Local output gamma feedback neural network

Abstract

A theory is introduced for a multi-layered local output gamma feedback neural network (LOGF-NN) within the locally recurrent globally feedforward neural networks paradigm. It is developed for the classification and prediction tasks for spatio-temporal systems, and allows the representation of different time scales through the incorporation of a gamma memory. The update equations for the feedforward and temporal weights and parameters are derived through the backpropagation through time (BTT) learning algorithm. As a demonstration, it is applied to the benchmark problem of single-step sunspot series prediction, and is compared to other neural network (weight elimination neural network: WNET) and statistical (linear and threshold autoregressive: TAR) methods. As a measure of prediction accuracy, the average relative variance (ARV) is used. The proposed LOGF-NN approach's performance is comparable to the TAR method and outperforms the linear AR and the WNET approaches.