Neural Networks最新文献

{"title":"Fine-grained hierarchical dynamics for image harmonization","authors":"Peng He ,&nbsp;Jun Yu ,&nbsp;Liuxue Ju ,&nbsp;Fang Gao","doi":"10.1016/j.neunet.2025.107618","DOIUrl":"10.1016/j.neunet.2025.107618","url":null,"abstract":"<div><div>Image harmonization aims to generate visually consistent composite images by ensuring compatibility between the foreground and background. Existing image harmonization strategies based on the global transformation emphasize using background information for foreground normalization, potentially overlooking significant variations in appearance among regions within various scenes. Simultaneously, the coherence of local information plays a critical role in generating visually consistent images as well. To address these issues, we propose the Hierarchical Dynamics Appearance Translation (HDAT) framework, enabling a seamless transition of features and parameters from local to global views in the network and adaptive adjustments of foreground appearance based on corresponding background information. Specifically, we introduce the dynamic region-aware convolution and fine-grained mixed attention mechanism to promote the harmonious coordination of global and local details. Among them, the dynamic region-aware convolution guided by foreground masks is utilized to learn adaptive representations and correlations of foreground and background elements based on global dynamics. Meanwhile, the fine-grained mixed attention dynamically adjusts features at different channels and positions to achieve local adaptations. Furthermore, we integrate a novel multi-scale feature calibration strategy to ensure information consistency across varying scales. Extensive experiments demonstrate that our HDAT framework significantly reduces the number of network parameters while outperforming existing methods both qualitatively and quantitatively.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107618"},"PeriodicalIF":6.0,"publicationDate":"2025-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144178711","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Optimizing connectivity through network gradients for Restricted Boltzmann Machines","authors":"Amanda Camacho Novaes de Oliveira,&nbsp;Daniel Ratton Figueiredo","doi":"10.1016/j.neunet.2025.107486","DOIUrl":"10.1016/j.neunet.2025.107486","url":null,"abstract":"<div><div>Leveraging sparse networks to connect successive layers in deep neural networks has recently been shown to provide benefits to large-scale state-of-the-art models. However, network connectivity also plays a significant role in the learning performance of shallow networks, such as the classic Restricted Boltzmann Machine (RBM). Efficiently finding sparse connectivity patterns that improve the learning performance of shallow networks is a fundamental problem. While recent principled approaches explicitly include network connections as model parameters that must be optimized, they often rely on explicit penalization or network sparsity as a hyperparameter. This work presents the Network Connectivity Gradients (NCG), an optimization method to find optimal connectivity patterns for RBMs. NCG leverages the idea of network gradients: given a specific connection pattern, it determines the gradient of every possible connection and uses the gradient to drive a continuous connection strength parameter that in turn is used to determine the connection pattern. Thus, learning RBM parameters and learning network connections is truly jointly performed, albeit with different learning rates, and without changes to the model’s classic energy-based objective function. The proposed method is applied to the MNIST and other data sets showing that better RBM models are found for the benchmark tasks of sample generation and classification. Results also show that NCG is robust to network initialization and is capable of both adding and removing network connections while learning.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107486"},"PeriodicalIF":6.0,"publicationDate":"2025-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144184485","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Frequency Self-Adaptation Graph Neural Network for Unsupervised Graph Anomaly Detection","authors":"Ming Gu ,&nbsp;Gaoming Yang ,&nbsp;Zhuonan Zheng ,&nbsp;Meihan Liu ,&nbsp;Haishuai Wang ,&nbsp;Jiawei Chen ,&nbsp;Sheng Zhou ,&nbsp;Jiajun Bu","doi":"10.1016/j.neunet.2025.107612","DOIUrl":"10.1016/j.neunet.2025.107612","url":null,"abstract":"<div><div>Unsupervised Graph Anomaly Detection (UGAD) seeks to identify abnormal patterns in graphs without relying on labeled data. Among existing UGAD methods, Graph Neural Networks (GNNs) have played a critical role in learning effective representation for detection by filtering low-frequency graph signals. However, the presence of anomalies can shift the frequency band of graph signals toward higher frequencies, thereby violating the fundamental assumptions underlying GNNs and anomaly detection frameworks. To address this challenge, the design of novel graph filters has garnered significant attention, with recent approaches leveraging anomaly labels in a semi-supervised manner. Nonetheless, the absence of anomaly labels in real-world scenarios has rendered these methods impractical, leaving the question of how to design effective filters in an unsupervised manner largely unexplored. To bridge this gap, we propose a novel <strong>F</strong>requency Self-<strong>A</strong>daptation <strong>G</strong>raph Neural Network for Unsupervised Graph <strong>A</strong>nomaly <strong>D</strong>etection (<strong>FAGAD</strong>). Specifically, FAGAD adaptively fuses signals across multiple frequency bands using full-pass signals as a reference. It is optimized via a self-supervised learning approach, enabling the generation of effective representations for unsupervised graph anomaly detection. Experimental results demonstrate that FAGAD achieves state-of-the-art performance on both artificially injected datasets and real-world datasets. The code and datasets are publicly available at <span><span>https://github.com/eaglelab-zju/FAGAD</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107612"},"PeriodicalIF":6.0,"publicationDate":"2025-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144178709","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Embedding Space Allocation with Angle-Norm Joint Classifiers for few-shot class-incremental learning","authors":"Dunwei Tu ,&nbsp;Huiyu Yi ,&nbsp;Tieyi Zhang ,&nbsp;Ruotong Li ,&nbsp;Furao Shen ,&nbsp;Jian Zhao","doi":"10.1016/j.neunet.2025.107608","DOIUrl":"10.1016/j.neunet.2025.107608","url":null,"abstract":"<div><div>Few-shot class-incremental learning (FSCIL) aims to continually learn new classes from only a few samples without forgetting previous ones, requiring intelligent agents to adapt to dynamic environments. FSCIL combines the characteristics and challenges of class-incremental learning and few-shot learning: (i) Current classes occupy the entire feature space, which is detrimental to learning new classes. (ii) The small number of samples in incremental rounds is insufficient for fully training. In existing mainstream virtual class methods, to address the challenge (i), they attempt to use virtual classes as placeholders. However, new classes may not necessarily align with the virtual classes. For challenge (ii), they replace trainable fully connected layers with Nearest Class Mean (NCM) classifiers based on cosine similarity, but NCM classifiers do not account for sample imbalance issues. To address these issues in previous methods, we propose the class-center guided embedding Space Allocation with Angle-Norm joint classifiers (SAAN) learning framework, which provides balanced space for all classes and leverages norm differences caused by sample imbalance to enhance classification criteria. Specifically, for challenge (i), SAAN divides the feature space into multiple subspaces and allocates a dedicated subspace for each session by guiding samples with the pre-set category centers. For challenge (ii), SAAN establishes a norm distribution for each class and generates angle-norm joint logits. Experiments demonstrate that SAAN can achieve state-of-the-art performance and it can be directly embedded into other SOTA methods as a plug-in, further enhancing their performance.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107608"},"PeriodicalIF":6.0,"publicationDate":"2025-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144178710","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Real-time fine finger motion decoding for transradial amputees with surface electromyography","authors":"Zihan Weng ,&nbsp;Yang Xiao ,&nbsp;Peiyang Li ,&nbsp;Chanlin Yi ,&nbsp;Pouya Bashivan ,&nbsp;Hailin Ma ,&nbsp;Guang Yao ,&nbsp;Yuan Lin ,&nbsp;Fali Li ,&nbsp;Dezhong Yao ,&nbsp;Jingming Hou ,&nbsp;Yangsong Zhang ,&nbsp;Peng Xu","doi":"10.1016/j.neunet.2025.107605","DOIUrl":"10.1016/j.neunet.2025.107605","url":null,"abstract":"<div><div>Advancements in human-machine interfaces (HMIs) are pivotal for enhancing rehabilitation technologies and improving the quality of life for individuals with limb loss. This paper presents a novel CNN-Transformer model for decoding continuous fine finger motions from surface electromyography (sEMG) signals by integrating the convolutional neural network (CNN) and Transformer architecture, focusing on applications for transradial amputees. This model leverages the strengths of both convolutional and Transformer architectures to effectively capture both local muscle activation patterns and global temporal dependencies within sEMG signals.</div><div>To achieve high-fidelity sEMG acquisition, we designed a flexible and stretchable epidermal array electrode sleeve (EAES) that conforms to the residual limb, ensuring comfortable long-term wear and robust signal capture, critical for amputees. Moreover, we presented a computer vision (CV) based multimodal data acquisition protocol that synchronizes sEMG recordings with video captures of finger movements, enabling the creation of a large, labeled dataset to train and evaluate the proposed model.</div><div>Given the challenges in acquiring reliable labeled data for transradial amputees, we adopted transfer learning and few-shot calibration to achieve fine finger motion decoding by leveraging datasets from non-amputated subjects. Extensive experimental results demonstrate the superior performance of the proposed model in various scenarios, including intra-session, inter-session, and inter-subject evaluations. Importantly, the system also exhibited promising zero-shot and few-shot learning capabilities for amputees, allowing for personalized calibration with minimal training data. The combined approach holds significant potential for advancing real-time, intuitive control of prostheses and other assistive technologies.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107605"},"PeriodicalIF":6.0,"publicationDate":"2025-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144185014","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 novel dynamic prescribed performance fuzzy-neural backstepping control for PMSM under step load","authors":"Xuechun Hu ,&nbsp;Yu Xia ,&nbsp;Zsófia Lendek ,&nbsp;Jinde Cao ,&nbsp;Radu-Emil Precup","doi":"10.1016/j.neunet.2025.107627","DOIUrl":"10.1016/j.neunet.2025.107627","url":null,"abstract":"<div><div>In order to meet the performance requirements of permanent magnet synchronous motor (PMSM) systems with time-varying model parameters and input constraints under step load, this paper proposes a dynamic prescribed performance fuzzy-neural backstepping control approach. Firstly, a novel finite-time asymmetric dynamic prescribed performance function (FADPPF) is proposed to tackle the issues of exceeding predefined error, control singularity, and system instability that arise in the traditional prescribed performance function under load changes. To address model accuracy degradation and control quality deterioration caused by nonlinear time-varying parameters and input constraints in the PMSM system, a backstepping controller is designed by combining the speed function (SF), fuzzy neural network (FNN), and the proposed FADPPF. The FNN approximates nonlinear uncertain functions in the system model; the SF, as an error amplification mechanism, works together with FADPPF to ensure the transient and steady-state performance of the system. The stability of the devised control strategy is proved using Lyapunov analysis. Finally, simulation results demonstrate the dynamic self-adjusting ability and effectiveness of FADPPF under step load. In addition, the feasibility and superiority of the proposed control scheme are validated.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107627"},"PeriodicalIF":6.0,"publicationDate":"2025-05-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144169815","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Cross-view self-supervised heterogeneous graph representation learning","authors":"Danfeng Zhao,&nbsp;Yanhao Chen,&nbsp;Wei Song,&nbsp;Qi He","doi":"10.1016/j.neunet.2025.107681","DOIUrl":"10.1016/j.neunet.2025.107681","url":null,"abstract":"<div><div>Heterogeneous graph neural networks (HGNNs) often face challenges in efficiently integrating information from multiple views, which hinders their ability to fully leverage complex data structures. To overcome this problem, we present an improved graph-level cross-attention mechanism specifically designed to enhance multi-view integration and improve the model's expressiveness in heterogeneous networks. By incorporating random walks, the Katz index, and Transformers, the model captures higher-order semantic relationships between nodes within the meta-path view. Node context information is extracted by decomposing the network and applying the attention mechanism within the network schema view. The improved graph-level cross-attention in the cross-view context adaptively fuses features from both views. Furthermore, a contrastive loss function is employed to select positive samples based on the local connection strength and global centrality of nodes, enhancing the model's robustness. The suggested self-supervised model performs exceptionally well in node classification and clustering tasks, according to experimental data, demonstrating the effectiveness of our method.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107681"},"PeriodicalIF":6.0,"publicationDate":"2025-05-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144185015","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 classification method of motor imagery based on brain functional networks by fusing PLV and ECSP","authors":"Chunling Fan,&nbsp;Yuebin Song,&nbsp;Xiaoqian Mao","doi":"10.1016/j.neunet.2025.107684","DOIUrl":"10.1016/j.neunet.2025.107684","url":null,"abstract":"<div><div>In order to enhance the decoding ability of brain states and evaluate the functional connection changes of relevant nodes in brain regions during motor imagery (MI), this paper proposes a brain functional network construction method which fuses edge features and node features. And we use deep learning methods to realize MI classification of left and right hand grasping tasks. Firstly, we use phase locking value (PLV) to extract edge features and input a weighted PLV to enhanced common space pattern (ECSP) to extract node features. Then, we fuse edge features and node features to construct a novel brain functional network. Finally, we construct an attention and multi-scale feature convolutional neural network (AMSF-CNN) to validate our method. The performance indicators of the brain functional network on the SHU_Dataset in the corresponding brain region will increase and be higher than those in the contralateral brain region when imagining one hand grasping. The average accuracy of our method reaches 79.65 %, which has a 25.85 % improvement compared to the accuracy provided by SHU_Dataset. By comparing with other methods on SHU_Dataset and BCI IV 2a Dataset, the average accuracies achieved by our method outperform other references. Therefore, our method provides theoretical support for exploring the working mechanism of the human brain during MI.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107684"},"PeriodicalIF":6.0,"publicationDate":"2025-05-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144185016","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"DiverseReID: Towards generalizable person re-identification via Dynamic Style Hallucination and decoupled domain experts","authors":"Jieru Jia,&nbsp;Huidi Xie,&nbsp;Qin Huang,&nbsp;Yantao Song,&nbsp;Peng Wu","doi":"10.1016/j.neunet.2025.107602","DOIUrl":"10.1016/j.neunet.2025.107602","url":null,"abstract":"<div><div>Person re-identification (re-ID) models often fail to generalize well when deployed to other camera networks with domain shift. A classical domain generalization (DG) solution is to enhance the diversity of source data so that a model can learn more domain-invariant, and hence generalizable representations. Existing methods typically mix images from different domains in a mini-batch to generate novel styles, but the mixing coefficient sampled from predefined Beta distribution requires careful manual tuning and may render sub-optimal performance. To this end, we propose a plug-and-play Dynamic Style Hallucination (DSH) module that adaptively adjusts the mixing weights based on the style distribution discrepancy between image pairs, which is dynamically measured with the reciprocal of Wasserstein distances. This approach not only reduces the tedious manual tuning of parameters but also significantly enriches style diversity by expanding the perturbation space to the utmost. In addition, to promote inter-domain diversity, we devise a Domain Experts Decoupling (DED) loss, which constrains features from one domain to go towards the orthogonal direction against features from other domains. The proposed approach, dubbed DiverseReID, is parameter-free and computationally efficient. Without bells and whistles, it outperforms the state-of-the-art on various DG re-ID benchmarks. Experiments verify that style diversity, not just the size of the training data, is crucial for enhancing generalization.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"189 ","pages":"Article 107602"},"PeriodicalIF":6.0,"publicationDate":"2025-05-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144135130","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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":"Overlapping community detection via Layer-Jaccard similarity incorporated nonnegative matrix factorization","authors":"Zhijian Zhuo ,&nbsp;Bilian Chen","doi":"10.1016/j.neunet.2025.107601","DOIUrl":"10.1016/j.neunet.2025.107601","url":null,"abstract":"<div><div>As information modernization progresses, the connections between entities become more elaborate, forming more intricate networks. Consequently, the emphasis on community detection has transitioned from discerning disjoint communities towards the identification of overlapping communities. A variety of algorithms based on the sparse adjacency matrix, which are sensitive to edge connections, are suitable for detecting edge-sparse areas between overlapping communities but lack the ability to detect edge-dense areas within the overlapping communities. Additionally, most algorithms do not take into account multihop information. To mitigate the aforementioned limitations, we propose an innovative approach termed Layer-Jaccard similarity incorporated nonnegative matrix factorization (LJSINMF), which utilizes both the adjacency matrix and the Layer-Jaccard similarity matrix. Our method initially employs a newly proposed Onion-shell method to decompose the network into layers. Subsequently, the layer of each node is used to construct a Layer-Jaccard similarity matrix, which facilitates the identification of edge-dense areas within the overlapping communities and serves as a general approach for enhancing other nonnegative matrix factorization-based algorithms. Ultimately, we integrate the adjacency matrix and the Layer-Jaccard similarity matrix into the nonnegative matrix factorization framework to determine the node-community membership matrix. Moreover, integrating the Layer-Jaccard similarity matrix into other algorithms is a promising approach to enhance their performance. Comprehensive experiments have been conducted on real-world networks and the results substantiate that the LJSINMF algorithm outperforms most state-of-the-art baseline methods in terms of three evaluation metrics.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"189 ","pages":"Article 107601"},"PeriodicalIF":6.0,"publicationDate":"2025-05-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144170220","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}