Spiking Neural Networks Vs Convolutional Neural Networks for Supervised Learning

Abstract

Deep learning has revolutionised the field of machine learning in near years, particularly for computer vision. Two methods are used to view supervised learning, SNN networks with the handwritten Digit Recognition Problem (NOD) and Normalized Normalized Approximate Descent (NORMAD). Experiments show that the identification accuracy of the prototype SNN does not deteriorate by more than 1% relative to the floating-point baseline, even with synaptic weights of 3-bit. In addition, the proposed SNN, which is trained on the basis of accurate spike timing data, outperforms the equivalent non-spiking artificial neural network (ANN) trained with back propagation, especially at low bit precision, and is in line with the convolutionary neural network that is normally used to train these system. Recent work shows the potential to use Spike-Based Data Encoding and learning for applications of the real world for positive neuromorphism.