Optimizing Neuromorphic Spike Encoding of Dynamic Stimulus Signals Using Information Theory

Abstract

Neuromorphic systems use spike representations of stimuli as inputs. These systems should ensure that the spikes carry a maximum amount of information on the signals that they encode. There is a pressing need to better understand how to maximize the information encoded into spikes, as it can have important implications for the outcome of the applications in which the spike representations are used. This work proposes the use of information theory, specifically the information rate, to maximize the information that a spike train carries on the signal that is encoded. The method consists of varying the encoding parameters to produce spike trains of different densities, and then estimating the information rate between the signal and the spike train over the entire range of spike densities. This allows to find an estimate of the spike density that maximizes the information rate, and therefore the optimal encoding parameters. The method is applied to the encoding of two stimuli (Brownian motion and speech) with a Leaky Integrate-and-Fire neuron. The proposed approach is fast and general, as it can be used with any dynamic stimulus input and any spike encoding technique. It offers a rigorous solution to the problem of spike encoding optimization and allows the separation of the encoding stage from task-specific applications that use spikes as inputs.