RGBT object tracking takes advantage of the complementary properties of RGB and thermal infrared (TIR) modalities. However, many existing methods focus on fusion within a single domain, either spatial or frequency, without fully exploiting the complementarity of multiple domains. This limits the interaction between domains and makes trackers less robust under severe conditions. To improve feature representation by leveraging both spatial and frequency information, we propose a Frequency domain information-Aided Spatial domain Feature Fusion framework for RGBT tracking (FASFF), which mainly consists of the Frequency domain Attention Enhancement Module (FAEM) and the Spatial-Frequency feature interaction Fusion Module (SFFM). The FAEM employs the Discrete Cosine Transform Attention block (DCTAttn), which highlights informative frequency components while integrating spatial and frequency domains information, and the Cross-modality Collaborative Attention block (CCAttn), which suppresses noise via cross-modal feature refinement to enhance intra-modality feature representation. The SFFM further harnesses the interaction between the spatial and frequency domains to collaboratively refine feature representation, leveraging complementary information from the frequency domain to enhance spatial domain representations. We evaluate the FASFF on three benchmark datasets: RGBT210, RGBT234, and LasHeR. The experimental results show that the FASFF maintains stable and reliable performance across different tracking scenarios, verifying its effectiveness for RGBT object tracking.

Recent progress in studying treelike committee machines (TCM) neural networks (NN) in Stojnic (2023c,g) showed that the Random Duality Theory (RDT) and its a partially lifted(pl RDT) variant are powerful tools that can be used for very precise networks capacity analysis. The initial considerations from Stojnic (2023c,g) regarding sign activations were extended to more general activations in Stojnic (2024), where particularly elegant results were obtained for any even number of quadratically and ReLU activated hidden layer neurons, d. While the results of Stojnic (2024) are applicable to any activation type in principle, a significant amount of numerical work is often required to make them practically usable. In this work, we examine wide hidden layer networks and uncover that certain aspects of such difficulties miraculously disappear. Specifically, we employ recently developed fully lifted (fl) RDT to characterize the capacity of wide (d → ∞) TCM nets. We obtain explicit, closed-form capacity characterizations for a generic class of hidden layer activations. While the utilized approach substantially reduces the necessary numerical evaluations, the ultimate success still requires a significant amount of residual numerical analysis. To determine the concrete capacity values, we analyzed four prominent activation examples: ReLU, quadratic, erf, and tanh. After successfully conducting all the residual numerical work for all of them, we have found that the lifting mechanism exhibits remarkably rapid convergence. Relative improvements do not exceed 0.01%-0.1% by the fourth level of lifting. As a convenient bonus, we also discover that the capacity characterizations obtained at the first and second lifting levels precisely match those derived through statistical physics replica theory methods in Zavatone-Veth and Pehlevan (2021) for generic activations and in Baldassi et al. (2019) for ReLU activations.

With the wide application of deep neural networks, the research on adversarial robustness has become one of hotspots, especially the issue of fairness in adversarial robustness has garnered the wide attention. Significant disparities in adversarial robustness across different classes could degrade the overall performance, because the effectiveness of the defense mechanism is constrained by the class with the lowest robustness called the worst class, that is the phenomenon known as the "wooden barrel effect". Most of the existing robust fairness algorithms were designed based on specific experimental metrics, neglecting analyzing the intrinsic causes of the robust fairness. And the trade-off of robust contribution between easy and hard classes was not considered. In addition, the robust balance between hard classes and their similar classes was ignored too. The above issues actually limited the further improvement of robust fairness. Therefore, A Frequency-Aware Adversarial Robust Fairness Algorithm (FAFA for short), which explores the inherent causes of fairness issues from frequency perspective, is proposed in the paper. And a Frequency-Aware Fair Adversarial Training method called FAAT based on the contribution of frequency components is introduced, and different loss weights and perturbation constraints are assigned to various classes to enhance the robustness of hard classes. Additionally, a Frequency-Aware Fair Fine-tuning Algorithm (FAFT) which redistributes features between hard classes and their similar classes is implemented, achieving a more balanced robustness distribution and improve the performance of the worst class. Extensive experiments were conducted on the CIFAR-10, CIFAR-100, and STL-10 datasets, and the results demonstrate that our proposed method significantly enhances the robustness of the worst class, the accuracy of the worst class reaches 34.76% using the ResNet18 model on the CIFAR-10 dataset under A3 attack, with an average improvement of approximately 5% compared to existing methods. Meanwhile, the overall robustness reaches 48.47%, and the comprehensive metric ρrob based on TRADES reaches 0.55.

Class imbalance presents a persistent bottleneck in machine learning, biasing models toward majority classes and degrading performance on rare events. Active learning partially alleviates this by focusing annotation efforts on underrepresented samples. However, this issue becomes more pronounced in multi-domain learning settings, where both intra-domain imbalances and shifts across domains complicate model training. Current multi-domain active learning techniques typically neglect the interaction of these dual imbalances, resulting in biased selection strategies and limited ability to generalize across domains. To tackle this, we introduce a comprehensive framework featuring a dynamic category-scaling weight mechanism that hierarchically addresses imbalance at the sample and domain levels. This mechanism assesses the relative difficulty of each class to guide effective sampling, while concurrently constructing a progressively refined "Ideal Domain" through the adaptive integration of well-performing domains. This ideal representation further supports a minimax optimization process to allocate annotation budgets both fairly and robustly. Extensive evaluations on various standard datasets confirm that our approach consistently surpasses leading methods in managing complex imbalanced scenarios across multiple domains.

This paper investigates the fixed-/preassigned-time synchronization control problem for delayed memristive neural networks (DMNNs) with external disturbance and discontinuous activation functions. Considering the inherent irregularity and aperiodicity in neural network scheduling, an aperiodic strict intermittent control strategy is adopted. Based on this framework, a class of aperiodic strict intermittent controllers is designed to achieve fixed-time synchronization (FxTS) for DMNNs. By appropriately tuning the controller parameters, a corresponding preassigned-time synchronization (PaTS) scheme is further constructed. Moreover, by fixing the control period and the active duration within each period, the proposed scheme reduces to periodic strict intermittent control and can be extended to MNNs with continuous activation functions. Furthermore, sufficient conditions are derived to guarantee both FxTS and PaTS, and explicit estimates of the settling time are provided. Finally, two numerical examples are presented to demonstrate the effectiveness of the proposed approach.

Policy gradient methods such as Proximal Policy Optimization (PPO) constrain policy updates only on sampled actions, leaving the unsampled action space entirely unconstrained-an issue we term Anchoring Blindness. This limitation induces uncontrolled drift in the policy distribution over unsampled regions, undermining training stability and often leading to suboptimal performance, particularly in continuous action spaces. To address this issue, we propose Anchored Policy Optimization (APO), a PPO variant equipped with Unsampled Action Ratios Regularization (UARR). UARR explicitly constrains the probability ratios between the new and old policies in unsampled action regions, preventing excessive deviation from 1.0. Experimental results on continuous-control tasks demonstrate that APO effectively anchors a broader range of action distributions, significantly improving optimization stability and avoiding convergence to suboptimal solutions. The code is available at https://github.com/wjl-bupt/APO.

Parameter-efficient fine-tuning (PEFT) is a key technique for adapting large pre-trained language models to downstream tasks with minimal parameter updates. However, existing PEFT methods often suffer from slow convergence, gradient noise, and weak alignment between learned features and task semantics. To this end, we develop a low-rank adaptation framework based on gradient spectral alignment. Specifically, we (1) perform spectral decomposition of the pre-trained gradient covariance matrix and apply eigenvalue scaling to retain 95% of the spectral energy, optimizing parameter initialization; (2) design a dual gradient projector that combines orthogonal and spherical projections to suppress gradient noise, reducing its variance to 38.2% of standard LoRA; and (3) introduce a dynamic eigenvalue scaling mechanism that adaptively recalibrates principal component weights via a nonlinear scaling function, enhancing representational capacity. Experiments on the GLUE benchmark show that our method achieves performance comparable to full fine-tuning while training only 0.1% of parameters, outperforming AdaLoRA and DoRA by 6.2% and 5.3%, respectively. Additionally, our approach improves the mathematical reasoning accuracy of LLaMA-7B on GSM8K by 7.8%. This work provides novel insights into improving the efficiency, stability, and adaptability of PEFT frameworks. The code is available at https://github.com/rainylover/GSLoRA.

Adverse weather image restoration aims to recover clean background scenes from images degraded by various weather conditions, such as haze, rain, and snow. With the rapid development of deep learning, single-task restoration methods targeting specific weather types have achieved remarkable progress and attracted increasing attention in recent years. More recently, to address the limited generalization of task-specific models, All-in-One (AiO) methods have emerged to handle multiple degradations within a unified framework. However, existing surveys mostly focus on individual degradation types or specific restoration paradigms, and unified reviews of deep learning-based adverse weather restoration are still limited. In this paper, we present a comprehensive survey that jointly organizes single-task and AiO restoration models from the perspectives of network architectures and learning paradigms. We further review widely used datasets, loss functions, and evaluation metrics across different restoration tasks. In addition, we summarize benchmark results of representative methods on public datasets to analyze their performance and generalization ability. Finally, we discuss key challenges and promising research directions to support future developments in this rapidly evolving field.

Universal Adversarial Perturbations (UAPs) differ from traditional image-specific perturbations in that they deceive target models across diverse instances using a single perturbation. Prior research has primarily focused on enhancing the transferability of non-targeted UAPs; however, these efforts fail to generate transferable UAPs capable of classifying images into a specific target class. To address this limitation, we propose a generative adversarial framework named CITP for generating cross-instance targeted perturbations. CITP leverages shared features among instances of the same class to produce perturbations that can be transferred to other instances within that class. The framework distinguishes between generated adversarial samples and images of the target class, enabling it to learn the label distribution of the target class. Additionally, CITP integrates a mid-level feature discriminator to improve the transferability of perturbations across different model architectures. Experimental results demonstrate that CITP exhibits exceptional transferability in cross-instance targeted attacks and achieves strong performance against four defense mechanisms. Notably, CITP extends beyond image data, enabling precise targeted attacks on video data as well.

Attention improves representation learning over RNNs, but its discrete nature limits continuous-time (CT) modeling. We introduce Neuronal Attention Circuit (NAC), a novel, biologically inspired CT-attention mechanism that reformulates attention logit computation as the solution to a linear first-order ODE with nonlinear interlinked gates derived from repurposing the wiring of C. elegans Neuronal Circuit Policies (NCPs). NAC replaces dense projections with sparse sensory gates for query-key projections and introduces a sparse backbone network with two heads for computing content-target and learnable time-constant gates, enabling efficient adaptive dynamics. To improve efficiency and memory consumption, we implement an adaptable, sparse, subquadratic Top-K pairwise concatenation mechanism that selectively curates query-key interactions. We provide rigorous theoretical guarantees, including state stability and bounded approximation errors. Empirically, we implement NAC in diverse domains, including irregular time-series, long-range forecasting, lane-keeping for autonomous vehicles, and industrial prognostics. We observe that NAC matches or outperforms in accuracy against several CT state-of-the-art baselines, while being interpretable at the neuron cell level.

Decentralized Federated Learning (DFL) has emerged as a key paradigm for collaborative model training across distributed edge devices while preserving data privacy and system autonomy. However, the performance of existing DFL frameworks is constrained by synchronization delays, high communication overhead, and poor scalability in heterogeneous environments. Conventional methods often rely on fixed-precision quantization and uniform synchronization, which overlook device diversity and data heterogeneity, leading to slower convergence and inefficient resource utilization. To address these limitations, this paper presents the Decentralized Adaptive Quantization-Aware Synchronous (DAQSyn) framework, a unified resource-aware learning strategy that integrates adaptive quantization, synchronization, and performance-weighted aggregation. In DAQSyn, each device autonomously adjusts its model precision according to its computational and communication capabilities, generating lightweight models that enhance convergence speed while preserving performance. The synchronization ensures efficient coordination among devices through a BSP-based barrier that harmonizes update timing, reducing idle waiting time by approximately 19% compared to FP16 and over 25% compared to low-bit quantization. The performance-weighted aggregation mechanism further enhances model stability by prioritizing high-quality local updates, leading to an accuracy improvement of 1% and a 2.5 ×  faster convergence relative to fixed-precision approaches. Combined, these mechanisms enable lightweight models with reduced communication cost and robust learning performance across heterogeneous devices and non-IID data settings.

Current vision-language multimodal long-term tracking methods are highly dependent on large-scale data training and complex cross-modal models. This not only leads to substantial computational overhead but also restricts their deployment and application in resource-constrained scenarios. To resolve this contradiction, we propose a training-free, fast-response post-processing tracking framework named VL-METrack, which dynamically fuses motion estimation and localization networks. This framework transforms the long-term tracking task from a data-driven model training problem into a heuristic decision-driven online adaptive decision-making problem. Specifically, we first adopt a general pre-trained network to provide initial target localization and confidence assessment. When the confidence indicates tracking ambiguity, instead of relying on deeper networks, our method leverages a collaborative processing mechanism consisting of optical flow and correlation filters. This mechanism dynamically fuses optical flow and the localization network through confidence based adaptive weighting, while incorporating an error correction mechanism that judges the rate of change in target orientation. Subsequently, we employ a dynamic template update strategy and establish a global re-detection mechanism, maintaining template timeliness through an adaptive update process and determining the optimal position via multi-template comparison within an expanded search region. Experimental results on the OTB, LaSOT, and TNL2K datasets demonstrate that our method achieves performance comparable to state-of-the-art approaches while maintaining low computational complexity and reduced resource consumption. The code is available at: https://github.com/yll1214/VL-METrack.

Unsupervised Person Re-Identification (ReID) is fundamentally challenged by a self-perpetuating vicious cycle, where camera-induced feature biases lead to unreliable pseudo-labels, which in turn corrupt the representation learning process. To break this cycle, this paper introduces the Synergistic Camera-aware Disentanglement and Consistency (SCC) optimization framework. Instead of tackling feature learning and label refinement in isolation, SCC establishes a virtuous cycle of mutual reinforcement between them. The framework begins by learning camera-robust representations through a camera-aware disentanglement mechanism, which mitigates camera bias by dynamically constructing disentangled prototypes. This process yields a more reliable feature space, providing a solid foundation for pseudo-labeling. Subsequently, a neighbor-center consistency optimization strategy leverages both local sample affinities and global cluster structures to significantly enhance the quality of these pseudo-labels. Critically, these refined labels serve as higher-quality supervision, guiding the representation learning in the next iteration to become even more discriminative. Through this synergistic co-optimization, SCC iteratively enhances both feature quality and label reliability. Extensive experiments on challenging datasets, including Market-1501 and MSMT17, demonstrate that our method significantly outperforms state-of-the-art approaches. Furthermore, quantitative analyses utilizing a camera-probe linear test, pseudo-label quality trajectories, and an early-vs-late stability study confirm that camera-invariant features and refined pseudo-labels reinforce each other across training, validating the effectiveness of our synergistic learning paradigm.

Robust tensor decomposition (RTD) is designed to distinguish low-rank and sparse tensors from noisy high-dimensional data, which holds fundamental significance in the fields of machine learning and computer vision. Nevertheless, current RTD-based methods fall short in addressing the issues of automatic noise adaptation and determination of model capacity. In response to these challenges, this paper introduces a robust non-parametric Bayesian method known as the Bayesian adaptive tensor ring decomposition (BATR) method. More precisely, BATR models unknown noise using a Dirichlet process Gaussian mixture model (DP-GMM), with automatic determination of the noise components. Besides, BATR accomplishes an adaptive low tensor ring (TR) rank model by incorporating a more advanced generalized hyperbolic (GH) prior into the probabilistic framework, thereby facilitating automatic TR rank determination. Furthermore, a variational Bayesian inference algorithm is employed to update the posteriors of the model. Extensive experiments on synthetic data, color images, face images, multispectral images, and hyperspectral images demonstrate the improved performance of BATR compared to other state-of-the-art methods.

Facial recognition (FR) models are vulnerable to adversarial attacks, in which attackers manipulate facial images to expose system vulnerabilities, underscoring the urgent need to improve the transferability of adversarial attacks. However, existing methods fail to fully leverage diverse initialization strategies for extending surrogate models, thereby limiting the transferability of adversarial samples. To address this, we propose the Multi-Initialization Enhanced Aggregation (MEA) attack method. This approach improves transferability by combining different parameter initializations to generate a diversified set of surrogate models. MEA consists of two stages: Multi-Initialization Adversarial Enhancement (MIAE) and Enhanced Adversarial Aggregation (EAA). In the MIAE stage, we enhance model diversity through checkpoint saving driven by diversity metrics and multi-layer initialization. In the EAA stage, we further enhance transferability by adding perturbations to high-level features. Additionally, we integrate an adversarial makeup technique that generates adversarial disguises from reference images, thereby further boosting the attack's effectiveness. Experimental results show that MEA outperforms the second-best input transformation attack by 20.35% and achieves a 9.22% improvement over existing facial adversarial attacks.