Fast inference with controlled uncertainty: application to astrophysical observations. – SHERLOCK

SHERLOCK is targeted as a balanced research and training project. Its originality lies in 3 main motivations: a promising research project in AI with interdisciplinary applications in astrophysics and chemistry, an ambitious training program connected to high-level research-oriented master programs, and a strong coupling research & training in AI.

SHERLOCK-research – The main purpose of SHERLOCK is to imagine new fast Bayesian inference methods that permit to quantify the uncertainty of their predictions, even in situations with no available ground truth. The proposed approach that we have recently initiated borrows ideas from Markov chain Monte Carlo (MCMC) methods combined with optimization and machine learning algorithms. On one hand, the quantification of uncertainty implies the exploration of the neighborhood of the proposed solution: this is what MCMC methods are good at. On the other hand, MCMC methods suffer from their limited ability to deal with data in large dimensions (tall data) or when an expensive likelihood must be evaluated a large number of times (big data). Then optimization methods may provide a solution that remains pointwise however. This project aims at bridging the gap between optimization and MCMC methods, for instance thanks to hybrid sampling methods and variable splitting. We are convinced that a compromise is possible and we want to explore this direction. Another crucial issue is scalability to which one answer is dimension reduction or compression. To this aim, we will investigate the potential of determinantal point processes (DPP). Our methodological contributions will concern both inverse problems and machine learning. We will consider applications to astrophysical data sets from the Orion-B consortium and gravitational waves (existing interdisciplinary collaboration). On a more prospective ground, we may consider problems in catalysis chemistry in collaboration with the Realcat equipex.

SHERLOCK-teaching & training - The “Data Analysis & Decision making” program that will become “Data science & artificial intelligence” in the last year program of Centrale Lille is under my supervision. This master level program favors a good opening to research. It is noticeable that about 5 students/year (close to 20%) on average pursue their studies in a PhD program: such profiles are highly demanded (e.g. last recruitments were in New-York University and Mc Gill/Facebook Research Montreal). In 2019, the new Master of Data Science will open in Lille, with the remarkable unified support of the University of Lille, Centrale Lille and IMT Lille-Douai. I have been leading this project over the last 2 years and will be responsible for this Master at his launch in September 2019. It is aimed at joining the emerging Graduate School of the University Lille-North Europe. Remarkably, it involves both the University of Lille, and therefore Polytech’Lille, and the engineering schools Centrale Lille, IMT Lille-Douai. SHERLOCK would impact 72 students every year by supporting invited professors and mobility grants for inward foreign students as well as for intership in foreign labs.