Spiking Neural Networks Using Backpropagation

Abstract

Brain-inspired Spiking Neural Networks (SNNs) occur with well-known neuromorphic hardware that delivers extra energy compared to conventional artificial neural networks (ANNs). Nevertheless, exploiting the same network layers as conventional ANNs to persevere a task appears unsuitable. Previous works employ similar architectures as Artificial Neural Networks and transform them into Spiking Neural Networks to attain the most exemplary performance as conventional ANNs. Nevertheless, this conversion technique needs greater timesteps for training spiking neural networks (SNNs). In this work, rather than using the ANN to SNN conversion method, we exploit the SNNs training directly using spike-based backpropagation. Since utilizing SNNs with the spike-based backpropagation requires fewer timesteps compared to ANN to SNN transformation approach. This work evaluates the classification performance on public and private (MNIST, Fashion MNIST, and KITTI) datasets.