NONLINEAR INDEPENDENT COMPONENT ANALYSIS USING ENSEMBLE LEARNING: THEORY

Harri Valpola
E­mail: Harri.Valpola@hut.fi URL: http://www.cis.hut.fi/

A nonlinear version of independent component analysis is presented. The mapping from sources to observations is modelled by a multi­layer perceptron network and the distributions of sources are modelled by mixtures of Gaussians. The posterior probability of all the un­ known parameters is estimated by ensemble learning. In this paper, we present the theory of the method, and in a companion paper experimental results.