Reflections on a Supervised Approach to Independent Component Analysis

This work focusses on a recent supervised approach to Independent Component Analysis, a linear transformation method that yields latent variables assumed to be non-gaussian and mutually independent. According to this approach the latent structure is identified by estimating the joint product density of independent components, using a technique that transforms the unsupervised learning problem into a supervised function approximation one. Like projection pursuit methodology, this procedure attempts to get interesting projections of the observed units, that seem to capture the latent clustered structure of the data.