Modeles Graphiques et Applications – MGA

SCIENTIFIC CONTEXT AND OBJECTIVES Probabilistic graphical models provide a flexible and powerful framework for capturing statistical dependencies in complex, multivariate data. They enable the building of large global models for complex phenomena out of smaller and more tractable local models. Throughout its rapid development since the 1990s, research on graphical models has always been very close to its application domains, constantly providing new set of tools and theories to practitioners from various fields such as audio processing, text processing, image processing, and bioinformatics. Graphical models have emerged as a common language between all these fields, which allows one field to readily absorb algorithmic advances from another field and allows the clear identification of common methodological bottlenecks shared by many application domains. The objectives of this project are to advance the methodological state-of-the-art of probabilistic modeling research, while applying the newly developed techniques to several domains. The composition of the team takes into account these two objectives: a strong methodological component, together with experts in the domains which both motivate and benefit immediately from theoretical and algorithmic advances. DESCRIPTION OF PROJECT Given a probabilistic model, the two main methodological problems are inference given the model and learning the parameters and/or structure of the model. Recent research has focused with success on those two problems, but only on domains with very specific assumptions (e.g., known graph structures, graph structures amenable to tractable exact inference, simple parametric probability distributions). However, the main difficulty in applying graphical models to real-life problems is that those assumptions are usually not met and the goal of the project is to deal with those situations (in particular densely connected graphs and heterogeneous data). In all the application domains that are represented in the project, several core applied problems have already been formalized in terms of probabilistic graphical models and the difficulty mentioned earlier has been identified. The project can be naturally divided in several separate workpackages, one methodological work package and one work package for each application domain. - Methodological package: recent advances were dedicated to specific instances of graphical models, in particular, simulation methods, variational methods and graph cuts. Those advances have often looked only at subproblems, have both advantages and drawbacks, and are usually developed separately (and often in separate scientific communities). The methodological objective is to build on these research ideas to develop new frameworks and algorithms dealing with difficulties encountered during the formulation of problems. In particular, there will be a strong focus on discriminative training methods, such as conditional random fields. - Application packages: application and valorisation of methodological advances to all application domains: computer vision (image segmentation, content-based image retrieval, object category identification), bioinformatics (gene networks modeling and functional prediction by phylogenomics), text processing (unsupervised semantic annotation and supervised tagging of documents). EXPECTED RESULTS The originality of this project will be the constant exchanges between methodological research and applied research, making sure that algorithms are developed with immediate potential applied impacts. We thus expect advances in terms of: -Methodology: in particular with respect to the main identified bottlenecks in learning and inference in graphical models. -Applications: the target problems we have selected in each domain are largely open and we expect that the project will lead to contributions to solving those problems.