Conversion of ConvNets to Spiking Neural Networks With Less Than One Spike per Neuron

Abstract

Spiking neural networks can leverage the high efficiency of temporal coding by converting architectures that were previously learnt with the backpropagation algorithm. In this work, we present the application of a time-coded neuron model for the conversion of classic artificial neural networks that re-duces the computational complexity in the synaptic connections. By adapting the ReLU activation function, the network achieved a sparsity of 0.142 spikes per neuron. The classifi-cation of handwritten digits from the MNIST dataset show that the neuron model is able to convert convolutional neural networks with several hidden layers.