A One-Bit-Matching Learning Algorithm for Independent Component Analysis

Independent component analysis (ICA) has many practical applications in the fields of signal and image processing and several ICA learning algorithms have been constructed via the selection of model probability density functions. However, there is still a lack of deep mathematical theory to validate these ICA algorithms, especially for the general case that superand sub-Gaussian sources coexist. In this paper, according to the one-bit-matching principle and by turning the de-mixing matrix into an orthogonal matrix via certain normalization, we propose a one-bit-matching ICA learning algorithm on the Stiefel manifold. It is shown by the simulated and audio experiments that our proposed learning algorithm works efficiently on the ICA problem with both superand sub-Gaussian sources and outperforms the extended Infomax and Fast-ICA algorithms.