On the Accuracy and Performance of Spiking Neural Network Simulations

Abstract

Spiking Neural Networks (SNNs) are a class of Artificial Neural Networks that show a time behaviour that cannot be computed with single one-shot functions. Therefore, to study their evolution over time, simulations are typically employed. Typical simulation approaches rely on time-stepped simulations, while more recent works have highlighted the opportunity to rely on Parallel Discrete Event Simulation (PDES) for improved accuracy. In particular, Speculative PDES has been shown to be a suitable simulation paradigm to deal with the peculiar temporal domain of SNNs. In this paper, we perform an experimental evaluation of these two different approaches, showing the implications on both simulation performance and accuracy. Our assessment showcases that Parallel Discrete Event Simulation can deliver good scaling on parallel architectures while offering more accurate results.