Application of the MEC network to principal component analysis and source separation
In this paper we present new developments of a previous work dealing with the problem of strongly-constrained orthonormal analysis of random signals. In the former work a neural learning rule arising from the study of the dynamics of a massive system in an abstract space was introduced, and the set of equations describing the motion of such a system was directly interpreted as a learning rule for neural layers. This adaptation rule can be used to solve several problems where orthonormal matrices are involved. Here we show two applications of such an approach: one dealing with PCA and one dealing with ICA.
Unable to display preview. Download preview PDF.
- 1.Comon P.: Independent Component Analysis, A New Concept ? Signal Processing 36 (1994) 287–314Google Scholar
- 2.Diamantaras K.I., Kung S.-Y.: Principal Component Neural Networks: Theory and Applications. J. Wiley, 1996Google Scholar
- 3.Fiori S., Uncini A., Piazza F.: A New Unsupervised Neural Algorithm for Orthonormal Signal Processing. Proc. of Int. Conf. Acoustic, Speech and Signal Processing — ICASSP (1997) 3349–3352Google Scholar
- 4.Laheld B., Cardoso J.F.: Adaptive Source Separation with Uniform Performances. Signal Processing VII: Theories and Applications 1 (1994) 183–186Google Scholar
- 5.Karhunen J., Joutsensalo J.: Learning of Robust Principal Component Subspace. Proc. of Int. Joint Conf. on N.N.-IJCNN 3 (1993) 2409–2412Google Scholar
- 6.Xu L., Oja E., Suen C.Y.: Modified Hebbian Learning for Curve and Surface Fitting. Neural Networks 5 (1992) 393–407Google Scholar