Partenaires

CNRS
Logo tutelle
Logo tutelle
Logo tutelle

Rechercher

Sur ce site

Sur le Web du CNRS

Accueil du site > Séminaires > Mathématiques des systèmes complexes > Learning networks and curves

Vendredi 30 novembre 2007 à 11h00

Learning networks and curves

Kevin Bleakley (Université Montpellier 2)

Résumé : A recent family of learning methods, based on what are called "kernels" and often used in tandem with the SVM algorithm, are causing a (not-so ?) quiet revolution in all things "prediction", such as voice, text and image recognition. Furthermore, they have found an important home in post-genomic biology, where masses of numerical information, if used well, can provide real biological insight. In this talk, I’ll give a (hopefully) not-too-technical introduction to these methods, highlighting their theoretical origins in functional analysis and statistics. Then I’ll introduce a way to use/choose kernels and the SVM algorithm to predict protein-protein interaction networks and metabolic networks in the cell, and show very encouraging results on two benchmark biological data sets. Next, I will change the subject and look at time-series prediction from a non-parametric point of view. The basic idea is to create a set of "experts", each of which uses data from the past in a different way, to predict the future. By aggregating the experts’ predictions (with coefficients based on an exponential weighting of each expert’s previous performance), it can be can show that several such strategies converge to the best possible strategy, with respect to certain classes of stationary and ergodic processes. I’ll conclude with a couple of comparisons with parametric methods on real-world data sets.

Dans la même rubrique :