A bigradient optimization approach for robust PCA, MCA, and source separation
L. Wang, Juha Karhunen, Erkki Oja · 2002
The authors earlier derived neural principal or minor component learning algorithms and their robust extensions by optimizing a generalized variance criterion under orthonormality constraints. In this paper, the authors propose an alternative approach, where the stochastic learning algorithm is derived by optimizing two criteria simultaneously. This yields a new bigradient algorithm, which can be used in slightly different forms for PCA, MCA, and their robust extensions in either symmetric (subspace) or hierarchic modes. The algorithm is successfully applied to separation of independent sources from their linear mixture.