SPARSIFICATION OF DYNAMIC BINARY NEURAL NETWORKS
森安, 淳吾MORIYASU, Jungo
562015-03-24 , 法政大学大学院理工学・工学研究科
This paper studies basic dynamics and learning capability of the dynamic binary neural network. The network has the signum activation function and can exhibit various binary periodic orbits. The network dynamics can be visualized by the Gray-code-based return map. We present a learning algorithm based on the correlation learning and the genetic algorithm. The purpose of the learning is not only storage of teacher signal but also enlargement of the domain of attraction to the teacher signal. As a typical example of the teacher signal, we use an artificial periodic orbit and a periodic orbit which corresponds to the control signal of the matrix converters. Performing basic numerical experiment, we have confirmed that the teacher signal can be stored successfully and the sparsification can be effective to reinforce the stability of the periodic orbit.