Frontiers in Computational Neuroscience最新文献

System-level brain modeling is a powerful method for building computational models of the brain and allows biologically motivated models to produce measurable behavior that can be tested against empirical data. System-level brain models occupy an intermediate position between detailed neuronal circuit models and abstract cognitive models. They are distinguished by their structural and functional resemblance to the brain, while also allowing for thorough testing and evaluation. In designing system-level brain models, several questions need to be addressed. What are the components of the system? At what level should these components be modeled? How are the components connected-that is, what is the structure of the system? What is the function of each component? What kind of information flows between the components, and how is that information coded? We mainly address models of cognitive abilities or subsystems that produce measurable behavior rather than models that to reproduce internal states, signals or activation patterns. In this method paper, we argue that system-level modeling is an excellent method for addressing complex cognitive and behavioral phenomena.

Introduction: The increasing integration of large language models (LLMs) into human-AI collaboration necessitates a deeper understanding of their cognitive impacts on users. Traditional evaluation methods have primarily focused on task performance, overlooking the underlying neural dynamics during interaction.

Methods: In this study, we introduce a novel framework that leverages electroencephalography (EEG) signals to assess how LLM interactions affect cognitive processes such as attention, cognitive load, and decision-making. Our framework integrates an Interaction-Aware Language Transformer (IALT), which enhances token-level modeling through dynamic attention mechanisms, and an Interaction-Optimized Reasoning Strategy (IORS), which employs reinforcement learning to refine reasoning paths in a cognitively aligned manner.

Results: By coupling these innovations with real-time neural data, the framework provides a fine-grained, interpretable assessment of LLM-induced cognitive changes. Extensive experiments on four benchmark EEG datasets Database for Emotion Analysis using Physiological Signals (DEAP), A Dataset for Affect, Personality and Mood Research on Individuals and Groups (AMIGOS), SJTU Emotion EEG Dataset (SEED), and Database for Emotion Recognition through EEG and ECG Signals (DREAMER) demonstrate that our method outperforms existing models in both emotion classification accuracy and alignment with cognitive signals. The architecture maintains high performance across varied EEG configurations, including low-density, noise-prone portable systems, highlighting its robustness and practical applicability.

Discussion: These findings offer actionable insights for designing more adaptive and cognitively aware LLM systems, and open new avenues for research at the intersection of artificial intelligence and neuroscience.

Spinal health forms the cornerstone of the overall human body functionality with the lumbar spine playing a critical role and prone to various types of injuries due to inflammation and diseases, including lumbar vertebral fractures. This paper proposes automated method for segmentation of lumbar vertebral body (VB) using image processing techniques such as shape features and morphological operations. This entails an initial phase of image preprocessing, followed by detection and localizing of vertebral regions. Subsequently, vertebral are segmented and labeled, with each classified into normal or fractured using classification techniques, k-nearest neighbors (KNN) and support vector machines (SVM). The methodology leverages unique vertebral characteristics like gray scales, shape features, and textural elements through a range of machine learning methods. The approach is assessed and validated on a clinical spine dataset dice score used for segmentation, achieving an average accuracy rate of 95%, and for classification, achieving average accuracy of 97.01%.

ECoG signals are widely used in Brain-Computer Interfaces (BCIs) due to their high spatial resolution and superior signal quality, particularly in the field of neural control. ECoG enables more accurate decoding of brain activity compared to traditional EEG. By obtaining cortical ECoG signals directly from the cerebral cortex, complex motor commands, such as finger movement trajectories, can be decoded more efficiently. However, existing studies still face significant challenges in accurately decoding finger movement trajectories. Specifically, current models tend to confuse the movement information of different fingers and fail to fully exploit the dependencies within time series when predicting long sequences, resulting in limited decoding performance. To address these challenges, this paper proposes a novel decoding method that transforms 2D ECoG data samples into 3D spatio-temporal spectrograms with time-stamped features via wavelet transform. The method further enables accurate decoding of finger bending by using a 1D convolutional network composed of Dilated-Transposed convolution, which together extract channel band features and temporal variations in tandem. The proposed method achieved the best performance among three subjects in BCI Competition IV. Compared with existing studies, our method made the correlation coefficient between the predicted multi-finger motion trajectory and the actual multi-finger motion trajectory exceed 80% for the first time, with the highest correlation coefficient reaching 82%. This approach provides new insights and solutions for high-precision decoding of brain-machine signals, particularly in precise command control tasks, and advances the application of BCI systems in real-world neuroprosthetic control.

Mind-wandering reflects a dynamic interplay between focused attention and off-task mental states. Despite its relevance in understanding fundamental cognitive processes, such as attention regulation, decision-making, and creativity, previous models have not yet provided an account of the neural mechanisms for autonomous shifts between focus state (FS) and mind-wandering (MW). To address this, we conduct model simulation experiments employing predictive coding as a theoretical framework of perception to investigate possible neural mechanisms underlying these autonomous shifts between the two states. In particular, we modeled perception processes of continuous sensory sequences using our previously proposed variational RNN model under free energy minimization. The current study extends this model by introducing an online adaptation mechanism of a meta-level parameter, referred to as the meta-prior w, which regulates the complexity term in the free energy minimization. Our simulation experiments demonstrated that autonomous shifts between FS and MW take place when w switches between low and high values responding to a decrease and increase of the average reconstruction error over a past time window. Particularly, high w prioritized top-down predictions while low w emphasized bottom-up sensations. In this work, we speculate that self-awareness of MW may occur when the error signal accumulated over time exceeds a certain threshold. Finally, this paper explores how our experiment results align with existing studies and highlights their potential for future research.

Here we present evidence for the ubiquity of fine spike timing and temporal coding broadly observed across sensory systems and widely conserved across diverse phyla, spanning invertebrates and vertebrates. A taxonomy of basic neural coding types includes channel activation patterns, temporal patterns of spikes, and patterns of spike latencies. Various examples and types of combination temporal-channel codes are discussed, including firing sequence codes. Multiplexing of temporal codes and mixed channel-temporal codes are considered. Neurophysiological and perceptual evidence for temporal coding in many sensory modalities is surveyed: audition, mechanoreception, electroreception, vision, gustation, olfaction, cutaneous senses, proprioception, and the vestibular sense. Precise phase-locked, phase-triggered, and spike latency codes can be found in many sensory systems. Temporal resolutions on millisecond and submillisecond scales are common. General correlation-based representations and operations are discussed. In almost every modality, there is some role for temporal coding, often in surprising places, such as color vision and taste. More investigations into temporal coding are well-warranted.

Contactless Human Activity Recognition (HAR) has played a critical role in smart healthcare and elderly care homes to monitor patient behavior, detect falls or abnormal activities in real time. The effectiveness of non-invasive HAR is often hindered by location-centric variations in Channel State Information (CSI). These variations limit the ability of HAR models to generalize across new unseen cross-domain environments, for instance, a model trained in one location might not perform well in another physical location. To address this challenge, in this study, we present a novel federated learning (FL) algorithm designed to train a robust global model from local datasets in different localizations. The proposed Federated Weighted Averaging for HAR (Fed-WAHAR) algorithm mitigates location-induced disparities, including heterogeneity and non-Independent and Identically Distributed (non-IID) data distributions. Fed-WAHAR employs a dynamic weighting approach based on local models' accuracy to improve global model classification accuracy and reduce convergence time effectively. We evaluated the performance of Fed-WAHAR using various metrics, including accuracy, precision, recall, F1 score, confusion matrix, and convergence analysis. Experimental results demonstrate that Fed-WAHAR achieves an accuracy of 85% in recognizing human activities across different locations, enhancing the ability of model to infer across new unseen locations.

Understanding how language and linguistic constructions are processed in the brain is a fundamental question in cognitive computational neuroscience. This study builds directly on our previous work analyzing Argument Structure Constructions (ASCs) in the BERT language model, extending the investigation to a simpler, brain-constrained architecture: a recurrent neural language model. Specifically, we explore the representation and processing of four ASCs-transitive, ditransitive, caused-motion, and resultative-in a Long Short-Term Memory (LSTM) network. We trained the LSTM on a custom GPT-4-generated dataset of 2,000 syntactically balanced sentences. We then analyzed the internal hidden layer activations using Multidimensional Scaling (MDS) and t-Distributed Stochastic Neighbor Embedding (t-SNE) to visualize sentence representations. The Generalized Discrimination Value (GDV) was calculated to quantify cluster separation. Our results show distinct clusters for the four ASCs across all hidden layers, with the strongest separation observed in the final layer. These findings are consistent with our earlier study based on a large language model and demonstrate that even relatively simple RNNs can form abstract, construction-level representations. This supports the hypothesis that hierarchical linguistic structure can emerge through prediction-based learning. In future work, we plan to compare these model-derived representations with neuroimaging data from continuous speech perception, further bridging computational and biological perspectives on language processing.

EEG emotion recognition has important applications in human-computer interaction and mental health assessment, but existing models have limitations in capturing the complex spatial and temporal features of EEG signals. To overcome this problem, we propose an innovative model that combines CNN-BiLSTM and DC-IGN and fused both outputs for sentiment classification via a fully connected layer. In addition, we use a piecewise exponential decay strategy to optimize the training process. We conducted a comprehensive comparative experiment on the SEED and DEAP datasets, it includes traditional models, existing advanced models, and different combination models (such as CNN + LSTM, CNN + LSTM+DC-IGN). The results show that our model achieves 94.35% accuracy on SEED dataset, 89.84% on DEAP-valence, 90.31% on DEAP-arousal, which is significantly better than other models. In addition, we further verified the superiority of the model through subject independent experiment and learning rate scheduling strategy comparison experiment. These results not only improve the performance of EEG emotion recognition, but also provide new ideas and methods for research in related fields, and prove the significant advantages of our model in capturing complex features and improving classification accuracy.

Mathematical analysis of biological neural networks, specifically inhibitory networks with all-to-all connections, is challenging due to their complexity and non-linearity. In examining the dynamics of individual neurons, many fast currents are involved solely in spike generation, while slower currents play a significant role in shaping a neuron's behavior. We propose a discrete map approach to analyze the behavior of inhibitory neurons that exhibit bursting modulated by slow calcium currents, leveraging the time-scale differences among neural currents. This discrete map tracks the number of spikes per burst for individual neurons. We compared the map's predictions for the number of spikes per burst and the long-term system behavior to data obtained from the continuous system. Our findings demonstrate that the discrete map can accurately predict the canonical behavioral signatures of bursting performance observed in the continuous system. Specifically, we show that the proposed map a) accounts for the dependence of the number of spikes per burst on initial calcium levels, b) explains the roles of individual currents in shaping the system's behavior, and c) can be explicitly analyzed to determine fixed points and assess their stability.