Pub Date : 2025-02-21DOI: 10.1016/j.neunet.2025.107273
René Larisch, Fred H. Hamker
Simple cells in the visual cortex process spatial as well as temporal information of the visual stream and enable the perception of motion information. Previous work suggests different mechanisms associated with direction selectivity, such as a temporal offset in thalamocortical input stream through lagged and non-lagged cells of the lateral geniculate nucleus (LGN), or solely from intercortical inhibition, or through a baseline selectivity provided by the thalamocortical connection tuned by intercortical inhibition.
While there exists a large corpus of models for spatiotemporal receptive fields, the majority of them built-in the spatiotemporal dynamics by utilizing a combination of spatial and temporal functions and thus, do not explain the emergence of spatiotemporal dynamics on basis of network dynamics emerging in the retina and the LGN. In order to better comprehend the emergence of spatiotemporal processing and direction selectivity, we used a spiking neural network to implement the visual pathway from the retina to the primary visual cortex. By varying different functional parts in our network, we demonstrate how the direction selectivity of simple cells emerges through the interplay between two components: tuned intercortical inhibition and a temporal offset in the feedforward path through lagged LGN cells. In contrast to previous findings, our model simulations suggest an alternative dynamic between these two mechanisms: While intercortical inhibition alone leads to bidirectional selectivity, a temporal shift in the thalamocortical pathway breaks this symmetry in favor of one direction, leading to unidirectional selectivity.
{"title":"A systematic analysis of the joint effects of ganglion cells, lagged LGN cells, and intercortical inhibition on spatiotemporal processing and direction selectivity","authors":"René Larisch, Fred H. Hamker","doi":"10.1016/j.neunet.2025.107273","DOIUrl":"10.1016/j.neunet.2025.107273","url":null,"abstract":"<div><div>Simple cells in the visual cortex process spatial as well as temporal information of the visual stream and enable the perception of motion information. Previous work suggests different mechanisms associated with direction selectivity, such as a temporal offset in thalamocortical input stream through lagged and non-lagged cells of the lateral geniculate nucleus (LGN), or solely from intercortical inhibition, or through a baseline selectivity provided by the thalamocortical connection tuned by intercortical inhibition.</div><div>While there exists a large corpus of models for spatiotemporal receptive fields, the majority of them built-in the spatiotemporal dynamics by utilizing a combination of spatial and temporal functions and thus, do not explain the emergence of spatiotemporal dynamics on basis of network dynamics emerging in the retina and the LGN. In order to better comprehend the emergence of spatiotemporal processing and direction selectivity, we used a spiking neural network to implement the visual pathway from the retina to the primary visual cortex. By varying different functional parts in our network, we demonstrate how the direction selectivity of simple cells emerges through the interplay between two components: tuned intercortical inhibition and a temporal offset in the feedforward path through lagged LGN cells. In contrast to previous findings, our model simulations suggest an alternative dynamic between these two mechanisms: While intercortical inhibition alone leads to bidirectional selectivity, a temporal shift in the thalamocortical pathway breaks this symmetry in favor of one direction, leading to unidirectional selectivity.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107273"},"PeriodicalIF":6.0,"publicationDate":"2025-02-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143509127","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-20DOI: 10.1016/j.neunet.2025.107266
Diange Zhou , Yilin Duan , Shengwen Li , Hong Yao
Semi-supervised Relation Extraction methods play an important role in extracting relationships from unstructured text, which can leverage both labeled and unlabeled data to improve extraction accuracy. However, these methods are grounded under the closed-world assumption, in which the relationship types of labeled and unlabeled data belong to the same closed set, that are not applicable to real-world scenarios that involve novel relationships. To address this issue, this paper proposes an open-world semi-supervised relation extraction task and a novel method, Seen relation Identification and Novel relation Discovery (SIND), to extract both seen and novel relations simultaneously. Specifically, SIND develops a contrastive learning strategy to improve the semantic representation of relations and incorporates a cluster-aware method for discovering novel relations by leveraging the pairwise similarity between samples in the feature space. Additionally, SIND utilizes the maximum entropy theory as the prior distribution to address the learning pace imbalance problem caused by the absence of labeled data for novel classes. Experimental results on three widely used benchmark datasets demonstrate that SIND achieves significant improvements over baseline models. This study provides an exploration to address the challenge of discovering relationships within unannotated data and presents a reference approach for various natural language processing tasks, such as text classification and named entity recognition, in open-world scenarios. The datasets and source code of this work are available at https://github.com/a-home-bird/SIND.
{"title":"Open-world semi-supervised relation extraction","authors":"Diange Zhou , Yilin Duan , Shengwen Li , Hong Yao","doi":"10.1016/j.neunet.2025.107266","DOIUrl":"10.1016/j.neunet.2025.107266","url":null,"abstract":"<div><div>Semi-supervised Relation Extraction methods play an important role in extracting relationships from unstructured text, which can leverage both labeled and unlabeled data to improve extraction accuracy. However, these methods are grounded under the closed-world assumption, in which the relationship types of labeled and unlabeled data belong to the same closed set, that are not applicable to real-world scenarios that involve novel relationships. To address this issue, this paper proposes an open-world semi-supervised relation extraction task and a novel method, Seen relation Identification and Novel relation Discovery (SIND), to extract both seen and novel relations simultaneously. Specifically, SIND develops a contrastive learning strategy to improve the semantic representation of relations and incorporates a cluster-aware method for discovering novel relations by leveraging the pairwise similarity between samples in the feature space. Additionally, SIND utilizes the maximum entropy theory as the prior distribution to address the learning pace imbalance problem caused by the absence of labeled data for novel classes. Experimental results on three widely used benchmark datasets demonstrate that SIND achieves significant improvements over baseline models. This study provides an exploration to address the challenge of discovering relationships within unannotated data and presents a reference approach for various natural language processing tasks, such as text classification and named entity recognition, in open-world scenarios. The datasets and source code of this work are available at <span><span>https://github.com/a-home-bird/SIND</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107266"},"PeriodicalIF":6.0,"publicationDate":"2025-02-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143463658","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}
Pub Date : 2025-02-20DOI: 10.1016/j.neunet.2025.107272
Yingqin Zhu , Wen Yu , Xiaoou Li
This paper introduces a novel transfer learning framework for time series forecasting that uses Concept Echo State Network (CESN) and a multi-objective optimization strategy. Our approach addresses the challenges of feature extraction and knowledge transfer in heterogeneous data environments. By optimizing CESN for each data source, we extract targeted features that capture the unique characteristics of individual datasets. Additionally, our multi-network architecture enables effective knowledge sharing among different ESNs, leading to improved forecasting performance. To further enhance efficiency, CESN reduces the need for extensive hyperparameter tuning by focusing on optimizing only the concept matrix and output weights. Our proposed framework offers a promising solution for forecasting problems where data is diverse, limited, or missing.
{"title":"A Multi-objective transfer learning framework for time series forecasting with Concept Echo State Networks","authors":"Yingqin Zhu , Wen Yu , Xiaoou Li","doi":"10.1016/j.neunet.2025.107272","DOIUrl":"10.1016/j.neunet.2025.107272","url":null,"abstract":"<div><div>This paper introduces a novel transfer learning framework for time series forecasting that uses Concept Echo State Network (CESN) and a multi-objective optimization strategy. Our approach addresses the challenges of feature extraction and knowledge transfer in heterogeneous data environments. By optimizing CESN for each data source, we extract targeted features that capture the unique characteristics of individual datasets. Additionally, our multi-network architecture enables effective knowledge sharing among different ESNs, leading to improved forecasting performance. To further enhance efficiency, CESN reduces the need for extensive hyperparameter tuning by focusing on optimizing only the concept matrix and output weights. Our proposed framework offers a promising solution for forecasting problems where data is diverse, limited, or missing.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107272"},"PeriodicalIF":6.0,"publicationDate":"2025-02-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143474196","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}
Pub Date : 2025-02-20DOI: 10.1016/j.neunet.2025.107274
You Zhao , Zhihua Allen-Zhao , Lei Wang , Xing He , Qin Mao
Second-order (inertial) neurodynamic approaches are excellent tools for solving convex optimization problems in an accelerated manner, while the majority of existing approaches to neurodynamic approaches focus on unconstrained and simple constrained convex optimization problems. This paper presents a centralized primal - dual projection neurodynamic approach with time scaling (CPDPNA-TS). Built upon the heavy - ball method, this approach is tailored for convex optimization problems characterized by set and affine constraints, which contains a second-order projection ODE (ordinary differential equation) with derivative feedback for the primal variables and a first-order ODE for the dual variables. We prove a strong global solution to CPDPNA-TS in terms of existence, uniqueness and feasibility. Subsequently, we demonstrate that CPDPNA-TS has a nonergodic exponential and an ergodic convergence properties when choosing suitable time scaling parameters, without strong convexity assumption on the objective functions. In addition, we extend the CPDPNA-TS to a case that CPDPNA-TS with a small perturbation and a case that has a distributed framework, and prove that two versions of the extension enjoy the similar convergence properties of CPDPNA-TS. Finally, we perform numerical experiments on sparse recovery in order to illustrate the effectiveness and superiority of the presented projection neurodynamic approaches.
{"title":"Inertial primal–dual projection neurodynamic approaches for constrained convex optimization problems and application to sparse recovery","authors":"You Zhao , Zhihua Allen-Zhao , Lei Wang , Xing He , Qin Mao","doi":"10.1016/j.neunet.2025.107274","DOIUrl":"10.1016/j.neunet.2025.107274","url":null,"abstract":"<div><div>Second-order (inertial) neurodynamic approaches are excellent tools for solving convex optimization problems in an accelerated manner, while the majority of existing approaches to neurodynamic approaches focus on unconstrained and simple constrained convex optimization problems. This paper presents a centralized primal - dual projection neurodynamic approach with time scaling (CPDPNA-TS). Built upon the heavy - ball method, this approach is tailored for convex optimization problems characterized by set and affine constraints, which contains a second-order projection ODE (ordinary differential equation) with derivative feedback for the primal variables and a first-order ODE for the dual variables. We prove a strong global solution to CPDPNA-TS in terms of existence, uniqueness and feasibility. Subsequently, we demonstrate that CPDPNA-TS has a nonergodic exponential and an ergodic <span><math><mrow><mi>O</mi><mfenced><mrow><mfrac><mrow><mn>1</mn></mrow><mrow><mi>t</mi></mrow></mfrac></mrow></mfenced></mrow></math></span> convergence properties when choosing suitable time scaling parameters, without strong convexity assumption on the objective functions. In addition, we extend the CPDPNA-TS to a case that CPDPNA-TS with a small perturbation and a case that has a distributed framework, and prove that two versions of the extension enjoy the similar convergence properties of CPDPNA-TS. Finally, we perform numerical experiments on sparse recovery in order to illustrate the effectiveness and superiority of the presented projection neurodynamic approaches.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107274"},"PeriodicalIF":6.0,"publicationDate":"2025-02-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143479114","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}
Pub Date : 2025-02-19DOI: 10.1016/j.neunet.2025.107269
Sugang Ma , Zhen Wan , Licheng Zhang , Bin Hu , Jinyu Zhang , Xiangmo Zhao
Numerous Transformer-based trackers have emerged due to the powerful global modeling capabilities of the Transformer. Nevertheless, the Transformer is a low-pass filter with insufficient capacity to extract high-frequency features of the target and these features are essential for target location in tracking tasks. To address this issue, this paper proposes a tracking algorithm that utilizes hybrid frequency features, which explores how to improve the performance of the tracker by fusing target multi-frequency features. Specifically, a novel feature extraction network is designed that uses CNN and Transformer to learn the multi-frequency features of the target in stages, taking advantage of both structures and balancing high- and low-frequency information. Secondly, a dual-branch encoder is designed to allow the tracker to capture global information while learning the local features of the target through another branch. Finally, a multi-frequency features fusion network is designed that uses wavelet transform and convolution to fuse high-frequency and low-frequency features. Extensive experimental results demonstrate that our tracker achieves superior tracking performance on six challenging benchmark datasets (i.e., LaSOT, TrackingNet, GOT-10k, TNL2K, UAV123, and OTB100).
{"title":"HFFTrack: Transformer tracking via hybrid frequency features","authors":"Sugang Ma , Zhen Wan , Licheng Zhang , Bin Hu , Jinyu Zhang , Xiangmo Zhao","doi":"10.1016/j.neunet.2025.107269","DOIUrl":"10.1016/j.neunet.2025.107269","url":null,"abstract":"<div><div>Numerous Transformer-based trackers have emerged due to the powerful global modeling capabilities of the Transformer. Nevertheless, the Transformer is a low-pass filter with insufficient capacity to extract high-frequency features of the target and these features are essential for target location in tracking tasks. To address this issue, this paper proposes a tracking algorithm that utilizes hybrid frequency features, which explores how to improve the performance of the tracker by fusing target multi-frequency features. Specifically, a novel feature extraction network is designed that uses CNN and Transformer to learn the multi-frequency features of the target in stages, taking advantage of both structures and balancing high- and low-frequency information. Secondly, a dual-branch encoder is designed to allow the tracker to capture global information while learning the local features of the target through another branch. Finally, a multi-frequency features fusion network is designed that uses wavelet transform and convolution to fuse high-frequency and low-frequency features. Extensive experimental results demonstrate that our tracker achieves superior tracking performance on six challenging benchmark datasets (i.e., LaSOT, TrackingNet, GOT-10k, TNL2K, UAV123, and OTB100).</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107269"},"PeriodicalIF":6.0,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143474195","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}
Pub Date : 2025-02-19DOI: 10.1016/j.neunet.2025.107271
Guangyu Gao , Zhuocheng Lv , Yan Zhang , A.K. Qin
Deep learning models are often vulnerable to adversarial attacks in both digital and physical environments. Particularly challenging are physical attacks that involve subtle, unobtrusive modifications to objects, such as patch-sticking or light-shooting, designed to maliciously alter the model’s output when the scene is captured and fed into the model. Developing physical adversarial attacks that are robust, flexible, inconspicuous, and difficult to trace remains a significant challenge. To address this issue, we propose an artistic-based camouflage named Adversarial Advertising Sign (AdvSign) for object detection task, especially in autonomous driving scenarios. Generally, artistic patterns, such as brand logos and advertisement signs, always have a high tolerance for visual incongruity and are widely exist with strong unobtrusiveness. We design these patterns into advertising signs that can be attached to various mobile carriers, such as carry-bags and vehicle stickers, to create adversarial camouflage with strong untraceability. This method is particularly effective at misleading self-driving cars, for instance, causing them to misidentify these signs as ‘stop’ signs. Our approach combines a trainable adversarial patch with various signs of artistic patterns to create advertising patches. By leveraging the diversity and flexibility of these patterns, we draw attention away from the conspicuous adversarial elements, enhancing the effectiveness and subtlety of our attacks. We then use the CARLA autonomous-driving simulator to place these synthesized patches onto 3D flat surfaces in different traffic scenes, rendering 2D composite scene images from various perspectives. These varied scene images are then input into the target detector for adversarial training, resulting in the final trained adversarial patch. In particular, we introduce a novel loss with artistic pattern constraints, designed to differentially adjust pixels within and outside the advertising sign during training. Extensive experiments in both simulated (composite scene images with AdvSign) and real-world (printed AdvSign images) environments demonstrate the effectiveness of AdvSign in executing physical attacks on state-of-the-art object detectors, such as YOLOv5. Our training strategy, leveraging diverse scene images and varied artistic transformations to adversarial patches, enables seamless integration with multiple patterns. This enhances attack effectiveness across various physical settings and allows easy adaptation to new environments and artistic patterns.
{"title":"Advertising or adversarial? AdvSign: Artistic advertising sign camouflage for target physical attacking to object detector","authors":"Guangyu Gao , Zhuocheng Lv , Yan Zhang , A.K. Qin","doi":"10.1016/j.neunet.2025.107271","DOIUrl":"10.1016/j.neunet.2025.107271","url":null,"abstract":"<div><div>Deep learning models are often vulnerable to adversarial attacks in both digital and physical environments. Particularly challenging are physical attacks that involve subtle, unobtrusive modifications to objects, such as patch-sticking or light-shooting, designed to maliciously alter the model’s output when the scene is captured and fed into the model. Developing physical adversarial attacks that are robust, flexible, inconspicuous, and difficult to trace remains a significant challenge. To address this issue, we propose an artistic-based camouflage named <em>Adv</em>ersarial <em>Adv</em>ertising <em>Sign</em> (<em>AdvSign</em>) for object detection task, especially in autonomous driving scenarios. Generally, artistic patterns, such as brand logos and advertisement signs, always have a high tolerance for visual incongruity and are widely exist with strong unobtrusiveness. We design these patterns into advertising signs that can be attached to various mobile carriers, such as carry-bags and vehicle stickers, to create adversarial camouflage with strong untraceability. This method is particularly effective at misleading self-driving cars, for instance, causing them to misidentify these signs as ‘stop’ signs. Our approach combines a trainable adversarial patch with various signs of artistic patterns to create advertising patches. By leveraging the diversity and flexibility of these patterns, we draw attention away from the conspicuous adversarial elements, enhancing the effectiveness and subtlety of our attacks. We then use the CARLA autonomous-driving simulator to place these synthesized patches onto 3D flat surfaces in different traffic scenes, rendering 2D composite scene images from various perspectives. These varied scene images are then input into the target detector for adversarial training, resulting in the final trained adversarial patch. In particular, we introduce a novel loss with artistic pattern constraints, designed to differentially adjust pixels within and outside the advertising sign during training. Extensive experiments in both simulated (composite scene images with AdvSign) and real-world (printed AdvSign images) environments demonstrate the effectiveness of AdvSign in executing physical attacks on state-of-the-art object detectors, such as YOLOv5. Our training strategy, leveraging diverse scene images and varied artistic transformations to adversarial patches, enables seamless integration with multiple patterns. This enhances attack effectiveness across various physical settings and allows easy adaptation to new environments and artistic patterns.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107271"},"PeriodicalIF":6.0,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143479116","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}
Pub Date : 2025-02-19DOI: 10.1016/j.neunet.2025.107265
Zhaowei Wang , Jun Meng , Haibin Li , Qiguo Dai , Xiaohui Lin , Yushi Luan
Accurate identification of molecular interactions is crucial for biological network analysis, which can provide valuable insights into fundamental regulatory mechanisms. Despite considerable progress driven by computational advancements, existing methods often rely on task-specific prior knowledge or inherent structural properties of molecules, which limits their generalizability and applicability. Recently, graph-based methods have emerged as a promising approach for predicting links in molecular networks. However, most of these methods focus primarily on aggregating topological information within individual domains, leading to an inadequate characterization of molecular interactions. To mitigate these challenges, we propose AMCGRL, a generalized multi-domain cooperative graph representation learning framework for multifarious molecular interaction prediction tasks. Concretely, AMCGRL incorporates multiple graph encoders to simultaneously learn molecular representations from both intra-domain and inter-domain graphs in a comprehensive manner. Then, the cross-domain decoder is employed to bridge these graph encoders to facilitate the extraction of task-relevant information across different domains. Furthermore, a hierarchical mutual attention mechanism is developed to capture complex pairwise interaction patterns between distinct types of molecules through inter-molecule communicative learning. Extensive experiments conducted on the various datasets demonstrate the superior representation learning capability of AMCGRL compared to the state-of-the-art methods, proving its effectiveness in advancing the prediction of molecular interactions.
{"title":"Attention-augmented multi-domain cooperative graph representation learning for molecular interaction prediction","authors":"Zhaowei Wang , Jun Meng , Haibin Li , Qiguo Dai , Xiaohui Lin , Yushi Luan","doi":"10.1016/j.neunet.2025.107265","DOIUrl":"10.1016/j.neunet.2025.107265","url":null,"abstract":"<div><div>Accurate identification of molecular interactions is crucial for biological network analysis, which can provide valuable insights into fundamental regulatory mechanisms. Despite considerable progress driven by computational advancements, existing methods often rely on task-specific prior knowledge or inherent structural properties of molecules, which limits their generalizability and applicability. Recently, graph-based methods have emerged as a promising approach for predicting links in molecular networks. However, most of these methods focus primarily on aggregating topological information within individual domains, leading to an inadequate characterization of molecular interactions. To mitigate these challenges, we propose AMCGRL, a generalized multi-domain cooperative graph representation learning framework for multifarious molecular interaction prediction tasks. Concretely, AMCGRL incorporates multiple graph encoders to simultaneously learn molecular representations from both intra-domain and inter-domain graphs in a comprehensive manner. Then, the cross-domain decoder is employed to bridge these graph encoders to facilitate the extraction of task-relevant information across different domains. Furthermore, a hierarchical mutual attention mechanism is developed to capture complex pairwise interaction patterns between distinct types of molecules through inter-molecule communicative learning. Extensive experiments conducted on the various datasets demonstrate the superior representation learning capability of AMCGRL compared to the state-of-the-art methods, proving its effectiveness in advancing the prediction of molecular interactions.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107265"},"PeriodicalIF":6.0,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143463657","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}
Pub Date : 2025-02-19DOI: 10.1016/j.neunet.2025.107298
Xiaochen Lai , Zheng Zhang , Liyong Zhang , Wei Lu , ZhuoHan Li
Multivariate time series imputation using graph neural networks (GNNs) has gained significant attention, where the variables and their correlations are depicted as the graph nodes and edges, offering a structured way to understand the intricacies of multivariate time series. On this basis, existing GNNs typically make the assumption of static correlations between variables, using a graph with fixed edge weights to model multivariate relationships. However, the static assumption is usually inconsistent with the dynamic nature of real-world data, where correlations between variables tend to change over time. In this paper, we propose a dynamic graph-based bilateral recurrent imputation network (DGBRIN) to address the above issue. Specifically, for each segment of a multivariate time series captured within a sliding window, we construct a specialized graph to capture the localized, dynamic correlations between variables. To this end, we design a dynamic adjacency matrix learning (DAML) module, which integrates temporal dependencies through an information fusion layer and mine localized monotonic correlations between variables using the Spearman rank correlation coefficient. These correlations are represented in segment-specific adjacency matrices. Subsequently, the adjacency matrices and time series are fed into a hybrid graph-based bilateral recurrent network for missing value imputation, which combines the advantages of recurrent neural networks and graph convolutional networks to effectively capture temporal dependencies and merge the correlation information between variables. We conduct experiments on eight real-world time series. The results demonstrate the effectiveness of the proposed model.
{"title":"Dynamic graph-based bilateral recurrent imputation network for multivariate time series","authors":"Xiaochen Lai , Zheng Zhang , Liyong Zhang , Wei Lu , ZhuoHan Li","doi":"10.1016/j.neunet.2025.107298","DOIUrl":"10.1016/j.neunet.2025.107298","url":null,"abstract":"<div><div>Multivariate time series imputation using graph neural networks (GNNs) has gained significant attention, where the variables and their correlations are depicted as the graph nodes and edges, offering a structured way to understand the intricacies of multivariate time series. On this basis, existing GNNs typically make the assumption of static correlations between variables, using a graph with fixed edge weights to model multivariate relationships. However, the static assumption is usually inconsistent with the dynamic nature of real-world data, where correlations between variables tend to change over time. In this paper, we propose a dynamic graph-based bilateral recurrent imputation network (DGBRIN) to address the above issue. Specifically, for each segment of a multivariate time series captured within a sliding window, we construct a specialized graph to capture the localized, dynamic correlations between variables. To this end, we design a dynamic adjacency matrix learning (DAML) module, which integrates temporal dependencies through an information fusion layer and mine localized monotonic correlations between variables using the Spearman rank correlation coefficient. These correlations are represented in segment-specific adjacency matrices. Subsequently, the adjacency matrices and time series are fed into a hybrid graph-based bilateral recurrent network for missing value imputation, which combines the advantages of recurrent neural networks and graph convolutional networks to effectively capture temporal dependencies and merge the correlation information between variables. We conduct experiments on eight real-world time series. The results demonstrate the effectiveness of the proposed model.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107298"},"PeriodicalIF":6.0,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143479117","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}
Pub Date : 2025-02-19DOI: 10.1016/j.neunet.2025.107268
Zhenghong Wang , Yi Wang , Furong Jia , Kun Liu , Yishan Zhang , Fan Zhang , Zhou Huang , Yu Liu
Electricity is generated through various resources and then flows between regions via a complex system (grid). Imbalances in electricity generation can lead to the waste of renewable energy. As renewable energy is becoming a larger part of the grid, it is crucial to balance generation across different resources due to the instability of renewable energy production, which depends on climate conditions. Long-term forecasting of electricity generation from multiple resources and regions can help achieve the balance and create sufficient buffers for targeted adjustments. This study revisits the cross-correlation among various energy sources across regions. Certain time-series within the grid that exhibit early fluctuations are identified as leading indicators for others. Based on the utilization of leading indicators, ALI-GC is proposed for the comprehensive modelling of global energy source interactions. Additionally, a novel deep learning model, ALI-GRU, is proposed for long-term (up to a month) collaborative electricity generation forecasting. We obtained regional-level hourly electricity generation data for the entire U.S. spanning from 2018 to 2024. In the context of hourly end-to-end forecasting and online learning scenarios, our ALI-GRU consistently outperforms state-of-the-art models by up to 11.63%. Our work demonstrates strong adaptability in large-scale, real-time forecasting scenarios, providing practical benefits for improving renewable energy management and utilization practices.
{"title":"Learning from leading indicators to predict long-term dynamics of hourly electricity generation from multiple resources","authors":"Zhenghong Wang , Yi Wang , Furong Jia , Kun Liu , Yishan Zhang , Fan Zhang , Zhou Huang , Yu Liu","doi":"10.1016/j.neunet.2025.107268","DOIUrl":"10.1016/j.neunet.2025.107268","url":null,"abstract":"<div><div>Electricity is generated through various resources and then flows between regions via a complex system (grid). Imbalances in electricity generation can lead to the waste of renewable energy. As renewable energy is becoming a larger part of the grid, it is crucial to balance generation across different resources due to the instability of renewable energy production, which depends on climate conditions. Long-term forecasting of electricity generation from multiple resources and regions can help achieve the balance and create sufficient buffers for targeted adjustments. This study revisits the cross-correlation among various energy sources across regions. Certain time-series within the grid that exhibit early fluctuations are identified as leading indicators for others. Based on the utilization of leading indicators, ALI-GC is proposed for the comprehensive modelling of global energy source interactions. Additionally, a novel deep learning model, ALI-GRU, is proposed for long-term (up to a month) collaborative electricity generation forecasting. We obtained regional-level hourly electricity generation data for the entire U.S. spanning from 2018 to 2024. In the context of hourly end-to-end forecasting and online learning scenarios, our ALI-GRU consistently outperforms state-of-the-art models by up to 11.63%. Our work demonstrates strong adaptability in large-scale, real-time forecasting scenarios, providing practical benefits for improving renewable energy management and utilization practices.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"186 ","pages":"Article 107268"},"PeriodicalIF":6.0,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143463659","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}
Pub Date : 2025-02-19DOI: 10.1016/j.neunet.2025.107270
Yuhui Quan , Xuyi He , Ruotao Xu , Yong Xu , Hui Ji
Banding artifacts in images stem from limitations in color bit depth, image compression, or over-editing, significantly degrades image quality, especially in regions with smooth gradients. Image debanding is about eliminating these artifacts while preserving the authenticity of image details. This paper introduces a novel approach to image debanding using a cross-scale invertible neural network (INN). The proposed INN is information-lossless and enhanced by a more effective cross-scale scheme. Additionally, we present a technique called banded deformable convolution, which fully leverages the anisotropic properties of banding artifacts. This technique is more compact, efficient, and exhibits better generalization compared to existing deformable convolution methods. Our proposed INN exhibits superior performance in both quantitative metrics and visual quality, as evidenced by the results of the experiments.
{"title":"Image debanding using cross-scale invertible networks with banded deformable convolutions","authors":"Yuhui Quan , Xuyi He , Ruotao Xu , Yong Xu , Hui Ji","doi":"10.1016/j.neunet.2025.107270","DOIUrl":"10.1016/j.neunet.2025.107270","url":null,"abstract":"<div><div>Banding artifacts in images stem from limitations in color bit depth, image compression, or over-editing, significantly degrades image quality, especially in regions with smooth gradients. Image debanding is about eliminating these artifacts while preserving the authenticity of image details. This paper introduces a novel approach to image debanding using a cross-scale invertible neural network (INN). The proposed INN is information-lossless and enhanced by a more effective cross-scale scheme. Additionally, we present a technique called banded deformable convolution, which fully leverages the anisotropic properties of banding artifacts. This technique is more compact, efficient, and exhibits better generalization compared to existing deformable convolution methods. Our proposed INN exhibits superior performance in both quantitative metrics and visual quality, as evidenced by the results of the experiments.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"187 ","pages":"Article 107270"},"PeriodicalIF":6.0,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143512347","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}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号