AAAI 2021 论文列表

Participatory budgeting is a method used by city governments to select public projects to fund based on residents' votes. Many cities use participatory budgeting at a district level. Typically, a budget is divided among districts proportionally to their population, and each district holds an election over local projects and then uses its budget to fund the projects most preferred by its voters. However, district-level participatory budgeting can yield poor social welfare because it does not necessarily fund projects supported across multiple districts. On the other hand, decision making that only takes global social welfare into account can be unfair to districts: A social-welfare-maximizing solution might not fund any of the projects preferred by a district, despite the fact that its constituents pay taxes to the city. Thus, we study how to fairly maximize social welfare in a participatory budgeting setting with a single city-wide election. We propose a notion of fairness that guarantees each district at least as much welfare as it would have received in a district-level election. We show that, although optimizing social welfare subject to this notion of fairness is NP-hard, we can efficiently construct a lottery over welfare-optimal outcomes that is fair in expectation. Moreover, we show that, when we are allowed to slightly relax fairness, we can efficiently compute a fair solution that is welfare-maximizing, but which may overspend the budget.

We consider the problem of learning a set of probability distributions from the empirical Bellman dynamics in distributional reinforcement learning (RL), a class of state-of-the-art methods that estimate the distribution, as opposed to only the expectation, of the total return. We formulate a method that learns a finite set of statistics from each return distribution via neural networks, as in the distributional RL literature. Existing distributional RL methods however constrain the learned statistics to predefined functional forms of the return distribution which is both restrictive in representation and difficult in maintaining the predefined statistics. Instead, we learn unrestricted statistics, i.e., deterministic (pseudo-)samples, of the return distribution by leveraging a technique from hypothesis testing known as maximum mean discrepancy (MMD), which leads to a simpler objective amenable to backpropagation. Our method can be interpreted as implicitly matching all orders of moments between a return distribution and its Bellman target. We establish sufficient conditions for the contraction of the distributional Bellman operator and provide finite-sample analysis for the deterministic samples in distribution approximation. Experiments on the suite of Atari games show that our method outperforms the standard distributional RL baselines and sets a new record in the Atari games for non-distributed agents.

The task of rationalization aims to extract pieces of input text as rationales to justify neural network predictions on text classification tasks. By definition, rationales represent key text pieces used for prediction and thus should have similar classification feature distribution compared to the original input text. However, previous methods mainly focused on maximizing the mutual information between rationales and labels while neglecting the relationship between rationales and input text. To address this issue, we propose a novel rationalization method that matches the distributions of rationales and input text in both the feature space and output space. Empirically, the proposed distribution matching approach consistently outperforms previous methods by a large margin. Our data and code are available.

Researches have demonstrated that low bit-width (e.g., INT8) quantization can be employed to accelerate the inference process. It makes the gradient quantization very promising since the backward propagation requires approximately twice more computation than forward one. Due to the variability and uncertainty of gradient distribution, a lot of methods have been proposed to attain training stability. However, most of them ignore the channel-wise gradient distributions and the impact of gradients with different magnitudes, resulting in the degradation of final accuracy. In this paper, we propose a novel INT8 quantization training framework for convolutional neural network to address the above issues. Specifically, we adopt Gradient Vectorized Quantization to quantize the gradient, based on the observation that layer-wise gradients contain multiple distributions along the channel dimension. Then, Magnitude-aware Clipping Strategy is introduced by taking the magnitudes of gradients into consideration when minimizing the quantization error, and we present a theoretical derivation to solve the quantization parameters of different distributions. Experimental results on broad range of computer vision tasks, such as image classification, object detection and video classification, demonstrate that the proposed Distribution Adaptive INT8 Quantization training method has achieved almost lossless training accuracy for different backbones, including ResNet, MobileNetV2, InceptionV3, VGG and AlexNet, which is superior to the state-of-the-art techniques. Moreover, we further implement the INT8 kernel that can accelerate the training iteration more than 200% under the latest Turing architecture, i.e., our method excels on both training accuracy and speed.

Learning theory of distributed algorithms has recently attracted enormous attention in the machine learning community. However, most of existing works focus on learning problem with pointwise loss and does not consider the communication among local processors. In this paper, we propose a new distributed pairwise ranking with communication (called DLSRank-C) based on the Newton-Raphson iteration, and establish its learning rate analysis in probability. Theoretical and empirical assessments demonstrate the effectiveness of DLSRank-C under mild conditions.

Recent progress in contrastive learning has revolutionized unsupervised representation learning. Concretely, multiple views (augmentations) from the same image are encouraged to map to close embeddings, while views from different images are pulled apart.In this paper, through visualizing and diagnosing classification errors, we observe that current contrastive models are ineffective at localizing the foreground object, limiting their ability to extract discriminative high-level features. This is due to the fact that view generation process considers pixels in an image uniformly.To address this problem, we propose a data-driven approach for learning invariance to backgrounds. It first estimates foreground saliency in images and then creates augmentations by copy-and-pasting the foreground onto a variety of back-grounds. The learning still follows an instance discrimination approach, so that the representation is trained to disregard background content and focus on the foreground. We study a variety of saliency estimation methods, and find that most methods lead to improvements for contrastive learning. With this approach, significant performance is achieved for self-supervised learning on ImageNet classification, and also for object detection on PASCAL VOC and MSCOCO.

Transfer learning, which is to improve the learning performance in the target domain by leveraging useful knowledge from the source domain, often requires that those two domains are very close, which limits its application scope. Recently, distant transfer learning has been studied to transfer knowledge between two distant or even totally unrelated domains via unlabeled auxiliary domains that act as a bridge in the spirit of human transitive inference that two completely unrelated concepts can be connected through gradual knowledge transfer. In this paper, we study distant transfer learning by proposing a DeEp Random Walk basEd distaNt Transfer (DERWENT) method. Different from existing distant transfer learning models that implicitly identify the path of knowledge transfer between the source and target instances through auxiliary instances, the proposed DERWENT model can explicitly learn such paths via the deep random walk technique. Specifically, based on sequences identified by the random walk technique on a data graph where source and target data have no direct connection, the proposed DERWENT model enforces adjacent data points in a sequence to be similar, makes the ending data point be represented by other data points in the same sequence, and considers weighted classification losses of source data. Empirical studies on several benchmark datasets demonstrate that the proposed DERWENT algorithm yields the state-of-the-art performance.

We study the classic stochastic linear bandit problem under the restriction that each arm may be selected for limited number of times. This simple constraint, which we call disposability, captures a common restriction that occurs in recommendation problems from a diverse array of applications ranging from personalized styling services to dating platforms. We show that the regret for this problem is characterized by a previously-unstudied function of the reward distribution among optimal arms. Algorithmically, our upper bound relies on an optimism-based policy which, while computationally intractable, lends itself to approximation via a fast alternating heuristic initialized with a classic similarity score. Experiments show that our policy dominates a set of benchmarks which includes algorithms known to be optimal for the linear bandit without disposability, along with natural modifications to these algorithms for the disposable setting.

The disjunctive temporal problem (DTP) is an expressive temporal formalism that extends Dechter et al.'s simple temporal problem. The DTP is well studied in the literature and has many important applications. It is known that deciding satisfiability of DTPs is NP-hard and that, in many cases, single-exponential algorithms (running in O(c^n) time) do not exist under the Exponential-Time Hypothesis. The computational hardness makes it worthwhile to identify restricted problems that are efficiently solvable. One way of doing this is to restrict the interactions of variables and constraints. We show that instances of DTP of any arity with integers bounded by poly(n) can be solved in n^{f(w)} time, where n denotes the problem size, w is the treewidth of the incidence graph and f is a computable function; in other words, this problem is in the complexity class XP and it can be solved in polynomial time whenever w is fixed. We complement this result by showing that binary DTPs that only involve the integers 0 and 1 are not fixed-parameter tractable with respect to treewidth, i.e. they do not admit a f(w)poly(n)$ time algorithm for any computable function f, under standard complexity assumptions. For instances with unbounded integers, we show that even binary DTPs parameterized by treewidth cannot be in XP, unless P = NP.

In the fight against the COVID-19 pandemic, many social activities have moved online; society's overwhelming reliance on the complex cyberspace makes its security more important than ever. In this paper, we propose and develop an intelligent system named Dr.HIN to protect users against the evolving Android malware attacks in the COVID-19 era and beyond. In Dr.HIN, besides app content, we propose to consider higher-level semantics and social relations among apps, developers and mobile devices to comprehensively depict Android apps; and then we introduce a structured heterogeneous information network (HIN) to model the complex relations and exploit meta-path guided strategy to learn node (i.e., app) representations from HIN. As the representations of malware could be highly entangled with benign apps in the complex ecosystem of development, it poses a new challenge of learning the latent explanatory factors hidden in the HIN embeddings to detect the evolving malware. To address this challenge, we propose to integrate domain priors generated from different views (i.e., app content, app authorship, app installation) to devise an adversarial disentangler to separate the distinct, informative factors of variations hidden in the HIN embeddings for large-scale Android malware detection. This is the first attempt of disentangled representation learning in HIN data. Promising experimental results based on the large-scale and real sample collections from security industry demonstrate the performance of Dr.HIN in evolving Android malware detection, by comparison with baselines and popular mobile security products.

Pedestrian trajectory prediction is one of the important tasks required for autonomous navigation and social robots in human environments. Previous studies focused on estimating social forces among individual pedestrians. However, they did not consider the social forces of groups on pedestrians, which results in over-collision avoidance problems. To address this problem, we present a Disentangled Multi-Relational Graph Convolutional Network (DMRGCN) for socially entangled pedestrian trajectory prediction. We first introduce a novel disentangled multi-scale aggregation to better represent social interactions, among pedestrians on a weighted graph. For the aggregation, we construct the multi-relational weighted graphs based on distances and relative displacements among pedestrians. In the prediction step, we propose a global temporal aggregation to alleviate accumulated errors for pedestrians changing their directions. Finally, we apply DropEdge into our DMRGCN to avoid the over-fitting issue on relatively small pedestrian trajectory datasets. Through the effective incorporation of the three parts within an end-to-end framework, DMRGCN achieves state-of-the-art performances on a variety of challenging trajectory prediction benchmarks.

In this paper, we propose a novel graph learning framework for phrase grounding in the image. Developing from the sequential to the dense graph model, existing works capture coarse-grained context but fail to distinguish the diversity of context among phrases and image regions. In contrast, we pay special attention to different motifs implied in the context of the scene graph and devise the disentangled graph network to integrate the motif-aware contextual information into representations. Besides, we adopt interventional strategies at the feature and the structure levels to consolidate and generalize representations. Finally, the cross-modal attention network is utilized to fuse intra-modal features, where each phrase can be computed similarity with regions to select the best-grounded one. We validate the efficiency of disentangled and interventional graph network (DIGN) through a series of ablation studies, and our model achieves state-of-the-art performance on Flickr30K Entities and ReferIt Game benchmarks.

The information bottleneck (IB) method is a technique for extracting information that is relevant for predicting the target random variable from the source random variable, which is typically implemented by optimizing the IB Lagrangian that balances the compression and prediction terms. However, the IB Lagrangian is hard to optimize, and multiple trials for tuning values of Lagrangian multiplier are required. Moreover, we show that the prediction performance strictly decreases as the compression gets stronger during optimizing the IB Lagrangian. In this paper, we implement the IB method from the perspective of supervised disentangling. Specifically, we introduce Disentangled Information Bottleneck (DisenIB) that is consistent on compressing source maximally without target prediction performance loss (maximum compression). Theoretical and experimental results demonstrate that our method is consistent on maximum compression, and performs well in terms of generalization, robustness to adversarial attack, out-of-distribution detection, and supervised disentangling.

Weakly-supervised semantic segmentation (WSSS) using image-level labels has recently attracted much attention for reducing annotation costs. Existing WSSS methods utilize localization maps from the classification network to generate pseudo segmentation labels. However, since localization maps obtained from the classifier focus only on sparse discriminative object regions, it is difficult to generate high-quality segmentation labels. To address this issue, we introduce discriminative region suppression (DRS) module that is a simple yet effective method to expand object activation regions. DRS suppresses the attention on discriminative regions and spreads it to adjacent non-discriminative regions, generating dense localization maps. DRS requires few or no additional parameters and can be plugged into any network. Furthermore, we introduce an additional learning strategy to give a self-enhancement of localization maps, named localization map refinement learning. Benefiting from this refinement learning, localization maps are refined and enhanced by recovering some missing parts or removing noise itself. Due to its simplicity and effectiveness, our approach achieves mIoU 71.4% on the PASCAL VOC 2012 segmentation benchmark using only image-level labels. Extensive experiments demonstrate the effectiveness of our approach.

Discovering new intents is a crucial task in dialogue systems. Most existing methods are limited in transferring the prior knowledge from known intents to new intents. These methods also have difficulties in providing high-quality supervised signals to learn clustering-friendly features for grouping unlabeled intents. In this work, we propose an effective method (Deep Aligned Clustering) to discover new intents with the aid of limited known intent data. Firstly, we leverage a few labeled known intent samples as prior knowledge to pre-train the model. Then, we perform k-means to produce cluster assignments as pseudo-labels. Moreover, we propose an alignment strategy to tackle the label inconsistency problem during clustering assignments. Finally, we learn the intent representations under the supervision of the aligned pseudo-labels. With an unknown number of new intents, we predict the number of intent categories by eliminating low-confidence intent-wise clusters. Extensive experiments on two benchmark datasets show that our method is more robust and achieves substantial improvements over the state-of-the-art methods.

We study the problem of inferring causal graphs from observational data. We are particularly interested in discovering graphs where all edges are oriented, as opposed to the partially directed graph that the state of the art discover. To this end, we base our approach on the algorithmic Markov condition. Unlike the statistical Markov condition, it uniquely identifies the true causal network as the one that provides the simplest— as measured in Kolmogorov complexity—factorization of the joint distribution. Although Kolmogorov complexity is not computable, we can approximate it from above via the Minimum Description Length principle, which allows us to define a consistent and computable score based on non-parametric multivariate regression. To efficiently discover causal networks in practice, we introduce the GLOBE algorithm, which greedily adds, removes, and orients edges such that it minimizes the overall cost. Through an extensive set of experiments, we show GLOBE performs very well in practice, beating the state of the art by a margin.

Machine Translation Quality Estimation (QE) is a task of predicting the quality of machine translations without relying on any reference. Recently, the predictor-estimator framework trains the predictor as a feature extractor, which leverages the extra parallel corpora without QE labels, achieving promising QE performance. However, we argue that there are gaps between the predictor and the estimator in both data quality and training objectives, which preclude QE models from benefiting from a large number of parallel corpora more directly. We propose a novel framework called DirectQE that provides a direct pretraining for QE tasks. In DirectQE, a generator is trained to produce pseudo data that is closer to the real QE data, and a detector is pretrained on these data with novel objectives that are akin to the QE task. Experiments on widely used benchmarks show that DirectQE outperforms existing methods, without using any pretraining models such as BERT. We also give extensive analyses showing how fixing the two gaps contributes to our improvements.

To infer the structure of a diffusion network from observed diffusion results, existing approaches customarily assume that observed data are complete and contain the final infection status of each node, as well as precise timestamps of node infections. Due to high cost and uncertainties in the monitoring of node infections, exact timestamps are often unavailable in practice, and even the final infection statuses of nodes are sometimes missing. In this work, we study how to carry out diffusion network inference without infection timestamps, using only partial observations of the final infection statuses of nodes. To this end, we iteratively infer the structure of the target diffusion network with observed data and imputed values for missing data, and learn the most likely infection transmission probabilities between nodes w.r.t. current inferred structure, which then help us update the imputation of missing data in turn. Extensive experimental results on both synthetic and real-world networks show that our approach can properly handle missing data and accurately uncover diffusion network structures.

We introduce a new (ϵₚ, δₚ)-differentially private algorithm for the k-means clustering problem. Given a dataset in Euclidean space, the k-means clustering problem requires one to find k points in that space such that the sum of squares of Euclidean distances between each data point and its closest respective point among the k returned is minimised. Although there exist privacy-preserving methods with good theoretical guarantees to solve this problem, in practice it is seen that it is the additive error which dictates the practical performance of these methods. By reducing the problem to a sequence of instances of maximum coverage on a grid, we are able to derive a new method that achieves lower additive error than previous works. For input datasets with cardinality n and diameter Δ, our algorithm has an O(Δ² (k log² n log(1/δₚ)/ϵₚ + k √(d log(1/δₚ))/ϵₚ)) additive error whilst maintaining constant multiplicative error. We conclude with some experiments and find an improvement over previously implemented work for this problem.