AAAI 2021 论文列表

Adversarial examples are input examples that are specifically crafted to deceive machine learning classifiers. State-of-the-art adversarial example detection methods characterize an input example as adversarial either by quantifying the magnitude of feature variations under multiple perturbations or by measuring its distance from estimated benign example distribution. Instead of using such metrics, the proposed method is based on the observation that the directions of adversarial gradients when crafting (new) adversarial examples play a key role in characterizing the adversarial space. Compared to detection methods that use multiple perturbations, the proposed method is efficient as it only applies a single random perturbation on the input example. Experiments conducted on two different databases, CIFAR-10 and ImageNet, show that the proposed detection method achieves, respectively, 97.9% and 98.6% AUC-ROC (on average) on five different adversarial attacks, and outperforms multiple state-of-the-art detection methods. Results demonstrate the effectiveness of using adversarial gradient directions for adversarial example detection.

Bayesian persuasion, as introduced by Kamenica and Gentzkow in 2011, is the study of information sharing policies among strategic agents. A prime example is signaling in online ad auctions: what information should a platform signal to an advertiser regarding a user when selling the opportunity to advertise to her? Practical considerations such as preventing discrimination, protecting privacy or acknowledging limited attention of the information receiver impose constraints on information sharing. We propose a simple way to mathematically model such constraints as restrictions on Receiver's admissible posterior beliefs. We consider two families of constraints - ex ante and ex post; the latter limits each instance of Sender-Receiver communication, while the former more general family can also pose restrictions in expectation. For the ex ante family, a result of Doval and Skreta (2018) establishes the existence of an optimal signaling scheme with a small number of signals - at most the number of constraints plus the number of states of nature - and we show this result is tight. For the ex post family, we tighten the previous bound of Vølund (2018), showing that the required number of signals is at most the number of states of nature, as in the original Kamenica-Gentzkow setting. As our main algorithmic result, we provide an additive bi-criteria FPTAS for an optimal constrained signaling scheme assuming a constant number of states of nature; we improve the approximation to single-criteria under a Slater-like regularity condition. The FPTAS holds under standard assumptions, and more relaxed assumptions yield a PTAS. We then establish a bound on the ratio between Sender's optimal utility under convex ex ante constraints and the corresponding ex post constraints. We demonstrate how this result can be applied to find an approximately welfare-maximizing constrained signaling scheme in ad auctions.

Online solvers for partially observable Markov decision processes have difficulty scaling to problems with large action spaces. Monte Carlo tree search with progressive widening attempts to improve scaling by sampling from the action space to construct a policy search tree. The performance of progressive widening search is dependent upon the action sampling policy, often requiring problem-specific samplers. In this work, we present a general method for efficient action sampling based on Bayesian optimization. The proposed method uses a Gaussian process to model a belief over the action-value function and selects the action that will maximize the expected improvement in the optimal action value. We implement the proposed approach in a new online tree search algorithm called Bayesian Optimized Monte Carlo Planning (BOMCP). Several experiments show that BOMCP is better able to scale to large action space POMDPs than existing state-of-the-art tree search solvers.

Dynamic mode decomposition (DMD) and its extensions are data-driven methods that have substantially contributed to our understanding of dynamical systems. However, because DMD and most of its extensions are deterministic, it is difficult to treat probabilistic representations of parameters and predictions. In this work, we propose a novel formulation of a Bayesian DMD model. Our Bayesian DMD model is consistent with the procedure of standard DMD, which is to first determine the subspace of observations, and then compute the modes on that subspace. Variational matrix factorization makes it possible to realize a fully-Bayesian scheme of DMD. Moreover, we derive a Bayesian DMD model for incomplete data, which demonstrates the advantage of probabilistic modeling. Finally, both of nonlinear simulated and real-world datasets are used to illustrate the potential of the proposed method.

Distributional Reinforcement Learning (RL) maintains the entire probability distribution of the reward-to-go, i.e. the return, providing more learning signals that account for the uncertainty associated with policy performance, which may be beneficial for trading off exploration and exploitation and policy learning in general. Previous works in distributional RL focused mainly on computing the state-action-return distributions, here we model the state-return distributions. This enables us to translate successful conventional RL algorithms that are based on state values into distributional RL. We formulate the distributional Bellman operation as an inference-based auto-encoding process that minimises Wasserstein metrics between target/model return distributions. The proposed algorithm, BDPG (Bayesian Distributional Policy Gradients), uses adversarial training in joint-contrastive learning to estimate a variational posterior from the returns. Moreover, we can now interpret the return prediction uncertainty as an information gain, which allows to obtain a new curiosity measure that helps BDPG steer exploration actively and efficiently. We demonstrate in a suite of Atari 2600 games and MuJoCo tasks, including well known hard-exploration challenges, how BDPG learns generally faster and with higher asymptotic performance than reference distributional RL algorithms.

Post-hoc explanation methods are gaining popularity for interpreting, understanding, and debugging neural networks. Most analyses using such methods explain decisions in response to inputs drawn from the test set. However, the test set may have few examples that trigger some model behaviors, such as high-confidence failures or ambiguous classifications. To address these challenges, we introduce a flexible model inspection framework: Bayes-TrEx. Given a data distribution, Bayes-TrEx finds in-distribution examples which trigger a specified prediction confidence. We demonstrate several use cases of Bayes-TrEx, including revealing highly confident (mis)classifications, visualizing class boundaries via ambiguous examples, understanding novel-class extrapolation behavior, and exposing neural network overconfidence. We use Bayes-TrEx to study classifiers trained on CLEVR, MNIST, and Fashion-MNIST, and we show that this framework enables more flexible holistic model analysis than just inspecting the test set. Code and supplemental material are available at https://github.com/serenabooth/Bayes-TrEx.

Randomized algorithms are used in many state-of-the-art solvers for constraint satisfaction problems (CSP) and Boolean satisfiability (SAT) problems. For many of these problems, there is no single solver which will dominate others. Having access to the underlying runtime distributions (RTD) of these solvers can allow for better use of algorithm selection, algorithm portfolios, and restart strategies. Previous state-of-the-art methods directly try to predict a fixed parametric distribution that the input instance follows. In this paper, we extend RTD prediction models into the Bayesian setting for the first time. This new model achieves robust predictive performance in the low observation setting, as well as handling censored observations. This technique also allows for richer representations which cannot be achieved by the classical models which restrict their output representations. Our model outperforms the previous state-of-the-art model in settings in which data is scarce, and can make use of censored data such as lower bound time estimates, where that type of data would otherwise be discarded. It can also quantify its uncertainty in its predictions, allowing for algorithm portfolio models to make better informed decisions about which algorithm to run on a particular instance.

Tree-form sequential decision making (TFSDM) extends classical one-shot decision making by modeling tree-form interactions between an agent and a potentially adversarial environment. It captures the online decision-making problems that each player faces in an extensive-form game, as well as Markov decision processes and partially-observable Markov decision processes where the agent conditions on observed history. Over the past decade, there has been considerable effort into designing online optimization methods for TFSDM. Virtually all of that work has been in the full-feedback setting, where the agent has access to counterfactuals, that is, information on what would have happened had the agent chosen a different action at any decision node. Little is known about the bandit setting, where that assumption is reversed (no counterfactual information is available), despite this latter setting being well understood for almost 20 years in one-shot decision making. In this paper, we give the first algorithm for the bandit linear optimization problem for TFSDM that offers both (i) linear-time iterations (in the size of the decision tree) and (ii) O(sqrt(T)) cumulative regret in expectation compared to any fixed strategy, at all times T. This is made possible by new results that we derive, which may have independent uses as well: 1) geometry of the dilated entropy regularizer, 2) autocorrelation matrix of the natural sampling scheme for sequence-form strategies, 3) construction of an unbiased estimator for linear losses for sequence-form strategies, and 4) a refined regret analysis for mirror descent when using the dilated entropy regularizer.

Open Set Domain Adaptation (OSDA) is a challenging domain adaptation setting which allows the existence of unknown classes on the target domain. Although existing OSDA methods are good at classifying samples of known classes, they ignore the classification ability for the unknown samples, making them unbalanced OSDA methods. To alleviate this problem, we propose a balanced OSDA methods which could recognize the unknown samples while maintain high classification performance for the known samples. Specifically, to reduce the domain gaps, we first project the features to a hyperspherical latent space. In this space, we propose to bound the centroid deviation angles to not only increase the intra-class compactness but also enlarge the inter-class margins. With the bounded centroid deviation angles, we employ the statistical Extreme Value Theory to recognize the unknown samples that are misclassified into known classes. In addition, to learn better centroids, we propose an improved centroid update strategy based on sample reweighting and adaptive update rate to cooperate with centroid alignment. Experimental results on three OSDA benchmarks verify that our method can significantly outperform the compared methods and reduce the proportion of the unknown samples being misclassified into known classes.

In recent years, visual recognition on challenging long-tailed distributions, where classes often exhibit extremely imbalanced frequencies, has made great progress mostly based on various complex paradigms (e.g., meta learning). Apart from these complex methods, simple refinements on training procedures also make contributions. These refinements, also called tricks, are minor but effective, such as adjustments in the data distribution or loss functions. However, different tricks might conflict with each other. If users apply these long-tail related tricks inappropriately, it could cause worse recognition accuracy than expected. Unfortunately, there has not been a scientific guideline of these tricks in the literature. In this paper, we first collect existing tricks in long-tailed visual recognition and then perform extensive and systematic experiments, in order to give a detailed experimental guideline and obtain an effective combination of these tricks. Furthermore, we also propose a novel data augmentation approach based on class activation maps for long-tailed recognition, which can be friendly combined with re-sampling methods and shows excellent results. By assembling these tricks scientifically, we can outperform state-of-the-art methods on four long-tailed benchmark datasets, including ImageNet-LT and iNaturalist 2018. Our code is open-source and available at https://github.com/zhangyongshun/BagofTricks-LT.

We describe a compilation language of backdoor decomposable monotone circuits (BDMCs) which generalizes several concepts appearing in the literature, e.g. DNNFs and backdoor trees. A C-BDMC sentence is a monotone circuit which satisfies decomposability property (such as in DNNF) in which the inputs (or leaves) are associated with CNF encodings from a given base class C. We consider the class of propagation complete (PC) encodings as a base class and we show that PC-BDMCs are polynomially equivalent to PC encodings. Additionally, we use this to determine the properties of PC-BDMCs and PC encodings with respect to the knowledge compilation map including the list of efficient operations on the languages.

Behavior Trees (BTs) have attracted much attention in the robotics field in recent years, which generalize existing control architectures and bring unique advantages for building robot systems. Automated synthesis of BTs can reduce human workload and build behavior models for complex tasks beyond the ability of human design, but theoretical studies are almost missing in existing methods because it is difficult to conduct formal analysis with the classic BT representations. As a result, they may fail in tasks that are actually solvable. This paper proposes BT expansion, an automated planning approach to building intelligent robot behaviors with BTs, and proves the soundness and completeness through the state-space formulation of BTs. The advantages of blended reactive planning and acting are formally discussed through the region of attraction of BTs, by which robots with BT expansion are robust to any resolvable external disturbances. Experiments with a mobile manipulator and test sets are simulated to validate the effectiveness and efficiency, where the proposed algorithm surpasses the baseline by virtue of its soundness and completeness. To the best of our knowledge, it is the first time to leverage the state-space formulation to synthesize BTs with a complete theoretical basis.

Generating human action proposals in untrimmed videos is an important yet challenging task with wide applications. Current methods often suffer from the noisy boundary locations and the inferior quality of confidence scores used for proposal retrieving. In this paper, we present BSN++, a new framework which exploits complementary boundary regressor and relation modeling for temporal proposal generation. First, we propose a novel boundary regressor based on the complementary characteristics of both starting and ending boundary classifiers. Specifically, we utilize the U-shaped architecture with nested skip connections to capture rich contexts and introduce bi-directional boundary matching mechanism to improve boundary precision. Second, to account for the proposal-proposal relations ignored in previous methods, we devise a proposal relation block to which includes two self-attention modules from the aspects of position and channel. Furthermore, we find that there inevitably exists data imbalanced problems in the positive/negative proposals and temporal durations, which harm the model performance on tail distributions. To relieve this issue, we introduce the scale-balanced re-sampling strategy. Extensive experiments are conducted on two popular benchmarks: ActivityNet-1.3 and THUMOS14, which demonstrate that BSN++ achieves the state-of-the-art performance. Not surprisingly, the proposed BSN++ ranked 1st place in the CVPR19 - ActivityNet challenge leaderboard on temporal action localization task.

In this paper, we study the response of large models from the BERT family to incoherent inputs that should confuse any model that claims to understand natural language. We define simple heuristics to construct such examples. Our experiments show that state-of-the-art models consistently fail to recognize them as ill-formed, and instead produce high confidence predictions on them. As a consequence of this phenomenon, models trained on sentences with randomly permuted word order perform close to state-of-the-art models. To alleviate these issues, we show that if models are explicitly trained to recognize invalid inputs, they can be robust to such attacks without a drop in performance.

Over the past half-decade, many methods have been considered for neural architecture search (NAS). Bayesian optimization (BO), which has long had success in hyperparameter optimization, has recently emerged as a very promising strategy for NAS when it is coupled with a neural predictor. Recent work has proposed different instantiations of this framework, for example, using Bayesian neural networks or graph convolutional networks as the predictive model within BO. However, the analyses in these papers often focus on the full-fledged NAS algorithm, so it is difficult to tell which individual components of the framework lead to the best performance. In this work, we give a thorough analysis of the "BO + neural predictor framework" by identifying five main components: the architecture encoding, neural predictor, uncertainty calibration method, acquisition function, and acquisition function optimization. We test several different methods for each component and also develop a novel path-based encoding scheme for neural architectures, which we show theoretically and empirically scales better than other encodings. Using all of our analyses, we develop a final algorithm called BANANAS, which achieves state-of-the-art performance on NAS search spaces. We adhere to the NAS research checklist (Lindauer and Hutter 2019) to facilitate best practices, and our code is available at https://github.com/naszilla/naszilla.

Analysing and computing with Gaussian processes arising from infinitely wide neural networks has recently seen a resurgence in popularity. Despite this, many explicit covariance functions of networks with activation functions used in modern networks remain unknown. Furthermore, while the kernels of deep networks can be computed iteratively, theoretical understanding of deep kernels is lacking, particularly with respect to fixed-point dynamics. Firstly, we derive the covariance functions of multi-layer perceptrons (MLPs) with exponential linear units (ELU) and Gaussian error linear units (GELU) and evaluate the performance of the limiting Gaussian processes on some benchmarks. Secondly, and more generally, we analyse the fixed-point dynamics of iterated kernels corresponding to a broad range of activation functions. We find that unlike some previously studied neural network kernels, these new kernels exhibit non-trivial fixed-point dynamics which are mirrored in finite-width neural networks. The fixed point behaviour present in some networks explains a mechanism for implicit regularisation in overparameterised deep models. Our results relate to both the static iid parameter conjugate kernel and the dynamic neural tangent kernel constructions

Robots operating in the real world must deal with uncertainty, be it due to working with humans who are unpredictable, or simply because they must operate in a dynamic environment. Ignoring the uncertainty is dangerous, while accounting for all possible outcomes is often computationally infeasible. One approach, which lies between ignoring the uncertainty completely and addressing it completely is using flexible plans with choice, formulated as Temporal Planning Networks (TPNs). This method has been successfully demonstrated to work in human-robot teamwork using the Pike executive, an online executive that unifies intent recognition and plan adaptation. However, one of the main challenges to using Pike is the need to manually specify the TPN. In this paper, we address this challenge by describing a technique for automatically synthesizing a TPN which covers multiple possible executions for a given temporal planning problem specified in PDDL 2.1. Our approach starts by using a diverse planner to generate multiple plans, and then merges them into a single TPN. As there were no available diverse planners for temporal planning, we first present a novel method for adapting an existing diverse planning method, based on top-k planning, to the temporal setting. We then describe how merging diverse plans into a single TPN is performed using constraint optimization. Finally, an empirical evaluation on a set of IPC benchmarks shows that our approach scales well, and generates TPNs which can generalize the set of plans they are generated from.

Dialogue policy learning based on reinforcement learning is difficult to be applied to real users to train dialogue agents from scratch because of the high cost. User simulators, which choose random user goals for the dialogue agent to train on, have been considered as an affordable substitute for real users. However, this random sampling method ignores the law of human learning, making the learned dialogue policy inefficient and unstable. We propose a novel framework, Automatic Curriculum Learning-based Deep Q-Network (ACL-DQN), which replaces the traditional random sampling method with a teacher policy model to realize the dialogue policy for automatic curriculum learning. The teacher model arranges a meaningful ordered curriculum and automatically adjusts it by monitoring the learning progress of the dialogue agent and the over-repetition penalty without any requirement of prior knowledge. The learning progress of the dialogue agent reflects the relationship between the dialogue agent's ability and the sampled goals' difficulty for sample efficiency. The over-repetition penalty guarantees the sampled diversity. Experiments show that the ACL-DQN significantly improves the effectiveness and stability of dialogue tasks with a statistically significant margin. Furthermore, the framework can be further improved by equipping with different curriculum schedules, which demonstrates that the framework has strong generalizability.

One of the most exciting applications of modern artificial intelligence is to automatically discover scientific laws from experimental data. This is not a trivial problem as it involves searching for a complex mathematical relationship over a large set of explanatory variables and operators that can be combined in an infinite number of ways. Inspired by the incredible success of deep learning in computer vision, we tackle this problem by adapting various successful network architectures into the symbolic law discovery pipeline. The novelty of our approach is in (1) encoding the input data as an image with super-resolution, (2) developing an appropriate deep network pipeline, and (3) predicting the importance of each mathematical operator from the relationship image. This allows us to prior the exponentially large search with the predicted importance of the symbolic operators, which can significantly accelerate the discovery process. We apply our model to a variety of plausible relationships---both simulated and from physics and mathematics domains---involving different dimensions and constituents. We show that our model is able to identify the underlying operators from data, achieving a high accuracy and AUC (91% and 0.96 on average resp.) for systems with as many as ten independent variables. Our method significantly outperforms the current state of the art in terms of data fitting (R^2), discovery rate (recovering the true relationship), and succinctness (output formula complexity). The discovered equations can be seen as first drafts of scientific laws that can be helpful to the scientists for (1) hypothesis building, and (2) understanding the complex underlying structure of the studied phenomena. Our approach holds a real promise to help speed up the rate of scientific discovery.