A Study on Spiking Neural Network Design

A Spiking Neural Network (SNN) is defined as an Artificial Neural Network (ANN) that has developed rapidly in the last decade. There are some difficulties in implementing these networks, especially in supervised machine learning. The main limitation is that classical learning approaches such as backpropagation cannot be directly applied to SNNs. We created a model to enable backpropagation, using a biological neuron model instead of the activation function of the ANN network. This study, a deep neural network that can work in both ANN and SNN modes was designed to classify the feature vectors obtained from the wavelet coefficient of magnetoencephalography signals taken from the human brain. Thus we propose a hybrid network that can operate in both conventional and firing modes by replacing the activation function of the traditional neural network with Izhikevich neurons. Our proposed network has been tested in classifying 4-class motor imaginary signals and the results are presented comparatively. We hope this work, which blends computational neuroscience and machine learning, will bring a different perspective to fired network design.