Neural implementation of the JADE-algorithm

Abstract

The Joint Approaximative Diagonalization of Eigenmatrices (JADE)-algorithm [6] is an algebraic approach for Indenpendent Component Analysis (ICA), a recent data analysis technique. The basic assumption of ICA is a linear superposition model where unknown source signals are mixed together by a mixing matrix. The aims is to recover the sources respectively the mixing matrix based upon the mixtures with only minimum or no knowledge about the sources. We will present a neural extension of the JADE-algorithm, discuss the properties of this new extension and apply it to an arbitrary mixture of real-world images.