Information Theory Limits of Neuromorphic Energy Efficiency

Abstract

The fundamental advantage of neuromorphic systems is their low power consumption, which emerges from their event-based computation implemented via spikes. However, we do not have a theory that explores the fundamental limits of the energy consumption that a neuromorphic system can achieve. In this work we present an approach to find those limitations using a mixture of principles from information theory and combinatorial techniques. We obtain a systematic way of finding the number of neurons and spikes per time unit that allow a required representational capacity.