Impact of the AER-induced timing distortion on Spiking Neural Networks implementing DSP

Spiking Neural Networks are considered to be the latest generation of artificial neural networks. They rely on principles inspired from brain operation and use coding strategies based on relative timings and/or rates to transport information. Implementing large hardware networks requires the use of Address Event Representation (AER) which affects the inter-spike timings. This paper focuses on various mathematical AER models and their influence on spike rate coding in Signal Processing applications implemented with SNN. For the most basic M/M/1 queue model, there is no effect on the results given by the considered benchmark multiplier operator, whereas for more realistic M/D/1 and M/G/1 models, the relative error comparing to Poisson process is respectively below 3% and 3.8% for realistic operations.