Information Geometry and Information Theory in Machine Learning

Abstract

Information geometry is a general framework of Riemannian manifolds with dual affine connections. Some manifolds (e.g. the manifold of an exponential family) have natural connections (e.g. e- and m-connections) with which the manifold is dually-flat. Conversely, a dually-flat structure can be introduced into a manifold from a potential function. This paper shows the case of quasi-additive algorithms as an example. Information theory is another important tool in machine learning. Many of its applications consider information-theoretic quantities such as the entropy and the mutual information, but few fully recognize the underlying essence of them. The asymptotic equipartition property is one of the essence in information theory. This paper gives an example of the property in a Markov decision process and shows how it is related to return maximization in reinforcement learning.