Towards Analog Implementation of Spiking Neural Networks for Audio Signals

This publication presents a novel approach to the training and deploying Spiking Neural Networks on analog hardware platforms. We proposed a scheme composed of steps involving architectural consideration, training, and mapping protocol. The method employs Genetic Algorithms and the Gradient-based method to train an architecture of single and multiple-layer SNN on AudioMnist data set to 85% of test accuracy. We also mapped the model to analog hardware, reaching an accuracy of 72% for selected runs of the mapping protocol. In the paper, we share insights and considerations which affect the complicated procedure of mapping SNN models to analog custom hardware.