Gaussian processes (GPs) are widely used for modeling dynamical systems due to their ability to provide principled uncertainty estimation and to incorporate prior knowledge via the mean or kernel function. However, GPs typically exhibit poor extrapolation performance outside the training region. To address this limitation while preserving the favorable uncertainty behavior of standard GPs, we propose a novel kernel design that integrates prior models-either analytical physical models or imperfect simulators-into the kernel function of the GP. Our approach is motivated by the assumption that similarity in the prior model implies similarity in the true system dynamics. Specifically, we transform the input space using the prior model and apply a base kernel (e.g., Radial Basis Function or Matérn) to construct a prior-model-informed kernel that reflects this assumption. To further enhance modeling flexibility, we add a standard residual kernel to correct discrepancies between the prior model and the true system. This yields our final model: a GP with prior-model-informed kernel (GP-PI-K). Through extensive experiments on benchmark dynamical systems, we demonstrate that GP-PI-K consistently outperforms existing baselines, including standard GPs, GPs with physics-informed or neural network mean functions, and deep kernel learning models, in terms of one-step and multi-step prediction accuracy, uncertainty estimation, and uncertainty calibration. Moreover, GP-PI-K achieves superior performance in downstream applications such as active learning and model-based reinforcement learning.
Wafer defect inspection is crucial for yield and reliability, but shrinking defect sizes demand higher imaging resolution. While high-magnification optics provide resolution, their narrow field of view limits inspection efficiency. To balance precision and throughput, we propose a solution that reconstructs high-resolution wafer images from large-field low-magnification captures via a super-resolution algorithm. This method can improve detection efficiency without affecting accuracy. We design a dual-domain fusion lightweight SR network (SFMambaSR) specifically for wafer microscopy images. In the spatial domain, a Visual State Space Model (VSSM) and Multi-Scale Feature Extraction (MSFE) module jointly fuse global and local representations, while in the frequency domain, a wavelet-based Frequency-Domain Transformation (FDT) module enhances high-frequency defect details. Experiments on our large-scale wafer microscopy dataset demonstrate that SFMambaSR achieves the best PSNR and competitive or best SSIM across 2 × , 3 × , and 4 × upscaling factors, while using only 807K parameters.
Knowledge tracing aims to model students' dynamic knowledge states based on their historical learning interactions and predict future learning performance. Existing sequence modeling methods often overlook individual differences and have limited semantic modeling capability. To address these issues, this paper proposes a knowledge tracing model that combines personalized modeling with knowledge distillation. The model introduces three personalized modules: (1) a personalized question understanding module that captures individual differences in students' understanding of the same question; (2) a personalized question-knowledge association module that models relationships between questions and relevant knowledge concepts; and (3) a personalized knowledge state forgetting module that simulates students' memory decay patterns. These modules allow for more accurate modeling of students' dynamic knowledge states. Furthermore, to overcome the semantic limitations of lightweight models, a large language model (LLM) is used as the teacher, and its semantic modeling capability is transferred to an LSTM-based student model via knowledge distillation. Experiments show that the proposed method consistently improves prediction performance on two benchmark datasets, demonstrating its effectiveness in modeling personalized learning and enhancing semantic representation.
Visual reinforcement learning (VRL) aims to extract effective visual information from high-dimensional observations to optimize decision-making policies. While existing VRL methods have achieved significant progress in various control tasks through data augmentation and auxiliary tasks, agents remain susceptible to distractions from redundant information and irrelevant factors, leading to overfitting and degraded generalization in unseen environments. To address this challenge, we propose a similarity-guided state attention (SSA) for VRL. In this method, a similarity guidance module (SGM) is designed to leverage the similarity between state embeddings from the original and augmented observations to guide the encoder to focus on task-relevant regions in the original observation, thereby producing a corresponding state attention map. Meanwhile, the state attention map of augmented observation is obtained by decoding its state embedding. Furthermore, cosine similarity is introduced to measure the global similarity between state attention maps of original and augmented observations, which is incorporated into self-supervised learning objective together with binary cross-entropy loss to encourage alignment of state attention maps in representation space. The combination of SGM and cosine similarity-based alignment of state attention maps facilitates self-supervised learning to obtain more robust state representations for downstream reinforcement learning, thereby enabling the agent to learn optimal policies and improve its generalization ability in unseen environments. Experimental results on the DeepMind Control Generalization Benchmark (DMControl-GB) demonstrate that SSA achieves superior robustness and generalization performance compared with representative VRL baselines.
In this research, we provide a convergence of the Adomian decomposition based neural network method for time fractional reaction diffusion initial boundary value problems in 2D, where the fractional time derivative is considered in the Caputo sense. We propose an idea of a proper combination of Adomian decomposition method (ADM) and Physics-informed neural network (PINN), after semi-discretization of the time fractional derivative. PINN uses the ADM based partial sum inside the loss function before generating the appropriate ADM-PINN approximation. In addition, we produce the required sufficient conditions on the given data under which the usual Adomian decomposition method converges in a bounded domain. Furthermore, we analyze the error bound of the proposed method for the time fractional model and demonstrate the convergence. Several numerical examples are produced to show the effectiveness of the present ADM-PINN approach. It is observed that the proposed method is highly effective for several two-dimensional time fractional problems including the Schrodinger equation for convergent approximation of the computed solution.
We establish compact-open topology uniform universal approximation driven by hypercyclic orbits of an affine composition operator. In one dimension, for suitable continuous activation functions, we show that a fixed weighted translation operator is hypercyclic on the closure, in the compact-open topology, of the class of functions represented by one-hidden-layer neural networks. Consequently, there exists a single seed function in this closure whose iterates approximate every target uniformly on compact sets. We also prove density of periodic points, and hence chaoticity of the operator. Under the corresponding compact-wise density assumption for deeper and higher-dimensional network classes, the same orbit-based approximation mechanism extends to arbitrary depth and dimension. Thus approximation is achieved by moving along the orbit of one fixed seed under one fixed operator, rather than by re-optimizing parameters for each target.
With the popularization of various sensors, multimodal Salient Object Detection (SOD) methods including Visual-Depth-Thermal Salient Object Detection (V-D-T SOD) have made remarkable developments. However, most existing V-D-T SOD methods usually achieve accurate detection with expensive computational cost, which limits their development at the edge application. To address this issue, we develop a Lightweight Attention-Aware Fusion Network (LAANet) based on the State Space Model (SSM). Specifically, inspired by the cross-attention mechanism, the Cross Mamba Fusion Module (CMFM) is proposed to realize attention perception among shallow high-resolution multimodal features by exploiting SSM with linear complexity. For low-resolution deep features, Attention Perception Fusion Module (APFM) is proposed based on the self-attention mechanism to mine semantic cues and complement the potential information lost in SSM amnestic memory. In addition, to response the problem of excessive decoding parameters and efficiently achieve feature reconstruction,we design a lightweight decoder (LDB). It is consist of simple dilated convolutions and linear operations with only 0.37M parameter. Extensive experiments on the VDT2048 dataset show that our method achieves performance close to that of SOTA, while having smaller number of parameters (6.78M), lower model complexity (5.23G), and faster inference speed (23.7FPS when the input size is 384*384). The code is available at https://github.com/GZNU-WJT/LAANet.
Webly Supervised Fine-Grained Recognition demands robust learning mechanisms to mitigate the impact of noisy labels and misleading annotations. To address this, current methods typically rely on prediction probabilities to filter noisy labels. However, such paradigms often struggle with mislabeled yet high-confidence samples, leading to confirmation bias. In this paper, we propose a novel framework named Decoupled Global-local Consensus Learning (DGCL) that identifies and leverages noisy-labeled samples through multi-perspective analysis, enhancing both robustness and representation power. Motivated by the high-quality dense representations of foundation models, we first introduce the Decoupled Global-Local Fusion (DGLF) framework, which leverages LoRA to calibrate the encoder's attention, enabling precise identification of discriminative regions. Subsequently, a dual-stream bilinear mechanism is designed to facilitate global-local interactions for enhanced detail capture. To mitigate confirmation bias, we introduce the Geometric-Semantic Consensus (GSC) strategy, which integrates geometric consistency and semantic confidence to partition samples into three mutually exclusive subsets. Finally, a Noise-Aware Supervised Contrastive (NASC) loss is introduced to convert filtered noise into repulsion signals, effectively compressing intra-class variance. Extensive experiments demonstrate that DGCL achieves a new state-of-the-art average accuracy of 91.73% on three web-supervised benchmark datasets, surpassing the previous best method by 1.63 percentage points. Our source code will be made publicly available at: https://github.com/YT3DVision/DGLF.
The depth map super-resolution guided by RGB images aims to reconstruct a high-resolution (HR) depth from a low-resolution (LR) depth map input with the assistance of a corresponding HR RGB image. Convolutional methods often struggle to maintain global consistency due to limited receptive fields, while Transformer-based models incur high computational costs on HR data. Furthermore, naive cross-modal fusion strategies frequently introduce texture copying artifacts or edge blurring when the RGB and depth modalities are misaligned. To tackle these challenges, we propose a multi-dimensional collaborative guidance network (MCGNet), a novel framework that combines the efficiency of State Space Models (SSM) with multi-dimensional guidance. First, we construct a spatial reconstruction module adopting the SSM as the reconstruction backbone. This module exploits the linear complexity of SSMs to model long-range dependencies, ensuring structurally consistent depth recovery across large smooth regions without the heavy burden of self-attention. Second, we introduce a spatial-frequency cooperative guidance module that leverages the Fourier transform to fuse RGB and depth features in the frequency domain, enabling explicit utilization of high-frequency RGB details to refine depth edges while preserving low-frequency depth structures. Third, an LR-HR mapping feature guidance module is developed to bridge the resolution gap, injecting scale-aware RGB priors into the depth decoding stream to facilitate precise LR-HR mapping learning. Extensive experiments on benchmark datasets demonstrate that our approach achieves state-of-the-art performance quantitatively and qualitatively.
Person search is challenging due to limitations in identity representation. Existing methods rely on one-hot encoding, ignoring semantic relationships among pedestrians. This leads to a fragmented feature space and reduces generalization ability, especially in large-scale scenarios with a significant proportion of unlabeled identities. For instance, in the CUHK-SYSU dataset, 72.7% of pedestrians lack identity annotations, limiting the effectiveness of supervised learning. To address these issues, we propose a novel Multimodal-Guided Self-Distillation (MGSD) method for Unified Person Search that leverages multimodal textual descriptions and self-distillation to enhance pedestrian representation learning. Specifically, we introduce three key innovations: (1) Multimodal LLM-Assisted Text Generation (MLTG) to provide fine-grained semantic context beyond discrete identity labels, enabling the model to capture inter-person relationships based on clothing attributes, appearance features, and environmental cues; (2) Semantic Structural Consistency Constraint (SSCC) to impose global structural constraints on the feature space, ensuring that distinct identities remain separable while preserving semantic similarities among visually similar individuals; and (3) Multimodal-Aware Self-Distillation Framework (MSDF), where the learnable visual encoder is progressively aligned with the pre-trained CLIP multimodal encoder, improving robustness to variations in illumination, occlusion, and background clutter. Extensive experiments demonstrate that our method significantly enhances retrieval accuracy and generalization, achieving state-of-the-art performance with an mAP of 56.1% on the PRW dataset while maintaining computational efficiency for large-scale real-world applications.
Visible-infrared person re-identification (VI-ReID) is an important technique for around-the-clock person matching, and its primary challenge arises from substantial cross-modal discrepancies. To address this challenge, we propose a Decoupled Information-Guided Cross-Modal Alignment (DIGCA) framework organized into three stages: multi-scale feature modeling, partial functional decoupling, and guided cross-modal alignment. First, a Hierarchical Context Extractor (HCE) aggregates multi-scale contextual information through dilated convolutions and residual connections to enrich the initial identity representation. Second, a Multi-Branch Unified Encoder (MBUE) organizes complementary feature streams through parallel global-relation and local-spatial modeling. The resulting representations exhibit different information tendencies and promote partial functional decoupling between identity-related and modality-related information. Finally, the Decoupled Information-Guided Cross-Modal Alignment (DIG-CMA) module refines the two streams with channel and spatial attention and integrates them through cross-attention. Under the joint optimization objective, the resulting representations support cross-modal alignment. Experiments on SYSU-MM01, RegDB, and LLCM show that DIGCA achieves competitive overall performance compared with recent VI-ReID methods, providing empirical support for the effectiveness of the proposed decoupling-guided alignment strategy in cross-modal identity matching.
Spatiotemporal Graph (STG) forecasting holds great significance in the field of urban computing. However, the challenge of data scarcity poses significant obstacles to this task. While cross-city few-shot learning offers a promising solution, existing methods face two fundamental challenges: 1) insufficient extraction of meta-knowledge from data-rich source cities, and 2) limited generality of the knowledge transfer mechanism. In this paper, we propose a novel STG few-shot learning framework named ST-MPPT, which addresses both challenges through masked pre-training and prompt tuning. In the pre-training stage, we perform spatiotemporal-decoupled masked pre-training on source cities with abundant data, enabling the model to learn long-term spatiotemporal patterns more comprehensively. In the downstream forecasting stage, we leverage the pre-trained encoders to acquire robust spatial and temporal representations. These representations are then used to construct a graph structure and enhance the downstream spatiotemporal predictor. To achieve a more general knowledge transfer, we introduce a novel prompt network. Instead of rigid pattern retrieval, this network dynamically generates input-specific prompts to steer the pre-trained encoders to adapt to different data distributions across diverse cities. Extensive experiments on four real-world spatiotemporal datasets demonstrate the superiority of ST-MPPT over strong and representative baselines.
Multimodal temporal alignment is a critical task for applications such as audiovisual understanding, lip reading, and instruction following. However, in weakly supervised settings, challenges like asynchronous sampling, irregular event triggers, and coarse labels hinder precise cross-modal alignment. Frame-based methods rely on fixed time grids, leading to redundant computation in sparse-event scenarios and reducing event-level precision. Differentiable time-warping methods typically require high-resolution inputs, resulting in high computational costs and sensitivity to numerical parameters. To address these challenges, we propose SynNeura, an event-driven continuous-time liquid-spiking neural framework for fine-grained alignment under weak supervision. SynNeura models alignment as a continuous-time latent-state process, analytically propagating states between events and updating only when spikes occur. This results in computational complexity that scales with the number of events rather than the sequence length. SynNeura introduces a piecewise-analytic update using matrix exponentials and trace variables, along with a hierarchical contrastive alignment objective at spike, trajectory, and state levels, enhancing robustness and consistency. Experiments with the AVE, LRS2, and YouCook2 datasets show SynNeura consistently outperforms frame-based and continuous-time baselines in alignment accuracy, temporal consistency, and efficiency. SynNeura achieves 0.291 MAE and 0.713 CAS on AVE, 0.648 TC on LRS2, and 0.829/0.794 EP/ER on YouCook2, with an overall score of 0.738. These results show SynNeura is a scalable, interpretable, and efficient solution for event-driven multimodal temporal alignment.
Quantization-Aware Fine-tuning (QAF) is critical for deploying large language models in resource-constrained scenarios. While in low-bit QAF, strong downstream performance hinges on both minimizing model quantization errors and providing robust initialization of low-rank adaptation. To further improve the model performance, we propose Structure-Aware Adaptation for low-bit QAF (Q-SA2) in this paper. By decomposing pre-trained weights into sparse components and principal subspaces, Q-SA2 provides a favorable initialization for downstream task adaptation. Specifically, to better preserve the representations of pre-trained weights and minimize errors, the sparse components are quantized, and the low-rank adapters are initialized from the principal subspaces. Then, we extend this framework to a weight-wise mixed-precision variant with a distribution-based outlier detection strategy on sparse components, enabling Q-SA2 to adaptively allocate higher precision to critical weights. Comprehensive experiments on WikiText-2, C4, GSM8K, and the GLUE benchmark demonstrate that the proposed method consistently outperforms state-of-the-art QAF methods, particularly on mathematical reasoning tasks, achieving an average improvement of 4% under 2-bit quantization and up to 7.78% on the Qwen2.5-7B model under 4-bit quantization. The code is available at https://github.com/Rainiay/Q-SA2.git.
Most existing ictal stereoelectroencephalography (SEEG)-based seizure onset zone (SOZ) localization methods rely on patient-specific training, limiting their clinical applicability due to the scarcity of seizure recordings and substantial inter-patient variability. Consequently, robust patient-independent SOZ localization remains a major challenge. In this work, we propose a deep learning approach for patient-independent SOZ localization using ictal SEEG recordings, aiming to improve cross-patient generalization while preserving seizure-related temporal characteristics. To mitigate domain shifts across subjects, we introduce a clinically guided feature learning strategy that combines a cross-frequency coupling (CFC) mechanism to capture SOZ-related abnormal interactions across frequency bands with a self-comparison (SC) mechanism to emphasize seizure-onset evolution patterns within SEEG channels. We further incorporate seizure detection as an auxiliary task within a multi-task learning framework to provide seizure-onset-related temporal supervision, thereby improving the temporal awareness and generalizability of SOZ localization. Experiments on the public OpenNeuro HUP dataset demonstrate substantial improvements over existing methods, while additional evaluations on a private clinical dataset further validate the robustness and cross-patient generalization capability of the proposed method. Moreover, comparisons between the learned CFC representations and clinically established phase-amplitude coupling (PAC) metrics reveal consistent physiological patterns, supporting the interpretability of the learned representations.
Anomaly detection is an essential part in industrial production which intends to identify anomaly samples and localize anomaly regions. Although current unsupervised anomaly detection methods have achieved excellent performance, they rely on clean and unlabeled normal samples. However, it is rarely satisfied in industrial scenarios and training with contaminated data is an inevitable challenge. To solve this problem, we propose an effective anomaly detection framework, named ContaminationAD. In our method, a pseudo-label generation strategy is introduced to evaluate the contamination levels of image patches and determines whether the sample is contaminated. Following this, a coarse localization of contaminated regions can be obtained and pseudo-labels are generated to guide text feature assignment for each sample. To concentrate on regions of interest while not neglecting fine-grained information, we design the Dual-Expert Network by incorporating routing mechanism to process visual feature adaptively. Based on this design, we propose semantic-driven segmentation weighting to facilitate anomaly localization. By applying routing weights and coarse localization when calculating the similarity between textual and local visual features, the training samples are dynamically weighted which enables the network to achieve a balance between the learning of normal and contaminated regions. Comprehensive experiments on the MVTecAD, VisA and BTAD datasets demonstrate that ContaminationAD outperforms the state-of-the-art anomaly detection methods. Notably, ContaminationAD exhibits strong robustness in the presence of data contamination, even when the contamination level is severe.
This paper investigates the synchronization control problem for a class of state-dependent switching neural networks with time-varying delay and semi-Markov jump parameters. A new intermittent discrete adaptive event-triggered control scheme is proposed to reduce control cost and communication burden. Specifically, a time-window mechanism is introduced to characterize the intermittent operation, where the activation width is determined by the synchronization error state at the beginning of each intermittent period. Furthermore, the triggering threshold is updated in a discrete adaptive manner according to the variation between the current sampled state and the most recently transmitted state. A switched Lyapunov-Krasovskii functional is constructed by incorporating the characteristics of semi-Markov jumps, time-varying delay, sampling period, and intermittent operation. Based on this framework, sufficient conditions are derived to guarantee exponential synchronization of the considered master-slave system. Finally, two numerical examples and two image encryption-decryption applications are provided to demonstrate the effectiveness of the proposed results.
The rapid advancement of large language models (LLMs) has revolutionized natural language processing, yet their enormous parameter size presents significant challenges for fine-tuning on downstream tasks. While parameter-efficient fine-tuning (PEFT) methods such as Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA) have substantially reduced computational requirements, they often introduce knowledge misalignment and representation degradation during adaptation. To bridge this gap, we propose GLA-LoRA, a novel PEFT framework centered on Global-Local Knowledge Alignment. Our approach establishes a unified learning strategy that synergistically integrates multi-granular contrastive learning with knowledge distillation. GLA-LoRA explicitly preserves dataset-wide semantic structures through global contrastive learning and maintains sample-level feature consistency via local invariance learning, while a teacher model guides the adaptation to retain the original model's knowledge. Extensive evaluations across eight benchmarks, spanning six GLUE tasks and two intent detection datasets, demonstrate that GLA-LoRA consistently outperforms strong baselines including standard LoRA and QLoRA. Under identical parameter budgets, GLA-LoRA achieves consistent and significant improvements over standard LoRA across three base models and eight benchmark datasets. On sentence-level classification and regression tasks, our method yields average accuracy gains of 1.4%, 1.1%, and 1.2% on LLaMA2-7B, RoBERTabase, and BERTLarge, respectively. For intent detection (CLINC and HWU), it achieves even more pronounced improvements, with up to 2.3% accuracy gain on BERTLarge. These results establish that explicit global-local knowledge alignment is essential for achieving high-fidelity, parameter-efficient fine-tuning across diverse language tasks.
Accurate identification of operating conditions in zinc oxide rotary kilns is essential to improving energy efficiency, product quality, and sustainability in non-ferrous metallurgical production. However, traditional monitoring approaches predominantly rely on operators’ empirical observation of kiln head flames, rendering the assessment subjective and limiting the effective integration of multimodal information. To overcome these constraints, this study develops a multimodal dataset tailored to zinc smelting scenarios and introduces a Cross-Modal Adaptive Fusion Network (CMAFNet). Specifically, the proposed framework incorporates a Cross-Modal Interaction Transformer (CMIT) to capture fine-grained dependencies across heterogeneous image, text, and process data. Subsequently, the Enhanced Adaptive Selection Fusion (EASF) module is developed to suppress redundancy, while the Dynamic Adaptive Fusion (DAF) module performs adaptive aggregation of the refined features at the relation level. Furthermore, a composite loss function combining alignment, center, and triplet losses is designed to mitigate cross-modal distribution discrepancies and enhance discriminative capability by reinforcing intra-class compactness and inter-class separability. Experimental results demonstrate that CMAFNet achieves an accuracy of 93.65%, an F1 score of 92.16%, and a Matthews correlation coefficient (MCC) of 0.894 on the seven-class operating condition identification task, demonstrating robust discriminative performance even for minority classes that are highly similar and prone to confusion. In the conventional three-class setting, the model attains an overall accuracy of 94.52%, with each class exceeding 90%, outperforming representative trimodal fusion baselines. Moreover, it shows competitive performance on a public dataset. Finally, ablation studies and visualizations substantiate the effectiveness and interpretability of the proposed framework.
Mean Teacher (MT)-based semi-supervised semantic segmentation faces two key limitations: (i) teacher-generated pseudo-labels are sensitive to scale variations and boundary details, leading to the accumulation of scale-induced uncertainty during iterative self-training; (ii) existing consistency regularization is largely restricted to single-scale or static alignment, lacking explicit modeling of class-level semantic structures across different views. To address these issues, we propose Cross-scale Dual-View Consistency Learning (CDV-CL), which explicitly models the complementarity between predictions from zoomed-in and zoomed-out views and jointly enforces pixel-level and class-level constraints to improve pseudo-label quality and stabilize semantic representation learning. Within this framework, we introduce a dual-view semantic boundary complementary fusion module (SEC-Fuse). It leverages cross-view confidence distributions and prediction discrepancies to estimate pixel-wise adaptive weights across dual-view teacher predictions, and then performs region-aware adaptive fusions to generate pseudo-labels that are more robust to scale variations. To complement the pixel-level pseudo-label refinement of SEC-Fuse, we further introduce cross-scale prototype alignment (CSP-Align), which aligns class prototypes across views in high-confidence non-boundary regions, thereby strengthening the class-level structure of the feature space and improving discriminability. Experiments show that CDV-CL achieves 84.72 % and 81.02 % mIoU on Pascal VOC and Cityscapes under the 1/16-label setting, respectively, outperforming state-of-the-art methods. It also demonstrates strong generalization to medical skin lesion segmentation on ISIC and remote sensing change detection.
