Independence is Far From Normal

摘要

Exploratory Projection Pursuit (EPP) is a statistical data analysis tool for identifying structure in high dimensional data. In this paper we consider two Neural implementations of EPP. The first; performed under orthonormal constraints; utilises criteria based on fourth order moments, this is shown to be a dual for Independent Component Analysis (ICA). The second is based on the Kullback divergence from normality (negentropy), and is seen to perform ICA on data which is a linear mixture of independent latent variables. Simulations are reported which show the exceptional convergence speed of the negentropy based algorithm when limited a priori knowledge of the source distributions and a simple momentum based acceleration scheme are employed.