Neurocomputing最新文献

{"title":"A centroid-based fine-tuning method for out-of-scope classification","authors":"Xinyi Cai ,&nbsp;Pei-Wei Tsai ,&nbsp;Youwen Zhang ,&nbsp;Jiao Tian ,&nbsp;Kai Zhang ,&nbsp;Ke Yu ,&nbsp;Hongwang Xiao ,&nbsp;Jinjun Chen","doi":"10.1016/j.neucom.2025.130828","DOIUrl":"10.1016/j.neucom.2025.130828","url":null,"abstract":"<div><div>Accurately detecting out-of-scope queries is a challenging task in task-oriented dialog systems. Most existing research focus on adding an outlier detector after classification or designing an open world classification to identify unknown intents. There is still a major performance gap on achieving high efficiency and accuracy based on above methods. In our research, we tend to solve this problem by constructing an out-of-scope class in the classification. We propose an explainable centroid-based fine-tuning method including a modified decision metric (MDM) and a centroid-based cosine loss (CCL) on Pre-trained Transformer models for optimization. This loss function builds on Copernican structure and assigns the same margin to each in-scope class to resolve an ambiguous configuration on out-of-scope detection. Moreover, cosine similarity is utilized to remove radial variations of centroids. Experimental results show that our proposed method achieves improvement compared to other baseline methods.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"650 ","pages":"Article 130828"},"PeriodicalIF":5.5,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144549499","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}

{"title":"WaveletFT: Discrete wavelet transform for parameter-efficient fine-tuning","authors":"Can Hu ,&nbsp;Jie Yang ,&nbsp;Shien Song ,&nbsp;Wentao Fan ,&nbsp;Tao Xie","doi":"10.1016/j.neucom.2025.130765","DOIUrl":"10.1016/j.neucom.2025.130765","url":null,"abstract":"<div><div>Recently, Low-rank adaptation (LoRA) has achieved significant popularity for fine-tuning foundational models owing to its ability to substantially reduce the number of trainable parameters and avoid additional inference costs. This reduction is achieved by introducing low-rank matrices A and B to represent weight update,defined as <span><math><mrow><mi>Δ</mi><mi>W</mi><mo>=</mo><mi>AB</mi></mrow></math></span>. Nonetheless, the accuracy gap frequently remains between LoRA and full fine-tuning (FT). Additionally, LoRA encounters difficulties concerning storage, particularly when extensive customization adaptations or larger base models are involved. In this work, we aim to resemble the learning capacity of FT and further reduce trainable parameters by leveraging the robust expressiveness of the wavelet transform (WT). We present a novel approach, named WaveletFT, which treats <span><math><mrow><mi>Δ</mi><mi>W</mi></mrow></math></span> as a matrix within the spatial domain and focuses on learning only a small subset of its coefficients. By employing the trained spectral coefficients, we utilize the inverse discrete WT to reconstruct <span><math><mrow><mi>Δ</mi><mi>W</mi></mrow></math></span>. Experimental results demonstrate that the proposed WaveletFT method offers comparable or superior performance with fewer parameters compared to LoRA across diverse tasks, such as natural language understanding, natural language generation, instruction tuning, image classification and text-to-image generation.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130765"},"PeriodicalIF":5.5,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144491679","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}

{"title":"Meta-learning-based delayless subband adaptive filter using complex self-attention for active noise control","authors":"Pengxing Feng,&nbsp;Hing Cheung So","doi":"10.1016/j.neucom.2025.130637","DOIUrl":"10.1016/j.neucom.2025.130637","url":null,"abstract":"<div><div>Active noise control typically employs adaptive filtering to generate secondary noise, where the least mean square algorithm is the most widely used. However, traditional updating rules are linear and exhibit limited effectiveness in addressing nonlinear environments and nonstationary noise. To tackle this challenge, we reformulate the active noise control problem as a meta-learning problem and propose a meta-learning-based delayless subband adaptive filter with deep neural networks. The core idea is to utilize a neural network as an adaptive algorithm that can adapt to different environments and types of noise. The neural network is trained under noisy observations, implying that it recognizes the optimized updating rule without true labels. A single-headed attention recurrent neural network is devised with learnable feature embedding to update the adaptive filter weight efficiently, enabling accurate computation of the secondary source to attenuate the unwanted primary noise. In order to relax the time constraint on updating the adaptive filter weights, the delayless subband architecture is employed, which will allow the system to be updated less frequently as the downsampling factor increases. In addition, the delayless subband architecture does not introduce additional time delays in active noise control systems. A skip updating strategy is introduced to decrease the updating frequency further, enabling machines with limited resources to adopt our meta-learning-based model. Extensive multi-condition training ensures generalization and robustness against various types of noise and environments. Simulation results demonstrate that the proposed model achieves lower steady-state error than competitors.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"650 ","pages":"Article 130637"},"PeriodicalIF":5.5,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144549693","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}

{"title":"A multimodal traffic scene understanding model integrated with optical flow maps","authors":"Yuanhang Wang ,&nbsp;Yonghua Zhou ,&nbsp;Min Zhong ,&nbsp;Yiduo Mei ,&nbsp;Hamido Fujita ,&nbsp;Hanan Aljuaid","doi":"10.1016/j.neucom.2025.130838","DOIUrl":"10.1016/j.neucom.2025.130838","url":null,"abstract":"<div><div>Autonomous driving technology has rapidly developed in recent years, promising wide-ranging applications in the transportation industry. However, the decision-making processes of autonomous driving systems are nontransparent and nonexplainable, hindering the broader adoption and industrial deployment of this technology. To overcome these challenges, this article proposes an end-to-end traffic scene understanding model (TSUM) integrated with various proven performance enhancement techniques derived from research on multimodal large models. First, features from Farneback optical flow maps and raw scene videos are simultaneously extracted using a shared-parameter network, which effectively captures the temporal dynamics of traffic scenes. By incorporating scene perception and causal analysis at the task layer, an adaptive task complexity progression strategy is designed, which reduces the task complexity. Finally, we establish a guiding library with a diverse array of guiding phrases to enhance the generalization capabilities of the TSUM. The model is experimentally validated on the publicly available Berkeley Deep Drive-X dataset. Comparison studies confirmed that the TSUM outperforms state-of-the-art methods across multiple evaluation metrics.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130838"},"PeriodicalIF":5.5,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144513827","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}

{"title":"CUR-Estimator: Towards reliable missing data imputation for aero-engine degradation process","authors":"Haoze Wu ,&nbsp;Shisheng Zhong ,&nbsp;Minghang Zhao ,&nbsp;Xuyun Fu ,&nbsp;Yongjian Zhang ,&nbsp;Song Fu","doi":"10.1016/j.neucom.2025.130876","DOIUrl":"10.1016/j.neucom.2025.130876","url":null,"abstract":"<div><div>In the process of missing data imputation for aero-engine life cycle degradation time series, two primary challenges arise. First, due to variations in the duration of different flight missions, and the fact that the same flight phase may also vary in duration across different flights, the time intervals for collecting key samples are not always consistent. This variability increases the difficulty of evaluating the impact of individual flights on overall performance changes. Second, when using neural networks for imputing missing data, issues such as significant noise or extended periods of missing data may arise, leading to unreasonable imputation results. To address the challenges, this paper proposes a Constrained Unseen Recovery Estimator (CUR-Estimator) for imputing missing data in aero-engine life cycle degradation datasets. Firstly, the time interval information is encoded via a transformer-enhanced gate recurrent unit. The results are then combined with missing masks to adjust the hidden states and input weights for the missing segments, forming the Interval-Aware Temporal Imputation Network. Secondly, this paper uses statistical interpolation methods to constrain the imputation results of the neural network, limiting the range of imputation outcomes and thereby reducing the possibility of unreasonable outputs. As an example, the Piecewise Cubic Hermite Interpolating Polynomial is applied to constrain the Interval-Aware Temporal Imputation Network in handling time interval information. Finally, experiments were conducted using a simulation dataset and a real civil aero-engine dataset, which showed that the proposed method has high accuracy and strong stability.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130876"},"PeriodicalIF":5.5,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144502553","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}

{"title":"LDM-KGC: A low-dimensional knowledge graph completion model based on multi-head attention mechanism","authors":"Chaoqun Zhang ,&nbsp;Bingjie Qiu ,&nbsp;Weidong Tang ,&nbsp;Bicheng Liang ,&nbsp;Danyang Cui ,&nbsp;Haisheng Luo ,&nbsp;Qiming Chen","doi":"10.1016/j.neucom.2025.130665","DOIUrl":"10.1016/j.neucom.2025.130665","url":null,"abstract":"<div><div>Existing Transformer-based knowledge graph completion methods often rely on high-dimensional embeddings to achieve competitive performance, which to some extent limits their scalability on large-scale knowledge graphs. To address this challenge, the LDM-KGC model based on the multi-head attention mechanism is proposed. By combining QKV-layer and Update-layer, LDM-KGC can not only learn rich information but also reduce information loss during training, thereby achieving superior embedding representations in low-dimensional spaces. Specifically, QKV-layer utilizes the multi-head attention mechanism to effectively capture interactions between entities and relations, while Update-layer further refines the resulting embeddings. Experimental results on the FB15k-237 and WN18RR datasets demonstrate that LDM-KGC outperforms 14 baseline models, significantly improving mean reciprocal rank (MRR) by 12.4 percentage points and 24.4 percentage points over the worst baseline, respectively. Notably, LDM-KGC achieves MRR of 36.5%, Hits@1 of 27.1%, Hits@3 of 40.2%, and Hits@10 of 55.2% on the FB15k-237 dataset. Furthermore, LDM-KGC reaches a Hits@10 score of 65.2% on the NELL-995 dataset. These results underscore the effectiveness of LDM-KGC in generating low-dimensional embeddings, thereby offering a scalable solution for large-scale knowledge graph completion.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130665"},"PeriodicalIF":5.5,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144518725","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}

{"title":"Adaptive decoupled strategy for robust and efficient low-rank matrix decomposition","authors":"Min Gan ,&nbsp;Peng Xue ,&nbsp;Fan Zhang ,&nbsp;Xiang-Xiang Su ,&nbsp;Xin Lin ,&nbsp;Guang-Yong Chen","doi":"10.1016/j.neucom.2025.130656","DOIUrl":"10.1016/j.neucom.2025.130656","url":null,"abstract":"<div><div>Low-rank matrix decomposition with missing values is vital in computer vision and pattern recognition, yet it presents significant challenges. This problem can be viewed as a separable nonlinear optimization, but traditional methods often fail to account for the coupling between parameters and the impact of solution properties on visual reconstruction. We observe that such separable nonlinear problems frequently encounters narrow ravines filled with sharp minima. Classic alternating optimization methods, the Wiberg algorithm and its variants tend to linger in these regions, converging to sharp minima, thereby slowing convergence and degrading reconstruction quality. This promotes us to introduce the Adaptive Decoupled Variable Projection algorithm (ADVP), which can adaptively handle the coupling of parameters, significantly accelerate the convergence rate, and dynamically adjust the parameter search subspace, helping algorithms avoid these ravines towards flatter local minima. These flat minima exhibit robustness against missing data, noise, and outliers, enhancing the quality of visual reconstruction. Extensive experiments on synthetic and real datasets have validated the efficiency of ADVP and its superior performance in visual reconstruction.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130656"},"PeriodicalIF":5.5,"publicationDate":"2025-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144518743","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}

{"title":"Domain generalization for person re-identification: A survey towards domain-agnostic person matching","authors":"Hyeonseo Lee,&nbsp;Juhyun Park,&nbsp;Jihyong Oh,&nbsp;Chanho Eom","doi":"10.1016/j.neucom.2025.130763","DOIUrl":"10.1016/j.neucom.2025.130763","url":null,"abstract":"<div><div>Person Re-identification (ReID) aims to retrieve images of the same individual captured across non-overlapping camera views, making it a critical component of intelligent surveillance systems. Traditional ReID methods assume that the training and test domains share similar characteristics and primarily focus on learning discriminative features within a given domain. However, they often fail to generalize to unseen domains due to domain shifts caused by variations in viewpoint, background, and lighting conditions. To address this issue, Domain-Adaptive ReID (DA-ReID) methods have been proposed. These approaches incorporate unlabeled target domain data during training and improve performance by aligning feature distributions between source and target domains. However, their reliance on access to target domain data limits their scalability and makes them less suitable for real-world deployments, where such data may not be available in advance. Domain-Generalizable ReID (DG-ReID) tackles a more realistic and challenging setting by aiming to learn domain-invariant features without relying on any target domain data. Recent methods have explored various strategies to enhance generalization across diverse environments, but the field remains relatively underexplored. In this paper, we present a comprehensive survey of DG-ReID. We first review the architectural components of DG-ReID including the overall setting, commonly used backbone networks and multi-source input configurations. Then, we categorize and analyze domain generalization modules that explicitly aim to learn domain-invariant and identity-discriminative representations. To examine the broader applicability of these techniques, we further conduct a case study on a related task that also involves distribution shifts. Finally, we discuss recent trends, open challenges, and promising directions for future research in DG-ReID. To the best of our knowledge, this is the first systematic survey dedicated to DG-ReID. A curated list of related resources and papers is also available at: <span><span>https://github.com/PerceptualAI-Lab/Awesome-Domain-Generalizable-Person-Re-ID</span><svg><path></path></svg></span></div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130763"},"PeriodicalIF":5.5,"publicationDate":"2025-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144502550","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}

{"title":"DualNet: A dual-path network with adaptive compensation for multivariate time series forecasting","authors":"Huashun Li,&nbsp;Weimin Wu,&nbsp;Wei Chen,&nbsp;Mei Zhang","doi":"10.1016/j.neucom.2025.130720","DOIUrl":"10.1016/j.neucom.2025.130720","url":null,"abstract":"<div><div>Multivariate time series prediction remains a fundamental challenge across various domains due to its complex temporal dynamics and inherent dependencies among variables. This paper introduces DualNet, an innovative architectural paradigm based on decoupling principles, which achieves effective decomposition and complementary enhancement of prediction tasks through spatial decomposition. The architecture synergistically integrates global pattern extraction and adaptive local compensation mechanisms, establishing two independent yet complementary learning spaces through functional decoupling. Our framework introduces a dual-path learning strategy that decomposes the prediction process into complementary components: a primary trajectory estimator focused on capturing inherent temporal evolution patterns, and a dedicated compensation mechanism that performs fine-grained calibration through residual learning. This decoupled design not only reduces the learning complexity of individual modules but also enhances overall prediction performance through complementary effects. We developed a temperature-scaled adaptive weighting scheme that dynamically adjusts compensation intensity based on temporal context, enabling the model to achieve optimal balance between prediction stability and adaptive refinement. Through the dual mechanisms of spatial decoupling and complementary enhancement, the architecture achieves organic unification of global feature extraction and local fine-tuning, while incorporating novel normalization strategies and hierarchical feature transformation mechanisms to enhance the model’s representational capacity. Comprehensive experiments conducted on various benchmark datasets demonstrate that this decoupling-based dual-channel architecture significantly improves the model’s capability to capture complex temporal patterns while maintaining prediction stability. The code is available at <span><span>https://github.com/ZS520L/DualNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130720"},"PeriodicalIF":5.5,"publicationDate":"2025-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144469897","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}

{"title":"Recommendation systems with LLM-based semantic embeddings and FAISS similarity search","authors":"Seema Safar ,&nbsp;Babita Roslind Jose ,&nbsp;Jimson Mathew ,&nbsp;T. Santhanakrishnan","doi":"10.1016/j.neucom.2025.130753","DOIUrl":"10.1016/j.neucom.2025.130753","url":null,"abstract":"<div><div>Content-based recommendation systems have gained significant attention for their ability to provide personalized suggestions by analyzing item descriptions. Leveraging the power of large language models (LLMs), this research introduces a novel recommendation approach that generates high-quality semantic embeddings to facilitate efficient similarity-based retrieval for Top-N recommendations. The proposed method capitalizes on the deep contextual understanding of LLMs to capture intricate semantic relationships within item content, thereby enhancing recommendation relevance. Furthermore, the system integrates FAISS (Facebook AI Similarity Search) to optimize similarity search, enabling faster and more scalable retrieval of relevant recommendations. To evaluate its effectiveness, the system is tested on four diverse real-world datasets: Yelp, Amazon Beauty, MovieLens, and LastFM, covering multiple domains. Performance is assessed using widely adopted evaluation metrics, including Normalized Discounted Cumulative Gain (NDCG), Precision, Recall, Hit Rate (HR), F1-Score and business-relevant evaluation measures. Extensive experimental results demonstrate that the proposed method, augmented with FAISS, consistently outperforms the existing state-of-the-art recommendation techniques. The code supporting this code is publicly available at: <span><span>https://github.com/seemasafar/Reco-System-Using-LLM</span><svg><path></path></svg></span></div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"649 ","pages":"Article 130753"},"PeriodicalIF":5.5,"publicationDate":"2025-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144469899","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}