Frontiers in Computational Neuroscience最新文献

The rapid growth of computational neuroscience and brain-computer interface (BCI) technologies require efficient, scalable, and biologically compatible approaches for neural data acquisition and interpretation. Traditional sensors and signal processing pipelines often struggle with the high dimensionality, temporal variability, and noise inherent in neural signals, particularly in elderly populations where continuous monitoring is essential. Triboelectric nanogenerators (TENGs), as self-powered and flexible multi-sensing devices, offer a promising avenue for capturing neural-related biophysical signals such as electroencephalography (EEG), electromyography (EMG), and cardiorespiratory dynamics. Their low-power and wearable characteristics make them suitable for long-term health and neurocognitive monitoring. When combined with deep learning models-including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and spiking neural networks (SNNs)-TENG-generated signals can be efficiently decoded, enabling insights into neural states, cognitive functions, and disease progression. Furthermore, neuromorphic computing paradigms provide an energy-efficient and biologically inspired framework that naturally aligns with the event-driven characteristics of TENG outputs. This mini review highlights the convergence of TENG-based sensing, deep learning algorithms, and neuromorphic systems for neural data interpretation. We discuss recent progress, challenges, and future perspectives, with an emphasis on applications in computational neuroscience, neurorehabilitation, and elderly health care.

The brain is a highly diverse and heterogeneous network, yet the functional role of this neural heterogeneity remains largely unclear. Despite growing interest in neural heterogeneity, a comprehensive understanding of how it influences computation across different neural levels and learning methods is still lacking. In this work, we systematically examine the neural computation of spiking neural networks (SNNs) in three key sources of neural heterogeneity: external, network, and intrinsic heterogeneity. We evaluate their impact using three distinct learning methods, which can carry out tasks ranging from simple curve fitting to complex network reconstruction and real-world applications. Our results show that while different types of neural heterogeneity contribute in distinct ways, they consistently improve learning accuracy and robustness. These findings suggest that neural heterogeneity across multiple levels improves learning capacity and robustness of neural computation, and should be considered a core design principle in the optimization of SNNs.

Considering biological constraints in artificial neural networks has led to dramatic improvements in performance. Nevertheless, to date, the positivity of long-range signals in the cortex has not been shown to yield improvements. While Non-negative matrix factorization (NMF) captures biological constraints of positive long-range interactions, deep convolutional neural networks with NMF modules do not match the performance of conventional neural networks (CNNs) of a similar size. This work shows that introducing intermediate modules that combine the NMF's positive activities, analogous to the processing in cortical columns, leads to improved performance on benchmark data that exceeds that of vanilla deep convolutional networks. This demonstrates that including positive long-range signaling together with local interactions of both signs in analogy to cortical hyper-columns has the potential to enhance the performance of deep networks.

The rapid growth of large-scale neuroscience datasets has spurred diverse modeling strategies, ranging from mechanistic models grounded in biophysics, to phenomenological descriptions of neural dynamics, to data-driven deep neural networks (DNNs). Each approach offers distinct strengths as mechanistic models provide interpretability, phenomenological models capture emergent dynamics, and DNNs excel at predictive accuracy but this also comes with limitations when applied in isolation. Universal differential equations (UDEs) offer a unifying modeling framework that integrates these complementary approaches. By treating differential equations as parameterizable, differentiable objects that can be combined with modern deep learning techniques, UDEs enable hybrid models that balance interpretability with predictive power. We provide a systematic overview of the UDE framework, covering its mathematical foundations, training methodologies, and recent innovations. We argue that UDEs fill a critical gap between mechanistic, phenomenological, and data-driven models in neuroscience, with potential to advance applications in neural computation, neural control, neural decoding, and normative modeling in neuroscience.

Sleep is known to support memory consolidation through a complex interplay of neural dynamics across multiple timescales. Using intracranial EEG (iEEG) recordings from patients undergoing clinical monitoring, we characterize spectral activity, neuronal avalanche dynamics, and temporal correlations across sleep-wake states, with a focus on their spatial distribution and potential functional relevance. We observe increased low-frequency power, larger avalanches, and enhanced long-range temporal correlations-quantified via Detrended Fluctuation Analysis-during N2 and N3 sleep. In contrast, REM sleep and wakefulness show reduced temporal persistence and fewer large-scale cascades, suggesting a shift toward more fragmented and flexible dynamics. These signatures vary across cortical regions, with distinctive patterns emerging in medial temporal and frontal areas-regions implicated in memory processing. Rather than providing direct evidence of consolidation, our results point to a functional neural landscape that may favor both stabilization and reconfiguration of internal representations during sleep. Overall, our findings highlight the utility of iEEG in revealing the multiscale spatio-temporal structure of sleep-related brain dynamics, offering insights into the physiological conditions that support memory-related processing.

While acquisition curves in human learning averaged at the group level display smooth, gradual changes in performance, individual learning curves across cognitive domains reveal sudden, discontinuous jumps in performance. Similar thresholding effects are a hallmark of a range of nonlinear systems which can be explored using simple, abstract models. Here, I investigate discontinuous changes in learning performance using Amari-Hopfield networks with Hebbian learning rules which are repeatedly exposed to a single stimulus. Simulations reveal that the attractor basin size for a target stimulus increases in discrete jumps rather than gradual changes with repeated stimulus exposure. The distribution of the size of these positive jumps in basin size is best approximated by a lognormal distribution, suggesting that the distribution is heavy-tailed. Examination of the transition graph structure for networks before and after basin size changes reveals that newly acquired states are often organized into hierarchically branching tree structures, and that the distribution of branch sizes is best approximated by a power law distribution. The findings suggest that even simple nonlinear network models of associative learning exhibit discontinuous changes in performance with repeated learning which mirror behavioral results observed in humans. Future work can investigate similar mechanisms in more biologically detailed network models, potentially offering insight into the network mechanisms of learning with repeated exposure or practice.

Massive computing systems are required to train neural networks. The prodigious amount of consumed energy makes the creation of AI applications significant polluters. Despite the enormous training effort, neural network error rates limit its use for medical applications, because errors can lead to intolerable morbidity and mortality. Two reasons contribute to the excessive training requirements and high error rates; an iterative reinforcement process (tuning) that does not guarantee convergence and the deployment of neuron models only capable of realizing linearly separable switching functions. tuning procedures require tens of thousands of training iterations. In addition, linearly separable neuron models have severely limited capability; which leads to large neural nets. For seven inputs, the ratio of total possible switching functions to linearly separable switching functions is 41 octillion. Addressed here is the creation of neuron models for the application of disease diagnosis. Algorithms are described that perform direct neuron creation. This results in far fewer training steps than that of current AI systems. The design algorithms result in neurons that do not manufacture errors (hallucinations). The algorithms utilize a template to create neuron models that are capable of performing any type of switching function. The algorithms show that a neuron model capable of performing both linearly and nonlinearly separable switching functions is vastly superior to the neuron models currently being used. Included examples illustrate use of the template for determining disease diagnoses (outputs) from symptoms (inputs). The examples show convergence with a single training iteration.

Introduction: It has become increasingly common to record brain activity simultaneously at more than one spatiotemporal scale. Here, we address a central question raised by such cross-scale datasets: do they reflect the same underlying dynamics observed in different ways, or different dynamics observed in the same way? In other words, to what extent can variation between modalities be attributed to system-level versus observer-level effects? System-level effects reflect genuine differences in neural dynamics at the resolution sampled by each device. Observer-level effects, by contrast, reflect artefactual differences introduced by the nonlinear transformations each device imposes on the signal. We demonstrate that noise, when incorporated into generative models, can help disentangle these two sources of variation.

Methods: We apply this noise-based approach to simultaneously recorded high-frequency broadband signals from macroelectrodes and microwires in the human hippocampus.

Results: Most subjects show a complex mixture of system- and observer-level contributions to their time series. However, in one subject, the cross-scale difference is statistically attributable to an observer-level effect-i.e., consistent with the same dynamics at both microwire and macroelectrode scales.

Discussion: This study shows that noise can be used in empirical datasets to determine whether cross-scale variation arises from differences in neural dynamics or differences in observer functions.

Seizure detection in a timely and accurate manner remains a primary challenge in clinical neurology, affecting diagnosis planning and patient management. Most of the traditional methods rely on feature extraction and traditional machine learning techniques, which are not efficient in capturing the dynamic characteristics of neural signals. It is the aim of this study to address such limitations by designing a deep learning model from bidirectional Long Short-Term Memory (BiLSTM) networks in a bid to enhance epileptic seizure identification reliability and accuracy. The dataset used, drawn from Kaggle's Epileptic Seizure Recognition challenge, consists of 11,500 samples with 179 features per sample corresponding to different electroencephalogram (EEG) readings. Data preprocessing was utilized to normalize and structure the input to the deep learning model. The proposed BiLSTM model employs sophisticated architecture to leverage temporal dependency and bidirectional data flows. It incorporates multiple dense and dropout layers alongside batch normalization to enhance the capability of the model in learning from the EEG data in an efficient manner. It supports end-to-end feature learning from the raw EEG signals without the need for intensive preprocessing and feature engineering. BiLSTM model performed better than others with 98.70% accuracy on the validation set and surpassed traditional techniques. The F1-score and other statistical metrics also validated the performance of the model as the confusion matrix achieved high values for recall and precision. The results confirm the capability of bidirectional LSTM networks to better identify seizures with significant improvements over conventional practices. Apart from facilitating seizure detection in a reliable fashion, the method improves the overall field of biomedical signal processing and can also be used in real-time observation and intervention protocols.