Search Papers | Poster Sessions | All Posters
Poster B92 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Replicating spectro-temporal dynamics in neurobiologically realistic neural networks via a self-supervised approach
Hamed Nejat1, Jason Sherfey2, Andre Bastos1; 1Vanderbilt University, 2Boston University
Computational modeling tools provide a precise platform to investigate theories and hypotheses in neuroscience. However, current neuronal circuit models fail to achieve realistic neural dynamics without non-physiological assumptions. One class of models can be trained to generate those dynamics with high computational performance but are biophysically unrealistic (e.g artificial neural networks). Another class of models are designed to be biophysically realistic yet most of these models heavily rely on manual tuning. In this study, we have implemented a self-supervised learning algorithm called generalized Stochastic Delta Rule (gSDR). With this rule, we have trained biophysical neural circuits to achieve specific responses, such as resting membrane potential, firing rate and oscillatory dynamics. These models can also be trained to reproduce observed neurophysiological data (e.g task modulated oscillatory dynamics). We test this by training the model to reproduce a visually evoked oscillation shift from alpha-beta (~10-30Hz) to gamma (~40-90Hz) based on high-density electrophysiological recordings. These gamma-beta interactions emerged by self-modulation of synaptic weights via gSDR. We demonstrated that this approach can be used to understand both neuronal circuit mechanisms as well as the computations they perform.
Keywords: Dynamics Modeling Predictive coding Electrophysiology