Frontiers in Computational Neuroscience最新文献

Introduction: People with post-stroke aphasia (PSA) exhibit significant interindividual variability attributed to distinctive network disruption patterns across individuals. This complexity limits the effectiveness of conventional one-size-fits-all brain stimulation approaches, but to date no individualized tACS targeting on functional network was studied in PSA. This two-phase study aimed to investigate the immediate network-modulation and language-facilitation effects of dual-site in-phase tACS utilizing a novel individualized targeting method based on individual's EEG dysfunctome.

Methods: In the first phase, network-based linear regression was used to identify aphasia-severity-predictive dysfunctome from the speech-production EEG data of 15 Cantonese-speaking people with aphasia (PWA). Individualized stimulation targets were determined using two targeting principles. Restoration-based targeting aims to restore a target edge which is centralized within the target dysfunctome but weakly-connected in the individual, whereas enhancement-based targeting selects a strongly-connected target edge. The second phase involved a single-session double-blinded sham-controlled trial with the same group to evaluate the immediate effects of dual-site 7-Hz 1-mA tACS under four conditions: Restoration In-phase (RI), Enhancement In-phase (EI), Enhancement Anti-phase (EA), and Sham (SH).

Results: In the first phase, we explored a range of frequency bands and EEG tasks and identified a left frontal-temporal theta network under divergent naming task that significantly predicted aphasia severity. The single-session clinical trial in the second phase demonstrated that RI condition produced increases in the target node strength, global network properties, and divergent naming performance, which were absent in sham and the other two real stimulation conditions.

Discussion: This was the first-of-its-kind dysfunctome-based data-driven individualized tACS demonstrated immediate neuromodulatory effects in PSA. The findings suggest that EEG dysfunctome can help pinpointing effective individualized targets for tACS to promote clinically-beneficial functional reorganization. Despite limited generalizability due to the small sample, this methodology holds significant potential for application in longer-term treatment and other network-based disorders.

Fitting models to experimental intracellular data is challenging. While detailed conductance-based models are difficult to train, phenomenological statistical models often fail to capture the rich intrinsic dynamics of circuits such as central pattern generators (CPGs). A recent trend has been to employ tools from deep learning to obtain data-driven models that can quantitatively learn intracellular dynamics from experimental data. This paper addresses the general questions of modeling, training, and interpreting a large class of such models in the context of estimating the dynamics of a neural circuit. In particular, we use recently introduced Recurrent Mechanistic Models to predict the dynamics of a Half-Center Oscillator (HCO), a type of CPG. We construct the HCO by interconnecting two neurons in the Stomatogastric Ganglion using the dynamic clamp experimental protocol. This allows us to gather ground truth synaptic currents, which the model is able to predict-even though these currents are not used during training. We empirically assess the speed and performance of the training methods of teacher forcing, multiple shooting, and generalized teacher forcing, which we present in a unified fashion tailored to data-driven models with explicit membrane voltage variables. From a theoretical perspective, we show that a key contraction condition in data-driven dynamics guarantees the applicability of these training methods. We also show that this condition enables the derivation of data-driven frequency-dependent conductances, making it possible to infer the excitability profile of a real neuronal circuit using a trained model.

The development of new studies that consider different structures of the hierarchical sensorimotor control system is essential to enable a more holistic understanding about movement. The incorporation of more biological proprioceptive and neuronal circuit models to muscles can turn neuromusculoskeletal systems more appropriate to investigate and elucidate motor control. Specifically, further studies that consider the closed-loop between proprioception and central nervous system may allow to better understand the yet open question about the importance of afferent feedback for sensorimotor learning and execution in the intact biological system. Therefore, this study aims to investigate the processing of spindle afferent firings by spiking neuronal network and their relevance for sensorimotor control. We integrated our previously published physiological model of the muscle spindle in a biological arm model, corresponding to a musculoskeletal system able to reproduce biological motion inside of the demoa multi-body simulation framework. We coupled this musculoskeletal system to physiologically-motivated neuronal spinal pathways, which were implemented based on literature in the NEST spiking neural network simulator, intended to perform human center-out reaching arising from spinal synaptic learning. As result, the spindle connections to the spinal neurons were strengthened for the more difficult targets (i.e. higher above placed targets) under perturbation, highlighting the importance of spindle proprioception to succeed in more difficult scenarios. Furthermore, an additionally-implemented simpler spinal network (that does not include the pathways with spindle proprioception) presented an inferior performance in the task by not being able to reach all the evaluated targets.

The brain's faculty to assimilate and retain information, continually updating its memory while limiting the loss of valuable past knowledge, remains largely a mystery. We address this challenge related to continuous learning in the context of associative memory networks, where the sequential storage of correlated patterns typically requires non-local learning rules or external memory systems. Our work demonstrates how incorporating biologically inspired inhibitory plasticity enables networks to autonomously explore their attractor landscape. The algorithm presented here allows for the autonomous retrieval of stored patterns, enabling the progressive incorporation of correlated memories. This mechanism is reminiscent of memory consolidation during sleep-like states in the mammalian central nervous system. The resulting framework provides insights into how neural circuits might maintain memories through purely local interactions and takes a step forward toward a more biologically plausible mechanism for memory rehearsal and continuous learning.

Artificial neural networks are limited in the number of patterns that they can store and accurately recall, with capacity constraints arising from factors such as network size, architectural structure, pattern sparsity, and pattern dissimilarity. Exceeding these limits leads to recall errors, eventually leading to catastrophic forgetting, which is a major challenge in continual learning. In this study, we characterize the theoretical maximum memory capacity of single-layer feedforward networks as a function of these parameters. We derive analytical expressions for maximum theoretical memory capacity and introduce a grid-based construction and sub-sampling method for pattern generation that takes advantage of the full storage potential of the network. Our findings indicate that maximum capacity scales as (N/S) ^S , where N is the number of input/output units and S the pattern sparsity, under threshold constraints related to minimum pattern differentiability. Simulation results validate these theoretical predictions and show that the optimal pattern set can be constructed deterministically for any given network size and pattern sparsity, systematically outperforming random pattern generation in terms of storage capacity. This work offers a foundational framework for maximizing storage efficiency in neural network systems and supports the development of data-efficient, sustainable AI.

As global life expectancy rises, a growing proportion of the population is affected by dementia, particularly Alzheimer's disease (AD) and Frontotemporal dementia (FTD). Electroencephalography (EEG) based diagnosis presents a non-invasive, cost effective alternative for early detection, yet existing methods are challenged by data scarcity, inter-subject variability, and privacy concerns. This study proposes lightweight and privacy-preserving EEG classification framework combining deep learning and Federated Learning (FL). Five convolutional neural networks (EEGNetv1, EEGNetv4, EEGITNet, EEGInception, EEGInceptionERP) have been evaluated on resting-state EEG dataset comprising 88 subjects. EEG signals are preprocessed using band-pass (1-45 Hz) and notch filtering, followed by exponential standardization and 4-second windowing. EEGNetv4 outperformed among other EEG tailored models, and upon utilizing the hybrid fusion techniques it achieves 97.1% accuracy using only 1,609 parameters and less than 1 MB of memory, demonstrating high efficiency. Moreover, FL using FedAvg is implemented across five stratified clients, achieving 96.9% accuracy on the hybrid fused EEGNetV4 model while preserving data privacy. This work establishes a scalable, resource-efficient, and privacy-compliant framework for EEG-based dementia diagnosis, suitable for deployment in real-world clinical and edge-device settings.

[This corrects the article DOI: 10.3389/fncom.2024.1404623.].