Bi-modal accuracy distribution in quantisation aware training of SNNs: an investigation

Neuromorphic technologies are being researched extensively due to their potential to achieve lower size, weight and power (SWaP) profiles in real-life applications by mimicking biological neurons. Understanding the caveats of deploying a Spiking Neural Network (SNN) in an embedded system is important, due to their potential to achieve lower size, weight and power (SWaP) profiles in real-life applications and its efficiency in applications using event-based data. This paper investigates the effects of the quantisation of SNNs from the perspective of deploying a model on FPGAs. Very few studies have been conducted to fully understand the significant drop in accuracy curves, which manifests itself as a bi-modal distribution. This paper attempts to identify whether the drop in accuracy is consistent across different models. And this precision-based characterisation of quantised SNNs will pave the way for the development of optimised-deployment procedure