International journal of neural systems最新文献

Designing effective neural architectures remains a central challenge in deep learning, and Neural Architecture Search (NAS) has become a popular tool for automating this process. However, many existing NAS approaches depend on hand-crafted architecture descriptors or shallow performance predictors, which fail to capture the structural complexity of candidate networks and often lead to unreliable search guidance. We introduce Graph Embedding Comparator with Isomorphic Multi-Comparison (GEC-IMC), an evolutionary NAS framework that learns architecture representations directly from their graph structure. A graph convolutional network encodes architectures into embeddings, while a contrastive learning strategy ensures that architectures with similar accuracy are mapped closer in the embedding space. On top of these embeddings, a comparator estimates the relative performance between two architectures, enabling more precise pairwise assessments during search. To further increase robustness, GEC-IMC incorporates an isomorphic multi-comparison mechanism, which evaluates multiple structurally equivalent variants of each architecture and aggregates their pairwise outcomes into a global score. This ranking score provides consistent feedback for evolutionary selection. Experiments on standard NAS benchmarks demonstrate that GEC-IMC achieves state-of-the-art performance with improved robustness over existing predictors. Ablation studies confirm the complementary roles of embedding learning and multi-comparison in enhancing search efficiency.

Generative adversarial network (GAN) architecture search aims to automate the discovery of high-performance network structures. While differentiable search methods like DAMGAN have shown promise, their reliance on inefficient SENet modules and adversarial loss for attention training limits both search efficiency and architecture quality. To address these limitations, we propose EAMGAN, an enhancing differentiable GAN architecture search with efficient attention and Fréchet Inception Distance (FID) guidance. Our approach introduces a lightweight attention module ECA-Net that replaces the fully-connected layers in SENet with a 1D convolution utilizing local context and weight sharing, thereby significantly reducing parameter count. Furthermore, we decouple the attention training from adversarial optimization and introduce a customized loss function based on the FID, which directly guides the architecture selection toward subnets that generate higher-quality images. Experiments on CIFAR-10 show that EAMGAN not only surpasses DAMGAN (Inception Score (IS): 9.03 versus 8.99, FID: 9.43 versus 10.27) but also achieves this with lower search cost (0.08 versus 0.09 GPU days). Competitive results on STL-10 further demonstrate its effectiveness and transferability.

Epileptic seizure prediction based on electroencephalogram (EEG) signals is one of the critical applications of medical artificial intelligence (AI), with considerable clinical potential for improving the quality of life of patients through early warnings. However, existing prediction models face dual challenges: insufficient feature representation and limited explainability of the decision. To address these challenges, this study proposes a dynamic multiscale cross-band fusion filter network (MCFNet) for end-to-end seizure prediction. Specifically, the model first decomposes EEG signals into multiscale components and incorporates a cross-band fusion attention mechanism to achieve multi-granularity signal fusion. Subsequently, the synchronous spectral filtering network, comprising both static and dynamic filtering modules, is designed to capture the periodic components and cross-channel dependencies in EEG signals. Notably, two explainable methods are introduced: a joint feature visualization strategy and an efficient feature ablation analysis, helping to bridge the gap between the "black-box" nature of deep learning and clinical needs. Evaluated on the CHB-MIT dataset, MCFNet achieves a sensitivity of 97.13%, a specificity of 97.22%, and a false positive rate (FPR) of 0.0326/h. Experimental results show that MCFNet not only exhibits superior predictive performance but also maintains a low FPR, offering a feasible scheme for clinical application of EEG-based seizure prediction.

The vagus nerve (VN) mediates bidirectional communication between the body and brain to maintain physiological homeostasis; likewise, alterations in ongoing vagal signaling may be indicators of disease and/or contribute to disease pathogenesis. Even though extensively documented in acute experiments, ongoing vagal activity has not been characterized longitudinally, over days or weeks, in mice, a preferred preclinical model. In addition, even though many VN recordings in mice occur during anesthesia, the effects of anesthesia on vagal signaling are unknown. This study uses a chronic implant mouse model to record vagal activity in anesthetized and awake, behaving animals for an average of 10 weeks and up to 6 months. Individual compound action potentials (CAPs) are tracked across multiple days by quantifying comparisons in features, including firing rates, waveform shape, inter-CAP interval histograms, and phase-locking to cardiac and respiratory signals, while demonstrating long-term electrode-nerve interface viability and stable signal-to-noise ratios. Additionally, cytokine challenge experiments produced detectable CAP responses up to 3 months after electrode implantation. Lastly, awake recordings incorporated video analysis to identify and remove motion artifacts to preserve and extract neural and cardiac recordings during daylight in-cage behavior. Results reveal diverse CAP populations with diverse physiological coupling and firing rates modulated by anesthesia. This work highlights the potential of chronic VN recordings to assess long-term changes in vagal activity in health and disease, with implications for the discovery of autonomic markers of disease and closed-loop VNS stimulation strategies.

The behavioral inhibition system (BIS), mediating responses to punishment cues and avoidance behaviors, is implicated in anxiety. However, the neural dynamics underpinning BIS, particularly regarding the temporal variability of brain network interactions, remain less explored. Using resting-state functional magnetic resonance imaging (rs-fMRI) of 181 healthy adults, this study investigated the association between BIS sensitivity and the temporal variability of functional connectivity within and between functional brain networks. This finding revealed a significant positive correlation between BIS scores and temporal variability, specifically in the connectivity involving subnetworks' sensory somatomotor hand network (SSHN)-ventral attention network (VAN), and sensory somatomotor mouth network (SSMN)-VAN. Notably, the high-BIS sensitivity group exhibited significantly greater temporal variability between VAN and SSMN/SSHN compared to the low-BIS sensitivity group. Furthermore, predicted BIS scores based on network variability showed a strong correlation with actual BIS scores (Pearson's [Formula: see text]). Moreover, significant mediation effects highlighted the bridging role of BIS scores between brain network variability and anxiety scale scores. This enhances the comprehension of the relationship between BIS, anxiety, and brain function, while also offering new insights into the pathogenesis of anxiety.

High-density surface electromyogram (HD-sEMG) has become a powerful signal source for hand gesture recognition. However, existing approaches suffer from limited feature diversity in hand-crafted methods and high data dependency in deep learning models, necessitating individual model calibration for each user due to neuromuscular differences. We propose EMG-ROCKET, an enhanced version of the RandOm Convolutional KErnel Transform (ROCKET), designed to extract diverse and robust HD-sEMG features without prior knowledge or extensive training. EMG-ROCKET integrates random channel fusion and enhanced aggregation functions to enhance robustness against cross-day signal variability in HD-sEMG applications. In cross-day evaluations of hand gesture recognition, a Ridge classifier using EMG-ROCKET features achieved 84.3% and 77.8% accuracy on two HD-sEMG datasets, outperforming all baseline methods. Furthermore, feature contribution analysis demonstrates the capability of EMG-ROCKET to capture spatial muscle activation patterns, offering insights into motion mechanisms. These results establish EMG-ROCKET as a promising, training-free solution for robust HD-sEMG feature extraction, facilitating practical human-machine interaction applications.

Seizure prediction from Electroencephalogram (EEG) signals is a critical task for proactive intervention in epilepsy management. Existing models often struggle to capture high-order inter-channel dependencies dynamically and adapt to the spectral variations preceding seizure onset, especially in cross-patient scenarios. To address these issues, a novel Unified Hypergraph-Mamba (UHM) framework, which for the first time integrates hypergraph-based spatial modeling with Mamba-based adaptive spectral modeling. Specifically, a hypergraph attention mechanism is designed to capture high-order spatial interactions among EEG channels, enabling dynamic representation of inter-channel dependencies. Concurrently, an adaptive spectral modeling module based on the Mamba architecture selectively emphasizes frequency components most indicative of preictal states. Together, these components form a unified architecture capable of jointly modeling spatiotemporal EEG dynamics. Extensive experiments conducted on both patient-specific and cross-patient settings demonstrate that our model consistently outperforms state-of-the-art baselines, achieving superior sensitivity and AUC.

Most existing studies analyzed the resting-state electroencephalogram (EEG) of DOC patients, and recent research demonstrated that the passive auditory paradigm was helpful for bedside detection of DOC and better captured sensory and cognitive responses. However, further studies of classification algorithms were needed for consciousness assessment in DOC based on task-state EEG data. In this study, EEG data from minimally conscious state (MCS) patients, vegetative state (VS) patients, and a healthy control group (HC) were collected using an auditory oddball paradigm. First, compared to the fragmented features adopted by most studies, multiple effective biomarkers for consciousness assessment in the time-frequency domains, connectivity and nonlinear dynamics were identified. Event-related potentials (ERP) results showed that MCS and VS patients exhibited lower N100 and MMN amplitudes than the HC group. Spectral analysis results indicated that VS patients had higher Delta power, and lower Alpha and Beta power than the MCS and HC groups. Second, different from insufficient classifiers in previous studies, this study systematically compared the performance of multiple machine learning and deep learning (DL) classifiers, including support vector machine (SVM), linear discriminant analysis (LDA), random forest (RF), eXtreme Gradient Boosting (XGBoost), decision tree (DT), EEGNet and ShallowConvNet. For machine learning methods, SVM and RF had an advantage in binary classification, and SVM had better performance in three-class classification. Among all individual classifiers, Shallow ConvNet had the best performance for binary and three-class classification. Moreover, an ensemble model incorporating all seven classifiers was proposed using a voting strategy, and further improved classification performance that was superior to existing studies. In addition, the importance of each feature was analyzed, identifying N100, MMN, Delta, Alpha, and Beta power as significant biomarkers of consciousness assessment.

Federated learning (FL), as a method that coordinates multiple clients to train models together without handing over local data, is naturally privacy-preserving for data. However, there is still a risk that malicious attackers can steal intermediate parameters and infer the user's original data during the model training, thereby leaking sensitive data privacy. To address the above problems, we propose an adaptive differential privacy blockchain federated learning (ADP-BCFL) method to accomplish the compliant use of distributed data while ensuring security. First, utilize blockchain to accomplish secure storage and valid querying of user summary data. Second, propose an adaptive DP mechanism to be applied in the process of federal learning, which adaptively adjusts the threshold size of parameter tailoring according to the parameter characteristics, controls the amount of introduced noise, and ensures a good global model accuracy while effectively solving the problem of inference attack. Finally, the ADP-BCFL method was validated on the MNIST, Fashion MNIST datasets and spatiotemporal dataset to effectively balance model performance and privacy.