IJCAI 2021 论文列表

This work presents a new fine-grained transparent object segmentation dataset, termed Trans10K-v2, extending Trans10K-v1, the first large-scale transparent object segmentation dataset. Unlike Trans10K-v1 that only has two limited categories, our new dataset has several appealing benefits. (1) It has 11 fine-grained categories of transparent objects, commonly occurring in the human domestic environment, making it more practical for real-world application. (2) Trans10K-v2 brings more challenges for the current advanced segmentation methods than its former version. Furthermore, a novel Transformer-based segmentation pipeline termed Trans2Seg is proposed. Firstly, the Transformer encoder of Trans2Seg provides the global receptive field in contrast to CNN's local receptive field, which shows excellent advantages over pure CNN architectures. Secondly, by formulating semantic segmentation as a problem of dictionary look-up, we design a set of learnable prototypes as the query of Trans2Seg's Transformer decoder, where each prototype learns the statistics of one category in the whole dataset. We benchmark more than 20 recent semantic segmentation methods, demonstrating that Trans2Seg significantly outperforms all the CNN-based methods, showing the proposed algorithm's potential ability to solve transparent object segmentation.Code is available in https://github.com/xieenze/Trans2Seg.

Many data mining and analytical tasks rely on the abstraction of networks (graphs) to summarize relational structures among individuals (nodes). Since relational data are often sensitive, we aim to seek effective approaches to generate utility-preserved yet privacy-protected structured data. In this paper, we leverage the differential privacy (DP) framework to formulate and enforce rigorous privacy constraints on deep graph generation models, with a focus on edge-DP to guarantee individual link privacy. In particular, we enforce edge-DP by injecting designated noise to the gradients of a link reconstruction based graph generation model, while ensuring data utility by improving structure learning with structure-oriented graph discrimination. Extensive experiments on two real-world network datasets show that our proposed DPGGAN model is able to generate graphs with effectively preserved global structure and rigorously protected individual link privacy.

We present a new and rich model of school choice with flexible diversity goals and specialized seats. The model also applies to other settings such as public housing allocation with diversity objectives. Our method of expressing flexible diversity goals is also applicable to other settings in moral multi-agent decision making where competing policies need to be balanced when allocating scarce resources. For our matching model, we present a polynomial-time algorithm that satisfies desirable properties, including strategyproofness and stability under several natural subdomains of our problem. We complement the results by providing a clear understanding about what results do not extend when considering the general model.

Recently, novel ILP systems under the answer set semantics have been proposed, some of which are robust to noise and scalable over large hypothesis spaces. One such system is FastLAS, which is significantly faster than other state-of-the-art ASP-based ILP systems. FastLAS is, however, only capable of Observational Predicate Learning (OPL), where the learned hypothesis defines predicates that are directly observed in the examples. It cannot learn knowledge that is indirectly observable, such as learning causes of observed events. This class of problems, known as non-OPL, is known to be difficult to handle in the context of non-monotonic semantics. Solving non-OPL learning tasks whilst preserving scalability is a challenging open problem. We address this problem with a new abductive method for translating examples of a non-OPL task to a set of examples, called possibilities, such that the original example is covered iff at least one of the possibilities is covered. This new method allows an ILP system capable of performing OPL tasks to be "upgraded" to solve non-OPL tasks. In particular, we present our new FastNonOPL system, which upgrades FastLAS with the new possibility generation. We compare it to other state-of-the-art ASP-based ILP systems capable of solving non-OPL tasks, showing that FastNonOPL is significantly faster, and in many cases more accurate, than these other systems.

Continuous DR-submodular maximization is an important machine learning problem, which covers numerous popular applications. With the emergence of large-scale distributed data, developing efficient algorithms for the continuous DR-submodular maximization, such as the decentralized Frank-Wolfe method, became an important challenge. However, existing decentralized Frank-Wolfe methods for this kind of problem have the sample complexity of $\mathcal{O}(1/\epsilon^3)$, incurring a large computational overhead. In this paper, we propose two novel sample efficient decentralized Frank-Wolfe methods to address this challenge. Our theoretical results demonstrate that the sample complexity of the two proposed methods is $\mathcal{O}(1/\epsilon^2)$, which is better than $\mathcal{O}(1/\epsilon^3)$ of the existing methods. As far as we know, this is the first published result achieving such a favorable sample complexity. Extensive experimental results confirm the effectiveness of the proposed methods.

Sleep staging is fundamental for sleep assessment and disease diagnosis. Although previous attempts to classify sleep stages have achieved high classification performance, several challenges remain open: 1) How to effectively extract salient waves in multimodal sleep data; 2) How to capture the multi-scale transition rules among sleep stages; 3) How to adaptively seize the key role of specific modality for sleep staging. To address these challenges, we propose SalientSleepNet, a multimodal salient wave detection network for sleep staging. Specifically, SalientSleepNet is a temporal fully convolutional network based on the $U^2$-Net architecture that is originally proposed for salient object detection in computer vision. It is mainly composed of two independent $U^2$-like streams to extract the salient features from multimodal data, respectively. Meanwhile, the multi-scale extraction module is designed to capture multi-scale transition rules among sleep stages. Besides, the multimodal attention module is proposed to adaptively capture valuable information from multimodal data for the specific sleep stage. Experiments on the two datasets demonstrate that SalientSleepNet outperforms the state-of-the-art baselines. It is worth noting that this model has the least amount of parameters compared with the existing deep neural network models.

Deep Neural Networks (DNNs) are popular machine learning models which have found successful application in many different domains across computer science. Nevertheless, providing formal guarantees on the behaviour of neural networks is hard and therefore their reliability in safety-critical domains is still a concern. Verification and repair emerged as promising solutions to address this issue. In the following, I will present some of my recent efforts in this area.

Medication recommendation is an essential task of AI for healthcare. Existing works focused on recommending drug combinations for patients with complex health conditions solely based on their electronic health records. Thus, they have the following limitations: (1) some important data such as drug molecule structures have not been utilized in the recommendation process. (2) drug-drug interactions (DDI) are modeled implicitly, which can lead to sub-optimal results. To address these limitations, we propose a DDI-controllable drug recommendation model named SafeDrug to leverage drugs’ molecule structures and model DDIs explicitly. SafeDrug is equipped with a global message passing neural network (MPNN) module and a local bipartite learning module to fully encode the connectivity and functionality of drug molecules. SafeDrug also has a controllable loss function to control DDI level in the recommended drug combinations effectively. On a benchmark dataset, our SafeDrug is relatively shown to reduce DDI by 19.43% and improves 2.88% on Jaccard similarity between recommended and actually prescribed drug combinations over previous approaches. Moreover, SafeDrug also requires much fewer parameters than previous deep learning based approaches, leading to faster training by about 14% and around 2× speed-up in inference.

Weakly supervised learning (WSL) refers to learning from a large amount of weak supervision data. This includes i) incomplete supervision (e.g., semi-supervised learning); ii) inexact supervision (e.g., multi-instance learning) and iii) inaccurate supervision (e.g., label noise learning). Unlike supervised learning which typically achieves performance improvement with more labeled data, WSL may sometimes even degenerate performance with more weak supervision data. It is thus desired to study safe WSL, which could robustly improve performance with weak supervision data. In this article, we share our understanding of the problem from in-distribution data to out-of-distribution data, and discuss possible ways to alleviate it, from the aspects of worst-case analysis, ensemble-learning, and bi-level optimization. We also share some open problems, to inspire future researches.

Information flow measures, over the duration of a game, the audience’s belief of who will win, and thus can reflect the amount of surprise in a game. To quantify the relationship between information flow and audiences' perceived quality, we conduct a case study where subjects watch one of the world’s biggest esports events, LOL S10. In addition to eliciting information flow, we also ask subjects to report their rating for each game. We find that the amount of surprise in the end of the game plays a dominant role in predicting the rating. This suggests the importance of incorporating when the surprise occurs, in addition to the amount of surprise, in perceived quality models. For content providers, it implies that everything else being equal, it is better for twists to be more likely to happen toward the end of a show rather than uniformly throughout.

Error detection is one of the most important steps in data cleaning and usually requires extensive human interaction to ensure quality. Existing supervised methods in error detection require a significant amount of training data while unsupervised methods rely on fixed inductive biases, which are usually hard to generalize, to solve the problem. In this paper, we present SPADE, a novel semi-supervised probabilistic approach for error detection. SPADE introduces a novel probabilistic active learning model, where the system suggests examples to be labeled based on the agreements between user labels and indicative signals, which are designed to capture potential errors. SPADE uses a two-phase data augmentation process to enrich a dataset before training a deep learning classifier to detect unlabeled errors. In our evaluation, SPADE achieves an average F1-score of 0.91 over five datasets and yields a 10% improvement compared with the state-of-the-art systems.

Recently, the Stochastic Particle Optimization Sampling (SPOS) method is proposed to solve the particle-collapsing pitfall of deterministic Particle Variational Inference methods by ultilizing the stochastic Overdamped Langevin dynamics to enhance exploration. In this paper, we propose an accelerated particle optimization sampling method called Stochastic Hamiltonian Particle Optimization Sampling (SHPOS). Compared to the first-order dynamics used in SPOS, SHPOS adopts an augmented second-order dynamics, which involves an extra momentum term to achieve acceleration. We establish a non-asymptotic convergence analysis for SHPOS, and show that it enjoys a faster convergence rate than SPOS. Besides, we also propose a variance-reduced stochastic gradient variant of SHPOS for tasks with large-scale datasets and complex models. Experiments on both synthetic and real data validate our theory and demonstrate the superiority of SHPOS over the state-of-the-art.

Hierarchical reinforcement learning (HRL) is only effective for long-horizon problems when high-level skills can be reliably sequentially executed. Unfortunately, learning reliably composable skills is difficult, because all the components of every skill are constantly changing during learning. We propose three methods for improving the composability of learned skills: representing skill initiation regions using a combination of pessimistic and optimistic classifiers; learning re-targetable policies that are robust to non-stationary subgoal regions; and learning robust option policies using model-based RL. We test these improvements on four sparse-reward maze navigation tasks involving a simulated quadrupedal robot. Each method successively improves the robustness of a baseline skill discovery method, substantially outperforming state-of-the-art flat and hierarchical methods.

Recent researches have suggested that the predictive accuracy of neural network may contend with its adversarial robustness. This presents challenges in designing effective regularization schemes that also provide strong adversarial robustness. Revisiting Vicinal Risk Minimization (VRM) as a unifying regularization principle, we propose Adversarial Labelling of Perturbed Samples (ALPS) as a regularization scheme that aims at improving the generalization ability and adversarial robustness of the trained model. ALPS trains neural networks with synthetic samples formed by perturbing each authentic input sample towards another one along with an adversarially assigned label. The ALPS regularization objective is formulated as a min-max problem, in which the outer problem is minimizing an upper-bound of the VRM loss, and the inner problem is L1-ball constrained adversarial labelling on perturbed sample. The analytic solution to the induced inner maximization problem is elegantly derived, which enables computational efficiency. Experiments on the SVHN, CIFAR-10, CIFAR-100 and Tiny-ImageNet datasets show that the ALPS has a state-of-the-art regularization performance while also serving as an effective adversarial training scheme.

Domain Invariant Representations (IR) has improved drastically the transferability of representations from a labelled source domain to a new and unlabelled target domain. Unsupervised Domain Adaptation (UDA) in presence of label shift remains an open problem. To this purpose, we present a bound of the target risk which incorporates both weights and invariant representations. Our theoretical analysis highlights the role of inductive bias in aligning distributions across domains. We illustrate it on standard benchmarks by proposing a new learning procedure for UDA. We observed empirically that weak inductive bias makes adaptation robust to label shift. The elaboration of stronger inductive bias is a promising direction for new UDA algorithms.

The agent in imitation learning (IL) is expected to mimic the behavior of the expert. Its performance relies highly on the quality of given expert demonstrations. However, the assumption that collected demonstrations are optimal cannot always hold in real-world tasks, which would seriously influence the performance of the learned agent. In this paper, we propose a robust method within the framework of Generative Adversarial Imitation Learning (GAIL) to address imperfect demonstration issue, in which good demonstrations can be adaptively selected for training while bad demonstrations are abandoned. Specifically, a binary weight is assigned to each expert demonstration to indicate whether to select it for training. The reward function in GAIL is employed to determine this weight (i.e. higher reward results in higher weight). Compared to some existing solutions that require some auxiliary information about this weight, we set up the connection between weight and model so that we can jointly optimize GAIL and learn the latent weight. Besides hard binary weighting, we also propose a soft weighting scheme. Experiments in the Mujoco demonstrate the proposed method outperforms other GAIL-based methods when dealing with imperfect demonstrations.

Imitation learning allows learning complex behaviors given demonstrations. Early approaches belonging to either Behavior Cloning or Inverse Reinforcement Learning were however of limited scalability to complex environments. A more promising approach termed as Generative Adversarial Imitation Learning tackles the imitation learning problem by drawing a connection with Generative Adversarial Networks. In this work, we advocate the use of this class of methods and investigate possible extensions by endowing them with global temporal consistency, in particular through a contrastive learning based approach.

We propose a stochastic recursive momentum method for Riemannian non-convex optimization that achieves a nearly-optimal complexity to find epsilon-approximate solution with one sample. The new algorithm requires one-sample gradient evaluations per iteration and does not require restarting with a large batch gradient, which is commonly used to obtain a faster rate. Extensive experiment results demonstrate the superiority of the proposed algorithm. Extensions to nonsmooth and constrained optimization settings are also discussed.

Deep reinforcement learning (RL) has demonstrated success in challenging decision-making/control tasks. However, RL methods, which solve tasks through maximizing the expected reward, may generate undesirable behaviors due to inferior local convergence or incompetent reward design. These undesirable behaviors of agents may not reduce the total reward but destroy the user experience of the application. For example, in the autonomous driving task, the policy actuated by speed reward behaves much more sudden brakes while human drivers generally don’t do that. To overcome this problem, we present a novel method named Reward-Constrained Behavior Cloning (RCBC) which synthesizes imitation learning and constrained reinforcement learning. RCBC leverages human demonstrations to induce desirable or human-like behaviors and employs lower-bound reward constraints for policy optimization to maximize the expected reward. Empirical results on popular benchmark environments show that RCBC learns significantly more human-desired policies with performance guarantees which meet the lower-bound reward constraints while performing better than or as well as baseline methods in terms of reward maximization.