A modular realization of adaptive PCA

We propose an adaptive PCA algorithm which alleviates suboptimality of the PCA method for nonstationary signals. A modular neural realization of adaptive PCA is considered and its design is formulated as an optimization problem, following the design of the vector quantizer. This formulation results in a competitive algorithm that learns data's local eigenstructures in an unsupervised way. The algorithm includes the recently proposed adaptive transform coding algorithm of R.D. Dony and S. Haykin (1995) as a special case and, as confirmed by simulation studies, the algorithm is better than their algorithm in mean square error (MSE).