Echo State Networks and Self-Prediction

Abstract

Prediction occurs in many biological nervous systems e.g. in the cortex [7]. We introduce a method of adapting the recurrent layer dynamics of an echo-state network (ESN) without attempting to train the weights directly. Initially a network is generated that fulfils the echo state – liquid state condition. A second network is then trained to predict the next internal state of the system. In simulation, the prediction of this module is then mixed with the actual activation of the internal state neurons, to produce dynamics that a partially driven by the network model, rather than the input data. The mixture is determined by a parameter α. The target function be produced by the network was sin ³(0.24t), given an input function sin(0.24t). White noise was added to the input signal at 15% of the amplitude of the signal. Preliminary results indicate that self prediction may improve performance of an ESN when performing signal mappings in the presence of additive noise.