Generalized Mixture Models for Blind Source Separation

Neural Independent Component Analysis (ICA) algorithms based on unimodal source distributions provide acceptable performances in the case of Blind Source Separation (BSS) of super-gaussian sources. However, their convergence profiles are significantly slower in the case of sub-gaussian sources. In some situations it is necessary to deal with sub-gaussian signals in the form of noise or others. In this case, one needs an algorithm that can deal efficiently with mixtures of both superand sub-gaussian signals. In this paper, we introduce generalized mixture models for super-and sub-gaussian sources based on the Exponential Power Distibution (EPD). The kurtosis and stability profiles of these models are investigated and the corresponding non-linearities are derived. A switching algorithm is designed for the blind source separation of mixtres of superand sub-gaussian sources. Experimental results are presented on the application of these models to homogeneous and mixed sources using a modified fast ICA algorithm.