NeurocomputingPub Date : 2025-07-14DOI: 10.1016/j.neucom.2025.131005
Yibo Cui , Shangsheng Li , Xin Yang , Gang Wang , Yizheng Wang
{"title":"Adaptive memory fusion for multi-frame optical flow estimation","authors":"Yibo Cui , Shangsheng Li , Xin Yang , Gang Wang , Yizheng Wang","doi":"10.1016/j.neucom.2025.131005","DOIUrl":"10.1016/j.neucom.2025.131005","url":null,"abstract":"<div><div>Multi-frame optical flow models have demonstrated superior performance compared to two-frame models. Nevertheless, they suffer from several drawbacks, such as the absence of optical flow for the initial frames, the sharp increase in computational load as the number of input frames increases, and performance fluctuations when <span><math><mo>≥</mo><mn>3</mn></math></span> frames are incorporated. To address these issues, we propose AMFFlow, based on the Adaptive Memory Fusion (AMF) module, which effectively utilizes multi-frame information through memory fusion, thereby overcoming the aforementioned limitations. By employing a memory fusion mechanism, our method efficiently utilizes multi-frame information to estimate more accurate optical flow while maintaining a constant computational cost per frame. To our knowledge, AMFFlow is a pioneering model that can effectively exploit multi-frame information using <span><math><mo>≥</mo><mn>5</mn></math></span> frames for optical flow estimation. The AMF module is characterized by its low latency and plug-and-play capability. Additionally, we propose the Proportionally Debiased Endpoint Error (PDEPE), a novel metric designed to address evaluation errors caused by dataset bias in multi-frame optical flow models. Extensive experimental results demonstrate that AMFFlow achieves state-of-the-art (SOTA) generalization performance. Compared to the competing MemFlow model, the performance improvements on Sintel Training Clean and Sintel Training Final are approximately 5.9 % and 3.9 %, respectively. The source code and pre-trained models will be made publicly available at <span><span>https://github.com/keacifer/AMFFlow</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 131005"},"PeriodicalIF":5.5,"publicationDate":"2025-07-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144657290","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-14DOI: 10.1016/j.neucom.2025.130814
Ruilan Gao , Changjian Jiang , Yu Zhang
{"title":"Recurrent spiking neural networks as models of the entorhinal–hippocampal system for path integration: Grid cells and beyond","authors":"Ruilan Gao , Changjian Jiang , Yu Zhang","doi":"10.1016/j.neucom.2025.130814","DOIUrl":"10.1016/j.neucom.2025.130814","url":null,"abstract":"<div><div>Grid cells in the mammalian medial entorhinal cortex (MEC) play a pivotal role in coding spatial information and integrating self-motion, functioning as a spatial metric through multiscale periodic representations. To investigate the properties and functions of these grid codes, mechanistic models employ continuous attractor neural networks (CANNs) with hand-tuned connectivity and dynamics, while normative models use recurrent neural networks (RNNs) for path integration where hexagonal grid patterns emerge spontaneously. In this work, we develop recurrent spiking neural networks (RSNNs) with biologically realistic structures for the path integration task, generating more biologically plausible representations resembling those in the entorhinal–hippocampal system. Leveraging various spiking neuron models including a novel adaptive neuron model, the RSNNs achieve accurate and generalizable path integration performance comparable to prior normative models. Besides, the RSNNs exhibit the inherent formation of multimodular hexagonal grid patterns with more biologically plausible grid scale ratios, as well as toroidal topology and low-dimensional neural dynamics consistent with biological observations. Through experiments and analyses, the path-integrating RSNNs offer new insights into the fundamental mechanisms underlying mammals’ exceptional navigation abilities, paving the way for future research in biologically inspired navigation systems.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 130814"},"PeriodicalIF":5.5,"publicationDate":"2025-07-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144657296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-14DOI: 10.1016/j.neucom.2025.131006
Mahroosh Banday, Brejesh Lall
{"title":"Multi spectral visible-thermal IR image translation using improved u-net & conditional diffusion","authors":"Mahroosh Banday, Brejesh Lall","doi":"10.1016/j.neucom.2025.131006","DOIUrl":"10.1016/j.neucom.2025.131006","url":null,"abstract":"<div><div>Translating images from visible spectrum to thermal IR (TIR) domain to achieve precise and realistic representations of TIR images is a challenging task. Thermal infrared imaging is of great significance in scenarios where vision is severely impaired especially in difficult lighting conditions such as night, haze, fog or cloudy weather. With these advantages, infrared imaging finds extensive applicability in navigation, surveillance, object detection, product inspection, agriculture as well as remote sensing. In order to build high performance deep models for such wide range of applications, it is necessary to have large amount of TIR data for training. However, there is unavailability of sufficient IR based datasets due to high cost of thermal infrared camera setups. While large number of visible image datasets are available, this scarcity of TIR datasets can be addressed by translating visible images to their TIR counterparts. In this paper, we leverage the widely available visible range data to propose two visible to TIR domain translation approaches, one is modified U-Net based non-generative approach called TIR-UNet and the other is conditional diffusion based generative approach that also uses U-Net as neural backbone for synthesizing TIR images. Both the proposed methods have been evaluated on four benchmark datasets and demonstrate high qualitative as well as quantitative performance in generating perceptually realistic, visually plausible and high quality TIR equivalents of given visible images. Compared to state-of-the-art methods which include U-Net and powerful GAN variants, our methods achieve remarkable performance increase on the metrics of MSE, PSNR and SSIM for both day and night images.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 131006"},"PeriodicalIF":5.5,"publicationDate":"2025-07-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144686698","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-14DOI: 10.1016/j.neucom.2025.131010
Deryk Willyan Biotto , Lucas Pascotti Valem , Daniel Carlos Guimarães Pedronette , Denis Henrique Pinheiro Salvadeo
{"title":"Transduction to induction: Unsupervised representation learning based on rank information","authors":"Deryk Willyan Biotto , Lucas Pascotti Valem , Daniel Carlos Guimarães Pedronette , Denis Henrique Pinheiro Salvadeo","doi":"10.1016/j.neucom.2025.131010","DOIUrl":"10.1016/j.neucom.2025.131010","url":null,"abstract":"<div><div>The use of deep learning in supervised scenarios has become well-established. However, there is growing interest in exploring unsupervised learning methods. Transductive approaches are promising for learning rich contextual relationships in unsupervised scenarios but face challenges when dealing with large amounts of data. The main motivation of this study is to investigate the feasibility of an inductive model, based on a neural network, learning representations from ranked lists generated by transductive methods in unsupervised scenarios. We propose an unsupervised approach called Inductive Ranking Learning (IRL), which leverages techniques to learn similarities and dissimilarities from pairs derived from ranked lists produced by transductive methods. This technique involves weighting the most relevant and irrelevant elements when calculating the error of likely positive and negative pairs, based on the position of the element in the ranked list relative to its pair. This allows learning without the need for labels. The proposed approach enables the use of transductive techniques to train inductive models, promoting generalization to unseen data, which is particularly important in scenarios where new data is constantly being introduced. Experimental results show promising performance, although the method may face challenges when dealing with ranked lists derived from large datasets. Overall, the proposed approach offers significant potential for both unsupervised learning and the exploration of transductive approaches in inductive models.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 131010"},"PeriodicalIF":5.5,"publicationDate":"2025-07-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144678903","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-13DOI: 10.1016/j.neucom.2025.130992
Josh Wooley , Ashley Zachery-Savella , Michelle Le , Sally Y. Scofield , Kishore Jay , Josh Mosse-Robinson , Peter J. West , Karen S. Wilcox , Daria Nesterovich Anderson
{"title":"DynamoSort: Using machine learning approaches for the automatic classification of seizure dynamotypes","authors":"Josh Wooley , Ashley Zachery-Savella , Michelle Le , Sally Y. Scofield , Kishore Jay , Josh Mosse-Robinson , Peter J. West , Karen S. Wilcox , Daria Nesterovich Anderson","doi":"10.1016/j.neucom.2025.130992","DOIUrl":"10.1016/j.neucom.2025.130992","url":null,"abstract":"<div><div>Epilepsy is characterised by unprovoked and recurring seizures, which can be electrically measured using electroencephalograms (EEG). To better understand the underlying mechanisms of seizures, researchers are exploring their temporal dynamics through the lens of dynamical systems modelling. Seizure initiation and termination patterns of spiking amplitude and frequency can be sorted into “dynamotypes”, which may be able to serve as biomarkers for intervention. However, manual classification of these dynamotypes requires trained raters and is prone to variability. To address this, we developed DynamoSort, a machine-learning algorithm for automatic seizure onset and offset classification. Dynamotype classification of real EEG data lacks a definitive ground truth, with mean inter-rater agreement at 73.4 % for onset and 64.2 % for offset types. Despite this, DynamoSort achieved a mean area under the curve (AUC) of 0.81 for onset and a mean AUC of 0.75 for offset types. Machine-human agreement was not significantly different from human-to-human agreement. To address the lack of ground truth in ratings, DynamoSort assigns probabilistic scores (-20−20), to indicate similarity to each seizure dynamotype based on spiking features, allowing for a characterization of seizure dynamics on a spectrum rather than the traditional qualitative taxonomy. DynamoSort is a straightforward, open-access tool that uses automatic labelling and probabilistic scoring to quantify subtle changes in seizure onset and offset dynamics.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 130992"},"PeriodicalIF":5.5,"publicationDate":"2025-07-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144678904","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-13DOI: 10.1016/j.neucom.2025.130906
Jiaming Zhang , Wenjun Zhang , Shaocheng Tong
{"title":"Adaptive neural network finite-time optimal control for unmanned surface vehicle system","authors":"Jiaming Zhang , Wenjun Zhang , Shaocheng Tong","doi":"10.1016/j.neucom.2025.130906","DOIUrl":"10.1016/j.neucom.2025.130906","url":null,"abstract":"<div><div>This article investigates the adaptive neural network (NN) optimal control design problem for unmanned surface vehicle (USV) systems by finite-time control theory. A new adaptive finite-time NN optimal control policy is developed, which is composed of a NN adaptive feed-forward controller and an optimal error feedback controller. The former is constructed by using backstepping recursive control design algorithm and the latter is designed by using adaptive dynamic programming (ADP) theory. It is demonstrated that developed finite-time optimal control strategy is able to ensure the USV system is stable in a finite-time interval and achieve optimal control performance. Moreover, it can handle the computational complexity problem existing in previous finite-time optimal control methods. Comparison and simulation results illustrate the validity and superiority of the developed optimal control concept.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"650 ","pages":"Article 130906"},"PeriodicalIF":5.5,"publicationDate":"2025-07-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144656742","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-13DOI: 10.1016/j.neucom.2025.131004
Khac-Anh Phu , Van-Dung Hoang , Van-Tuong-Lan Le
{"title":"Predicting occluded skeletal joints via tracking-based feature extraction","authors":"Khac-Anh Phu , Van-Dung Hoang , Van-Tuong-Lan Le","doi":"10.1016/j.neucom.2025.131004","DOIUrl":"10.1016/j.neucom.2025.131004","url":null,"abstract":"<div><div>A rapidly growing research area in computer vision is the recognition of human poses, which imposes strong occlusion problems in subsequent tracking. Therefore, tracking mechanisms should contain algorithms for handling occlusions such that the skeletal data is continuous between frames. This paper introduces a new skeletal tracking approach where a deep-learning-based model for a Skeleton Feature Extractor is embedded into the tracking algorithm. This differs from the common tracking methods that, in the process of image feature extraction, consider feature extraction at joints and the spatial relation between them, thus making occlusion scenarios highly detectable. We further make a comparison with our model and MotioNet, which is a broadly applied model for 3D motion reconstruction. MotioNet can interpolate the missing joints based on the information both spatial and temporal. It, however, does not work when actual joints are occluded for some frames. Our model predicts the skeletal joints that are missing. Experiments on the JHMDB and Penn_Action dataset were meant to show that the method improves the accuracy of forecasting occluded skeletal joint positions by the same PCK metric.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 131004"},"PeriodicalIF":5.5,"publicationDate":"2025-07-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144662121","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-12DOI: 10.1016/j.neucom.2025.130898
Nguyen Van Thieu , Nguyen Thanh Hoang , Hossam Faris
{"title":"GrafoRVFL: A gradient-free optimization framework for boosting random vector functional link network","authors":"Nguyen Van Thieu , Nguyen Thanh Hoang , Hossam Faris","doi":"10.1016/j.neucom.2025.130898","DOIUrl":"10.1016/j.neucom.2025.130898","url":null,"abstract":"<div><div>Random Vector Functional Link (RVFL) networks have garnered attention as a rapid and efficient neural network model due to their simplified architecture and reduced training complexity. Nevertheless, the hyperparameter tuning of this network remains a substantial obstacle in the pursuit of enhanced performance across many applications. In this study, we present GrafoRVFL, an open-source framework that employs gradient-free algorithms to optimize RVFL networks’ hyperparameters. GrafoRVFL is a system that is adaptable and helps to enhance the performance of RVFL models. It is constructed on top of Numpy, Mealpy, and Scikit-Learn. We evaluate the proposed framework by comparing 14 hybrid gradient-free trained RVFL models on a variety of regression and classification datasets. The best-performing models achieve classification accuracies of 96%, 92%, and 85% on the breast cancer, waveform, and magic telescope datasets, respectively. For regression, R-scores of 0.70, 0.89, and 0.80 are observed on the diabetes, Boston housing, and California housing datasets. Additionally, we compare three hybrid RVFL models with GridSearchCV and RandomizedSearchCV on the digits dataset. The results show that our hybrid models yield better performance while requiring significantly less computational time. This suggests that our proposed framework can serve as a critical resource for researchers and practitioners who are seeking practical and resilient approaches to real-world issues. The source code of the library is accessible to the public on the GitHub repository: <span><span>https://github.com/thieu1995/GrafoRVFL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 130898"},"PeriodicalIF":5.5,"publicationDate":"2025-07-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144634486","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-12DOI: 10.1016/j.neucom.2025.130930
Bing Zhang , Xinglong Chen , Yuhua Zheng , Shuai Li , Duc Truong Pham , Yao Mao
{"title":"Two novel harmonic-resistant zeroing neural networks for time-varying problems in robotic manipulators","authors":"Bing Zhang , Xinglong Chen , Yuhua Zheng , Shuai Li , Duc Truong Pham , Yao Mao","doi":"10.1016/j.neucom.2025.130930","DOIUrl":"10.1016/j.neucom.2025.130930","url":null,"abstract":"<div><div>This paper presents two novel harmonically disturbance-resistant zeroing neural network (ZNN) models: the known frequency harmonic-resistant ZNN (KFHRZNN) and the unknown frequency harmonic-resistant ZNN (UFHRZNN). These models are designed to tackle the pseudoinverse of time-varying matrices and inverse kinematics challenges in robotic manipulators. By precisely accounting for the derivatives of harmonic disturbances, they significantly mitigate these interferences, thereby improving the control efficacy of robots in high-speed, dynamic settings. The study elucidates the design rationale, convergence characteristics, and stability assessments for both KFHRZNN and UFHRZNN. Numerical simulations and physical experiments validate the effectiveness and advantages of these models in resolving time-varying issues within robotic manipulators, highlighting their precision and robustness against harmonic disturbances.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 130930"},"PeriodicalIF":5.5,"publicationDate":"2025-07-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144672056","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
NeurocomputingPub Date : 2025-07-12DOI: 10.1016/j.neucom.2025.130895
Xiaowan Ren, Youlong Yang
{"title":"Adaptive hypergraph structure regularized semi-supervised non-negative matrix factorization for image clustering","authors":"Xiaowan Ren, Youlong Yang","doi":"10.1016/j.neucom.2025.130895","DOIUrl":"10.1016/j.neucom.2025.130895","url":null,"abstract":"<div><div>Semi-supervised non-negative matrix factorization (SNMF) is a powerful technique used in image clustering. Unlike traditional graphs, hypergraphs can capture higher-order geometric relationships among data samples. However, many existing hypergraph learning methods use a fixed hypergraph structure, which fails to dynamically optimize the hypergraph during the learning process. As a result, these methods may not accurately represent the true relationships between samples. To address this issue, this paper introduces a novel semi-supervised SNMF-based method called Adaptive Hypergraph Structure Regularized Semi-supervised Non-negative Matrix Factorization (AHS-SNMF). This method enhances the capture of high-order similarity between data samples by adaptively adjusting the hypergraph structure and iteratively improving it throughout the learning process. Additionally, we integrate labeled projection matrix with hypergraph regularization to minimize common errors found in traditional clustering methods. This approach strengthens the representation of labeled data in relation to unlabeled data, thereby boosting the model’s robustness and performance. Experimental results from comparative studies on six datasets confirm that the AHS-SNMF method significantly improves performance in semi-supervised learning tasks.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"651 ","pages":"Article 130895"},"PeriodicalIF":5.5,"publicationDate":"2025-07-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144672053","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}