AAAI 2024 论文列表

Few-shot Class-incremental learning (FSCIL) is a challenging task in machine learning that aims to recognize new classes from a limited number of instances while preserving the ability to classify previously learned classes without retraining the entire model. This presents challenges in updating the model with new classes using limited training data, particularly in balancing acquiring new knowledge while retaining the old. We propose a novel method named Multiple Mxing Self-Distillation (M2SD) during the training phase to address these issues. Specifically, we propose a dual-branch structure that facilitates the expansion of the entire feature space to accommodate new classes. Furthermore, we introduce a feature enhancement component that can pass additional enhanced information back to the base network by self-distillation, resulting in improved classification performance upon adding new classes. After training, we discard both structures, leaving only the primary network to classify new class instances. Extensive experiments demonstrate that our approach achieves superior performance over previous state-of-the-art methods.

Document layout analysis is a crucial step for intelligent document understanding. However, many existing methods primarily focus on the visual aspects and overlook the textual features of documents. Although document pre-trained models utilize multi-modal features during the pre-training phase, they tend to operate as a unimodal pipeline when it comes to layout analysis tasks. Furthermore, current multi-modal methods perform worse than unimodal detectors on complex layout analysis datasets. To address these limitations, we propose an effective and pluggable multi-modal fusion approach named M2Doc, which fuses visual and textual features for better layout detection. M2Doc contains two pluggable multi-modal fusion modules, early-fusion and late-fusion, which align and fuse visual and textual features at the pixel level and block level. Benefitting from the concision and effectiveness of M2Doc, it can be easily applied to various detectors for better layout detection, including two-stage and end-to-end object detectors. Our experimental results demonstrate significant performance improvements in detectors equipped with M2Doc on datasets such as DocLayNet (+11.3 mAP) and M6Doc (+1.9 mAP). Furthermore, through the integration of the DINO detector with M2Doc, we achieve state-of-the-art results on DocLayNet (89.0 mAP), M6Doc (69.9 mAP), and PubLayNet (95.5 mAP). The code will be publicly released at https://github.com/johnning2333/M2Doc.

3D perception is a critical problem in autonomous driving. Recently, the Bird’s-Eye-View (BEV) approach has attracted extensive attention, due to low-cost deployment and desirable vision detection capacity. However, the existing models ignore a realistic scenario during the driving procedure, i.e., one or more view cameras may be failed, which largely deteriorates their performance. To tackle this problem, we propose a generic Masked BEV (M-BEV) perception framework, which can effectively improve robustness to this challenging scenario, by random masking and reconstructing camera views in the end-to-end training. More specifically, we develop a novel Masked View Reconstruction (MVR) module in our M-BEV. It mimics various missing cases by randomly masking features of different camera views, then leverages the original features of these views as self-supervision and reconstructs the masked ones with the distinct spatio-temporal context across camera views. Via such a plug-and-play MVR, our M-BEV is capable of learning the missing views from the resting ones, and thus well generalized for robust view recovery and accurate perception in the testing. We perform extensive experiments on the popular NuScenes benchmark, where our framework can significantly boost 3D perception performance of the state-of-the-art models on various missing view cases, e.g., for the absence of back view, our M-BEV promotes the PETRv2 model with 10.3% mAP gain.

Deep equilibrium (DEQ) models have emerged as a promising class of implicit layer models, which abandon traditional depth by solving for the fixed points of a single nonlinear layer. Despite their success, the stability of the fixed points for these models remains poorly understood. By considering DEQ models as nonlinear dynamic systems, we propose a robust DEQ model named LyaDEQ with guaranteed provable stability via Lyapunov theory. The crux of our method is ensuring the Lyapunov stability of the DEQ model's fixed points, which enables the proposed model to resist minor initial perturbations. To avoid poor adversarial defense due to Lyapunov-stable fixed points being located near each other, we orthogonalize the layers after the Lyapunov stability module to separate different fixed points. We evaluate LyaDEQ models under well-known adversarial attacks, and experimental results demonstrate significant improvement in robustness. Furthermore, we show that the LyaDEQ model can be combined with other defense methods, such as adversarial training, to achieve even better adversarial robustness.

Incomplete Multiple Kernel Clustering algorithms, which aim to learn a common latent representation from pre-constructed incomplete multiple kernels from the original data, followed by k-means for clustering. They have attracted intensive attention due to their high computational efficiency. However, our observation reveals that the imputation of these approaches for each kernel ignores the influence of other incomplete kernels. In light of this, we present a novel method called Low-Rank Kernel Tensor Learning for Incomplete Multiple Views Clustering (LRKT-IMVC) to address the above issue. Specifically, LRKT-IMVC first introduces the concept of kernel tensor to explore the inter-view correlations, and then the low-rank kernel tensor constraint is used to further capture the consistency information to impute missing kernel elements, thereby improving the quality of clustering. Moreover, we carefully design an alternative optimization method with promising convergence to solve the resulting optimization problem. The proposed method is compared with recent advances in experiments with different missing ratios on seven well-known datasets, demonstrating its effectiveness and the advantages of the proposed interpolation method.

Human face captured at night or in dimly lit environments has become a common practice, accompanied by complex low-light and low-resolution degradations. However, the existing face super-resolution (FSR) technologies and derived cascaded schemes are inadequate to recover credible textures. In this paper, we propose a novel approach that decomposes the restoration task into face structural fidelity maintaining and texture consistency learning. The former aims to enhance the quality of face images while improving the structural fidelity, while the latter focuses on eliminating perturbations and artifacts caused by low-light degradation and reconstruction. Based on this, we develop a novel low-light low-resolution face super-resolution framework. Our method consists of two steps: an illumination correction face super-resolution network (IC-FSRNet) for lighting the face and recovering the structural information, and a detail enhancement model (DENet) for improving facial details, thus making them more visually appealing and easier to analyze. As the relighted regions could provide complementary information to boost face super-resolution and vice versa, we introduce the mutual learning to harness the informative components from relighted regions and reconstruction, and achieve the iterative refinement. In addition, DENet equipped with diffusion probabilistic model is built to further improve face image visual quality. Experiments demonstrate that the proposed joint optimization framework achieves significant improvements in reconstruction quality and perceptual quality over existing two-stage sequential solutions. Code is available at https://github.com/wcy-cs/IC-FSRDENet.

With the rise of real-time rendering and the evolution of display devices, there is a growing demand for post-processing methods that offer high-resolution content in a high frame rate. Existing techniques often suffer from quality and latency issues due to the disjointed treatment of frame supersampling and extrapolation. In this paper, we recognize the shared context and mechanisms between frame supersampling and extrapolation, and present a novel framework, Space-time Supersampling (STSS). By integrating them into a unified framework, STSS can improve the overall quality with lower latency. To implement an efficient architecture, we treat the aliasing and warping holes unified as reshading regions and put forth two key components to compensate the regions, namely Random Reshading Masking (RRM) and Efficient Reshading Module (ERM). Extensive experiments demonstrate that our approach achieves superior visual fidelity compared to state-of-the-art (SOTA) methods. Notably, the performance is achieved within only 4ms, saving up to 75\% of time against the conventional two-stage pipeline that necessitates 17ms.

Motivated by recent work in computational social choice, we extend the metric distortion framework to clustering problems. Given a set of n agents located in an underlying metric space, our goal is to partition them into k clusters, optimizing some social cost objective. The metric space is defined by a distance function d between the agent locations. Information about d is available only implicitly via n rankings, through which each agent ranks all other agents in terms of their distance from her. Still, even though no cardinal information (i.e., the exact distance values) is available, we would like to evaluate clustering algorithms in terms of social cost objectives that are defined using d. This is done using the notion of distortion, which measures how far from optimality a clustering can be, taking into account all underlying metrics that are consistent with the ordinal information available. Unfortunately, the most important clustering objectives (e.g., those used in the well-known k-median and k-center problems) do not admit algorithms with finite distortion. To sidestep this disappointing fact, we follow two alternative approaches: We first explore whether resource augmentation can be beneficial. We consider algorithms that use more than k clusters but compare their social cost to that of the optimal k-clusterings. We show that using exponentially (in terms of k) many clusters, we can get low (constant or logarithmic) distortion for the k-center and k-median objectives. Interestingly, such an exponential blowup is shown to be necessary. More importantly, we explore whether limited cardinal information can be used to obtain better results. Somewhat surprisingly, for k-median and k-center, we show that a number of queries that is polynomial in k and only logarithmic in n (i.e., only sublinear in the number of agents for the most relevant scenarios in practice) is enough to get constant distortion.

Recently, instance contrastive learning achieves good results in unsupervised domain adaptation. It reduces the distances between positive samples and the anchor, increases the distances between negative samples and the anchor, and learns discriminative feature representations for target samples. However, most recent methods for identifying positive and negative samples are based on whether the pseudo-labels of samples and the pseudo-label of the anchor correspond to the same class. Due to the lack of target labels, many uncertain data are mistakenly labeled during the training process, and many low training potential data are also utilized. To address these problems, we propose Low Category Uncertainty and High Training Potential Instance Learning for Unsupervised Domain Adaptation (LUHP). We first propose a weight to measure the category uncertainty of the target sample. We can effectively filter the samples near the decision boundary through category uncertainty thresholds which are calculated by weights. Then we propose a new loss to focus on samples with high training potential. Finally, for anchors with low category uncertainty, we propose a sample reuse strategy to make the model more robust. We demonstrate the effectiveness of LUHP by showing the results of four datasets widely used in unsupervised domain adaptation.

We show a hardness result for the number of training domains required to achieve a small population error in the test domain. Although many domain generalization algorithms have been developed under various domain-invariance assumptions, there is significant evidence to indicate that out-of-distribution (o.o.d.) test accuracy of state-of-the-art o.o.d. algorithms is on par with empirical risk minimization and random guess on the domain generalization benchmarks such as DomainBed. In this work, we analyze its cause and attribute the lost domain generalization to the lack of training domains. We show that, in a minimax lower bound fashion, any learning algorithm that outputs a classifier with an ε excess error to the Bayes optimal classifier requires at least poly(1/ε) number of training domains, even though the number of training data sampled from each training domain is large. Experiments on the DomainBed benchmark demonstrate that o.o.d. test accuracy is monotonically increasing as the number of training domains increases. Our result sheds light on the intrinsic hardness of domain generalization and suggests benchmarking o.o.d. algorithms by the datasets with a sufficient number of training domains.

Safety is an indispensable requirement for applying reinforcement learning (RL) to real problems. Although there has been a surge of safe RL algorithms proposed in recent years, most existing work typically 1) relies on receiving numeric safety feedback; 2) does not guarantee safety during the learning process; 3) limits the problem to a priori known, deterministic transition dynamics; and/or 4) assume the existence of a known safe policy for any states. Addressing the issues mentioned above, we thus propose Long-term Binary-feedback Safe RL (LoBiSaRL), a safe RL algorithm for constrained Markov decision processes (CMDPs) with binary safety feedback and an unknown, stochastic state transition function. LoBiSaRL optimizes a policy to maximize rewards while guaranteeing long-term safety that an agent executes only safe state-action pairs throughout each episode with high probability. Specifically, LoBiSaRL models the binary safety function via a generalized linear model (GLM) and conservatively takes only a safe action at every time step while inferring its effect on future safety under proper assumptions. Our theoretical results show that LoBiSaRL guarantees the long-term safety constraint, with high probability. Finally, our empirical results demonstrate that our algorithm is safer than existing methods without significantly compromising performance in terms of reward.

This paper studies long-term fair machine learning which aims to mitigate group disparity over the long term in sequential decision-making systems. To define long-term fairness, we leverage the temporal causal graph and use the 1-Wasserstein distance between the interventional distributions of different demographic groups at a sufficiently large time step as the quantitative metric. Then, we propose a three-phase learning framework where the decision model is trained on high-fidelity data generated by a deep generative model. We formulate the optimization problem as a performative risk minimization and adopt the repeated gradient descent algorithm for learning. The empirical evaluation shows the efficacy of the proposed method using both synthetic and semi-synthetic datasets.

In partial label learning (PLL), each instance is associated with a set of candidate labels, among which only one is correct. The traditional PLL almost all implicitly assume that the distribution of the classes is balanced. However, in real-world applications, the distribution of the classes is imbalanced or long-tailed, leading to the long-tailed partial label learning problem. The previous methods solve this problem mainly by ameliorating the ability to learn in the tail classes, which will sacrifice the performance of the head classes. While keeping the performance of the head classes may degrade the performance of the tail classes. Therefore, in this paper, we construct two classifiers, i.e., a head classifier for keeping the performance of dominant classes and a tail classifier for improving the performance of the tail classes. Then, we propose a classifier weight estimation module to automatically estimate the shot belongingness (head class or tail class) of the samples and allocate the weights for the head classifier and tail classifier when making prediction. This cooperation improves the prediction ability for both the head classes and the tail classes. The experiments on the benchmarks demonstrate the proposed approach improves the accuracy of the SOTA methods by a substantial margin. Code and data are available at: https://github.com/pruirui/HTC-LTPLL.

Real-world data is extremely imbalanced and presents a long-tailed distribution, resulting in models biased towards classes with sufficient samples and performing poorly on rare classes. Recent methods propose to rebalance classes but they undertake the seesaw dilemma (what is increasing performance on tail classes may decrease that of head classes, and vice versa). In this paper, we argue that the seesaw dilemma is derived from the gradient imbalance of different classes, in which gradients of inappropriate classes are set to important for updating, thus prone to overcompensation or undercompensation on tail classes. To achieve ideal compensation, we formulate long-tailed recognition as a multi-objective optimization problem, which fairly respects the contributions of head and tail classes simultaneously. For efficiency, we propose a Gradient-Balancing Grouping (GBG) strategy to gather the classes with similar gradient directions, thus approximately making every update under a Pareto descent direction. Our GBG method drives classes with similar gradient directions to form a more representative gradient and provides ideal compensation to the tail classes. Moreover, we conduct extensive experiments on commonly used benchmarks in long-tailed learning and demonstrate the superiority of our method over existing SOTA methods. Our code is released at https://github.com/WickyLee1998/GBG_v1.

Video recognition systems are vulnerable to adversarial examples. Recent studies show that style transfer-based and patch-based unrestricted perturbations can effectively improve attack efficiency. These attacks, however, face two main challenges: 1) Adding large stylized perturbations to all pixels reduces the naturalness of the video and such perturbations can be easily detected. 2) Patch-based video attacks are not extensible to targeted attacks due to the limited search space of reinforcement learning that has been widely used in video attacks recently. In this paper, we focus on the video black-box setting and propose a novel attack framework named LogoStyleFool by adding a stylized logo to the clean video. We separate the attack into three stages: style reference selection, reinforcement-learning-based logo style transfer, and perturbation optimization. We solve the first challenge by scaling down the perturbation range to a regional logo, while the second challenge is addressed by complementing an optimization stage after reinforcement learning. Experimental results substantiate the overall superiority of LogoStyleFool over three state-of-the-art patch-based attacks in terms of attack performance and semantic preservation. Meanwhile, LogoStyleFool still maintains its performance against two existing patch-based defense methods. We believe that our research is beneficial in increasing the attention of the security community to such subregional style transfer attacks.

Log anomaly detection is a key component in the field of artificial intelligence for IT operations (AIOps). Considering log data of variant domains, retraining the whole network for unknown domains is inefficient in real industrial scenarios. However, previous deep models merely focused on extracting the semantics of log sequences in the same domain, leading to poor generalization on multi-domain logs. To alleviate this issue, we propose a unified Transformer-based framework for Log anomaly detection (LogFormer) to improve the generalization ability across different domains, where we establish a two-stage process including the pre-training and adapter-based tuning stage. Specifically, our model is first pre-trained on the source domain to obtain shared semantic knowledge of log data. Then, we transfer such knowledge to the target domain via shared parameters. Besides, the Log-Attention module is proposed to supplement the information ignored by the log-paring. The proposed method is evaluated on three public datasets and one real-world dataset. Experimental results on multiple benchmarks demonstrate the effectiveness of our LogFormer with fewer trainable parameters and lower training costs.

We introduce the algorithmic problem of finding a locally rainbow path of length l connecting two distinguished vertices s and t in a vertex-colored directed graph. Herein, a path is locally rainbow if between any two visits of equally colored vertices, the path traverses consecutively at leaset r differently colored vertices. This problem generalizes the well-known problem of finding a rainbow path. It finds natural applications whenever there are different types of resources that must be protected from overuse, such as crop sequence optimization or production process scheduling. We show that the problem is computationally intractable even if r=2 or if one looks for a locally rainbow among the shortest paths. On the positive side, if one looks for a path that takes only a short detour (i.e., it is slightly longer than the shortest path) and if r is small, the problem can be solved efficiently. Indeed, the running time of the respective algorithm is near-optimal unless the ETH fails.

In this paper, we address the nonrigid shape matching with outliers by a novel and effective pointwise map refinement method, termed Locality Preserving Refinement. For accurate pointwise conversion from a given functional map, our method formulates a two-step procedure. Firstly, starting with noisy point-to-point correspondences, we identify inliers by leveraging the neighborhood support, which yields a closed-form solution with linear time complexity. After obtained the reliable correspondences of inliers, we refine the pointwise correspondences for outliers using local linear embedding, which operates in an adaptive spectral similarity space to further eliminate the ambiguities that are difficult to handle in the functional space. By refining pointwise correspondences with local consistency thus embedding geometric constraints into functional spaces, our method achieves considerable improvement in accuracy with linearithmic time and space cost. Extensive experiments on public benchmarks demonstrate the superiority of our method over the state-of-the-art methods. Our code is publicly available at https://github.com/XiaYifan1999/LOPR.

This paper for the first time leverages multi-modal videos for weakly-supervised temporal video grounding. As labeling the video moment is labor-intensive and subjective, the weakly-supervised approaches have gained increasing attention in recent years. However, these approaches could inherently compromise performance due to inadequate supervision. Therefore, to tackle this challenge, we for the first time pay attention to exploiting complementary information extracted from multi-modal videos (e.g., RGB frames, optical flows), where richer supervision is naturally introduced in the weaklysupervised context. Our motivation is that by integrating different modalities of the videos, the model is learned from synergic supervision and thereby can attain superior generalization capability. However, addressing multiple modalities† would also inevitably introduce additional computational overhead, and might become inapplicable if a particular modality is inaccessible. To solve this issue, we adopt a novel route: building a multi-modal distillation algorithm to capitalize on the multi-modal knowledge as supervision for model training, while still being able to work with only the single modal input during inference. As such, we can utilize the benefits brought by the supplementary nature of multiple modalities, without compromising the applicability in practical scenarios. Specifically, we first propose a cross-modal mutual learning framework and train a sophisticated teacher model to learn collaboratively from the multi-modal videos. Then we identify two sorts of knowledge from the teacher model, i.e., temporal boundaries and semantic activation map. And we devise a local-global distillation algorithm to transfer this knowledge to a student model of single-modal input at both local and global levels. Extensive experiments on large-scale datasets demonstrate that our method achieves state-of-the-art performance with/without multi-modal inputs.