Information transmission with renewal neurons
Carl van Vreeswijk
Gatsby Computational Neuroscience Unit
University College London
Submitted to Computational Neuroscience 2000
Information theory has become popular for illucidating cortical
function, both in experimental and theoretical studies. however, these studies make
unrealistic assumptions about the noise. Generally experimental studies assume Poisson
statistics, the theoretical models use rate-based neurons with rate-independent noise.
These simplifications qualitatively affect the results. Generalizing, from Poisson, to
arbitrary rate-dependent renewal neurons I shown that: 1) For constant rate, one loses
information if one only the total spike count is used, unlike in Poisson processes. 2) An
optimal encoder has a sparse representation of the input, unlike the whitening obtained in
models with rate-based neurons. Thus, when using information theory, a sufficiently good
description of the noise is vital.