An Introduction to Kernel Methods.

Kernel methods give a systematic and principled approach to training learning machines and the good generalisation performance achieved can be readily justified using statistical learning theory or Bayesian arguments. We describe how to use kernel methods for classification, regression and novelty detection and in each case we find that training can be reduced to optimisation of a convex cost function. We describe algorithmic approaches for training these systems including model selection strategies and techniques for handling unlabelled data. Finally we present some recent applications. The emphasis will be on using RBF kernels which generate RBF networks but the approach is general since other types of learning machines (e.g. feed­forward neural networks or polynomial classifiers) can be readily generated with different choices of kernel.