Multilayer reservoir computing to overcome the memory-nonlinearity trade-off

Abstract

[eng] Machine learning, and more precisely reservoir computing, is a state-of-the-art research field because of its successful application to time-dependent computationallyhard tasks. Along this MSc Thesis, we will extend the basic concepts and features of Echo State Networks, which are a specific type of reservoir computers. The main goal of this work will be finding out a reservoir configuration on which the so-called memory-nonlinearity trade-off is overcome. In order to achieve such an enhancement, we will explore changes in the main properties of the reservoir, including its parameters, its degree of linearity and its topology