FROM PRINCIPAL COMPONENT ANALYSIS TO UNSUPERVISED LEARNING WITH KERNELS
Dr Alexander J Smola, GMD First, Berlin

Abstract: Principal Component Analysis is one of the most popular techniques in unsupervised learning. It offers both means of data description (by fitting a Gaussian distribution to the observations) and of feature extraction (by providing linear forms via projections on the principal axes). In my talk I will show how this linear technique can be extended to nonlinear versions by using kernel functions to achieve algorithms such as regularized principal manifolds (similar to the generative topographic map and principal curves) or kernel feature analysis. The latter, in particular can be formulated as kernel principal component analysis, or, as carried out recently, in the more computationally efficient form of sparse kernel feature analysis, leading to settings similar to nonlinear projection pursuit. In particular it offers computation up to an order faster of both feature extractors and of the features itself than in standard kernel pca.

This seminar was held at the Department of Computer Science, Royal Holloway, University of London on 13 September 1999.

back