Abstract

Reservoir computing has recently been put forward as the generic name of a new research line in machine learning [1]. This field combines Echo State Networks, introduced by Jaeger [2], and Liquid State Machines, introduced by Maass [3]. These systems, that can be used to solve complex classification problems in a computationally efficient way, were proposed as an alternative to the recurrent neural networks that depend on very intensive and time-consuming training algorithms. Reservoir computing systems have two distinct parts: the reservoir, and an output layer. Traditionally, the reservoir consists of a vast number of randomly interconnected nodes, receiving input signals. The output layer is a linear weighted sum of the node states with the weights optimised by learning. In practice, networks of several hundred nonlinear nodes often have a good performance.

© 2011 IEEE