Convex Optimization
Approaches for Model Selection
Kristiaan Pelckmans and Johan Suykens
K.U. Leuven, ESAT-SCD
In this talk we review recent progress in solving model selection tasks by
using convex optimization techniques. We present different ways of characterizing
the training problem of a regularized learning machine with respect to the
choice of tuning parameters, and the subsequent optimization of these tuning
parameters with respect to a model selection criterion. To make this approach
practically feasible, relaxations of the resulting multi-level inference
problems are derived in terms of a convex optimization problem. The ideas
are illustrated for the use of least squares support vector machines, and
extensions are given for other learning machines.