A genetic algorithm for self-supervised models of oscillatory neurodynamics.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Predictive coding theories suggest alpha/beta rhythms (8-30 Hz) prepare neural pathways for predicted inputs, leading to inhibition and reduced gamma rhythms (40-90 Hz).
• The circuit mechanisms underlying this predictive routing and the alpha/beta-gamma push-pull interaction remain unclear.

Purpose of the Study:

• To explore the circuit mechanisms implementing predictive routing and the alpha/beta-gamma push-pull interaction.
• To develop a novel computational tool for training biophysical neural circuits without manual parameter tuning.

Main Methods:

• Developed a self-supervised learning algorithm, generalized Stochastic Delta Rule (gSDR), to train biophysical neural circuits.
• Applied gSDR to model neurophysiology, specifically the shift from baseline to stimulus-induced gamma oscillations in macaque visual cortex.
• Investigated how gSDR self-modulates synaptic weights and how push-pull dynamics emerge from local and top-down network modulation.

Main Results:

• gSDR successfully trained neural circuits to meet defined objectives, demonstrating its utility in computational modeling.
• The model reproduced stimulus-induced gamma oscillations, showing self-modulation of synaptic weights via gSDR.
• Demonstrated that gamma-beta push-pull interactions can arise from stochastic local circuitry modulation and top-down inputs.

Conclusions:

• gSDR is an effective algorithm for training biophysical neural circuits for complex neuronal objectives.
• The study reveals inhibitory neuron mechanisms responsible for gamma-beta push-pull dynamics in predictive processing.
• This work provides insights into the neural implementation of predictive coding and routing.