YGER Florian
florian.yger [at] dauphine.fr
Short bio
Associate professor at Université Paris-Dauphine since 2015. JSPS fellow in the laboratory of Pr. Sugiyama (from 2014 to 2015) and visiting researcher, RIKEN AIP (summer 2017-).
Topics of interest
Trustworthy machine learning, Causal inference, interpretable AI
Project in Prairie
Florian Yger will address the questions of trust, explainability and interpretability in machine learning models (including deep learning) with a focus on the robustness to adversarial examples and counterfactual reasoning on data. This project has natural and practical applications in the medical field.
Quote
In the last decade, deep learning has made possible breakthrouhgts in several domains (e.g. computer vision, machine translation, games, …). Yet those hardly interpretable algorithms are fed with huge amounts of -sometimes sensitive- data and can suffer from malicious attacks: attacks on the privacy of the data and attacks on the robustness where adversarial examples are generated to fool the algorithm. This is a critical issue (especially in medical applications) and we feel that an effort toward a deeper theoretical analysis is needed.
RUDI Alessandro
alessandro.rudi [at] inria.fr
Short bio
Alessandro Rudi is Researcher at INRIA, Paris from 2017. He received his PhD in 2014 from the University of Genova, after being a visiting student at the Center for Biological and Computational Learning at Massachusetts Institute of Technology. Between 2014 and 2017 he has been a postdoctoral fellow at Laboratory of Computational and Statistical Learning at Italian Institute of Technology and University of Genova.
Topics of interest
Large scale machine learning, structured prediction
Project in Prairie
My main research interest is in machine learning. In particular my focus is on theoretical and algorithmic aspect of statistical machine learning, with the goal of devising algorithms that at the same time can (a) scale on big data (b) be easily applied in practice (c) have strong theoretical guarantees in terms of statistical and computational aspects (d) achieve state of the art error on the prediction, with reduced computational costs.
Quote
Large Scale ML with statistical guarantees. ML algorithms can be divided in (a) non-parametric, with strong theoretical guarantees, but high computational requirements especially in terms of memory footprint, as kernel methods (b) parametric, as deep nets, with small computational complexity and effective results in practice, but without theoretical guarantees. The goal of my work is to develop hybrid methods that take the best of both worlds: fast, effective and with theoretical guarantees. Structured Prediction. Nowadays data are very often more complex than vectors. In many fields learning problems consist in predicting structured/complex objects from other structured objects. Using the power of infinite-dimensional implicit embeddings, my goal in this direction consists in developing a unified theoretical and algorithmic framework able to deal effectively with a wide family of structured inputs and outputs.
MALLAT Stéphane
stephane.mallat [at] ens.fr
Short bio
Professor at NYU from 1988 to 1994. Professor at Ecole Polytechnique, from 1994 to 2012. Co-founder and CEO of a semiconductor start-up from 2001 to 2007. Professor in Computer Science at Ecole Normale Supérieure from 2012-to 2017. Professor at the Collège de France in Data Sciences since 2017. Member of the French Academy of sciences, of the French Academy of Technologies and foreign member of the US National Academy of Engineering. IEEE and EUSIPCO Fellow. Recipient of the SPIE 2007 Outstanding Achievement Award, of the 2004 European IST Grand prize, of the 2004 INIST-CNRS prize for most cited French Researcher in engineering, of the 2015 IEEE Signal Processing best sustaining paper award, of the 2017 IEEE Freidrich Gauss Prize.
Topics of interest
Harmonic analysis, machine learning, signal processing
Project in Prairie
Stéphane Mallat will be working on the mathematical understanding and interpretability of deep neural networks with applications to images, audio, financial data, quantum chemistry and cosmology. He teaches Data Sciences and will promote the interface between industry, academia and students through the organization of data challenges in http://challengedata.ens.fr
Quote
Deep neural networks have considerable applications to services, industry, and science but remain a black box whose properties are not well understood. Robustness and interpretability of deep neural networks become a major issue for their applications. Understanding deep networks involves many branches of mathematics, including statistics, harmonic analysis, geometry, and optimization in high dimension, together with algorithmic experiments on real data. Working on very different types of data and applications, gives an access to generic mathematical and algorithmic properties of these networks. Simplifying network architectures, while preserving performance is an important direction of this research.
GAILLARD Pierre
pierre.gaillard [at] inria.fr
Short bio
Researcher at INRIA Paris within the SIERRA project team. Recipient of the PhD dissertation awards Paul Caseau (Awarded by the French Academy of Technologies and EDF) and AMIES on industrial mathematical PhD.
Topics of interest
Online learning, adversarial learning, machine learning
Project in Prairie
Pierre Gaillard will consider fundamental machine learning problems with a particular focus on online learning. He will pursue to mix theoretical research with applications on real data through interdisciplinary collaborations. All along his research, he plans to design algorithms with both robust theoretical guarantees and good practical performance. He will be involved in the PSL AI graduate school.
Quote
Nowadays, the volume and speed of data flows are constantly increasing. Many applications need to move from offline methods to sequential methods that can acquire data, adapt to it and process it on the fly. At the same time, the data are becoming more and more sophisticated. Traditional statistical assumptions such as stationarity are no longer satisfied. Designing efficient algorithms that can learn from data as one goes along with as few assumptions as possible is a major challenge of today’s machine learning.
Short bio
After dual PhDs from Ecole Polytechnique and Stanford University in optimisation and finance, followed by a postdoc at U.C. Berkeley, Alexandre d’Aspremont joined the faculty at Princeton University as an assistant then associate professor. He returned to Europe in 2011 and is now a research director at CNRS, attached to Ecole Normale Supérieure in Paris. He received the SIAM Optimization prize, a NSF CAREER award, and an ERC starting grant. He co-founded and is scientific director of the MASH Msc degree at PSL. He also co-founded Kayrros SAS, which focuses on energy markets and earth observation.
Topics of interest
Optimisation, machine learning
Project in Prairie
Alexandre d’Aspremont’s work is focused on optimization and applications in machine learning, statistics, bioinformatics, signal processing and finance. He collaborates with several companies on projects linked to earth observation, insurance pricing, statistical arbitrage, etc. He is also co scientific director of MASH, a Msc program focused on machine learning and its applications in digital marketing, journalism, public policy, etc.
Quote
Optimization plays a central role in modern statistics and machine learning
in particular. Beyond direct applications of optimization algorithms, convex duality results and complexity theory underpin many recent developments in statistical learning (e.g. compressed sensing or matrix completion problems). Yet the link and clear empirical tradeoff between statistical performance and computational complexity has yet to be fully explained, which are crucial in most domains, in particular safety/health critical ones.
Bacry Emmanuel
bacry [at] ceremade.dauphine.fr
Short bio
Senior Researcher CNRS (2016-), Chief Scientific Officer of the Health Data Hub (2019-), Head of Health/Data projects, Ecole Polytechnique (2019- ), Head of Data Science & Big Data Initiative and Associate professor at Ecole Polytechnique (2014-2019).
Topics of interest
Machine learning, Point processes, Big data, Health
Project in Prairie
Emmanuel Bacry will work on AI algorithms applied to longitudinal data with a focus on healthcare data of the SNDS database. He will work on various applications such as detecting weak signals in pharmacoepidemiology or optimization of health pathways of a given pathology, developed in close collaboration with CNAM and the Health Data Hub.
Quote
The SNDS database is a unique medico-administrative health database. It is one of the largest in the world (>65 million people, >200Tb). Though it contains no clinical data, it is very rich, and can contribute to major applications with strong potential impacts on health and/or on economy of health. Moreover, thanks to the Health Data Hub, this database will be soon enriched by large amount of clinical databases. Working on these very large databases is complex but extremely exciting.
Bach Francis
francis.bach [at] inria.fr
Short bio
Researcher at Inria, leading since 2011 the machine learning team which is part of the Computer Science department at Ecole Normale Supérieure. Ph.D. Berkeley (2005). ERC Starting grant (2009) and Consolidator Grant (2016), Inria young researcher prize (2012), ICML test-of-time award (2014), Lagrange prize in continuous optimization (2018). Co-editor-in-chief of the Journal of Machine Learning Research. Member of the Academy of Sciences.
Topics of interest
Machine learning, Optimization, Statistics
Project in Prairie
Francis Bach will address fundamental problems in machine learning, using optimization methods that come with theoretical guarantees and can help solve challenging problems in computer vision and natural language processing. He will participate in the teaching effort through his existing classes and in collaborations with PRAIRIE industry partners.
Quote
Machine learning algorithms are ubiquitous in most scientific, industrial and personal domains, with many successful applications. As a scientific field, machine learning has always been characterized by the constant exchanges between theory and practice, with a stream of algorithms that exhibit both good empirical performance on real-world problems and some form of theoretical guarantees. Many of the recent and well publicized applications come from deep neural networks, where these exchanges are harder to make. Bridging this new gap between theory and practice is key to obtaining performance guarantees and uncertainty estimates, which are crucial in most domains, in particular safety/
health critical ones.
