The impact of conversion from structural brain connectivity to reservoir weights on the learning performance of connectome-based reservoir computing

Abstract

Recent studies have used reservoir computing approaches to investigate the functional properties of a network of structural connections between all brain regions (i.e., the connectome). In these studies, the echo state network is used as a reservoir computing model, and the weights between nodes in the reservoir layer, typically set at random, are set based on structural brain connectivity. Various methods have been used to convert structural brain connectivity into reservoir weights. However, it has been unclear how different methods affect the learning performance of reservoir computing. This study evaluates and compares the learning performance of connectome-based reservoir computing on a memory capacity task using various methods for converting structural brain connectivity into reservoir weights. We found that using structural brain connectivity weights as reservoir weights, as they are, resulted in relatively inferior performance, and that the performance was improved by randomly multiplying the converted weights with a negative sign.