10 Mar 2004; Mat-Nat Faculty (UITO), room U7. Wednesday 12:15-14:00

Eirik Fossgaard "Code length principles as a tool in model selection".

We will give an introduction to binary prefix coding and the universal integer code of Rissanen. We will define the Minimum Description Length Principle as developed by Rissanen, explore its connection to model selection problems and provide a known result on optimality of the Normalized Maximum Likelihood (NML) Principle.

We will then (in a later talk) turn to discuss some of the main aspects of the work in my newly submitted Ph.D thesis: "An invariant bayesian model selection principle for gaussian data in a sparse representation"  and investigate invariant code length principles and provide an asymptotic result on code lengths for gaussian models. The necessity of, and problem of how to, incorporate prior information into our models will be discussed, and we will provide a result connecting discretization of model parameters to the problem of selecting a suitable prior distribution on model classes.