Songcan Chen

Aiyang Han, Chuanxing Geng, Songcan Chen

As an effective data augmentation method, Mixup synthesizes an extra amount of samples through linear interpolations. Despite its theoretical dependency on data properties, Mixup reportedly performs well as a regularizer and calibrator contributing reliable robustness and generalization to deep model training. In this paper, inspired by Universum Learning which uses out-of-class samples to assist the target tasks, we investigate Mixup from a largely under-explored perspective - the potential to generate in-domain samples that belong to none of the target classes, that is, universum. We find that in the framework of supervised contrastive learning, Mixup-induced universum can serve as surprisingly high-quality hard negatives, greatly relieving the need for large batch sizes in contrastive learning. With these findings, we propose Universum-inspired supervised Contrastive learning (UniCon), which incorporates Mixup strategy to generate Mixup-induced universum as universum negatives and pushes them apart from anchor samples of the target classes. We extend our method to the unsupervised setting, proposing Unsupervised Universum-inspired contrastive model (Un-Uni). Our approach not only improves Mixup with hard labels, but also innovates a novel measure to generate universum data. With a linear classifier on the learned representations, UniCon shows state-of-the-art performance on various datasets. Specially, UniCon achieves 81.7% top-1 accuracy on CIFAR-100, surpassing the state of art by a significant margin of 5.2% with a much smaller batch size, typically, 256 in UniCon vs. 1024 in SupCon using ResNet-50. Un-Uni also outperforms SOTA methods on CIFAR-100. The code of this paper is released on https://github.com/hannaiiyanggit/UniCon.

Zhongchen Ma, Lisha Li, Qirong Mao et al.

Contrastive learning (CL) has shown impressive advances in image representation learning in whichever supervised multi-class classification or unsupervised learning. However, these CL methods fail to be directly adapted to multi-label image classification due to the difficulty in defining the positive and negative instances to contrast a given anchor image in multi-label scenario, let the label missing one alone, implying that borrowing a commonly-used way from contrastive multi-class learning to define them will incur a lot of false negative instances unfavorable for learning. In this paper, with the introduction of a label correction mechanism to identify missing labels, we first elegantly generate positives and negatives for individual semantic labels of an anchor image, then define a unique contrastive loss for multi-label image classification with missing labels (CLML), the loss is able to accurately bring images close to their true positive images and false negative images, far away from their true negative images. Different from existing multi-label CL losses, CLML also preserves low-rank global and local label dependencies in the latent representation space where such dependencies have been shown to be helpful in dealing with missing labels. To the best of our knowledge, this is the first general multi-label CL loss in the missing-label scenario and thus can seamlessly be paired with those losses of any existing multi-label learning methods just via a single hyperparameter. The proposed strategy has been shown to improve the classification performance of the Resnet101 model by margins of 1.2%, 1.6%, and 1.3% respectively on three standard datasets, MSCOCO, VOC, and NUS-WIDE. Code is available at https://github.com/chuangua/ContrastiveLossMLML.

Jiaqiang Zhang, Senzhang Wang, Songcan Chen

Detecting abnormal nodes from attributed networks is of great importance in many real applications, such as financial fraud detection and cyber security. This task is challenging due to both the complex interactions between the anomalous nodes with other counterparts and their inconsistency in terms of attributes. This paper proposes a self-supervised learning framework that jointly optimizes a multi-view contrastive learning-based module and an attribute reconstruction-based module to more accurately detect anomalies on attributed networks. Specifically, two contrastive learning views are firstly established, which allow the model to better encode rich local and global information related to the abnormality. Motivated by the attribute consistency principle between neighboring nodes, a masked autoencoder-based reconstruction module is also introduced to identify the nodes which have large reconstruction errors, then are regarded as anomalies. Finally, the two complementary modules are integrated for more accurately detecting the anomalous nodes. Extensive experiments conducted on five benchmark datasets show our model outperforms current state-of-the-art models.

Yuyan Zhou, Dong Liang, Songcan Chen et al.

When taking images against strong light sources, the resulting images often contain heterogeneous flare artifacts. These artifacts can importantly affect image visual quality and downstream computer vision tasks. While collecting real data pairs of flare-corrupted/flare-free images for training flare removal models is challenging, current methods utilize the direct-add approach to synthesize data. However, these methods do not consider automatic exposure and tone mapping in image signal processing pipeline (ISP), leading to the limited generalization capability of deep models training using such data. Besides, existing methods struggle to handle multiple light sources due to the different sizes, shapes and illuminance of various light sources. In this paper, we propose a solution to improve the performance of lens flare removal by revisiting the ISP and remodeling the principle of automatic exposure in the synthesis pipeline and design a more reliable light sources recovery strategy. The new pipeline approaches realistic imaging by discriminating the local and global illumination through convex combination, avoiding global illumination shifting and local over-saturation. Our strategy for recovering multiple light sources convexly averages the input and output of the neural network based on illuminance levels, thereby avoiding the need for a hard threshold in identifying light sources. We also contribute a new flare removal testing dataset containing the flare-corrupted images captured by ten types of consumer electronics. The dataset facilitates the verification of the generalization capability of flare removal methods. Extensive experiments show that our solution can effectively improve the performance of lens flare removal and push the frontier toward more general situations.

Xinrui Wang, Wenhai Wan, Chuanxin Geng et al.

Learning binary classifiers from positive and unlabeled data (PUL) is vital in many real-world applications, especially when verifying negative examples is difficult. Despite the impressive empirical performance of recent PUL methods, challenges like accumulated errors and increased estimation bias persist due to the absence of negative labels. In this paper, we unveil an intriguing yet long-overlooked observation in PUL: \textit{resampling the positive data in each training iteration to ensure a balanced distribution between positive and unlabeled examples results in strong early-stage performance. Furthermore, predictive trends for positive and negative classes display distinctly different patterns.} Specifically, the scores (output probability) of unlabeled negative examples consistently decrease, while those of unlabeled positive examples show largely chaotic trends. Instead of focusing on classification within individual time frames, we innovatively adopt a holistic approach, interpreting the scores of each example as a temporal point process (TPP). This reformulates the core problem of PUL as recognizing trends in these scores. We then propose a novel TPP-inspired measure for trend detection and prove its asymptotic unbiasedness in predicting changes. Notably, our method accomplishes PUL without requiring additional parameter tuning or prior assumptions, offering an alternative perspective for tackling this problem. Extensive experiments verify the superiority of our method, particularly in a highly imbalanced real-world setting, where it achieves improvements of up to $11.3\%$ in key metrics. The code is available at \href{https://github.com/wxr99/HolisticPU}{https://github.com/wxr99/HolisticPU}.

Xiang Li, Xinrui Wang, Songcan Chen · pku

In Multi-Label Learning (MLL), it is extremely challenging to accurately annotate every appearing object due to expensive costs and limited knowledge. When facing such a challenge, a more practical and cheaper alternative should be Single Positive Multi-Label Learning (SPMLL), where only one positive label needs to be provided per sample. Existing SPMLL methods usually assume unknown labels as negatives, which inevitably introduces false negatives as noisy labels. More seriously, Binary Cross Entropy (BCE) loss is often used for training, which is notoriously not robust to noisy labels. To mitigate this issue, we customize an objective function for SPMLL by pushing only one pair of labels apart each time to prevent the domination of negative labels, which is the main culprit of fitting noisy labels in SPMLL. To further combat such noisy labels, we explore the high-rankness of label matrix, which can also push apart different labels. By directly extending from SPMLL to MLL with full labels, a unified loss applicable to both settings is derived. Experiments on real datasets demonstrate that the proposed loss not only performs more robustly to noisy labels for SPMLL but also works well for full labels. Besides, we empirically discover that high-rankness can mitigate the dramatic performance drop in SPMLL. Most surprisingly, even without any regularization or fine-tuned label correction, only adopting our loss defeats state-of-the-art SPMLL methods on CUB, a dataset that severely lacks labels.

Ling Li, Dong Liang, Yuanhang Gao et al.

Evaluating the performance of low-light image enhancement (LLE) is highly subjective, thus making integrating human preferences into image enhancement a necessity. Existing methods fail to consider this and present a series of potentially valid heuristic criteria for training enhancement models. In this paper, we propose a new paradigm, i.e., aesthetics-guided low-light image enhancement (ALL-E), which introduces aesthetic preferences to LLE and motivates training in a reinforcement learning framework with an aesthetic reward. Each pixel, functioning as an agent, refines itself by recursive actions, i.e., its corresponding adjustment curve is estimated sequentially. Extensive experiments show that integrating aesthetic assessment improves both subjective experience and objective evaluation. Our results on various benchmarks demonstrate the superiority of ALL-E over state-of-the-art methods.

Feihu Huang, Songcan Chen

Minimax optimization plays an important role in many machine learning tasks such as generative adversarial networks (GANs) and adversarial training. Although recently a wide variety of optimization methods have been proposed to solve the minimax problems, most of them ignore the distributed setting where the data is distributed on multiple workers. Meanwhile, the existing decentralized minimax optimization methods rely on the strictly assumptions such as (strongly) concavity and variational inequality conditions. In the paper, thus, we propose an efficient decentralized momentum-based gradient descent ascent (DM-GDA) method for the distributed nonconvex-PL minimax optimization, which is nonconvex in primal variable and is nonconcave in dual variable and satisfies the Polyak-Lojasiewicz (PL) condition. In particular, our DM-GDA method simultaneously uses the momentum-based techniques to update variables and estimate the stochastic gradients. Moreover, we provide a solid convergence analysis for our DM-GDA method, and prove that it obtains a near-optimal gradient complexity of $O(ε^{-3})$ for finding an $ε$-stationary solution of the nonconvex-PL stochastic minimax problems, which reaches the lower bound of nonconvex stochastic optimization. To the best of our knowledge, we first study the decentralized algorithm for Nonconvex-PL stochastic minimax optimization over a network.

Xin Li, Mingqiang Wei, Songcan Chen

Self-supervised learning is attracting wide attention in point cloud processing. However, it is still not well-solved to gain discriminative and transferable features of point clouds for efficient training on downstream tasks, due to their natural sparsity and irregularity. We propose PointSmile, a reconstruction-free self-supervised learning paradigm by maximizing curriculum mutual information (CMI) across the replicas of point cloud objects. From the perspective of how-and-what-to-learn, PointSmile is designed to imitate human curriculum learning, i.e., starting with an easy curriculum and gradually increasing the difficulty of that curriculum. To solve "how-to-learn", we introduce curriculum data augmentation (CDA) of point clouds. CDA encourages PointSmile to learn from easy samples to hard ones, such that the latent space can be dynamically affected to create better embeddings. To solve "what-to-learn", we propose to maximize both feature- and class-wise CMI, for better extracting discriminative features of point clouds. Unlike most of existing methods, PointSmile does not require a pretext task, nor does it require cross-modal data to yield rich latent representations. We demonstrate the effectiveness and robustness of PointSmile in downstream tasks including object classification and segmentation. Extensive results show that our PointSmile outperforms existing self-supervised methods, and compares favorably with popular fully-supervised methods on various standard architectures.

Weikai Li, Meng Cao, Songcan Chen

Unsupervised Source (data) Free domain adaptation (USFDA) aims to transfer knowledge from a well-trained source model to a related but unlabeled target domain. In such a scenario, all conventional adaptation methods that require source data fail. To combat this challenge, existing USFDAs turn to transfer knowledge by aligning the target feature to the latent distribution hidden in the source model. However, such information is naturally limited. Thus, the alignment in such a scenario is not only difficult but also insufficient, which degrades the target generalization performance. To relieve this dilemma in current USFDAs, we are motivated to explore a new perspective to boost their performance. For this purpose and gaining necessary insight, we look back upon the origin of the domain adaptation and first theoretically derive a new-brand target generalization error bound based on the model smoothness. Then, following the theoretical insight, a general and model-smoothness-guided Jacobian norm (JN) regularizer is designed and imposed on the target domain to mitigate this dilemma. Extensive experiments are conducted to validate its effectiveness. In its implementation, just with a few lines of codes added to the existing USFDAs, we achieve superior results on various benchmark datasets.

Enhao Zhang, Chuanxing Geng, Songcan Chen

Data augmentation for minority classes is an effective strategy for long-tailed recognition, thus developing a large number of methods. Although these methods all ensure the balance in sample quantity, the quality of the augmented samples is not always satisfactory for recognition, being prone to such problems as over-fitting, lack of diversity, semantic drift, etc. For these issues, we propose the Class-aware Universum Inspired Re-balance Learning(CaUIRL) for long-tailed recognition, which endows the Universum with class-aware ability to re-balance individual minority classes from both sample quantity and quality. In particular, we theoretically prove that the classifiers learned by CaUIRL are consistent with those learned under the balanced condition from a Bayesian perspective. In addition, we further develop a higher-order mixup approach, which can automatically generate class-aware Universum(CaU) data without resorting to any external data. Unlike the traditional Universum, such generated Universum additionally takes the domain similarity, class separability, and sample diversity into account. Extensive experiments on benchmark datasets demonstrate the surprising advantages of our method, especially the top1 accuracy in minority classes is improved by 1.9% 6% compared to the state-of-the-art method.

Zhongchen Ma, Songcan Chen

Similarity-based method gives rise to a new class of methods for multi-label learning and also achieves promising performance. In this paper, we generalize this method, resulting in a new framework for classification task. Specifically, we unite similarity-based learning and generalized linear models to achieve the best of both worlds. This allows us to capture interdependencies between classes and prevent from impairing performance of noisy classes. Each learned parameter of the model can reveal the contribution of one class to another, providing interpretability to some extent. Experiment results show the effectiveness of the proposed approach on multi-class and multi-label datasets

Chuanxing Geng, Aiyang Han, Songcan Chen

Consistency and complementarity are two key ingredients for boosting multi-view clustering (MVC). Recently with the introduction of popular contrastive learning, the consistency learning of views has been further enhanced in MVC, leading to promising performance. However, by contrast, the complementarity has not received sufficient attention except just in the feature facet, where the Hilbert Schmidt Independence Criterion term or the independent encoder-decoder network is usually adopted to capture view-specific information. This motivates us to reconsider the complementarity learning of views comprehensively from multiple facets including the feature-, view-label- and contrast- facets, while maintaining the view consistency. We empirically find that all the facets contribute to the complementarity learning, especially the view-label facet, which is usually neglected by existing methods. Based on this, a simple yet effective \underline{M}ultifacet \underline{C}omplementarity learning framework for \underline{M}ulti-\underline{V}iew \underline{C}lustering (MCMVC) is naturally developed, which fuses multifacet complementarity information, especially explicitly embedding the view-label information. To our best knowledge, it is the first time to use view-labels explicitly to guide the complementarity learning of views. Compared with the SOTA baselines, MCMVC achieves remarkable improvements, e.g., by average margins over $5.00\%$ and $7.00\%$ respectively in complete and incomplete MVC settings on Caltech101-20 in terms of three evaluation metrics.

Yunyun Wang, Weiwen Zheng, Songcan Chen

Previous unsupervised domain adaptation (UDA) methods aim to promote target learning via a single-directional knowledge transfer from label-rich source domain to unlabeled target domain, while its reverse adaption from target to source has not jointly been considered yet so far. In fact, in some real teaching practice, a teacher helps students learn while also gets promotion from students to some extent, which inspires us to explore a dual-directional knowledge transfer between domains, and thus propose a Dual-Correction Adaptation Network (DualCAN) in this paper. However, due to the asymmetrical label knowledge across domains, transfer from unlabeled target to labeled source poses a more difficult challenge than the common source-to-target counterpart. First, the target pseudo-labels predicted by source commonly involve noises due to model bias, hence in the reverse adaptation, they may hurt the source performance and bring a negative target-to-source transfer. Secondly, source domain usually contains innate noises, which will inevitably aggravate the target noises, leading to noise amplification across domains. To this end, we further introduce a Noise Identification and Correction (NIC) module to correct and recycle noises in both domains. To our best knowledge, this is the first naive attempt of dual-directional adaptation for noisy UDA, and naturally applicable to noise-free UDA. A theory justification is given to state the rationality of our intuition. Empirical results confirm the effectiveness of DualCAN with remarkable performance gains over state-of-the-arts, particularly for extreme noisy tasks (e.g., ~+ 15% on Pw->Pr and Pr->Rw of Office-Home).

Xiang Li, Songcan Chen

Label Distribution Learning (LDL) assigns soft labels, a.k.a. degrees, to a sample. In reality, it is always laborious to obtain complete degrees, giving birth to the Incomplete LDL (InLDL). However, InLDL often suffers from performance degeneration. To remedy it, existing methods need one or more explicit regularizations, leading to burdensome parameter tuning and extra computation. We argue that label distribution itself may provide useful prior, when used appropriately, the InLDL problem can be solved without any explicit regularization. In this paper, we offer a rational alternative to use such a prior. Our intuition is that large degrees are likely to get more concern, the small ones are easily overlooked, whereas the missing degrees are completely neglected in InLDL. To learn an accurate label distribution, it is crucial not to ignore the small observed degrees but to give them properly large weights, while gradually increasing the weights of the missing degrees. To this end, we first define a weighted empirical risk and derive upper bounds between the expected risk and the weighted empirical risk, which reveals in principle that weighting plays an implicit regularization role. Then, by using the prior of degrees, we design a weighted scheme and verify its effectiveness. To sum up, our model has four advantages, it is 1) model selection free, as no explicit regularization is imposed; 2) with closed form solution (sub-problem) and easy-to-implement (a few lines of codes); 3) with linear computational complexity in the number of samples, thus scalable to large datasets; 4) competitive with state-of-the-arts even without any explicit regularization.

Yunyun Wang, Yao Liu, Songcan Chen

Universal domain adaptation (UniDA) is a general unsupervised domain adaptation setting, which addresses both domain and label shifts in adaptation. Its main challenge lies in how to identify target samples in unshared or unknown classes. Previous methods commonly strive to depict sample "confidence" along with a threshold for rejecting unknowns, and align feature distributions of shared classes across domains. However, it is still hard to pre-specify a "confidence" criterion and threshold which are adaptive to various real tasks, and a mis-prediction of unknowns further incurs misalignment of features in shared classes. In this paper, we propose a new UniDA method with adaptive Unknown Authentication by Classifier Paradox (UACP), considering that samples with paradoxical predictions are probably unknowns belonging to none of the source classes. In UACP, a composite classifier is jointly designed with two types of predictors. That is, a multi-class (MC) predictor classifies samples to one of the multiple source classes, while a binary one-vs-all (OVA) predictor further verifies the prediction by MC predictor. Samples with verification failure or paradox are identified as unknowns. Further, instead of feature alignment for shared classes, implicit domain alignment is conducted in output space such that samples across domains share the same decision boundary, though with feature discrepancy. Empirical results validate UACP under both open-set and universal UDA settings.

Dan Li, Songcan Chen

Decision tree (DT) attracts persistent research attention due to its impressive empirical performance and interpretability in numerous applications. However, the growth of traditional yet widely-used univariate decision trees (UDTs) is quite time-consuming as they need to traverse all the features to find the splitting value with the maximal reduction of the impurity at each internal node. In this paper, we newly design a splitting criterion to speed up the growth. The criterion is induced from Geometric Mean Metric Learning (GMML) and then optimized under its diagonalized metric matrix constraint, consequently, a closed-form rank of feature discriminant abilities can at once be obtained and the top 1 feature at each node used to grow an intent DT (called as dGMML-DT, where d is an abbreviation for diagonalization). We evaluated the performance of the proposed methods and their corresponding ensembles on benchmark datasets. The experiment shows that dGMML-DT achieves comparable or better classification results more efficiently than the UDTs with 10x average speedup. Furthermore, dGMML-DT can straightforwardly be extended to its multivariable counterpart (dGMML-MDT) without needing laborious operations.

Jiayu Yao, Qingyuan Wu, Quan Feng et al.

Self-supervised learning (SSL), as a newly emerging unsupervised representation learning paradigm, generally follows a two-stage learning pipeline: 1) learning invariant and discriminative representations with auto-annotation pretext(s), then 2) transferring the representations to assist downstream task(s). Such two stages are usually implemented separately, making the learned representation learned agnostic to the downstream tasks. Currently, most works are devoted to exploring the first stage. Whereas, it is less studied on how to learn downstream tasks with limited labeled data using the already learned representations. Especially, it is crucial and challenging to selectively utilize the complementary representations from diverse pretexts for a downstream task. In this paper, we technically propose a novel solution by leveraging the attention mechanism to adaptively squeeze suitable representations for the tasks. Meanwhile, resorting to information theory, we theoretically prove that gathering representation from diverse pretexts is more effective than a single one. Extensive experiments validate that our scheme significantly exceeds current popular pretext-matching based methods in gathering knowledge and relieving negative transfer in downstream tasks.

Feihu Huang, Chunyu Xuan, Xinrui Wang et al.

Minimax optimization recently is widely applied in many machine learning tasks such as generative adversarial networks, robust learning and reinforcement learning. In the paper, we study a class of nonconvex-nonconcave minimax optimization with nonsmooth regularization, where the objective function is possibly nonconvex on primal variable $x$, and it is nonconcave and satisfies the Polyak-Lojasiewicz (PL) condition on dual variable $y$. Moreover, we propose a class of enhanced momentum-based gradient descent ascent methods (i.e., MSGDA and AdaMSGDA) to solve these stochastic nonconvex-PL minimax problems. In particular, our AdaMSGDA algorithm can use various adaptive learning rates in updating the variables $x$ and $y$ without relying on any specifical types. Theoretically, we prove that our methods have the best known sample complexity of $\tilde{O}(ε^{-3})$ only requiring one sample at each loop in finding an $ε$-stationary solution. Some numerical experiments on PL-game and Wasserstein-GAN demonstrate the efficiency of our proposed methods.

Junwei Zeng, Dong Liang, Sheng-Jun Huang et al.

Atmospheric turbulence significantly degrades long-range imaging by introducing geometric warping and exposure-time-dependent blur, which adversely affects both visual quality and the performance of high-level vision tasks. Existing methods for synthesizing turbulence effects often oversimplify the relationship between blur and exposure-time, typically assuming fixed or binary exposure settings. This leads to unrealistic synthetic data and limited generalization capability of trained models. To address this gap, we revisit the modulation transfer function (MTF) formulation and propose a novel Exposure-Time-dependent MTF (ET-MTF) that models blur as a continuous function of exposure-time. For blur synthesis, we derive a tilt-invariant point spread function (PSF) from the ET-MTF, which, when integrated with a spatially varying blur-width field, provides a comprehensive and physically accurate characterization of turbulence-induced blur. Building on this synthesis pipeline, we construct ET-Turb, a large-scale synthetic turbulence dataset that explicitly incorporates continuous exposure-time modeling across diverse optical and atmospheric conditions. The dataset comprises 5,083 videos (2,005,835 frames), partitioned into 3,988 training and 1,095 test videos. Extensive experiments demonstrate that models trained on ET-Turb produce more realistic restorations and achieve superior generalization on real-world turbulence data compared to those trained on other datasets. The dataset is publicly available at: github.com/Jun-Wei-Zeng/ET-Turb.

Feihu Huang, Xinrui Wang, Junyi Li et al.

Federated learning is a popular distributed and privacy-preserving learning paradigm in machine learning. Recently, some federated learning algorithms have been proposed to solve the distributed minimax problems. However, these federated minimax algorithms still suffer from high gradient or communication complexity. Meanwhile, few algorithm focuses on using adaptive learning rate to accelerate these algorithms. To fill this gap, in the paper, we study a class of nonconvex minimax optimization, and propose an efficient adaptive federated minimax optimization algorithm (i.e., AdaFGDA) to solve these distributed minimax problems. Specifically, our AdaFGDA builds on the momentum-based variance reduced and local-SGD techniques, and it can flexibly incorporate various adaptive learning rates by using the unified adaptive matrices. Theoretically, we provide a solid convergence analysis framework for our AdaFGDA algorithm under non-i.i.d. setting. Moreover, we prove our AdaFGDA algorithm obtains a lower gradient (i.e., stochastic first-order oracle, SFO) complexity of $\tilde{O}(ε^{-3})$ with lower communication complexity of $\tilde{O}(ε^{-2})$ in finding $ε$-stationary point of the nonconvex minimax problems. Experimentally, we conduct some experiments on the deep AUC maximization and robust neural network training tasks to verify efficiency of our algorithms.

Meng Cao, Jiexi Liu, Songcan Chen

Imbalanced Domain Generalization (IDG) focuses on mitigating both domain and label shifts, both of which fundamentally shape the model's decision boundaries, particularly under heterogeneous long-tailed distributions across domains. Despite its practical significance, it remains underexplored, primarily due to the technical complexity of handling their entanglement and the paucity of theoretical foundations. In this paper, we begin by theoretically establishing the generalization bound for IDG, highlighting the role of posterior discrepancy and decision margin. This bound motivates us to focus on directly steering decision boundaries, marking a clear departure from existing methods. Subsequently, we technically propose a novel Negative-Dominant Contrastive Learning (NDCL) for IDG to enhance discriminability while enforce posterior consistency across domains. Specifically, inter-class decision-boundary separation is enhanced by placing greater emphasis on negatives as the primary signal in our contrastive learning, naturally amplifying gradient signals for minority classes to avoid the decision boundary being biased toward majority classes. Meanwhile, intra-class compactness is encouraged through a re-weighted cross-entropy strategy, and posterior consistency across domains is enforced through a prediction-central alignment strategy. Finally, rigorous yet challenging experiments on benchmarks validate the effectiveness of our NDCL. The code is available at https://github.com/Alrash/NDCL.

Xinrui Wang, Chuanxing Geng, Wenhai Wan et al.

Online continual learning requires the models to learn from constant, endless streams of data. While significant efforts have been made in this field, most were focused on mitigating the catastrophic forgetting issue to achieve better classification ability, at the cost of a much heavier training workload. They overlooked that in real-world scenarios, e.g., in high-speed data stream environments, data do not pause to accommodate slow models. In this paper, we emphasize that model throughput -- defined as the maximum number of training samples that a model can process within a unit of time -- is equally important. It directly limits how much data a model can utilize and presents a challenging dilemma for current methods. With this understanding, we revisit key challenges in OCL from both empirical and theoretical perspectives, highlighting two critical issues beyond the well-documented catastrophic forgetting: Model's ignorance: the single-pass nature of OCL challenges models to learn effective features within constrained training time and storage capacity, leading to a trade-off between effective learning and model throughput; Model's myopia: the local learning nature of OCL on the current task leads the model to adopt overly simplified, task-specific features and excessively sparse classifier, resulting in the gap between the optimal solution for the current task and the global objective. To tackle these issues, we propose the Non-sparse Classifier Evolution framework (NsCE) to facilitate effective global discriminative feature learning with minimal time cost. NsCE integrates non-sparse maximum separation regularization and targeted experience replay techniques with the help of pre-trained models, enabling rapid acquisition of new globally discriminative features.

Feihu Huang, Yuning Luo, Songcan Chen

Matrix-structured parameters frequently appear in many artificial intelligence models such as large language models. More recently, an efficient Muon optimizer is designed for matrix parameters of large-scale models, and shows markedly faster convergence than the vector-wise algorithms. Although some works have begun to study convergence properties (i.e., optimization error) of the Muon optimizer, its generalization properties (i.e., generalization error) is still not established. Thus, in this paper, we study generalization error of the Muon optimizer based on algorithmic stability and mathematical induction, and prove that the Muon has a generalization error of $O\big(\frac{1}{Nκ^{T}}\big)$, where $N$ is training sample size, and $T$ denotes iteration number, and $κ>0$ denotes minimum difference between singular values of gradient estimate. To enhance generalization of the Muon, we propose an effective mixed Muon (MiMuon) optimizer by cautiously using orthogonalization of gradient, which is a hybrid of Muon and momentum-based SGD optimizers. Then we prove that our MiMuon optimizer has a lower generalization error of $O\big(\frac{1}{N}\big)$ than $O\big(\frac{1}{Nκ^{T}}\big)$ of Muon optimizer, since $κ$ generally is very small. Meanwhile, we also studied the convergence properties of our MiMuon algorithm, and prove that our MiMuon algorithm has the same convergence rate of $O(\frac{1}{T^{1/4}})$ as the Muon algorithm. Some numerical experimental results on training large models including Qwen3-0.6B and YOLO26m demonstrate efficiency of the MiMuon optimizer.

Enhao Zhang, Chaohua Li, Chuanxing Geng et al.

Neural Collapse (NC) presents an elegant geometric structure that enables individual activations (features), class means and classifier (weights) vectors to reach \textit{optimal} inter-class separability during the terminal phase of training on a \textit{balanced} dataset. Once shifted to imbalanced classification, such an optimal structure of NC can be readily destroyed by the notorious \textit{minority collapse}, where the classifier vectors corresponding to the minority classes are squeezed. In response, existing works endeavor to recover NC typically by optimizing classifiers. However, we discover that this squeezing phenomenon is not only confined to classifier vectors but also occurs with class means. Consequently, reconstructing NC solely at the classifier aspect may be futile, as the feature means remain compressed, leading to the violation of inherent \textit{self-duality} in NC (\textit{i.e.}, class means and classifier vectors converge mutually) and incidentally, resulting in an unsatisfactory collapse of individual activations towards the corresponding class means. To shake off these dilemmas, we present a unified \textbf{All}-around \textbf{N}eural \textbf{C}ollapse framework (AllNC), aiming to comprehensively restore NC across multiple aspects including individual activations, class means and classifier vectors. We thoroughly analyze its effectiveness and verify on multiple benchmark datasets that it achieves state-of-the-art in both balanced and imbalanced settings.

Dong Liang, Yue Sun, Yun Du et al.

Current knowledge distillation (KD) methods primarily focus on transferring various structured knowledge and designing corresponding optimization goals to encourage the student network to imitate the output of the teacher network. However, introducing too many additional optimization objectives may lead to unstable training, such as gradient conflicts. Moreover, these methods ignored the guidelines of relative learning difficulty between the teacher and student networks. Inspired by human cognitive science, in this paper, we redefine knowledge from a new perspective -- the student and teacher networks' relative difficulty of samples, and propose a pixel-level KD paradigm for semantic segmentation named Relative Difficulty Distillation (RDD). We propose a two-stage RDD framework: Teacher-Full Evaluated RDD (TFE-RDD) and Teacher-Student Evaluated RDD (TSE-RDD). RDD allows the teacher network to provide effective guidance on learning focus without additional optimization goals, thus avoiding adjusting learning weights for multiple losses. Extensive experimental evaluations using a general distillation loss function on popular datasets such as Cityscapes, CamVid, Pascal VOC, and ADE20k demonstrate the effectiveness of RDD against state-of-the-art KD methods. Additionally, our research showcases that RDD can integrate with existing KD methods to improve their upper performance bound.

Feihu Huang, Guanyi Zhang, Songcan Chen

Lion optimizer is a popular learning-based optimization algorithm in machine learning, which shows impressive performance in training many deep learning models. Although convergence property of the Lion optimizer has been studied, its generalization analysis is still missing. To fill this gap, we study generalization property of the Lion via algorithmic stability based on the mathematical induction. Specifically, we prove that the Lion has a generalization error of $O(\frac{1}{Nτ^T})$, where $N$ is training sample size, and $τ>0$ denotes the smallest absolute value of non-zero element in gradient estimator, and $T$ is the total iteration number. In addition, we obtain an interesting byproduct that the SignSGD algorithm has the same generalization error as the Lion. To enhance generalization of the Lion, we design a novel efficient Cautious Lion (i.e., CLion) optimizer by cautiously using sign function. Moreover, we prove that our CLion has a lower generalization error of $O(\frac{1}{N})$ than $O(\frac{1}{Nτ^T})$ of the Lion, since the parameter $τ$ generally is very small. Meanwhile, we study convergence property of our CLion optimizer, and prove that our CLion has a fast convergence rate of $O(\frac{\sqrt{d}}{T^{1/4}})$ under $\ell_1$-norm of gradient for nonconvex stochastic optimization, where $d$ denotes the model dimension. Extensive numerical experiments demonstrate effectiveness of our CLion optimizer.

Xinrui Wang, Shao-Yuan Li, Bartłomiej Twardowski et al.

Online Continual Learning (OCL) aims to learn from endless non\text{-}stationary data streams, yet most existing methods assume a flat label space and overlook the hierarchical organization of real\text{-}world concepts that evolves both horizontally (sibling classes) and vertically (coarse or fine categories). To better reflect this context, we introduce a new problem setting, DHOCL (Online Continual Learning from Dynamic Hierarchies), where taxonomies evolve across granularities and each sample provides supervision at a single hierarchical level. In this setting, we find two fundamental issues: (i) partial supervision under mixed granularities provides only point-wise signals over an evolving path-wise hierarchy, which constrains plasticity and undermines cross-level semantic consistency, and (ii) the dynamically evolving hierarchies induce granularity-dependent interference, destabilizing popular replay and regularization mechanisms and thereby exacerbating catastrophic forgetting. To tackle these issues, we propose HALO (Hierarchical Adaptive Learning with Organized Prototypes), which adaptively combines complementary classification heads, regularized by organized learnable hierarchical prototypes, enabling rapid adaptation, hierarchical consistency, and structured knowledge consolidation as the taxonomy evolves. Extensive experiments on multiple benchmarks demonstrate that HALO consistently outperforms existing methods across hierarchical accuracy, mistake severity, and continual performance.

Feihu Huang, Guanyi Zhang, Songcan Chen

Adam and AdamW are a class of default optimizers for training deep learning models in machine learning. These adaptive algorithms converge faster but generalize worse compared to SGD. In fact, their proved generalization error $O(\frac{1}{\sqrt{N}})$ also is larger than $O(\frac{1}{N})$ of SGD, where $N$ denotes training sample size. Recently, although some variants of Adam have been proposed to improve its generalization, their improved generalizations are still unexplored in theory. To fill this gap, in the paper, we restudy generalization of Adam and AdamW via algorithmic stability, and first prove that Adam and AdamW without square-root (i.e., Adam(W)-srf) have a generalization error $O(\frac{\hatρ^{-2T}}{N})$, where $T$ denotes iteration number and $\hatρ>0$ denotes the smallest element of second-order momentum plus a small positive number. To improve generalization, we propose a class of efficient clever Adam (i.e., HomeAdam(W)) algorithms via sometimes returning momentum-based SGD. Moreover, we prove that our HomeAdam(W) have a smaller generalization error $O(\frac{1}{N})$ than $O(\frac{\hatρ^{-2T}}{N})$ of Adam(W)-srf, since $\hatρ$ is generally very small. In particular, it is also smaller than the existing $O(\frac{1}{\sqrt{N}})$ of Adam(W). Meanwhile, we prove our HomeAdam(W) have a faster convergence rate of $O(\frac{1}{T^{1/4}})$ than $O(\frac{\breveρ^{-1}}{T^{1/4}})$ of the Adam(W)-srf, where $\breveρ\leq\hatρ$ also is very small. Extensive numerical experiments demonstrate efficiency of our HomeAdam(W) algorithms.

Jiaqiang Zhang, Songcan Chen

Graph contrastive learning (GCL) is an effective paradigm for node representation learning in graphs. The key components hidden behind GCL are data augmentation and positive-negative pair selection. Typical data augmentations in GCL, such as uniform deletion of edges, are generally blind and resort to local perturbation, which is prone to producing under-diversity views. Additionally, there is a risk of making the augmented data traverse to other classes. Moreover, most methods always treat all other samples as negatives. Such a negative pairing naturally results in sampling bias and likewise may make the learned representation suffer from semantic drift. Therefore, to increase the diversity of the contrastive view, we propose two simple and effective global topological augmentations to compensate current GCL. One is to mine the semantic correlation between nodes in the feature space. The other is to utilize the algebraic properties of the adjacency matrix to characterize the topology by eigen-decomposition. With the help of both, we can retain important edges to build a better view. To reduce the risk of semantic drift, a prototype-based negative pair selection is further designed which can filter false negative samples. Extensive experiments on various tasks demonstrate the advantages of the model compared to the state-of-the-art methods.

Lue Tao, Lei Feng, Hongxin Wei et al.

Adversarial training, originally designed to resist test-time adversarial examples, has shown to be promising in mitigating training-time availability attacks. This defense ability, however, is challenged in this paper. We identify a novel threat model named stability attack, which aims to hinder robust availability by slightly manipulating the training data. Under this threat, we show that adversarial training using a conventional defense budget $ε$ provably fails to provide test robustness in a simple statistical setting, where the non-robust features of the training data can be reinforced by $ε$-bounded perturbation. Further, we analyze the necessity of enlarging the defense budget to counter stability attacks. Finally, comprehensive experiments demonstrate that stability attacks are harmful on benchmark datasets, and thus the adaptive defense is necessary to maintain robustness. Our code is available at https://github.com/TLMichael/Hypocritical-Perturbation.

Jiexi Liu, Songcan Chen

Learning universal time series representations applicable to various types of downstream tasks is challenging but valuable in real applications. Recently, researchers have attempted to leverage the success of self-supervised contrastive learning (SSCL) in Computer Vision(CV) and Natural Language Processing(NLP) to tackle time series representation. Nevertheless, due to the special temporal characteristics, relying solely on empirical guidance from other domains may be ineffective for time series and difficult to adapt to multiple downstream tasks. To this end, we review three parts involved in SSCL including 1) designing augmentation methods for positive pairs, 2) constructing (hard) negative pairs, and 3) designing SSCL loss. For 1) and 2), we find that unsuitable positive and negative pair construction may introduce inappropriate inductive biases, which neither preserve temporal properties nor provide sufficient discriminative features. For 3), just exploring segment- or instance-level semantics information is not enough for learning universal representation. To remedy the above issues, we propose a novel self-supervised framework named TimesURL. Specifically, we first introduce a frequency-temporal-based augmentation to keep the temporal property unchanged. And then, we construct double Universums as a special kind of hard negative to guide better contrastive learning. Additionally, we introduce time reconstruction as a joint optimization objective with contrastive learning to capture both segment-level and instance-level information. As a result, TimesURL can learn high-quality universal representations and achieve state-of-the-art performance in 6 different downstream tasks, including short- and long-term forecasting, imputation, classification, anomaly detection and transfer learning.

Jiexi Liu, Meng Cao, Songcan Chen

Irregularly sampled time series (ISTS), characterized by non-uniform time intervals with natural missingness, are prevalent in real-world applications. Existing approaches for ISTS modeling primarily rely on observed values to impute unobserved ones or infer latent dynamics. However, these methods overlook a critical source of learning signal: the reconstruction error inherently produced during model training. Such error implicitly reflects how well a model captures the underlying data structure and can serve as an informative proxy for unobserved values. To exploit this insight, we propose iTimER, a simple yet effective self-supervised pre-training framework for ISTS representation learning. iTimER models the distribution of reconstruction errors over observed values and generates pseudo-observations for unobserved timestamps through a mixup strategy between sampled errors and the last available observations. This transforms unobserved timestamps into noise-aware training targets, enabling meaningful reconstruction signals. A Wasserstein metric aligns reconstruction error distributions between observed and pseudo-observed regions, while a contrastive learning objective enhances the discriminability of learned representations. Extensive experiments on classification, interpolation, and forecasting tasks demonstrate that iTimER consistently outperforms state-of-the-art methods under the ISTS setting.

Dong Liang, Zhengyan Xu, Ling Li et al.

In this paper, we propose a physics-inspired contrastive learning paradigm for low-light enhancement, called PIE. PIE primarily addresses three issues: (i) To resolve the problem of existing learning-based methods often training a LLE model with strict pixel-correspondence image pairs, we eliminate the need for pixel-correspondence paired training data and instead train with unpaired images. (ii) To address the disregard for negative samples and the inadequacy of their generation in existing methods, we incorporate physics-inspired contrastive learning for LLE and design the Bag of Curves (BoC) method to generate more reasonable negative samples that closely adhere to the underlying physical imaging principle. (iii) To overcome the reliance on semantic ground truths in existing methods, we propose an unsupervised regional segmentation module, ensuring regional brightness consistency while eliminating the dependency on semantic ground truths. Overall, the proposed PIE can effectively learn from unpaired positive/negative samples and smoothly realize non-semantic regional enhancement, which is clearly different from existing LLE efforts. Besides the novel architecture of PIE, we explore the gain of PIE on downstream tasks such as semantic segmentation and face detection. Training on readily available open data and extensive experiments demonstrate that our method surpasses the state-of-the-art LLE models over six independent cross-scenes datasets. PIE runs fast with reasonable GFLOPs in test time, making it easy to use on mobile devices.

Chaohua Li, Enhao Zhang, Chuanxing Geng et al.

Out-of-distribution detection (OOD) is a pivotal task for real-world applications that trains models to identify samples that are distributionally different from the in-distribution (ID) data during testing. Recent advances in AI, particularly Vision-Language Models (VLMs) like CLIP, have revolutionized OOD detection by shifting from traditional unimodal image detectors to multimodal image-text detectors. This shift has inspired extensive research; however, existing categorization schemes (e.g., few- or zero-shot types) still rely solely on the availability of ID images, adhering to a unimodal paradigm. To better align with CLIP's cross-modal nature, we propose a new categorization framework rooted in both image and text modalities. Specifically, we categorize existing methods based on how visual and textual information of OOD data is utilized within image + text modalities, and further divide them into four groups: OOD Images (i.e., outliers) Seen or Unseen, and OOD Texts (i.e., learnable vectors or class names) Known or Unknown, across two training strategies (i.e., train-free or training-required). More importantly, we discuss open problems in CLIP-like OOD detection and highlight promising directions for future research, including cross-domain integration, practical applications, and theoretical understanding.

Jiexi Liu, Meng Cao, Songcan Chen

Irregularly sampled multivariate time series (ISMTS) are prevalent in reality. Due to their non-uniform intervals between successive observations and varying sampling rates among series, the channel-independent (CI) strategy, which has been demonstrated more desirable for complete multivariate time series forecasting in recent studies, has failed. This failure can be further attributed to the sampling sparsity, which provides insufficient information for effective CI learning, thereby reducing its capacity. When we resort to the channel-dependent (CD) strategy, even higher capacity cannot mitigate the potential loss of diversity in learning similar embedding patterns across different channels. We find that existing work considers CI and CD strategies to be mutually exclusive, primarily because they apply these strategies to the global channel. However, we hold the view that channel strategies do not necessarily have to be used globally. Instead, by appropriately applying them locally and globally, we can create an opportunity to take full advantage of both strategies. This leads us to introduce the Channel Harmony ISMTS Transformer (TimeCHEAT), which utilizes the CD locally and the CI globally. Specifically, we segment the ISMTS into sub-series level patches. Locally, the CD strategy aggregates information within each patch for time embedding learning, maximizing the use of relevant observations while reducing long-range irrelevant interference. Here, we enhance generality by transforming embedding learning into an edge weight prediction task using bipartite graphs, eliminating the need for special prior knowledge. Globally, the CI strategy is applied across patches, allowing the Transformer to learn individualized attention patterns for each channel. Experimental results indicate our proposed TimeCHEAT demonstrates competitive SOTA performance across three mainstream tasks.

Chaohua Li, Enhao Zhang, Chuanxing Geng et al.

In open-set recognition (OSR), a promising strategy is exploiting pseudo-unknown data outside given $K$ known classes as an additional $K$+$1$-th class to explicitly model potential open space. However, treating unknown classes without distinction is unequal for them relative to known classes due to the category-agnostic and scale-agnostic of the unknowns. This inevitably not only disrupts the inherent distributions of unknown classes but also incurs both class-wise and instance-wise imbalances between known and unknown classes. Ideally, the OSR problem should model the whole class space as $K$+$\infty$, but enumerating all unknowns is impractical. Since the core of OSR is to effectively model the boundaries of known classes, this means just focusing on the unknowns nearing the boundaries of targeted known classes seems sufficient. Thus, as a compromise, we convert the open classes from infinite to $K$, with a novel concept Target-Aware Universum (TAU) and propose a simple yet effective framework Dual Contrastive Learning with Target-Aware Universum (DCTAU). In details, guided by the targeted known classes, TAU automatically expands the unknown classes from the previous $1$ to $K$, effectively alleviating the distribution disruption and the imbalance issues mentioned above. Then, a novel Dual Contrastive (DC) loss is designed, where all instances irrespective of known or TAU are considered as positives to contrast with their respective negatives. Experimental results indicate DCTAU sets a new state-of-the-art.

Feihu Huang, Yuning Luo, Songcan Chen

Large models recently are widely applied in artificial intelligence, so efficient training of large models has received widespread attention. More recently, a useful Muon optimizer is specifically designed for matrix-structured parameters of large models. Although some works have begun to studying Muon optimizer, the existing Muon and its variants still suffer from high sample complexity or high memory for large models. To fill this gap, we propose a light and fast Muon (LiMuon) optimizer for training large models, which builds on the momentum-based variance reduced technique and randomized Singular Value Decomposition (SVD). Our LiMuon optimizer has a lower memory than the current Muon and its variants. Moreover, we prove that our LiMuon has a lower sample complexity of $O(ε^{-3})$ for finding an $ε$-stationary solution of non-convex stochastic optimization under the smooth condition. Recently, the existing convergence analysis of Muon optimizer mainly relies on the strict Lipschitz smooth assumption, while some artificial intelligence tasks such as training large language models (LLMs) do not satisfy this condition. We also proved that our LiMuon optimizer has a sample complexity of $O(ε^{-3})$ under the generalized smooth condition. Numerical experimental results on training DistilGPT2 and ViT models verify efficiency of our LiMuon optimizer.

Chuanxing Geng, Qifei Li, Xinrui Wang et al.

Using unlabeled wild data containing both in-distribution (ID) and out-of-distribution (OOD) data to improve the safety and reliability of models has recently received increasing attention. Existing methods either design customized losses for labeled ID and unlabeled wild data then perform joint optimization, or first filter out OOD data from the latter then learn an OOD detector. While achieving varying degrees of success, two potential issues remain: (i) Labeled ID data typically dominates the learning of models, inevitably making models tend to fit OOD data as IDs; (ii) The selection of thresholds for identifying OOD data in unlabeled wild data usually faces dilemma due to the unavailability of pure OOD samples. To address these issues, we propose a novel loss-difference OOD detection framework (LoD) by \textit{intentionally label-noisifying} unlabeled wild data. Such operations not only enable labeled ID data and OOD data in unlabeled wild data to jointly dominate the models' learning but also ensure the distinguishability of the losses between ID and OOD samples in unlabeled wild data, allowing the classic clustering technique (e.g., K-means) to filter these OOD samples without requiring thresholds any longer. We also provide theoretical foundation for LoD's viability, and extensive experiments verify its superiority.

Haifeng Yang, Chuanxing Geng, Pong C. Yuen et al.

In open-set recognition, existing methods generally learn statically fixed decision boundaries using known classes to reject unknown classes. Though they have achieved promising results, such decision boundaries are evidently insufficient for universal unknown classes in dynamic and open scenarios as they can potentially appear at any position in the feature space. Moreover, these methods just simply reject unknown class samples during testing without any effective utilization for them. In fact, such samples completely can constitute the true instantiated representation of the unknown classes to further enhance the model's performance. To address these issues, this paper proposes a novel dynamic against dynamic idea, i.e., dynamic method against dynamic changing open-set world, where an open-set self-learning (OSSL) framework is correspondingly developed. OSSL starts with a good closed-set classifier trained by known classes and utilizes available test samples for model adaptation during testing, thus gaining the adaptability to changing data distributions. In particular, a novel self-matching module is designed for OSSL, which can achieve the adaptation in automatically identifying known class samples while rejecting unknown class samples which are further utilized to enhance the discriminability of the model as the instantiated representation of unknown classes. Our method establishes new performance milestones respectively in almost all standard and cross-data benchmarks.

Jiexi Liu, Meng Cao, Songcan Chen

Irregularly sampled multivariate time series (ISMTS) are prevalent in reality. Most existing methods treat ISMTS as synchronized regularly sampled time series with missing values, neglecting that the irregularities are primarily attributed to variations in sampling rates. In this paper, we introduce a novel perspective that irregularity is essentially relative in some senses. With sampling rates artificially determined from low to high, an irregularly sampled time series can be transformed into a hierarchical set of relatively regular time series from coarse to fine. We observe that additional coarse-grained relatively regular series not only mitigate the irregularly sampled challenges to some extent but also incorporate broad-view temporal information, thereby serving as a valuable asset for representation learning. Therefore, following the philosophy of learning that Seeing the big picture first, then delving into the details, we present the Multi-Scale and Multi-Correlation Attention Network (MuSiCNet) combining multiple scales to iteratively refine the ISMTS representation. Specifically, within each scale, we explore time attention and frequency correlation matrices to aggregate intra- and inter-series information, naturally enhancing the representation quality with richer and more intrinsic details. While across adjacent scales, we employ a representation rectification method containing contrastive learning and reconstruction results adjustment to further improve representation consistency. MuSiCNet is an ISMTS analysis framework that competitive with SOTA in three mainstream tasks consistently, including classification, interpolation, and forecasting.

Yunyun Wang, Zheng Duan, Xinyue Liao et al.

Open-Set Domain Generalization (OSDG) tackles the realistic scenario where deployed models encounter both domain shifts and novel object categories. Despite impressive progress with vision-language models like CLIP, existing methods still fall into the dilemma between structural risk of known-classes and open-space risk from unknown-classes, and easily suffers from over-confidence, especially when distinguishing ``hard unknowns" that share fine-grained visual similarities with known classes. To this end, we propose a Semantic-enhanced CLIP (SeeCLIP) framework that explicitly addresses this dilemma through fine-grained semantic enhancement. In SeeCLIP, we propose a semantic-aware prompt enhancement module to decompose images into discriminative semantic tokens, enabling nuanced vision-language alignment beyond coarse category labels. To position unknown prompts effectively, we introduce duplex contrastive learning with complementary objectives, that is, repulsion to maintain separability from known classes, and cohesion to preserve semantic proximity. Further, our semantic-guided diffusion module synthesizes pseudo-unknowns by perturbing extracted semantic tokens, generating challenging samples that are visually similar to known classes yet exhibit key local differences. These hard negatives force the model to learn finer decision boundaries. Extensive experiments across five benchmarks demonstrate consistent improvements of 3% accuracy and 5% H-score over state-of-the-art methods.

Jia Tang, Xinrui Wang, Songcan Chen

Continual learning (CL) involves acquiring and accumulating knowledge from evolving tasks while alleviating catastrophic forgetting. Recently, leveraging contrastive loss to construct more transferable and less forgetful representations has been a promising direction in CL. Despite advancements, their performance is still limited due to confusion arising from both inter-task and intra-task features. To address the problem, we propose a simple yet effective contrastive strategy named \textbf{G}lobal \textbf{P}re-fixing, \textbf{L}ocal \textbf{A}djusting for \textbf{S}upervised \textbf{C}ontrastive learning (GPLASC). Specifically, to avoid task-level confusion, we divide the entire unit hypersphere of representations into non-overlapping regions, with the centers of the regions forming an inter-task pre-fixed \textbf{E}quiangular \textbf{T}ight \textbf{F}rame (ETF). Meanwhile, for individual tasks, our method helps regulate the feature structure and form intra-task adjustable ETFs within their respective allocated regions. As a result, our method \textit{simultaneously} ensures discriminative feature structures both between tasks and within tasks and can be seamlessly integrated into any existing contrastive continual learning framework. Extensive experiments validate its effectiveness.

Enhao Zhang, Chaohua Li, Chuanxing Geng et al.

Based on the success of large-scale visual foundation models like CLIP in various downstream tasks, this paper initially attempts to explore their impact on Long-Tailed Semi-Supervised Learning (LTSSL) by employing the foundation model with three strategies: Linear Probing (LP), Lightweight Fine-Tuning (LFT), and Full Fine-Tuning (FFT). Our analysis presents the following insights: i) Compared to LTSSL algorithms trained from scratch, FFT results in a decline in model performance, whereas LP and LFT, although boosting overall model performance, exhibit negligible benefits to tail classes. ii) LP produces numerous false pseudo-labels due to \textit{underlearned} training data, while LFT can reduce the number of these false labels but becomes overconfident about them owing to \textit{biased fitting} training data. This exacerbates the pseudo-labeled and classifier biases inherent in LTSSL, limiting performance improvement in the tail classes. With these insights, we propose a Unbiased Lightweight Fine-tuning strategy, \textbf{ULFine}, which mitigates the overconfidence via confidence-aware adaptive fitting of textual prototypes and counteracts the pseudo-labeled and classifier biases via complementary fusion of dual logits. Extensive experiments demonstrate that ULFine markedly decreases training costs by over ten times and substantially increases prediction accuracies compared to state-of-the-art methods.

Xinrui Wang, Chuanxing Geng, Wenhai Wan et al.

Recent studies have verified that semi-supervised learning (SSL) is vulnerable to data poisoning backdoor attacks. Even a tiny fraction of contaminated training data is sufficient for adversaries to manipulate up to 90\% of the test outputs in existing SSL methods. Given the emerging threat of backdoor attacks designed for SSL, this work aims to protect SSL against such risks, marking it as one of the few known efforts in this area. Specifically, we begin by identifying that the spurious correlations between the backdoor triggers and the target class implanted by adversaries are the primary cause of manipulated model predictions during the test phase. To disrupt these correlations, we utilize three key techniques: Gaussian Filter, complementary learning and trigger mix-up, which collectively filter, obstruct and dilute the influence of backdoor attacks in both data pre-processing and feature learning. Experimental results demonstrate that our proposed method, Backdoor Invalidator (BI), significantly reduces the average attack success rate from 84.7\% to 1.8\% across different state-of-the-art backdoor attacks. It is also worth mentioning that BI does not sacrifice accuracy on clean data and is supported by a theoretical guarantee of its generalization capability.

Meng Cao, Songcan Chen

Domain generalization addresses domain shift in real-world applications. Most approaches adopt a domain angle, seeking invariant representation across domains by aligning their marginal distributions, irrespective of individual classes, naturally leading to insufficient exploration of discriminative information. Switching to a class angle, we find that multiple domain-related peaks or clusters within the same individual classes must emerge due to distribution shift. In other words, marginal alignment does not guarantee conditional alignment, leading to suboptimal generalization. Therefore, we argue that acquiring discriminative generalization between classes within domains is crucial. In contrast to seeking distribution alignment, we endeavor to safeguard domain-related between-class discrimination. To this end, we devise a novel Conjugate Consistent Enhanced Module, namely Con2EM, based on a distribution over domains, i.e., a meta-distribution. Specifically, we employ a novel distribution-level Universum strategy to generate supplementary diverse domain-related class-conditional distributions, thereby enhancing generalization. This allows us to resample from these generated distributions to provide feedback to the primordial instance-level classifier, further improving its adaptability to the target-agnostic. To ensure generation accuracy, we establish an additional distribution-level classifier to regularize these conditional distributions. Extensive experiments have been conducted to demonstrate its effectiveness and low computational cost compared to SOTAs.

Xiang Li, Jiexi Liu, Xinrui Wang et al.

In reality, data often exhibit associations with multiple labels, making multi-label learning (MLL) become a prominent research topic. The last two decades have witnessed the success of MLL, which is indispensable from complete and accurate supervised information. However, obtaining such information in practice is always laborious and sometimes even impossible. To circumvent this dilemma, incomplete multi-label learning (InMLL) has emerged, aiming to learn from incomplete labeled data. To date, enormous InMLL works have been proposed to narrow the performance gap with complete MLL, whereas a systematic review for InMLL is still absent. In this paper, we not only attempt to fill the lacuna but also strive to pave the way for innovative research. Specifically, we retrospect the origin of InMLL, analyze the challenges of InMLL, and make a taxonomy of InMLL from the data-oriented and algorithm-oriented perspectives, respectively. Besides, we also present real applications of InMLL in various domains. More importantly, we highlight several potential future trends, including four open problems that are more in line with practice and three under-explored/unexplored techniques in addressing the challenges of InMLL, which may shed new light on developing novel research directions in the field of InMLL.

Wenhai Wan, Xinrui Wang, Ming-Kun Xie et al.

Learning from noisy data has attracted much attention, where most methods focus on closed-set label noise. However, a more common scenario in the real world is the presence of both open-set and closed-set noise. Existing methods typically identify and handle these two types of label noise separately by designing a specific strategy for each type. However, in many real-world scenarios, it would be challenging to identify open-set examples, especially when the dataset has been severely corrupted. Unlike the previous works, we explore how models behave when faced with open-set examples, and find that \emph{a part of open-set examples gradually get integrated into certain known classes}, which is beneficial for the separation among known classes. Motivated by the phenomenon, we propose a novel two-step contrastive learning method CECL (Class Expansion Contrastive Learning) which aims to deal with both types of label noise by exploiting the useful information of open-set examples. Specifically, we incorporate some open-set examples into closed-set classes to enhance performance while treating others as delimiters to improve representative ability. Extensive experiments on synthetic and real-world datasets with diverse label noise demonstrate the effectiveness of CECL.

Quan Feng, Songcan Chen

Multi-task learning is to improve the performance of the model by transferring and exploiting common knowledge among tasks. Existing MTL works mainly focus on the scenario where label sets among multiple tasks (MTs) are usually the same, thus they can be utilized for learning across the tasks. While almost rare works explore the scenario where each task only has a small amount of training samples, and their label sets are just partially overlapped or even not. Learning such MTs is more challenging because of less correlation information available among these tasks. For this, we propose a framework to learn these tasks by jointly leveraging both abundant information from a learnt auxiliary big task with sufficiently many classes to cover those of all these tasks and the information shared among those partially-overlapped tasks. In our implementation of using the same neural network architecture of the learnt auxiliary task to learn individual tasks, the key idea is to utilize available label information to adaptively prune the hidden layer neurons of the auxiliary network to construct corresponding network for each task, while accompanying a joint learning across individual tasks. Our experimental results demonstrate its effectiveness in comparison with the state-of-the-art approaches.

Weikai Li, Songcan Chen

Unsupervised domain adaptation (UDA) aims to transfer knowledge from a well-labeled source domain to a different but related unlabeled target domain with identical label space. Currently, the main workhorse for solving UDA is domain alignment, which has proven successful. However, it is often difficult to find an appropriate source domain with identical label space. A more practical scenario is so-called partial domain adaptation (PDA) in which the source label set or space subsumes the target one. Unfortunately, in PDA, due to the existence of the irrelevant categories in the source domain, it is quite hard to obtain a perfect alignment, thus resulting in mode collapse and negative transfer. Although several efforts have been made by down-weighting the irrelevant source categories, the strategies used tend to be burdensome and risky since exactly which irrelevant categories are unknown. These challenges motivate us to find a relatively simpler alternative to solve PDA. To achieve this, we first provide a thorough theoretical analysis, which illustrates that the target risk is bounded by both model smoothness and between-domain discrepancy. Considering the difficulty of perfect alignment in solving PDA, we turn to focus on the model smoothness while discard the riskier domain alignment to enhance the adaptability of the model. Specifically, we instantiate the model smoothness as a quite simple intra-domain structure preserving (IDSP). To our best knowledge, this is the first naive attempt to address the PDA without domain alignment. Finally, our empirical results on multiple benchmark datasets demonstrate that IDSP is not only superior to the PDA SOTAs by a significant margin on some benchmarks (e.g., +10% on Cl->Rw and +8% on Ar->Rw ), but also complementary to domain alignment in the standard UDA