Neurocomputing最新文献_第10页

Dynamic proximal policy optimization: Enhancing PPO with adaptive entropy and smooth clipping 动态近端策略优化：自适应熵和平滑裁剪增强PPO

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

Neurocomputing

Pub Date : 2026-01-30 DOI: 10.1016/j.neucom.2026.132861

Shiyu Sha, Yanhong Liu, Benyan Huo

Proximal Policy Optimization (PPO) is widely adopted in reinforcement learning but suffers from two fundamental limitations: its static entropy coefficient fails to adapt to evolving exploration–exploitation requirements, and its hard clipping mechanism introduces gradient discontinuities that can destabilize policy updates. This paper proposes Dynamic Proximal Policy Optimization (DPPO), which systematically addresses these challenges through adaptive entropy regulation and smooth policy ratio clipping. DPPO introduces two key innovations: (1) a dynamic entropy coefficient adjustment mechanism that modulates exploration based on training performance, implemented via two strategies—Surrogate Loss-Based Entropy Adjustment (SLEA) for epoch-level stability and Batch-Wise Entropy Adjustment (BWEA) for fine-grained responsiveness; (2) a smooth clipping function combining Taylor expansion with piecewise exponential decay, ensuring

C^{1}

continuity and eliminating gradient discontinuities. Extensive experiments on six continuous control tasks in the PyBullet environment demonstrate that DPPO consistently outperforms PPO and three state-of-the-art baselines (PPO-

λ

, TrulyPPO, ESPO), achieving higher sample efficiency, faster convergence, and improved stability. SLEA excels in high-dimensional tasks requiring robust exploration strategies, while BWEA achieves faster convergence in lower-complexity environments. Ablation studies confirm the individual contributions of both mechanisms, highlighting DPPO’s potential for diverse reinforcement learning applications.

近端策略优化（PPO）在强化学习中被广泛采用，但它有两个基本的局限性：它的静态熵系数不能适应不断变化的勘探开发需求，它的硬剪切机制引入了梯度不连续，这可能会破坏策略更新的稳定性。本文提出动态近端策略优化（DPPO），通过自适应熵调节和平滑策略比例裁剪，系统地解决了这些挑战。DPPO引入了两个关键创新：(1)一种动态熵系数调整机制，该机制基于训练性能调节探索，通过两种策略实现-基于代理损失的熵调整（SLEA）用于时代级稳定性和批量熵调整（BWEA）用于细粒度响应；(2)结合Taylor展开和分段指数衰减的光滑裁剪函数，保证C1连续，消除梯度不连续。PyBullet环境中六个连续控制任务的广泛实验表明，DPPO始终优于PPO和三个最先进的基线（PPO-λ, TrulyPPO， ESPO），实现更高的样本效率，更快的收敛和更好的稳定性。SLEA在需要稳健的勘探策略的高维任务中表现出色，而BWEA在低复杂性环境中实现更快的收敛。消融研究证实了这两种机制的个体贡献，突出了DPPO在多种强化学习应用中的潜力。

{"title":"Dynamic proximal policy optimization: Enhancing PPO with adaptive entropy and smooth clipping","authors":"Shiyu Sha, Yanhong Liu, Benyan Huo","doi":"10.1016/j.neucom.2026.132861","DOIUrl":"10.1016/j.neucom.2026.132861","url":null,"abstract":"<div><div>Proximal Policy Optimization (PPO) is widely adopted in reinforcement learning but suffers from two fundamental limitations: its static entropy coefficient fails to adapt to evolving exploration–exploitation requirements, and its hard clipping mechanism introduces gradient discontinuities that can destabilize policy updates. This paper proposes Dynamic Proximal Policy Optimization (DPPO), which systematically addresses these challenges through adaptive entropy regulation and smooth policy ratio clipping. DPPO introduces two key innovations: (1) a dynamic entropy coefficient adjustment mechanism that modulates exploration based on training performance, implemented via two strategies—Surrogate Loss-Based Entropy Adjustment (SLEA) for epoch-level stability and Batch-Wise Entropy Adjustment (BWEA) for fine-grained responsiveness; (2) a smooth clipping function combining Taylor expansion with piecewise exponential decay, ensuring <span><math><msup><mi>C</mi><mn>1</mn></msup></math></span> continuity and eliminating gradient discontinuities. Extensive experiments on six continuous control tasks in the PyBullet environment demonstrate that DPPO consistently outperforms PPO and three state-of-the-art baselines (PPO-<span><math><mi>λ</mi></math></span>, TrulyPPO, ESPO), achieving higher sample efficiency, faster convergence, and improved stability. SLEA excels in high-dimensional tasks requiring robust exploration strategies, while BWEA achieves faster convergence in lower-complexity environments. Ablation studies confirm the individual contributions of both mechanisms, highlighting DPPO’s potential for diverse reinforcement learning applications.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"674 ","pages":"Article 132861"},"PeriodicalIF":6.5,"publicationDate":"2026-01-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192050","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

DiCo: Disentangled concept representation for text-to-image person re-identification 文本到图像人物再识别的解纠缠概念表示

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

Neurocomputing

Pub Date : 2026-01-30 DOI: 10.1016/j.neucom.2026.132885

Giyeol Kim , Chanho Eom

Text-to-image person re-identification (TIReID) aims to retrieve person images from a large gallery given free-form textual descriptions. TIReID is challenging due to the substantial modality gap between visual appearances and textual expressions, as well as the need to model fine-grained correspondences that distinguish individuals with similar attributes such as clothing color, texture, or outfit style. To address these issues, we propose DiCo (Disentangled Concept Representation), a novel framework that achieves hierarchical and disentangled cross-modal alignment. DiCo introduces a shared slot-based representation, where each slot acts as a part-level anchor across modalities and is further decomposed into multiple concept blocks. This design enables the disentanglement of complementary attributes (e.g., color, texture, shape) while maintaining consistent part-level correspondence between image and text. Extensive experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate that our framework achieves competitive performance with state-of-the-art methods, while also enhancing interpretability through explicit slot- and block-level representations for more fine-grained retrieval results.

文本到图像的人物再识别（TIReID）旨在从给定自由格式文本描述的大型图库中检索人物图像。TIReID具有挑战性，因为视觉外观和文本表达之间存在巨大的模态差异，并且需要对细粒度对应进行建模，以区分具有相似属性（如服装颜色、质地或服装风格）的个体。为了解决这些问题，我们提出了DiCo（解纠缠概念表示），这是一个实现分层和解纠缠跨模态对齐的新框架。DiCo引入了一种基于插槽的共享表示，其中每个插槽充当跨模式的部件级锚，并进一步分解为多个概念块。这种设计使互补属性（如颜色、纹理、形状）的分离成为可能，同时保持图像和文本之间一致的部分级对应关系。在中大- pedes、ICFG-PEDES和RSTPReid上进行的大量实验表明，我们的框架使用最先进的方法实现了具有竞争力的性能，同时还通过显式的槽级和块级表示增强了可解释性，从而获得更细粒度的检索结果。

{"title":"DiCo: Disentangled concept representation for text-to-image person re-identification","authors":"Giyeol Kim , Chanho Eom","doi":"10.1016/j.neucom.2026.132885","DOIUrl":"10.1016/j.neucom.2026.132885","url":null,"abstract":"<div><div>Text-to-image person re-identification (TIReID) aims to retrieve person images from a large gallery given free-form textual descriptions. TIReID is challenging due to the substantial modality gap between visual appearances and textual expressions, as well as the need to model fine-grained correspondences that distinguish individuals with similar attributes such as clothing color, texture, or outfit style. To address these issues, we propose DiCo (Disentangled Concept Representation), a novel framework that achieves hierarchical and disentangled cross-modal alignment. DiCo introduces a shared slot-based representation, where each slot acts as a part-level anchor across modalities and is further decomposed into multiple concept blocks. This design enables the disentanglement of complementary attributes (<em>e.g.</em>, color, texture, shape) while maintaining consistent part-level correspondence between image and text. Extensive experiments on CUHK-PEDES, ICFG-PEDES, and RSTPReid demonstrate that our framework achieves competitive performance with state-of-the-art methods, while also enhancing interpretability through explicit slot- and block-level representations for more fine-grained retrieval results.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"675 ","pages":"Article 132885"},"PeriodicalIF":6.5,"publicationDate":"2026-01-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146147235","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

Unsupervised logical anomaly detection in industry: Mamba optimized by nested pyramid fusion and parameter-shared state space enhancement 工业中无监督逻辑异常检测：嵌套金字塔融合和参数共享状态空间增强优化的Mamba

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

Neurocomputing

Pub Date : 2026-01-30 DOI: 10.1016/j.neucom.2026.132824

Yueming Hu , Hanfei Lin , Huiling Huang , Yuhang Guo , Feibin Wu , Jinhuan Long , Jun Han

In recent years, unsupervised learning-based anomaly detection algorithms have provided significant support in the field of industrial inspection. Although current unsupervised algorithms can achieve excellent performance in detection tasks targeting structural anomaly scenarios, their effectiveness is relatively poor when dealing with detection tasks for logical anomaly scenarios. To overcome the aforementioned limitations, this study proposes an unsupervised logic anomaly detection algorithm based on the Mamba model framework combined with a nested pyramid feature fusion module and a parameter-shared state space enhancement module. First, a feature pyramid network is used as the fusion method to construct a nested feature pyramid fusion module, thus fully considering the special detection information at each level. Then, an improved feature space module is combined to solve the problem of high computational complexity caused by long-distance modeling. These two modules work together, addressing the problems of single-scale features failing to capture logical correlations and the difficulty of long-distance modeling of component-related regions in images. When tested on the MVTec-LOCO dataset, specifically designed for logical anomaly detection, the proposed method achieved an accuracy of 79.41 at the image level and 75.31 at the pixel level in the AU-ROC metric yielded, representing improvements of 1.42 and 3.36 respectively compared to baseline methods. While maintaining high efficiency in structural anomaly detection, the proposed method also achieves state-of-the-art performance (SoTA) in logical anomaly detection.

近年来，基于无监督学习的异常检测算法在工业检测领域提供了重要的支持。虽然目前的无监督算法在针对结构异常场景的检测任务中可以取得优异的性能，但在处理针对逻辑异常场景的检测任务时，其有效性相对较差。为了克服上述局限性，本研究提出了一种基于Mamba模型框架的无监督逻辑异常检测算法，该算法结合嵌套金字塔特征融合模块和参数共享状态空间增强模块。首先，采用特征金字塔网络作为融合方法，构建嵌套的特征金字塔融合模块，充分考虑每一层的特殊检测信息；然后，结合改进的特征空间模块，解决了远程建模带来的计算复杂度高的问题。这两个模块一起工作，解决了单尺度特征无法捕获逻辑相关性和图像中组件相关区域的远程建模困难的问题。在专门为逻辑异常检测设计的MVTec-LOCO数据集上进行测试时，所提出的方法在图像水平上的准确率为79.41，在AU-ROC度量中在像素水平上的准确率为75.31，与基线方法相比分别提高了1.42和3.36。在保持结构异常检测效率的同时，该方法在逻辑异常检测方面也达到了最先进的性能（SoTA）。

{"title":"Unsupervised logical anomaly detection in industry: Mamba optimized by nested pyramid fusion and parameter-shared state space enhancement","authors":"Yueming Hu , Hanfei Lin , Huiling Huang , Yuhang Guo , Feibin Wu , Jinhuan Long , Jun Han","doi":"10.1016/j.neucom.2026.132824","DOIUrl":"10.1016/j.neucom.2026.132824","url":null,"abstract":"<div><div>In recent years, unsupervised learning-based anomaly detection algorithms have provided significant support in the field of industrial inspection. Although current unsupervised algorithms can achieve excellent performance in detection tasks targeting structural anomaly scenarios, their effectiveness is relatively poor when dealing with detection tasks for logical anomaly scenarios. To overcome the aforementioned limitations, this study proposes an unsupervised logic anomaly detection algorithm based on the Mamba model framework combined with a nested pyramid feature fusion module and a parameter-shared state space enhancement module. First, a feature pyramid network is used as the fusion method to construct a nested feature pyramid fusion module, thus fully considering the special detection information at each level. Then, an improved feature space module is combined to solve the problem of high computational complexity caused by long-distance modeling. These two modules work together, addressing the problems of single-scale features failing to capture logical correlations and the difficulty of long-distance modeling of component-related regions in images. When tested on the MVTec-LOCO dataset, specifically designed for logical anomaly detection, the proposed method achieved an accuracy of <strong>79.41</strong> at the image level and <strong>75.31</strong> at the pixel level in the AU-ROC metric yielded, representing improvements of <strong>1.42</strong> and <strong>3.36</strong> respectively compared to baseline methods. While maintaining high efficiency in structural anomaly detection, the proposed method also achieves state-of-the-art performance (SoTA) in logical anomaly detection.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"674 ","pages":"Article 132824"},"PeriodicalIF":6.5,"publicationDate":"2026-01-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191844","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

SCL-SOD: A hybrid self-supervised contrastive learning framework for salient object detection SCL-SOD：一种用于显著目标检测的混合自监督对比学习框架

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

Neurocomputing

Pub Date : 2026-01-30 DOI: 10.1016/j.neucom.2026.132889

Zhengda Wu , Jinbao Wang , Yingchun Cui , Jinghua Zhu

Salient Object Detection (SOD) aims to identify the most visually distinctive objects in images, with broad applications in object detection, image classification, and image synthesis. Most existing SOD methods adopt supervised learning frameworks that heavily rely on labeled images as supervision signals. However, these methods often underperform in complex scenarios where camouflaged objects and backgrounds exhibit high similarity, primarily due to two limitations: (1) Insufficient supervision from labels fails to capture holistic salient regions, and (2) Task-driven supervised learning overly focuses on target objects while neglecting contextual receptive fields, resulting in elevated false-positive rates. To address these challenges, we propose a novel hybrid model, SCL-SOD, that integrates self-supervised contrastive representation learning with supervised learning in an encoder-decoder architecture with a T2T-ViT backbone. Specifically, our model has two key components: Image-wise Contrastive Learning Encoder (ICLE) that enhances global feature discriminability by learning invariant representations across different augmented views; Pixel-wise Contrastive Learning Decoder (PCLD) that refines local prediction accuracy by enforcing feature consistency at the pixel level. The final optimization combines the weighted supervised detection loss and the self-supervised contrastive loss. Extensive experiments on six standard RGB benchmarks across five evaluation metrics demonstrate that our proposed SCL-SOD model outperforms 11 state-of-the-art SOD methods, particularly in challenging scenarios with cluttered backgrounds.

显著目标检测（SOD）旨在识别图像中视觉上最显著的目标，在目标检测、图像分类和图像合成等领域有着广泛的应用。现有的SOD方法大多采用监督学习框架，严重依赖标记图像作为监督信号。然而，这些方法在伪装对象和背景具有高度相似性的复杂场景中往往表现不佳，主要是由于两个限制：(1)标签监督不足，无法捕获整体显著区域；(2)任务驱动的监督学习过度关注目标对象，而忽略了上下文接受域，导致误报率升高。为了解决这些挑战，我们提出了一种新的混合模型，SCL-SOD，它在具有T2T-ViT主干的编码器-解码器架构中集成了自监督对比表征学习和监督学习。具体来说，我们的模型有两个关键组成部分：图像智能对比学习编码器（ICLE），它通过学习不同增强视图之间的不变表示来增强全局特征的可判别性；像素级对比学习解码器（PCLD），通过在像素级加强特征一致性来改进局部预测精度。最后的优化结合了加权监督检测损失和自监督对比损失。在五个评估指标的六个标准RGB基准上进行的大量实验表明，我们提出的SCL-SOD模型优于11种最先进的SOD方法，特别是在具有杂乱背景的挑战性场景中。

{"title":"SCL-SOD: A hybrid self-supervised contrastive learning framework for salient object detection","authors":"Zhengda Wu , Jinbao Wang , Yingchun Cui , Jinghua Zhu","doi":"10.1016/j.neucom.2026.132889","DOIUrl":"10.1016/j.neucom.2026.132889","url":null,"abstract":"<div><div>Salient Object Detection (SOD) aims to identify the most visually distinctive objects in images, with broad applications in object detection, image classification, and image synthesis. Most existing SOD methods adopt supervised learning frameworks that heavily rely on labeled images as supervision signals. However, these methods often underperform in complex scenarios where camouflaged objects and backgrounds exhibit high similarity, primarily due to two limitations: (1) Insufficient supervision from labels fails to capture holistic salient regions, and (2) Task-driven supervised learning overly focuses on target objects while neglecting contextual receptive fields, resulting in elevated false-positive rates. To address these challenges, we propose a novel hybrid model, <strong>SCL-SOD</strong>, that integrates self-supervised contrastive representation learning with supervised learning in an encoder-decoder architecture with a T2T-ViT backbone. Specifically, our model has two key components: Image-wise Contrastive Learning Encoder (<strong>ICLE</strong>) that enhances global feature discriminability by learning invariant representations across different augmented views; Pixel-wise Contrastive Learning Decoder (<strong>PCLD</strong>) that refines local prediction accuracy by enforcing feature consistency at the pixel level. The final optimization combines the weighted supervised detection loss and the self-supervised contrastive loss. Extensive experiments on six standard RGB benchmarks across five evaluation metrics demonstrate that our proposed SCL-SOD model outperforms 11 state-of-the-art SOD methods, particularly in challenging scenarios with cluttered backgrounds.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"674 ","pages":"Article 132889"},"PeriodicalIF":6.5,"publicationDate":"2026-01-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192053","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

Deep fuzzy clustering inference network and its application to non-destructively estimating strength of cement microstructure 深度模糊聚类推理网络及其在水泥微观结构强度无损估计中的应用

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

Neurocomputing

Pub Date : 2026-01-30 DOI: 10.1016/j.neucom.2026.132900

Xu Wu , Shuangrong Liu , Wenwei Wang , Lin Wang , Xinbo Deng , Cong Liu , Bo Yang

A novel deep fuzzy clustering-based inference neural network (DFCINN) is proposed to develop the design methodology of the fuzzy clustering-based neural networks (FCNNs). The conventional FCNNs, while offering solutions to the rule explosion problem in neuro-fuzzy models through high-level information granularity, often falter in generalization with data of complex structures or high dimensionality. This limitation stems from their clustering-based rule generation strategy, which struggles to produce the expected rule base that has the potential to establish distinct boundaries for decision-making among different classes under such conditions. To overcome these challenges, the DFCINN integrates a deep structural framework with a fuzzy clustering strategy, effectively capturing both inter-class heterogeneity and intra-class homogeneity, which aids in constructing the desired rule base. Moreover, the cascade learning method is developed to train the parameters of the fuzzy rules, considering both clustering-based and classification losses. The performance of the DFCINN is assessed using the datasets with varying characteristics and its results are compared with those of various other methods. Moreover, the DFCINN is applied to non-destructively estimate the strength grade of cement microstructure. Experimental results reveal that the performance of the DFCINN is superior to that of its competitors.

提出了一种新的基于深度模糊聚类的推理神经网络（DFCINN），以发展基于模糊聚类的神经网络（FCNNs）的设计方法。传统的fcnn虽然通过高信息粒度解决了神经模糊模型中的规则爆炸问题，但对于结构复杂或高维的数据，其泛化能力往往会有所下降。这种限制源于它们基于聚类的规则生成策略，该策略努力生成预期的规则库，这些规则库有可能在这种条件下为不同类别之间的决策建立明显的边界。为了克服这些挑战，DFCINN将深层结构框架与模糊聚类策略相结合，有效地捕获类间异质性和类内同质性，这有助于构建所需的规则库。此外，考虑聚类损失和分类损失，采用级联学习方法训练模糊规则的参数。使用具有不同特征的数据集对DFCINN的性能进行了评估，并将其结果与其他各种方法的结果进行了比较。此外，还将DFCINN应用于水泥微观结构强度等级的无损估计。实验结果表明，DFCINN的性能优于竞争对手。

{"title":"Deep fuzzy clustering inference network and its application to non-destructively estimating strength of cement microstructure","authors":"Xu Wu , Shuangrong Liu , Wenwei Wang , Lin Wang , Xinbo Deng , Cong Liu , Bo Yang","doi":"10.1016/j.neucom.2026.132900","DOIUrl":"10.1016/j.neucom.2026.132900","url":null,"abstract":"<div><div>A novel deep fuzzy clustering-based inference neural network (DFCINN) is proposed to develop the design methodology of the fuzzy clustering-based neural networks (FCNNs). The conventional FCNNs, while offering solutions to the rule explosion problem in neuro-fuzzy models through high-level information granularity, often falter in generalization with data of complex structures or high dimensionality. This limitation stems from their clustering-based rule generation strategy, which struggles to produce the expected rule base that has the potential to establish distinct boundaries for decision-making among different classes under such conditions. To overcome these challenges, the DFCINN integrates a deep structural framework with a fuzzy clustering strategy, effectively capturing both inter-class heterogeneity and intra-class homogeneity, which aids in constructing the desired rule base. Moreover, the cascade learning method is developed to train the parameters of the fuzzy rules, considering both clustering-based and classification losses. The performance of the DFCINN is assessed using the datasets with varying characteristics and its results are compared with those of various other methods. Moreover, the DFCINN is applied to non-destructively estimate the strength grade of cement microstructure. Experimental results reveal that the performance of the DFCINN is superior to that of its competitors.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"674 ","pages":"Article 132900"},"PeriodicalIF":6.5,"publicationDate":"2026-01-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146192219","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

Complex evidential reasoning-based multi-source information fusion for pattern classification 基于复杂证据推理的多源信息融合模式分类

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

Neurocomputing

Pub Date : 2026-01-30 DOI: 10.1016/j.neucom.2026.132891

Kang Sun, Daijun Wei, Mingli Lei, Ningkui Wang

Complex evidence theory (CET) extends the description space of the Dempster-Shafer (D-S) evidence theory from the real domain to the complex domain, which has greatly improved its decision-making ability. In the decision-making process, the weight and reliability of the evidence sources are two indispensable core elements. Their synergistic effect forms an adaptive mechanism that serves as the key to eliminating interference from untrustworthy information, directly determining the scientific rigor and credibility of decisions. However, CET fails to account for these two critical factors, and this deficiency severely limits its ability to handle complex multi-source information scenarios while undermining the persuasiveness of its decision outcomes. Therefore, this article proposes a reasoning model called complex evidence reasoning (CER) based on CET, which integrates multiple pieces of evidence by considering their weights and reliability. We propose a method for generating complex mass function (CMF) with evidence weights, define a complex weighted belief distribution with reliability (CWRBD), and propose a new combination rule called the multi-angle collaborative (MAC) combination rule. Numerical examples are provided to verify the effectiveness of the proposed CER. Moreover, the proposed CER is applied to classification problems, and the results show that the CER-based classification algorithm is superior to other methods in terms of accuracy and stability.

复杂证据理论（CET）将Dempster-Shafer （D-S）证据理论的描述空间从实域扩展到复域，极大地提高了其决策能力。在决策过程中，证据来源的权重和可靠性是不可或缺的两个核心要素。它们的协同效应形成了一种适应机制，是消除不可信信息干扰的关键，直接决定了决策的科学严密性和可信度。然而，CET没有考虑到这两个关键因素，这一缺陷严重限制了其处理复杂的多源信息场景的能力，同时削弱了其决策结果的说服力。因此，本文提出了一种基于CET的复杂证据推理（CER）推理模型，该模型通过考虑多个证据的权重和可靠性来整合多个证据。提出了一种带证据权的复杂质量函数（CMF）生成方法，定义了带可靠性的复杂加权信念分布（CWRBD），并提出了一种新的组合规则多角度协同（MAC）组合规则。通过数值算例验证了该方法的有效性。将所提出的CER应用于分类问题，结果表明基于CER的分类算法在准确率和稳定性方面都优于其他方法。

{"title":"Complex evidential reasoning-based multi-source information fusion for pattern classification","authors":"Kang Sun, Daijun Wei, Mingli Lei, Ningkui Wang","doi":"10.1016/j.neucom.2026.132891","DOIUrl":"10.1016/j.neucom.2026.132891","url":null,"abstract":"<div><div>Complex evidence theory (CET) extends the description space of the Dempster-Shafer (D-S) evidence theory from the real domain to the complex domain, which has greatly improved its decision-making ability. In the decision-making process, the weight and reliability of the evidence sources are two indispensable core elements. Their synergistic effect forms an adaptive mechanism that serves as the key to eliminating interference from untrustworthy information, directly determining the scientific rigor and credibility of decisions. However, CET fails to account for these two critical factors, and this deficiency severely limits its ability to handle complex multi-source information scenarios while undermining the persuasiveness of its decision outcomes. Therefore, this article proposes a reasoning model called complex evidence reasoning (CER) based on CET, which integrates multiple pieces of evidence by considering their weights and reliability. We propose a method for generating complex mass function (CMF) with evidence weights, define a complex weighted belief distribution with reliability (CWRBD), and propose a new combination rule called the multi-angle collaborative (MAC) combination rule. Numerical examples are provided to verify the effectiveness of the proposed CER. Moreover, the proposed CER is applied to classification problems, and the results show that the CER-based classification algorithm is superior to other methods in terms of accuracy and stability.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"674 ","pages":"Article 132891"},"PeriodicalIF":6.5,"publicationDate":"2026-01-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146102537","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

Pinning-based memory state-feedback control for complex dynamical networks subject to double random deception attacks 双重随机欺骗攻击下复杂动态网络的pin -based记忆状态反馈控制

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

Neurocomputing

Pub Date : 2026-01-30 DOI: 10.1016/j.neucom.2026.132888

Jinyuan Zhang , Yuechao Ma

This research studies the topic of safe synchronization for complex dynamical networks (CDNs) under double random deception attacks. To begin with, a new memory state-feedback pinning controller is proposed for conserving resources and achieving better control performance. Moreover, a novel double random deception attack model taking into account the double communication channels is developed. In particular, an attack acting on both the non-delayed state and the time-varying delayed state is analyzed by the non-delayed and time-varying delayed forms of the random variable. Besides, the research constructs a novel synchronization criterion using the appropriate Lyapunov-Krasovskii function to ensure the

H_{\infty}

performance of CDNs subject to double random deception attacks. Ultimately, a simulation instance is exhibited to validate the efficacy and advantage of the presented theories.

研究了双重随机欺骗攻击下复杂动态网络的安全同步问题。首先，提出了一种新的记忆状态反馈固定控制器，以节省资源并获得更好的控制性能。在此基础上，提出了一种考虑双通信通道的双随机欺骗攻击模型。特别地，通过随机变量的非延迟和时变延迟形式，分析了同时作用于非延迟状态和时变延迟状态的攻击。此外，利用适当的Lyapunov-Krasovskii函数构造了一种新的同步准则，以保证cdn在双重随机欺骗攻击下的H∞性能。最后，通过一个仿真实例验证了所提理论的有效性和优越性。

引用次数: 0

IPMT++: Improving few-shot semantic segmentation with contrastive learning IPMT++：利用对比学习改进少镜头语义分割

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

Neurocomputing

Pub Date : 2026-01-29 DOI: 10.1016/j.neucom.2026.132887

Ching Chen , Hsin-Lung Wu

Few-shot semantic segmentation (FSS) aims to segment target objects in a query image using only a few annotated support images. This paper focuses on enhancing a previous attention-based FSS model, called IPMT, which faces two key challenges: (1) significant intra-class variation among pixels of the same category, and (2) high visual similarity between foreground and background pixels. These issues often lead to mismatches when computing pairwise pixel-level correlations. To address these limitations, we propose an extended version of IPMT, termed IPMT

+ +

, which integrates a novel contrastive transformer. This transformer employs contrastive learning to refine IPMT’s prototype representations, thereby improving pixel-level alignment between support and query features. Extensive experiments on the PASCAL-

5^{i}

and COCO-

20^{i}

datasets demonstrate that IPMT

+ +

consistently outperforms recent state-of-the-art methods and shows strong generalization across different backbones and few-shot settings. Our code is available at https://github.com/Alisa1114/IPMT2Plus.

少镜头语义分割（FSS）的目的是在查询图像中只使用少量注释的支持图像来分割目标对象。本文的重点是增强先前基于注意力的FSS模型，称为IPMT，该模型面临两个关键挑战：(1)同一类别像素之间的类内差异显著；(2)前景和背景像素之间的视觉相似性高。在计算成对像素级相关性时，这些问题通常会导致不匹配。为了解决这些限制，我们提出了IPMT的扩展版本，称为IPMT++，它集成了一个新的对比变压器。该转换器使用对比学习来改进IPMT的原型表示，从而改进支持和查询特征之间的像素级对齐。在PASCAL-5i和COCO-20i数据集上进行的大量实验表明，IPMT++始终优于最近最先进的方法，并且在不同的主干和少量射击设置中表现出很强的泛化性。我们的代码可在https://github.com/Alisa1114/IPMT2Plus上获得。

{"title":"IPMT++: Improving few-shot semantic segmentation with contrastive learning","authors":"Ching Chen , Hsin-Lung Wu","doi":"10.1016/j.neucom.2026.132887","DOIUrl":"10.1016/j.neucom.2026.132887","url":null,"abstract":"<div><div>Few-shot semantic segmentation (FSS) aims to segment target objects in a query image using only a few annotated support images. This paper focuses on enhancing a previous attention-based FSS model, called IPMT, which faces two key challenges: (1) significant intra-class variation among pixels of the same category, and (2) high visual similarity between foreground and background pixels. These issues often lead to mismatches when computing pairwise pixel-level correlations. To address these limitations, we propose an extended version of IPMT, termed IPMT<span><math><mo>+</mo><mo>+</mo></math></span>, which integrates a novel contrastive transformer. This transformer employs contrastive learning to refine IPMT’s prototype representations, thereby improving pixel-level alignment between support and query features. Extensive experiments on the PASCAL-<span><math><msup><mn>5</mn><mi>i</mi></msup></math></span> and COCO-<span><math><msup><mn>20</mn><mi>i</mi></msup></math></span> datasets demonstrate that IPMT<span><math><mo>+</mo><mo>+</mo></math></span> consistently outperforms recent state-of-the-art methods and shows strong generalization across different backbones and few-shot settings. Our code is available at <span><span>https://github.com/Alisa1114/IPMT2Plus</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"673 ","pages":"Article 132887"},"PeriodicalIF":6.5,"publicationDate":"2026-01-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146171443","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

Global Mittag-Leffler stability of fractional-order quaternion-valued neural networks with neutral delays on time scales 具有中性时滞的分数阶四元数值神经网络的全局Mittag-Leffler稳定性

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

Neurocomputing

Pub Date : 2026-01-29 DOI: 10.1016/j.neucom.2026.132812

Cheng Huang , Qiankun Song , Yurong Liu

This paper establishes a criterion for the global Mittag-Leffler stability of equilibrium points in fractional-order quaternion-valued neural networks (FOQVNNs) featuring neutral delays on time scales. Although conventional decomposition methods can address quaternion-valued systems by transforming them into real-valued or complex-valued equivalents, they inevitably lead to an increase in the dimensionality of linear matrix inequalities (LMIs). Furthermore, such decomposition often destroys the inherent algebraic structure of quaternions, potentially leading to conservative stability criteria. To overcome these limitations, the proposed approach avoids decomposition and derives stability criteria directly within the quaternion domain. The derivation synthesizes the Lyapunov stability principle, the free-weighting matrix method, and matrix inequality techniques. To validate the theoretical results, an illustrative example accompanied by numerical simulations is presented.

建立了具有中性时滞的分数阶四元数值神经网络平衡点全局Mittag-Leffler稳定性判据。虽然传统的分解方法可以通过将四元数值系统转换为实值或复值等价物来解决四元数值系统，但它们不可避免地导致线性矩阵不等式（lmi）的维数增加。此外，这种分解经常破坏四元数固有的代数结构，可能导致保守的稳定性准则。为了克服这些限制，该方法避免了分解，并直接在四元数域内推导出稳定性准则。推导综合了李雅普诺夫稳定性原理、自由加权矩阵法和矩阵不等式技术。为了验证理论结果，给出了一个实例并进行了数值模拟。

引用次数: 0

Towards enhancing learning on imbalanced data: A novel adaptive weighting strategy 增强对不平衡数据的学习：一种新的自适应加权策略

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

Neurocomputing

Pub Date : 2026-01-29 DOI: 10.1016/j.neucom.2026.132886

Dung Nguyen , Van-Dung Hoang , Van-Tuong-Lan Le

In recent years, deep learning has made significant progress, contributing to the improved performance of computer vision tasks such as classification and object detection. Representative models such as Convolutional Neural Networks (CNNs) and Transformers have played a key role in this development. These models have been widely applied and have shown high effectiveness in image recognition tasks. However, data imbalance remains a considerable challenge, as models tend to focus excessively on easy samples while overlooking hard or infrequent ones. To address this issue, the Focal Loss function was introduced to down-weight easy samples and enhance the learning of harder ones through predefined parameters. Nevertheless, keeping these parameters fixed throughout the training process may not be optimal, as their suitable values can vary depending on the training stage and dataset characteristics. To overcome this limitation, this paper proposes an adaptive weighting strategy in which the parameters are dynamically adjusted across different training stages. Experimental results show that the proposed method not only improves the detection of hard samples but also enhances the overall performance of the model on imbalanced datasets. These results suggest that dynamically adjusting loss parameters is an effective approach for addressing data imbalance in deep learning models.

近年来，深度学习取得了重大进展，有助于提高计算机视觉任务的性能，如分类和目标检测。卷积神经网络（cnn）和变形金刚等代表性模型在这一发展中发挥了关键作用。这些模型在图像识别任务中得到了广泛的应用，并显示出很高的有效性。然而，数据不平衡仍然是一个相当大的挑战，因为模型往往过于关注简单的样本，而忽略了困难或不常见的样本。为了解决这一问题，引入了Focal Loss函数，通过预定义参数降低了简单样本的权重，增强了难样本的学习能力。然而，在整个训练过程中保持这些参数固定可能不是最优的，因为它们的合适值可能会根据训练阶段和数据集特征而变化。为了克服这一限制，本文提出了一种自适应加权策略，该策略在不同的训练阶段动态调整参数。实验结果表明，该方法不仅改善了硬样本的检测，而且提高了模型在不平衡数据集上的整体性能。这些结果表明，动态调整损失参数是解决深度学习模型中数据不平衡的有效方法。

{"title":"Towards enhancing learning on imbalanced data: A novel adaptive weighting strategy","authors":"Dung Nguyen , Van-Dung Hoang , Van-Tuong-Lan Le","doi":"10.1016/j.neucom.2026.132886","DOIUrl":"10.1016/j.neucom.2026.132886","url":null,"abstract":"<div><div>In recent years, deep learning has made significant progress, contributing to the improved performance of computer vision tasks such as classification and object detection. Representative models such as Convolutional Neural Networks (CNNs) and Transformers have played a key role in this development. These models have been widely applied and have shown high effectiveness in image recognition tasks. However, data imbalance remains a considerable challenge, as models tend to focus excessively on easy samples while overlooking hard or infrequent ones. To address this issue, the Focal Loss function was introduced to down-weight easy samples and enhance the learning of harder ones through predefined parameters. Nevertheless, keeping these parameters fixed throughout the training process may not be optimal, as their suitable values can vary depending on the training stage and dataset characteristics. To overcome this limitation, this paper proposes an adaptive weighting strategy in which the parameters are dynamically adjusted across different training stages. Experimental results show that the proposed method not only improves the detection of hard samples but also enhances the overall performance of the model on imbalanced datasets. These results suggest that dynamically adjusting loss parameters is an effective approach for addressing data imbalance in deep learning models.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"673 ","pages":"Article 132886"},"PeriodicalIF":6.5,"publicationDate":"2026-01-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146171455","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