Bayesian Optimisation for a Biologically Inspired Population Neural Network

TL;DR

This paper employs Bayesian Optimization to efficiently identify hyper-parameters for a biologically inspired neural network, enabling the simulation of specific brain rhythms like alpha, theta, and beta bands with improved accuracy and automation.

Contribution

It introduces an automated hyper-parameter tuning method using Bayesian Optimization for a biologically plausible neural network, improving over trial-and-error approaches.

Findings

Successfully optimized parameters for alpha, theta, and beta rhythms.

Automated hyper-parameter search reduces manual trial-and-error.

Constraints based on network output improve rhythm accuracy.

Abstract

We have used Bayesian Optimisation (BO) to find hyper-parameters in an existing biologically plausible population neural network. The 8-dimensional optimal hyper-parameter combination should be such that the network dynamics simulate the resting state alpha rhythm (8 - 13 Hz rhythms in brain signals). Each combination of these eight hyper-parameters constitutes a 'datapoint' in the parameter space. The best combination of these parameters leads to the neural network's output power spectral peak being constraint within the alpha band. Further, constraints were introduced to the BO algorithm based on qualitative observation of the network output time series, so that high amplitude pseudo-periodic oscillations are removed. Upon successful implementation for alpha band, we further optimised the network to oscillate within the theta (4 - 8 Hz) and beta (13 - 30 Hz) bands. The changing rhythms in the model can now be studied using the identified optimal hyper-parameters for the respective frequency bands. We have previously tuned parameters in the existing neural network by the trial-and-error approach; however, due to time and computational constraints, we could not vary more than three parameters at once. The approach detailed here, allows an automatic hyper-parameter search, producing reliable parameter sets for the network.