Neural Networks最新文献

{"title":"Bidirectional Semantic Consistency Guided Contrastive Embedding for Generative Zero-Shot Learning","authors":"Zhengzhang Hou ,&nbsp;Zhanshan Li ,&nbsp;Jingyao Li","doi":"10.1016/j.neunet.2025.107423","DOIUrl":"10.1016/j.neunet.2025.107423","url":null,"abstract":"<div><div>Generative zero-shot learning methods synthesize features for unseen classes by learning from image features and class semantic vectors, effectively addressing bias in transferring knowledge from seen to unseen classes. However, existing methods directly employ global image features without incorporating semantic information, failing to ensure that synthesized features for unseen classes maintain semantic consistency. This results in a lack of discriminative power for these synthesized features. To address these limitations, we propose a Bidirectional Semantic Consistency Guided (BSCG) generation model. The BSCG model utilizes a Bidirectional Semantic Guidance Framework (BSGF) that combines Attribute-to-Visual Guidance (AVG) and Visual-to-Attribute Guidance (VAG) to enhance interaction and mutual learning between visual features and attribute semantics. Additionally, we propose a Contrastive Consistency Space (CCS) to optimize feature quality further by improving intra-class compactness and inter-class separability. This approach ensures robust knowledge transfer and enhances the model’s generalization ability. Extensive experiments on three benchmark datasets show that the BSCG model significantly outperforms existing state-of-the-art approaches in both conventional and generalized zero-shot learning settings. The codes are available at: <span><span>https://github.com/ithicker/BSCG</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107423"},"PeriodicalIF":6.0,"publicationDate":"2025-03-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143746937","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":"Adaptive bigraph-based multi-view unsupervised dimensionality reduction","authors":"Qianyao Qiang ,&nbsp;Bin Zhang ,&nbsp;Chen Jason Zhang ,&nbsp;Feiping Nie","doi":"10.1016/j.neunet.2025.107424","DOIUrl":"10.1016/j.neunet.2025.107424","url":null,"abstract":"<div><div>As a crucial machine learning technology, graph-based multi-view unsupervised dimensionality reduction aims to learn compact low-dimensional representations for unlabeled multi-view data using graph structures. However, it faces several challenges, including the integration of multiple heterogeneous views, the absence of label guidance, the rigidity of predefined similarity graphs, and high computational intensity. To address these issues, we propose a novel method called adaptive Bigraph-based Multi-view Unsupervised Dimensionality Reduction (BMUDR). BMUDR dynamically learns view-specific anchor sets and adaptively constructs a bigraph shared by multiple views, facilitating the discovery of low-dimensional representations through sample-anchor relationships. The generation of anchors and the construction of anchor similarity matrices are integrated into the dimensionality reduction process. Diverse contributions of different views are automatically weighed to leverage their complementary and consistent properties. In addition, an optimization algorithm is designed to enhance computational efficiency and scalability, and it provides impressive performance in low-dimensional representation learning, as demonstrated by extensive experiments on various benchmark datasets.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107424"},"PeriodicalIF":6.0,"publicationDate":"2025-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143725256","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":"Using samples with label noise for robust continual learning","authors":"Hongyi Nie ,&nbsp;Shiqi Fan ,&nbsp;Yang Liu ,&nbsp;Quanming Yao ,&nbsp;Zhen Wang","doi":"10.1016/j.neunet.2025.107422","DOIUrl":"10.1016/j.neunet.2025.107422","url":null,"abstract":"<div><div>Recent studies have shown that effectively leveraging samples with label noise can enhance model robustness by uncovering more reliable feature patterns. While existing methods, such as label correction methods and loss correction techniques, have demonstrated success in utilizing noisy labels, they assume that noisy and clean samples (samples with correct annotations) share the same label space.However, this assumption does not hold in continual machine learning, where new categories and tasks emerge over time, leading to label shift problems that are specific to this setting. As a result, existing methods may struggle to accurately estimate the ground truth labels for noisy samples in such dynamic environments, potentially exacerbating label noise and further degrading performance. To address this critical gap, we propose a <strong>S</strong>hift-<strong>A</strong>daptive <strong>N</strong>oise <strong>U</strong>tilization (<strong>SANU</strong>) method, designed to transform samples with label noise into usable samples for continual learning. SANU introduces a novel source detection mechanism that identifies the appropriate label space for noisy samples, leveraging a meta-knowledge representation module to improve the generalization of the detection process. By re-annotating noisy samples through label guessing and label generation strategies, SANU adapts to label shifts, turning noisy data into useful inputs for training. Experimental results across three continual learning datasets demonstrate that SANU effectively mitigates the label shift problem, significantly enhancing model performance by utilizing re-annotated samples with label noise.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107422"},"PeriodicalIF":6.0,"publicationDate":"2025-03-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143759540","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":"CCA: Contrastive cluster assignment for supervised and semi-supervised medical image segmentation","authors":"Jinghua Zhu ,&nbsp;Chengying Huang ,&nbsp;Heran Xi ,&nbsp;Hui Cui","doi":"10.1016/j.neunet.2025.107415","DOIUrl":"10.1016/j.neunet.2025.107415","url":null,"abstract":"<div><div>Transformers have shown great potential in vision tasks such as semantic segmentation. However, most of the existing transformer-based segmentation models neglect the cross-attention between pixel features and class features which impedes the application of transformers. Inspired by the concept of object queries in k-means Mask Transformer, we develop cluster learning and contrastive cluster assignment (CCA) for medical image segmentation in this paper. The cluster learning leverages the object queries to fit the feature-level cluster centers. The contrastive cluster assignment is introduced to guide the pixel class prediction using the cluster centers. Our method is a plug-in and can be integrated into any model. We design two networks for supervised segmentation tasks and semi-supervised segmentation tasks respectively. We equip the decoder with our proposed modules for the supervised segmentation to improve the pixel-level predictions. For the semi-supervised segmentation, we enhance the feature extraction capability of the encoder by using our proposed modules. We conduct comprehensive comparison and ablation experiments on public medical image datasets (ACDC, LA, Synapse, and ISIC2018), the results demonstrate that our proposed models outperform state-of-the-art models consistently, validating the effectiveness of our proposed method. The source code is accessible at <span><span>https://github.com/zhujinghua1234/CCA-Seg</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107415"},"PeriodicalIF":6.0,"publicationDate":"2025-03-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143714781","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":"SuperM2M: Supervised and mixture-to-mixture co-learning for speech enhancement and noise-robust ASR","authors":"Zhong-Qiu Wang","doi":"10.1016/j.neunet.2025.107408","DOIUrl":"10.1016/j.neunet.2025.107408","url":null,"abstract":"<div><div>The current dominant approach for neural speech enhancement is based on supervised learning by using simulated training data. The trained models, however, often exhibit limited generalizability to real-recorded data. To address this, this paper investigates training enhancement models directly on real target-domain data. We propose to adapt mixture-to-mixture (M2M) training, originally designed for speaker separation, for speech enhancement, by modeling multi-source noise signals as a single, combined source. In addition, we propose a co-learning algorithm that improves M2M with the help of supervised algorithms. When paired close-talk and far-field mixtures are available for training, M2M realizes speech enhancement by training a deep neural network (DNN) to produce speech and noise estimates in a way such that they can be linearly filtered to reconstruct the close-talk and far-field mixtures. This way, the DNN can be trained directly on real mixtures, and can leverage close-talk and far-field mixtures as a weak supervision to enhance far-field mixtures. To improve M2M, we combine it with supervised approaches to co-train the DNN, where mini-batches of real close-talk and far-field mixture pairs and mini-batches of simulated mixture and clean speech pairs are alternately fed to the DNN, and the loss functions are respectively (a) the mixture reconstruction loss on the real close-talk and far-field mixtures and (b) the regular enhancement loss on the simulated clean speech and noise. We find that, this way, the DNN can learn from real and simulated data to achieve better generalization to real data. We name this algorithm SuperM2M (supervised and mixture-to-mixture co-learning). Evaluation results on the CHiME-4 dataset show its effectiveness and potential.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107408"},"PeriodicalIF":6.0,"publicationDate":"2025-03-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143714780","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":"FE reduced-order model-informed neural operator for structural dynamic response prediction","authors":"Lai-Hao Yang ,&nbsp;Xu-Liang Luo ,&nbsp;Zhi-Bo Yang ,&nbsp;Chang-Feng Nan ,&nbsp;Xue-Feng Chen ,&nbsp;Yu Sun","doi":"10.1016/j.neunet.2025.107437","DOIUrl":"10.1016/j.neunet.2025.107437","url":null,"abstract":"<div><div>Physics-Informed Neural Networks (PINN) have achieved remarkable advancements in recent years and have been extensively used in solving differential equations across various disciplines. However, when predicting structural dynamic responses, directly applying them to solve partial differential equations of structural dynamic models encounters challenges like inadequate result accuracy, inefficient training processes, and limited versatility. Furthermore, embedding large-scale structural dynamic models as physical constraints for neural networks can lead to poor trainability and low precision accuracy. To address the above issues, in this paper, we propose a novel FE reduced-order model-informed neural operator (FRINO) for structural dynamic response prediction with high precision, low computational cost, and broad versatility. Specifically, the Fourier neural operator (FNO) is employed to capture the dominant features of structural dynamic responses in the frequency domain, facilitating accurate and efficient solutions. Additionally, a reduced-order model derived using proper orthogonal decomposition is integrated to constrain the FNO. This ensures that the predicted solutions conform to physical differential equations, while also mitigating the high computational costs typically associated with large-dimensional physical equations. Special cantilever beam cases are designed to validate and evaluate the performance of the proposed FRINO. The comparative results demonstrate that FRINO can learn not only the responses of structural dynamic models but also the inherent dynamic characteristics of mechanical structure, allowing for precise predictions of structural responses under diverse unknown excitations. The results demonstrate that, compared with the PINN method, FRINO enhances prediction accuracy by up to two orders of magnitude and computation speed by up to three orders of magnitude. Besides, for practical use of FRINO, one should comprehensively consider the factors such as physical loss, training data resolution, and network width to obtain optimal performance of FRINO.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107437"},"PeriodicalIF":6.0,"publicationDate":"2025-03-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143737799","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":"Fixed-time synchronization of proportional delay memristive complex-valued competitive neural networks","authors":"Jiapeng Han,&nbsp;Liqun Zhou","doi":"10.1016/j.neunet.2025.107411","DOIUrl":"10.1016/j.neunet.2025.107411","url":null,"abstract":"<div><div>The fixed-time synchronization (FXS) is considered for memristive complex-valued competitive neural networks (MCVCNNs) with proportional delays. Two less conservative criteria supporting the FXS of MCVCNNs are founded by involving Lyapunov method and inequality techniques. Suitable switch controllers are designed by defining different norms of complex numbers instead of treating complex-valued neural networks as two real-valued systems. Furthermore, the settling time (ST) has been approximated. Finally, two simulations are shown to confirm the effectiveness of criteria in this paper and the outcomes of practical application in image protection.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107411"},"PeriodicalIF":6.0,"publicationDate":"2025-03-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143704895","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":"IBPL: Information Bottleneck-based Prompt Learning for graph out-of-distribution detection","authors":"Yanan Cao ,&nbsp;Fengzhao Shi ,&nbsp;Qing Yu ,&nbsp;Xixun Lin ,&nbsp;Chuan Zhou ,&nbsp;Lixin Zou ,&nbsp;Peng Zhang ,&nbsp;Zhao Li ,&nbsp;Dawei Yin","doi":"10.1016/j.neunet.2025.107381","DOIUrl":"10.1016/j.neunet.2025.107381","url":null,"abstract":"<div><div>When training and test graph samples follow different data distributions, graph out-of-distribution (OOD) detection becomes an indispensable component of constructing the reliable and safe graph learning systems. Motivated by the significant progress on prompt learning, graph prompt-based methods, which enable a well-trained graph neural network to detect OOD graphs without modifying any model parameters, have been a standard benchmark with promising computational efficiency and model effectiveness. However, these methods ignore the influence of overlapping features existed in both in-distribution (ID) and OOD graphs, which weakens the difference between them and leads to sub-optimal detection results. In this paper, we present the <strong>I</strong>nformation <strong>B</strong>ottleneck-based <strong>P</strong>rompt <strong>L</strong>earning (IBPL) to overcome this challenging problem. Specifically, IBPL includes a new graph prompt that jointly performs the mask operation on node features and the graph structure. Building upon this, we develop an information bottleneck (IB)-based objective to optimize the proposed graph prompt. Since the overlapping features are inaccessible, IBPL introduces the noise data augmentation which generates a series of perturbed graphs to fully covering the overlapping features. Through minimizing the mutual information between the prompt graph and the perturbed graphs, our objective can eliminate the overlapping features effectively. In order to avoid the negative impact of perturbed graphs, IBPL simultaneously maximizes the mutual information between the prompt graph and the category label for better extracting the ID features. We conduct experiments on multiple real-world datasets in both supervised and unsupervised scenarios. The empirical results and extensive model analyses demonstrate the superior performance of IBPL over several competitive baselines.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107381"},"PeriodicalIF":6.0,"publicationDate":"2025-03-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143714778","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":"AESeg: Affinity-enhanced segmenter using feature class mapping knowledge distillation for efficient RGB-D semantic segmentation of indoor scenes","authors":"Wujie Zhou ,&nbsp;Yuxiang Xiao ,&nbsp;Fangfang Qiang ,&nbsp;Xiena Dong ,&nbsp;Caie Xu ,&nbsp;Lu Yu","doi":"10.1016/j.neunet.2025.107438","DOIUrl":"10.1016/j.neunet.2025.107438","url":null,"abstract":"<div><div>Recent advances in deep learning for semantic segmentation models have introduced dynamic segmentation methods as opposed to static segmentation methods represented by full convolutional networks. Dynamic prediction methods replace static classifiers with learnable class embeddings to achieve global semantic awareness. Although dynamic methods excel in accuracy, the learning and inference of class embeddings is usually accompanied by a tedious computational burden. To address this challenge, we propose an affinity-enhanced semantic segmentation framework that synergistically combines the strengths of static and dynamic methodologies. Specifically, our approach leverages semantic features to obtain preliminary static segmentation results and constructs a binary affinity matrix that explicitly encodes pixel-wise category relationships. This affinity matrix serves as a dynamic classification kernel, effectively integrating global context awareness with static features, achieving comparable performance to purely dynamic approaches but with a substantially reduced computational overhead. Furthermore, we introduce a novel feature-to-category mapping refinement technique. This technique performs feature knowledge migration by learning a linear transformation between the semantic feature space and the segmentation probability space, resulting in improved accuracy without increasing model complexity. Numerous experiments demonstrated that the proposed method achieves the best performance on the widely used NYUv2 and SUN-RGBD datasets. And the effectiveness of our method in different scenes is verified on the outdoor scene dataset CamVid.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"188 ","pages":"Article 107438"},"PeriodicalIF":6.0,"publicationDate":"2025-03-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143759543","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":"Noise-resistant predefined-time convergent ZNN models for dynamic least squares and multi-agent systems","authors":"Yiwei Li ,&nbsp;Jiaxin Liu ,&nbsp;Lei Jia ,&nbsp;Liangze Yin ,&nbsp;Xingpei Li ,&nbsp;Yong Zhang","doi":"10.1016/j.neunet.2025.107412","DOIUrl":"10.1016/j.neunet.2025.107412","url":null,"abstract":"<div><div>Zeroing neural networks (ZNNs) are commonly used for dynamic matrix equations, but their performance under numerically unstable conditions has not been thoroughly explored, especially in situations involving unequal row-column matrices. The challenge is further aggravated by noise, particularly in dynamic least squares (DLS) problems. To address these issues, we propose the QR decomposition-driven noise-resistant ZNN (QRDN-ZNN) model, specifically designed for DLS problems. By integrating QR decomposition into the ZNN framework, QRDN-ZNN enhances numerical stability and guarantees both precise and rapid convergence through a novel activation function (N-Af). As validated by theoretical analysis and experiments, the model can effectively counter disturbances and enhance solution accuracy in dynamic environments. Experimental results show that, in terms of noise resistance, the QRDN-ZNN model outperforms existing mainstream ZNN models, including the original ZNN, integral-enhanced ZNN, double-integral enhanced ZNN, and super-twisting ZNN. Furthermore, the N-Af offers higher accuracy and faster convergence than other state-of-the-art activation functions. To demonstrate the practical utility of the method, We develop a new noise-resistant consensus protocol inspired by QRDN-ZNN, which enables multi-agent systems to reach consensus even in noisy conditions.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"187 ","pages":"Article 107412"},"PeriodicalIF":6.0,"publicationDate":"2025-03-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143696832","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}