Efficiently fine-tuning pre-trained models for downstream tasks is a key challenge in the era of foundation models. Parameter-efficient fine-tuning (PEFT) presents a promising solution, achieving performance comparable to full fine-tuning by updating only a small number of adaptation weights per layer. Traditional PEFT methods typically rely on a single expert, where the adaptation weight is a low-rank matrix. However, for complex tasks, the data's inherent diversity poses a significant challenge for such models, as a single adaptation weight cannot adequately capture the features of all samples. To address this limitation, we explore how to integrate multiple small adaptation experts into a compact structure to defeat a large adapter. Specifically, we propose Tucker Adaptation (TuckA), a method with four key properties: (i) We use Tucker decomposition to create a compact 3D tensor where each slice naturally serves as an expert. The low-rank nature of this decomposition ensures that the number of parameters scales efficiently as more experts are added. (ii) We introduce a hierarchical strategy that organizes these experts into groups at different granularities, allowing the model to capture both local and global data patterns. (iii) We develop an efficient batch-level routing mechanism, which reduces the router's parameter size by a factor of L compared to routing at every adapted layer (where L is the number of adapted layers) (iv) We propose data-aware initialization to achieve loss-free expert load balancing based on theoretical analysis. Extensive experiments on benchmarks in natural language understanding, image classification, and mathematical reasoning speak to the efficacy of TuckA, offering a new and effective solution to the PEFT problem.

Generative diffusion models show promise for data augmentation. However, applying them to fine-grained tasks presents a significant challenge: ensuring synthetic images accurately capture the subtle, category-defining features critical for high fidelity. Standard approaches, such as text-based Classifier-Free Guidance (CFG), often lack the required specificity, potentially generating misleading examples that degrade fine-grained classifier performance. To address this, we propose Hierarchically Guided Fine-grained Augmentation (HiGFA). HiGFA leverages the temporal dynamics of the diffusion sampling process. It employs strong text and transformed contour guidance with fixed strengths in the early-to-mid sampling stages to establish overall scene, style, and structure. In the final sampling stages, HiGFA activates a specialized fine-grained classifier guidance and dynamically modulates the strength of all guidance signals based on prediction confidence. This hierarchical, confidence-driven orchestration enables HiGFA to generate diverse yet faithful synthetic images by intelligently balancing global structure formation with precise detail refinement. Experiments on several FGVC datasets demonstrate the effectiveness of HiGFA.

Recent advances in image-level self-supervised learning (SSL) have made significant progress, yet learning dense representations for patches remains challenging. Mainstream methods encounter an over-dispersion phenomenon that patches from the same instance/category scatter, harming downstream performance on dense tasks. This work reveals that image-level SSL avoids over-dispersion by involving implicit semantic concentration. Specifically, the non-strict spatial alignment ensures intra-instance consistency, while shared patterns, i.e., similar parts of within-class instances in the input space, ensure inter-image consistency. Unfortunately, these approaches are infeasible for dense SSL due to their spatial sensitivity and complicated scene-centric data. These observations motivate us to explore explicit semantic concentration for dense SSL. First, to break the strict spatial alignment, we propose to distill the patch correspondences. Facing noisy and imbalanced pseudo labels, we propose a noise-tolerant ranking loss. The core idea is extending the Average Precision (AP) loss to continuous targets, such that its decision-agnostic and adaptive focusing properties prevent the student model from being misled. Second, to discriminate the shared patterns from complicated scenes, we propose the object-aware filter to map the output space to an object-based space. Specifically, patches are represented by learnable prototypes of objects via cross-attention. Last but not least, empirical studies across various tasks soundly support the effectiveness of our method.

The Area Under the Receiver Operating Characteristics Curve (AUC) is a widely used metric for evaluating model performance across all possible decision thresholds. Existing methods for AUC optimization typically assume a predefined parametric distribution of thresholds. However, the optimal decision threshold depends on the misclassification costs, which follow a non-parametric distribution.This motivates us to introduce a variant of AUC, termed Cost-aware AUC (CAUC), where the thresholds are conditioned on an empirically determined cost distribution. Unfortunately, as a bilevel problem, it is challenging to directly optimize the CAUC: 1) The inner problem of finding the optimal thresholds is non-convex, leading to potential issues with convergence; 2) The outer problem involves the derivative of False Positive Rate (FPR) w.r.t. the threshold, which is unavailable without an explicit formulation of threshold distribution. To address challenge 1), we utilize the convex relaxation technique to reshape the inner problem into a convex one. Facing challenge 2), we propose an adaptive kernel density estimation framework. Specifically, the derivative of FPR is considered an aggregation of various kernel functions. To avoid manually crafting the aggregation function, we propose a finite-difference-based stochastic algorithm to optimize the model without explicit aggregation function. Theoretically, the proposed algorithm enjoys a convergence rate of $\mathcal {O}(\epsilon ^{-4})$O(ε-4). Empirical studies across various datasets and cost distributions speak to the effectiveness and soundness of our framework.

The past decade has witnessed notable achievements in self-supervised learning for video tasks. Recent efforts typically adopt the Masked Video Modeling (MVM) paradigm, leading to significant progress on multiple video tasks. However, two critical challenges remain: 1) Without human annotations, the random temporal sampling introduces uncertainty, increasing the difficulty of model training. 2) Previous MVM methods primarily recover the masked patches in the pixel space, leading to insufficient information compression for downstream tasks. To address these challenges jointly, we propose a self-supervised framework that leverages Temporal Correspondence for video Representation learning (T-CoRe). For challenge 1), we propose a sandwich sampling strategy that selects two auxiliary frames to reduce reconstruction uncertainty in a two-side-squeezing manner. Addressing challenge 2), we introduce an auxiliary branch into a self-distillation architecture to restore representations in the latent space, generating high-level semantic representations enriched with temporal information. Experiments of T-CoRe consistently present superior performance across several downstream tasks, demonstrating its effectiveness for video representation learning. The code is available at https://github.com/yafeng19/T-CORE.

In this work, we observe a counterintuitive phenomenon in self-supervised learning (SSL): longer training may impair the performance of dense prediction tasks (e.g., semantic segmentation). We refer to this phenomenon as Self-supervised Dense Degradation (SDD) and demonstrate its consistent presence across sixteen state-of-the-art SSL methods with various losses, architectures, and datasets. When the model performs suboptimally on dense tasks at the end of training, measuring the performance during training becomes essential. However, evaluating dense performance effectively without annotations remains an open challenge. To tackle this issue, we introduce a Dense representation Structure Estimator (DSE), composed of a class-relevance measure and an effective dimensionality measure. The proposed DSE is both theoretically grounded and empirically validated to be closely correlated with the downstream performance. Based on this metric, we introduce a straightforward yet effective model selection strategy and a DSE-based regularization method. Experiments on sixteen SSL methods across four benchmarks confirm that model selection improves mIoU by $3.0\\%$ on average with negligible computational cost. Additionally, DSE regularization consistently mitigates the effects of dense degradation. Code is available at \url{https://github.com/EldercatSAM/SSL-Degradation}.

The rapid growth of online video resources has significantly promoted the development of video retrieval methods. As a standard evaluation metric for video retrieval, Average Precision (AP) assesses the overall rankings of relevant videos at the top list, making the predicted scores a reliable reference for the users. However, recent video retrieval methods utilize pair-wise losses that treat all sample pairs equally, leading to an evident gap between the training objective and evaluation metric. To effectively bridge this gap, in this work, we aim to address two primary challenges: a) The current similarity measure and AP-based loss are suboptimal for video retrieval; b) The noticeable noise from frame-to-frame matching introduces ambiguity in estimating the AP loss. In response to these challenges, we propose the Hierarchical learning framework for Average-Precision-oriented Video Retrieval (HAP-VR). For the former challenge, we develop the TopK-Chamfer Similarity and QuadLinear-AP loss to measure and optimize video-level similarities in terms of AP. For the latter challenge, we suggest constraining the frame-level similarities to achieve an accurate AP loss estimation. Experimental results present that HAP-VR outperforms existing methods on several benchmark datasets, providing a feasible solution for video retrieval tasks and thus offering potential benefits for the multi-media application.

The Area Under the ROC Curve (AUC) is a well-known metric for evaluating instance-level long-tail learning problems. In the past two decades, many AUC optimization methods have been proposed to improve model performance under long-tail distributions. In this paper, we explore AUC optimization methods in the context of pixel-level long-tail semantic segmentation, a much more complicated scenario. This task introduces two major challenges for AUC optimization techniques. On one hand, AUC optimization in a pixel-level task involves complex coupling across loss terms, with structured inner-image and pairwise inter-image dependencies, complicating theoretical analysis. On the other hand, we find that mini-batch estimation of AUC loss in this case requires a larger batch size, resulting in an unaffordable space complexity. To address these issues, we develop a pixel-level AUC loss function and conduct a dependency-graph-based theoretical analysis of the algorithm's generalization ability. Additionally, we design a Tail-Classes Memory Bank (T-Memory Bank) to manage the significant memory demand. Finally, comprehensive experiments across various benchmarks confirm the effectiveness of our proposed AUCSeg method. The code is available at https://github.com/boyuh/AUCSeg.

Stochastic optimization of the Area Under the Precision-Recall Curve (AUPRC) is a crucial problem for machine learning. Despite extensive studies on AUPRC optimization, generalization is still an open problem. In this work, we present the first trial in the algorithm-dependent generalization of stochastic AUPRC optimization. The obstacles to our destination are three-fold. First, according to the consistency analysis, the majority of existing stochastic estimators are biased with biased sampling strategies. To address this issue, we propose a stochastic estimator with sampling-rate-invariant consistency and reduce the consistency error by estimating the full-batch scores with score memory. Second, standard techniques for algorithm-dependent generalization analysis cannot be directly applied to listwise losses. To fill this gap, we extend the model stability from instance-wise losses to listwise losses. Third, AUPRC optimization involves a compositional optimization problem, which brings complicated computations. In this work, we propose to reduce the computational complexity by matrix spectral decomposition. Based on these techniques, we derive the first algorithm-dependent generalization bound for AUPRC optimization. Motivated by theoretical results, we propose a generalization-induced learning framework, which improves the AUPRC generalization by equivalently increasing the batch size and the number of valid training examples. Practically, experiments on image retrieval and long-tailed classification speak to the effectiveness and soundness of our framework.

Area Under the ROC Curve (AUC) is a widely used ranking metric in imbalanced learning due to its insensitivity to label distributions. As a well-known multiclass extension of AUC, Multiclass AUC (MAUC, a.k.a. M-metric) measures the average AUC of multiple binary classifiers. In this paper, we argue that simply optimizing MAUC is far from enough for imbalanced multi-classification. More precisely, MAUC only focuses on learning scoring functions via ranking optimization, while leaving the decision process unconsidered. Therefore, scoring functions being able to make good decisions might suffer from low performance in terms of MAUC. To overcome this issue, we turn to explore AUCµ, another multiclass variant of AUC, which further takes the decision process into consideration. Motivated by this fact, we propose a surrogate risk optimization framework to improve model performance from the perspective of AUCµ. Practically, we propose a two-stage training framework for multi-classification, where at the first stage a scoring function is learned maximizing AUCµ, and at the second stage we seek for a decision function to improve the F1-metric via our proposed soft F1. Theoretically, we first provide sufficient conditions that optimizing the surrogate losses could lead to the Bayes optimal scoring function. Afterward, we show that the proposed surrogate risk enjoys a generalization bound in order of O(1/√N). Experimental results on four benchmark datasets demonstrate the effectiveness of our proposed method in both AUCµ and F1-metric.

In this paper, we focus on the mural-restoration task, which aims to detect damaged regions in the mural and repaint them automatically. Different from traditional image restoration tasks like in/out/blind-painting and image renovation, the corrupted mural suffers from more complicated degradation. However, existing mural-restoration methods and datasets still focus on simple degradation like masking. Such a significant gap prevents mural-restoration from being applied to real scenarios. To fill this gap, in this work, we propose a systematic framework to simulate the physical process for damaged murals and provide a new benchmark dataset for mural-restoration. Limited by the simplification of the data synthesis process, the previous mural-restoration methods suffer from poor performance in our proposed dataset. To handle this problem, we propose the Attention Diffusion Framework (ADF) for this challenging task. Within the framework, a damage attention map module is proposed to estimate the damage extent. Facing the diversity of defects, we propose a series of loss functions to choose repair strategies adaptively. Finally, experimental results support the effectiveness of the proposed framework in terms of both mural synthesis and restoration.

Long-tailed learning is one of the most challenging problems in visual recognition. There are some studies aiming to solve long-tailed classification from the perspective of feature learning. Recent work proposes to learn the balanced representation by fixing the linear classifier as Equiangular Tight Frame (ETF), since they argue what matters in classification is the structure of the feature, instead of their directions. Holding a different view, in this paper, we show that features with fixed directions may be harmful to the generalization of models, even if it is completely symmetric. To avoid this issue, we propose Representation-Balanced Learning Framework (RBL), which introduces orthogonal matrices to learn directions while maintaining the geometric structure of ETF. Theoretically, our contributions are two-fold: 1). we point out that the feature learning of RBL is insensitive toward training set label distribution, it always learns a balanced representation space. 2). we provide a generalization analysis of proposed RBL through training stability. To analyze the stability of the parameter with orthogonal constraint, we propose a novel training stability analysis paradigm, Two-Parameter Model Stability. Practically, our method is extremely simple in implementation but shows great superiority on several benchmark datasets.

In this paper, we aim to tackle flexible cost requirements for long-tail datasets, where we need to construct a (a) cost-sensitive and (b) class-distribution robust learning framework. The misclassification cost and the area under the ROC curve (AUC) are popular metrics for (a) and (b), respectively. However, limited by their formulations, models trained with AUC cannot be applied to cost-sensitive decision problems, and models trained with fixed costs are sensitive to the class distribution shift. To address this issue, we present a new setting where costs are treated like a dataset to deal with arbitrarily unknown cost distributions. Moreover, we propose a novel weighted version of AUC where the cost distribution can be integrated into its calculation through decision thresholds. To formulate this setting, we propose a novel bilevel paradigm to bridge weighted AUC (WAUC) and cost. The inner-level problem approximates the optimal threshold from sampling costs, and the outer-level problem minimizes the WAUC loss over the optimal threshold distribution. To optimize this bilevel paradigm, we employ a stochastic optimization algorithm (SACCL) to optimize it. Finally, experiment results show that our algorithm performs better than existing cost-sensitive learning methods and two-stage AUC decisions approach.

Positive-Unlabeled (PU) data arise frequently in a wide range of fields such as medical diagnosis, anomaly analysis and personalized advertising. The absence of any known negative labels makes it very challenging to learn binary classifiers from such data. Many state-of-the-art methods reformulate the original classification risk with individual risks over positive and unlabeled data, and explicitly minimize the risk of classifying unlabeled data as negative. This, however, usually leads to classifiers with a bias toward negative predictions, i.e., they tend to recognize most unlabeled data as negative. In this paper, we propose a label distribution alignment formulation for PU learning to alleviate this issue. Specifically, we align the distribution of predicted labels with the ground-truth, which is constant for a given class prior. In this way, the proportion of samples predicted as negative is explicitly controlled from a global perspective, and thus the bias toward negative predictions could be intrinsically eliminated. On top of this, we further introduce the idea of functional margins to enhance the model's discriminability, and derive a margin-based learning framework named Positive-Unlabeled learning with Label Distribution Alignment (PULDA). This framework is also combined with the class prior estimation process for practical scenarios, and theoretically supported by a generalization analysis. Moreover, a stochastic mini-batch optimization algorithm based on the exponential moving average strategy is tailored for this problem with a convergence guarantee. Finally, comprehensive empirical results demonstrate the effectiveness of the proposed method.

The Area Under the ROC curve (AUC) is a crucial metric for machine learning, which is often a reasonable choice for applications like disease prediction and fraud detection where the datasets often exhibit a long-tail nature. However, most of the existing AUC-oriented learning methods assume that the training data and test data are drawn from the same distribution. How to deal with domain shift remains widely open. This paper presents an early trial to attack AUC-oriented Unsupervised Domain Adaptation (UDA) (denoted as AUCUDA hence after). Specifically, we first construct a generalization bound that exploits a new distributional discrepancy for AUC. The critical challenge is that the AUC risk could not be expressed as a sum of independent loss terms, making the standard theoretical technique unavailable. We propose a new result that not only addresses the interdependency issue but also brings a much sharper bound with weaker assumptions about the loss function. Turning theory into practice, the original discrepancy requires complete annotations on the target domain, which is incompatible with UDA. To fix this issue, we propose a pseudo-labeling strategy and present an end-to-end training framework. Finally, empirical studies over five real-world datasets speak to the efficacy of our framework.

Positive-Unlabeled (PU) learning tries to learn binary classifiers from a few labeled positive examples with many unlabeled ones. Compared with ordinary semi-supervised learning, this task is much more challenging due to the ab-sence of any known negative labels. While existing cost-sensitive-based methods have achieved state-of-the-art per-formances, they explicitly minimize the risk of classifying unlabeled data as negative samples, which might result in a negative-prediction preference of the classifier. To allevi-ate this issue, we resort to a label distribution perspective for PU learning in this paper. Noticing that the label distribution of unlabeled data is fixed when the class prior is known, it can be naturally used as learning supervision for the model. Motivated by this, we propose to pursue the la-bel distribution consistency between predicted and ground-truth label distributions, which is formulated by aligning their expectations. Moreover, we further adopt the entropy minimization and Mixup regularization to avoid the trivial solution of the label distribution consistency on unlabeled data and mitigate the consequent confirmation bias. Exper-iments on three benchmark datasets validate the effective-ness of the proposed method.

The over-smoothing issue is a well-known challenge for Graph Convolutional Networks (GCN). Specifically, it is often observed that increasing the depth of GCN ends up in a trivial embedding subspace where the difference among node embeddings belonging to the same cluster tends to vanish. This paper believes that the main cause lies in the limited diversity along the message passing pipeline. Inspired by this, we propose a Sparse-Motif Ensemble Graph Convolutional Network (SMEGCN). We argue that merely employing the original graph Laplacian as the spectrum of the graph cannot capture the diversified local structure of complex graphs. Hence, to improve the diversity of the graph spectrum, we introduce local topological structures of complex graphs into GCN by employing the so-called graph motifs or the small network subgraphs. Moreover, we find that the motif connections are much denser than the edge connections, which might converge to an all-one matrix within a few times of message-passing. To fix this, we first propose the notion of sparse motif to avoid spurious motif connections. Subsequently, we propose a hierarchical motif aggregation mechanism to integrate the graph spectral information from a series of different sparse-motif message passing paths. Finally, we conduct a series of theoretical and experimental analyses to demonstrate the superiority of the proposed method.

Stochastic optimization of the Area Under the Precision-Recall Curve (AUPRC) is a crucial problem for machine learning. Although various algorithms have been extensively studied for AUPRC optimization, the generalization is only guaranteed in the multi-query case. In this work, we present the first trial in the single-query generalization of stochastic AUPRC optimization. For sharper generalization bounds, we focus on algorithm-dependent generalization. There are both algorithmic and theoretical obstacles to our destination. From an algorithmic perspective, we notice that the majority of existing stochastic estimators are biased when the sampling strategy is biased, and is leave-one-out unstable due to the non-decomposability. To address these issues, we propose a sampling-rate-invariant unbiased stochastic estimator with superior stability. On top of this, the AUPRC optimization is formulated as a composition optimization problem, and a stochastic algorithm is proposed to solve this problem. From a theoretical perspective, standard techniques of the algorithm-dependent generalization analysis cannot be directly applied to such a listwise compositional optimization problem. To fill this gap, we extend the model stability from instancewise losses to listwise losses and bridge the corresponding generalization and stability. Additionally, we construct state transition matrices to describe the recurrence of the stability, and simplify calculations by matrix spectrum. Practically, experimental results on three image retrieval datasets on speak to the effectiveness and soundness of our framework.

Weakly supervised semantic instance segmentation with only image-level supervision, instead of relying on expensive pixel-wise masks or bounding box annotations, is an important problem to alleviate the data-hungry nature of deep learning. In this article, we tackle this challenging problem by aggregating the image-level information of all training images into a large knowledge graph and exploiting semantic relationships from this graph. Specifically, our effort starts with some generic segment-based object proposals (SOP) without category priors. We propose a multiple instance learning (MIL) framework, which can be trained in an end-to-end manner using training images with image-level labels. For each proposal, this MIL framework can simultaneously compute probability distributions and category-aware semantic features, with which we can formulate a large undirected graph. The category of background is also included in this graph to remove the massive noisy object proposals. An optimal multi-way cut of this graph can thus assign a reliable category label to each proposal. The denoised SOP with assigned category labels can be viewed as pseudo instance segmentation of training images, which are used to train fully supervised models. The proposed approach achieves state-of-the-art performance for both weakly supervised instance segmentation and semantic segmentation. The code is available at https://github.com/yun-liu/LIID.

Nowadays, we have witnessed the early progress on learning the association between voice and face automatically, which brings a new wave of studies to the computer vision community. However, most of the prior arts along this line (a) merely adopt local information to perform modality alignment and (b) ignore the diversity of learning difficulty across different subjects. In this paper, we propose a novel framework to jointly address the above-mentioned issues. Targeting at (a), we propose a two-level modality alignment loss where both global and local information are considered. Compared with the existing methods, we introduce a global loss into the modality alignment process. The global component of the loss is driven by the identity classification. Theoretically, we show that minimizing the loss could maximize the distance between embeddings across different identities while minimizing the distance between embeddings belonging to the same identity, in a global sense (instead of a mini-batch). Targeting at (b), we propose a dynamic reweighting scheme to better explore the hard but valuable identities while filtering out the unlearnable identities. Experiments show that the proposed method outperforms the previous methods in multiple settings, including voice-face matching, verification and retrieval.

Multipartite ranking is a basic task in machine learning, where the Area Under the receiver operating characteristics Curve (AUC) is generally applied as the evaluation metric. Despite that AUC reflects the overall performance of the model, it is inconsistent with the expected performance in some application scenarios, where only a low False Positive Rate (FPR) is meaningful. To leverage high performance under low FPRs, we consider an alternative metric for multipartite ranking evaluating the True Positive Rate (TPR) at a given FPR, denoted as TPR@FPR. Unfortunately, the key challenge of direct TPR@FPR optimization is two-fold: \textbf{a)} the original objective function is not differentiable, making gradient backpropagation impossible; \textbf{b)} the loss function could not be written as a sum of independent instance-wise terms, making mini-batch based optimization infeasible. To address these issues, we propose a novel framework on top of the deep learning framework named \textit{Cross-Batch Approximation for Multipartite Ranking (CBA-MR)}. In face of \textbf{a)}, we propose a differentiable surrogate optimization problem where the instances having a short-time effect on FPR are rendered with different weights based on the random walk hypothesis. To tackle \textbf{b)}, we propose a fast ranking estimation method, where the full-batch loss evaluation is replaced by a delayed update scheme with the help of an embedding cache. Finally, experimental results on four real-world benchmarks are provided to demonstrate the effectiveness of the proposed method.

Though recent methods on semi-supervised video object segmentation (VOS) have achieved an appreciable improvement of segmentation accuracy, it is still hard to get an adequate speed-accuracy balance when facing real-world application scenarios. In this work, we propose Discriminative Matching for real-time Video Object Segmentation (DMVOS), a real-time VOS framework with high-accuracy to fill this gap. Based on the matching mechanism, our framework introduces discriminative information through the Isometric Correlation module and the Instance Center Offset module. Specifically, the isometric correlation module learns a pixel-level similarity map with semantic discriminability, and the instance center offset module is applied to exploit the instance-level spatial discriminability. Experiments on two benchmark datasets show that our model achieves state-of-the-art performance with extremely fast speed, for example, J&F of 87.8% on DAVIS-2016 validation set with 35 milliseconds per frame.