Techniques for accurate and scalable simulation of spiking neural networks using speculative discrete event simulation

In the era of ubiquitous Artificial Intelligence and power-hungry Neural Networks, the brain offers prime inspiration for a faster, greener, more efficient, and arguably more effective alternative: Spiking Neural Networks (SNNs). This thesis explores simulating SNNs using Parallel Discrete Event Simulation (PDES) with Time Warp. We present the motivations for this approach, the challenges that using it poses, and illustrate our solutions in-depth from both a theoretical and technical point of view, with emphasis on the latter. With the ability to execute SNN simulation on a PDES support, we show how this simulation method allows for achieving significantly higher simulation accuracy with respect to the traditional Time-Stepped approach. In our experimentation, the traditional approach was shown to suffer from substantial drift from the expected network activity due to the compounding effect of inaccuracies. Higher accuracy is crucial to properly simulate and thus study biological neural networks in silico, as well as simulate analogical neuromorphic chips, but we also show it plays a fundamental role when using SNNs for AI by replicating recognition experiments and achieving higher classification accuracy, all while using simpler network topologies, with lower energy consumption. Finally, our experimentation also highlights the high scalability of our approach thanks to effective utilisation of both parallel and distributed computing.