Neural Networks最新文献

{"title":"LUFormer : A luminance-informed localized transformer with frequency augmentation for nighttime flare removal","authors":"Wei Lu ,&nbsp;He Zhao ,&nbsp;Dubuke Ma ,&nbsp;Peiguang Jing","doi":"10.1016/j.neunet.2025.107660","DOIUrl":"10.1016/j.neunet.2025.107660","url":null,"abstract":"<div><div>Flare caused by unintended light scattering or reflection in night scenes significantly degrades image quality. Existing methods explore frequency factors and semantic priors but fail to comprehensively integrate all relevant information. To address this, we propose LUFormer, a luminance-informed Transformer network with localized frequency augmentation. Central to our approach are two key modules: the luminance-guided branch (LGB) and the dual domain hybrid attention (DDHA) unit. The LGB provides global brightness semantic priors, emphasizing the disruption of luminance distribution caused by flare. The DDHA improves deep flare representation in both the spatial and frequency domains. In the spatial domain, it broadens the receptive field through pixel rearrangement and cross-window dilation, while in the frequency domain, it emphasizes and amplifies low-frequency components via a compound attention mechanism. Our approach leverages the LGB, which globally guides semantic refinement, to construct a U-shaped progressive focusing framework. In this architecture, the DDHA locally augments multi-domain features across multiple scales. Extensive experiments on real-world benchmarks demonstrate that the proposed LUFormer outperforms state-of-the-art methods. The code is publicly available at: https://github.com/HeZhao0725/LUFormer.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107660"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144271955","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":"Predicting the directed acyclic graph based on feature extraction","authors":"Qiying Wu ,&nbsp;Huiwen Wang","doi":"10.1016/j.neunet.2025.107661","DOIUrl":"10.1016/j.neunet.2025.107661","url":null,"abstract":"<div><div>Directed acyclic graphs (DAGs) are important tools for causal discovery. However, the existing methods mainly focus on estimating DAGs from observed cross-sectional or time series data, and less attention is given to the prediction of DAGs. We introduce a novel DAG prediction method that transforms the DAG prediction problem into a matrix prediction problem. This approach obtains causal order and conditional independence information by extracting the demixing matrices and correlation coefficient matrices at different time points and predicts future DAGs by modeling these matrices. This method provides a versatile framework that can be adapted to include a range of time series forecasting techniques according to specific needs. Numerical simulations demonstrate the effectiveness of the proposed method in terms of predicting both feature matrices and the final DAG. A real-world application involving financial market data successfully predicts risk spillover relationship changes. The flexibility of the method and its ability to forecast the future relationships between variables have significant implications for fields such as economics, management, and social sciences.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107661"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144239525","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":"Enhancing partition distinction: A contrastive policy to recommendation unlearning","authors":"Lin Li ,&nbsp;Shengda Zhuo ,&nbsp;Hongguang Lin ,&nbsp;Jinchun He ,&nbsp;Wangjie Qiu ,&nbsp;Qinnan Zhang ,&nbsp;Changdong Wang ,&nbsp;Shuqiang Huang","doi":"10.1016/j.neunet.2025.107667","DOIUrl":"10.1016/j.neunet.2025.107667","url":null,"abstract":"<div><div>With the growing privacy and data contamination concerns in recommendation systems, recommendation unlearning, <em>i.e.</em>, unlearning the impact of specific learned data, has garnered more attention. Unfortunately, existing research primarily focuses on the complete unlearning of target data, neglecting the balance between unlearning integrity, practicality, and efficiency. Two major restrictions hinder the widespread application of this unlearning paradigm in practice. First, while prior studies often assume consistent similarity among samples, they overly emphasize the local collaborative relationships between samples and central nodes, leading to an imbalance between local and global collaborative information. Second, while data partition appears to be a default setup, this evidently exacerbates the sparsity of recommendation data, which can have a potentially negative impact on recommendation quality. To fill these gaps, this paper proposes a data partitioning and submodel training strategy, named <em>Partition Distinction with Contrastive Recommendation Unlearning</em> (PDCRU), which aims to balance data partitioning and feature sparsity. The key idea is to extract structural features as global collaborative information for samples and introduce structural feature constraints based on sample similarity during the partitioning process. For submodel training, we leverage contrastive learning to introduce additional high-quality training signals to enhance model embeddings. Extensive experiments validate the feasibility and consistent superiority of our method over existing recommendation unlearning models in learning and unlearning. Specifically, our model achieves a 4.83% improvement in performance and a 4.64x enhancement in unlearning efficiency compared to baseline methods. The code is released at <span><span>https://github.com/linli0818/PDCRU</span><svg><path></path></svg></span></div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107667"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144263866","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":"LaTP: LiDAR-aided multimodal token pruning for efficient trajectory prediction of autonomous driving","authors":"Yantao Lu ,&nbsp;Shiqi Sun ,&nbsp;Ning Liu ,&nbsp;Bo Jiang ,&nbsp;Yilan Li ,&nbsp;Jinchao Chen ,&nbsp;Ying Zhang ,&nbsp;Yichen Zhu ,&nbsp;Senem Velipasalar","doi":"10.1016/j.neunet.2025.107673","DOIUrl":"10.1016/j.neunet.2025.107673","url":null,"abstract":"<div><div>The rapid advancement of Large Vision Language Models (LVLMs) has spurred significant progress in autonomous driving, especially in end-to-end trajectory prediction, which is crucial for enabling autonomous driving across diverse traffic scenarios. Nevertheless, the onboard computational requirements of autonomous vehicles present challenges for deploying LVLMs on resource-constrained devices, as they demand substantial processing power. Token pruning is one of the most promising approach that achieves considerable inference speed gains without requiring additional model training. While token pruning has demonstrated its efficacy in various domains, it appears that the current approaches are designed for generalized tasks and have not been tailored to address the unique demands of trajectory prediction in autonomous driving. Specifically, within the context of trajectory prediction of autonomous driving, there are two considerations that have not been adequately addressed: (i) content information, where irrelevant visual elements, despite their complex features, cannot be pruned effectively due to their non-trivial appearance; (ii) distance information, which is critical for accurate trajectory prediction but often overlooked by conventional pruning approaches. As a result, directly applying existing pruning methods to LVLMs without considering these crucial differences may lead to a degradation in performance. To overcome these challenges, we propose a novel token pruning method, LiDAR-aided Token Prune (LaTP), specifically designed for LVLM-based trajectory prediction in autonomous driving. LaTP efficiently integrates LiDAR points to provide distance information for camera inputs and uses a content- and distance-aware token importance indicator to discard visual tokens that are inconsequential for driving. This approach significantly improves inference speed without compromising control accuracy. Experiments on the nuScenes dataset validate the effectiveness of our method, showing superior performance compared to general token pruning baselines. Specifically, LaTP achieves a pruning ratio of up to 75% while maintaining an Average Displacement Error (ADE) of 2.03 meters and a Collision Rate (col.) of 2.35%, demonstrating its ability to significantly reduce computational load without sacrificing prediction accuracy.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107673"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144263668","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":"Rect-ViT: Rectified attention via feature attribution can improve the adversarial robustness of Vision Transformers","authors":"Xu Kang,&nbsp;Bin Song","doi":"10.1016/j.neunet.2025.107666","DOIUrl":"10.1016/j.neunet.2025.107666","url":null,"abstract":"<div><div>Deep neural networks (DNNs) have suffered from input perturbations and adversarial examples (AEs) for a long time, mainly caused by the distribution difference between robust and non-robust features. Recent research shows that Vision Transformers (ViTs) are more robust than traditional convolutional neural networks (CNNs). We studied the relationship between the activation distribution and robust features in the attention mechanism in ViTs, coming up with a discrepancy in the token distribution between natural and adversarial examples during adversarial training (AT). When predicting AEs, some tokens irrelevant to the targets are still activated, giving rise to the extraction of non-robust features, which reduces the robustness of ViTs. Therefore, we propose Rect-ViT, which can rectify robust features based on class-relevant gradients. Performing the relevance back-propagation of auxiliary tokens during forward prediction can achieve rectification and alignment of token activation distributions, thereby improving the robustness of ViTs during AT. The proposed rectified attention mechanism can be adapted to a variety of mainstream ViT architectures. Along with traditional AT, Rect-ViT can also be effective in other AT modes like TRADES and MART, even for state-of-the-art AT approaches. Experimental results reveal that Rect-ViT improves average robust accuracy by 0.64% and 1.72% on CIFAR10 and Imagenette against four classic attack methods. These modest gains have significant practical implications in safety-critical applications and suggest potential effectiveness for complex visual tasks.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107666"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144279370","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":"Dynamic periodic event-triggered control of stochastic complex networks with time-varying delays","authors":"Xuetao Yang ,&nbsp;Anjie Li ,&nbsp;Quanxin Zhu","doi":"10.1016/j.neunet.2025.107659","DOIUrl":"10.1016/j.neunet.2025.107659","url":null,"abstract":"<div><div>This paper focuses on the exponential stabilization of stochastic complex network systems with time-varying delays. A novel dynamic periodic event-triggered control (ETC) with the graph theory and the Lyapunov–Razumikhin method is proposed. First, different from continuous ETCs, periodic sampling inherently prevents the occurrence of Zeno phenomenon. Second, compared with the traditional static ETCs, our dynamic ETC can reduce the update frequency of the controller and save communication resources by designing a proper dynamic function. Moreover, the graph theory is employed to handle the coupling relationships among nodes in complex networks and the Lyapunov–Razumikhin method is employed to address the difficulties caused by time delays in stochastic complex systems. Then, the mean-square exponential stabilization for stochastic complex network systems is obtained. Finally, a numerical example of single-link robot arm with multiple nodes is performed in a stochastic complex network system to verify the effectiveness of theoretical results.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107659"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144239874","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":"HCD: A Hierarchy Constraint-Aware Neural Cognitive Diagnosis Framework","authors":"Xinjie Sun ,&nbsp;Qi Liu ,&nbsp;Kai Zhang ,&nbsp;Shuanghong Shen ,&nbsp;Fei Wang ,&nbsp;Yan Zhuang ,&nbsp;Zheng Zhang ,&nbsp;Weiyin Gong ,&nbsp;Shijin Wang ,&nbsp;Lina Yang ,&nbsp;Xingying Huo","doi":"10.1016/j.neunet.2025.107668","DOIUrl":"10.1016/j.neunet.2025.107668","url":null,"abstract":"<div><div>Cognitive diagnosis (CD) aims to reveal students’ proficiency in specific knowledge concepts. With the increasing adoption of intelligent education applications, accurately assessing students’ knowledge mastery has become an urgent challenge. Although existing cognitive diagnosis frameworks enhance diagnostic accuracy by analyzing students’ explicit response records, they primarily focus on individual knowledge state, failing to adequately reflect the relative ability performance of students within hierarchies. To address this, we propose the <em><strong>H</strong>ierarchy Constraint-Aware Neural <strong>C</strong>ognitive <strong>D</strong>iagnosis Framework (<strong>HCD</strong>),</em> designed to more accurately represent student ability performance within real educational contexts. Specifically, the framework introduces a hierarchy mapping layer to identify students’ levels. It then employs a hierarchy convolution-enhanced attention layer for in-depth analysis of knowledge concepts performance among students at the same level, uncovering nuanced differences. A hierarchy inter-sampling attention layer captures performance differences across hierarchies, offering a comprehensive understanding of the relationships among students’ knowledge state. Finally, through personalized diagnostic enhancement, the framework seamlessly integrates hierarchy constraint-aware features with existing typical diagnostic methods, significantly improving the precision of student knowledge state representation and enhancing the adaptability and diagnostic performance of existing frameworks. Research shows that this framework not only reasonably constrains changes in students’ knowledge state to align with real educational contexts, but also supports the scientific rigor and fairness of educational assessments, thereby advancing the field of cognitive diagnosis. To support reproducible research, we have published the data and code at <span><span>https://github.com/xinjiesun-ustc/HCD</span><svg><path></path></svg></span>, encouraging further innovation in this field.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107668"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144239872","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":"BEVFix: Deep feature enhancement for robust 3D object detection","authors":"Wenxuan Li ,&nbsp;Jian Zhou ,&nbsp;Chi Chen ,&nbsp;Hongkai Yu ,&nbsp;Bo Du ,&nbsp;Qin Zou","doi":"10.1016/j.neunet.2025.107675","DOIUrl":"10.1016/j.neunet.2025.107675","url":null,"abstract":"<div><div>Recent advancements in Bird’s Eye View (BEV)-based 3D object detection have highlighted its potential to enhance scene understanding in autonomous driving applications. However, existing BEV-based methods utilizing point clouds for 3D object detection face significant challenges due to inherent sparsity and noise, which often compromise the accuracy of BEV representations. Furthermore, in multimodal 3D object detection, the lack of depth information in images can lead to distortions in the image BEV features generated through view transformations, further leading to inaccuracies in the fused BEV representation. To overcome these limitations, we introduce BEVFix, an innovative end-to-end 3D object detection method designed to refine BEV representations. BEVFix starts by generating a mask based on the point cloud distribution to identify specific regions requiring repair. This is followed by our WaveRefiner, which employs Discrete Wavelet Transform (DWT) for multi-frequency decomposition and utilizes a Feed-Forward Network (FFN) to isolate noise while selectively retaining critical features. These components work synergistically to reduce noise and enhance BEV representations. Experiments on the nuScenes and Waymo datasets demonstrate that BEVFix significantly improves performance, achieving state-of-the-art results. The source code will be publicly available at <span><span>https://github.com/WenxuanLi-whu/Co-Fix3d</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107675"},"PeriodicalIF":6.0,"publicationDate":"2025-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144253456","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 generalized defect-data-free defect inspection method based on image reconstruction and anomaly detection","authors":"Minjie Du ,&nbsp;Siqi Gu ,&nbsp;Zihan Qin ,&nbsp;Lizhe Xie ,&nbsp;Zheng Wang ,&nbsp;Yining Hu","doi":"10.1016/j.neunet.2025.107662","DOIUrl":"10.1016/j.neunet.2025.107662","url":null,"abstract":"<div><div>This paper presents a novel framework based on hierarchical image reconstruction, employing image reconstruction and anomaly detection techniques. Unlike traditional supervised methods, our approach operates without the need for defect-specific training data, enabling generalization across diverse product types. Using hierarchical reconstruction modules and a self-attention mechanism, our method achieves an average precision of 97.83% on the MVTec AD 2D dataset, surpassing the U-Net model by 11.1% and the U-Transformer by 12.9%. Furthermore, the model inference speed reaches 24.1 FPS, representing a 48.1% increase over U-Transformer models. These results demonstrate the framework’s effectiveness in enhancing both detection accuracy and speed, providing a robust solution for real-time industrial defect inspection.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107662"},"PeriodicalIF":6.0,"publicationDate":"2025-06-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144263670","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 receptive field graph neural networks","authors":"Hepeng Gao ,&nbsp;Funing Yang ,&nbsp;Yongjian Yang ,&nbsp;Yuanbo Xu ,&nbsp;Yijun Su","doi":"10.1016/j.neunet.2025.107658","DOIUrl":"10.1016/j.neunet.2025.107658","url":null,"abstract":"<div><div>Graph Neural Networks (GNNs) have drawn increasing attention in recent years and achieved outstanding success in many scenarios and tasks. However, existing methods indicate that the performance of representation learning drops dramatically as GNNs deepen, which is attributed to <strong>over-smoothing representation</strong>. To handle the above issue, we propose an adaptive receptive field graph neural network (ADRP-GNN) that aggregates information by adaptively expanding receptive fields with a monolayer graph convolution layer, avoiding deepening to result in the over-smoothing issue. Specifically, we first present a Multi-hop Graph Convolution Network (MuGC) that captures the information of the nodes and their multi-hop neighbors with only one layer, preventing frequent passing messages between nodes from the over-smoothing issue. Then, we design a Meta Learner that realizes the adaptive receptive field for each node to select related neighbor information. Finally, a Backbone Network is employed to enhance the architecture’s learning ability. In addition, our architecture adaptively generates receptive fields instead of handcrafting stacked layers, which can integrate existing GNN frameworks to fit various scenarios. Extensive experiments indicate that our architecture is effective for the over-smoothing issue and improves accuracy by 0.52% to 6.88% compared to state-of-the-art methods on node classification tasks on eight datasets.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107658"},"PeriodicalIF":6.0,"publicationDate":"2025-06-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144253457","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}