The 22nd Finnish Summer School on Probability Theory

The 22nd Finnish Summer School on Probability Theory

22. todennäköisyyslaskennan kesäkoulu pidetään Lahden Mukkulassa 5.-9.6.2000 (tiedotus suomeksi)

22 sommarskolan i sannolikhetsteori hålls på Mokulla gård i Lahtis den 5-9 juni 2000 (information på svenska)

The 22nd Finnish Summer School on Probability Theory will be arranged at Lahti from June 5 to June 9, 2000 (more information in English).

Speakers: Søren Asmussen (Lund University), Hans Bühlmann (Eidgenössische Technische Hochschule, Zurich), Yuri Kabanov (Université de Besançon) and Gábor Lugosi (Universitat Pompeu Fabra, Barcelona)

Søren Asmussen (Lund University, Sweden) will speak about Markov additive processes, with applications to queueing theory
Hans Bühlmann (Eidgenössische Technische Hochschule, Zurich) will speak about Credibility Theory
Yuri Kabanov (Université de Franche-Comté, Besançon) will speak on Arbitrage Theory
Gábor Lugosi (Universitat Pompeu Fabra, Barcelona) will speak about Pattern recognition

More information and lecture abstracts will be distributed in April - May. There is no participation fee, but we need to limit the number of participants to about 30.

Course description

Gàbor Lugosi: Pattern recognition Abstract

In these lectures we review some topics of statistical learning theory, also known as pattern recognition, or classification. The basic problem is to classify an observation into one of finitely many categories when the available information - the training data -, is a set of classified observations.

A common approach is to set up a model class (i.e., a class of classification rules) and select one by minimizing the empirical error. This empirical error minimization principle is the starting point of a rich theory pioneered by Vapnik and Chervonenkis which we summarize here. Among the basic tools of analysis we find inequalities from empirical process theory and concentration-of-measure inequalities. We survey some of these results.

We also discuss model selection issues, that is, methods for setting up a model class. Some of the most successful methods, known as complexity regularization, or structural risk minimization, use penalized error estimates to find the right-sized model class.