A Kernel Classifier for Distributions

This paper presents a new algorithm for classifying distributions. The algorithm combines the principle of margin maximization and a kernel trick, applied to distributions. Thus, it combines the discriminative power of support vector machines and the well-developed framework of generative models. It can be applied to a number of real-life tasks which include data represented as distributions. The algorithm can also be applied for introducing some prior knowledge on invariances into a discriminative model. We illustrate this approach in details for the case of Gaussian distributions, using a toy problem. We also present experiments devoted to the real-life problem of invariant image classification.

Related material


EPFL authors