Kernelized Renyi distance for subset selection and similarity scoring

Abstract

Renyi entropy refers to a generalized class of entropies that have been used in several applications. In this work, we derive a non-parametric distance between distributions based on the quadratic Renyi entropy. The distributions are estimated via Parzen density estimates. The quadratic complexity of the distance evaluation is mitigated with GPU-based parallelization. This results in an efficiently evaluated non-parametric entropic distance - the kernelized Renyi distance or the KRD. We adapt the KRD into a similarity measure and show its application to speaker recognition. We further extend KRD to measure dissimilarities between distributions and illustrate its applications to statistical subset selection and dictionary learning for object recognition and pose estimation.