Simple deterministically constructed recurrent neural networks

Abstract

A large number of models for time series processing, forecasting or modeling follows a state-space formulation. Models in the specific class of state-space approaches, referred to as Reservoir Computing, fix their state-transition function. The state space with the associated state transition structure forms a reservoir, which is supposed to be sufficiently complex so as to capture a large number of features of the input stream that can be potentially exploited by the reservoir-to-output readout mapping. The largely "black box" character of reservoirs prevents us from performing a deeper theoretical investigation of the dynamical properties of successful reservoirs. Reservoir construction is largely driven by a series of (more-or-less) ad-hoc randomized model building stages, with both the researchers and practitioners having to rely on a series of trials and errors. We show that a very simple deterministically constructed reservoir with simple cycle topology gives performances comparable to those of the Echo State Network (ESN) on a number of time series benchmarks. Moreover, we argue that the memory capacity of such a model can be made arbitrarily close to the proved theoretical limit. © 2010 Springer-Verlag Berlin Heidelberg.