Spiking neural network encoding optimization via reinforcement learning and evolution strategy

Abstract

Spiking Neural Network, a neural network model used in Neuromorphic chips, is attracting attention as a next generation neural network model, but it is difficult to overcome low accuracy because of lack of a powerful learning algorithm like backpropagation. Recently, approach that converts Analog-valued Neural Networks trained using backpropagation to Spiking Neural Networks showed that Spiking Neural Network can obtain high accuracy comparable to Analog-valued Neural Network. However, this approach has a disadvantage of slowing the inference speed by encoding activation information into spike firing rates to obtain high accuracy. In order to tackle this problem, we converted the rate-encoded network into a network with efficient encoding scheme, Time-To-First encoding. The results of our study were as accurate as or better than those of the previous studies and the speed was faster than that of the rate-encoded networks. In addition, our model is a hardware-friendly model, which means this model is efficient to be implemented in Neuromorphic chips.