AAAI 2022 论文列表

We consider the problem of acquiring causal representations and concepts in a reinforcement learning setting. Our approach defines a causal variable as being both manipulable by a policy, and able to predict the outcome. We thereby obtain a parsimonious causal graph in which interventions occur at the level of policies. The approach avoids defining a generative model of the data, prior pre-processing, or learning the transition kernel of the Markov decision process. Instead, causal variables and policies are determined by maximizing a new optimization target inspired by mediation analysis, which differs from the expected return. The maximization is accomplished using a generalization of Bellman's equation which is shown to converge, and the method finds meaningful causal representations in a simulated environment.

Time series data appears in many real-world fields such as energy, transportation, communication systems. Accurate modelling and forecasting of time series data can be of significant importance to improve the efficiency of these systems. Extensive research efforts have been taken for time series problems. Different types of approaches, including both statistical-based methods and machine learning-based methods, have been investigated. Among these methods, ensemble learning has shown to be effective and robust. However, it is still an open question that how we should determine weights for base models in the ensemble. Sub-optimal weights may prevent the final model from reaching its full potential. To deal with this challenge, we propose a reinforcement learning (RL) based model combination (RLMC) framework for determining model weights in an ensemble for time series forecasting tasks. By formulating model selection as a sequential decision-making problem, RLMC learns a deterministic policy to output dynamic model weights for non-stationary time series data. RLMC further leverages deep learning to learn hidden features from raw time series data to adapt fast to the changing data distribution. Extensive experiments on multiple real-world datasets have been implemented to showcase the effectiveness of the proposed method.

We study a finite-horizon restless multi-armed bandit problem with multiple actions, dubbed as R(MA)^2B. The state of each arm evolves according to a controlled Markov decision process (MDP), and the reward of pulling an arm depends on both the current state and action of the corresponding MDP. Since finding the optimal policy is typically intractable, we propose a computationally appealing index policy entitled Occupancy-Measured-Reward Index Policy for the finite-horizon R(MA)^2B. Our index policy is well-defined without the requirement of indexability condition and is provably asymptotically optimal as the number of arms tends to infinity. We then adopt a learning perspective where the system parameters are unknown, and propose R(MA)^2B-UCB, a generative model based reinforcement learning augmented algorithm that can fully exploit the structure of Occupancy-Measured-Reward Index Policy. Compared to existing algorithms, R(MA)^2B-UCB performs close to offline optimum, and achieves a sub-linear regret and a low computational complexity all at once. Experimental results show that R(MA)^2B-UCB outperforms existing algorithms in both regret and running time.

Despite the remarkable performance of graph neural networks (GNNs) in semi-supervised learning, it is criticized for not making full use of unlabeled data and suffering from over-fitting. Recently, graph data augmentation, used to improve both accuracy and generalization of GNNs, has received considerable attentions. However, one fundamental question is how to evaluate the quality of graph augmentations in principle? In this paper, we propose two metrics, Consistency and Diversity, from the aspects of augmentation correctness and generalization. Moreover, we discover that existing augmentations fall into a dilemma between these two metrics. Can we find a graph augmentation satisfying both consistency and diversity? A well-informed answer can help us understand the mechanism behind graph augmentation and improve the performance of GNNs. To tackle this challenge, we analyze two representative semi-supervised learning algorithms: label propagation (LP) and consistency regularization (CR). We find that LP utilizes the prior knowledge of graphs to improve consistency and CR adopts variable augmentations to promote diversity. Based on this discovery, we treat neighbors as augmentations to capture the prior knowledge embodying homophily assumption, which promises a high consistency of augmentations. To further promote diversity, we randomly replace the immediate neighbors of each node with its remote neighbors. After that, a neighbor-constrained regularization is proposed to enforce the predictions of the augmented neighbors to be consistent with each other. Extensive experiments on five real-world graphs validate the superiority of our method in improving the accuracy and generalization of GNNs.

End-to-end speech-to-text translation (E2E-ST) is becoming increasingly popular due to the potential of its less error propagation, lower latency, and fewer parameters. Given the triplet training corpus〈speech, transcription, translation〉, the conventional high-quality E2E-ST system leverages the〈speech, transcription〉pair to pre-train the model and then utilizes the〈speech, translation〉pair to optimize it further. However, this process only involves two-tuple data at each stage, and this loose coupling fails to fully exploit the association between triplet data. In this paper, we attempt to model the joint probability of transcription and translation based on the speech input to directly leverage such triplet data. Based on that, we propose a novel regularization method for model training to improve the agreement of dual-path decomposition within triplet data, which should be equal in theory. To achieve this goal, we introduce two Kullback-Leibler divergence regularization terms into the model training objective to reduce the mismatch between output probabilities of dual-path. Then the well-trained model can be naturally transformed as the E2E-ST models by a pre-defined early stop tag. Experiments on the MuST-C benchmark demonstrate that our proposed approach significantly outperforms state-of-the-art E2E-ST baselines on all 8 language pairs while achieving better performance in the automatic speech recognition task.

Modal regression, a widely used regression protocol, has been extensively investigated in statistical and machine learning communities due to its robustness to outlier and heavy-tailed noises. Understanding modal regression's theoretical behavior can be fundamental in learning theory. Despite significant progress in characterizing its statistical property, the majority results are based on the assumption that samples are independent and identical distributed (i.i.d.), which is too restrictive for real-world applications. This paper concerns about the statistical property of regularized modal regression (RMR) within an important dependence structure - Markov dependent. Specifically, we establish the upper bound for RMR estimator under moderate conditions and give an explicit learning rate. Our results show that the Markov dependence impacts on the generalization error in the way that sample size would be discounted by a multiplicative factor depending on the spectral gap of the underlying Markov chain. This result shed a new light on characterizing the theoretical underpinning for robust regression.

Due to the poor generalization performance of traditional empirical risk minimization (ERM) in the case of distributional shift, Out-of-Distribution (OoD) generalization algorithms receive increasing attention. However, OoD generalization algorithms overlook the great variance in the quality of training data, which significantly compromises the accuracy of these methods. In this paper, we theoretically reveal the relationship between training data quality and algorithm performance, and analyze the optimal regularization scheme for Lipschitz regularized invariant risk minimization. A novel algorithm is proposed based on the theoretical results to alleviate the influence of low quality data at both the sample level and the domain level. The experiments on both the regression and classification benchmarks validate the effectiveness of our method with statistical significance.

In the Bayesian reinforcement learning (RL) setting, a prior distribution over the unknown problem parameters -- the rewards and transitions -- is assumed, and a policy that optimizes the (posterior) expected return is sought. A common approximation, which has been recently popularized as meta-RL, is to train the agent on a sample of N problem instances from the prior, with the hope that for large enough N, good generalization behavior to an unseen test instance will be obtained. In this work, we study generalization in Bayesian RL under the probably approximately correct (PAC) framework, using the method of algorithmic stability. Our main contribution is showing that by adding regularization, the optimal policy becomes uniformly stable in an appropriate sense. Most stability results in the literature build on strong convexity of the regularized loss -- an approach that is not suitable for RL as Markov decision processes (MDPs) are not convex. Instead, building on recent results of fast convergence rates for mirror descent in regularized MDPs, we show that regularized MDPs satisfy a certain quadratic growth criterion, which is sufficient to establish stability. This result, which may be of independent interest, allows us to study the effect of regularization on generalization in the Bayesian RL setting.

We consider the problem of reforming an envy-free matching when each agent is assigned a single item. Given an envy-free matching, we consider an operation to exchange the item of an agent with an unassigned item preferred by the agent that results in another envy-free matching. We repeat this operation as long as we can. We prove that the resulting envy-free matching is uniquely determined up to the choice of an initial envy-free matching, and can be found in polynomial time. We call the resulting matching a reformist envy-free matching, and then we study a shortest sequence to obtain the reformist envy-free matching from an initial envy-free matching. We prove that a shortest sequence is computationally hard to obtain even when each agent accepts at most four items and each item is accepted by at most three agents. On the other hand, we give polynomial-time algorithms when each agent accepts at most three items or each item is accepted by at most two agents. Inapproximability and fixed-parameter (in)tractability are also discussed.

Producing densely annotated data is a difficult and tedious task for medical imaging applications. To address this problem, we propose a novel approach to generate supervision for semi-supervised semantic segmentation. We argue that visually similar regions between labeled and unlabeled images likely contain the same semantics and therefore should share their label. Following this thought, we use a small number of labeled images as reference material and match pixels in an unlabeled image to the semantic of the best fitting pixel in a reference set. This way, we avoid pitfalls such as confirmation bias, common in purely prediction-based pseudo-labeling. Since our method does not require any architectural changes or accompanying networks, one can easily insert it into existing frameworks. We achieve the same performance as a standard fully supervised model on X-ray anatomy segmentation, albeit using 95% fewer labeled images. Aside from an in-depth analysis of different aspects of our proposed method, we further demonstrate the effectiveness of our reference-guided learning paradigm by comparing our approach against existing methods for retinal fluid segmentation with competitive performance as we improve upon recent work by up to 15% mean IoU.

Speech enhancement aims at recovering a clean speech from a noisy input, which can be classified into single speech enhancement and personalized speech enhancement. Personalized speech enhancement usually utilizes the speaker identity extracted from the noisy speech itself (or a clean reference speech) as a global embedding to guide the enhancement process. Different from them, we observe that the speeches of the same speaker are correlated in terms of frame-level short-time Fourier Transform (STFT) spectrogram. Therefore, we propose reference-based speech enhancement via a feature alignment and fusion network (FAF-Net). Given a noisy speech and a clean reference speech spoken by the same speaker, we first propose a feature level alignment strategy to warp the clean reference with the noisy speech in frame level. Then, we fuse the reference feature with the noisy feature via a similarity-based fusion strategy. Finally, the fused features are skipped connected to the decoder, which generates the enhanced results. Experimental results demonstrate that the performance of the proposed FAF-Net is close to state-of-the-art speech enhancement methods on both DNS and Voice Bank+DEMAND datasets. Our code is available at https://github.com/HieDean/FAF-Net.

Deep neural networks (DNNs) have been widely applied in various domains in artificial intelligence including computer vision and natural language processing. A DNN is typically trained for many epochs and then a validation dataset is used to select the DNN in an epoch (we simply call this epoch ``the last epoch") as the final model for making predictions on unseen samples, while it usually cannot achieve a perfect accuracy on unseen samples. An interesting question is ``how many test (unseen) samples that a DNN misclassifies in the last epoch were ever correctly classified by the DNN before the last epoch?". In this paper, we empirically study this question and find on several benchmark datasets that the vast majority of the misclassified samples in the last epoch were ever classified correctly before the last epoch, which means that the predictions for these samples were flipped from ``correct" to ``wrong". Motivated by this observation, we propose to restrict the behavior changes of a DNN on the correctly-classified samples so that the correct local boundaries can be maintained and the flipping error on unseen samples can be largely reduced. Extensive experiments on different benchmark datasets with different modern network architectures demonstrate that the proposed flipping error reduction (FER) approach can substantially improve the generalization, the robustness, and the transferability of DNNs without introducing any additional network parameters or inference cost, only with a negligible training overhead.

Multi-agent reinforcement learning (MARL) provides an efficient way for simultaneously learning policies for multiple agents interacting with each other. However, in scenarios requiring complex interactions, existing algorithms can suffer from an inability to accurately anticipate the influence of self-actions on other agents. Incorporating an ability to reason about other agents' potential responses can allow an agent to formulate more effective strategies. This paper adopts a recursive reasoning model in a centralized-training-decentralized-execution framework to help learning agents better cooperate with or compete against others. The proposed algorithm, referred to as the Recursive Reasoning Graph (R2G), shows state-of-the-art performance on multiple multi-agent particle and robotics games.

Neural network controllers have become popular in control tasks thanks to their flexibility and expressivity. Stability is a crucial property for safety-critical dynamical systems, while stabilization of partially observed systems, in many cases, requires controllers to retain and process long-term memories of the past. We consider the important class of recurrent neural networks (RNN) as dynamic controllers for nonlinear uncertain partially-observed systems, and derive convex stability conditions based on integral quadratic constraints, S-lemma and sequential convexification. To ensure stability during the learning and control process, we propose a projected policy gradient method that iteratively enforces the stability conditions in the reparametrized space taking advantage of mild additional information on system dynamics. Numerical experiments show that our method learns stabilizing controllers with fewer samples and achieves higher final performance compared with policy gradient.

Positive-Unlabeled (PU) learning methods train a classifier to distinguish between the positive and negative classes given only positive and unlabeled data. While traditional PU methods require the labeled positive samples to be an unbiased sample of the positive distribution, in practice the labeled sample is often a biased draw from the true distribution. Prior work shows that if we know the likelihood that each positive instance will be selected for labeling, referred to as the propensity score, then the biased sample can be used for PU learning. Unfortunately, no prior work has been proposed an inference strategy for which the propensity score is identifiable. In this work, we propose two sets of assumptions under which the propensity score can be uniquely determined: one in which no assumption is made on the functional form of the propensity score (requiring assumptions on the data distribution), and the second which loosens the data assumptions while assuming a functional form for the propensity score. We then propose inference strategies for each case. Our empirical study shows that our approach significantly outperforms the state-of-the-art propensity estimation methods on a rich variety of benchmark datasets.

Reconfiguring two shortest paths in a graph means modifying one shortest path to the other by changing one vertex at a time, so that all the intermediate paths are also shortest paths. This problem has several natural applications, namely: (a) revamping road networks, (b) rerouting data packets in a synchronous multiprocessing setting, (c) the shipping container stowage problem, and (d) the train marshalling problem. When modelled as graph problems, (a) is the most general case while (b), (c) and (d) are restrictions to different graph classes. We show that (a) is intractable, even for relaxed variants of the problem. For (b), (c) and (d), we present efficient algorithms to solve the respective problems. We also generalise the problem to when at most k (for some k >= 2) contiguous vertices on a shortest path can be changed at a time.

Generalized structural equations models (GSEMs) (Peters and Halpern 2021), are, as the name suggests, a generalization of structural equations models (SEMs). They can deal with (among other things) infinitely many variables with infinite ranges, which is critical for capturing dynamical systems. We provide a sound and complete axiomatization of causal reasoning in GSEMs that is an extension of the sound and complete axiomatization provided by Halpern (2000) for SEMs. Considering GSEMs helps clarify what properties Halpern's axioms capture.

Online on-demand ridesourcing service has played a huge role in transforming urban transportation. A central function in most on-demand ridesourcing platforms is to dynamically assign drivers to rider requests that could balance the request waiting times and the driver pick-up distances. To deal with the online nature of this problem, existing literature either divides the time horizon into short windows and applies a static offline assignment algorithm within each window or assumes a fully online setting that makes decisions for each request immediately upon its arrival. In this paper, we propose a more realistic model for the driver-request assignment that bridges the above two settings together. Our model allows the requests to wait after their arrival but assumes that they may leave at any time following a quitting function. Under this model, we design an efficient algorithm for assigning available drivers to requests in real-time. Our algorithm is able to incorporate future estimated driver arrivals into consideration and make strategic waiting and matching decisions that could balance the waiting time and pick-up distance of the assignment. We prove that our algorithm is optimal ex-ante in the single-request setting, and demonstrate its effectiveness in the general multi-request setting through experiments on both synthetic and real-world datasets.

Exiting simple samples in adaptive multi-exit networks through early modules is an effective way to achieve high computational efficiency. One can observe that deployments of multi-exit architectures on resource-constrained devices are easily limited by high memory footprint of early modules. In this paper, we propose a novel approach named recurrent aggregation operator (ReX), which uses recurrent neural networks (RNNs) to effectively aggregate intra-patch features within a large receptive field to get delicate local representations, while bypassing large early activations. The resulting model, named ReXNet, can be easily extended to dynamic inference by introducing a novel consistency-based early exit criteria, which is based on the consistency of classification decisions over several modules, rather than the entropy of the prediction distribution. Extensive experiments on two benchmark datasets, i.e., Visual Wake Words, ImageNet-1k, demonstrate that our method consistently reduces the peak RAM and average latency of a wide variety of adaptive models on low-power devices.