Metastability in a Stochastic System of Spiking Neurons with Leakage

TL;DR

This paper studies a large network of spiking neurons with leakage, showing that as the network size grows, the system exhibits metastable behavior where it remains in a state of activity for a random exponential time before settling into inactivity.

Contribution

It introduces two models of leakage in spiking neuron networks and proves their metastable behavior as the network size increases.

Findings

System exhibits metastability as population size diverges.

The trapping time converges to an exponential distribution.

Models differ in how leakage affects membrane potentials.

Abstract

We consider a finite system of interacting point processes with memory of variable length modeling a finite but large network of spiking neurons with two different leakage mechanisms. Associated to each neuron there are two point processes, describing its successive spiking and leakage times. For each neuron, the rate of the spiking point process is an exponential function of its membrane potential, with the restriction that the rate takes the value 0 when the membrane potential is 0. At each spiking time, the membrane potential of the neuron resets to 0, and simultaneously, the membrane potentials of the other neurons increase by one unit. The leakage can be modeled in two different ways. In the first way, at each occurrence time of the leakage point process associated to a neuron, the membrane potential of that neuron is reset to 0, with no effect on the other neurons. In the second way, if the membrane potential of the neuron is strictly positive, at each occurrence time of the leakage point process associated to that neuron, its membrane potential decreases by one unit, with no effect on the other neurons. In both cases, the leakage point process of the neurons has constant rate. For both models, we prove that the system has a metastable behavior as the population size diverges. This means that the time at which the system gets trapped by the list of null membrane potentials suitably re-scaled converges to a mean one exponential random time.