Subspace blind extraction by less-complete ICA

A new ICA paradigm based on the constrained ICA (cICA) framework is proposed to extract a subspace of underlying sources mixed in an observed signal. Its algorithm achieves both separating independent components and reducing output dimension simultaneously by deriving a single-stage learning rule. Particularly, this technique is able to extract the interesting ICs according to their density types in super- or sub-Gaussian. Experiments with the synthetic audio and image data demonstrate the proposed algorithms superior than other existing methods.