Comparing Connectivity-To-Reservoir Conversion Methods for Connectome-Based Reservoir Computing

The reservoir computing approach has been used in recent studies to investigate the functional aspects of an empirical network consisting of a complete set of structural connections in the brain (i.e., the connectome). These studies use the echo state network as a model for reservoir computing and determine the connections between nodes in the reservoir layer based on the wiring patterns in the connectome. Various methods have been used to convert structural brain connectivity weights to node-to-node weights in the connectome-based reservoir. Comparison of the conversion methods has just begun, and it is not well understood which factors in the conversion methods are the keys to better learning performance. Here, we compare a variety of conversion methods with respect to the learning performance of the connectome-based reservoir in a standard memory capacity task. As main results, we find that the directionality of the weights and their signs in the connectome-based reservoir are crucial for achieving higher performance and that randomly multiplying the reservoir weights by 1 or −1 improves the inferior performance of the method in which the structural connectivity weights purely determine the reservoir weights. This study provides useful insights for those who wish to improve the learning performance while preserving the information of the structural connectivity weights in the connectome-based reservoir.