Paper 161 Details

Title

An unexpected discrepancy in a well-known problem: Kraskov estimators applied to spiking neural networks

Abstract

While revisiting a problem in spiking neural networks (Shanahan, 2008) we discovered an interesting anomaly that we feel deserves further investigation as it could signify a useful contribution to the field of complex systems analysis. Robust methods for estimating information-theoretic measures like the two mutual information estimators in Kraskov et al. (2004) normally yield similar results, but we have found a case where they do not. Through extensive testing and the use of software libraries that are known to be reliable, we are confident that this seemingly contradictory outcome is not an error, and that it is a new discovery. We advocate the need for more thorough testing and benchmarking of estimators, and warn against the risks of drawing premature conclusions when using only one test method.