Neurocomputing最新文献

{"title":"Discovering causality for efficient cooperation in multi-agent environments","authors":"Rafael Pina,&nbsp;Varuna De Silva,&nbsp;Corentin Artaud","doi":"10.1016/j.neucom.2025.130358","DOIUrl":"10.1016/j.neucom.2025.130358","url":null,"abstract":"<div><div>In cooperative Multi-Agent Reinforcement Learning (MARL) agents are required to learn behaviours as a team to achieve a common goal. However, while learning a task, some agents may end up learning sub-optimal policies, not contributing to the objective of the team. Such agents are called lazy agents due to their non-cooperative behaviours that may arise from failing to understand whether they caused the rewards. As a consequence, we observe that the emergence of cooperative behaviours is not necessarily a byproduct of being able to solve a task as a team. In this paper, we investigate the applications of causality in MARL and how it can be applied in MARL to penalise these lazy agents. We observe that causality estimations can be used to improve the credit assignment to the agents and show how it can be leveraged to improve independent learning in MARL. Furthermore, we investigate how Amortised Causal Discovery can be used to automate causality detection within MARL environments. The results demonstrate that causality relations between individual observations and the team reward can be used to detect and punish lazy agents, making them develop more intelligent behaviours. This results in improvements not only in the overall performances of the team but also in their individual capabilities. In addition, results show that Amortised Causal Discovery can be used efficiently to find causal relations in MARL.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"642 ","pages":"Article 130358"},"PeriodicalIF":5.5,"publicationDate":"2025-05-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144069583","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":"Recursive wavelet transform network for robust copy-move forgery detection","authors":"Yakun Niu ,&nbsp;Xinjie Wu ,&nbsp;Cheng Liu","doi":"10.1016/j.neucom.2025.130373","DOIUrl":"10.1016/j.neucom.2025.130373","url":null,"abstract":"<div><div>In RWTN-Net, the Frequency Rotational-Invariant Feature Extractor (FRFE) firstly performs multi-stage wavelet transform and Sorted Convolution to extract multi-scale rotational invariant low-frequency and high-frequency features, which are robust to geometric transformations. Then, an Adaptive Multi-Scale Attention Fusion (AMAF) is designed to fuse features of different scales with an adaptive attention. The channel weights of low-resolution features are used to guide the weights allocation of high-resolution features, thereby enhancing the network’s understanding of geometric details and semantic information. Moreover, the Local Average Self-Correlation Calculation (LASCC) adopts a diagonal-guided sparse sampling strategy to select key feature points along the diagonal of each patch in the feature map for correlation calculation, which effectively improves the computational efficiency. Finally, a localization module is deployed to combine the matching maps of different receptive fields in an accumulative manner, and an adaptive U-net is further employed to obtain accurate localization results. Experimental results on public datasets demonstrate the effectiveness of the proposed RWTN-Net. The source code is available at <span><span>https://github.com/studyimg/RWTN-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"641 ","pages":"Article 130373"},"PeriodicalIF":5.5,"publicationDate":"2025-05-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143942171","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":"Over-squashing in Graph Neural Networks: A comprehensive survey","authors":"S. Akansha","doi":"10.1016/j.neucom.2025.130389","DOIUrl":"10.1016/j.neucom.2025.130389","url":null,"abstract":"<div><div>Graph Neural Networks (GNNs) revolutionize machine learning for graph-structured data, effectively capturing complex relationships. They disseminate information through interconnected nodes, but long-range interactions face challenges known as ”over-squashing”. This survey delves into the challenge of over-squashing in GNNs, where long-range information dissemination is hindered, impacting tasks reliant on intricate long-distance interactions. It comprehensively explores the causes, consequences, and mitigation strategies for over-squashing. Various methodologies are reviewed, including graph rewiring, novel normalization, spectral analysis, and curvature-based strategies, with a focus on their trade-offs and effectiveness. The survey also discusses the interplay between over-squashing and other GNN limitations, such as over-smoothing, and provides a taxonomy of models designed to address these issues in node and graph-level tasks. Benchmark datasets for performance evaluation are also detailed, making this survey a valuable resource for researchers and practitioners in the GNN field.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"642 ","pages":"Article 130389"},"PeriodicalIF":5.5,"publicationDate":"2025-05-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144069532","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":"MDiffSR: Mutual information and diffusion model in image super-resolution","authors":"Mingze Jiang ,&nbsp;Jinfu Fan ,&nbsp;Linqing Huang ,&nbsp;Zhencun Jiang ,&nbsp;Qingkai Bu","doi":"10.1016/j.neucom.2025.130372","DOIUrl":"10.1016/j.neucom.2025.130372","url":null,"abstract":"<div><div>In recent years, diffusion model has shown powerful functions and great potential in image <em>super-resolution</em>(SR) tasks due to its superiority in imaging capability and stability. Although it is increasingly used in image SR tasks and has achieved many gratifying results, it still has challenges such as insufficient utilization of original image information, resulting in blurred image details. In addition, the inference process for diffusion models can be time-consuming and difficult to train. To solve these problems, we propose the MDiffSR model, which improves the speed of image detail generation and training. Specifically, we introduce <em>mutual information</em>(MI) loss and contrast the idea of learning to improve the ability of model to utilize raw image information. In addition, in order to enhance the generation of image details, we integrate multi-head attention mechanism into the model, which enhances the ability of the model to capture global image features. Experimental results on multiple datasets show that our model outperforms existing methods both visually and quantitatively.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"642 ","pages":"Article 130372"},"PeriodicalIF":5.5,"publicationDate":"2025-05-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144069534","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 multi-scale vision mixture-of-experts for salient object detection with Kolmogorov–Arnold adapter","authors":"Chaojun Cen ,&nbsp;Fei Li ,&nbsp;Ping Hu ,&nbsp;Zhenbo Li","doi":"10.1016/j.neucom.2025.130349","DOIUrl":"10.1016/j.neucom.2025.130349","url":null,"abstract":"<div><div>Diverse domains and object variations make salient object detection (SOD) a challenging task in computer vision. Many previous studies have adopted multi-scale neural networks with attention mechanisms. Although they are popular, the design of their networks lacks sufficient flexibility, which hinders their generalization across objects of different scales and domains. To address the above issue, we propose a novel mixture-of-experts salient object detection (MoESOD) approach. We design a multi-scale mixture-of-experts (MMoE) module, essentially large neural networks, to improve the model’s expressive power and generalization ability without significantly increasing computational cost. By leveraging expert competition and collaboration strategies, the MMoE module effectively integrates contributions from different experts. The MMoE module not only captures multi-scale features but also effectively fuses semantic information across scales through the expert gating mechanism. Additionally, the novel Kolmogorov–Arnold adapter (KAA) is designed to enhance the model’s flexibility, allowing it to adapt easily to SOD tasks across different domains. Comprehensive experiments show that MoESOD consistently achieves higher performance than, or at least comparable performance to, state-of-the-art methods on <em>17</em> different SOD benchmarks and <em>1</em> downstream tasks. To the best of our knowledge, this is the first study to explore Kolmogorov–Arnold network within the SOD community.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"644 ","pages":"Article 130349"},"PeriodicalIF":5.5,"publicationDate":"2025-05-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144090650","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":"Text similarity based on two independent channels: Siamese Convolutional Neural Networks and Siamese Recurrent Neural Networks","authors":"Zhengfang He","doi":"10.1016/j.neucom.2025.130355","DOIUrl":"10.1016/j.neucom.2025.130355","url":null,"abstract":"<div><div>In the present-day context, a large amount of information exists in text. It is hard to extract meaningful and potential information from the text. From the current research, text similarity provides a method applied in many practical scenarios. Traditional text similarity algorithms are easy to implement, but more than these algorithms are needed to extract text features. At present, most text similarity algorithms are based on deep learning. However, these algorithms often struggle with adequately extracting both local and context text features, and they typically do not differentiate between the effectiveness of these two types of feature extraction. To address these shortcomings, this paper proposes Two Independent Channels: Siamese Convolutional Neural Networks and Siamese Recurrent Neural Networks (TIC-SCNN-SRNN). This approach is designed for binary classification, where ‘1’ indicates similarity and ‘0’ indicates dissimilarity. In detail, this paper proposes the Siamese Convolutional Neural Networks (SCNN) model to address the issue of insufficient extraction of local text features. Additionally, it introduces the Siamese Recurrent Neural Networks (SRNN) model to tackle the problem of insufficient extraction of context text features. Due to the issue of not distinguishing the effects of local and context text features extractions, this paper conducts independent weight learning on these two models to research which is more effective for text similarity tasks. In order to verify the effectiveness of the model, this paper experiments on SciTail, TwitterPPD, and QQP datasets. The experimental results show that SCNN and SRNN influence the text similarity tasks, but SRNN is more effective than SCNN. Furthermore, to verify the advantages of the TIC-SCNN-SRNN model, this paper tests the performance of the state-of-the-art CNN&amp;RNN models. The test results show that the TIC-SCNN-SRNN model performs the best, indicating that the model proposed in this paper is more effective for the text similarity tasks.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"643 ","pages":"Article 130355"},"PeriodicalIF":5.5,"publicationDate":"2025-05-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144072703","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":"Multi-scale spatio-temporal memory network for semi-supervised video object segmentation","authors":"Hui Wang ,&nbsp;Yuqian Zhao ,&nbsp;Fan Zhang ,&nbsp;Lingli Yu ,&nbsp;Chunhua Yang","doi":"10.1016/j.neucom.2025.130487","DOIUrl":"10.1016/j.neucom.2025.130487","url":null,"abstract":"<div><div>Recently, video object segmentation has received great attention in the computer vision community. Most of the existing methods exhibit poor segmentation performance when dealing with complex scenes with local information confusion such as foreground-foreground similarity and foreground-background similarity. To tackle this problem, this study proposes a semi-supervised video target segmentation (semi-VOS) framework multi-scale memory (MSM) based on spatio-temporal memory, aiming to solve the problem of insufficient local attention by constructing the correlation of space and time in a given video. The spatio-temporal memory network is used as the basic framework to display and store the target appearance feature calculated from the historical frames in the external memory, resulting in employing the historical frame feature information over a long period of time. Considering the difficulty of computing local correlations, a filtering mechanism is designed to remove the global noise in the memory reading stage. An atrous spatial pyramid pooling module is added before decoding to prevent the local information loss induced by the downsampling operation. Extensive experiments are conducted on video object segmentation benchmarks including DAVIS-16 validation, DAVIS-17 validation and test, and YouTube-2018 validation datasets. The experimental results demonstrate the feasibility and effectiveness of the proposed framework on various complex scenarios compared with previous methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"642 ","pages":"Article 130487"},"PeriodicalIF":5.5,"publicationDate":"2025-05-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144069533","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":"Theoretical bases for the relation between excitability, variability and synchronization in sequential neural dynamics","authors":"Blanca Berbel,&nbsp;Roberto Latorre,&nbsp;Pablo Varona","doi":"10.1016/j.neucom.2025.130218","DOIUrl":"10.1016/j.neucom.2025.130218","url":null,"abstract":"<div><div>Functional neural rhythms arise from the excitability of individual neurons and the collective dynamics of the circuits in which they are embedded. The self-organization of neural networks can be best investigated in circuits where these two aspects can be elucidated. In this context, Central Pattern Generators (CPGs) are an ideal model system where the interplay between synaptic and ionic currents shapes robust yet flexible rhythmic sequences underlying a wide range of motor functions. CPGs often integrate both chemical and electrical synapses, forming closed-loop network topologies that autonomously generate and coordinate functional rhythms. Recent findings in CPGs show the existence of sequential dynamical invariants, which are robust relationships between specific time intervals that build the rhythmic sequence constraining their cycle-by-cycle variability. However, the precise mechanisms by which electrical and chemical synapses interact to modulate sequential neural dynamics remain largely unknown. This modeling study examines the role of electrical currents in rhythm modulation within CPGs, with a particular focus on how these currents influence neural excitability and, in turn, modulate the variability of the rhythmic pattern on a cycle-by-cycle basis. We show that changes in neuronal excitability induced by electrical currents can modulate dynamical invariants, leading to the emergence of distinct sequence interval relationships that facilitate the rhythm adaptability. Synchronization between electrically coupled neurons is thus a key factor in regulating these effects. Finally, our results show that sequential dynamical invariants sustain the balance between excitability, variability, and synchronization, underscoring their essential role in the adaptive coordination of functional neural rhythmic patterns.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"645 ","pages":"Article 130218"},"PeriodicalIF":5.5,"publicationDate":"2025-05-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144116547","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":"LAA: Local Awareness Attention for point cloud self-supervised representation learning","authors":"Jiawei Yu ,&nbsp;Hongqiang Wu ,&nbsp;Wen Shangguan ,&nbsp;Yanchang Niu ,&nbsp;Biqing Huang","doi":"10.1016/j.neucom.2025.130365","DOIUrl":"10.1016/j.neucom.2025.130365","url":null,"abstract":"<div><div>Local awareness is essential for point cloud representation learning. In recent times, due to the increase in the amount of point cloud data and the success of the self-supervised learning paradigm in other domains, there has been an increase in the number of studies investigating point cloud self-supervised representation learning. However, the majority of current methods for implementing local awareness are incompatible with the paradigm of point cloud self-supervised pre-training, which makes it difficult for pre-trained models to benefit from it. Consequently, previous point cloud pre-training models have predominantly resulted in a global effective receptive field, with less focus on local awareness. A Gaussian-distributed, larger, more natural effective receptive field without artifacts will result in a superior representation of point cloud features. To address this issue, this paper proposes <strong>Local Awareness Attention (LAA)</strong>, a plug-and-play module that enables local geometric perception while at the same time capturing global features. LAA consists of two branches. The first obtains local geometric information through the attention of each query and its neighborhood. The remaining branch learns global features through self-attention. The LAA module then fuses the features captured by the two branches through a single softmax, resulting in a competitive mechanism that achieves adaptive and multi-scale self-attention. Extensive experiments in indoor environments demonstrate that our LAA obtains stable effect enhancement in multiple transformer-based point cloud self-supervised pretraining networks, specifically outperforming multiple baselines by 0.1%–0.2% in ModelNet40 and by 0.2%–0.3% in ScanObjectNN.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"640 ","pages":"Article 130365"},"PeriodicalIF":5.5,"publicationDate":"2025-05-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143936379","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":"FSCMF: A Dual-Branch Frequency-Spatial Joint Perception Cross-Modality Network for visible and infrared image fusion","authors":"Xiaoyang Zhang ,&nbsp;Chengpei Xu ,&nbsp;Guodong Fan ,&nbsp;Zhen Hua ,&nbsp;Jinjiang Li ,&nbsp;Jingchun Zhou","doi":"10.1016/j.neucom.2025.130376","DOIUrl":"10.1016/j.neucom.2025.130376","url":null,"abstract":"<div><div>Existing image fusion methods face limitations in deep feature modeling and high-frequency information enhancement, leading to detail loss and reduced target saliency in complex scenarios. To address these issues, this paper proposes a Dual-Branch Frequency-Spatial Joint Perception Cross-Modality Network (FSCMF), which integrates local details, global context, and frequency-domain information through a dual-branch architecture to enhance multimodal feature complementarity. Specifically, FSCMF combines CNN and Transformer in a dual-branch design, where the CNN branch focuses on extracting local structures and texture details, while the Transformer branch captures long-range dependencies to improve global consistency. To further optimize feature representation, we introduce the Frequency-Spatial Adaptive Attention Module (FSAA), in which the frequency domain branch enhances high-frequency components to improve edge sharpness, while the spatial domain branch adaptively refines salient region features, ensuring a dynamic balance between global and local information. Additionally, we propose the Weighted Cross-Spectral Feature Fusion Module (WCSFF) to enhance cross-modality feature interaction through adaptive weighting, thereby improving detail integrity and semantic consistency in the fused image. A maximum frequency loss function is further incorporated to ensure the preservation of critical frequency components. Extensive experiments on three public datasets — MSRS, M<span><math><msup><mrow></mrow><mrow><mn>3</mn></mrow></msup></math></span>FD, and LLVIP — demonstrate that FSCMF outperforms existing methods in both qualitative and quantitative evaluations, producing fusion results with higher visual consistency and better information retention. Furthermore, additional experiments on object detection and semantic segmentation validate FSCMF’s potential in high-level computer vision tasks, highlighting its broad application value. The code of FSCMF is available at <span><span>https://github.com/boshizhang123/FSCMF</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"641 ","pages":"Article 130376"},"PeriodicalIF":5.5,"publicationDate":"2025-05-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143937296","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}