Spiking neural networks optimization

Abstract

Spiking Neural Networks offer a new approach to tackle the efficiency issues of conventional artificial intelligence (AI), but there is a big performance gap between Spiking Neural Networks and Artificial Neural Networks. This work contributes to the optimization of Spiking Neural Networks in order to reduce this gap. The proposed methods are focused on architecture optimization and the development of a new spike-based loss functions, to prevent the dead neuron issue. The proposed techniques are implemented using PyTorch and SnnTorch frameworks, and benchmarked using two different data sets. The evaluation of the three approaches has been done on two different data sets, using six different loss functions, and using accuracy as a main metric. Hence, new horizons of applications are opened up, as well as many possible optimization techniques.