Efficiency metrics for auditory neuromorphic spike encoding techniques using information theory

Abstract

Spike encoding of sound consists in converting a sound waveform into spikes. It is of interest in many domains, including the development of audio-based spiking neural network applications, where it is the first and a crucial stage of processing. Many spike encoding techniques exist, but there is no systematic approach to quantitatively evaluate their performance. This work proposes the use of three efficiency metrics based on information theory to solve this problem. The first, coding efficiency, measures the fraction of information that the spikes encode on the amplitude of the input signal. The second, computational efficiency, measures the information encoded subject to abstract computational costs imposed on the algorithmic operations of the spike encoding technique. The third, energy efficiency, measures the actual energy expended in the implementation of a spike encoding task. These three efficiency metrics are used to evaluate the performance of four spike encoding techniques for sound on the encoding of a cochleagram representation of speech data. The spike encoding techniques are: Independent Spike Coding, Send-on-Delta coding, Ben’s Spiker Algorithm, and Leaky Integrate-and-Fire (LIF) coding. The results show that LIF coding has the overall best performance in terms of coding, computational, and energy efficiency.