A Kernel Classifier for Distributions
Alexei Pozdnoukov and Samy Bengio
IDIAP, Martigny, CH.
This paper presents a new algorithm for classifying distributions.
It can be applied to a number of real-life tasks which include data
represented as distributions.
While classical SVMs discriminate between example points of two classes, we propose a novel SVM formulation to discriminate between example distributions of two classes, while still keeping advantages of SVMs such as margin maximization and kernel trick. This extension can be used for many different settings, including principled incorporation of invariances described by distributions, which was illustrated in this paper. Other possible uses of this model include the possibility to maximize the margin for problems that were traditionally solved by generative models and log likelihood ratios such as speech processing.