Neural Networks最新文献

{"title":"Self identity mapping","authors":"Xiuding Cai ,&nbsp;Yaoyao Zhu ,&nbsp;Linjie Fu ,&nbsp;Dong Miao ,&nbsp;Yu Yao","doi":"10.1016/j.neunet.2025.108132","DOIUrl":"10.1016/j.neunet.2025.108132","url":null,"abstract":"<div><div>Regularization is essential in deep learning to enhance generalization and mitigate overfitting. However, conventional techniques often rely on heuristics, making them less reliable or effective across diverse settings. We propose Self Identity Mapping (SIM), a simple yet effective, data-intrinsic regularization framework that leverages an inverse mapping mechanism to enhance representation learning. By reconstructing the input from its transformed output, SIM reduces information loss during forward propagation and facilitates smoother gradient flow. To address computational inefficiencies, We instantiate SIM as <span><math><mrow><mi>ρ</mi><mtext>SIM</mtext></mrow></math></span> by incorporating patch-level feature sampling and projection-based method to reconstruct latent features, effectively lowering complexity. As a model-agnostic, task-agnostic regularizer, SIM can be seamlessly integrated as a plug-and-play module, making it applicable to different network architectures and tasks. We extensively evaluate <span><math><mrow><mi>ρ</mi><mtext>SIM</mtext></mrow></math></span> across three tasks: image classification, few-shot prompt learning, and domain generalization. Experimental results show consistent improvements over baseline methods, highlighting <span><math><mrow><mi>ρ</mi><mtext>SIM</mtext></mrow></math></span>’s ability to enhance representation learning across various tasks. We also demonstrate that <span><math><mrow><mi>ρ</mi><mtext>SIM</mtext></mrow></math></span> is orthogonal to existing regularization methods, boosting their effectiveness. Moreover, our results confirm that <span><math><mrow><mi>ρ</mi><mtext>SIM</mtext></mrow></math></span> effectively preserves semantic information and enhances performance in dense-to-dense tasks, such as semantic segmentation and image translation, as well as in non-visual domains including audio classification and time series anomaly detection.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108132"},"PeriodicalIF":6.3,"publicationDate":"2025-10-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145268078","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":"FTA2C: Achieving superior trade-off between accuracy and robustness in adversarial training","authors":"Zhenghan Gao ,&nbsp;Chengming Liu ,&nbsp;Yucheng Shi ,&nbsp;Xin Guo ,&nbsp;Jing Xu ,&nbsp;Hong Zhang ,&nbsp;Lei Shi","doi":"10.1016/j.neunet.2025.108176","DOIUrl":"10.1016/j.neunet.2025.108176","url":null,"abstract":"<div><div>Deep neural networks are notoriously vulnerable to adversarial perturbations, largely due to the presence of non-robust features that destabilize model performance. Traditional Adversarial Training (AT) methods on feature space typically operate on one part of features individually, resulting in the loss of useful information in them, and improve robustness at the expense of accuracy, making it difficult to optimize the inherent trade-off between the two. To address this challenge, we propose a novel plug-in method termed Feature Transformation Alignment and Compression (FTA2C). FTA2C comprises three key components. First, a compression network constrains the perturbation space to reduce the vulnerability of non-robust features. Second, a feature transformation network enhances the expressiveness of robust features. Third, an alignment mechanism enforces consistency between adversarial and natural samples in the robust feature space. The above mechanism achieves co-processing of the two parts of the feature. Additionally, we propose the Defense Efficiency Metric (DEM) to evaluate defense methods. DEM quantifies the trade-off between maintaining natural accuracy and enhancing adversarial robustness, offering a unified and interpretable standard for comparing defense strategies. Extensive experiments conducted on four benchmark datasets demonstrate that FTA2C significantly improvements robustness under the high-level accuracy, resulting in superior trade-off performance. Our code is available at <span><span>https://github.com/HymanGao31/FTA2C</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108176"},"PeriodicalIF":6.3,"publicationDate":"2025-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145268322","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":"Elevating adversarial robustness by contrastive multitasking defence in medical image segmentation","authors":"Sneha Shukla,&nbsp;Puneet Gupta","doi":"10.1016/j.neunet.2025.108182","DOIUrl":"10.1016/j.neunet.2025.108182","url":null,"abstract":"<div><div>Although Deep Learning (DL)-based Medical Image Segmentation (MIS) models are critically important, adversarial attacks substantially diminish their efficacy. Such attacks subtly perturb inputs, causing the model to produce inaccurate predictions. This problem is more prevalent in medical images, as their intricate textures can mislead the model to focus on irrelevant regions, undermining performance and robustness. Thus, defending against adversarial attacks is crucial for a robust DL-based MIS model. While existing defences have proven effective in non-medical domains, their impact in medical domains remains limited. To bridge this gap, we propose a novel defence, <strong>CEASE</strong> (<strong>C</strong>ontrastiv<strong>E</strong> Multit<strong>AS</strong>king D<strong>E</strong>fence), to significantly enhance the adversarial resilience of MIS models, delivering notable performance gain. <em>CEASE</em> exhibits contrastive learning, multitask learning, and their consolidation-based defence. Initially, we investigate the importance of contrastive learning in a DL-based MIS model. It leverages the observation that learning similar features for clean, adversarial, and augmented samples during training significantly enhances adversarial robustness. Subsequently, our proposed multitask learning-based defence provides generic feature representation and selects auxiliary tasks based on their weak relevance to the main task, improving model robustness. Eventually, we leverage the advantages of contrastive and multitask learning to propose their fusion-based defence. It employs contrastive learning specifically for MIS tasks and follows the proposed multitask model architecture. Experiments on publicly available datasets across several state-of-the-art MIS models reveal that <em>CEASE</em> surpasses the well-known defences by mitigating the efficacy of adversarial attacks up to 0% attack success rate on maximum average distortion with modest performance advancement.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108182"},"PeriodicalIF":6.3,"publicationDate":"2025-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145268827","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":"WaveNet-SF: A hybrid network for retinal disease detection based on wavelet transform in spatial-frequency domain","authors":"Jilan Cheng ,&nbsp;Guoli Long ,&nbsp;Zeyu Zhang ,&nbsp;Zhenjia Qi ,&nbsp;Hanyu Wang ,&nbsp;Libin Lu ,&nbsp;Shuihua Wang ,&nbsp;Yudong Zhang ,&nbsp;Jin Hong","doi":"10.1016/j.neunet.2025.108189","DOIUrl":"10.1016/j.neunet.2025.108189","url":null,"abstract":"<div><div>Retinal diseases are a leading cause of vision impairment and blindness, with timely diagnosis being critical for effective treatment. Optical Coherence Tomography (OCT) has become a standard imaging modality for retinal disease diagnosis, but OCT images often suffer from issues such as speckle noise, complex lesion shapes, and varying lesion sizes, making interpretation challenging. In this paper, we propose a novel model, WaveNet-SF, to enhance retinal disease detection by integrating the spatial-domain and frequency-domain learning. The framework utilizes wavelet transforms to decompose OCT images into low- and high-frequency components, enabling the model to extract both global structural features and fine-grained details. To improve lesion detection, we introduce a Multi-Scale Wavelet Spatial Attention (MSW-SA) module, which enhances the model's focus on regions of interest at multiple scales. Additionally, a High-Frequency Feature Compensation (HFFC) block is incorporated to recover edge information lost during wavelet decomposition, suppress noise, and preserve fine details crucial for lesion detection. Our approach achieves state-of-the-art (SOTA) classification accuracies of 97.82 % and 99.58 % on the OCT-C8 and OCT2017 datasets, respectively, surpassing existing methods. These results demonstrate the efficacy of WaveNet-SF in addressing the challenges of OCT image analysis and its potential as a powerful tool for retinal disease diagnosis.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108189"},"PeriodicalIF":6.3,"publicationDate":"2025-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145259713","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":"Linear convergence of proximal gradient method for linear sparse SVM","authors":"Xiaoqi Jiao ,&nbsp;Heng Lian ,&nbsp;Jiamin Liu ,&nbsp;Yingying Zhang","doi":"10.1016/j.neunet.2025.108162","DOIUrl":"10.1016/j.neunet.2025.108162","url":null,"abstract":"<div><div>Despite the hinge loss function being non-strongly-convex and non-strongly smooth, we establish the linear rate of convergence for sparse linear support vector machines (SVM) up to its statistical accuracy. The algorithm we use is the proximal gradient method for composite functions, applied to a sequence of regularization parameters to compute the approximate solution path on a grid. Unlike works on loss functions that are strongly convex and strongly smooth, here we do not have linear convergence to the exact solution, but we can demonstrate <em>linear convergence to the population truth up to the statistical error</em> (in particular, we simultaneously consider numerical convergence and statistical convergence). For any regularization parameter in the chosen decreasing sequence, we show that the estimator is in a small neighborhood of the exact solution after <span><math><mrow><mi>O</mi><mo>(</mo><mi>log</mi><msup><mi>s</mi><mo>*</mo></msup><mo>)</mo></mrow></math></span> iterations, where <span><math><msup><mi>s</mi><mo>*</mo></msup></math></span> is the sparsity of the true coefficient in the model, and a total number of <span><math><mrow><mi>O</mi><mo>(</mo><mi>log</mi><mi>n</mi><mo>)</mo></mrow></math></span> stages (i.e., using a sequence of regularization parameters of length <span><math><mrow><mi>O</mi><mo>(</mo><mi>log</mi><mi>n</mi><mo>)</mo></mrow></math></span>) are required to achieve the near-oracle statistical rate, with <span><math><mi>n</mi></math></span> the sample size.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108162"},"PeriodicalIF":6.3,"publicationDate":"2025-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145268317","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":"Uncertainty propagation in feed-forward neural network models","authors":"Jeremy Diamzon ,&nbsp;Daniele Venturi","doi":"10.1016/j.neunet.2025.108178","DOIUrl":"10.1016/j.neunet.2025.108178","url":null,"abstract":"<div><div>We develop new uncertainty propagation methods for feed-forward neural network architectures with leaky ReLU activation functions subject to random perturbations in the input vectors. In particular, we derive analytical expressions for the probability density function (PDF) of the neural network output and its statistical moments as a function of the input uncertainty and the parameters of the network, i.e., weights and biases. A key finding is that an appropriate linearization of the leaky ReLU activation function yields accurate statistical results even for large perturbations in the input vectors. This can be attributed to the way information propagates through the network. We also propose new analytically tractable Gaussian copula surrogate models to approximate the full joint PDF of the neural network output. To validate our theoretical results, we conduct Monte Carlo simulations and a thorough error analysis on a multi-layer neural network representing a nonlinear integro-differential operator between two polynomial function spaces. Our findings demonstrate excellent agreement between the theoretical predictions and Monte Carlo simulations.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108178"},"PeriodicalIF":6.3,"publicationDate":"2025-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145268831","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":"Autoencoder-based contrastive learning for next basket recommendation","authors":"Ling Huang ,&nbsp;Zhe-Yuan Li ,&nbsp;Xiao-Dong Huang ,&nbsp;Yuefang Gao ,&nbsp;Chang-Dong Wang ,&nbsp;Philip S. Yu","doi":"10.1016/j.neunet.2025.108166","DOIUrl":"10.1016/j.neunet.2025.108166","url":null,"abstract":"<div><div>Next Basket Recommendation (NBR) aims to predict the items in the next basket a user will interact with, based on the user’s basket interaction history. However, data sparsity has been a significant challenge in this area. Contrastive Learning (CL) leverages data augmentation and constructs contrastive losses to enhance the embeddings quality, thus effectively addressing the issue of data sparsity. However, the existing methods rely on adding information to basket embedding or segmenting baskets for contrastive learning, which tend to disrupt the original embedding and have limited applicability in the NBR scenarios with diverse data characteristics. To address the above problems, we propose a novel model called Autoencoder-based Contrastive learning for Next Basket Recommendation (AC-NBR). The proposed method mainly consists of three modules, namely AE-based Basket Augmentation, AE-based Contrastive Learning, and Next-Basket Predictor. In the first module, two different basket augmentation methods are designed to provide sufficient and diverse positive pairs for CL. Specifically, we leverage an encoder-decoder structure with appropriate Gaussian noise to extract key features. This process not only helps mitigate noise interference but also improves the robustness of the embedding representation. In addition, the mean and standard deviation of the embedding representation space are learned separately. Then, Gaussian sampling is performed and the sampled latent representation is reconstructed through the decoder to achieve basket augmentation. This approach preserves core information while enhancing the embedding’s diversity and adaptability. In the second module, based on the two basket augmentations and the initial basket embeddings, three sets of positive pairs are constructed for CL. In the third module, we first encode the optimized basket sequence through a Gated Recurrent Unit (GRU) and then employ two Multi-Layer Perceptrons (MLPs) to predict the items likely to be contained in the next basket, thereby obtaining the final prediction results. The effectiveness of AC-NBR is confirmed through comprehensive experiments on three real-world datasets.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108166"},"PeriodicalIF":6.3,"publicationDate":"2025-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145268321","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":"Multi-view spectral clustering algorithm based on bipartite graph and multi-feature similarity fusion","authors":"Shunyong Li ,&nbsp;Kun Liu ,&nbsp;Mengjiao Zheng ,&nbsp;Liang Bai","doi":"10.1016/j.neunet.2025.108177","DOIUrl":"10.1016/j.neunet.2025.108177","url":null,"abstract":"<div><div>Multi-view clustering remains a challenging task due to the heterogeneity and inconsistency across multiple views. Most esisting multi-view spectral clustering methods adopt a two-stage approch–constructing fused spectral embeddings matrix followed by k-means clustering–which often leads to information loss and suboptimal performance. Moreover, current graph and feature fusion strategies struggle to address view-specific discrepancies and label misalignment, while their high computational complexity hinders scalability to large datasets. To overcome these limitations, we propose a unified Multi-view Spectral Clustering algorithm based on Bipartite Graph and Multi-feature Similarity Fusion (BG-MFS). The proposed framework jointly integrates bipartite graph construction, multi-feature similarity fusion, and discrete clustering within a single optimization model, enabling mutual reinforcement among components. Furthermore, an entropy-based weighting mechanism is introduced to adaptively assess the contribution of each view. Extensive experiments demonstrate that BG-MFS consistently outperforms state-of-the-art methods in both clustering accuracy and computational efficiency.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108177"},"PeriodicalIF":6.3,"publicationDate":"2025-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145259462","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":"CEVG-RTNet: A real-time architecture for robust forest fire smoke detection in complex environments","authors":"Jun Wang ,&nbsp;Chunman Yan","doi":"10.1016/j.neunet.2025.108187","DOIUrl":"10.1016/j.neunet.2025.108187","url":null,"abstract":"<div><div>Forest fire smoke detection is crucial for early warning and emergency management, especially under complex environmental conditions such as low contrast, high transparency, background interference, low illumination, occlusion, and overlapping smoke sources. These factors significantly hinder detection accuracy in real-world scenarios. To address these challenges, we propose CEVG-RTNet, a real-time forest fire smoke detection architecture designed to enhance robustness under such complex conditions. CEVG-RTNet incorporates several novel components. The Spatial-Channel Priori Perceptual Convolution (SCPP-Conv) module improves the model's ability to localize smoke and perceive its morphology, even in low-contrast and high-transparency environments. The Hierarchical Residual Feature Alignment (HRFA) module addresses the challenge of multi-scale feature extraction by aligning local and large-scale smoke features through a residual-guided alignment strategy and multi-layer perceptron (MLP)-based aggregation. To further refine dynamic smoke detection, the Dynamic Recursive Feature Enhancement (DRFE) module applies recursive channel adaptive enhancement and cross-channel attention strategies. Additionally, Polygonal-Intersection over Union (PolyIoU) Loss, a novel loss function, is introduced to handle the morphological complexity of smoke regions. The architecture leverages a graph sparse attention mechanism to enhance accuracy without excessive computational cost. Experimental results demonstrate the effectiveness of CEVG-RTNet, with the variant CEVG-RTNet-n achieving 89.1% precision, 82.9% recall, [email protected] of 89%, and [email protected]:0.95 of 58.9%. The model operates with 3.04M parameters, 6.6G FLOPs, and 99.42 FPS, showcasing its strong generalization, anti-interference capabilities, and suitability for complex forest fire smoke detection. The source code is available at: <span><span>https://github.com/CNNanmuzi/CEVG-RTNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108187"},"PeriodicalIF":6.3,"publicationDate":"2025-10-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145253397","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":"Anchor point segmentation based multi-view clustering","authors":"Wenhua Dong ,&nbsp;Xiao-Jun Wu ,&nbsp;Bo Fan","doi":"10.1016/j.neunet.2025.108175","DOIUrl":"10.1016/j.neunet.2025.108175","url":null,"abstract":"<div><div>Existing bipartite graph based methods commonly learn a consistent anchor graph across multiple views utilizing various optimization techniques to determine clustering assignments, maintaining linear complexity w.r.t. the number of samples. Owing to their efficiency and effectiveness, these approaches have attracted significant attention. However, the inherent geometric relationship in which anchors and the raw data share common centroids remains under-explored, leaving room for potential improvements in algorithm efficiency. This relationship enables the use of anchors to efficiently learn clustering centroids. In this paper, we propose a novel multi-view clustering approach termed anchor point segmentation based multi-view clustering (APS-MVC). Specifically, we group the raw data by first assigning each data point to an anchor point, then to a centroid. This process is modeled as a two-step transition within a Markov chain, where the optimal centroids and the soft partition of anchors are learned simultaneously by encoding the graph structure information of the anchor points. Furthermore, the proposed APS-MVC effectively tackles the out-of-sample issue. The resultant optimization problem is solved efficiently, exhibiting square complexity w.r.t. the number of anchors. Experimental results on six benchmark datasets validate the effectiveness of the proposed method. The source code is available at: <span><span>https://github.com/Wenhua-Dong/APS-MVC</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"194 ","pages":"Article 108175"},"PeriodicalIF":6.3,"publicationDate":"2025-10-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145268832","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}