Neural Networks最新文献

{"title":"MIIGAN: Mambas make strong GAN for infrared image generation","authors":"Fuchao Wang ,&nbsp;Huaici Zhao ,&nbsp;Yuhuai Peng ,&nbsp;Jian Fang ,&nbsp;Pengfei Liu ,&nbsp;Ronghua Zhang","doi":"10.1016/j.neunet.2025.108021","DOIUrl":"10.1016/j.neunet.2025.108021","url":null,"abstract":"<div><div>Collecting infrared images on-site is the most direct and realistic approach. However, it is costly, and due to varying environmental conditions, replicating the same conditions for comparative experiments is challenging. This presents significant obstacles for research in infrared technology. To target this issue, we propose MIIGAN, a Visible-to-Infrared Image Generation model that achieves SOTA performance. MIIGAN employs a GAN based on U-Net, with Mamba blocks serving as the core module to improve generation quality. Additionally, we develop a Spatial and Channel Attention Module (SCAM) and integrate it into the skip connections of U-Net to enhance feature extraction. We also design a Dual-encoder combining Transformer and Mamba to improve the discriminator’s performance. Furthermore, we introduce the Difference and Product learning Module (DPM) into the Dual-encoder to enhance differential and consistency feature extraction. Finally, we integrate multi-layer feature differential and consistency losses into the objective function of the discriminator, providing comprehensive pixel-level feedback across multiple scales. We conduct extensive comparative and ablation studies across four datasets and perform downstream object detection tasks on the generated infrared images to validate MIIGAN’s performance. The source code is available at <span><span>https://github.com/wangfc0913/miigan.git</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 108021"},"PeriodicalIF":6.3,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144902652","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":"Rethinking softmax in incremental learning","authors":"Zheng Zhai ,&nbsp;Jiali Zhang ,&nbsp;Haiyu Wang ,&nbsp;Mingxin Wu ,&nbsp;Keshun Yang ,&nbsp;Xiaoyan Qiao ,&nbsp;Qiang Sun","doi":"10.1016/j.neunet.2025.108017","DOIUrl":"10.1016/j.neunet.2025.108017","url":null,"abstract":"<div><div>Mitigating catastrophic forgetting remains a fundamental challenge in incremental learning. This paper identifies a key limitation of the widely used softmax cross-entropy loss: the non-identifiability inherent in the standard softmax cross-entropy distillation loss. To address this issue, we propose two complementary strategies: (1) adopting an imbalance-invariant distillation loss to mitigate the adverse effect of imbalanced weights during distillation, and (2) regularizing the original prediction/distillation loss with shift-sensitive alternatives, which render the optimization problem identifiable and proactively prevent imbalance from arising. These strategies form the foundation of five novel approaches that can be seamlessly integrated into existing distillation-based incremental learning frameworks such as LWF, LWM, and LUCIR.</div><div>We validate the effectiveness of our approaches through extensive numerical experiments, demonstrating consistent improvements in predictive accuracy and substantial reductions in forgetting. For example, in a 10-task incremental learning setting on CIFAR-100, our methods improve the average accuracy of three widely used approaches - LWF, LWM, and LUCIR - by 11.8 %, 11.5 %, and 12.8 %, respectively, while reducing their average forgetting rates by 16.5 %, 16.8 %, and 13.8 %, respectively. Our code is publicly available at <span><span>https://github.com/nexais/RethinkSoftmax</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 108017"},"PeriodicalIF":6.3,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144919660","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":"Spatial-frequency domain aggregation upsampling for pan-sharpening","authors":"Yilong Liu ,&nbsp;Kai Sun ,&nbsp;Yuan Liu ,&nbsp;Junying Hu ,&nbsp;Junmin Liu ,&nbsp;Jiangshe Zhang","doi":"10.1016/j.neunet.2025.108007","DOIUrl":"10.1016/j.neunet.2025.108007","url":null,"abstract":"<div><div>Pan-sharpening, fusing high-resolution panchromatic (PAN) images with low-resolution multispectral (LRMS) to generate high-resolution multispectral (HRMS) images, is critical for enhancing remote sensing image quality. Despite significant advancements in deep learning methods, research on the image upsampling process remains limited. Existing approaches either fail to effectively utilize the information from PAN images or struggle to balance spectral and spatial information, thereby constraining the performance of these models. To alleviate these problems, we propose a novel Spatial-Frequency Domain Aggregation Upsampling (SFAU) method. Our method consists of three core modules: the Dual-Domain Nonlinear Fusion (DDNF), Region-Specific Attention Mechanism (RSAM), and Adaptive Feature Fusion Gate (AFFG). The DDNF module integrates Frequency-Aware Feature Aggregation (FAFA) and Spatial Domain Enhancement techniques, enabling the capture of high-frequency features while refining local structural details. The RSAM module adaptively refines feature representations and preserves spatial-spectral correlations. Finally, the AFFG module effectively combines the outputs from the DDNF and RSAM modules, ensuring a balanced integration of spatial and spectral information. Extensive experiments demonstrate that our method outperforms other popular upsampling techniques and significantly enhances the performance of many leading pan-sharpening models, particularly in high-contrast and spectrally complex regions. Additionally, our approach shows strong generalization in real-world scenarios, highlighting its potential for practical remote sensing applications. Code is available at <span><span>https://github.com/zacianfans/SFAU</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 108007"},"PeriodicalIF":6.3,"publicationDate":"2025-08-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144912897","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":"UTN: Unsupervised optical flow estimation network based on transformer","authors":"Xiaochen Liu ,&nbsp;Tao Zhang ,&nbsp;Mingming Liu","doi":"10.1016/j.neunet.2025.108015","DOIUrl":"10.1016/j.neunet.2025.108015","url":null,"abstract":"<div><div>With the aim of enabling unsupervised optical flow estimation, we propose a scalable framework based on a transformer and a feature pyramid network (FPN). Central to our approach is the incorporation of a transformer-CNN based structure within the encoder, designed to capture global and local dependency features from input image pairs—a crucial element for precise pixel-wise flow estimation. Subsequently, we integrate a normalized cross-correlation module (NCCM) and an attention-based intermediate flow estimation (AIFE) module into the FPN-based decoder. The NCCM enhances the decoder's focus on the saliency of shared foreground objects through correlation operations, while the AIFE refines flow estimation using an auxiliary positional mask and intermediate flow matrix. Furthermore, we propose a static optical flow loss, providing a distinct training clue that effectively boosts flow accuracy. Comprehensive experiments, including comparisons with state-of-the-art methods and ablation studies, were conducted across benchmark datasets such as FlyingChairs, MPI-Sintel, KITTI-2012, and KITTI-2015. Notably, our method achieved substantial performance gains. For instance, on the MPI-Sintel dataset, we observed a reduction in End-Point-Error (EPE) of 24.27 % on the clean dataset and 28.01 % on the final dataset compared to ARFlow. Ablation studies corroborated the efficacy of the NCCM, AIFE, and static optical flow loss in enhancing estimation accuracy.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 108015"},"PeriodicalIF":6.3,"publicationDate":"2025-08-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144893601","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":"Joint noise detection and L2,p-norm metric in least squares twin SVM for robust multiclass classification","authors":"Chao Yuan ,&nbsp;Xiaoyuan Xu ,&nbsp;Farshad Arvin ,&nbsp;Huiyu Mu ,&nbsp;Haiyang Li ,&nbsp;Jigen Peng","doi":"10.1016/j.neunet.2025.107991","DOIUrl":"10.1016/j.neunet.2025.107991","url":null,"abstract":"<div><div>The least squares twin support vector machine (LSTSVM) serves as a foundational framework for binary classification and is widely applied in statistical learning due to its solid theoretical foundation. It also plays a crucial role in advancing research in multiclass classification. However, the presence of noise in real-world datasets often leads to substantial performance degradation, compromising the reliability and generalizability of this model. Given the ubiquitous presence of noise, its influence on the learning of classification hyperplanes warrants rigorous attention. In this paper, we propose a robust multiclass classification model grounded in LSTSVM, designed to mitigate the influence of noisy data. The proposed framework replaces the conventional squared <span><math><msub><mi>L</mi><mn>2</mn></msub></math></span>-norm with the more robust <span><math><msub><mi>L</mi><mrow><mn>2</mn><mo>,</mo><mi>p</mi></mrow></msub></math></span>-norm (<span><math><mrow><mn>0</mn><mo>&lt;</mo><mi>p</mi><mo>≤</mo><mn>2</mn></mrow></math></span>), which enhances resilience against noise. Furthermore, we introduce an innovative noise detection mechanism with a transparent physical interpretation, whereby a probabilistic weight is assigned to each sample to quantify its likelihood of being a normal observation. Specifically, normal samples receive a weight of 1, whereas suspected noisy samples receive a weight of 0. To solve the resulting non-convex optimization problem efficiently, we develop an iterative algorithm that adaptively penalizes normal samples exhibiting substantial errors. The convergence property of the algorithm is rigorously analyzed and theoretically supported. Moreover, the model is extended to semi-supervised learning, enabling the effective exploitation of both a limited set of labeled samples and the structural information inherent in numerous unlabeled samples. Finally, extensive experiments on benchmark and image datasets under varying noise levels demonstrate that the proposed approach consistently outperforms existing methods in terms of classification accuracy and robustness, validating its practical effectiveness in noisy multiclass settings.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 107991"},"PeriodicalIF":6.3,"publicationDate":"2025-08-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144902651","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":"Granular ball twin support vector machine with Universum data","authors":"M.A. Ganaie,&nbsp;Vrushank Ahire","doi":"10.1016/j.neunet.2025.107974","DOIUrl":"10.1016/j.neunet.2025.107974","url":null,"abstract":"<div><div>Support vector machines often underperform when limited to labelled target class data and demonstrate sensitivity to noise and outliers. To address these limitations, we propose the Granular Ball Twin Support Vector Machine with Universum Data (GBU-TSVM), which uniquely integrates Universum samples with granular ball computing in the TSVM framework. Unlike conventional TSVMs representing data as points in feature space, the proposed GBU-TSVM models instances as hyperballs, significantly improving robustness against noise while enhancing computational efficiency. Granular representation enables effective data grouping, reducing processing complexity while preserving critical structural information. Incorporating Universum data, consisting of samples outside the target classes, provides additional contextual information that refines decision boundaries and improves generalization. Experiments on UCI benchmark datasets demonstrate GBU-TSVM’s superior performance, measured in terms of accuracy and training time. It achieves 92.38 % accuracy on the Molec Biol Promoter dataset under optimal conditions and maintains 89.17 % accuracy even with 20 % noise contamination. It consistently outperforms baseline models such as GBSVM, TSVM, GBTSVM, Pin-GTSVM, and UTSVM. These results establish GBU-TSVM as an advanced framework for robust classification in challenging data environments.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 107974"},"PeriodicalIF":6.3,"publicationDate":"2025-08-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144926175","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":"Dual-channel hierarchical interactive learning for the prediction of Protein-Ligand binding affinity","authors":"Zheyu Wu ,&nbsp;Huifang Ma ,&nbsp;Bin Deng ,&nbsp;Zhixin Li ,&nbsp;Liang Chang","doi":"10.1016/j.neunet.2025.107982","DOIUrl":"10.1016/j.neunet.2025.107982","url":null,"abstract":"<div><div>Protein-ligand binding affinity (PLBA) is a crucial metric in drug screening for identifying potential candidate compounds. In recent years, deep learning-based methods have used representation learning to model interactions within protein-ligand complexes, demonstrating great promise in affinity prediction tasks. Existing studies have considered both intramolecular (covalent) and intermolecular (non-covalent) interactions to some extent. However, these interactions are often treated as independent features, lacking explicit hierarchical dependency modeling, which may lead to insufficient representation of interaction information and ultimately limit the accuracy of affinity predictions. To address this issue, we propose a novel approach—Dual-channel Hierarchical Interactive Learning (DHIL)—to achieve a more comprehensive modeling of protein-ligand interactions. DHIL employs a dual-channel encoding structure to simultaneously learn intramolecular and intermolecular interactions, ensuring the completeness of interaction features. Additionally, we design a hierarchical interactive learning paradigm to facilitate information exchange between these two interaction types at multiple levels, promoting their collaborative modeling. This mechanism mimics the local-to-global working principles of biological systems, enabling a more detailed and holistic representation of protein-ligand interactions. We conduct extensive and comprehensive experiments on a diverse set of benchmark datasets, rigorously evaluating the effectiveness of DHIL. The results demonstrate that DHIL significantly improves PLBA prediction accuracy, outperforming existing methods and further validating its potential in drug discovery and screening tasks.</div><div>Nevertheless, the proposed framework introduces notable computational overhead due to multi-scale graph construction and cross-level message passing. It also exhibits sensitivity to the quality of input 3D binding conformations, which may affect its robustness in practical applications. These limitations suggest future directions for improving model efficiency and generalizability. To facilitate reproducibility and further research, the complete source code of DHIL has been released at: <span><span>https://github.com/WZY-0814/DHIL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 107982"},"PeriodicalIF":6.3,"publicationDate":"2025-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144879968","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":"Enabling generalized zero-shot learning towards unseen domains by intrinsic learning from redundant LLM semantics","authors":"Jiaqi Yue ,&nbsp;Chunhui Zhao ,&nbsp;Jiancheng Zhao ,&nbsp;Biao Huang","doi":"10.1016/j.neunet.2025.107997","DOIUrl":"10.1016/j.neunet.2025.107997","url":null,"abstract":"<div><div>Generalized zero-shot learning (GZSL) focuses on recognizing seen and unseen classes against domain shift problem where data of unseen classes may be misclassified as seen classes. However, existing GZSL is still limited to seen domains. In the current work, we study cross-domain GZSL (CDGZSL) which addresses GZSL towards unseen domains. Different from existing GZSL methods, CDGZSL constructs a common feature space across domains and acquires the corresponding intrinsic semantics shared among domains to transfer from seen to unseen domains. Considering the information asymmetry problem caused by redundant class semantics annotated with large language models (LLMs), we present Meta Domain Alignment Semantic Refinement (MDASR). Technically, MDASR consists of two parts: Inter-class similarity alignment, which eliminates the non-intrinsic semantics not shared across all domains under the guidance of inter-class feature relationships, and unseen-class meta generation, which preserves intrinsic semantics to maintain connectivity between seen and unseen classes by simulating feature generation. MDASR effectively aligns the redundant semantic space with the common feature space, mitigating the information asymmetry in CDGZSL. The effectiveness of MDASR is demonstrated on two public datasets, Office-Home and Mini-DomainNet, as well as on a self-constructed multi-domain rare animal dataset. We have shared the LLM-based semantics for these datasets as a benchmark.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 107997"},"PeriodicalIF":6.3,"publicationDate":"2025-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144903223","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":"Unified auxiliary restoration network for robust multimodal 3D object detection in adverse conditions","authors":"Jae Hyun Yoon ,&nbsp;Jong Won Jung ,&nbsp;Seok Bong Yoo","doi":"10.1016/j.neunet.2025.107992","DOIUrl":"10.1016/j.neunet.2025.107992","url":null,"abstract":"<div><div>The fusion of LiDAR and camera sensors offers remarkable results in multimodal 3D object detection with enhanced performance. However, existing fusion methods are primarily designed considering ideal data, ignoring the practical challenges of sensor specification and environmental variations encountered in autonomous driving. Thus, these methods often exhibit a significant performance degradation when faced with adverse conditions, such as sparse point cloud and inclement weather. To address these multiple adverse conditions simultaneously, we present the first attempt to apply auxiliary restoration networks in multimodal 3D object detection. These networks restore degraded point cloud and image, ensuring the primary multimodal detection network obtains higher quality features in a unified form. Especially, we propose a spherical domain point upsampler based on bilateral point generation and an adjustment network with a horizontal alignment block. Additionally, for efficient fusion with restored point cloud and image, we suggest a graph detector with a unified loss function, including auxiliary, contrastive, and difficulty losses. The experimental results demonstrate that the proposed approach prevents a performance decline in adverse conditions and outperforms state-of-the-art methods. The source code with pretrained weights for the proposed model is available at <span><span>https://github.com/jhyoon964/auxphere</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 107992"},"PeriodicalIF":6.3,"publicationDate":"2025-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144893604","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":"GraphGuard: An adaptive approach for restoring accuracy in backdoor-compromised GNNs","authors":"Adil Ahmad ,&nbsp;Anwar Shah ,&nbsp;Waleed Alnumay ,&nbsp;Bahar Ali","doi":"10.1016/j.neunet.2025.107990","DOIUrl":"10.1016/j.neunet.2025.107990","url":null,"abstract":"<div><div>Backdoor attacks present a significant threat to the reliability of machine learning models, including Graph Neural Networks (GNNs), by embedding triggers that manipulate model behavior. While many existing defenses focus on identifying these vulnerabilities, few address restoring model accuracy after an attack. This paper introduces a method for restoring the original accuracy of GNNs affected by backdoor attacks, a task complicated by the complex structure of graph data. Our approach combines advanced filtering and augmentation techniques that enhance the GNN’s resilience against hidden triggers. The filtering mechanisms remove suspicious data points to minimize the influence of poisoned inputs, while augmentation introduces controlled variation to strengthen the model against backdoor triggers. To optimize restoration, we present an adaptive framework that adjusts the balance between filtering and augmentation based on model sensitivity and attack severity, reducing both false positives and negatives. Additionally, we incorporate Explainable AI (XAI) techniques to improve the interpretability of the model’s decision-making process, enabling transparent detection and understanding of backdoor triggers. Results demonstrate that our method achieves an average accuracy restoration of 97–99 % across various backdoor attack scenarios, providing an effective solution to maintain the performance and integrity of GNNs in sensitive applications.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"193 ","pages":"Article 107990"},"PeriodicalIF":6.3,"publicationDate":"2025-08-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144902653","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}