Computing with noise in spiking neural networks

Abstract

Trial-to-trial variability is an ubiquitous characteristic in neural firing patterns and is often regarded as a side-effect of intrinsic noise. Increasing evidence indicates that this variability is a signature of network computation. The computational role of noise is not yet clear and existing frameworks use abstract models for stochastic computation. In this work, we use networks of spiking neurons to perform stochastic inference by sam- pling. We provide a novel analytical description of the neural response function with an unprecedented range of validity. This description enables an implementation of spiking networks in simulations to sample from Boltzmann distributions. We show the robust- ness of these networks to parameter variations and highlight the substantial advantages of short-term plasticity in our framework. We demonstrate accelerated inference on neu- romorphic hardware with a speed-up of 10^4 compared to biological networks, regardless of network size. We further explore the role of noise as a computational component in our sampling networks and identify the functional equivalence between synaptic connec- tions and mutually shared noise. Based on this, we implement interconnected sampling ensembles which exploit their activity as noise resource to maintain a stochastic firing regime.