Neurocomputing最新文献

{"title":"Adaptive neural network fault-tolerant control for stochastic nonlinear systems based on reinforcement learning","authors":"Hao Jiang ,&nbsp;Liping Yin ,&nbsp;Zong-Yao Sun ,&nbsp;Chaoxu Mu ,&nbsp;Junsheng Zhao","doi":"10.1016/j.neucom.2026.133035","DOIUrl":"10.1016/j.neucom.2026.133035","url":null,"abstract":"<div><div>This paper presents a reinforcement learning-based adaptive fault-tolerant optimal control strategy for a class of stochastic nonlinear systems subject to dead-zone input, sensor faults, and actuator faults. Initially, a reinforcement learning algorithm with an identifier-critic-actor neural networks structure is introduced, and an innovative optimal control scheme featuring prescribed-time convergence is designed. To tackle the issue of unmeasurable states due to sensor faults, a fault-tolerant coordinate transformation technique is employed. By leveraging the universal approximation property of neural networks, the unknown nonlinear terms in the system, as well as the uncertainties arising from faults and dead-zone, are effectively estimated. Moreover, to optimize the utilization of communication resources, an event-triggered fault-tolerant optimal controller is designed. Theoretical analysis demonstrates that the proposed algorithm ensures that all signals in the closed-loop system remain bounded in probability and that the tracking error converges to a neighborhood of the origin within the prescribed time. Finally, simulation results further validate the effectiveness of the designed control strategy.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133035"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386790","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"STGNet: A spatio-temporal graph neural network for motion prediction in autonomous driving","authors":"Xinyu Wang,&nbsp;Li Liu","doi":"10.1016/j.neucom.2026.133063","DOIUrl":"10.1016/j.neucom.2026.133063","url":null,"abstract":"<div><div>Predicting the future motion of agents in a scene is vital for safe autonomous driving. Current motion prediction methods typically utilize graph neural networks (GNNs) to model the interactions between scene entities retrieved from a fixed radius, but fail to consider potential correlations that extend beyond this radius. Moreover, existing methods only obtain a relative positional encoding between scene entities from a single viewpoint, ignoring spatio-temporal features from mutual viewpoints. Targeting these problems, this paper proposes a GNN-based motion prediction method referred to as STGNet. STGNet first exploits the spatio-temporal correlations in a scene to construct an interaction graph between entities based on their similarities in a feature space. Then, it adopts a dynamic neighboring strategy to adaptively select the radius of the neighborhood in the graph according to the scene context. Moreover, STGNet employs an expressive relative positional encoding to represent the pairwise relationships in the graph, which enhances the quality of the positional encoding by using the spatio-temporal features in mutual viewpoints. Experiments on two public motion prediction datasets, along with an extensive analysis, validate the effectiveness of STGNet and its superiority to other motion prediction methods in the literature.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133063"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386792","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A unified graph neural network-based approach for few-shot learning with task nodes and DiffPool abstraction","authors":"Poupak Azad ,&nbsp;Arash Heidari ,&nbsp;Cuneyt Gurcan Akcora ,&nbsp;Ahmad Khonsari ,&nbsp;Seyed Hamed Rastegar","doi":"10.1016/j.neucom.2026.133003","DOIUrl":"10.1016/j.neucom.2026.133003","url":null,"abstract":"<div><div>Deep learning models usually work best when there is a lot of labeled data available. Few-shot node classification, on the other hand, has big problems in many real-world sectors, such as medical diagnostics, anomaly detection, social networks, and blockchain monitoring. These include samples that are not all the same and are hard to find, graph structures that change over time, node representations that are not balanced, and patterns that change over time. These make traditional training approaches less effective. For this reason, we suggest a new Few-Shot Learning (FSL) framework based on Graph Neural Networks (GNNs), especially Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs). This framework is meant to improve local geometric connectivity and optimize task graph topologies. Our solution adds three important parts to the architecture: (1) class nodes, which are made by averaging feature representations of support samples with the same label and connecting them with K-Nearest Neighbor (KNN) edges to make the semantic alignment between support and query instances better; (2) a task node, which captures global task-level context to make individual node representations richer; and (3) a graph pooling module (DiffPool), which hierarchically abstracts task graphs into task embeddings, and helps generalization across tasks. GCNs combine features from nearby nodes, while GATs employ attention techniques to make sure that the most important nodes get passed messages, which improves the graph structure even more. We also use approximation KNN methods to build graphs more quickly, which cuts down on runtime a lot without losing accuracy in predictions. Traditional meta-learning methods need complicated episodic training, but our technique makes learning easier by using improved task graphs directly for classification. We did a lot of tests on seven benchmark datasets, including Cora, Citeseer, Pubmed, and Corafull, and our technique almost did better than the others, getting state-of-the-art accuracy (up to 98.25 % on Corafull).</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133003"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386794","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MID-YOLO: An enhanced YOLOv8-based method for multi-type insulator defect detection","authors":"Lin Zhu ,&nbsp;Yuxing Mao ,&nbsp;Xingji Huang ,&nbsp;Chunxu Chen ,&nbsp;Wenchao Yang ,&nbsp;Bozheng Lei","doi":"10.1016/j.neucom.2026.132959","DOIUrl":"10.1016/j.neucom.2026.132959","url":null,"abstract":"<div><div>Insulators in overhead transmission systems are susceptible to the adverse effects of harsh weather and other conditions in outdoor environments, posing threats to the safety and stability of the transmission systems. Currently, deep learning-based object detection algorithms have made significant strides in the field of insulator defect detection. However, existing methods mainly tackled insulator broken, overlooking the nuanced detection of complex flashover issues, which require a robust detection network, this research advances insulator defect detection by proposing a novel algorithm based on the YOLOv8 framework, aimed at accurately localizing diverse insulator faults within complex natural scenes. Key innovations include a new neck feature fusion structure and a detail-aware module, which together enhance the model’s proficiency in identifying small defects. These modifications have led to notable improvements: a 3.1% increase in average accuracy and an 8.2% uplift in recall rate. Specifically, the CFM module’s efficient fusion and TCB’s channel weighting significantly boost feature extraction, the MRB’s multi-branch strategy increases recall while reducing parameters, and the DAB’s attention mechanism sharpens detail perception. The developed MID-YOLO model matches the performance of more parameter-heavy counterparts, offering a practical solution for insulator defect detection.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 132959"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386381","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"WSTC: Task-adaptive medical vision–language model with semantic tokens and dynamic alignment","authors":"Xiaolan Gao ,&nbsp;Jiaorao Wang ,&nbsp;Dan Yang","doi":"10.1016/j.neucom.2026.133008","DOIUrl":"10.1016/j.neucom.2026.133008","url":null,"abstract":"<div><div>Recent medical vision–language models enable zero- and few-shot transfer, yet still depend on handcrafted prompts and task-specific heads. To address these limitations, we introduce Weakly Semantic-aware Task Conditioning (WSTC), a lightweight, plug-and-play framework that endows frozen vision–language models with strong adaptability across classification, retrieval, and segmentation. WSTC contains two modules: (i) the Weakly Semantic-Aware Module, which distills task cues from image patches into image-derived semantic tokens, without text-side prompt engineering or prompt tuning, and (ii) the Task-Conditioned Dynamic Alignment Module, which dynamically generates projection matrices to align vision–language embeddings under task guidance. Built on a frozen UniMedCLIP backbone, WSTC adapts to new tasks without tuning the backbone and requires a moderate trainable parameter budget for adaptation. Across both zero-shot and few-shot regimes, our method UniMedCLIP augmented with WSTC consistently outperforms strong frozen vision–language models (CLIP, MedCLIP, BioViL, PubMedCLIP, UniMedCLIP) and further surpasses adapter baselines such as Tip-Adapter and Meta-Adapter. These results highlight WSTC as a practical solution for scalable medical vision–language models under extreme data scarcity.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133008"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386317","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DS-MVC: A deep multi-view clustering method based on dynamic confidence fusion and differential guidance","authors":"Guanghong Zhou ,&nbsp;Yong Wang ,&nbsp;Xingyuan Ji ,&nbsp;Chao Kong ,&nbsp;Guifu Lu","doi":"10.1016/j.neucom.2026.133079","DOIUrl":"10.1016/j.neucom.2026.133079","url":null,"abstract":"<div><div>Multi-view clustering aims to uncover discriminative data structures by integrating complementary information from multiple feature views. However, existing approaches often encounter several limitations: they struggle to handle high-dimensional and heterogeneous features, tend to assign suboptimal weights to different views, suffer from severe feature redundancy, and fail to effectively account for variations in view quality. These issues can substantially degrade clustering performance. To address these challenges, we propose a novel deep multi-view clustering framework, named DS-MVC. The proposed approach incorporates an enhanced feature fusion strategy and a multi-scale view contrastive learning scheme. First, we propose Dynamic View-Confidence Fusion, operating at the feature level. Specifically, we estimate the prediction confidence of each sample in each view and assign adaptive, sample-specific weights accordingly. This mechanism effectively emphasizes high-quality views while suppressing the influence of noisy or low-quality views, thereby enabling more accurate and fine-grained feature integration. Second, we propose Multi-Scale View Contrastive Learning, which leverages inter-view discrepancies to guide representation learning. By constructing hierarchical contrastive objectives based on prediction discrepancies between samples, the model is able to capture underlying structural relationships and contextual dependencies across views, leading to richer and more discriminative representations. Extensive experiments on multiple benchmark datasets demonstrate that DS-MVC achieves superior performance in terms of clustering accuracy and robustness. Furthermore, ablation studies validate the effectiveness of each component and confirm the generalization capability of the proposed framework.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133079"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386323","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Label distribution learning via implicit distribution representation","authors":"Zhuoran Zheng ,&nbsp;Han Hu ,&nbsp;Xin Su ,&nbsp;Chen Lyu","doi":"10.1016/j.neucom.2026.133029","DOIUrl":"10.1016/j.neucom.2026.133029","url":null,"abstract":"<div><div>In contrast to multi-label learning, label distribution learning characterizes the polysemy of examples by a label distribution to represent richer semantics. In the learning process of label distribution, the training data is collected mainly through manual annotation or label enhancement algorithms to generate label distribution. Unfortunately, the complexity of the manual annotation task or the inaccuracy of the label enhancement algorithm leads to noise and uncertainty in the label distribution training set. To alleviate this problem, we introduce the implicit distribution in the label distribution learning framework to characterize the uncertainty of each label value. Specifically, we use deep implicit representation learning to construct a label distribution matrix with Gaussian prior constraints, where each row component corresponds to the distribution estimate of each label value, and this row component is constrained by a prior Gaussian distribution to moderate the noise and uncertainty interference in the label distribution dataset. Finally, each row component of the label distribution matrix is transformed into a standard label distribution form by using the self-attention algorithm. We evaluate our model using several representative metrics, such as Chebyshev distance (0.0779 <span><math><mo>±</mo></math></span> 0.0021) and KL divergence (0.0404 <span><math><mo>±</mo></math></span> 0.0020), and demonstrate that our method achieves significant improvements in performance, mitigating noise and enhancing label distribution accuracy. The code is publicly available at: <span><span>https://github.com/WaterHQH/SNNGCN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133029"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386725","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Finite-time bipartite output synchronization and H∞ bipartite output synchronization for multi-weighted coupled memristive neural networks","authors":"Hong-An Tang ,&nbsp;Zi-Yi Xia ,&nbsp;Xiaofang Hu ,&nbsp;Shukai Duan ,&nbsp;Lidan Wang","doi":"10.1016/j.neucom.2026.132721","DOIUrl":"10.1016/j.neucom.2026.132721","url":null,"abstract":"<div><div>This article focuses on the finite-time bipartite output synchronization (FTBOS) and finite-time <span><math><msub><mi>H</mi><mrow><mi>∞</mi></mrow></msub></math></span> bipartite output synchronization (FTHBOS) of multi-weighted coupled memristive neural networks (CMNNs). Firstly, as an extension of the concept of finite-time bipartite synchronization, a new definition of FTBOS for CMNNs is proposed. Compared to bipartite synchronization, FTBOS in CMNNs means that only partial states are required to achieve bipartite synchronization within a finite time interval. Secondly, considering that many researchers have studied the state synchronization of single weighted CMNNs but have not yet discussed the FTBOS of multi-weighted CMNNs, this article uses the proposed definition to solve the FTBOS and adaptive FTBOS problems for multi-weighted CMNNs. Thirdly, an output feedback control scheme is utilized to investigate the FTHBOS of multi-weighted CMNNs, and some different adaptive laws are designed according to different coupling weights to obtain the adaptive FTHBOS criterion of multi-weighted CMNNs. Eventually, the feasibility of the theoretical results is illustrated by two numerical examples.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 132721"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386737","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Wrinkles in time: Multi-scale patching and super-resolution for efficient time series forecasting","authors":"Yuwei Chen,&nbsp;Wenjing Jia,&nbsp;Qiang Wu","doi":"10.1016/j.neucom.2026.133021","DOIUrl":"10.1016/j.neucom.2026.133021","url":null,"abstract":"<div><div>Time series forecasting has become an indispensable part of many fields, from weather prediction to equipment failure forecasting, where accuracy and efficiency directly impact the quality and speed of decision-making. Recently, models with self-attention mechanisms, such as Transformers, have gained widespread attention due to their strong performance in natural language processing and computer vision. However, self-attention mechanisms face certain limitations when applied to time series forecasting tasks, such as high computational complexity and inefficiency in handling long sequences. In contrast, Multilayer Perceptrons (MLPs) exhibit superior efficiency compared to self-attention mechanisms, allowing for faster processing times, especially when dealing with long sequences. However, MLPs have limitations in capturing long-range dependencies and complex patterns within time series data, which can hinder their predictive performance. To this end, we propose a hybrid Convolutional Neural Network–Multilayer Perceptron (CNN–MLP) framework designed to retain the efficiency of MLP-based forecasting while addressing its limitations. Our approach begins by decomposing the input series into its seasonal and trend components. We then apply two targeted enhancements: (i) a multi-scale patching scheme that reshapes the seasonal component into 2D patches processed by a 2D CNN to extract intra- and inter-period patterns; and (ii) a diffusion-based super-resolution module that enriches the trend component with fine-scale temporal detail. Collectively, these innovations introduce wrinkles in time, improving feature extraction at both global and local scales and yielding a model that achieves competitive accuracy while maintaining low parameter counts, low inference latency, and stable performance across diverse forecasting scenarios. Our source code is available at <span><span>https://github.com/sonnybjs/Wrinkles-in-time.</span><svg><path></path></svg></span></div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133021"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386739","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"ReCoD: Enhancing image description for cross-modal understanding via retrieval and comparison feedback mechanism","authors":"Geunyoung Jung ,&nbsp;Jun Park ,&nbsp;Hankyeol Lee ,&nbsp;Kyungwoo Song ,&nbsp;Jiyoung Jung","doi":"10.1016/j.neucom.2026.133025","DOIUrl":"10.1016/j.neucom.2026.133025","url":null,"abstract":"<div><div>To effectively utilize the large language models (LLMs) in the vision domain, it is essential to establish a strong connection between the visual and textual modalities. While deep embeddings can facilitate this connection, representing images as detailed textual descriptions offers significant advantages in terms of the usability and interpretability inherent in natural language. In this paper, we introduce a method of image description enhancement designed to generate highly detailed descriptions that include discriminative attributes of the given image, without requiring additional training. Our method, <span>ReCoD</span>, consists of two main components: 1) <em>“image retrieval”</em>, which retrieves the image most similar to the descriptions of the target image, and 2) <em>“comparison”</em>, which identifies and describes the differences between the target image and the retrieved image. These two components are complementary and form an iterative feedback mechanism. As this process iterates, the retrieved image becomes visually closer to the target image, and the descriptions become progressively more informative. Extensive experiments demonstrate the effectiveness of bridging the gap between the two modalities and the quality of our enhanced descriptions. The code is available at <span><span>https://github.com/gyjung975/ReCoD</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"676 ","pages":"Article 133025"},"PeriodicalIF":6.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147386779","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}