AAAI 2021 Papers

We study the problem of obtaining accurate policy gradient estimates using a finite number of samples. Monte-Carlo methods have been the default choice for policy gradient estimation, despite suffering from high variance in the gradient estimates. On the other hand, more sample efficient alternatives like Bayesian quadrature methods have received little attention due to their high computational complexity. In this work, we propose deep Bayesian quadrature policy gradient (DBQPG), a computationally efficient high-dimensional generalization of Bayesian quadrature, for policy gradient estimation. We show that DBQPG can substitute Monte-Carlo estimation in policy gradient methods, and demonstrate its effectiveness on a set of continuous control benchmarks. In comparison to Monte-Carlo estimation, DBQPG provides (i) more accurate gradient estimates with a significantly lower variance, (ii) a consistent improvement in the sample complexity and average return for several deep policy gradient algorithms, and, (iii) the uncertainty in gradient estimation that can be incorporated to further improve the performance.

Without prohibitive and laborious 3D annotations, weakly-supervised 3D human pose methods mainly employ the model regularization with geometric projection consistency or geometry estimation from multi-view images. Nevertheless, those approaches explicitly need known parameters of calibrated cameras, exhibiting a limited model generalization in various realistic scenarios. To mitigate this issue, in this paper, we propose a Deductive Weakly-Supervised Learning (DWSL) for 3D human pose machine. Our DWSL firstly learns latent representations on depth and camera pose for 3D pose reconstruction. Since weak supervision usually causes ill-conditioned learning or inferior estimation, our DWSL introduces deductive reasoning to make an inference for the human pose from a view to another and develops a reconstruction loss to demonstrate what the model learns and infers is reliable. This learning by deduction strategy employs the view-transform demonstration and structural rules derived from depth, geometry and angle constraints, which improves the reliability of the model training with weak supervision. On three 3D human pose benchmarks, we conduct extensive experiments to evaluate our proposed method, which achieves superior performance in comparison with state-of-the-art weak-supervised methods. Particularly, our model shows an appealing potential for learning from 2D data captured in dynamic outdoor scenes, which demonstrates promising robustness and generalization in realistic scenarios. Our code is publicly available at https://github.com/Xipeng-Chen/DWSL-3D-pose.

The goal of person search is to localize and match query persons from scene images. For high efficiency, one-step methods have been developed to jointly handle the pedestrian detection and identification sub-tasks using a single network. There are two major challenges in the current one-step approaches. One is the mutual interference between the optimization objectives of multiple sub-tasks. The other is the sub-optimal identification feature learning caused by small batch size when end-to-end training. To overcome these problems, we propose a decoupled and memory-reinforced network (DMRNet). Specifically, to reconcile the conflicts of multiple objectives, we simplify the standard tightly coupled pipelines and establish a deeply decoupled multi-task learning framework. Further, we build a memory-reinforced mechanism to boost the identification feature learning. By queuing the identification features of recently accessed instances into a memory bank, the mechanism augments the similarity pair construction for pairwise metric learning. For better encoding consistency of the stored features, a slow-moving average of the network is applied for extracting these features. In this way, the dual networks reinforce each other and converge to robust solution states. Experimentally, the proposed method obtains 93.2% and 46.9% mAP on CUHK-SYSU and PRW datasets, which exceeds all the existing one-step methods.

Recurrent Neural Networks (RNNs) have demonstrated their effectiveness in learning and processing sequential data (e.g., speech and natural language). However, due to the black-box nature of neural networks, understanding the decision logic of RNNs is quite challenging. Some recent progress has been made to approximate the behavior of an RNN by weighted automata. They provide better interpretability, but still suffer from poor scalability. In this paper, we propose a novel approach to extracting weighted automata with the guidance of a target RNN's decision and context information. In particular, we identify the patterns of RNN's step-wise predictive decisions to instruct the formation of automata states. Further, we propose a state composition method to enhance the context-awareness of the extracted model. Our in-depth evaluations on typical RNN tasks, including language model and classification, demonstrate the effectiveness and advantage of our method over the state-of-the-arts. The evaluation results show that our method can achieve accurate approximation of an RNN even on large-scale tasks.

This paper deals with distributed reinforcement learning problems with safety constraints. In particular, we consider that a team of agents cooperate in a shared environment, where each agent has its individual reward function and safety constraints that involve all agents' joint actions. As such, the agents aim to maximize the team-average long-term return, subject to all the safety constraints. More intriguingly, no central controller is assumed to coordinate the agents, and both the rewards and constraints are only known to each agent locally/privately. Instead, the agents are connected by a peer-to-peer communication network to share information with their neighbors. In this work, we first formulate this problem as a distributed constrained Markov decision process (D-CMDP) with networked agents. Then, we propose a decentralized policy gradient (PG) method, Safe Dec-PG, to perform policy optimization based on this D-CMDP model over a network. Convergence guarantees, together with numerical results, showcase the superiority of the proposed algorithm. To the best of our knowledge, this is the first decentralized PG algorithm that accounts for the coupled safety constraints with a quantifiable convergence rate in multi-agent reinforcement learning. Finally, we emphasize that our algorithm is also novel in solving a class of decentralized stochastic nonconvex-concave minimax optimization problems, where both the algorithm design and corresponding theoretical analysis are of independent interest.

We study decentralized stochastic linear bandits, where a network of N agents acts cooperatively to efficiently solve a linear bandit-optimization problem over a d-dimensional space. For this problem, we propose DLUCB: a fully decentralized algorithm that minimizes the cumulative regret over the entire network. At each round of the algorithm each agent chooses its actions following an upper confidence bound (UCB) strategy and agents share information with their immediate neighbors through a carefully designed consensus procedure that repeats over cycles. Our analysis adjusts the duration of these communication cycles ensuring near-optimal regret performance O(d log{NT}sqrt{NT}) at a communication rate of O(dN^2) per round. The structure of the network affects the regret performance via a small additive term – coined the regret of delay – that depends on the spectral gap of the underlying graph. Notably, our results apply to arbitrary network topologies without a requirement for a dedicated agent acting as a server. In consideration of situations with high communication cost, we propose RC-DLUCB: a modification of DLUCB with rare communication among agents. The new algorithm trades off regret performance for a significantly reduced total communication cost of O(d^3N^5/2) over all T rounds. Finally, we show that our ideas extend naturally to the emerging, albeit more challenging, setting of safe bandits. For the recently studied problem of linear bandits with unknown linear safety constraints, we propose the first safe decentralized algorithm. Our study contributes towards applying bandit techniques in safety-critical distributed systems that repeatedly deal with unknown stochastic environments. We present numerical simulations for various network topologies that corroborate our theoretical findings.

Deep learning has been successful for many computer vision tasks due to the availability of shared and centralised large-scale training data. However, increasing awareness of privacy concerns poses new challenges to deep learning, especially for human subject related recognition such as person re-identification (Re-ID). In this work, we solve the Re-ID problem by decentralised learning from non-shared private training data distributed at multiple user sites of independent multi-domain label spaces. We propose a novel paradigm called Federated Person Re-Identification (FedReID) to construct a generalisable global model (a central server) by simultaneously learning with multiple privacy-preserved local models (local clients). Specifically, each local client receives global model updates from the server and trains a local model using its local data independent from all the other clients. Then, the central server aggregates transferrable local model updates to construct a generalisable global feature embedding model without accessing local data so to preserve local privacy. This client-server collaborative learning process is iteratively performed under privacy control, enabling FedReID to realise decentralised learning without sharing distributed data nor collecting any centralised data. Extensive experiments on ten Re-ID benchmarks show that FedReID achieves compelling generalisation performance beyond any locally trained models without using shared training data, whilst inherently protects the privacy of each local client. This is uniquely advantageous over contemporary Re-ID methods.

While deep learning demonstrates its strong ability to handle independent and identically distributed (IID) data, it often suffers from out-of-distribution (OoD) generalization, where the test data come from another distribution (w.r.t. the training one). Designing a general OoD generalization framework for a wide range of applications is challenging, mainly due to different kinds of distribution shifts in the real world, such as the shift across domains or the extrapolation of correlation. Most of the previous approaches can only solve one specific distribution shift, leading to unsatisfactory performance when applied to various OoD benchmarks. In this work, we propose DecAug, a novel decomposed feature representation and semantic augmentation approach for OoD generalization. Specifically, DecAug disentangles the category-related and context-related features by orthogonalizing the two gradients (w.r.t. intermediate features) of losses for predicting category and context labels, where category-related features contain causal information of the target object, while context-related features cause distribution shifts between training and test data. Furthermore, we perform gradient-based augmentation on context-related features to improve the robustness of learned representations. Experimental results show that DecAug outperforms other state-of-the-art methods on various OoD datasets, which is among the very few methods that can deal with different types of OoD generalization challenges.

Human-object interaction (HOI) detection requires a large amount of annotated data. Current algorithms suffer from insufficient training samples and category imbalance within datasets. To increase data efficiency, in this paper, we propose an efficient and effective data augmentation method called DecAug for HOI detection. Based on our proposed object state similarity metric, object patterns across different HOIs are shared to augment local object appearance features without changing their states. Further, we shift spatial correlation between humans and objects to other feasible configurations with the aid of a pose-guided Gaussian Mixture Model while preserving their interactions. Experiments show that our method brings up to 3.3 mAP and 1.6 mAP improvements on V-COCO and HICO-DET dataset for two advanced models. Specifically, interactions with fewer samples enjoy more notable improvement. Our method can be easily integrated into various HOI detection models with negligible extra computational consumption.

Multi-agent coordination tends to benefit from efficient communication, where cooperation often happens based on exchanging information about what the agents intend to do, i.e. intention sharing. It becomes a key problem to model the intention by some proper abstraction. Currently, it is either too coarse such as final goals or too fined as primitive steps, which is inefficient due to the lack of modularity and semantics. In this paper, we design a novel multi-agent coordination protocol based on subgoal intentions, defined as the probability distribution over feasible subgoal sequences. The subgoal intentions encode macro-action behaviors with modularity so as to facilitate joint decision making at higher abstraction. Built over the proposed protocol, we present Dec-SGTS (Decentralized Sub-Goal Tree Search) to solve decentralized online multi-agent planning hierarchically and efficiently. Each agent runs Dec-SGTS asynchronously by iteratively performing three phases including local sub-goal tree search, local subgoal intention update and global subgoal intention sharing. We conduct the experiments on courier dispatching problem, and the results show that Dec-SGTS achieves much better reward while enjoying a significant reduction of planning time and communication cost compared with Dec-MCTS (Decentralized Monte Carlo Tree Search).

It is common to evaluate a set of items by soliciting people to rate them. For example, universities ask students to rate the teaching quality of their instructors, and conference organizers ask authors of submissions to evaluate the quality of the reviews. However, in these applications, students often give a higher rating to a course if they receive higher grades in a course, and authors often give a higher rating to the reviews if their papers are accepted to the conference. In this work, we call these external factors the "outcome" experienced by people, and consider the problem of mitigating these outcome-induced biases in the given ratings when some information about the outcome is available. We formulate the information about the outcome as a known partial ordering on the bias. We propose a debiasing method by solving a regularized optimization problem under this ordering constraint, and also provide a carefully designed cross-validation method that adaptively chooses the appropriate amount of regularization. We provide theoretical guarantees on the performance of our algorithm, as well as experimental evaluations.

The threat of data-poisoning backdoor attacks on learning algorithms typically comes from the labeled data. However, in deep semi-supervised learning (SSL), unknown threats mainly stem from the unlabeled data. In this paper, we propose a novel deep hidden backdoor (DeHiB) attack scheme for SSL-based systems. In contrast to the conventional attacking methods, the DeHiB can inject malicious unlabeled training data to the semi-supervised learner so as to enable the SSL model to output premeditated results. In particular, a robust adversarial perturbation generator regularized by a unified objective function is proposed to generate poisoned data. To alleviate the negative impact of the trigger patterns on model accuracy and improve the attack success rate, a novel contrastive data poisoning strategy is designed. Using the proposed data poisoning scheme, one can implant the backdoor into the SSL model using the raw data without hand-crafted labels. Extensive experiments based on CIFAR10 and CIFAR100 datasets demonstrated the effectiveness and crypticity of the proposed scheme.

The design of online algorithms has tended to focus on algorithms with worst-case guarantees, e.g., bounds on the competitive ratio. However, it is well-known that such algorithms are often overly pessimistic, performing sub-optimally on non-worst-case inputs. In this paper, we develop an approach for data-driven design of online algorithms that maintain near-optimal worst-case guarantees while also performing learning in order to perform well for typical inputs. Our approach is to identify policy classes that admit global worst-case guarantees, and then perform learning using historical data within the policy classes. We demonstrate the approach in the context of two classical problems, online knapsack and online set cover, proving competitive bounds for rich policy classes in each case. Additionally, we illustrate the practical implications via a case study on electric vehicle charging.

Knowledge distillation (KD) has proved to be an effective approach for deep neural network compression, which learns a compact network (student) by transferring the knowledge from a pre-trained, over-parameterized network (teacher). In traditional KD, the transferred knowledge is usually obtained by feeding training samples to the teacher network to obtain the class probabilities. However, the original training dataset is not always available due to storage costs or privacy issues. In this study, we propose a novel data-free KD approach by modeling the intermediate feature space of the teacher with a multivariate normal distribution and leveraging the soft targeted labels generated by the distribution to synthesize pseudo samples as the transfer set. Several student networks trained with these synthesized transfer sets present competitive performance compared to the networks trained with the original training set and other data-free KD approaches.

Data augmentation has been widely used to improve generalizability of machine learning models. However, comparatively little work studies data augmentation for graphs. This is largely due to the complex, non-Euclidean structure of graphs, which limits possible manipulation operations. Augmentation operations commonly used in vision and language have no analogs for graphs. Our work studies graph data augmentation for graph neural networks (GNNs) in the context of improving semi-supervised node-classification. We discuss practical and theoretical motivations, considerations and strategies for graph data augmentation. Our work shows that neural edge predictors can effectively encode class-homophilic structure to promote intra-class edges and demote inter-class edges in given graph structure, and our main contribution introduces the GAug graph data augmentation framework, which leverages these insights to improve performance in GNN-based node classification via edge prediction. Extensive experiments on multiple benchmarks show that augmentation via GAug improves performance across GNN architectures and datasets.

The progress in Query-focused Multi-Document Summarization (QMDS) has been limited by the lack of sufficient largescale high-quality training datasets. We present two QMDS training datasets, which we construct using two data augmentation methods: (1) transferring the commonly used single-document CNN/Daily Mail summarization dataset to create the QMDSCNN dataset, and (2) mining search-query logs to create the QMDSIR dataset. These two datasets have complementary properties, i.e., QMDSCNN has real summaries but queries are simulated, while QMDSIR has real queries but simulated summaries. To cover both these real summary and query aspects, we build abstractive end-to-end neural network models on the combined datasets that yield new state-of-the-art transfer results on DUC datasets. We also introduce new hierarchical encoders that enable a more efficient encoding of the query together with multiple documents. Empirical results demonstrate that our data augmentation and encoding methods outperform baseline models on automatic metrics, as well as on human evaluations along multiple attributes.

We present a method for learning dynamics of complex physical processes described by time-dependent nonlinear partial differential equations (PDEs). Our particular interest lies in extrapolating solutions in time beyond the range of temporal domain used in training. Our choice for a baseline method is physics-informed neural network (PINN) because the method parameterizes not only the solutions, but also the equations that describe the dynamics of physical processes. We demonstrate that PINN performs poorly on extrapolation tasks in many benchmark problems. To address this, we propose a novel method for better training PINN and demonstrate that our newly enhanced PINNs can accurately extrapolate solutions in time. Our method shows up to 72% smaller errors than state-of-the-art methods in terms of the standard L2-norm metric.

Filter pruning is a commonly used method for compressing Convolutional Neural Networks (ConvNets), due to its friendly hardware supporting and flexibility. However, existing methods mostly need a cumbersome procedure, which brings many extra hyper-parameters and training epochs. This is because only using sparsity and pruning stages cannot obtain a satisfying performance. Besides, many works do not consider the difference of pruning ratio across different layers. To overcome these limitations, we propose a novel dynamic and progressive filter pruning (DPFPS) scheme that directly learns a structured sparsity network from Scratch. In particular, DPFPS imposes a new structured sparsity-inducing regularization specifically upon the expected pruning parameters in a dynamic sparsity manner. The dynamic sparsity scheme determines sparsity allocation ratios of different layers and a Taylor series based channel sensitivity criteria is presented to identify the expected pruning parameters. Moreover, we increase the structured sparsity-inducing penalty in a progressive manner. This helps the model to be sparse gradually instead of forcing the model to be sparse at the beginning. Our method solves the pruning ratio based optimization problem by an iterative soft-thresholding algorithm (ISTA) with dynamic sparsity. At the end of the training, we only need to remove the redundant parameters without other stages, such as fine-tuning. Extensive experimental results show that the proposed method is competitive with 11 state-of-the-art methods on both small-scale and large-scale datasets (i.e., CIFAR and ImageNet). Specifically, on ImageNet, we achieve a 44.97% pruning ratio of FLOPs by compressing ResNet-101, even with an increase of 0.12% Top-5 accuracy. Our pruned models and codes are released at https://github.com/taoxvzi/DPFPS.

Recent years, human-object interaction (HOI) detection has achieved impressive advances. However, conventional two-stage methods are usually slow in inference. On the other hand, existing one-stage methods mainly focus on the union regions of interactions, which introduce unnecessary visual information as disturbances to HOI detection. To tackle the problems above, we propose a novel one-stage HOI detection approach DIRV in this paper, based on a new concept called interaction region for the HOI problem. Unlike previous methods, our approach concentrates on the densely sampled interaction regions across different scales for each human-object pair, so as to capture the subtle visual features that is most essential to the interaction. Moreover, in order to compensate for the detection flaws of a single interaction region, we introduce a novel voting strategy that makes full use of those overlapped interaction regions in place of conventional Non-Maximal Suppression (NMS). Extensive experiments on two popular benchmarks: V-COCO and HICO-DET show that our approach outperforms existing state-of-the-arts by a large margin with the highest inference speed and lightest network architecture. Our code is publicly available at www.github.com/MVIG-SJTU/DIRV.