AAAI 2021 Papers

The reinforcement learning (RL) framework formalizes the notion of learning with interactions. Many real-world problems have large state-spaces and/or action-spaces such as in Go, StarCraft, protein folding, and robotics or are non-Markovian, which cause significant challenges to RL algorithms. In this work we address the large action-space problem by sequentializing actions, which can reduce the action-space size significantly, even down to two actions at the expense of an increased planning horizon. We provide explicit and exact constructions and equivalence proofs for all quantities of interest for arbitrary history-based processes. In the case of MDPs, this could help RL algorithms that bootstrap. In this work we show how action-binarization in the non-MDP case can significantly improve Extreme State Aggregation (ESA) bounds. ESA allows casting any (non-MDP, non-ergodic, history-based) RL problem into a fixed-sized non-Markovian state-space with the help of a surrogate Markovian process. On the upside, ESA enjoys similar optimality guarantees as Markovian models do. But a downside is that the size of the aggregated state-space becomes exponential in the size of the action-space. In this work, we patch this issue by binarizing the action-space. We provide an upper bound on the number of states of this binarized ESA that is logarithmic in the original action-space size, a double-exponential improvement.

Algorithmic fairness has attracted significant attention in recent years, with many quantitative measures suggested for characterizing the fairness of different machine learning algorithms. Despite this interest, the robustness of those fairness measures with respect to an intentional adversarial attack has not been properly addressed. Indeed, most adversarial machine learning has focused on the impact of malicious attacks on the accuracy of the system, without any regard to the system's fairness. We propose new types of data poisoning attacks where an adversary intentionally targets the fairness of a system. Specifically, we propose two families of attacks that target fairness measures. In the anchoring attack, we skew the decision boundary by placing poisoned points near specific target points to bias the outcome. In the influence attack on fairness, we aim to maximize the covariance between the sensitive attributes and the decision outcome and affect the fairness of the model. We conduct extensive experiments that indicate the effectiveness of our proposed attacks.

Mitigating the risk arising from extreme events is a fundamental goal with many applications, such as the modelling of natural disasters, financial crashes, epidemics, and many others. To manage this risk, a vital step is to be able to understand or generate a wide range of extreme scenarios. Existing approaches based on Generative Adversarial Networks (GANs) excel at generating realistic samples, but seek to generate typical samples, rather than extreme samples. Hence, in this work, we propose ExGAN, a GAN-based approach to generate realistic and extreme samples. To model the extremes of the training distribution in a principled way, our work draws from Extreme Value Theory (EVT), a probabilistic approach for modelling the extreme tails of distributions. For practical utility, our framework allows the user to specify both the desired extremeness measure, as well as the desired extremeness probability they wish to sample at. Experiments on real US Precipitation data show that our method generates realistic samples, based on visual inspection and quantitative measures, in an efficient manner. Moreover, generating increasingly extreme examples using ExGAN can be done in constant time (with respect to the extremeness probability τ), as opposed to the O(1/τ) time required by the baseline approach.

The predominant paradigm in evolutionary game theory and more generally online learning in games is based on a clear distinction between a population of dynamic agents that interact given a fixed, static game. In this paper, we move away from the artificial divide between dynamic agents and static games, to introduce and analyze a large class of competitive settings where both the agents and the games they play evolve strategically over time. We focus on arguably the most archetypal game-theoretic setting---zero-sum games (as well as network generalizations)---and the most studied evolutionary learning dynamic---replicator, the continuous-time analogue of multiplicative weights. Populations of agents compete against each other in a zero-sum competition that itself evolves adversarially to the current population mixture. Remarkably, despite the chaotic coevolution of agents and games, we prove that the system exhibits a number of regularities. First, the system has conservation laws of an information-theoretic flavor that couple the behavior of all agents and games. Secondly, the system is Poincare recurrent, with effectively all possible initializations of agents and games lying on recurrent orbits that come arbitrarily close to their initial conditions infinitely often. Thirdly, the time-average agent behavior and utility converge to the Nash equilibrium values of the time-average game. Finally, we provide a polynomial time algorithm to efficiently predict this time-average behavior for any such coevolving network game.

Data augmentation is one of the most effective ways to stabilize learning by improving the generalization of machine-learning models. In recent years, automatic data augmentation methods, such as AutoAugment or Fast AutoAugment have been attracting attention; and these methods improved the results of image classification and object detection tasks. However, several problems remain. Most notably, a larger training dataset requires higher computational costs. When searching with a small dataset in an attempt to determine the data augmentation approach, the true data space and sampling data space do not fully correspond with each other, thereby causing the generalization performance to deteriorate. Moreover, in the existing automatic augmentation methods, the search phase is often dominated by an exceptional sub-policy, which results in a loss of diversity of operations. In this study, we solved these problems by introducing evolutionary computation to previous methods. As mentioned earlier, maintaining diversity is essential. Therefore, we adopted the thermodynamical genetic algorithm (TDGA), which can control the population diversity with a specific genetic operator, known as the thermodynamical selection rule. To confirm the effectiveness of the proposed method, computational experiments were conducted using two benchmark datasets, CIFAR-10 and SVHN, as examples. The experimental results show that the proposed method can obtain various useful augmentation sub-policies for the problems while reducing the computational cost.

We address the problem of solving complex Bayesian games, characterized by high-dimensional type and action spaces, many (> 2) players, and general-sum payoffs. Our approach applies to symmetric one-shot Bayesian games, with no given analytic structure. We represent agent strategies in parametric form as neural networks, and apply natural evolution strategies (NES) [wierstra2014natural] for deep model optimization. For pure equilibrium computation, we formulate the problem as bi-level optimization, and employ NES in an iterative algorithm to implement both inner-loop best response optimization and outer-loop regret minimization. In simple games including first- and second-price auctions, it is capable of recovering known analytic solutions. For mixed equilibrium computation, we adopt an incremental strategy generation framework, with NES as strategy generator producing a finite sequence of approximate best-response strategies. We then calculate equilibria over this finite strategy set via a model-based optimization process. Both our pure and mixed equilibrium computation methods employ NES to efficiently search for strategies over the functional space, given only black-box simulation access to noisy payoff samples. We experimentally demonstrate the efficacy of all methods on two simultaneous sealed-bid auction games with distinct type distributions, and observe that the solutions exhibit qualitatively different behavior in these two environments.

Existing approaches construct appropriate interaction models to explore semantic conflicts between claims and relevant articles, which provides practical solutions for interpretable claim verification. However, these conflicts are not necessarily all about questioning the false part of claims, which makes considerable semantic conflicts difficult to be used as evidence to explain the results of claim verification. In this paper, we propose evidence inference networks (EVIN), which focus on the conflicts questioning the core semantics of claims and serve as evidence for interpretable claim verification. Specifically, EVIN first captures the core semantic segments of claims and the users' principal opinions in relevant articles. Then, it finely-grained identifies the semantic conflicts contained in each relevant article from these opinions. Finally, it constructs coherence modeling to match the conflicts that queries the core semantic fragments of claims as explainable evidence. Experiments on two widely used datasets demonstrate that EVIN not only achieves satisfactory performance but also provides explainable evidence for end-users.

As one of the most powerful topic models, Latent Dirichlet Allocation (LDA) has been used in a vast range of tasks, including document understanding, information retrieval and peer-reviewer assignment. Despite its tremendous popularity, the security of LDA has rarely been studied. This poses severe risks to security-critical tasks such as sentiment analysis and peer-reviewer assignment that are based on LDA. In this paper, we are interested in knowing whether LDA models are vulnerable to adversarial perturbations of benign document examples during inference time. We formalize the evasion attack to LDA models as an optimization problem and prove it to be NP-hard. We then propose a novel and efficient algorithm, EvaLDA to solve it. We show the effectiveness of EvaLDA via extensive empirical evaluations. For instance, in the NIPS dataset, EvaLDA can averagely promote the rank of a target topic from 10 to around 7 by only replacing 1% of the words with similar words in a victim document. Our work provides significant insights into the power and limitations of evasion attacks to LDA models.

Enabling autonomous systems to comply with an ethical theory is critical given their accelerating deployment in domains that impact society. While many ethical theories have been studied extensively in moral philosophy, they are still challenging to implement by developers who build autonomous systems. This paper proposes a novel approach for building ethically compliant autonomous systems that optimize completing a task while following an ethical framework. First, we introduce a definition of an ethically compliant autonomous system and its properties. Next, we offer a range of ethical frameworks for divine command theory, prima facie duties, and virtue ethics. Finally, we demonstrate the accuracy and usability of our approach in a set of autonomous driving simulations and a user study of planning and robotics experts.

An agent, or a coalition of agents, faces an ethical dilemma between several statements if she is forced to make a conscious choice between which of these statements will be true. This paper proposes to capture ethical dilemmas as a modality in strategic game settings with and without limit on sacrifice and for perfect and imperfect information games. The authors show that the dilemma modality cannot be defined through the earlier proposed blameworthiness modality. The main technical result is a sound and complete axiomatization of the properties of this modality with sacrifice in games with perfect information.

We consider the problem of estimating a spectral risk measure (SRM) from i.i.d. samples, and propose a novel method that is based on numerical integration. We show that our SRM estimate concentrates exponentially, when the underlying distribution has bounded support. Further, we also consider the case when the underlying distribution satisfies an exponential moment bound, which includes sub-Gaussian and subexponential distributions. For these distributions, we derive a concentration bound for our estimation scheme. We validate the theoretical findings on a synthetic setup, and in a vehicular traffic routing application.

Game theory has been increasingly applied in settings where the game is not known outright, but has to be estimated by sampling. For example, meta-games that arise in multi-agent evaluation can only be accessed by running a succession of expensive experiments that may involve simultaneous deployment of several agents. In this paper, we focus on α-rank, a popular game-theoretic solution concept designed to perform well in such scenarios. We aim to estimate the α-rank of the game using as few samples as possible. Our algorithm maximizes information gain between an epistemic belief over the α-ranks and the observed payoff. This approach has two main benefits. First, it allows us to focus our sampling on the entries that matter the most for identifying the α-rank. Second, the Bayesian formulation provides a facility to build in modeling assumptions by using a prior over game payoffs. We show the benefits of using information gain as compared to the confidence interval criterion of ResponseGraphUCB, and provide theoretical results justifying our method.

We consider a natural setting where network parameters are estimated from noisy and incomplete information about the network. More specifically, we investigate how we can efficiently estimate the number of small subgraphs (e.g., edges, triangles, etc.) based on full access to one or two noisy and incomplete samples of a large underlying network and on few queries revealing the neighborhood of carefully selected vertices. After specifying a random generator which removes edges from the underlying graph, we present estimators with strong provable performance guarantees, which exploit information from the noisy network samples and query a constant number of the most important vertices for the estimation. Our experimental evaluation shows that, in practice, a single noisy network sample and a couple of hundreds neighborhood queries suffice for accurately estimating the number of triangles in networks with millions of vertices and edges.

Identifying causal effects from observational data is a pervasive challenge found throughout the empirical sciences. Very general methods have been developed to decide the identifiability of a causal quantity from a combination of observational data and causal knowledge about the underlying system. In practice, however, there are still challenges to estimating identifiable causal functionals from finite samples. Recently, a method known as double/debiased machine learning (DML) (Chernozhukov et al. 2018) has been proposed to learn parameters leveraging modern machine learning techniques, which is both robust to model misspecification and bias-reducing. Still, DML has only been used for causal estimation in settings when the back-door condition (also known as conditional ignorability) holds. In this paper, we develop a new, general class of estimators for any identifiable causal functionals that exhibit DML properties, which we name DML-ID. In particular, we introduce a complete identification algorithm that returns an influence function (IF) for any identifiable causal functional. We then construct the DML estimator based on the derived IF. We show that DML-ID estimators hold the key properties of debiasedness and doubly robustness. Simulation results corroborate with the theory.

In survival analysis, deep learning approaches have been proposed for estimating an individual's probability of survival over some time horizon. Such approaches can capture complex non-linear relationships, without relying on restrictive assumptions regarding the relationship between an individual's characteristics and their underlying survival process. To date, however, these methods have focused primarily on optimizing discriminative performance and have ignored model calibration. Well-calibrated survival curves present realistic and meaningful probabilistic estimates of the true underlying survival process for an individual. However, due to the lack of ground-truth regarding the underlying stochastic process of survival for an individual, optimizing and measuring calibration in survival analysis is an inherently difficult task. In this work, we i) highlight the shortcomings of existing approaches in terms of calibration and ii) propose a new training scheme for optimizing deep survival analysis models that maximizes discriminative performance, subject to good calibration. Compared to state-of-the-art approaches across two publicly available datasets, our proposed training scheme leads to significant improvements in calibration, while maintaining good discriminative performance.

Most discrete evolutionary algorithms (EAs) implement elitism, meaning that they make the biologically implausible assumption that the fittest individuals never die. While elitism favours exploitation and ensures that the best seen solutions are not lost, it has been widely conjectured that non-elitism is necessary to explore promising fitness valleys without getting stuck in local optima. Determining when non-elitist EAs outperform elitist EAs has been one of the most fundamental open problems in evolutionary computation. A recent analysis of a non-elitist EA shows that this algorithm does not outperform its elitist counterparts on the benchmark problem JUMP. We solve this open problem through rigorous runtime analysis of elitist and non-elitist population-based EAs on a class of multi-modal problems. We show that with 3-tournament selection and appropriate mutation rates, the non-elitist EA optimises the multi-modal problem in expected polynomial time, while an elitist EA requires exponential time with overwhelmingly high probability. A key insight in our analysis is the non-linear selection profile of the tournament selection mechanism which, with appropriate mutation rates, allows a small sub-population to reside on the local optimum while the rest of the population explores the fitness valley. In contrast, we show that the comma-selection mechanism which does not have this non-linear profile, fails to optimise this problem in polynomial time. The theoretical analysis is complemented with an empirical investigation on instances of the set cover problem, showing that non-elitist EAs can perform better than the elitist ones. We also provide examples where usage of mutation rates close to the error thresholds is beneficial when employing non-elitist population-based EAs.

Though deep learning has been applied successfully in many scenarios, malicious inputs with human-imperceptible perturbations can make it vulnerable in real applications. This paper proposes an error-correcting neural network (ECNN) that combines a set of binary classifiers to combat adversarial examples in the multi-class classification problem. To build an ECNN, we propose to design a code matrix so that the minimum Hamming distance between any two rows (i.e., two codewords) and the minimum shared information distance between any two columns (i.e., two partitions of class labels) are simultaneously maximized. Maximizing row distances can increase the system fault tolerance while maximizing column distances helps increase the diversity between binary classifiers. We propose an end-to-end training method for our ECNN, which allows further improvement of the diversity between binary classifiers. The end-to-end training renders our proposed ECNN different from the traditional error-correcting output code (ECOC) based methods that train binary classifiers independently. ECNN is complementary to other existing defense approaches such as adversarial training and can be applied in conjunction with them. We empirically demonstrate that our proposed ECNN is effective against the state-of-the-art white-box and black-box attacks on several datasets while maintaining good classification accuracy on normal examples.

This paper focuses on the unsupervised domain adaptation problem for video-based crowd counting, in which we use labeled data as source domain and unlabelled video data as target domain. It is challenging as there is a huge gap between the source and the target domain and no annotations of samples are available in the target domain. The key issue is how to utilize unlabelled videos in the target domain for knowledge learning and transferring from the source domain. To tackle this problem, we propose a novel Error-aware Density Isomorphism REConstruction Network (EDIREC-Net) for cross-domain crowd counting. EDIREC-Net jointly transfers a pre-trained counting model to target domains using a density isomorphism reconstruction objective and models the reconstruction erroneousness by error reasoning. Specifically, as crowd flows in videos are consecutive, the density maps in adjacent frames turn out to be isomorphic. On this basis, we regard the density isomorphism reconstruction error as a self-supervised signal to transfer the pre-trained counting models to different target domains. Moreover, we leverage an estimation-reconstruction consistency to monitor the density reconstruction erroneousness and suppress unreliable density reconstructions during training. Experimental results on four benchmark datasets demonstrate the superiority of the proposed method and ablation studies investigate the efficiency and robustness. The source code is available at https://github.com/GehenHe/EDIREC-Net.

The aim of this paper is to offer the first systematic exploration and definition of equivalent causal models in the context where both models are not made up of the same variables. The idea is that two models are equivalent when they agree on all "essential" causal information that can be expressed using their common variables. I do so by focussing on the two main features of causal models, namely their structural relations and their functional relations. In particular, I define several relations of causal ancestry and several relations of causal sufficiency, and require that the most general of these relations are preserved across equivalent models.