AN ELECTRIC FIELD APPROACH TO INDEPENDENT COMPONENT ANALYSIS

Sepp Hochreiter and Michael C. Mozer
fhochreit,mozerg@cs.colorado.edu

We propose a novel algorithm for Independent Compo- nent Analysis (ICA) that is based on an electric eld metaphor. As with all ICA techniques, the algorithm searches for a demixing model that produces compo- nents whose joint distribution matches the factorial distribution (i.e., the product of the marginal distri- butions). The joint and factorial distributions are rep- resented as positively and negatively charged particles, respectively, and the dynamics of the search are based on the interactions among particles. The algorithm can deal with arbitrary distributions for the sources, non- linear mixing functions, noisy observations, and an un- equal number of source and mixture components. The limitation of the algorithm is that it does not scale with the number of sources. Nonetheless, we demonstrate that the algorithm can solve challenging ICA problems that are beyond the capabilities of other ICA methods.