Performance Comparison of Deep Spiking CNN with Artificial Deep CNN for Image Classification Tasks

Recent advancement in neuromorphic devices has created new opportunities to develop and deploy the algorithms which mimic human-like information processing. Spiking neural networks (SNNs) are considered as third generation neural networks that produce sparse and asynchronous representation of spikes similar to the processing of the primate brain. SNNs offer advantages of low power requirement, event driven processing and fast inference compared to the state-of-the art second generation artificial neural networks. In order to understand the performance comparison of spiking Convolutional Neural Networks (Spiking-CNN), we adopted the population time encoding technique with at most one spike at a time on well-known CNN architecture. The network is trained using spike time dependent plasticity (STDP) and reward-modulated STDP learning rules with the use of SpykeTorch library in python. The simulation results considering image classification task on MNIST dataset indicate the comparable performance of the spiking CNN with traditional CNN architecture. The performance of the adopted spiking-CNN shows almost 97% accuracy on the selected dataset. The study suggests the need for further development of methodologies which can outperform the traditional CNN using spiking-CNN.