Consideration on liquid structure contributing to discrimination capability of Liquid State Machine

Abstract

Liquid State Machine (LSM) is one of reservoir computing models and due to universal computing capability its application and improvement attract researchers. Meanwhile development of measurement technology reveals the existence of specific structure in brain networks, such as scale-free, small-world, and modular properties, which contribute to higher-order brain functions such as cognition and memory. In this paper, we apply various network models to a recurrent neural network of an LSM and investigate the relationship between structural properties and accuracy of discrimination. Results suggest that modularity of a recurrent neural network enhances discrimination capability of LSM.