Analog Spiking Neural Network Synthesis for the MNIST

Abstract

Different from classical artificial neural network which processes digital data, the spiking neural network (SNN) processes spike trains, and hence can bring new advantages to the information processing. Indeed, its event-driven property helps to capture the rich dynamics the neurons have within the brain, and the sparsity of collected spikes helps reducing computational power. Besides, the nonlinear behavior of neurons offers easier solutions for non-linearity problems than classical methods and is well suited to stochastic resonance-based systems. An analog SNN composed of 86 electronic neurons (eNeuron) and 1238 synapses interacting through two hidden layers is proposed. It is made from two different types of eNeurons based on different versions of Leaky integrate-and-fire (LIF) or Morris-Lecar (ML) models. The proposed neural network, coupling deep learning and ultra-low power, is trained using a common machine learning system (TensorFlow) for the MNIST. LIF eNeurons implementations present some limitations in terms of dynamic range, while, considering different activation functions, ML eNeurons achieve robust accuracy which is approximately of 0.82.