Lecture by Remy Bardenet, AI Researcher, during the May 19 Colicium Polaris

The scientific community of researchers in computer science and automation in Lille, which has been growing rapidly in recent years, has come together around the constitution of the “Colocium Polaris”. Eight to ten times a year researchers come there to present their work. On May 19, Remy Bardenet, a CNRS researcher at Lille University, will give a lecture at the Colloquium Polaris at the Erica Amphitheater in Villeneuve d’Ask. Her presentation will focus on improving random patterns by applying variety.

Coloccium polaris has two entities: Inria Lille – Nord Europe Research Center and Crystal University Laboratory: Lily Computer, Signal and Automation Research Center.

Their researchers seek to enrich the scientific life of their communities through these international-level conversations where quality work is shared in the fields of digital science, computer science and automation, which transcend the boundaries of their respective organizations.

Remy Bardenett, Artificial Intelligence Researcher

After training in mathematics at the University of Strasbourg and a master’s degree in machine learning at ENS Kachan, Remy Bardenet earned a doctorate in computer science from Paris-Sud University in 2012. His thesis focuses on the development of statistical methods for analysis. Data from one of today’s major particle physics tests, the Pierre Agar test in Argentina.

He declared:

“In short, I received my PhD in November 2012 from the University of Paris-Sud, France, working with Balaz Kegel on the Monte Carlo method and Bayesian optimization, applied to particle physics and machine learning. I was a significant collaborator of Pierre Agar.

During his postdoctoral fellowship at Oxford University in the United Kingdom, Remy Bardnet studied the mathematical limitations of Bayesian statistics when data was too much to store in memory.

He adds:

“I joined Chris Holmes’ group at Oxford University in the United Kingdom to work as a postdocker for the Marko Chain Monte Carlo method for big data. Since then, I have also been working on computational biology applications within the 2020 Science Network, of which I am now an Emeritus member. A

He then joined Lille University’s Crystal Laboratory in early 2015, on the Sigma team (“Signals, Models and Applications”) where he worked with Paul and Adrian Hardy on a research program on the statistical use of repulsive point processes. Lily University’s Painlev Mathematics Laboratory.

Together with her colleagues at the SIGMA team, Rémi Bardenet recognizes repulsive processes arising from quantum optics in traditional signal processing equipment, whose statistical features make it possible to develop new denomination algorithms. He explored connections with signal processing on the one hand and quantum optics on the other.

The study funded him from the ERC starting grant “Blackjack” from 2020 to 2025, as well as a national chair called “Bacart” in artificial intelligence from 2020 to 2024. He also received the CNRS bronze medal last year.

Variety of presentation, improving random sampling by imposing themes

Sampling involves selecting a small number of items to represent a large, perhaps infinite, set of items.

Thus, when statisticians encounter a dataset where each person is described by a complex number of features, they can sometimes retain only the most informative features in their dataset. The items are then columns in a fat matrix and the sample is a small set of column indicators.

For a numerical integration, a function is summarized by a finite number of evaluations of this function, which will then be combined to make an integral estimate of it. In this case, the base set is a function, an infinite collection of input-output pairs, and the sample is a small set of function evaluations.

Remy Bardenet is interested in random sampling, in other words sampling algorithms that describe the probability distribution on a subset of components. Monte Carlo Integration, for example, is a numerical integration technique using random numbers. Although many basic random sampling algorithms draw items independently of each other, Remy Bardenet focuses on sample distribution where individual items are sampled together, making the sample as varied as possible.

During the conversation, he will reveal some sampling problems for which he was able to convert a natural concept of diversity into a sampling algorithm that is enriched with computational and sophisticated statistical performance guarantees.

Leave a Comment