Gaussianizing Transformations for ICA

Nonlinear principal components analysis is shown to generate some of the most common criteria for solving the linear independent components analysis problem. These include minimum kurtosis, maximum likelihood and the contrast score functions. In this paper, a topology that can separate the independent sources from a linear mixture by specifically utilizing a Gaussianizing nonlinearity is demonstrated. The link between the proposed topology and nonlinear principal components is established. Possible extensions to nonlinear mixtures and several implementation issues are also discussed.