Neurocomputing最新文献

A functional neuron with thermal perception and energy regulation 具有热感知和能量调节功能的神经元

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-18 DOI: 10.1016/j.neucom.2025.132453

Zhao Lei , Yitong Guo , Jun Ma

Biological neurons and muscle cells often generate thermal energy during their discharge process. The occurrence of electrical activity, such as action potentials, is typically accompanied by a measurable release of heat. In this study, a thermosensitive neuron model is derived from a memristor-coupled neural circuit, which integrates capacitor, inductor, memristor, and thermistor, and thermal effect and interaction with memristive regulation are discussed. The model can reproduce typical firing behaviors including spiking, periodic oscillations, and bursting, and it reveals the emergence of chaotic discharges induced by hidden attractors. From the perspective of energy analysis, both the Hamilton energy and thermal energy are adopted as quantitative metrics, and then the correlation between energy level and firing modes in electrical activities is explained. The results show that chaotic discharges are associated with the lowest average Hamilton energy <H> yet the highest consumption of thermal energy, while periodic discharges exhibit an opposite behavior. Furthermore, by tuning some thermal parameters (e.g., B′ and λ), environment–related factors (e.g., k₂), and the external stimulation frequency ω, desired control over different discharge patterns can be achieved. Additive Gaussian white noise is also independently introduced into each circuit branch to explore stochastic resonance. The findings demonstrate that noise intensity significantly influences both energy levels and rhythmic behavior. This work provides a comprehensive theoretical framework for understanding thermally coupled neural dynamics and offers a novel approach for designing energy–efficient and highly tunable neuromorphic systems.

生物神经元和肌肉细胞在放电过程中经常产生热能。电活动的发生，如动作电位，通常伴随着可测量的热量释放。在这项研究中，从一个集成电容、电感、忆阻器和热敏电阻的忆阻耦合神经电路中导出了一个热敏神经元模型，并讨论了热效应和与忆阻调节的相互作用。该模型可以再现典型的放电行为，包括尖峰、周期振荡和爆破，并揭示了隐藏吸引子诱导的混沌放电的出现。从能量分析的角度，采用汉密尔顿能和热能作为定量指标，解释了电活动中能级与发射模式之间的关系。结果表明，混沌放电具有最低的平均哈密尔顿能<；H>； 和最高的热能消耗，而周期性放电则表现出相反的行为。此外，通过调整一些热参数（如B′和λ）、环境相关因素（如k2）和外部刺激频率ω，可以实现对不同放电模式的所需控制。加性高斯白噪声也被独立地引入到每个电路支路中以探索随机共振。研究结果表明，噪声强度对能量水平和节律行为都有显著影响。这项工作为理解热耦合神经动力学提供了一个全面的理论框架，并为设计节能和高可调的神经形态系统提供了一种新的方法。

{"title":"A functional neuron with thermal perception and energy regulation","authors":"Zhao Lei , Yitong Guo , Jun Ma","doi":"10.1016/j.neucom.2025.132453","DOIUrl":"10.1016/j.neucom.2025.132453","url":null,"abstract":"<div><div>Biological neurons and muscle cells often generate thermal energy during their discharge process. The occurrence of electrical activity, such as action potentials, is typically accompanied by a measurable release of heat. In this study, a thermosensitive neuron model is derived from a memristor-coupled neural circuit, which integrates capacitor, inductor, memristor, and thermistor, and thermal effect and interaction with memristive regulation are discussed. The model can reproduce typical firing behaviors including spiking, periodic oscillations, and bursting, and it reveals the emergence of chaotic discharges induced by hidden attractors. From the perspective of energy analysis, both the Hamilton energy and thermal energy are adopted as quantitative metrics, and then the correlation between energy level and firing modes in electrical activities is explained. The results show that chaotic discharges are associated with the lowest average Hamilton energy <<em>H</em>> yet the highest consumption of thermal energy, while periodic discharges exhibit an opposite behavior. Furthermore, by tuning some thermal parameters (e.g., <em>B′</em> and <em>λ</em>), environment–related factors (e.g., <em>k</em><sub>2</sub>), and the external stimulation frequency <em>ω</em>, desired control over different discharge patterns can be achieved. Additive Gaussian white noise is also independently introduced into each circuit branch to explore stochastic resonance. The findings demonstrate that noise intensity significantly influences both energy levels and rhythmic behavior. This work provides a comprehensive theoretical framework for understanding thermally coupled neural dynamics and offers a novel approach for designing energy–efficient and highly tunable neuromorphic systems.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132453"},"PeriodicalIF":6.5,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789407","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}

引用次数: 0

Dual perspective hierarchical alignment network for joint multimodal aspect-based sentiment analysis 面向联合多模态面向方面情感分析的双视角层次对齐网络

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-18 DOI: 10.1016/j.neucom.2025.132466

Juan Yang , Wanli Zou , Xu Du , Jun Shen

With the accelerating growth of multimodal user-generated content, effectively aligning image and text representations to establish deep interactions has become a key challenge in Multimodal Aspect-Based Sentiment Analysis (MABSA). To address the limitations in existing methods, such as the lack of cross-modal structural alignment, the introduction of potential noise by simple semantic alignment, and shallow fusion strategies, we propose a novel Dual Perspective Hierarchical Alignment (DPHA) framework for joint MABSA. Specifically, DPHA leverages the denoised image descriptions and aligns them with the textual content at the structural level. To further enhance structural alignment, an unsupervised contrastive learning is proposed to minimize the distance between the dependency structures of image descriptions and the corresponding text. Meanwhile, DPHA performs a cascaded fusion of object regions and salient patches to make a balance between sufficient visual information and potential noise. The fused visual representations are then semantically aligned with textual information. Next, the alignment matrices computed from the two previous modules are used to guide attention flow across multiple Transformer layers, thereby achieving deep multimodal fusion. During training, a task-specific curriculum learning strategy is incorporated to enable dynamic sample scheduling to more effectively facilitate the acquisition of the model’s capability. Experiments on benchmark datasets have verified that DPHA significantly outperforms existing baselines. Ablation studies have also demonstrated the effectiveness of each component in the proposed DPHA.

随着多模态用户生成内容的加速增长，有效地对齐图像和文本表示以建立深度交互已成为多模态基于方面的情感分析（MABSA）的关键挑战。为了解决现有方法缺乏跨模态结构对齐、简单语义对齐引入潜在噪声以及浅融合策略等局限性，我们提出了一种新的双视角分层对齐（Dual Perspective Hierarchical alignment， DPHA）框架。具体来说，DPHA利用去噪的图像描述，并在结构级别上将它们与文本内容对齐。为了进一步增强结构一致性，提出了一种无监督对比学习方法，以最小化图像描述的依赖结构与相应文本之间的距离。同时，DPHA对目标区域和显著斑块进行级联融合，在充分的视觉信息和潜在的噪声之间取得平衡。然后将融合的视觉表示与文本信息在语义上对齐。接下来，使用从前两个模块计算的对齐矩阵来引导跨多个Transformer层的注意力流，从而实现深度多模态融合。在训练过程中，采用特定于任务的课程学习策略来实现动态样本调度，从而更有效地促进模型能力的获取。在基准数据集上的实验验证了DPHA显著优于现有基线。消融研究也证明了提议的DPHA中每个成分的有效性。

{"title":"Dual perspective hierarchical alignment network for joint multimodal aspect-based sentiment analysis","authors":"Juan Yang , Wanli Zou , Xu Du , Jun Shen","doi":"10.1016/j.neucom.2025.132466","DOIUrl":"10.1016/j.neucom.2025.132466","url":null,"abstract":"<div><div>With the accelerating growth of multimodal user-generated content, effectively aligning image and text representations to establish deep interactions has become a key challenge in Multimodal Aspect-Based Sentiment Analysis (MABSA). To address the limitations in existing methods, such as the lack of cross-modal structural alignment, the introduction of potential noise by simple semantic alignment, and shallow fusion strategies, we propose a novel Dual Perspective Hierarchical Alignment (DPHA) framework for joint MABSA. Specifically, DPHA leverages the denoised image descriptions and aligns them with the textual content at the structural level. To further enhance structural alignment, an unsupervised contrastive learning is proposed to minimize the distance between the dependency structures of image descriptions and the corresponding text. Meanwhile, DPHA performs a cascaded fusion of object regions and salient patches to make a balance between sufficient visual information and potential noise. The fused visual representations are then semantically aligned with textual information. Next, the alignment matrices computed from the two previous modules are used to guide attention flow across multiple Transformer layers, thereby achieving deep multimodal fusion. During training, a task-specific curriculum learning strategy is incorporated to enable dynamic sample scheduling to more effectively facilitate the acquisition of the model’s capability. Experiments on benchmark datasets have verified that DPHA significantly outperforms existing baselines. Ablation studies have also demonstrated the effectiveness of each component in the proposed DPHA.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132466"},"PeriodicalIF":6.5,"publicationDate":"2025-12-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789398","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}

引用次数: 0

MPD-GS: Mask-guided point densification for Gaussian splatting 高斯溅射的掩模引导点致密化

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-16 DOI: 10.1016/j.neucom.2025.132438

Junhui He , Wen Xiao , Guilong Wang , Jiteng Cheng , Jiaxing Zhang , Chao Yang

3D Gaussian Splatting (3DGS) enables high-quality novel view synthesis by combining 3D Gaussian primitives with differentiable rendering, achieving impressive real-time performance. However, its Adaptive Density Control (ADC) mechanism struggles in regions with high-frequency details, leading to oversmoothing and loss of fine textures. This issue arises from the sparse distribution of initial Gaussian points, which hinders effective refinement. To address this limitation, we propose Mask-guided Point Densification Gaussian Splatting (MPD-GS), which enhances point distribution by selectively densifying Gaussian points in masked regions. We identify critical projection regions based on pixel-level error metrics during training and edge pixels extracted using Canny detection. These masked pixels are then back-projected into 3D space using depth information, improving detail preservation and rendering quality. Experiments on public benchmarks demonstrate that MPD-GS effectively recovers fine textures while maintaining computational efficiency, making it a versatile enhancement for both 2D and 3D Gaussian Splatting-based methods. The source code is available at https://github.com/Geo3DSmart/MPD-3DGS.

3D高斯飞溅（3DGS）通过将3D高斯原语与可微分渲染相结合，实现了高质量的新颖视图合成，实现了令人印象深刻的实时性能。然而，它的自适应密度控制（ADC）机制在高频细节区域挣扎，导致过平滑和精细纹理的丢失。这个问题源于初始高斯点的稀疏分布，这阻碍了有效的细化。为了解决这一限制，我们提出了掩模引导点致密化高斯溅射（MPD-GS），它通过选择性地致密高斯点来增强点分布。我们基于训练过程中的像素级误差度量和使用Canny检测提取的边缘像素来识别关键投影区域。然后使用深度信息将这些被掩盖的像素反向投影到3D空间中，从而提高细节保存和渲染质量。公共基准实验表明，MPD-GS在保持计算效率的同时有效地恢复了精细纹理，使其成为基于二维和三维高斯喷溅方法的多功能增强。源代码可从https://github.com/Geo3DSmart/MPD-3DGS获得。

{"title":"MPD-GS: Mask-guided point densification for Gaussian splatting","authors":"Junhui He , Wen Xiao , Guilong Wang , Jiteng Cheng , Jiaxing Zhang , Chao Yang","doi":"10.1016/j.neucom.2025.132438","DOIUrl":"10.1016/j.neucom.2025.132438","url":null,"abstract":"<div><div>3D Gaussian Splatting (3DGS) enables high-quality novel view synthesis by combining 3D Gaussian primitives with differentiable rendering, achieving impressive real-time performance. However, its Adaptive Density Control (ADC) mechanism struggles in regions with high-frequency details, leading to oversmoothing and loss of fine textures. This issue arises from the sparse distribution of initial Gaussian points, which hinders effective refinement. To address this limitation, we propose Mask-guided Point Densification Gaussian Splatting (MPD-GS), which enhances point distribution by selectively densifying Gaussian points in masked regions. We identify critical projection regions based on pixel-level error metrics during training and edge pixels extracted using Canny detection. These masked pixels are then back-projected into 3D space using depth information, improving detail preservation and rendering quality. Experiments on public benchmarks demonstrate that MPD-GS effectively recovers fine textures while maintaining computational efficiency, making it a versatile enhancement for both 2D and 3D Gaussian Splatting-based methods. The source code is available at <span><span>https://github.com/Geo3DSmart/MPD-3DGS</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132438"},"PeriodicalIF":6.5,"publicationDate":"2025-12-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789322","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}

引用次数: 0

NVFusion: Lightweight infrared and low-light night vision image fusion in dark environments NVFusion：在黑暗环境中进行轻型红外和低光夜视图像融合

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-16 DOI: 10.1016/j.neucom.2025.132449

Chenchen Li, Fu Li, Yi Niu, Zhifu Zhao

Current infrared and visible image fusion methods perform poorly in dark environments where visible images lack meaningful information. Low-light cameras can capture high-visibility images in the dark, making infrared and low-light night vision image fusion more meaningful, but there is no specific dataset to support this research. Therefore, we propose a new dataset of paired infrared and low-light night vision images captured using synchronous binocular cameras, which fills the gap in datasets in the field of low-light image research. Our dataset has a wide variety of scenes and objects, and its quantity and quality are suitable for the training and testing of image fusion algorithms. For this dataset, we propose NVFusion, a lightweight model designed for fusing infrared and low-light night vision images with wavelet transform. Our fusion framework leverages adaptive spatial weights based on local statistics to preserve high-brightness and high-contrast regions, while maximum gradient masks ensure rich texture details are retained. Additionally, we integrate a lightweight CNN with wavelet transform for multi-scale feature fusion, enhancing computational efficiency. Extensive experiments demonstrate that NVFusion outperforms the State-Of-The-Art(SOTA) approaches in preserving details and highlighting targets while maintaining high efficiency.

目前的红外和可见光图像融合方法在黑暗环境中表现不佳，可见图像缺乏有意义的信息。低照度摄像机可以在黑暗中捕获高能见度图像，使红外与低照度夜视图像融合更有意义，但目前没有具体的数据集支持本研究。为此，我们提出了一种利用同步双目相机捕获的红外和微光夜视图像配对数据集，填补了微光图像研究领域数据集的空白。我们的数据集具有广泛的场景和对象，其数量和质量适合于图像融合算法的训练和测试。针对该数据集，我们提出了一种基于小波变换的轻型模型NVFusion，用于融合红外和微光夜视图像。我们的融合框架利用基于局部统计的自适应空间权重来保留高亮度和高对比度区域，而最大梯度掩模确保保留丰富的纹理细节。此外，我们将轻量级CNN与小波变换相结合进行多尺度特征融合，提高了计算效率。大量的实验表明，NVFusion在保持高效率的同时，在保留细节和突出目标方面优于最先进的（SOTA）方法。

{"title":"NVFusion: Lightweight infrared and low-light night vision image fusion in dark environments","authors":"Chenchen Li, Fu Li, Yi Niu, Zhifu Zhao","doi":"10.1016/j.neucom.2025.132449","DOIUrl":"10.1016/j.neucom.2025.132449","url":null,"abstract":"<div><div>Current infrared and visible image fusion methods perform poorly in dark environments where visible images lack meaningful information. Low-light cameras can capture high-visibility images in the dark, making infrared and low-light night vision image fusion more meaningful, but there is no specific dataset to support this research. Therefore, we propose a new dataset of paired infrared and low-light night vision images captured using synchronous binocular cameras, which fills the gap in datasets in the field of low-light image research. Our dataset has a wide variety of scenes and objects, and its quantity and quality are suitable for the training and testing of image fusion algorithms. For this dataset, we propose NVFusion, a lightweight model designed for fusing infrared and low-light night vision images with wavelet transform. Our fusion framework leverages adaptive spatial weights based on local statistics to preserve high-brightness and high-contrast regions, while maximum gradient masks ensure rich texture details are retained. Additionally, we integrate a lightweight CNN with wavelet transform for multi-scale feature fusion, enhancing computational efficiency. Extensive experiments demonstrate that NVFusion outperforms the State-Of-The-Art(SOTA) approaches in preserving details and highlighting targets while maintaining high efficiency.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132449"},"PeriodicalIF":6.5,"publicationDate":"2025-12-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789428","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}

引用次数: 0

Generalizable AI-driven cross-domain breast cancer diagnosis based on mammograms by solving Fourier transformation-based jigsaw puzzles 通过解决基于傅里叶变换的拼图游戏，基于乳房x光片的通用ai驱动跨域乳腺癌诊断

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-15 DOI: 10.1016/j.neucom.2025.132439

Wanfang Xie , Zhenyu Liu , Xu Fu , Meiyun Wang , Xiaoping Yin , Jiangang Liu

The performance of deep learning models for the early diagnosis of breast cancer based on mammograms often degrades when domain shifts occur. Although certain domain generalization methods can produce images with new styles for alleviating domain shifts, each produced image exhibits a single style. The present study introduces a Fourier transformation-based jigsaw puzzle (F-Jip) method which incorporates a jigsaw puzzle generation (JPG) module, a new style generation (NSG) module, and an enhanced jigsaw puzzle generation (EJPG) process to produce the enhanced jigsaw puzzles with multi-domain information. The enhanced jigsaw puzzles can guide the model to learn domain-invariant features to alleviate the influence of domain shifts for cross-domain breast cancer diagnosis. A leave-one-domain-out cross validation using six datasets which can simulate varying domain shifts encountered in clinical scenarios is used to evaluate the performance of the models. In each fold, the AUC and accuracy of the proposed model are higher than those of several state-of-the-art domain generalization models. Additionally, when compared to BI-RADS assessments of the radiologists in one dataset, the proposed model exhibits better performance for discriminating between the benign and malignant lesions, and further identifies 81.3 % patients with benign lesions in the subgroup analysis of patients with BI-RADS 4 lesions to avoid unnecessary biopsies. The experimental results indicate that the proposed model has the potentials for assisting doctors in diagnosing breast cancer.

当域移位发生时，基于乳房x光片的乳腺癌早期诊断的深度学习模型的性能往往会下降。虽然某些领域泛化方法可以产生具有新风格的图像以减轻领域转移，但每个生成的图像都显示单一风格。本文提出了一种基于傅里叶变换的拼图生成方法，该方法将拼图生成（JPG）模块、新样式生成（NSG）模块和增强拼图生成（EJPG）过程结合起来，生成具有多域信息的增强拼图。增强的拼图可以引导模型学习域不变特征，减轻域漂移对跨域乳腺癌诊断的影响。使用六个数据集进行留一个域的交叉验证，可以模拟临床场景中遇到的不同域转移，用于评估模型的性能。在每个层次上，该模型的AUC和精度都高于目前几种领域泛化模型。此外，当与一个数据集中放射科医生的BI-RADS评估相比，所提出的模型在区分良性和恶性病变方面表现更好，并进一步在BI-RADS 4病变患者的亚组分析中识别出81.3 %的良性病变患者，以避免不必要的活检。实验结果表明，该模型具有辅助医生诊断乳腺癌的潜力。

{"title":"Generalizable AI-driven cross-domain breast cancer diagnosis based on mammograms by solving Fourier transformation-based jigsaw puzzles","authors":"Wanfang Xie , Zhenyu Liu , Xu Fu , Meiyun Wang , Xiaoping Yin , Jiangang Liu","doi":"10.1016/j.neucom.2025.132439","DOIUrl":"10.1016/j.neucom.2025.132439","url":null,"abstract":"<div><div>The performance of deep learning models for the early diagnosis of breast cancer based on mammograms often degrades when domain shifts occur. Although certain domain generalization methods can produce images with new styles for alleviating domain shifts, each produced image exhibits a single style. The present study introduces a Fourier transformation-based jigsaw puzzle (F-Jip) method which incorporates a jigsaw puzzle generation (JPG) module, a new style generation (NSG) module, and an enhanced jigsaw puzzle generation (EJPG) process to produce the enhanced jigsaw puzzles with multi-domain information. The enhanced jigsaw puzzles can guide the model to learn domain-invariant features to alleviate the influence of domain shifts for cross-domain breast cancer diagnosis. A leave-one-domain-out cross validation using six datasets which can simulate varying domain shifts encountered in clinical scenarios is used to evaluate the performance of the models. In each fold, the AUC and accuracy of the proposed model are higher than those of several state-of-the-art domain generalization models. Additionally, when compared to BI-RADS assessments of the radiologists in one dataset, the proposed model exhibits better performance for discriminating between the benign and malignant lesions, and further identifies 81.3 % patients with benign lesions in the subgroup analysis of patients with BI-RADS 4 lesions to avoid unnecessary biopsies. The experimental results indicate that the proposed model has the potentials for assisting doctors in diagnosing breast cancer.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132439"},"PeriodicalIF":6.5,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789402","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}

引用次数: 0

RotatQ: Knowledge graph embedding based on quaternion unit RotatQ：基于四元数单位的知识图嵌入

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-15 DOI: 10.1016/j.neucom.2025.132413

Shiwen Xie , Yongfang Xie , Cheng Hu , Tingwen Huang

Knowledge graph embedding (KGE) aims to own the ability of automatic prediction for missing entity in knowledge graph, which has become a research hotspot due to the incompleteness of knowledge graph. Although existing KGE models, such as TransE and RotatE, achieve good performance, there are still great difficulties in modelling relation patterns, i.e., symmetry/antisymmetry, inversion, and composition, and inferring side-side complex relations prediction, i.e., 1–1, 1-N, N-1, and N-N. Therefore, motivated by Rodrigues' rotation formula, we propose a novel KGE model, called RotatQ, which models entities and relations as a quaternion and utilizes a unit quaternion to realize the rotation from head entity to tail entity. Besides, to implement the multiple representations of entities in different types, we take type information into consideration with hierarchical type encoders. Due to some properties of quaternion, it has three free degrees to model relation patterns and complex relations simply. To demonstrate the effectiveness of our model RotatQ, we validate our model on four prevalent public databases by comparing with other state-of-the-art models. Experimental results on link prediction tasks show that RotatQ not only has good performance in most metrics, but also has good stability in knowledge graphs.

知识图嵌入（KGE）旨在拥有对知识图中缺失实体的自动预测能力，由于知识图的不完备性，知识图嵌入已成为一个研究热点。虽然现有的KGE模型（如TransE和RotatE）取得了良好的性能，但在对称/反对称、反演和组成等关系模式的建模以及推断1-1、1-N、N-1和N-N等侧-侧复杂关系预测方面仍然存在很大的困难。因此，在Rodrigues旋转公式的激励下，我们提出了一种新的KGE模型RotatQ，该模型将实体和关系建模为四元数，并利用单位四元数实现从头部实体到尾部实体的旋转。此外，为了实现不同类型实体的多重表示，我们使用分层类型编码器考虑了类型信息。由于四元数的一些性质，它具有三个自由度，可以对关系模式和复杂关系进行简单的建模。为了证明我们的模型RotatQ的有效性，我们通过与其他最先进的模型进行比较，在四个流行的公共数据库上验证了我们的模型。链路预测任务的实验结果表明，RotatQ不仅在大多数指标上具有良好的性能，而且在知识图上具有良好的稳定性。

{"title":"RotatQ: Knowledge graph embedding based on quaternion unit","authors":"Shiwen Xie , Yongfang Xie , Cheng Hu , Tingwen Huang","doi":"10.1016/j.neucom.2025.132413","DOIUrl":"10.1016/j.neucom.2025.132413","url":null,"abstract":"<div><div>Knowledge graph embedding (KGE) aims to own the ability of automatic prediction for missing entity in knowledge graph, which has become a research hotspot due to the incompleteness of knowledge graph. Although existing KGE models, such as TransE and RotatE, achieve good performance, there are still great difficulties in modelling relation patterns, i.e., symmetry/antisymmetry, inversion, and composition, and inferring side-side complex relations prediction, i.e., 1–1, 1-N, N-1, and N-N. Therefore, motivated by Rodrigues' rotation formula, we propose a novel KGE model, called RotatQ, which models entities and relations as a quaternion and utilizes a unit quaternion to realize the rotation from head entity to tail entity. Besides, to implement the multiple representations of entities in different types, we take type information into consideration with hierarchical type encoders. Due to some properties of quaternion, it has three free degrees to model relation patterns and complex relations simply. To demonstrate the effectiveness of our model RotatQ, we validate our model on four prevalent public databases by comparing with other state-of-the-art models. Experimental results on link prediction tasks show that RotatQ not only has good performance in most metrics, but also has good stability in knowledge graphs.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132413"},"PeriodicalIF":6.5,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789321","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}

引用次数: 0

A bio-inspired module for improving semantic segmentation model’s robustness 一种改进语义分割模型鲁棒性的仿生模块

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-15 DOI: 10.1016/j.neucom.2025.132276

Yu Wang , Bin Li , Yuki Todo , Zheng Tang

The biological primary visual system has greatly contributed to the design of artificial neural networks, with its mode of information transmission serving as an important reference. Simulating the primary visual pathway not only helps us better understand the biological visual system but also holds significant value for the development of the computer vision field. However, few recent studies have incorporated this simulation into rapidly evolving deep neural networks. Existing works mainly focus on electrophysiological investigations of visual feature selectivity and the simulation of neural impulses across different cortical areas. Enhancing deep neural networks through insights from visual neuroscience remains a highly important area of research. In this study, we propose an Antagonistic Modulation Module (AMM) by exploring the structure of the primary visual pathway and the functional roles of each cortical region. By simulating the primary visual pathway, the AMM extracts orientation information and performs antagonistic modulation of neural stimulation. This module can be incorporated into deep neural networks for semantic segmentation as a front-end component, effectively improving the model’s robustness. We believe that our proposed antagonistic modulation approach offers a meaningful contribution to the development of robust network architectures and holds significant implications for research in visual neuroscience.

生物初级视觉系统对人工神经网络的设计有很大的贡献，其信息传递方式是一个重要的参考。模拟初级视觉通路不仅有助于我们更好地理解生物视觉系统，而且对计算机视觉领域的发展具有重要的价值。然而，最近很少有研究将这种模拟结合到快速发展的深度神经网络中。现有的工作主要集中在视觉特征选择性的电生理研究和不同皮层区域神经脉冲的模拟。通过视觉神经科学的见解来增强深度神经网络仍然是一个非常重要的研究领域。在这项研究中，我们通过探索主要视觉通路的结构和每个皮质区域的功能作用，提出了拮抗调制模块（AMM）。通过模拟初级视觉通路，AMM提取方向信息并对神经刺激进行拮抗调节。该模块可以作为前端组件集成到深度神经网络中进行语义分割，有效提高了模型的鲁棒性。我们相信，我们提出的拮抗调制方法为鲁棒网络架构的发展提供了有意义的贡献，并对视觉神经科学的研究具有重要意义。

{"title":"A bio-inspired module for improving semantic segmentation model’s robustness","authors":"Yu Wang , Bin Li , Yuki Todo , Zheng Tang","doi":"10.1016/j.neucom.2025.132276","DOIUrl":"10.1016/j.neucom.2025.132276","url":null,"abstract":"<div><div>The biological primary visual system has greatly contributed to the design of artificial neural networks, with its mode of information transmission serving as an important reference. Simulating the primary visual pathway not only helps us better understand the biological visual system but also holds significant value for the development of the computer vision field. However, few recent studies have incorporated this simulation into rapidly evolving deep neural networks. Existing works mainly focus on electrophysiological investigations of visual feature selectivity and the simulation of neural impulses across different cortical areas. Enhancing deep neural networks through insights from visual neuroscience remains a highly important area of research. In this study, we propose an Antagonistic Modulation Module (AMM) by exploring the structure of the primary visual pathway and the functional roles of each cortical region. By simulating the primary visual pathway, the AMM extracts orientation information and performs antagonistic modulation of neural stimulation. This module can be incorporated into deep neural networks for semantic segmentation as a front-end component, effectively improving the model’s robustness. We believe that our proposed antagonistic modulation approach offers a meaningful contribution to the development of robust network architectures and holds significant implications for research in visual neuroscience.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132276"},"PeriodicalIF":6.5,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789429","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}

引用次数: 0

Knowledge interaction graph attention network for multi-behavior recommendation 多行为推荐的知识交互图注意网络

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-15 DOI: 10.1016/j.neucom.2025.132419

Yulu Du, Zili Guan

Multi-behavior recommendation aims to improve the performance on the target behavior by jointly considering multiple auxiliary behaviors, such as clicks and favourites. The recent efforts in multi-behavior recommendation take additional side information into account, such as knowledge graphs, to enhance item representation and alleviate sparsity of target signals. However, these methods are insufficient to capture the complex dependencies between multiple behaviors and knowledge semantics. To tackle this challenge, we propose a Knowledge Interaction Graph Attention Network (KIGAT) for multi-behavior recommendation. Specifically, we integrate multi-behavior interaction graph with knowledge graph as a unified knowledge interaction graph, and adopt knowledge graph embedding techniques to initialize embeddings for each entity and relation. Moreover, a knowledge interaction graph attention module is proposed to learn the complex dependencies between multiple behaviors and knowledge relations in knowledge interaction graph, as well as to model cross-behavior dependencies influenced by knowledge relations and finally capture high-order collaborative signals of multiple behaviors. Extensive experiments on three real-world datasets demonstrate that KIGAT significantly outperforms various state-of-the-art baselines and verify the effectiveness of our method.

多行为推荐是通过综合考虑点击、收藏等多种辅助行为来提高目标行为的性能。最近在多行为推荐方面的努力考虑了额外的侧信息，如知识图，以增强项目表示和减轻目标信号的稀疏性。然而，这些方法不足以捕获多个行为和知识语义之间的复杂依赖关系。为了解决这一挑战，我们提出了一个多行为推荐的知识交互图注意网络（KIGAT）。具体而言，我们将多行为交互图与知识图整合为一个统一的知识交互图，并采用知识图嵌入技术对每个实体和关系进行初始化嵌入。此外，提出了知识交互图注意模块，学习知识交互图中多种行为与知识关系之间的复杂依赖关系，并对受知识关系影响的跨行为依赖关系进行建模，最终捕获多行为的高阶协同信号。在三个真实世界数据集上进行的大量实验表明，KIGAT显著优于各种最先进的基线，并验证了我们方法的有效性。

{"title":"Knowledge interaction graph attention network for multi-behavior recommendation","authors":"Yulu Du, Zili Guan","doi":"10.1016/j.neucom.2025.132419","DOIUrl":"10.1016/j.neucom.2025.132419","url":null,"abstract":"<div><div>Multi-behavior recommendation aims to improve the performance on the target behavior by jointly considering multiple auxiliary behaviors, such as clicks and favourites. The recent efforts in multi-behavior recommendation take additional side information into account, such as knowledge graphs, to enhance item representation and alleviate sparsity of target signals. However, these methods are insufficient to capture the complex dependencies between multiple behaviors and knowledge semantics. To tackle this challenge, we propose a Knowledge Interaction Graph Attention Network (KIGAT) for multi-behavior recommendation. Specifically, we integrate multi-behavior interaction graph with knowledge graph as a unified knowledge interaction graph, and adopt knowledge graph embedding techniques to initialize embeddings for each entity and relation. Moreover, a knowledge interaction graph attention module is proposed to learn the complex dependencies between multiple behaviors and knowledge relations in knowledge interaction graph, as well as to model cross-behavior dependencies influenced by knowledge relations and finally capture high-order collaborative signals of multiple behaviors. Extensive experiments on three real-world datasets demonstrate that KIGAT significantly outperforms various state-of-the-art baselines and verify the effectiveness of our method.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132419"},"PeriodicalIF":6.5,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789527","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}

引用次数: 0

Unrolling operator splitting in learning PDEs for object detection 目标检测中学习偏微分方程的展开算子分割

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-15 DOI: 10.1016/j.neucom.2025.132172

Banu Wirawan Yohanes, Philip O. Ogunbona, Wanqing Li

Object detection presents significant challenges due to the variability in object scale, location, and orientation within images. Most state-of-the-art detectors are based on convolutional or Transformer architectures, which, while effective, often result in deep, opaque models that generalise poorly and lack interpretability. In contrast, iterative algorithms offer greater transparency and generalisation, albeit at the cost of efficiency and accuracy. In this work, we reformulate object detection as a partial differential equation (PDE)-constrained optimal control problem. This formulation exploits linear combinations of fundamental differential invariants—such as translation and rotation invariance—to embed structural priors into the learning process. We solve this problem using operator splitting via the Alternating Direction Method of Multipliers (ADMM), and unroll each optimisation step into a network layer, yielding a novel architecture: ADMM-ODNet. This approach provides a principled and interpretable alternative to conventional deep networks. Experimental results on the Corel, Pascal VOC and COCO datasets demonstrate that ADMM-ODNet outperforms leading models such as Cascade Mask R-CNN, Swin Transformer, Deformable DETR, DINO, DN-and RT-DETR, and achieves performance comparable to Plain DETR and YOLO, while requiring significantly fewer parameters.

由于图像中物体的尺度、位置和方向的可变性，物体检测提出了重大挑战。大多数最先进的检测器都是基于卷积或Transformer架构，这些架构虽然有效，但往往会导致深度、不透明的模型，泛化能力差，缺乏可解释性。相比之下，迭代算法提供了更大的透明度和泛化，尽管以效率和准确性为代价。在这项工作中，我们将目标检测重新表述为偏微分方程（PDE）约束的最优控制问题。该公式利用基本微分不变量的线性组合，如平移和旋转不变量，将结构先验嵌入到学习过程中。我们通过乘数交替方向法（ADMM）使用算子分裂解决了这个问题，并将每个优化步骤展开到一个网络层，产生了一个新的体系结构：ADMM- odnet。这种方法为传统深度网络提供了一种原则性和可解释性的替代方案。在Corel、Pascal VOC和COCO数据集上的实验结果表明，ADMM-ODNet优于Cascade Mask R-CNN、Swin Transformer、Deformable DETR、DINO、dn -和RT-DETR等领先的模型，性能可与Plain DETR和YOLO相媲美，同时所需参数明显减少。

{"title":"Unrolling operator splitting in learning PDEs for object detection","authors":"Banu Wirawan Yohanes, Philip O. Ogunbona, Wanqing Li","doi":"10.1016/j.neucom.2025.132172","DOIUrl":"10.1016/j.neucom.2025.132172","url":null,"abstract":"<div><div>Object detection presents significant challenges due to the variability in object scale, location, and orientation within images. Most state-of-the-art detectors are based on convolutional or Transformer architectures, which, while effective, often result in deep, opaque models that generalise poorly and lack interpretability. In contrast, iterative algorithms offer greater transparency and generalisation, albeit at the cost of efficiency and accuracy. In this work, we reformulate object detection as a partial differential equation (PDE)-constrained optimal control problem. This formulation exploits linear combinations of fundamental differential invariants—such as translation and rotation invariance—to embed structural priors into the learning process. We solve this problem using operator splitting via the Alternating Direction Method of Multipliers (ADMM), and unroll each optimisation step into a network layer, yielding a novel architecture: ADMM-ODNet. This approach provides a principled and interpretable alternative to conventional deep networks. Experimental results on the Corel, Pascal VOC and COCO datasets demonstrate that ADMM-ODNet outperforms leading models such as Cascade Mask R-CNN, Swin Transformer, Deformable DETR, DINO, DN-and RT-DETR, and achieves performance comparable to Plain DETR and YOLO, while requiring significantly fewer parameters.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132172"},"PeriodicalIF":6.5,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789319","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}

引用次数: 0

Graph learning and its advancements on large language models: A holistic survey 图学习及其在大型语言模型上的进展：一个全面的调查

IF 6.5 2区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Neurocomputing

Pub Date : 2025-12-15 DOI: 10.1016/j.neucom.2025.132396

Shaopeng Wei , Jun Wang , Yu Zhao , Xingyan Chen , Xiaochun Hu , Qing Li , Fuzhen Zhuang , Fuji Ren , Gang Kou

Graph learning (GL) is a prominent field that seeks to capture the intricate relationships among nodes and the topological structures of graphs. Over time, GL has evolved from the foundations of graph theory to the broader domain of graph data mining. With the advent of representation learning, it has achieved remarkable performance across a wide range of scenarios. Due to its vast potential for practical applications, GL has attracted substantial attention. While several surveys have provided valuable overviews of GL, they often fail to present a coherent connection among related objectives, methods, and applications. Consequently, they have not fully covered the breadth of emerging scenarios and pressing challenges arising from the rapid expansion of GL. In particular, large language models (LLMs) have recently exerted a transformative impact on human life, yet they exhibit notable limitations in structured scenarios. How to enhance the capabilities of these models through GL remains an open question. This survey addresses this gap by focusing on recent advancements in integrating GL with pre-trained language models, with a special emphasis on applications in the context of LLMs. Distinct from prior surveys on GL, our work offers a comprehensive review that examines current research through the lens of graph structure, highlighting recent applications, trends, and challenges. Specifically, we begin by introducing a taxonomy of GL methods, followed by a summary of the techniques employed in this field. We then present a detailed discussion of mainstream applications and conclude by outlining promising future research directions.

图学习（GL）是一个突出的领域，旨在捕捉节点之间的复杂关系和图的拓扑结构。随着时间的推移，GL已经从图论的基础发展到更广泛的图数据挖掘领域。随着表征学习的出现，它在广泛的场景中取得了显着的性能。由于其巨大的实际应用潜力，GL已经引起了人们的广泛关注。虽然一些调查提供了有价值的GL概述，但它们往往无法在相关目标、方法和应用之间呈现连贯的联系。因此，它们并没有完全涵盖新兴场景的广度和GL快速扩张所带来的紧迫挑战。特别是，大型语言模型（llm）最近对人类生活产生了变革性的影响，但它们在结构化场景中表现出明显的局限性。如何通过GL增强这些模型的能力仍然是一个悬而未决的问题。本调查通过关注将GL与预训练语言模型集成的最新进展来解决这一差距，并特别强调法学硕士背景下的应用。与之前关于GL的调查不同，我们的工作提供了一个全面的回顾，通过图表结构的视角审视当前的研究，突出了最近的应用、趋势和挑战。具体来说，我们首先介绍GL方法的分类，然后总结该领域中使用的技术。然后，我们对主流应用进行了详细的讨论，并概述了未来的研究方向。

{"title":"Graph learning and its advancements on large language models: A holistic survey","authors":"Shaopeng Wei , Jun Wang , Yu Zhao , Xingyan Chen , Xiaochun Hu , Qing Li , Fuzhen Zhuang , Fuji Ren , Gang Kou","doi":"10.1016/j.neucom.2025.132396","DOIUrl":"10.1016/j.neucom.2025.132396","url":null,"abstract":"<div><div>Graph learning (GL) is a prominent field that seeks to capture the intricate relationships among nodes and the topological structures of graphs. Over time, GL has evolved from the foundations of graph theory to the broader domain of graph data mining. With the advent of representation learning, it has achieved remarkable performance across a wide range of scenarios. Due to its vast potential for practical applications, GL has attracted substantial attention. While several surveys have provided valuable overviews of GL, they often fail to present a coherent connection among related objectives, methods, and applications. Consequently, they have not fully covered the breadth of emerging scenarios and pressing challenges arising from the rapid expansion of GL. In particular, large language models (LLMs) have recently exerted a transformative impact on human life, yet they exhibit notable limitations in structured scenarios. How to enhance the capabilities of these models through GL remains an open question. This survey addresses this gap by focusing on recent advancements in integrating GL with pre-trained language models, with a special emphasis on applications in the context of LLMs. Distinct from prior surveys on GL, our work offers a comprehensive review that examines current research through the lens of graph structure, highlighting recent applications, trends, and challenges. Specifically, we begin by introducing a taxonomy of GL methods, followed by a summary of the techniques employed in this field. We then present a detailed discussion of mainstream applications and conclude by outlining promising future research directions.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"668 ","pages":"Article 132396"},"PeriodicalIF":6.5,"publicationDate":"2025-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145789526","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}

引用次数: 0