A new information-theoretic based ICA algorithm for blind signal separation

The typical information-theoretic approaches such as INFOMAX and MMI perform independent component analysis (ICA) by using a fixed nonlinearity function. Consequently, they can only separate either sub-Gaussian or super-Gaussian source signals, but not both. This article considers a flexible nonlinearity function that is a single polynomial term with the exponent learnable. The separation ability of this function is analysed, and a new ICA algorithm is proposed. The experiments have shown that this algorithm can successfully separate the mixture of sub-Gaussian and super-Gaussian sources.