Estimating information quantities from biological data

Ilya Nemenman
UCSB/Institute for Theoretical Physics

In many biological applications, notably in computational neuroscience and bioinformatics, information theoretic methods are now routinely being used to solve diverse problems that range from the search for transcription factor binding sites to studying sensory adaptation in animals. The major hurdle in applying such methods is the reliable estimation of entropic quantities from small samples. I will discuss this difficult problem, approaches that the others have taken, and then present a Bayesian estimator of entropies introduced by us recently. I will analyze properties of the estimator and explain when it is expected to work well and how to diagnose errors. Throughout the talk I will focus on applications to biological examples (mainly to a fly visual system neuroscience experiment) to illustrate the potentials of the method.