The scientific community of researchers in computer science and automation in Lille, which has been growing rapidly in recent years, has come together to form the “Colloquium Polaris”. Eight to ten times a year, researchers come to present their work. On May 19, Rémi Bardenet, a CNRS researcher at the University of Lille, will speak at the Colloquium Polaris, which will take place at the Ircica amphitheater in Villeneuve d’Ascq. His talk will focus on improving random sampling by imposing diversity.

The Polaris Colloquium has two entities: the Inria Lille – Nord Europe research center and the university laboratory CRIStAL: Centre de Recherche en Informatique, Signal et Automatique de Lille.

Their researchers intend to enrich the scientific life of their community by proposing this colloquium of international dimension where quality work is shared in the field of digital sciences, computer science and automation, which go beyond the borders of their respective institutions.

Rémi Bardenet, researcher in artificial intelligence

After studying mathematics at the University of Strasbourg and obtaining a master’s degree in Machine Learning at ENS Cachan, Rémi Bardenet obtained a PhD in computer science from the University of Paris-Sud in 2012. His thesis focuses on the development of statistical methods for the analysis of data from one of the current large particle physics experiments, the Pierre Auger experiment in Argentina.

He states:

“In a few words, I received my PhD in November 2012 at the University of Paris-Sud, France, working with Balázs Kégl on Monte Carlo methods and Bayesian optimization, applied to particle physics and machine learning. I was notably a member of the Pierre Auger collaboration.”

During his postdoctoral stay at the University of Oxford, UK, Rémi Bardenet is studying the computational limits of Bayesian statistics when the data are too large to be stored in memory.

He adds:

“I then joined Chris Holmes’ group at the University of Oxford, UK, to work as a postdoc on Markov chain Monte Carlo for big data. Since then, I have also been working on applications to computational biology in the 2020 Science Network, of which I am now an emeritus member.”

He then joined the CRIStAL laboratory at the University of Lille in early 2015, in the SigMA (“Signals, models and applications”) team, where he is working on a research program on the statistical use of repulsive point processes with Adrien Hardy, from the Laboratoire de Mathématiques Paul Painlevé at the University of Lille.

With his colleagues of the SIGMA team, Rémi Bardenet has recognized in traditional signal processing tools repulsive processes from quantum optics, whose statistical properties allow to develop new denoising algorithms. He continues to explore the connections with signal processing on the one hand, and with quantum optics on the other hand.

This research work has earned him the ERC Starting Grant “Blackjack” from 2020 to 2025, as well as a national chair in artificial intelligence called “Baccarat” from 2020 to 2024. He also received the CNRS Bronze Medal last year.

Improving random sampling by imposing diversity, the theme of the presentation

Sampling consists in selecting a small number of elements to represent a large, possibly infinite, set of elements.

Thus, when statisticians are faced with a data set where each individual is described by an intractable number of characteristics, they may sometimes retain only the most informative characteristics in their data set. The features are then columns in a fat matrix and the sample is a small set of column indices.

For numerical integration, a function is summarized by a finite number of evaluations of that function, which are then combined to estimate its integral. In this case, the basis set is the function, an infinite collection of input-output pairs, and the sample is the small set of evaluations of the function.

Rémi Bardenet is interested in random sampling, that is, in sampling algorithms that are described by drawing a probability distribution on subsets of elements. For example, Monte Carlo integration is a numerical integration technique using random numbers. While many fundamental random sampling algorithms draw elements independently of each other, Rémi Bardenet focuses on sampling distributions where individual elements are sampled jointly, with the constraint that the resulting sample be as diverse as possible.

During the colloquium, he will present some sampling problems for which he was able to transform a natural notion of diversity into a sampling algorithm that is both computationally feasible and has state-of-the-art statistical performance guarantees.

Translated from Intervention de Rémi Bardenet, chercheur en IA, ce 19 mai lors du Colloquium Polaris