Influence of Response Function Approximation on Learning Performance of Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) are neural networks (NNs) that use spikes (pulses) as input and output signals. Since their units have the ability to handle time-series information, SNNs are expected to be applied to complex time-series signal processing. However, the complexity of the unit operation makes it difficult to scale up the network compared to non-spiking NNs. Therefore, we pay attention to the fact that simplified activation functions (such as ReLU) improve the performance of large scale NNs. We discuss the effect of simplifying the spike response function, which corresponds to the activation function in SNN, from the viewpoint of learning performance, and report the results.