Adaptively extendable multi-stage spiking neural network

Abstract

Recently, a significant improvement has been observed in the recognition rate in deep neural networks (DNNs). However, as the number of layers increases, additional computations and significant power consumption are required by the DNN. In this study, we propose a novel spiking neural network (SNN) that exhibits high recognition rate and reduced computational cost. If the reliability of the output of the current neural network (NN) is decided to be low, we feed forward the result to the input of the next NN. We use backpropagation learning algorithm to train the component NN. Since most of the decisions are made in the early stage, the proposed method shows approximately 83% reduction of the computational cost compared with the conventional SNN with the same recognition rate. (C) 2021 The Korean Institute of Communications and Information Sciences (KICS). Publishing services by Elsevier B.V.