Theories for unsupervised learning: PCA and its nonlinear extensions
Lei Xu · 2002
Several theories are proposed for unsupervised learning in one layer nonlinear network. It has been shown that all the learning rules developed under the theories merge at performing principal component analysis (PCA) type tasks when the network reduces into linear one. However, for nonlinear networks the performances of these rules become different, which indicates many possibilities for nonlinear extensions of PCA. These theories provide a number of potential guidelines for further explorations on nonlinear PCA type learning. Moreover, the relations between these proposed theories as well as to some existing theories have also been discussed.>