Init ML - ClipDrop March 2021 -
Lead machine learning scientist @ ClipDrop, computer vision deep models for matting, inpainting, super-resolution, depth and surface-normals estimation, text-to-image generative models, ...
I have recently obtained my Ph.D from Institut de Physique Théorique
(IPhT - CEA Saclay, Université Paris-Saclay).
Prior to this, I obtained a Master of Science in Theoretical physics at Ecole Normale
Supérieure de Paris and a Master of Science and Engineering from Ecole
polytechnique. I obtained as well a Bachelor of Science from Ecole
polytechnique and was student in Classe Préparatoire aux Grandes Ecoles
at Lycee Henri IV in Paris.
I am interested in statistical physics of disordered systems and I worked during my Ph.D
with Lenka Zdeborová and Florent Krzakala on the application
of these methods to provide a theoretical understanding of deep neural
networks and more classical machine learning models and algorithms.
In my spare time, I used to play rugby 🏈 and I still practise skydiving, windsurfing, kitesurfing,
surfing 🏄, hiking, biking 🚵♂️ and
a lot of sport in general!
Lead machine learning scientist @ ClipDrop, computer vision deep models for matting, inpainting, super-resolution, depth and surface-normals estimation, text-to-image generative models, ...
As suggested by the formulation "most cows appear in grass and most camels appear in sand", Empirical Risk Minimization relies essentially on the dominant background color pixels to produce a prediction. To address this issue, the idea is to learn invariant features across multiple training distributions and to use those correlations as a proxy for out-of distribution generalization. To clarify the situation, as a core project we started to design simple, memory cheap, linear 'unit-test' that already capture OoD generalization failures. We designed and proposed three linear problems (with three scrambled versions) that contain invariant causal correlations that we would like to learn, as well as a spurious correlation that we would like to discard.
I established the phase diagrams of various machine learning models, using the replica method and message passing algorithms from statistical physics of disordered systems, with an emphasis on the potential differences between statistical and algorithmic thresholds, for simple synthetic and theoretically tractable tasks.
Questionner in physics and chemistry for the preparation of competitive exams.
Yves Couder, Emmanuel Fort and coworkers recently discovered that a millimetric droplet sustained on the surface of a vibrating fluid bath may self-propel through a resonant interaction with its own wave field. This internship within the department of Mathematics, aimed to study such a droplet, called a walker, crossing a submerged obstacle. This difference of depth, due to the obstacle, is analog to a changement of media and optical index in optics and thus we showed a Snell-Descartes law for walking droplets.
The advent of graphene and topological insulators in 2005 and 2006 was a revolution in condensed matter physics. The main specificity of these two-dimensional materials have probed some theories established in very remote contexts in condensed matter, such as quantum mechanics and relativistic field theory, rather encountered in high-energy physics. It is about understanding the appearance of topological band structures in two-dimensional materials, semiconductors or semi-metal, covered by models called " strong bonds " and the study of electrical and thermal transport properties these particular materials by means of statistical physics .
I was responsible of developing a measuring bench of the influence of external disturbances including sensitivity to mechanical vibration of new optoelectronic oscilateurs (OEO), which allow the generation of signals with very low noise for radar applications.