Correlation coding in stochastic neural networks

Abstract

Stimulus-dependent changes have been observed in the correlations between the spike trains of simultaneously-recorded pairs of neurons from the auditory cortex of marmosets even when there was no change in the average firing rates. A simple neural model can reproduce most of the characteristics of these experimental observations based on model neurons having leaky integration and fire-and-reset spikes and with Poisson-distributed, balanced input. The source of the synchrony in the model was common sensory input. The outputs of neurons in the model appear noisy (almost Poisson) owing to the stochastic nature of the input signal, but there is nevertheless a strong central peak in the correlation of the output spike trains. The experimental data and this simple model clearly demonstrate how even a noisy-looking spike train can convey basic information about a sensory stimulus in the relative spike timing between neurons.