A deep dynamic binary neural network and its application to matrix converters

Abstract

This paper studies the deep dynamic binary neural network that is characterized by the signum activation function, ternary weighting parameters and integer threshold parameters. In order to store a desired binary periodic orbit, we present a simple learning method based on the correlation learning. The method is applied to a teacher signal that corresponds to control signal of the matrix converter in power electronics. Performing numerical experiments, we investigate storage of the teacher signal and its stability as the depth of the network varies. © 2014 Springer International Publishing Switzerland.