I hold a Research Scientist position at Criteo AI Lab. Previously, I was a PhD student at Sorbonne Université in the LIP6 and ISIR laboratories within team MLIA, under the direction of Sylvain Lamprier and Patrick Gallinari.
I am interested in machine learning and more specifically deep learning. Since the beginning of my PhD studies, I have been working on dynamical systems, i.e. temporal differential equations, for deep generative modeling.
One of the main directions that I have explored was the modeling spatiotemporal data by differential equations. In particular, I have studied how to leverage differential equations and their links with neural networks in order to create performant prediction models for complex structured data such as videos. Inspired by such connections, I designed with colleagues a novel temporal model based on residual connections instead of recurrent neural networks, and that lead to state-of-the-art results for stochastic video prediction. We also leveraged a resolution technique for partial differential equations, the functional separation of variables, to propose a novel interpretation of spatiotemporal disentanglement, leading to a simple but performant disentangled predictive model.
Another promising direction is to employ dynamical systems to directly study the training dynamics of deep generative models. Indeed, we analyzed differential equations describing the temporal evolution of neural networks during training. In particular, thanks to the study of these differential equations, and leveraging the neural tangent kernel theory, we proposed a novel theoretical framework for the understanding of generative adversarial networks, which are performant generative model for both static and sequential data.
Other Research Areas
During the second half of my studies at Ecole Normale Supérieure de Lyon, I became interested in machine leaning and artificial intelligence, and have done three related long research internships since then.
I worked more particularly on:
- fairness and accountability of automated decision-making processes;
- robustness of classifiers to adversarial and random examples;
- convex optimization;
- unsupervised representation learning for time series.
Previous Research Experience
Here is a list of my research internships (in reverse chronological order), that were done within the scope of my studies at Ecole Normale Supérieure de Lyon:
- EPFL, MLO laboratory, Lausanne, Switzerland, supervised by Martin Jaggi (5 months): unsupervised general purpose scalable representation learning for time series;
- ENS de Lyon, LIP laboratory, team MC2, Lyon, France, supervised by Omar Fawzi (6 months): robustness of classifiers to random and adversarial perturbations, convex optimization;
- Inria, team Privatics, Lyon, France, supervised by Daniel Le Métayer: fairness and accountability of automated decision-making processes;
- University of Konstanz, Computer Graphics and Media Informatics working group, Konstanz, Germany, supervised by Abdalla G. M. Ahmed and Oliver Deussen: blue-noise sampling;
- Inria, team Marelle, Valbonne, France, supervised by Yves Bertot: integration of a logical system in the higher order proof system Coq.
You can find more information about my research experience in my CV.