IJCAI 2022 论文列表

In the semi-supervised object detection task, due to the scarcity of labeled data and the diversity and complexity of objects to be detected, the quality of pseudo-labels generated by existing methods for unlabeled data is relatively low, which severely restricts the performance of semi-supervised object detection. In this paper, we revisit the pseudo-labeling based Teacher-Student mutual learning framework for semi-supervised object detection and identify that the inconsistency of the location and feature of the candidate object proposals between the Teacher and the Student branches are the fatal cause of the low quality of the pseudo labels. To address this issue, we propose a simple yet effective technique within the mainstream teacher-student framework, called Double Check Soft Teacher, to overcome the harm caused by insufficient quality of pseudo labels. Specifically, our proposed method leverages teacher model to generate pseudo labels for the student model. Especially, the candidate boxes generated by the student model based on the pseudo label will be sent to the teacher model for "double check", and then the teacher model will output probabilistic soft label with background class for those candidate boxes, which will be used to train the student model. Together with a pseudo labeling mechanism based on the sum of the TOP-K prediction score, which improves the recall rate of pseudo labels, Double Check Soft Teacher consistently surpasses state-of-the-art methods by significant margins on the MS-COCO benchmark, pushing the new state-of-the-art. Source codes are available at https://github.com/wkfdb/DCST.

One of the key challenges in visual Reinforcement Learning (RL) is to learn policies that can generalize to unseen environments. Recently, data augmentation techniques aiming at enhancing data diversity have demonstrated proven performance in improving the generalization ability of learned policies. However, due to the sensitivity of RL training, naively applying data augmentation, which transforms each pixel in a task-agnostic manner, may suffer from instability and damage the sample efficiency, thus further exacerbating the generalization performance. At the heart of this phenomenon is the diverged action distribution and high-variance value estimation in the face of augmented images. To alleviate this issue, we propose Task-aware Lipschitz Data Augmentation (TLDA) for visual RL, which explicitly identifies the task-correlated pixels with large Lipschitz constants, and only augments the task-irrelevant pixels for stability. We verify the effectiveness of our approach on DeepMind Control suite, CARLA and DeepMind Manipulation tasks. The extensive empirical results show that TLDA improves both sample efficiency and generalization; it outperforms previous state-of-the-art methods across 3 different visual control benchmarks.

Domain adaptive text classification is a challenging problem for the large-scale pretrained language models because they often require expensive additional labeled data to adapt to new domains. Existing works usually fails to leverage the implicit relationships among words across domains. In this paper, we propose a novel method, called Domain Adaptation with Structured Knowledge (DASK), to enhance domain adaptation by exploiting word-level semantic relationships. DASK first builds a knowledge graph to capture the relationship between pivot terms (domain-independent words) and non-pivot terms in the target domain. Then during training, DASK injects pivot-related knowledge graph information into source domain texts. For the downstream task, these knowledge-injected texts are fed into a BERT variant capable of processing knowledge-injected textual data. Thanks to the knowledge injection, our model learns domain-invariant features for non-pivots according to their relationships with pivots. DASK ensures the pivots to have domain-invariant behaviors by dynamically inferring via the polarity scores of candidate pivots during training with pseudo-labels. We validate DASK on a wide range of cross-domain sentiment classification tasks and observe up to 2.9% absolute performance improvement over baselines for 20 different domain pairs. Code is available at https://github.com/hikaru-nara/DASK.

Domain generalization (DG) aims at generalizing a classifier trained on multiple source domains to an unseen target domain with domain shift. A common pervasive theme in existing DG literature is domain-invariant representation learning with various invariance assumptions. However, prior works restrict themselves to an impractical assumption for real-world challenges: If a mapping induced by a deep neural network (DNN) could align the source domains well, then such a mapping aligns a target domain as well. In this paper, we simply take DNNs as feature extractors to relax the requirement of distribution alignment. Specifically, we put forward a novel angular invariance and the accompanied norm shift assumption. Based on the proposed term of invariance, we propose a novel deep DG method dubbed Angular Invariance Domain Generalization Network (AIDGN). The optimization objective of AIDGN is developed with a von-Mises Fisher (vMF) mixture model. Extensive experiments on multiple DG benchmark datasets validate the effectiveness of the proposed AIDGN method.

Color Constancy aims to eliminate the color cast of RAW images caused by non-neutral illuminants. Though contemporary approaches based on convolutional neural networks significantly improve illuminant estimation, they suffer from the seriously insufficient data problem. To solve this problem by effectively utilizing multi-domain data, we propose the Domain Adversarial Learning Color Constancy (DALCC) which consists of the Domain Adversarial Learning Branch (DALB) and the Feature Reweighting Module (FRM). In DALB, the Camera Domain Classifier and the feature extractor compete against each other in an adversarial way to encourage the emergence of domain-invariant features. At the same time, the Illuminant Transformation Module performs color space conversion to solve the inconsistent color space problem caused by those domain-invariant features. They collaboratively avoid model degradation of multi-device training caused by the domain discrepancy of feature distribution, which enables our DALCC to benefit from multi-domain data. Besides, to better utilize multi-domain data, we propose the FRM that reweights the feature map to suppress Non-Primary Illuminant regions, which reduces the influence of misleading illuminant information. Experiments show that the proposed DALCC can more effectively take advantage of multi-domain data and thus achieve state-of-the-art performance on commonly used benchmark datasets.

Unsupervised domain adaptation (UDA), which is an important topic in transfer learning, aims to predict unlabeled data from target domain with access to labeled data from the source domain. In this work, we propose a novel framework called SIDA (Surrogate Mutual Information Maximization Domain Adaptation) with strong theoretical guarantees. To be specific, SIDA implements adaptation by maximizing mutual information (MI) between features. In the framework, a surrogate joint distribution models the underlying joint distribution of the unlabeled target domain. Our theoretical analysis validates SIDA by bounding the expected risk on target domain with MI and surrogate distribution bias. Experiments show that our approach is comparable with state-of-the-art unsupervised adaptation methods on standard UDA tasks.

Document-level relation extraction aims to extract relations between entities in a document. In contrast to sentence-level relation extraction, it deals with longer texts and more complex entity interactions, which requires reasoning over multiple sentences with rich reasoning skills. Most current researches construct a document-level graph first, and then focus on the overall graph structure or the paths between the target entity pair in the graph. In this paper, we propose a novel subgraph reasoning (SGR) framework for document-level relation extraction. SGR combines the advantages of both graph-based models and path-based models, integrating various paths between the target entity pair into a much simpler subgraph structure to perform relational reasoning. Moreover, the paths generated by our designed heuristic strategy explicitly model the requisite reasoning skills and roughly cover the supporting sentences for each relation instance. Experimental results on DocRED show that SGR outperforms existing models, and further analyses demonstrate that our method is both effective and explainable. Our code is available at https://github.com/Crysta1ovo/SGR.

Document-level Event Factuality Identification (DEFI) predicts the event factuality according to the current document, and mainly depends on event-related tokens and sentences. However, previous studies relied on annotated information and did not filter irrelevant and noisy texts. Therefore, this paper proposes a novel end-to-end model, i.e., Reinforced Multi-Granularity Hierarchical Attention Network (RMHAN), which can learn information at different levels of granularity from tokens and sentences hierarchically. Moreover, with hierarchical reinforcement learning, RMHAN first selects relevant and meaningful tokens, and then selects useful sentences for document-level encoding. Experimental results on DLEF-v2 corpus show that RMHAN model outperforms several state-of-the-art baselines and achieves the best performance.

Few-shot Intent Detection (FSID) is a challenging task in dialogue systems due to the scarcity of available annotated utterances. Although existing few-shot learning approaches have made remarkable progress, they fall short in adapting to the Generalized Few-shot Intent Detection (GFSID) task where both seen and unseen classes are present. A core problem of the simultaneous existence of these two tasks is that limited training samples fail to cover the diversity of user expressions. In this paper, we propose an effective Diversity Features Enhanced Prototypical Network (DFEPN) to enhance diversity features for novel intents by fully exploiting the diversity of known intent samples. Specially, DFEPN generates diversity features of samples in the hidden space via a diversity feature generator module and then fuses these features with original support vectors to get a more suitable prototype vector of each class. To evaluate the effectiveness of our model on both FSID and GFSID tasks, we carry out sufficient experiments on two benchmark intent detection datasets. Results demonstrate that our proposed model outperforms existing state-of-the-art methods and keeps stable performance on both two tasks.

Gram-based and patch-based approaches are two important research lines of style transfer. Recent diversified Gram-based methods have been able to produce multiple and diverse stylized outputs for the same content and style images. However, as another widespread research interest, the diversity of patch-based methods remains challenging due to the stereotyped style swapping process based on nearest patch matching. To resolve this dilemma, in this paper, we dive into the crux of existing patch-based methods and propose a universal and efficient module, termed DivSwapper, for diversified patch-based arbitrary style transfer. The key insight is to use an essential intuition that neural patches with higher activation values could contribute more to diversity. Our DivSwapper is plug-and-play and can be easily integrated into existing patch-based and Gram-based methods to generate diverse results for arbitrary styles. We conduct theoretical analyses and extensive experiments to demonstrate the effectiveness of our method, and compared with state-of-the-art algorithms, it shows superiority in diversity, quality, and efficiency.

A high-resolution network exhibits remarkable capability in extracting multi-scale features for human pose estimation, but fails to capture long-range interactions between joints and has high computational complexity. To address these problems, we present a Dynamic lightweight High-Resolution Network (Dite-HRNet), which can efficiently extract multi-scale contextual information and model long-range spatial dependency for human pose estimation. Specifically, we propose two methods, dynamic split convolution and adaptive context modeling, and embed them into two novel lightweight blocks, which are named dynamic multi-scale context block and dynamic global context block. These two blocks, as the basic component units of our Dite-HRNet, are specially designed for the high-resolution networks to make full use of the parallel multi-resolution architecture. Experimental results show that the proposed network achieves superior performance on both COCO and MPII human pose estimation datasets, surpassing the state-of-the-art lightweight networks. Code is available at: https://github.com/ZiyiZhang27/Dite-HRNet.

A fundamental question in social choice and multi-agent systems is aggregating ordinal preferences expressed by agents into a measurably prudent collective choice. A promising line of recent work views ordinal preferences as a proxy for underlying cardinal preferences. It aims to optimize distortion, the worst-case approximation ratio of the (utilitarian) social welfare. When agents rank the set of alternatives, prior work identifies near-optimal voting rules for selecting one or more alternatives. However, ranking all the alternatives is prohibitive when there are many alternatives. In this work, we consider the setting where each agent ranks only her t favorite alternatives and identify almost tight bounds on the best possible distortion when selecting a single alternative or a committee of alternatives of a given size k. Our results also extend to approximating higher moments of social welfare. Along the way, we close a gap left open in prior work by identifying asymptotically tight distortion bounds for committee selection given full rankings.

Knowledge distillation is widely adopted in semantic segmentation to reduce the computation cost. The previous knowledge distillation methods for semantic segmentation focus on pixel-wise feature alignment and intra-class feature variation distillation, neglecting to transfer the knowledge of the inter-class distance in the feature space, which is important for semantic segmentation such a pixel-wise classification task. To address this issue, we propose an Inter-class Distance Distillation (IDD) method to transfer the inter-class distance in the feature space from the teacher network to the student network. Furthermore, semantic segmentation is a position-dependent task, thus we exploit a position information distillation module to help the student network encode more position information. Extensive experiments on three popular datasets: Cityscapes, Pascal VOC and ADE20K show that our method is helpful to improve the accuracy of semantic segmentation models and achieves the state-of-the-art performance. E.g. it boosts the benchmark model (``PSPNet+ResNet18") by 7.50% in accuracy on the Cityscapes dataset.

Distilling interpretable physical laws from videos has led to expanded interest in the computer vision community recently thanks to the advances in deep learning, but still remains a great challenge. This paper introduces an end-to-end unsupervised deep learning framework to uncover the explicit governing equations of dynamics presented by moving object(s), based on recorded videos. Instead in the pixel (spatial) coordinate system of image space, the physical law is modeled in a regressed underlying physical coordinate system where the physical states follow potential explicit governing equations. A numerical integrator-based sparse regression module is designed and serves as a physical constraint to the autoencoder and coordinate system regression, and, in the meanwhile, uncover the parsimonious closed-form governing equations from the learned physical states. Experiments on simulated dynamical scenes show that the proposed method is able to distill closed-form governing equations and simultaneously identify unknown excitation input for several dynamical systems recorded by videos, which fills in the gap in literature where no existing methods are available and applicable for solving this type of problem.

We present an online demonstration tool illustrating a general approach to computing justifications for accepting a given decision when confronted with the preferences of several agents. Such a justification consists of a set of axioms providing a normative basis for the decision, together with a step-by-step explanation of how those axioms determine the decision. Our open-source implementation may also prove useful for realising other kinds of projects in computational social choice, particularly those requiring access to a SAT solver.

We study the recently introduced network untangling problem, a variant of Vertex Cover on temporal graphs---graphs whose edge set changes over discrete time steps. There are two versions of this problem. The goal is to select at most k time intervals for each vertex such that all time-edges are covered and (depending on the problem variant) either the maximum interval length or the total sum of interval lengths is minimized. This problem has data mining applications in finding activity timelines that explain the interactions of entities in complex networks. Both variants of the problem are NP-hard. In this paper, we initiate a multivariate complexity analysis involving the following parameters: number of vertices, lifetime of the temporal graph, number of intervals per vertex, and the interval length bound. For both problem versions, we (almost) completely settle the parameterized complexity for all combinations of those four parameters, thereby delineating the border of fixed-parameter tractability.

We address the efficiency problem of personalized ranking from implicit feedback by hashing users and items with binary codes, so that top-N recommendation can be fast executed in a Hamming space by bit operations. However, current hashing methods for top-N recommendation fail to align their learning objectives (such as pointwise or pairwise loss) with the benchmark metrics for ranking quality (e.g. Average Precision, AP), resulting in sub-optimal accuracy. To this end, we propose a Discrete Listwise Personalized Ranking (DLPR) model that optimizes AP under discrete constraints for fast and accurate top-N recommendation. To resolve the challenging DLPR problem, we devise an efficient algorithm that can directly learn binary codes in a relaxed continuous solution space. Specifically, theoretical analysis shows that the optimal solution to the relaxed continuous optimization problem is exactly the same as that of the original discrete DLPR problem. Through extensive experiments on two real-world datasets, we show that DLPR consistently surpasses state-of-the-art hashing methods for top-N recommendation.

We define and investigate new inconsistency measures that are particularly suitable for dealing with inconsistent spatio-temporal information, as they explicitly take into account the spatial and temporal dimensions, as well as the dimension concerning the identifiers of the monitored objects. Specifically, we first define natural measures that look at individual dimensions (time, space, and objects), and then propose measures based on the notion of a repair. We then analyze their behavior w.r.t. common postulates defined for classical propositional knowledge bases, and find that the latter are not suitable for spatio-temporal databases, in that the proposed inconsistency measures do not often satisfy them. In light of this, we argue that also postulates should explicitly take into account the spatial, temporal, and object dimensions, and thus define ``dimension-aware'' counterparts of common postulates, which are indeed often satisfied by the new inconsistency measures. Finally, we study the complexity of the proposed inconsistency measures.

We study a cooperative game setting where we want to gather more players through their social connections. Social connections can be modeled as a graph, and initially, only a subset of the players are in the game. We want to introduce diffusion incentives in such a cooperative game, i.e., incentivize the players to use their connections to invite more players to join the game. Our goal cannot be achieved by existing classical solutions, such as the Shapley value. Hence, to combat this problem, we have already proposed a solution called weighted permission Shapley value. Under this solution, for each player, inviting all of her neighbors is a dominant strategy in all monotone games. As one special application of the diffusion cooperative game, we also considered the diffusion incentives in query networks and the weighted permission Shapley value successfully characterizes the solution to the query network. Furthermore, we also characterize a Sybil-proof solution to the query network called the double geometric mechanism.