Statistical learning theory

The methods of statistical physics and numerical simulations are used to determine the typical properties of learning algorithms. The most interesting results recently obtained by our group are (1) the prediction of a first order phase transition as a function of the number of training examples in the problem of determining the anisotropy axis of a ditribution, (2) the properties of SVMs and more precisely the role of the data normalization, (3) the introduction of a new distance in ILP (Inductive Logic Programming) allowing concept learning in the vicinity of the phase transition.

Computer science applications

 Modeling of biological activity of molecules issued from large chemolibraries, based on results of high throughput data. Our methods of analysis involve machine learning tools.
 Extraction of information from texts using automatic language treatments, in the domain of functional genomics and modeling of gene interactions.
 Conception of the AIRA system, a software platform that creates personnal agents for information seeking in the web and the comparative evaluation of existing tools.

Cognitive modeling

 Students modeling based on empirical data of interactions with a teaching software. Application to the Algebra domain, with data from interactions with the software Aplusix.
 Modeling and simulation of the construction of semantic associations from exposure to texts. Coupling with vision models.
 Models of complex systems with learning capabilities : (1) interaction between learning and evolution (Baldwin effect) and (2) models of social systems of interacting agents.