AAAI 2021 Papers

Point processes are a useful mathematical tool for describing events over time, and so there are many recent approaches for representing and learning them. One notable open question is how to precisely describe the flexibility of point process models and whether there exists a general model that can represent all point processes. Our work bridges this gap. Focusing on the widely used event intensity function representation of point processes, we provide a proof that a class of learnable functions can universally approximate any valid intensity function. The proof connects the well known Stone-Weierstrass Theorem for function approximation, the uniform density of non-negative continuous functions using a transfer functions, the formulation of the parameters of a piece-wise continuous functions as a dynamic system, and a recurrent neural network implementation for capturing the dynamics. Using these insights, we design and implement UNIPoint, a novel neural point process model, using recurrent neural networks to parameterise sums of basis function upon each event. Evaluations on synthetic and real world datasets show that this simpler representation performs better than Hawkes process variants and more complex neural network-based approaches. We expect this result will provide a practical basis for selecting and tuning models, as well as furthering theoretical work on representational complexity and learnability.

Commonsense AI has long been seen as a near impossible goal---until recently. Now, research interest has sharply increased with an influx of new benchmarks and models. We propose two new ways to evaluate commonsense models, emphasizing their generality on new tasks and building on diverse, recently introduced benchmarks. First, we propose a new multitask benchmark, Rainbow, to promote research on commonsense models that generalize well over multiple tasks and datasets. Second, we propose a novel evaluation, the cost equivalent curve, that sheds new insight on how the choice of source datasets, pretrained language models, and transfer learning methods impacts performance and data efficiency. We perform extensive experiments---over 200 experiments encompassing 4800 models---and report multiple valuable and sometimes surprising findings, e.g., that transfer almost always leads to better or equivalent performance if following a particular recipe, that QA-based commonsense datasets transfer well with each other, while commonsense knowledge graphs do not, and that perhaps counter-intuitively, larger models benefit more from transfer than smaller ones. Last but not least, we introduce a new universal commonsense reasoning model, UNICORN, that establishes new state-of-the-art performance across 8 popular commonsense benchmarks, aNLI (87.3%), CosmosQA (91.8%), HellaSWAG (93.9%), PIQA (90.1%), SocialIQa (83.2%), WinoGrande (86.6%), CycIC (94.0%) and CommonsenseQA (79.3%).

This paper presents our task-oriented dialog system UBAR which models task-oriented dialogs on a dialog session level. Specifically, UBAR is acquired by fine-tuning the large pre-trained unidirectional language model GPT-2 on the sequence of the entire dialog session which is composed of user utterance, belief state, database result, system act, and system response of every dialog turn. Additionally, UBAR is evaluated in a more realistic setting, where its dialog context has access to user utterances and all content it generated such as belief states, system acts, and system responses. Experimental results on the MultiWOZ datasets show that UBAR achieves state-of-the-art performances in multiple settings, improving the combined score of response generation, policy optimization, and end-to-end modeling by 4.7, 3.5, and 9.4 points respectively. Thorough analyses demonstrate that the session-level training sequence formulation and the generated dialog context are essential for UBAR to operate as a fully end-to-end task-oriented dialog system in real life. We also examine the transfer ability of UBAR to new domains with limited data and provide visualization and a case study to illustrate the advantages of UBAR in modeling on a dialog session level.

With the increasing popularity of graph-based learning, graph neural networks (GNNs) emerge as the essential tool for gaining insights from graphs. However, unlike the conventional CNNs that have been extensively explored and exhaustively tested, people are still worrying about the GNNs' robustness under the critical settings, such as financial services. The main reason is that existing GNNs usually serve as a black-box in predicting and do not provide the uncertainty on the predictions. On the other side, the recent advancement of Bayesian deep learning on CNNs has demonstrated its success of quantifying and explaining such uncertainties to fortify CNN models. Motivated by these observations, we propose UAG, the first systematic solution to defend adversarial attacks on GNNs through identifying and exploiting hierarchical uncertainties in GNNs. UAG develops a Bayesian uncertainty technique to explicitly capture uncertainties in GNNs and further employs an uncertainty-aware attention technique to defend adversarial attacks on GNNs. Intensive experiments show that our proposed defense approach outperforms the state-of-the-art solutions by a significant margin.

Learning user representation is a critical task for recommendation systems as it can encode user preference for personalized services. User representation is generally learned from behavior data, such as clicking interactions and review comments. However, for less popular domains, the behavior data is insufficient to learn precise user representations. To deal with this problem, a natural thought is to leverage content-rich domains to complement user representations. Inspired by the recent success of BERT in NLP, we propose a novel pre-training and fine-tuning based approach U-BERT. Different from typical BERT applications, U-BERT is customized for recommendation and utilizes different frameworks in pre-training and fine-tuning. In pre-training, U-BERT focuses on content-rich domains and introduces a user encoder and a review encoder to model users' behaviors. Two pre-training strategies are proposed to learn the general user representations; In fine-tuning, U-BERT focuses on the target content-insufficient domains. In addition to the user and review encoders inherited from the pre-training stage, U-BERT further introduces an item encoder to model item representations. Besides, a review co-matching layer is proposed to capture more semantic interactions between the reviews of the user and item. Finally, U-BERT combines user representations, item representations and review interaction information to improve recommendation performance. Experiments on six benchmark datasets from different domains demonstrate the state-of-the-art performance of U-BERT.

Knowledge graphs (KGs) are of great importance to many real world applications, but they generally suffer from incomplete information in the form of missing relations between entities. Knowledge graph completion (also known as relation prediction) is the task of inferring missing facts given existing ones. Most of the existing work is proposed by maximizing the likelihood of observed instance-level triples. Not much attention, however, is paid to the ontological information, such as type information of entities and relations. In this work, we propose a type-augmented relation prediction (TaRP) method, where we apply both the type information and instance-level information for the relation prediction. In particular, type information and instance-level information are encoded as prior probabilities and likelihoods of relations respectively, and are combined by following the Bayes' rule. Our proposed TaRP method achieves significantly better performance than state-of-the-art methods on four benchmark datasets: FB15K, FB15K-237, YAGO26K-906, and DB111K-174. In addition, we show that the TaRP achieves the significantly improved data efficiency. More importantly, the type information extracted from a specific dataset can generalize well to different datasets through the proposed TaRP model.

Recognition of human activities is an important task due to its far-reaching applications such as healthcare system, context-aware applications, and security monitoring. Recently, WiFi based human activity recognition (HAR) is becoming ubiquitous due to its non-invasiveness. Existing WiFi-based HAR methods regard WiFi signals as a temporal sequence of channel state information (CSI), and employ deep sequential models (e.g., RNN, LSTM) to automatically capture channel-over-time features. Although being remarkably effective, they suffer from two major drawbacks. Firstly, the granularity of a single temporal point is blindly elementary for representing meaningful CSI patterns. Secondly, the time-over-channel features are also important, and could be a natural data augmentation. To address the drawbacks, we propose a novel Two-stream Convolution Augmented Human Activity Transformer (THAT) model. Our model proposes to utilize a two-stream structure to capture both time-over-channel and channel-over-time features, and use the multi-scale convolution augmented transformer to capture range-based patterns. Extensive experiments on four real experiment datasets demonstrate that our model outperforms state-of-the-art models in terms of both effectiveness and efficiency.

We present a new approach for learning the structure of a treewidth-bounded Bayesian Network (BN). The key to our approach is applying an exact method (based on MaxSAT) locally, to improve the score of a heuristically computed BN. This approach allows us to scale the power of exact methods—so far only applicable to BNs with several dozens of random variables—to large BNs with several thousands of random variables. Our experiments show that our method improves the score of BNs provided by state-of-the-art heuristic methods, often significantly.

We study the cocktail party problem and propose a novel attention network called Tune-In, abbreviated for training under negative environments with interference. It firstly learns two separate spaces of speaker-knowledge and speech-stimuli based on a shared feature space, where a new block structure is designed as the building block for all spaces, and then cooperatively solves different tasks. Between the two spaces, information is cast towards each other via a novel cross- and dual-attention mechanism, mimicking the bottom-up and top-down processes of a human's cocktail party effect. It turns out that substantially discriminative and generalizable speaker representations can be learnt in severely interfered conditions via our self-supervised training. The experimental results verify this seeming paradox. The learnt speaker embedding has superior discriminative power than a standard speaker verification method; meanwhile, Tune-In achieves remarkably better speech separation performances in terms of SI-SNRi and SDRi consistently in all test modes, and especially at lower memory and computational consumption, than state-of-the-art benchmark systems.

In explainable artificial intelligence, there is increasing interest in understanding the behaviour of autonomous agents to build trust and validate performance. Modern agent architectures, such as those trained by deep reinforcement learning, are currently so lacking in interpretable structure as to effectively be black boxes, but insights may still be gained from an external, behaviourist perspective. Inspired by conceptual spaces theory, we suggest that a versatile first step towards general understanding is to discretise the state space into convex regions, jointly capturing similarities over the agent's action, value function and temporal dynamics within a dataset of observations. We create such a representation using a novel variant of the CART decision tree algorithm, and demonstrate how it facilitates practical understanding of black box agents through prediction, visualisation and rule-based explanation.

Most collaborative filtering (CF) models estimate missing ratings with an implicit assumption that the ratings are missing-at-random, which may cause the biased rating estimation and degraded performance since recent deep exploration shows that ratings may likely be missing-not-at-random (MNAR). To debias MNAR rating estimation, we introduce item observability and user selection to depict the generation of MNAR ratings and propose a tripartite CF (TCF) framework to jointly model the triple aspects of rating generation: item observability, user selection, and ratings, and to estimate the MNAR ratings. An item observability variable is introduced to a complete observability model to infer whether an item is observable to a user. TCF also conducts a complete rating model for rating generation and utilizes a user selection model dependent on the item observability and rating values to model user selection of the observable items. We further elaborately instantiate TCF as a Tripartite Probabilistic Matrix Factorization model (TPMF) by leveraging the probabilistic matrix factorization. Besides, TPMF introduces multifaceted dependency between user selection and ratings to model the influence of user selection on ratings. Extensive experiments on synthetic and real-world datasets show that modeling item observability and user selection effectively debias MNAR rating estimation, and TPMF outperforms the state-of-the-art methods in estimating the MNAR ratings.

Clustering ensemble generates a consensus clustering result by integrating multiple weak base clustering results. Although it often provides more robust results compared with single clustering methods, it still suffers from the robustness problem if it does not treat the unreliability of base results carefully. Conventional clustering ensemble methods often use all data for ensemble, while ignoring the noises or outliers on the data. Although some robust clustering ensemble methods are proposed, which extract the noises on the data, they still characterize the robustness in a single level, and thus they cannot comprehensively handle the complicated robustness problem. In this paper, to address this problem, we propose a novel Tri-level Robust Clustering Ensemble (TRCE) method by transforming the clustering ensemble problem to a multiple graph learning problem. Just as its name implies, the proposed method tackles robustness problem in three levels: base clustering level, graph level and instance level. By considering the robustness problem in a more comprehensive way, the proposed TRCE can achieve a more robust consensus clustering result. Experimental results on benchmark datasets also demonstrate it. Our method often outperforms other state-of-the-art clustering ensemble methods. Even compared with the robust ensemble methods, ours also performs better.

We initiate the study of trembling-hand perfection in sequential (i.e., extensive-form) games with correlation. We introduce the extensive-form perfect correlated equilibrium (EFPCE) as a refinement of the classical extensive-form correlated equilibrium (EFCE) that amends its weaknesses off the equilibrium path. This is achieved by accounting for the possibility that the players may make mistakes while following recommendations independently at each information set of the game. After providing an axiomatic definition of EFPCE, we show that one always exists since any perfect (Nash) equilibrium constitutes an EFPCE, and that it is a refinement of EFCE, as any EFPCE is also an EFCE. Then, we prove that, surprisingly, computing an EFPCE is not harder than finding an EFCE, since the problem can be solved in polynomial time for general n-player extensive-form games (also with chance). This is achieved by formulating the problem as that of finding a limit solution (as $epsilon rightarrow 0$) to a suitably defined trembling LP parametrized by $epsilon$, featuring exponentially-many variables and polynomially-many constraints. To this end, we show how a recently developed polynomial-time algorithm for trembling LPs can be adapted to deal with problems having an exponential number of variables. This calls for the solution of a sequence of (non-trembling) LPs with exponentially-many variables and polynomially-many constraints, which is possible in polynomial time by applying an ellipsoid against hope approach.

It is well-known that deciding consistency for normal answer set programs (ASP) is NP-complete, thus, as hard as the satisfaction problem for propositional logic (SAT). The exponential time hypothesis (ETH) implies that the best algorithms to solve these problems take exponential time in the worst case. However, accounting for the treewidth, the consistency problem for ASP is slightly harder than SAT: while SAT can be solved by an algorithm that runs in exponential time in the treewidth k, ASP requires exponential time in k · log(k). This extra cost is due to checking that there are no self-supported true atoms due to positive cycles in the program. In this paper, we refine this recent result and show that consistency for ASP can be decided in exponential time in k · log(ι) where ι is a novel measure, bounded by both treewidth k and the size of the largest strongly-connected component of the positive dependency graph of the program. We provide a treewidth-aware reduction from ASP to SAT that adheres to the above limit.

Recently program learning techniques have been proposed to process source code based on syntactical structures (e.g., abstract syntax trees) and/or semantic information (e.g., dependency graphs). While graphs may be better than trees at capturing code semantics, constructing the graphs from code inputs through the semantic analysis of multiple viewpoints can lead to inaccurate noises for a specific software engineering task. Compared to graphs, syntax trees are more precisely defined on the grammar and easier to parse; unfortunately, previous tree-based learning techniques have not been able to learn semantic information from trees to achieve better accuracy than graph-based techniques. We have proposed a new learning technique, named TreeCaps, by fusing together capsule networks with tree-based convolutional neural networks to achieve a learning accuracy higher than some existing graph-based techniques while it is based only on trees. TreeCaps introduces novel variable-to-static routing algorithms into the capsule networks to compensate for the loss of previous routing algorithms. Aside from accuracy, we also find that TreeCaps is the most robust to withstand those semantic-preserving program transformations that change code syntax without modifying the semantics. Evaluated on a large number of Java and C/C++ programs, TreeCaps models outperform prior deep learning models of program source code, in terms of both accuracy and robustness for program comprehension tasks such as code functionality classification and function name prediction. Our implementation is publicly available at: https://github.com/bdqnghi/treecaps.

Much research has been devoted to the problem of estimating treatment effects from observational data; however, most methods assume that the observed variables only contain confounders, i.e., variables that affect both the treatment and the outcome. Unfortunately, this assumption is frequently violated in real-world applications, since some variables only affect the treatment but not the outcome, and vice versa. Moreover, in many cases only the proxy variables of the underlying confounding factors can be observed. In this work, we first show the importance of differentiating confounding factors from instrumental and risk factors for both average and conditional average treatment effect estimation, and then we propose a variational inference approach to simultaneously infer latent factors from the observed variables, disentangle the factors into three disjoint sets corresponding to the instrumental, confounding, and risk factors, and use the disentangled factors for treatment effect estimation. Experimental results demonstrate the effectiveness of the proposed method on a wide range of synthetic, benchmark, and real-world datasets.

With the progression of Generative Adversarial Networks (GANs), image translation methods has achieved increasingly remarkable performance. However, most available methods can only achieve image level translation, which is unable to precisely control the regions to be translated. In this paper, we propose a novel self-adaptive region translation network (SART) for region-level translation, which uses region-adaptive instance normalization (RIN) and a region matching loss (RML) for this task. We first encode the style and content image for each region with style and content encoder. To translate both shape and texture of the target region, we inject region-adaptive style features into the decoder by RIN. To ensure independent translation among different regions, RML is proposed to measure the similarity between the non-translated/translated regions of content and translated images. Extensive experiments on three publicly available datasets, i.e. Morph, RaFD and CelebAMask-HQ, suggest that our approach demonstrate obvious improvement over state-of-the-art methods like StarGAN, SEAN and FUNIT. Our approach has further advantages in precise control of the regions to be translated. As a result, region level expression changes and step-by-step make-up can be achieved. The video demo is available at (https://youtu.be/DvIdmcR2LEc).

Network modeling aims to learn the latent representations of nodes such that the representations preserve both network structures and node attribute information. This problem is fundamental due to its prevalence in numerous domains. However, existing approaches either target the static networks or struggle to capture the complicated temporal dependency, while most real-world networks evolve over time and the success of network modeling hinges on the understanding of how entities are temporally connected. In this paper, we present TRRN, a transformer-style relational reasoning network with dynamic memory updating, to deal with the above challenges. TRRN employs multi-head self-attention to reason over a set of memories, which provides a multitude of shortcut paths for information to flow from past observations to the current latent representations. By utilizing the policy networks augmented with differentiable binary routers, TRRN estimates the possibility of each memory being activated and dynamically updates the memories at the time steps when they are most relevant. We evaluate TRRN with the tasks of node classification and link prediction on four real temporal network datasets. Experimental results demonstrate the consistent performance gains for TRRN over the leading competitors.

Safety validation is important during the development of safety-critical autonomous systems but can require significant computational effort. Existing algorithms often start from scratch each time the system under test changes. We apply transfer learning to improve the efficiency of reinforcement learning based safety validation algorithms when applied to related systems. Knowledge from previous safety validation tasks is encoded through the action value function and transferred to future tasks with a learned set of attention weights. Including a learned state and action value transformation for each source task can improve performance even when systems have substantially different failure modes. We conduct experiments on safety validation tasks in gridworld and autonomous driving scenarios. We show that transfer learning can improve the initial and final performance of validation algorithms and reduce the number of training steps.