Neural Networks最新文献

{"title":"Backdoor attacks against Hybrid Classical-Quantum Neural Networks","authors":"Ji Guo ,&nbsp;Wenbo Jiang ,&nbsp;Rui Zhang ,&nbsp;Wenshu Fan ,&nbsp;Jiachen Li ,&nbsp;Guoming Lu ,&nbsp;Hongwei Li","doi":"10.1016/j.neunet.2025.107776","DOIUrl":"10.1016/j.neunet.2025.107776","url":null,"abstract":"<div><div>Hybrid Classical-Quantum Neural Networks (HQNNs) represent a promising advancement in Quantum Machine Learning (QML), yet their security has been rarely explored. In this paper, we present the first systematic study of backdoor attacks on HQNNs. We begin by proposing an attack framework and providing a theoretical analysis of the generalization bounds and minimum perturbation requirements for backdoor attacks on HQNNs. Next, we employ two classic backdoor attack methods on HQNNs and Convolutional Neural Networks (CNNs) to further investigate the robustness of HQNNs. Our experimental results demonstrate that HQNNs are more robust than CNNs, requiring more significant image modifications for successful attacks. Additionally, we introduce the Qcolor backdoor, which utilizes color shifts as triggers and employs the Non-dominated Sorting Genetic Algorithm II (NSGA-II) to optimize hyperparameters. Through extensive experiments, we demonstrate the effectiveness, stealthiness, and robustness of the Qcolor backdoor.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107776"},"PeriodicalIF":6.0,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144510735","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A link prediction method for multi-modal knowledge graphs based on Adaptive Fusion and Modality Information Enhancement","authors":"Zenglong Wang,&nbsp;Xuan Liu,&nbsp;Zheng Liu,&nbsp;Yu Weng,&nbsp;Chaomurilige","doi":"10.1016/j.neunet.2025.107771","DOIUrl":"10.1016/j.neunet.2025.107771","url":null,"abstract":"<div><div>Multi-modal knowledge graphs (MMKGs) enrich the semantic expression capabilities of traditional knowledge graphs by incorporating diverse modal information, showcasing immense potential in various knowledge reasoning tasks. However, existing MMKGs encounter numerous challenges in the link prediction task (i.e., knowledge graph completion reasoning), primarily due to the complexity and diversity of modal information and the imbalance in quality. These challenges make the efficient fusion and enhancement of multi-modal information difficult to achieve. Most existing methods adopt simple concatenation or weighted fusion of modal features, but such approaches fail to fully capture the deep semantic interactions between modalities and perform poorly when confronted with modal noise or missing information. To address these issues, this paper proposes a novel framework model—Adaptive Fusion and Modality Information Enhancement(AFME). This framework consists of two parts: the Modal Information Fusion module (MoIFu) and the Modal Information Enhancement module (MoIEn). By introducing a relationship-driven denoising mechanism and a dynamic weight allocation mechanism, the framework achieves efficient adaptive fusion of multi-modal information. It employs a generative adversarial network (GAN) structure to enable global guidance of structural modalities over feature modalities and adopts a multi-layer self-attention mechanism to optimize both intra- and inter-modal features. Finally, it jointly optimizes the losses of the triple prediction task and the adversarial generation task. Experimental results demonstrate that the AFME framework significantly improves multi-modal feature utilization and knowledge reasoning capabilities on multiple benchmark datasets, validating its efficiency and superiority in complex multi-modal scenarios.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107771"},"PeriodicalIF":6.0,"publicationDate":"2025-06-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144534933","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Refining visual token sequence for efficient image captioning","authors":"Tiantao Xian ,&nbsp;Zhiheng Zhou ,&nbsp;Wenlve Zhou ,&nbsp;Zhipeng Zhang","doi":"10.1016/j.neunet.2025.107759","DOIUrl":"10.1016/j.neunet.2025.107759","url":null,"abstract":"<div><div>In practical applications, both accuracy and speed are critical for image captioning (IC) systems. Recently, transformer-based architectures have significantly advanced the field of IC; however, these improvements often come at the cost of increased computational complexity and slower inference speeds. In this paper, we conduct a comprehensive analysis of the computational overhead of IC models and find that the visual encoding process accounts for the majority of this overhead. Considering the redundancy in visual information — where many regions are irrelevant or provide low information for prediction — we propose a knowledge-injection-based visual token <strong>R</strong>eduction module. This module estimates the importance of each token and retains only a subset of them. To minimize visual semantic loss, we introduce token <strong>F</strong>usion and <strong>I</strong>nsertion modules that supplement visual semantics by reusing discarded tokens and capturing global semantics. Based on this, our visual token sequence refinement strategy, referred to as RFI, is deployed at specific positions in the visual backbone to hierarchically compress the visual token sequence, thereby reducing the overall computational overhead of the model at its source. Extensive experiments demonstrate the effectiveness of the proposed method, showing that it can accelerate model inference without sacrificing performance. Additionally, the method allows for flexible trade-offs between accuracy and speed under different settings.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107759"},"PeriodicalIF":6.0,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144534934","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Neural networks with low-resolution parameters","authors":"Eduardo Lobo Lustosa Cabral ,&nbsp;Larissa Driemeier","doi":"10.1016/j.neunet.2025.107763","DOIUrl":"10.1016/j.neunet.2025.107763","url":null,"abstract":"<div><div>The expanding scale of large neural network models introduces significant challenges, driving efforts to reduce memory usage and enhance computational efficiency. Such measures are crucial to ensure the practical implementation and effective application of these sophisticated models across a wide array of use cases. This study examines the impact of parameter bit precision on model performance compared to standard 32-bit models, with a focus on multiclass object classification in images. The models analyzed include those with fully connected layers, convolutional layers, and transformer blocks, with model weight resolution ranging from 1 bit to 4.08 bits. The findings indicate that models with lower parameter bit precision achieve results comparable to 32-bit models, showing promise for use in memory-constrained devices. While low-resolution models with a small number of parameters require more training epochs to achieve accuracy comparable to 32-bit models, those with a large number of parameters achieve similar performance within the same number of epochs. Additionally, data augmentation can destabilize training in low-resolution models, but including zero as a potential value in the weight parameters helps maintain stability and prevents performance degradation. Overall, 2.32-bit weights offer the optimal balance of memory reduction, performance, and efficiency. However, further research should explore other dataset types and more complex and larger models. These findings suggest a potential new era for optimized neural network models with reduced memory requirements and improved computational efficiency, though advancements in dedicated hardware are necessary to fully realize this potential.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107763"},"PeriodicalIF":6.0,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144534937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Operating performance assessment of industrial process based on MIC-graph convolutional networks with local slow feature analysis","authors":"Lili Hao ,&nbsp;Fei Chu ,&nbsp;Tao Chen ,&nbsp;Mingxing Jia ,&nbsp;Fuli Wang","doi":"10.1016/j.neunet.2025.107773","DOIUrl":"10.1016/j.neunet.2025.107773","url":null,"abstract":"<div><div>To ensure the safe and stable operation of industrial processes, deep neural network-based operational performance assessment methods have been extensively adopted according to the latest research findings. However, existing industrial process performance assessment models often fail to account for the local spatial structure features and the slowly varying features from time series samples. Such limitations result in the suboptimal exploitation of spatial interaction information and hinder the models’ responsiveness to complex system state transitions, thereby impeding the precise assessment of industrial process performance. To this end, a maximum information coefficient-based graph convolutional networks (MIC-GCN) is proposed for operational performance assessment, which aims to effectively capture the intricate interactions of latent spatial structures embedded in temporal process data. First, a MIC-based graph construction method is employed to transform time series data into graph-structured data with nodes and edges, thereby preserving the local geometric structure of the original data and revealing high-dimensional spatial interaction information among data samples. Second, local slow feature analysis (SFA) is utilized to extract fine-grained dynamic correlation information from the spatial structure of the data. Furthermore, the Siamese GCNs are designed to simultaneously process graph-structured data samples at two consecutive time steps, which facilitates the capture of slowly varying feature representations embedded in the evolving topological structures. The proposed method can precisely extract and deeply mine spatiotemporal interactive features information, thereby enhancing the accuracy of performance assessment. Experimental validation on coal slurry flotation and dense medium coal preparation platforms confirms the method’s efficacy and reliability.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107773"},"PeriodicalIF":6.0,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144502053","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DiffCNN: A collaborative framework of diffusion model and CNN for semi-supervised medical image segmentation","authors":"Shanshan Xu ,&nbsp;Lixia Tian","doi":"10.1016/j.neunet.2025.107813","DOIUrl":"10.1016/j.neunet.2025.107813","url":null,"abstract":"<div><div>The highly prevalent teacher-student architecture has demonstrated great success in semi-supervised medical image segmentation. Despite its excellent performance, the architecture still faces two challenges: 1) the optimization of the teacher subnet relies heavily on the student subnet, and this greatly limits the capability of the teacher subnet; 2) the commonly used CNN-based structure for the construction of the teacher and student subnets cannot deal well with noisy medical images. To address these challenges, we propose DiffCNN, a collaborative framework of diffusion model and CNN for semi-supervised medical image segmentation. Unlike classic approaches that use two subnets of the same structure, our proposed DiffCNN employs two subnets of quite different structures. Specifically, in addition to a CNN subnet, DiffCNN also employs a diffusion subnet to alleviate the influences of noises through learning the underlying distribution of the mask. Collaborative training of the diffusion and CNN subnets makes it possible for the two subnets to learn from each other and accordingly extract complementary information from the input images more effectively. Furthermore, adversarial learning is involved to further enhance the capability of the diffusion subnet through forcing the diffusion-based segmentations to access real masks. We evaluate the performance of the proposed DiffCNN on three datasets, and the results demonstrate the superior performance of the DiffCNN over the state-of-the-art semi-supervised segmentation methods.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107813"},"PeriodicalIF":6.0,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144534932","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Intracortical functional connectivity during deep sleep reveals prosocial preferences","authors":"Andjela Markovic ,&nbsp;Lorena R.R. Gianotti ,&nbsp;Mirjam Studler ,&nbsp;Daria Knoch","doi":"10.1016/j.neunet.2025.107803","DOIUrl":"10.1016/j.neunet.2025.107803","url":null,"abstract":"<div><div>Prosociality is crucial for cohesive community functioning. Recently, we found that increased slow-wave activity during deep sleep in the right temporoparietal junction (TPJ), a key component of the social brain, is associated with stronger prosocial preferences. Building on this finding, we here investigate connectivity between the right TPJ and other regions of the social brain as a marker of inter-individual differences in prosocial preferences. Using whole-night high-density sleep electroencephalography (EEG) recordings from 54 participants (mean age = 21.5 ± 2 years; 28 females), we employed source localization to derive intracortical EEG functional connectivity during deep sleep in the slow-wave frequency range (0.8 to 4.6 Hz). We correlated connectivity between the right TPJ and 23 other key social brain regions with prosocial preferences, assessed in an incentivized public goods game with real monetary consequences. Our results revealed a negative correlation between prosocial preferences and connectivity for 22 of the 23 analyzed connections. Six of these connections demonstrated a significant negative correlation with prosocial preferences after adjusting for multiple testing (-0.36 ≤ rho ≤ -0.30; 0.006 ≤ <em>p</em> ≤ 0.038), indicating lower functional connectivity of the right TPJ with other social brain regions during deep sleep in individuals with stronger prosocial preferences. These results suggest that the benefits of deep sleep for prosocial decision-making may be enhanced when the right TPJ reduces its interactions within the social brain during this sleep stage.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107803"},"PeriodicalIF":6.0,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144548591","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MSC-transformer-based 3D-attention with knowledge distillation for multi-action classification of separate lower limbs","authors":"Heng Yan ,&nbsp;Zilu Wang ,&nbsp;Junhua Li","doi":"10.1016/j.neunet.2025.107806","DOIUrl":"10.1016/j.neunet.2025.107806","url":null,"abstract":"<div><div>Deep learning has been extensively applied to motor imagery (MI) classification using electroencephalogram (EEG). However, most existing deep learning models do not extract features from EEG using dimension-specific attention mechanisms based on the characteristics of each dimension (e.g., spatial dimension), while effectively integrate local and global features. Furthermore, implicit information generated by the models has been ignored, leading to underutilization of essential information of EEG. Although MI classification has been relatively thoroughly investigated, the exploration of classification including real movement (RM) and motor observation (MO) is very limited, especially for separate lower limbs. To address the above problems and limitations, we proposed a multi-scale separable convolutional Transformer-based filter-spatial-temporal attention model (MSC-T3AM) to classify multiple lower limb actions. In MSC-T3AM, spatial attention, filter and temporal attention modules are embedded to allocate appropriate attention to each dimension. Multi-scale separable convolutions (MSC) are separately applied after the projections of query, key, and value in self-attention module to improve computational efficiency and classification performance. Furthermore, knowledge distillation (KD) was utilized to help model learn suitable probability distribution. The comparison results demonstrated that MSC-T3AM with online KD achieved best performance in classification accuracy, exhibiting an elevation of 2 %-19 % compared to a few counterpart models. The visualization of features extracted by MSC-T3AM with online KD reiterated the superiority of the proposed model. The ablation results showed that filter and temporal attention modules contributed most for performance improvement (improved by 2.8 %), followed by spatial attention module (1.2 %) and MSC module (1 %). Our study also suggested that online KD was better than offline KD and the case without KD. The code of MSC-T3AM is available at: <span><span>https://github.com/BICN001/MSC-T3AM</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107806"},"PeriodicalIF":6.0,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144548590","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"PRformer: Pyramidal recurrent transformer for multivariate time series forecasting","authors":"Yongbo Yu ,&nbsp;Weizhong Yu ,&nbsp;Feiping Nie ,&nbsp;Zongcheng Miao ,&nbsp;Ya Liu ,&nbsp;Xuelong Li","doi":"10.1016/j.neunet.2025.107769","DOIUrl":"10.1016/j.neunet.2025.107769","url":null,"abstract":"<div><div>The self-attention mechanism in Transformer architecture, invariant to sequence order, necessitates positional embeddings to encode temporal order in time series prediction. We argue that this reliance on positional embeddings restricts the Transformer’s ability to effectively represent temporal sequences, particularly when employing longer lookback windows. To address this, we introduce an innovative approach that combines Pyramid RNN embeddings (PRE) for univariate time series with the Transformer’s capability to model multivariate dependencies. PRE, utilizing pyramidal one-dimensional convolutional layers, constructs multiscale convolutional features that preserve temporal order. Additionally, RNNs, layered atop these features, learn multiscale time series representations sensitive to sequence order. This integration into Transformer models with attention mechanisms results in significant performance enhancements. We present the PRformer, a model integrating PRE with a standard Transformer encoder, demonstrating state-of-the-art performance on various real-world datasets. This performance highlights the effectiveness of our approach in leveraging longer lookback windows and underscores the critical role of robust temporal representations in maximizing Transformer’s potential for prediction tasks.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107769"},"PeriodicalIF":6.0,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144489828","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Bayesian inverse problems with conditional Sinkhorn generative adversarial networks in least volume latent spaces","authors":"Qiuyi Chen ,&nbsp;Panagiotis Tsilifis ,&nbsp;Mark Fuge","doi":"10.1016/j.neunet.2025.107740","DOIUrl":"10.1016/j.neunet.2025.107740","url":null,"abstract":"<div><div>Solving inverse problems in scientific and engineering fields has long been intriguing and holds great potential for many applications, yet most techniques still struggle to address issues such as high dimensionality, nonlinearity and model uncertainty inherent in these problems. Recently, generative models such as Generative Adversarial Networks (GANs) have shown great potential in approximating complex high dimensional conditional distributions and have paved the way for characterizing posterior densities in Bayesian inverse problems, yet the problems’ high dimensionality and high nonlinearity often impedes the model’s training. In this paper we show how to tackle these issues with Least Volume—a novel unsupervised nonlinear dimension reduction method—that can learn to represent the given datasets with the minimum number of latent variables while estimating their intrinsic dimensions. Once the low dimensional latent spaces are identified, efficient and accurate training of conditional generative models becomes feasible, resulting in a latent conditional GAN framework for posterior inference. We demonstrate the power of the proposed methodology on a variety of applications including inversion of parameters in systems of ODEs and high dimensional hydraulic conductivities in subsurface flow problems, and reveal the impact of the observables’ and unobservables’ intrinsic dimensions on inverse problems.</div></div>","PeriodicalId":49763,"journal":{"name":"Neural Networks","volume":"191 ","pages":"Article 107740"},"PeriodicalIF":6.0,"publicationDate":"2025-06-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144530734","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}