A Comparative Study of Supervised Learning Algorithms for Spiking Neural Networks

Spiking neural networks (SNNs), often referred to as third-generation neural networks, more closely mimic the activity of the human brain compared to traditional artificial neural networks (ANNs), offering potential advantages in energy efficiency and performance. However, classical learning algorithms cannot be applied directly for SNN training since the spikes are not differentiable. In recent years, several supervised learning algorithms have been developed for SNNs, which can be broadly categorized into three groups: ANN-to-SNN (ANN2SNN) conversion algorithms, backpropagation through time (BPTT) algorithms, and spike-timing-dependent plasticity (STDP) algorithms. In this paper, we make a comparative study on these three approaches and evaluate their performance on the CIFAR-10 dataset. We discuss future research perspectives focusing on improving spike encoding strategies and developing more robust training protocols for SNNs.