Tikhonov Regularization and the Representer Theorem

Description

We discuss a fundamental class of learning algorithms the can be interpreted as a form of Tikhonov regularization. Different algorithms correspond to different choices of loss function. Tikhonov regularization provides an estimator which solves a suitable minimization problem. In this class we discuss the properties of this minimization problem, showing in particular that a minimizer can be found by solving a finite dimensional problem

Slides

Slides for this lecture (none)

Suggested Reading

• De Vito E., Rosasco L., Caponnetto A., Piana M. and Verri A. Some Properties of Regularized Kernel Methods. JMLR 2004
• Scholkopf B., Herbrich R., Smola AJ. A Generalized Representer Theorem COLT, 2001.