { * }$ , the learned language is no longer coherent (it maps different words to the same type) and doesn’t solve the task. On the other hand, agents trained first with supervised learning are able to learn a language that both solves the task and is consistent with $\\mathcal { D }$ . ",
|
| 722 |
+
"bbox": [
|
| 723 |
+
174,
|
| 724 |
+
103,
|
| 725 |
+
825,
|
| 726 |
+
200
|
| 727 |
+
],
|
| 728 |
+
"page_idx": 7
|
| 729 |
+
},
|
| 730 |
+
{
|
| 731 |
+
"type": "text",
|
| 732 |
+
"text": "Intuitively, what’s happening is that the samples in $\\mathcal { D }$ are also valid for solving the task, since we assume agents speaking $L ^ { * }$ can solve the task. Thus, self-play after supervised learning simply ‘fills in the gaps’ for examples not in $\\mathcal { D }$ .4 Emergent languages that start with self-play, on the other hand, contain input-output mappings that are inconsistent with $L ^ { * }$ , which must be un-learned during subsequent supervised learning. ",
|
| 733 |
+
"bbox": [
|
| 734 |
+
174,
|
| 735 |
+
208,
|
| 736 |
+
825,
|
| 737 |
+
279
|
| 738 |
+
],
|
| 739 |
+
"page_idx": 7
|
| 740 |
+
},
|
| 741 |
+
{
|
| 742 |
+
"type": "text",
|
| 743 |
+
"text": "In theory, the above issue could be resolved using Pop-S2P; if the distilled agent could use the population of emergent languages to discover structural rules (e.g. discovering that the languages in the OR game in Figure 4c are compositional), it could use the samples from $\\mathcal { D }$ to refine a posterior distribution over target languages that is consistent with these rules (e.g. learning the distribution of compositional languages consistent with $\\mathcal { D }$ ). Current approaches to supervised fine-tuning in language, though, do not do this (Lazaridou et al., 2017; Lewis et al., 2017). An interesting direction for future work is examining how to apply Bayesian techniques to S2P. ",
|
| 744 |
+
"bbox": [
|
| 745 |
+
174,
|
| 746 |
+
285,
|
| 747 |
+
825,
|
| 748 |
+
382
|
| 749 |
+
],
|
| 750 |
+
"page_idx": 7
|
| 751 |
+
},
|
| 752 |
+
{
|
| 753 |
+
"type": "text",
|
| 754 |
+
"text": "6 EXPLORING VARIANTS OF S2P ",
|
| 755 |
+
"text_level": 1,
|
| 756 |
+
"bbox": [
|
| 757 |
+
176,
|
| 758 |
+
405,
|
| 759 |
+
459,
|
| 760 |
+
420
|
| 761 |
+
],
|
| 762 |
+
"page_idx": 7
|
| 763 |
+
},
|
| 764 |
+
{
|
| 765 |
+
"type": "text",
|
| 766 |
+
"text": "6.1 POPULATION-BASED S2P ",
|
| 767 |
+
"text_level": 1,
|
| 768 |
+
"bbox": [
|
| 769 |
+
174,
|
| 770 |
+
438,
|
| 771 |
+
390,
|
| 772 |
+
452
|
| 773 |
+
],
|
| 774 |
+
"page_idx": 7
|
| 775 |
+
},
|
| 776 |
+
{
|
| 777 |
+
"type": "text",
|
| 778 |
+
"text": "In this section, we aim to show that (1) S2P outperforms the supervised learning baseline, and (2) Pop-S2P outperforms S2P. We conduct our experiments in the more complex IBR game, since the agents must communicate in English, and measure performance by calculating the accuracy at different (fixed) numbers of samples. Our baseline is then the performance of a supervised learner on a fixed number of samples. ",
|
| 779 |
+
"bbox": [
|
| 780 |
+
174,
|
| 781 |
+
464,
|
| 782 |
+
614,
|
| 783 |
+
561
|
| 784 |
+
],
|
| 785 |
+
"page_idx": 7
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"type": "text",
|
| 789 |
+
"text": "We show the results in Figure 6. We first note that, when both 1k and 10k samples are used for supervised learning, S2P (sched) outperforms the supervised learning baseline. We can also see that the population-based approach outperforms single agent S2P (sched) by a significant margin. We also compare our distillation method to an ensembling method that keeps all 50 populations at test time, and find that ensembling performs significantly better, although it is much less efficient. This suggests that there is room to push distilled Pop-S2P to even better performance. ",
|
| 790 |
+
"bbox": [
|
| 791 |
+
174,
|
| 792 |
+
569,
|
| 793 |
+
614,
|
| 794 |
+
694
|
| 795 |
+
],
|
| 796 |
+
"page_idx": 7
|
| 797 |
+
},
|
| 798 |
+
{
|
| 799 |
+
"type": "image",
|
| 800 |
+
"img_path": "images/3959aab1406eead38a90d70e1606acf6e18b83853052ebb0c66dfe530e5fcf8c.jpg",
|
| 801 |
+
"image_caption": [
|
| 802 |
+
"Figure 6: S2P (sched) outperforms the supervised baseline in the IBR game, and is in turn outperformed by PopS2P. "
|
| 803 |
+
],
|
| 804 |
+
"image_footnote": [],
|
| 805 |
+
"bbox": [
|
| 806 |
+
629,
|
| 807 |
+
465,
|
| 808 |
+
818,
|
| 809 |
+
599
|
| 810 |
+
],
|
| 811 |
+
"page_idx": 7
|
| 812 |
+
},
|
| 813 |
+
{
|
| 814 |
+
"type": "text",
|
| 815 |
+
"text": "6.2 EXAMINING S2P SCHEDULES ",
|
| 816 |
+
"text_level": 1,
|
| 817 |
+
"bbox": [
|
| 818 |
+
176,
|
| 819 |
+
713,
|
| 820 |
+
418,
|
| 821 |
+
727
|
| 822 |
+
],
|
| 823 |
+
"page_idx": 7
|
| 824 |
+
},
|
| 825 |
+
{
|
| 826 |
+
"type": "text",
|
| 827 |
+
"text": "In this section, we aim to: (1) evaluate several S2P schedules empirically on the IBR game; and (2) attain a better understanding of S2P through quantitative experiments. ",
|
| 828 |
+
"bbox": [
|
| 829 |
+
173,
|
| 830 |
+
739,
|
| 831 |
+
823,
|
| 832 |
+
768
|
| 833 |
+
],
|
| 834 |
+
"page_idx": 7
|
| 835 |
+
},
|
| 836 |
+
{
|
| 837 |
+
"type": "text",
|
| 838 |
+
"text": "Parameter freezing improves S2P We show the results comparing different S2P schedules in Figure 7a. We find that in this more complex game, the sup2sp S2P performs much worse than the other options. We also see that adding freezing slightly improves the performance on the target language (Figure 8 in the Appendix also shows that it converges more quickly). We hypothesize that this is because it reduces the language drift that is experienced during each round of self-play updates (Lee et al., 2019). Overall, however, the difference between different S2P schedules is relatively small, and it’s unclear if the same ordering will hold in a different domain. ",
|
| 839 |
+
"bbox": [
|
| 840 |
+
173,
|
| 841 |
+
785,
|
| 842 |
+
825,
|
| 843 |
+
883
|
| 844 |
+
],
|
| 845 |
+
"page_idx": 7
|
| 846 |
+
},
|
| 847 |
+
{
|
| 848 |
+
"type": "image",
|
| 849 |
+
"img_path": "images/272763dd45fb79799b9b252a4b4f58c316e494907a089adbb5d69ecdf6589dae.jpg",
|
| 850 |
+
"image_caption": [
|
| 851 |
+
"Figure 7: (a) Comparing test performances of different S2P methods on the IBR game. For each method, we picked the model that gave the best performance on $\\mathcal { D } _ { v a l }$ . (b) 2D visualization of S2P (sched) performance over the course of training, in terms of performance on $L ^ { * }$ (vertical axis) and performance in self-play (horizontal axis). The zig-zag patterns indicates that most self-play updates result in a short-term decrease in target language performance. (c) Visualization of the role of the supervised and self-play updates in sched S2P. "
|
| 852 |
+
],
|
| 853 |
+
"image_footnote": [],
|
| 854 |
+
"bbox": [
|
| 855 |
+
181,
|
| 856 |
+
103,
|
| 857 |
+
812,
|
| 858 |
+
292
|
| 859 |
+
],
|
| 860 |
+
"page_idx": 8
|
| 861 |
+
},
|
| 862 |
+
{
|
| 863 |
+
"type": "text",
|
| 864 |
+
"text": "Self-play acts as a regularizer What is the role of self-play in S2P? We can start to decipher this by taking a closer look at the sched S2P. We plot the training performance of this method in Figure 7b. Interestingly, we notice from the zig-zag pattern that the validation performance usually goes down after every set of self-play updates. However, the overall validation performance goes up after the next round of supervised updates. This is also reflected in the poor performance of the sup2sp S2P in Figure 6. ",
|
| 865 |
+
"bbox": [
|
| 866 |
+
173,
|
| 867 |
+
438,
|
| 868 |
+
825,
|
| 869 |
+
522
|
| 870 |
+
],
|
| 871 |
+
"page_idx": 8
|
| 872 |
+
},
|
| 873 |
+
{
|
| 874 |
+
"type": "text",
|
| 875 |
+
"text": "This phenomenon can be explained by framing self-play as a form of regularization: alternating between supervised and self-play updates is a way to satisfy the parallel constraints of ‘is consistent with the dataset $\\mathcal { D } ^ { \\ast }$ and ‘performs well on the task’. We visualize this pictorially in Figure 7b: while a set of self-play updates results in poor performance on $\\mathcal { D }$ , eventually the learned language moves closer to satisfying both constraints. ",
|
| 876 |
+
"bbox": [
|
| 877 |
+
174,
|
| 878 |
+
529,
|
| 879 |
+
825,
|
| 880 |
+
599
|
| 881 |
+
],
|
| 882 |
+
"page_idx": 8
|
| 883 |
+
},
|
| 884 |
+
{
|
| 885 |
+
"type": "text",
|
| 886 |
+
"text": "7 DISCUSSION ",
|
| 887 |
+
"text_level": 1,
|
| 888 |
+
"bbox": [
|
| 889 |
+
176,
|
| 890 |
+
646,
|
| 891 |
+
310,
|
| 892 |
+
662
|
| 893 |
+
],
|
| 894 |
+
"page_idx": 8
|
| 895 |
+
},
|
| 896 |
+
{
|
| 897 |
+
"type": "text",
|
| 898 |
+
"text": "In this work, we investigated the research question of how to combine supervised and self-play updates, with a focus on training agents to learn a language. However, this research question is not only important for language learning; it is also a important in equilibrium selection and learning social conventions (Lerer & Peysakhovich, 2019) in general games. For example, in robotics there may be a trade-off between performing a task well (moving an object to a certain place) and having your policy be interpretable by humans (so that they will not stumble over you). Examining how to combine supervised and self-play updates in these settings is an exciting direction for future work. ",
|
| 899 |
+
"bbox": [
|
| 900 |
+
174,
|
| 901 |
+
694,
|
| 902 |
+
825,
|
| 903 |
+
791
|
| 904 |
+
],
|
| 905 |
+
"page_idx": 8
|
| 906 |
+
},
|
| 907 |
+
{
|
| 908 |
+
"type": "text",
|
| 909 |
+
"text": "There are several axes of complexity not addressed in our environments and problem set-up. First, we consider only single-state environments, and agents don’t have to make temporally extended decisions. Second, we do not consider pre-training on large text corpora that are separate from the desired task (Radford et al., 2019; Devlin et al., 2018). Third, we limit our exploration of self-play to the multi-agent setting, which is not the case in works such as instruction following (Andreas & Klein, 2015). Introducing these elements may result in additional practical considerations for S2P learning, which we leave for future work. Our goal in this paper is not to determine the best method of S2P in all of these settings, but rather to inspire others to use the framing of ‘supervised self-play algorithms’ to make progress on sample efficient human-in-the-loop language learning. ",
|
| 910 |
+
"bbox": [
|
| 911 |
+
174,
|
| 912 |
+
797,
|
| 913 |
+
825,
|
| 914 |
+
924
|
| 915 |
+
],
|
| 916 |
+
"page_idx": 8
|
| 917 |
+
},
|
| 918 |
+
{
|
| 919 |
+
"type": "text",
|
| 920 |
+
"text": "ACKNOWLEDGEMENTS ",
|
| 921 |
+
"text_level": 1,
|
| 922 |
+
"bbox": [
|
| 923 |
+
176,
|
| 924 |
+
103,
|
| 925 |
+
367,
|
| 926 |
+
117
|
| 927 |
+
],
|
| 928 |
+
"page_idx": 9
|
| 929 |
+
},
|
| 930 |
+
{
|
| 931 |
+
"type": "text",
|
| 932 |
+
"text": "We are very grateful to Angeliki Lazaridou, with whom discussions at ICML 2019 and her simultaneous work (Lazaridou et al., 2020) shifted the direction of this work considerably. We also thank Jean Harb, Liam Fedus, Amy Zhang, Evgeny Naumov, Cinjon Resnick, Igor Mordatch, and others at MILA and Facebook AI Research for discussions related to the ideas in this paper. Special thanks to Arthur Szlam and Kavya Srinet for discussing their ongoing work with us. RL is supported in part by a Vanier Scholarship. ",
|
| 933 |
+
"bbox": [
|
| 934 |
+
174,
|
| 935 |
+
133,
|
| 936 |
+
826,
|
| 937 |
+
218
|
| 938 |
+
],
|
| 939 |
+
"page_idx": 9
|
| 940 |
+
},
|
| 941 |
+
{
|
| 942 |
+
"type": "text",
|
| 943 |
+
"text": "REFERENCES ",
|
| 944 |
+
"text_level": 1,
|
| 945 |
+
"bbox": [
|
| 946 |
+
174,
|
| 947 |
+
239,
|
| 948 |
+
285,
|
| 949 |
+
255
|
| 950 |
+
],
|
| 951 |
+
"page_idx": 9
|
| 952 |
+
},
|
| 953 |
+
{
|
| 954 |
+
"type": "text",
|
| 955 |
+
"text": "Jacob Andreas and Dan Klein. Alignment-based compositional semantics for instruction following. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pp. 1165–1174, Lisbon, Portugal, September 2015. Association for Computational Linguistics. doi: 10.18653/v1/D15-1138. URL https://www.aclweb.org/anthology/D15-1138. ",
|
| 956 |
+
"bbox": [
|
| 957 |
+
174,
|
| 958 |
+
265,
|
| 959 |
+
826,
|
| 960 |
+
320
|
| 961 |
+
],
|
| 962 |
+
"page_idx": 9
|
| 963 |
+
},
|
| 964 |
+
{
|
| 965 |
+
"type": "text",
|
| 966 |
+
"text": "Kris Cao, Angeliki Lazaridou, Marc Lanctot, Joel Z Leibo, Karl Tuyls, and Stephen Clark. Emergent communication through negotiation. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id ${ . } =$ Hk6WhagRW. ",
|
| 967 |
+
"bbox": [
|
| 968 |
+
174,
|
| 969 |
+
330,
|
| 970 |
+
825,
|
| 971 |
+
373
|
| 972 |
+
],
|
| 973 |
+
"page_idx": 9
|
| 974 |
+
},
|
| 975 |
+
{
|
| 976 |
+
"type": "text",
|
| 977 |
+
"text": "Rahma Chaabouni, Eugene Kharitonov, Emmanuel Dupoux, and Marco Baroni. Anti-efficient encoding in emergent communication. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d\\textquotesingle Alché-Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 6290–6300. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/ 8859-anti-efficient-encoding-in-emergent-communication.pdf. ",
|
| 978 |
+
"bbox": [
|
| 979 |
+
174,
|
| 980 |
+
383,
|
| 981 |
+
826,
|
| 982 |
+
454
|
| 983 |
+
],
|
| 984 |
+
"page_idx": 9
|
| 985 |
+
},
|
| 986 |
+
{
|
| 987 |
+
"type": "text",
|
| 988 |
+
"text": "Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Salem Lahlou, Lucas Willems, Chitwan Saharia, Thien Huu Nguyen, and Yoshua Bengio. BabyAI: First steps towards grounded language learning with a human in the loop. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ rJeXCo0cYX. ",
|
| 989 |
+
"bbox": [
|
| 990 |
+
174,
|
| 991 |
+
465,
|
| 992 |
+
826,
|
| 993 |
+
522
|
| 994 |
+
],
|
| 995 |
+
"page_idx": 9
|
| 996 |
+
},
|
| 997 |
+
{
|
| 998 |
+
"type": "text",
|
| 999 |
+
"text": "Kyunghyun Cho, Bart van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using RNN encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1724–1734, Doha, Qatar, October 2014. Association for Computational Linguistics. doi: 10.3115/v1/D14-1179. URL https://www.aclweb.org/ anthology/D14-1179. ",
|
| 1000 |
+
"bbox": [
|
| 1001 |
+
174,
|
| 1002 |
+
532,
|
| 1003 |
+
826,
|
| 1004 |
+
617
|
| 1005 |
+
],
|
| 1006 |
+
"page_idx": 9
|
| 1007 |
+
},
|
| 1008 |
+
{
|
| 1009 |
+
"type": "text",
|
| 1010 |
+
"text": "Edward Choi, Angeliki Lazaridou, and Nando de Freitas. Multi-agent compositional communication learning from raw visual input. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $=$ rknt2Be0-. ",
|
| 1011 |
+
"bbox": [
|
| 1012 |
+
174,
|
| 1013 |
+
627,
|
| 1014 |
+
823,
|
| 1015 |
+
670
|
| 1016 |
+
],
|
| 1017 |
+
"page_idx": 9
|
| 1018 |
+
},
|
| 1019 |
+
{
|
| 1020 |
+
"type": "text",
|
| 1021 |
+
"text": "Michael Cogswell, Jiasen Lu, Stefan Lee, Devi Parikh, and Dhruv Batra. Emergence of Compositional Language with Deep Generational Transmission. arXiv:1904.09067 [cs, stat], April 2019. arXiv: 1904.09067. ",
|
| 1022 |
+
"bbox": [
|
| 1023 |
+
174,
|
| 1024 |
+
681,
|
| 1025 |
+
826,
|
| 1026 |
+
723
|
| 1027 |
+
],
|
| 1028 |
+
"page_idx": 9
|
| 1029 |
+
},
|
| 1030 |
+
{
|
| 1031 |
+
"type": "text",
|
| 1032 |
+
"text": "Harm de Vries, Kurt Shuster, Dhruv Batra, Devi Parikh, Jason Weston, and Douwe Kiela. Talk the walk: Navigating new york city through grounded dialogue. arXiv preprint arXiv:1807.03367, 2018. ",
|
| 1033 |
+
"bbox": [
|
| 1034 |
+
173,
|
| 1035 |
+
734,
|
| 1036 |
+
826,
|
| 1037 |
+
776
|
| 1038 |
+
],
|
| 1039 |
+
"page_idx": 9
|
| 1040 |
+
},
|
| 1041 |
+
{
|
| 1042 |
+
"type": "text",
|
| 1043 |
+
"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. ",
|
| 1044 |
+
"bbox": [
|
| 1045 |
+
171,
|
| 1046 |
+
787,
|
| 1047 |
+
825,
|
| 1048 |
+
818
|
| 1049 |
+
],
|
| 1050 |
+
"page_idx": 9
|
| 1051 |
+
},
|
| 1052 |
+
{
|
| 1053 |
+
"type": "text",
|
| 1054 |
+
"text": "Katrina Evtimova, Andrew Drozdov, Douwe Kiela, and Kyunghyun Cho. Emergent communication in a multi-modal, multi-step referential game. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } }$ rJGZq6g0-. ",
|
| 1055 |
+
"bbox": [
|
| 1056 |
+
176,
|
| 1057 |
+
828,
|
| 1058 |
+
825,
|
| 1059 |
+
871
|
| 1060 |
+
],
|
| 1061 |
+
"page_idx": 9
|
| 1062 |
+
},
|
| 1063 |
+
{
|
| 1064 |
+
"type": "text",
|
| 1065 |
+
"text": "Joseph Farrell and Matthew Rabin. Cheap talk. Journal of Economic Perspectives, 10(3): 103–118, September 1996. doi: 10.1257/jep.10.3.103. URL http://www.aeaweb.org/ articles?id=10.1257/jep.10.3.103. ",
|
| 1066 |
+
"bbox": [
|
| 1067 |
+
174,
|
| 1068 |
+
882,
|
| 1069 |
+
825,
|
| 1070 |
+
924
|
| 1071 |
+
],
|
| 1072 |
+
"page_idx": 9
|
| 1073 |
+
},
|
| 1074 |
+
{
|
| 1075 |
+
"type": "text",
|
| 1076 |
+
"text": "Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon Whiteson. Learning to Communicate with Deep Multi-Agent Reinforcement Learning. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp. 2137–2145. Curran Associates, Inc., 2016. ",
|
| 1077 |
+
"bbox": [
|
| 1078 |
+
174,
|
| 1079 |
+
103,
|
| 1080 |
+
826,
|
| 1081 |
+
160
|
| 1082 |
+
],
|
| 1083 |
+
"page_idx": 10
|
| 1084 |
+
},
|
| 1085 |
+
{
|
| 1086 |
+
"type": "text",
|
| 1087 |
+
"text": "Laura Harding Graesser, Kyunghyun Cho, and Douwe Kiela. Emergent linguistic phenomena in multi-agent communication games. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 3691–3701, Hong Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1384. URL https:// www.aclweb.org/anthology/D19-1384. ",
|
| 1088 |
+
"bbox": [
|
| 1089 |
+
173,
|
| 1090 |
+
167,
|
| 1091 |
+
826,
|
| 1092 |
+
252
|
| 1093 |
+
],
|
| 1094 |
+
"page_idx": 10
|
| 1095 |
+
},
|
| 1096 |
+
{
|
| 1097 |
+
"type": "text",
|
| 1098 |
+
"text": "Serhii Havrylov and Ivan Titov. Emergence of Language with Multi-agent Games: Learning to Communicate with Sequences of Symbols. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems 30, pp. 2149–2159. Curran Associates, Inc., 2017. ",
|
| 1099 |
+
"bbox": [
|
| 1100 |
+
174,
|
| 1101 |
+
261,
|
| 1102 |
+
826,
|
| 1103 |
+
318
|
| 1104 |
+
],
|
| 1105 |
+
"page_idx": 10
|
| 1106 |
+
},
|
| 1107 |
+
{
|
| 1108 |
+
"type": "text",
|
| 1109 |
+
"text": "Eric Jang, Shixiang Gu, and Ben Poole. Categorical Reparameterization with GumbelSoftmax. In International Conference on Learning Representations, 2017. URL https: //openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ rkE3y85ee. ",
|
| 1110 |
+
"bbox": [
|
| 1111 |
+
174,
|
| 1112 |
+
325,
|
| 1113 |
+
823,
|
| 1114 |
+
368
|
| 1115 |
+
],
|
| 1116 |
+
"page_idx": 10
|
| 1117 |
+
},
|
| 1118 |
+
{
|
| 1119 |
+
"type": "text",
|
| 1120 |
+
"text": "Natasha Jaques, Angeliki Lazaridou, Edward Hughes, Caglar Gulcehre, Pedro Ortega, Dj Strouse, Joel Z. Leibo, and Nando De Freitas. Social Influence as Intrinsic Motivation for Multi-Agent Deep Reinforcement Learning. In International Conference on Machine Learning, pp. 3040–3049, May 2019. URL http://proceedings.mlr.press/v97/jaques19a.html. ",
|
| 1121 |
+
"bbox": [
|
| 1122 |
+
173,
|
| 1123 |
+
376,
|
| 1124 |
+
826,
|
| 1125 |
+
434
|
| 1126 |
+
],
|
| 1127 |
+
"page_idx": 10
|
| 1128 |
+
},
|
| 1129 |
+
{
|
| 1130 |
+
"type": "text",
|
| 1131 |
+
"text": "Simon Kirby. Iterated learning and the evolution of language. Current Opinion in Neurobiology, pp. 7, 2014. ",
|
| 1132 |
+
"bbox": [
|
| 1133 |
+
173,
|
| 1134 |
+
441,
|
| 1135 |
+
825,
|
| 1136 |
+
470
|
| 1137 |
+
],
|
| 1138 |
+
"page_idx": 10
|
| 1139 |
+
},
|
| 1140 |
+
{
|
| 1141 |
+
"type": "text",
|
| 1142 |
+
"text": "Satwik Kottur, José Moura, Stefan Lee, and Dhruv Batra. Natural language does not emerge ‘naturally’ in multi-agent dialog. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 2962–2967, Copenhagen, Denmark, September 2017. Association for Computational Linguistics. doi: 10.18653/v1/D17-1321. URL https://www.aclweb.org/ anthology/D17-1321. ",
|
| 1143 |
+
"bbox": [
|
| 1144 |
+
174,
|
| 1145 |
+
478,
|
| 1146 |
+
826,
|
| 1147 |
+
549
|
| 1148 |
+
],
|
| 1149 |
+
"page_idx": 10
|
| 1150 |
+
},
|
| 1151 |
+
{
|
| 1152 |
+
"type": "text",
|
| 1153 |
+
"text": "Angeliki Lazaridou, Alexander Peysakhovich, and Marco Baroni. Multi-Agent Cooperation and the Emergence of (Natural) Language. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } }$ Hk8N3Sclg. ",
|
| 1154 |
+
"bbox": [
|
| 1155 |
+
174,
|
| 1156 |
+
556,
|
| 1157 |
+
826,
|
| 1158 |
+
599
|
| 1159 |
+
],
|
| 1160 |
+
"page_idx": 10
|
| 1161 |
+
},
|
| 1162 |
+
{
|
| 1163 |
+
"type": "text",
|
| 1164 |
+
"text": "Angeliki Lazaridou, Karl Moritz Hermann, Karl Tuyls, and Stephen Clark. Emergence of linguistic communication from referential games with symbolic and pixel input. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id= HJGv1Z-AW. ",
|
| 1165 |
+
"bbox": [
|
| 1166 |
+
174,
|
| 1167 |
+
608,
|
| 1168 |
+
825,
|
| 1169 |
+
664
|
| 1170 |
+
],
|
| 1171 |
+
"page_idx": 10
|
| 1172 |
+
},
|
| 1173 |
+
{
|
| 1174 |
+
"type": "text",
|
| 1175 |
+
"text": "Angeliki Lazaridou, Anna Potapenko, and Olivier Tieleman. Multi-agent communication meets natural language: Synergies between functional and structural language learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 7663–7674, Online, July 2020. Association for Computational Linguistics. URL https://www.aclweb.org/ anthology/2020.acl-main.685. ",
|
| 1176 |
+
"bbox": [
|
| 1177 |
+
174,
|
| 1178 |
+
672,
|
| 1179 |
+
826,
|
| 1180 |
+
743
|
| 1181 |
+
],
|
| 1182 |
+
"page_idx": 10
|
| 1183 |
+
},
|
| 1184 |
+
{
|
| 1185 |
+
"type": "text",
|
| 1186 |
+
"text": "Jason Lee, Kyunghyun Cho, Jason Weston, and Douwe Kiela. Emergent translation in multiagent communication. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id $_ { \\cdot } =$ H1vEXaxA-. ",
|
| 1187 |
+
"bbox": [
|
| 1188 |
+
176,
|
| 1189 |
+
751,
|
| 1190 |
+
823,
|
| 1191 |
+
795
|
| 1192 |
+
],
|
| 1193 |
+
"page_idx": 10
|
| 1194 |
+
},
|
| 1195 |
+
{
|
| 1196 |
+
"type": "text",
|
| 1197 |
+
"text": "Jason Lee, Kyunghyun Cho, and Douwe Kiela. Countering language drift via visual grounding. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 4376–4386, Hong Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1447. URL https://www.aclweb.org/anthology/D19-1447. ",
|
| 1198 |
+
"bbox": [
|
| 1199 |
+
174,
|
| 1200 |
+
803,
|
| 1201 |
+
826,
|
| 1202 |
+
872
|
| 1203 |
+
],
|
| 1204 |
+
"page_idx": 10
|
| 1205 |
+
},
|
| 1206 |
+
{
|
| 1207 |
+
"type": "text",
|
| 1208 |
+
"text": "Adam Lerer and Alexander Peysakhovich. Learning existing social conventions via observationally augmented self-play. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pp. 107–114. ACM, 2019. ",
|
| 1209 |
+
"bbox": [
|
| 1210 |
+
176,
|
| 1211 |
+
882,
|
| 1212 |
+
825,
|
| 1213 |
+
924
|
| 1214 |
+
],
|
| 1215 |
+
"page_idx": 10
|
| 1216 |
+
},
|
| 1217 |
+
{
|
| 1218 |
+
"type": "text",
|
| 1219 |
+
"text": "David Lewis. Convention: A philosophical study. Harvard University Press, 1969. ",
|
| 1220 |
+
"bbox": [
|
| 1221 |
+
173,
|
| 1222 |
+
103,
|
| 1223 |
+
712,
|
| 1224 |
+
118
|
| 1225 |
+
],
|
| 1226 |
+
"page_idx": 11
|
| 1227 |
+
},
|
| 1228 |
+
{
|
| 1229 |
+
"type": "text",
|
| 1230 |
+
"text": "Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, and Dhruv Batra. Deal or no deal? endto-end learning of negotiation dialogues. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 2443–2453, Copenhagen, Denmark, September 2017. Association for Computational Linguistics. doi: 10.18653/v1/D17-1259. URL https: //www.aclweb.org/anthology/D17-1259. ",
|
| 1231 |
+
"bbox": [
|
| 1232 |
+
174,
|
| 1233 |
+
127,
|
| 1234 |
+
826,
|
| 1235 |
+
199
|
| 1236 |
+
],
|
| 1237 |
+
"page_idx": 11
|
| 1238 |
+
},
|
| 1239 |
+
{
|
| 1240 |
+
"type": "text",
|
| 1241 |
+
"text": "Fushan Li and Michael Bowling. Ease-of-Teaching and Language Structure from Emergent Communication. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d\\textquotesingle Alché-Buc, E. Fox, and R. Garnett (eds.), Advances in Neural Information Processing Systems 32, pp. 15825–15835. Curran Associates, Inc., 2019. URL http://papers.nips.cc/paper/9714-ease-ofteaching-and-language-structure-from-emergent-communication.pdf. ",
|
| 1242 |
+
"bbox": [
|
| 1243 |
+
173,
|
| 1244 |
+
208,
|
| 1245 |
+
826,
|
| 1246 |
+
279
|
| 1247 |
+
],
|
| 1248 |
+
"page_idx": 11
|
| 1249 |
+
},
|
| 1250 |
+
{
|
| 1251 |
+
"type": "text",
|
| 1252 |
+
"text": "Michael L Littman. Markov games as a framework for multi-agent reinforcement learning. In International Conference on Machine Learning, volume 157, pp. 157–163, 1994. ",
|
| 1253 |
+
"bbox": [
|
| 1254 |
+
171,
|
| 1255 |
+
287,
|
| 1256 |
+
825,
|
| 1257 |
+
318
|
| 1258 |
+
],
|
| 1259 |
+
"page_idx": 11
|
| 1260 |
+
},
|
| 1261 |
+
{
|
| 1262 |
+
"type": "text",
|
| 1263 |
+
"text": "Ryan Lowe, YI WU, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. MultiAgent Actor-Critic for Mixed Cooperative-Competitive Environments. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems 30, pp. 6379–6390. Curran Associates, Inc., 2017. URL http://papers.nips.cc/paper/7217-multi-agent-actor-critic-formixed-cooperative-competitive-environments.pdf. ",
|
| 1264 |
+
"bbox": [
|
| 1265 |
+
174,
|
| 1266 |
+
327,
|
| 1267 |
+
826,
|
| 1268 |
+
411
|
| 1269 |
+
],
|
| 1270 |
+
"page_idx": 11
|
| 1271 |
+
},
|
| 1272 |
+
{
|
| 1273 |
+
"type": "text",
|
| 1274 |
+
"text": "Ryan Lowe, Jakob Foerster, Y-Lan Boureau, Joelle Pineau, and Yann Dauphin. On the pitfalls of measuring emergent communication. In Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems, AAMAS ’19, pp. 693–701. International Foundation for Autonomous Agents and Multiagent Systems, 2019. ISBN 978-1-4503-6309-9. URL http: //www.ifaamas.org/Proceedings/aamas2019/pdfs/p693.pdf. ",
|
| 1275 |
+
"bbox": [
|
| 1276 |
+
173,
|
| 1277 |
+
420,
|
| 1278 |
+
826,
|
| 1279 |
+
491
|
| 1280 |
+
],
|
| 1281 |
+
"page_idx": 11
|
| 1282 |
+
},
|
| 1283 |
+
{
|
| 1284 |
+
"type": "text",
|
| 1285 |
+
"text": "Chris J. Maddison, Andriy Mnih, and Yee Whye Teh. The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id $\\scriptstyle 1 = \\ S 1$ jE5L5gl. ",
|
| 1286 |
+
"bbox": [
|
| 1287 |
+
176,
|
| 1288 |
+
500,
|
| 1289 |
+
825,
|
| 1290 |
+
544
|
| 1291 |
+
],
|
| 1292 |
+
"page_idx": 11
|
| 1293 |
+
},
|
| 1294 |
+
{
|
| 1295 |
+
"type": "text",
|
| 1296 |
+
"text": "Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi-agent populations. In AAAI Conference on Artificial Intelligence, 2018. URL https://aaai.org/ ocs/index.php/AAAI/AAAI18/paper/view/17007. ",
|
| 1297 |
+
"bbox": [
|
| 1298 |
+
174,
|
| 1299 |
+
553,
|
| 1300 |
+
823,
|
| 1301 |
+
595
|
| 1302 |
+
],
|
| 1303 |
+
"page_idx": 11
|
| 1304 |
+
},
|
| 1305 |
+
{
|
| 1306 |
+
"type": "text",
|
| 1307 |
+
"text": "Mathijs Mul, Diane Bouchacourt, and Elia Bruni. Mastering emergent language: learning to guide in simulated navigation. arXiv:1908.05135 [cs], August 2019. arXiv: 1908.05135. ",
|
| 1308 |
+
"bbox": [
|
| 1309 |
+
171,
|
| 1310 |
+
606,
|
| 1311 |
+
823,
|
| 1312 |
+
635
|
| 1313 |
+
],
|
| 1314 |
+
"page_idx": 11
|
| 1315 |
+
},
|
| 1316 |
+
{
|
| 1317 |
+
"type": "text",
|
| 1318 |
+
"text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. URL https://d4mucfpksywv.cloudfront.net/better-language-models/ language_models_are_unsupervised_multitask_learners.pdf. ",
|
| 1319 |
+
"bbox": [
|
| 1320 |
+
174,
|
| 1321 |
+
643,
|
| 1322 |
+
825,
|
| 1323 |
+
700
|
| 1324 |
+
],
|
| 1325 |
+
"page_idx": 11
|
| 1326 |
+
},
|
| 1327 |
+
{
|
| 1328 |
+
"type": "text",
|
| 1329 |
+
"text": "Limor Raviv and Inbal Arnon. Systematicity, but not compositionality: Examining the emergence of linguistic structure in children and adults using iterated learning. Cognition, 181:160–173, December 2018. ISSN 0010-0277. ",
|
| 1330 |
+
"bbox": [
|
| 1331 |
+
174,
|
| 1332 |
+
710,
|
| 1333 |
+
826,
|
| 1334 |
+
752
|
| 1335 |
+
],
|
| 1336 |
+
"page_idx": 11
|
| 1337 |
+
},
|
| 1338 |
+
{
|
| 1339 |
+
"type": "text",
|
| 1340 |
+
"text": "Cinjon Resnick\\*, Abhinav Gupta\\*, Jakob N. Foerster, Andrew M. Dai, and Kyunghyun Cho. Capacity, bandwidth, and compositionality in emergent language learning. ArXiv, abs/1910.11424, 2019. ",
|
| 1341 |
+
"bbox": [
|
| 1342 |
+
173,
|
| 1343 |
+
762,
|
| 1344 |
+
825,
|
| 1345 |
+
791
|
| 1346 |
+
],
|
| 1347 |
+
"page_idx": 11
|
| 1348 |
+
},
|
| 1349 |
+
{
|
| 1350 |
+
"type": "text",
|
| 1351 |
+
"text": "Andrei A Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy distillation. In International Conference on Learning Representations, 2016. URL https: //arxiv.org/pdf/1511.06295.pdf. ",
|
| 1352 |
+
"bbox": [
|
| 1353 |
+
174,
|
| 1354 |
+
801,
|
| 1355 |
+
826,
|
| 1356 |
+
858
|
| 1357 |
+
],
|
| 1358 |
+
"page_idx": 11
|
| 1359 |
+
},
|
| 1360 |
+
{
|
| 1361 |
+
"type": "text",
|
| 1362 |
+
"text": "Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Individualized controlled continuous communication model for multiagent cooperative and competitive tasks. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id= rye7knCqK7. ",
|
| 1363 |
+
"bbox": [
|
| 1364 |
+
174,
|
| 1365 |
+
867,
|
| 1366 |
+
825,
|
| 1367 |
+
924
|
| 1368 |
+
],
|
| 1369 |
+
"page_idx": 11
|
| 1370 |
+
},
|
| 1371 |
+
{
|
| 1372 |
+
"type": "text",
|
| 1373 |
+
"text": "Kenny Smith, Henry Brighton, and Simon Kirby. Complex Systems In Language Evolution: The Cultural Emergence Of Compositional Structure. Advances in Complex Systems (ACS), 6(04): 537–558, 2003. doi: 10.1142/S0219525903001055. URL https://ideas.repec.org/a/ wsi/acsxxx/v06y2003i04ns0219525903001055.html. ",
|
| 1374 |
+
"bbox": [
|
| 1375 |
+
176,
|
| 1376 |
+
103,
|
| 1377 |
+
825,
|
| 1378 |
+
159
|
| 1379 |
+
],
|
| 1380 |
+
"page_idx": 12
|
| 1381 |
+
},
|
| 1382 |
+
{
|
| 1383 |
+
"type": "text",
|
| 1384 |
+
"text": "Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learning multiagent communication with backpropagation. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29, pp. 2244–2252. Curran Associates, Inc., 2016. URL http://papers.nips.cc/paper/6398-learning-multiagentcommunication-with-backpropagation.pdf. ",
|
| 1385 |
+
"bbox": [
|
| 1386 |
+
174,
|
| 1387 |
+
169,
|
| 1388 |
+
826,
|
| 1389 |
+
239
|
| 1390 |
+
],
|
| 1391 |
+
"page_idx": 12
|
| 1392 |
+
},
|
| 1393 |
+
{
|
| 1394 |
+
"type": "text",
|
| 1395 |
+
"text": "Olivier Tieleman, Angeliki Lazaridou, Shibl Mourad, Charles Blundell, and Doina Precup. Shaping representations through communication. 2018. URL https://openreview.net/pdf?id $\\underline { { \\underline { { \\mathbf { \\Pi } } } } } =$ HkzL4hR9Ym. ",
|
| 1396 |
+
"bbox": [
|
| 1397 |
+
174,
|
| 1398 |
+
248,
|
| 1399 |
+
825,
|
| 1400 |
+
290
|
| 1401 |
+
],
|
| 1402 |
+
"page_idx": 12
|
| 1403 |
+
},
|
| 1404 |
+
{
|
| 1405 |
+
"type": "text",
|
| 1406 |
+
"text": "A HYPERPARAMETERS ",
|
| 1407 |
+
"text_level": 1,
|
| 1408 |
+
"bbox": [
|
| 1409 |
+
176,
|
| 1410 |
+
102,
|
| 1411 |
+
382,
|
| 1412 |
+
118
|
| 1413 |
+
],
|
| 1414 |
+
"page_idx": 13
|
| 1415 |
+
},
|
| 1416 |
+
{
|
| 1417 |
+
"type": "text",
|
| 1418 |
+
"text": "We provide hyperparameter details in Table 1. ",
|
| 1419 |
+
"bbox": [
|
| 1420 |
+
174,
|
| 1421 |
+
133,
|
| 1422 |
+
473,
|
| 1423 |
+
148
|
| 1424 |
+
],
|
| 1425 |
+
"page_idx": 13
|
| 1426 |
+
},
|
| 1427 |
+
{
|
| 1428 |
+
"type": "table",
|
| 1429 |
+
"img_path": "images/564a2779d7753bd29bf3fe55dfbfad84a350622995a0015a3073d1c48e76ea41.jpg",
|
| 1430 |
+
"table_caption": [
|
| 1431 |
+
"Table 1: Hyperparameters considered in S2P training. "
|
| 1432 |
+
],
|
| 1433 |
+
"table_footnote": [],
|
| 1434 |
+
"table_body": "<table><tr><td>Hyperparameter</td><td>Values</td></tr><tr><td>Learning rate</td><td>1e-2,1e-3,2e-3, 6e-3,1e-4, 5e-4,6e-4</td></tr><tr><td>Model architecture</td><td>Linear, Bilinear, Non-Linear</td></tr><tr><td>Number of encoders (perfect emcomm)</td><td>1,2,5,10,20,50,100,200,500,1000</td></tr><tr><td>Hidden layer size (Linear)</td><td>200,500,1000</td></tr><tr><td>Number of encoders (Pop-S2P)</td><td>20,40,50,60,80,100</td></tr><tr><td>Number of distractors</td><td>1,4,9</td></tr><tr><td>GRU hidden size Word embedding size</td><td>256</td></tr><tr><td>Image embedding size (from pretrained Resnet50)</td><td>512 2048</td></tr><tr><td>Batch size</td><td>1,512,1000</td></tr><tr><td>Random seeds</td><td></td></tr><tr><td>Optimizer</td><td>0,1, 2,3,4</td></tr><tr><td>Dropout</td><td>Adam, SGD</td></tr><tr><td>Gumbel relaxation temperature</td><td>0, 0.3</td></tr><tr><td>Vocabulary size</td><td>1</td></tr><tr><td>Max sentence length</td><td>100,200,500,1000,5000</td></tr><tr><td>m in sched</td><td>12,15,20,30, 50</td></tr><tr><td>lin sched</td><td>0,1, 30,40, 50, 70</td></tr><tr><td></td><td>0,30,40,50</td></tr><tr><td>qin rand</td><td>0.75</td></tr><tr><td>rinsched_rand_frz</td><td>0.5</td></tr><tr><td>Number of initial supervised steps (pretraining)</td><td>0,1000,2000,3000,5000</td></tr></table>",
|
| 1435 |
+
"bbox": [
|
| 1436 |
+
181,
|
| 1437 |
+
160,
|
| 1438 |
+
805,
|
| 1439 |
+
468
|
| 1440 |
+
],
|
| 1441 |
+
"page_idx": 13
|
| 1442 |
+
},
|
| 1443 |
+
{
|
| 1444 |
+
"type": "text",
|
| 1445 |
+
"text": "B CALCULATION OF OPTIMAL SAMPLE COMPLEXITY IN OR GAME ",
|
| 1446 |
+
"text_level": 1,
|
| 1447 |
+
"bbox": [
|
| 1448 |
+
173,
|
| 1449 |
+
523,
|
| 1450 |
+
738,
|
| 1451 |
+
539
|
| 1452 |
+
],
|
| 1453 |
+
"page_idx": 13
|
| 1454 |
+
},
|
| 1455 |
+
{
|
| 1456 |
+
"type": "text",
|
| 1457 |
+
"text": "Here we provide a quick calculation for how quickly a human might learn a new compositional language $L$ in the OR game in as few examples as possible, which we use as a baseline in Figure 4a. We assume a OR game with $p = 6$ properties, $t = 1 0$ types, $T = 6$ words sent per message (concatenated together), and $| V | = 6 0$ vocabulary size. If this language $L$ is compositional, then each word in the vocabulary is assigned to 1 type. Thus, we need to learn 60 total assignments. In this analysis we assume we can construct (i.e. hand-design) the samples seen by the human, and thus the final number should be considered something like a lower bound. ",
|
| 1458 |
+
"bbox": [
|
| 1459 |
+
173,
|
| 1460 |
+
554,
|
| 1461 |
+
825,
|
| 1462 |
+
651
|
| 1463 |
+
],
|
| 1464 |
+
"page_idx": 13
|
| 1465 |
+
},
|
| 1466 |
+
{
|
| 1467 |
+
"type": "text",
|
| 1468 |
+
"text": "Since $T = 6$ , we get information about 6 word type assignments for every sample. However, this information is entangled as we don’t know which word corresponded to which type. Thus, we (1) divide the problem up by first constructing 9 (word sequence, object) sample pairs where none of the object types overlap between each sample. With this information, we are able to narrow down the word $\\gets$ type assignments into 10 groups of 6 (that is, in each group we have 6 words corresponding to 6 types, but we don’t know which type belongs to which word). Note we don’t need 10 samples as the last one can be inferred by exclusion. (2) We then construct 5 more samples where each type belongs to a separate group. We can do this because $t > p$ . Because each type belongs to a separate group, cross-referencing the words observed from samples in (1) and (2) uniquely defines each word $\\gets$ type assignment. Note again we don’t need 6 samples as the last one can be inferred by exclusion. This gives us a total of $9 + 5 = 1 4$ samples. ",
|
| 1469 |
+
"bbox": [
|
| 1470 |
+
173,
|
| 1471 |
+
659,
|
| 1472 |
+
825,
|
| 1473 |
+
811
|
| 1474 |
+
],
|
| 1475 |
+
"page_idx": 13
|
| 1476 |
+
},
|
| 1477 |
+
{
|
| 1478 |
+
"type": "text",
|
| 1479 |
+
"text": "C ADDITIONAL PLOTS ",
|
| 1480 |
+
"text_level": 1,
|
| 1481 |
+
"bbox": [
|
| 1482 |
+
174,
|
| 1483 |
+
103,
|
| 1484 |
+
375,
|
| 1485 |
+
117
|
| 1486 |
+
],
|
| 1487 |
+
"page_idx": 14
|
| 1488 |
+
},
|
| 1489 |
+
{
|
| 1490 |
+
"type": "text",
|
| 1491 |
+
"text": "We show training curves for various S2P schedules. ",
|
| 1492 |
+
"bbox": [
|
| 1493 |
+
173,
|
| 1494 |
+
133,
|
| 1495 |
+
513,
|
| 1496 |
+
148
|
| 1497 |
+
],
|
| 1498 |
+
"page_idx": 14
|
| 1499 |
+
},
|
| 1500 |
+
{
|
| 1501 |
+
"type": "image",
|
| 1502 |
+
"img_path": "images/c4c1eb663c3375de094fa31fe6c4a2f7ae3798f2e0646233d035a38a703b1fd7.jpg",
|
| 1503 |
+
"image_caption": [
|
| 1504 |
+
"Figure 8: Training curves for various S2P methods in the IBR game described in $\\ S 4$ "
|
| 1505 |
+
],
|
| 1506 |
+
"image_footnote": [],
|
| 1507 |
+
"bbox": [
|
| 1508 |
+
176,
|
| 1509 |
+
161,
|
| 1510 |
+
823,
|
| 1511 |
+
433
|
| 1512 |
+
],
|
| 1513 |
+
"page_idx": 14
|
| 1514 |
+
}
|
| 1515 |
+
]
|
parse/train/rJxGLlBtwH/rJxGLlBtwH_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/rJxGLlBtwH/rJxGLlBtwH_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|