Spike train patterning and forecastability

Theories of neural coding rely on a knowledge of correlations between firing events. These correlations are also useful to validate biophysical models for the neural activity. We present a methodology for validating models based on the assessment of linear and non-linear correlations between variables derived from the spike train. The firing pattern of an electroreceptor is analyzed in this framework. We show that a purely stochastic model fails to capture the essential correlations between interspike intervals, even though it reproduces the interval histogram and certain spike train spectral features. However, a biophysical model, based on the Fitzhugh-Nagumo equations with noise, does exhibit many of the correlations seen in the data, including those between successive firing phases.