Neural Networks最新文献

{"title":"ESCAN: Efficient GPU sharing for cascade neural network inference","authors":"Jianan Wang,&nbsp;Yang Shi,&nbsp;Zhaoyun Chen,&nbsp;Mei Wen","doi":"10.1016/j.neunet.2025.107703","DOIUrl":"10.1016/j.neunet.2025.107703","url":null,"abstract":"<div><div>Cascading, as a multi-model combination approach, balances model execution efficiency and accuracy. This excellent method is widely used in various industrial production and commercial deployments, particularly in cloud-based inference services. With the increasing demand for low-latency services, researchers are more focused on the execution efficiency of these models, especially device utilization. It is highly desirable to fully utilize GPU resources by multiplexing different inference tasks on the same GPU through device-sharing techniques such as Multiprocessing Services (MPS). However, we find it struggling when applying MPS to cascade neural networks consisting of multiple related submodels. These difficulties arise primarily from the early-exit mechanism and the execution order of the submodels. To address these obstacles, we analyzed the characteristics of cascade neural networks and combined them with device-sharing optimization techniques. Our findings indicate that improving the efficiency of cascade models through device sharing requires a balance between the gains from sharing devices and the potential wastage of computation resources due to the early-exit mechanism.</div><div>Based on our analysis, we proposed ESCAN, a GPU-sharing optimization framework for online inference of cascade neural networks. This framework includes exit-ratio-aware batch-parallel execution strategies and the corresponding resource allocation algorithms, all integrated into PyTorch. Experiments show that ESCAN improves inference efficiency by an average of 19.53% compared to the execution strategy with all cascade submodels running in parallel. Additionally, ESCAN significantly improves the efficiency of searching for computation resource allocation schemes. ESCAN optimizes the utilization of computational resources through effective GPU-sharing, greatly enhancing the efficiency of online inference for cascade models. This approach delivers a low-latency, high-precision optimization solution for interactive online services based on cascade neural networks.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107703"},"PeriodicalIF":6.0,"publicationDate":"2025-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144312738","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":"Foster noisy label learning by exploiting noise-induced distortion in foreground localization","authors":"Ang Chen ,&nbsp;Feng Xu ,&nbsp;Xin Lyu ,&nbsp;Tao Zeng ,&nbsp;Xin Li","doi":"10.1016/j.neunet.2025.107712","DOIUrl":"10.1016/j.neunet.2025.107712","url":null,"abstract":"<div><div>Large-scale, well-annotated datasets are crucial for training deep neural networks. However, the prevalence of noisy-labeled samples can cause irreversible impairment to the generalization of models. Existing approaches have attempted to mitigate the impact of noisy labels by exploiting the different loss or confidence distributions between clean and noisy data to detect and correct noisy labels. This paper investigates the noise-induced distorting effect on foreground localization by tracking the model’s spatial attention distribution on visual activation maps. We observe that in clean samples, highly responsive regions usually focus on label-relevant foreground regions, whereas in noisy data, the model accidentally attends to uninformative background regions or cluttered object edges due to interference from label noise. Inspired by the observations, we propose a novel two-stage foreground localization-augmented noisy label learning framework named FLSC to concurrently boost the accuracy of sample selection and label correction for robust training. Specifically, FLSC first quantifies noise-induced distortion in foreground localization to foster conventional loss-based selection criteria by calculating the information reduction when deriving the foreground images from the original images based on the attention distribution. Next, we propose a noise-adaptive adversarial erasing strategy, which suppresses background activation by imposing adaptive erasure regularization, to eliminate overfitting to noisy samples while enhancing the learning of robust representations. To the best of our knowledge, it is the first attempt to exploit localization quality evaluation based on feature activation to address the label noise problem. Extensive experiments on synthetic and real-world datasets validate the superior performance of FLSC compared to state-of-the-art methods.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107712"},"PeriodicalIF":6.0,"publicationDate":"2025-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144365715","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":"Adversarial regularized diffusion model for fair recommendations","authors":"Ran Yang ,&nbsp;Yihao Zhang ,&nbsp;Kaibei Li ,&nbsp;Qinyang He ,&nbsp;Xiaokang Li ,&nbsp;Wei Zhou","doi":"10.1016/j.neunet.2025.107695","DOIUrl":"10.1016/j.neunet.2025.107695","url":null,"abstract":"<div><div>With the widespread deployment of recommendation systems, concerns have grown over algorithmic fairness and representation bias in recommendation outcomes. Existing debiasing methods primarily suffer from two critical limitations: (1) Explicit feature removal strategies risk eliminating semantic signals entangled with sensitive attributes, inevitably degrading recommendation performance. (2) Conventional adversarial learning frameworks impose rigid gradient reversal to enforce independence from sensitive attributes, yet cause semantic distortion in latent representations through uncontrolled adversarial conflicts between fairness objectives and recommendation goals.</div><div>To address these challenges, we propose a fairness-aware recommendation framework leveraging the dynamic equilibrium of diffusion model. During the forward diffusion process, we introduce adaptive gradient-aware noise injection, where fairness discriminators from the reverse denoising process guide Gaussian perturbations through their aggregated gradient statistics, achieving feature-aware bias dissociation while preserving user interest semantics. The reverse denoising process employs adversarial regularization with sensitivity-aware gradient constraints, iteratively purifying recommendation-oriented embeddings through alternating optimization of denoising prediction and fairness discrimination tasks. To further enhance fairness-utility tradeoffs, we design an interest fusion mechanism at denoising initialization and develop a bias-controlled rounding function for candidate generation. Extensive experiments on three real-world datasets with sensitive attributes demonstrate that our model outperforms state-of-the-art methods in recommendation accuracy and fairness. We publish the source code at <span><span>https://github.com/YangRan993/DiffuFair</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107695"},"PeriodicalIF":6.0,"publicationDate":"2025-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144306860","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":"MTCM: Multi-context temporal consistent modeling for referring video object segmentation","authors":"Sun-Hyuk Choi,&nbsp;Hayoung Jo,&nbsp;Seong-Whan Lee","doi":"10.1016/j.neunet.2025.107701","DOIUrl":"10.1016/j.neunet.2025.107701","url":null,"abstract":"<div><div>Referring Video Object Segmentation (RVOS) focuses on segmenting objects in a video that is based on a provided text description. With recent advancements in transformers, many transformer-based RVOS methods have emerged to enhance interactions between the two modalities. However, these methods often struggle with temporal modeling due to issues with query consistency and limited context awareness. Query inconsistency could result in unstable masks that switch between different objects in the middle of the video, and insufficient context consideration could cause incorrect object segmentation due to a poor alignment with the textual description. To overcome the above challenges, we propose the Multi-context Temporal Consistency Module (MTCM), which integrates an Aligner and a Multi-Context Enhancer (MCE). The Aligner enhances query consistency by filtering out noise and aligning queries, while the MCE selects text-relevant queries through comprehensive context analysis. We applied MTCM to four distinct models, achieving performance improvements across all of them, including a <span><math><mi>J&amp;F</mi></math></span> score of 47.6 on the MeViS dataset. The code is available in <span><span>https://github.com/Choi58/MTCM</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107701"},"PeriodicalIF":6.0,"publicationDate":"2025-06-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144306933","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":"MGNet: RGBT tracking via cross-modality cross-region mutual guidance","authors":"Jianming Zhang ,&nbsp;Jing Yang ,&nbsp;Yu Qin ,&nbsp;Zhu Xiao ,&nbsp;Jin Wang","doi":"10.1016/j.neunet.2025.107707","DOIUrl":"10.1016/j.neunet.2025.107707","url":null,"abstract":"<div><div>Compared to single modal object tracking, the main challenge in RGBT tracking lies in effectively fusing features from both modalities. However, many existing methods neglect the dependence of distinct regions from different modalities, instead only considering that of identical regions, which fails to capture the cross-modal cross-regional relationships. In other words, they do not leverage the mutual guidance between different regions of different modalities. To address this limitation, we propose a novel RGBT tracking network, MGNet, which employs dual-stage attention and multi-scale feature fusion. The network includes the Cross-modality Cross-region Dual-stage Attention (CCDA) module and the Multi-scale Intra-region Feature Fusion (MIFF) module. The CCDA module processes features in two stages to preserve the unique features of identical region of different modalities, and then achieves mutual guidance across them. Specifically, in the first stage, features from different regions of different modalities are combined into a mixed representation, maintaining the distinct features of each region. In the second stage, attention mechanisms are applied to the mixed representation, facilitating cross-modality cross-region mutual guidance. Additionally, the MIFF module can perceive feature changes at multiple scales, ensuring effective fusion within each region. Our method achieves superior performance on three RGBT benchmark datasets (GTOT, RGBT234, and LasHeR) while running at 75 FPS, demonstrating both high accuracy and real-time performance.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107707"},"PeriodicalIF":6.0,"publicationDate":"2025-06-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144307407","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":"Physics-informed neural networks for solving inverse problems in phase field models","authors":"B.R. Zhao ,&nbsp;D.K. Sun ,&nbsp;H. Wu ,&nbsp;C.J. Qin ,&nbsp;Q.G. Fei","doi":"10.1016/j.neunet.2025.107665","DOIUrl":"10.1016/j.neunet.2025.107665","url":null,"abstract":"<div><div>The integration of materials science with Physics-Informed Neural Networks (PINNs) is critical for understanding and predicting material properties, especially through the study of inverse problems. However, much of the current research in materials science primarily focuses on applying PINNs to forward problems or improving prediction accuracy. This paper shifts the focus to inverse problems related to numerical simulation modeling, encompassing diffusion, flow, and phase transition problems through PINNs. By constructing a neural network that integrates data-driven and physics-driven modules, this study uncovers the underlying physical laws embedded within the data. More importantly, this work further validates the applicability of PINNs in the inversion of key anisotropic material parameters, with benchmark anisotropic function inversion results demonstrating a high degree of consistency between predicted and theoretical values. Additionally, this study extends the application of PINNs to multi-physics coupled systems by addressing inverse problems associated with the governing equations of phase field, temperature field, and flow field, thereby enabling parameter inversion under multi-physics conditions. This novel approach addressing inverse problems and the inversion of critical material parameters provides new perspectives, demonstrating the potential of integrating numerical simulation data and deep learning, further deepening the research on PINNs in material science.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107665"},"PeriodicalIF":6.0,"publicationDate":"2025-06-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144306861","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":"Scalable one-pass multi-view clustering with tensorized multiscale bipartite graphs fusion","authors":"Fei Wang,&nbsp;Gui-Fu Lu","doi":"10.1016/j.neunet.2025.107669","DOIUrl":"10.1016/j.neunet.2025.107669","url":null,"abstract":"<div><div>In the existing multi-view clustering task, anchor-based methods are widely used for large-scale data processing to reduce computational complexity and achieve satisfactory results. However, most existing anchor-based algorithms generate a single-scale bipartite graph for each view, limiting a more accurate representation of the original data. Moreover, these algorithms typically require further clustering processing, and the contribution of each view to the final clustering result is static, lacking dynamic adjustment based on the data characteristics. To address the above issues, we introduce an innovative multi-view clustering method called Scalable One-pass Multi-View Clustering with Tensorized Multiscale Bipartite Graphs Fusion (SOMVC/TMBGF). Specifically, we initially generate multiple scales of bipartite graphs for each view and adaptively fuse them to obtain a partition matrix, thereby fully leveraging the structural information of the original data for a more accurate representation. Subsequently, we combine the partition matrices from each view into a tensor constrained with Tensor Schatten <span><math><mi>p</mi></math></span>-norm, capturing the higher-order correlations and complementary information between views. Finally, to enhance clustering performance, we integrate partition matrix learning and clustering into a unified framework, dynamically adjusting the contribution of each view’s partition matrix through weighted spectral rotation to obtain the final clustering result. Experimental results show that SOMVC/TMBGF outperforms existing methods significantly in both clustering performance and computational efficiency, demonstrating its advantage in handling large-scale multi-view data.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107669"},"PeriodicalIF":6.0,"publicationDate":"2025-06-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144307406","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":"Diffusion policy distillation for offline reinforcement learning","authors":"Jiazhi Zhang ,&nbsp;Yuhu Cheng ,&nbsp;C.L. Philip Chen ,&nbsp;Hengrui Zhang ,&nbsp;Xuesong Wang","doi":"10.1016/j.neunet.2025.107694","DOIUrl":"10.1016/j.neunet.2025.107694","url":null,"abstract":"<div><div>Offline reinforcement learning aims to learn a well-performing target policy from a static empirical dataset. Leveraging its powerful distribution expression capabilities, the diffusion model has been widely adopted as a type of policy in offline reinforcement learning. However, sampling a single action from diffusion policy necessitates a multi-step denoise process, which results in slow decision-making speed and poses challenges for application in real-time control tasks. Inspired by the teacher–student mechanism in human learning, this paper proposes a diffusion policy distillation (DPD) framework, which employs a deterministic policy to distill the target policy induced by the diffusion model. Although the deterministic policy cannot express the complex behavior policy induced by the empirical dataset properly, it can effectively learn a relevant target policy. Moreover, since the distillated deterministic policy is one-step, it avoids the need for iterative denoising, thereby inheriting the performance of the target policy while effectively improving the decision-making speed. DPD is plug-and-play and thus can be combined with offline reinforcement learning methods based on diffusion policy. Experimental results on D4RL Gym-MuJoCo datasets indicate that the distillation policy can achieve a higher normalized score than the original policy with a lower standard deviation, and improve the decision-making speed by over 10 times.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107694"},"PeriodicalIF":6.0,"publicationDate":"2025-06-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144306931","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":"Convergent adaptive control based prescribed-time synchronization of switched fuzzy competitive network systems with time-varying delays","authors":"Dongdong Gao ,&nbsp;Fanchao Kong ,&nbsp;Tingwen Huang","doi":"10.1016/j.neunet.2025.107691","DOIUrl":"10.1016/j.neunet.2025.107691","url":null,"abstract":"<div><div>This paper addresses the prescribed-time control problem for discontinuous fuzzy neutral-type competitive neural networks (FNTCNNs) featuring switchings and time-varying delays. Notably, FNTCNNs constitute a generalized class of singularly perturbed Filippov systems. The establishment of a prescribed-time stability lemma for time-varying delay singularly perturbed systems remains a critical yet unresolved challenge. To address this, we first develop a novel prescribed-time stability lemma for singularly perturbed Filippov systems using adjustment functions, the comparison principle, and inequality techniques. This is achieved through the application of the one-norm and the introduction of a new stability definition for such systems. Considering the switching law inherent in FNTCNNs, we achieve prescribed-time stabilization control by designing adaptive prescribed-time control strategies, employing differential inclusion theory and Filippov’s solution framework. The proposed adaptive control strategies demonstrate convergence properties, ensuring that both the control strategies and system state variables converge to zero within the same prescribed-time interval. These newly developed strategies offer significant advantages over existing approaches. Finally, we validate our principal results through numerical simulations of second-order multi-agent systems subject to discontinuous disturbances.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107691"},"PeriodicalIF":6.0,"publicationDate":"2025-06-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144297358","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 general analog solver of linear and quadratic programming in one step","authors":"Sichun Du,&nbsp;Yu Dong,&nbsp;Pingdan Xiao,&nbsp;Zhengmiao Wei,&nbsp;Qinghui Hong","doi":"10.1016/j.neunet.2025.107698","DOIUrl":"10.1016/j.neunet.2025.107698","url":null,"abstract":"<div><div>Real-time solving of linear programming (LP) and quadratic programming (QP) problems faces critical demand across engineering and scientific domains. Conventional numerical approaches suffer from exponential growth in computational complexity as problem dimensionality and structural complexity increase. To address this challenge, we present a general analog solver grounded in neurodynamic principles, achieving closed-form solutions for both LP and QP through physical-level computation in one step. The proposed solver achieves the solution of LP/QP problems under diverse constraints through configurable interconnections of modular analog circuits. The analog computing architecture based on continuous-time dynamics leverages its inherent parallelism and sub-microsecond convergence properties to enhance the efficiency of optimization problem solving. Through five PSPICE simulation test experiments, the proposed QP solver achieved an average solution accuracy exceeding 99.9%, with robustness metrics maintaining over 93% precision when subjected to circuit nonidealities, including noise, parasitic resistance, and device deviation. Comparative analysis shows that the proposed solver demonstrates 173.572<span><math><mo>×</mo></math></span>, 115.871<span><math><mo>×</mo></math></span>, 8.387<span><math><mo>×</mo></math></span>, 3.241<span><math><mo>×</mo></math></span>, 21.623<span><math><mo>×</mo></math></span>, respectively, acceleration over traditional QP solvers.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"190 ","pages":"Article 107698"},"PeriodicalIF":6.0,"publicationDate":"2025-06-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144312914","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}