AAAI 2021 Papers

AI-systems that model and interact with their users can up-date their models over time to reflect new information and changes in the environment. Although these updates may improve the overall performance of the AI-system, they may actually hurt the performance with respect to individual users. Prior work has studied the tradeoff between improving the system’s performance following an update and the compatibility of the updated system with prior user experience. The more the model is forced to be compatible with a prior version, the higher loss in performance it will incur. This paper challenges this assumption by showing that by personalizing the loss function to specific users, it is possible to increase the prediction performance of the AI-system while sacrificing less compatibility for these users. Our approach updates the sample weights to reflect their contribution to the compatibility of the model for a particular user following the update. We construct a portfolio of different models that vary in how they personalize the loss function for a user. We select the best model to use for a target user based on a validation set. We apply this approach to three supervised learning tasks commonly used in the human-computer decision-making literature. We show that using our approach leads to significant improvements in the performance-compatibility tradeoff over the non-personalized approach of Bansal et al., achieving up to 300% improvement for certain users. We present several use cases that illustrate the difference between the personalized and non-personalized approach for two of our domains.

BERT has set a new state-of-the-art performance on entity resolution (ER) task, largely owed to fine-tuning pre-trained language models and the deep pair-wise interaction. Albeit being remarkably effective, it comes with a steep increase in computational cost, as the deep-interaction requires to exhaustively compute every tuple pair to search for co-references. For ER task, it is often prohibitively expensive due to the large cardinality to be matched. To tackle this, we introduce a siamese network structure that independently encodes tuples using BERT but delays the pair-wise interaction via an enhanced alignment network. This siamese structure enables a dedicated blocking module to quickly filter out obviously dissimilar tuple pairs, and thus drastically reduces the cardinality of fine-grained matching. Further, the blocking and entity matching are integrated into a multi-task learning framework for facilitating both tasks. Extensive experiments on multiple datasets demonstrate that our model significantly outperforms state-of-the-art models (including BERT) in both efficiency and effectiveness.

Code generation aims to automatically generate a piece of code given an input natural language utterance. Currently, among dominant models, it is treated as a sequence-to-tree task, where a decoder outputs a sequence of actions corresponding to the pre-order traversal of an Abstract Syntax Tree. However, such a decoder only exploits the pre-order traversal based preceding actions, which are insufficient to ensure correct action predictions. In this paper, we first throughly analyze the context modeling difference between neural code generation models with different traversals based decodings (preorder traversal vs breadth-first traversal), and then propose to introduce a mutual learning framework to jointly train these models. Under this framework, we continuously enhance both two models via mutual distillation, which involves synchronous executions of two one-to-one knowledge transfers at each training step. More specifically, we alternately choose one model as the student and the other as its teacher, and require the student to fit the training data and the action prediction distributions of its teacher. By doing so, both models can fully absorb the knowledge from each other and thus could be improved simultaneously. Experimental results and in-depth analysis on several benchmark datasets demonstrate the effectiveness of our approach. We release our code at https://github.com/DeepLearnXMU/CGML.

Training an agent to solve control tasks directly from high-dimensional images with model-free reinforcement learning (RL) has proven difficult. A promising approach is to learn a latent representation together with the control policy. However, fitting a high-capacity encoder using a scarce reward signal is sample inefficient and leads to poor performance. Prior work has shown that auxiliary losses, such as image reconstruction, can aid efficient representation learning. However, incorporating reconstruction loss into an off-policy learning algorithm often leads to training instability. We explore the underlying reasons and identify variational autoencoders, used by previous investigations, as the cause of the divergence. Following these findings, we propose effective techniques to improve training stability. This results in a simple approach capable of matching state-of-the-art model-free and model-based algorithms on MuJoCo control tasks. Furthermore, our approach demonstrates robustness to observational noise, surpassing existing approaches in this setting. Code, results, and videos are anonymously available at https://sites.google.com/view/sac-ae/home.

This paper studies defense mechanisms against model inversion (MI) attacks -- a type of privacy attacks aimed at inferring information about the training data distribution given the access to a target machine learning model. Existing defense mechanisms rely on model-specific heuristics or noise injection. While being able to mitigate attacks, existing methods significantly hinder model performance. There remains a question of how to design a defense mechanism that is applicable to a variety of models and achieves better utility-privacy tradeoff. In this paper, we propose the Mutual Information Regularization based Defense (MID) against MI attacks. The key idea is to limit the information about the model input contained in the prediction, thereby limiting the ability of an adversary to infer the private training attributes from the model prediction. Our defense principle is model-agnostic and we present tractable approximations to the regularizer for linear regression, decision trees, and neural networks, which have been successfully attacked by prior work if not attached with any defenses. We present a formal study of MI attacks by devising a rigorous game-based definition and quantifying the associated information leakage. Our theoretical analysis sheds light on the inefficacy of DP in defending against MI attacks, which has been empirically observed in several prior works. Our experiments demonstrate that MID leads to state-of-the-art performance for a variety of MI attacks, target models and datasets.

In addition to high accuracy, robustness is becoming increasingly important for machine learning models in various applications. Recently, much research has been devoted to improving the model robustness by training with noise perturbations. Most existing studies assume a fixed perturbation level for all training examples, which however hardly holds in real tasks. In fact, excessive perturbations may destroy the discriminative content of an example, while deficient perturbations may fail to provide helpful information for improving the robustness. Motivated by this observation, we propose to adaptively adjust the perturbation levels for each example in the training process. Specifically, a novel active learning framework is proposed to allow the model interactively querying the correct perturbation level from human experts. By designing a cost-effective sampling strategy along with a new query type, the robustness can be significantly improved with a few queries. Both theoretical analysis and experimental studies validate the effectiveness of the proposed approach.

In a graph, a k-plex is a vertex set in which every vertex is not adjacent to at most k vertices of this set. The maximum k-plex problem, which asks for the largest k-plex from the given graph, is a key primitive in a variety of real-world applications like community detection and so on. In the paper, we develop an exact algorithm, Maplex, for solving this problem in real world graphs practically. Based on the existing first-order and the novel second-order reduction rules, we design a powerful preprocessing method which efficiently removes redundant vertices and edges for Maplex. Also, the graph color heuristic is widely used for overestimating the maximum clique of a graph. For the first time, we generalize this technique for bounding the size of maximum k-plex in Maplex. Experiments are carried out to compare our algorithm with other state-of-the-art solvers on a wide range of publicly available graphs. Maplex outperforms all other algorithms on large real world graphs and is competitive with existing solvers on artificial dense graphs. Finally, we shed light on the effectiveness of each key component of Maplex.

Transformer-based architectures have shown great success in image captioning, where object regions are encoded and then attended into the vectorial representations to guide the caption decoding. However, such vectorial representations only contain region-level information without considering the global information reflecting the entire image, which fails to expand the capability of complex multi-modal reasoning in image captioning. In this paper, we introduce a Global Enhanced Transformer (termed GET) to enable the extraction of a more comprehensive global representation, and then adaptively guide the decoder to generate high-quality captions. In GET, a Global Enhanced Encoder is designed for the embedding of the global feature, and a Global Adaptive Decoder are designed for the guidance of the caption generation. The former models intra- and inter-layer global representation by taking advantage of the proposed Global Enhanced Attention and a layer-wise fusion module. The latter contains a Global Adaptive Controller that can adaptively fuse the global information into the decoder to guide the caption generation. Extensive experiments on MS COCO dataset demonstrate the superiority of our GET over many state-of-the-arts.

Integrating model-based machine learning methods into deep neural architectures allows one to leverage both the expressive power of deep neural nets and the ability of model-based methods to incorporate domain-specific knowledge. In particular, many works have employed the expectation maximization (EM) algorithm in the form of an unrolled layer-wise structure that is jointly trained with a backbone neural network. However, it is difficult to discriminatively train the backbone network by backpropagating through the EM iterations as they are prone to the vanishing gradient problem. To address this issue, we propose Highway Expectation Maximization Networks (HEMNet), which is comprised of unrolled iterations of the generalized EM (GEM) algorithm based on the Newton-Rahpson method. HEMNet features scaled skip connections, or highways, along the depths of the unrolled architecture, resulting in improved gradient flow during backpropagation while incurring negligible additional computation and memory costs compared to standard unrolled EM. Furthermore, HEMNet preserves the underlying EM procedure, thereby fully retaining the convergence properties of the original EM algorithm. We achieve significant improvement in performance on several semantic segmentation benchmarks and empirically show that HEMNet effectively alleviates gradient decay.

Generative moment matching networks (GMMN) present a theoretically sound approach to learning deep generative mod-els. However, such methods are typically limited by the high sample complexity, thereby impractical in generating complex data. In this paper, we present a new strategy to train GMMN with a low sample complexity while retaining the theoretical soundness. Our method introduces some auxiliary variables, whose values are provided by a pre-trained model such as an encoder network in practice. Conditioned on these variables, we partition the distribution into a set of conditional distributions, which can be effectively matched with a low sample complexity. We instantiate this strategy by presenting an amortized network called GMMN-DP with shared auxiliary variable information for the data generation task, as well as developing an efficient stochastic training algorithm.The experimental results show that GMMN-DP can generate complex samples on datasets such as CelebA and CIFAR-10, where the vanilla GMMN fails.

Supervised learning models have been increasingly used for making decisions about individuals in applications such as hiring, lending, and college admission. These models may inherit pre-existing biases from training datasets and discriminate against protected attributes (e.g., race or gender). In addition to unfairness, privacy concerns also arise when the use of models reveals sensitive personal information. Among various privacy notions, differential privacy has become popular in recent years. In this work, we study the possibility of using a differentially private exponential mechanism as a post-processing step to improve both fairness and privacy of supervised learning models. Unlike many existing works, we consider a scenario where a supervised model is used to select a limited number of applicants as the number of available positions is limited. This assumption is well-suited for various scenarios, such as job application and college admission. We use ``equal opportunity'' as the fairness notion and show that the exponential mechanisms can make the decision-making process perfectly fair. Moreover, the experiments on real-world datasets show that the exponential mechanism can improve both privacy and fairness, with a slight decrease in accuracy compared to the model without post-processing.

Ensemble-based Adversarial Training is a principled approach to achieve robustness against adversarial attacks. An important technicality of this approach is to control the transferability of adversarial examples between ensemble members. We propose in this work a simple, but effective strategy to collaborate among committee models of an ensemble model. This is achieved via the secure and insecure sets defined for each model member on a given sample, hence help us to quantify and regularize the transferability. Consequently, our proposed framework provides the flexibility to reduce the adversarial transferability as well as promote the diversity of ensemble members, which are two crucial factors for better robustness in our ensemble approach. We conduct extensive and comprehensive experiments to demonstrate that our proposed method outperforms the state-of-the-art ensemble baselines, at the same time can detect a wide range of adversarial examples with a near perfect accuracy.

Conflict-Based Search (CBS) is a powerful algorithmic framework for optimally solving classical multi-agent path finding (MAPF) problems, where time is discretized into the time steps. Continuous-time CBS (CCBS) is a recently proposed version of CBS that guarantees optimal solutions without the need to discretize time. However, the scalability of CCBS is limited because it does not include any known improvements of CBS. In this paper, we begin to close this gap and explore how to adapt successful CBS improvements, namely, prioritizing conflicts (PC), disjoint splitting (DS), and high-level heuristics, to the continuous time setting of CCBS. These adaptions are not trivial, and require careful handling of different types of constraints, applying a generalized version of the Safe interval path planning (SIPP) algorithm, and extending the notion of cardinal conflicts. We evaluate the effect of the suggested enhancements by running experiments both on general graphs and 2^k-neighborhood grids. CCBS with these improvements significantly outperforms vanilla CCBS, solving problems with almost twice as many agents in some cases and pushing the limits of multi-agent path finding in continuous-time domains.

Determining the plausibility of causal relations between clauses is a commonsense reasoning task that requires complex inference ability. The general approach to this task is to train a large pretrained language model on a specific dataset. However, the available training data for the task is often scarce, which leads to instability of model training or reliance on the shallow features of the dataset. This paper presents a number of techniques for making models more robust in the domain of causal reasoning. Firstly, we perform adversarial training by generating perturbed inputs through synonym substitution. Secondly, based on a linguistic theory of discourse connectives, we perform data augmentation using a discourse parser for detecting causally linked clauses in large text, and a generative language model for generating distractors. Both methods boost model performance on the Choice of Plausible Alternatives (COPA) dataset, as well as on a Balanced COPA dataset, which is a modified version of the original data that has been developed to avoid superficial cues, leading to a more challenging benchmark. We show a statistically significant improvement in performance and robustness on both datasets, even with only a small number of additionally generated data points.

Many widely-used causal discovery methods such as Greedy Equivalent Search (GES), although with asymptotic correctness guarantees, have been reported to produce sub-optimal solutions on finite data, or when the causal faithfulness condition is violated. The constraint-based procedure with Boolean satisfiability (SAT) solver, and the recently proposed Sparsest Permutation (SP) algorithm have shown superb performance, but currently they do not scale well. In this work, we demonstrate that optimal score-based exhaustive search is remarkably useful for causal discovery: it requires weaker conditions to guarantee asymptotic correctness, and outperforms well-known methods including PC, GES, GSP, and NOTEARS. In order to achieve scalability, we also develop an approximation algorithm for larger systems based on the A* method, which scales up to 60+ variables and obtains better results than existing greedy algorithms such as GES, MMHC, and GSP. Our results illustrate the risk of assuming the faithfulness assumption, the advantages of exhaustive search methods, and the limitations of greedy search methods, and shed light on the computational challenges and techniques in scaling up to larger networks and handling unfaithful data.

Convolutional neural networks (CNNs) have achieved state-of-the-art performance on various tasks in computer vision. However, recent studies demonstrate that these models are vulnerable to carefully crafted adversarial samples and suffer from a significant performance drop when predicting them. Many methods have been proposed to improve adversarial robustness (e.g., adversarial training and new loss functions to learn adversarially robust feature representations). Here we offer a unique insight into the predictive behavior of CNNs that they tend to misclassify adversarial samples into the most probable false classes. This inspires us to propose a new Probabilistically Compact (PC) loss with logit constraints which can be used as a drop-in replacement for cross-entropy (CE) loss to improve CNN's adversarial robustness. Specifically, PC loss enlarges the probability gaps between true class and false classes meanwhile the logit constraints prevent the gaps from being melted by a small perturbation. We extensively compare our method with the state-of-the-art using large scale datasets under both white-box and black-box attacks to demonstrate its effectiveness. The source codes are available at https://github.com/xinli0928/PC-LC.

This paper studies regret minimization with randomized value functions in reinforcement learning. In tabular finite-horizon Markov Decision Processes, we introduce a clipping variant of one classical Thompson Sampling (TS)-like algorithm, randomized least-squares value iteration (RLSVI). Our $tilde{mathrm{O}}(H^2Ssqrt{AT})$ high-probability worst-case regret bound improves the previous sharpest worst-case regret bounds for RLSVI and matches the existing state-of-the-art worst-case TS-based regret bounds.

Hyperparameter optimization (HO) is an important problem in machine learning which is normally formulated as a bilevel optimization problem. Gradient-based methods are dominant in bilevel optimization due to their high scalability to the number of hyperparameters, especially in a deep learning problem. However, traditional gradient-based bilevel optimization methods need intermediate steps to obtain the exact or approximate gradient of hyperparameters, namely hypergradient, for the upper-level objective, whose complexity is high especially for high dimensional datasets. Recently, a penalty method has been proposed to avoid the computation of the hypergradient, which speeds up the gradient-based BHO methods. However, the penalty method may result in a very large number of constraints, which greatly limits the efficiency of this method, especially for high dimensional data problems. To address this limitation, in this paper, we propose a doubly stochastic gradient descent algorithm (DSGPHO) to improve the efficiency of the penalty method. Importantly, we not only prove the proposed method can converge to the KKT condition of the original problem in a convex setting, but also provide the convergence rate of DSGPHO which is the first result in the references of gradient-based bilevel optimization as far as we know. We compare our method with three state-of-the-art gradient-based methods in three tasks, i.e., data denoising, few-shot learning, and training data poisoning, using several large-scale benchmark datasets. All the results demonstrate that our method outperforms or is comparable to the existing methods in terms of accuracy and efficiency.

Online solvers for partially observable Markov decision processes have difficulty scaling to problems with large action spaces. This paper proposes a method called PA-POMCPOW to sample a subset of the action space that provides varying mixtures of exploitation and exploration for inclusion in a search tree. The proposed method first evaluates the action space according to a score function that is a linear combination of expected reward and expected information gain. The actions with the highest score are then added to the search tree during tree expansion. Experiments show that PA-POMCPOW is able to outperform existing state-of-the-art solvers on problems with large discrete action spaces.