Louis THIRY, PhD

• 2024-2025: Chercheur sur les agents logiciels autonomes basés sur LLMs et VLMs, H Company, Paris.
• 2024: Postdoc sur la modelisation hybride physique/apprentissage de l'ocean, équipe ANGE, INRIA Paris.
• 2021-2023: Postdoc sur la modélisation stochastique et l'assimilation de donnés pour les écoulement océaniques au sein de l'équipe ODYSSEY de l'INRIA Rennes.
• 2018-2021: Doctorat sur "L'efficacité des méthodes locales pour la classification d'images et la regression d'énergie en physique", effectué au sein de l'équipe DATA du département d'informatique de l'ENS Paris, sous la direction de Stéphane Mallat.

Enseignement/Formations (Teaching)

• Cours statistiques et probabilités, 1A Info, ENS Rennes, 2022.
• TPs Fondamentaux du Deep Learning, Slides 1, Slides 2, Slides 3.

Vulgarisation scientifique (Outreach)

• Guest Speaker on AI and Art, P versus NP Event, Futurium Museum, Berlin. Live demo of AI-generated portraits drawn by a robotic arm.

Publications

• Mécanique des fluides océanique/atomsphérique (Ocean/atmosphere fluid dynamics)

• A unified formulation of quasi‐geostrophic and shallow water equations via projection, Journal of Advances in Earth Modeling Systems, 2024.
• MQGeometry-1.0: a multi-layer quasi-geostrophic solver on non-rectangular geometries, Geoscientific Model Development, 2024.
• Transport Noise Defined from Wavelet Transform for Model-based Stochastic Ocean Models, STUOD, Spinger, 2023
• Modified hyper-viscosity for coarse resolution ocean models, Stochastic Transport in Upper Ocean Dynamics Workshop, Imperial College London, UK, 2023.
• Finite-volume discretization of the quasi-geostrophic equations with implicit dissipation, authorea preprint.

• Classification d'images (Image classification)

• Classification-Denoising Networks, Preprint, 2024.
• The Unreasonable Effectiveness of Patches in Deep Convolutional Kernels Methods, International Conference of Learing Representation (ICLR) 2021.
• Deep network classification by Scattering and homotopy dictionary learning, International Conference of Learing Representation (ICLR) 2020.

• Apprentissage en chimie quantique et physique des materiaux (Machine learning applied to quantum chemistry and condensed matter physics)

• Pushing the frontiers of density functionals by solving the fractional electron problem, Science, 9 Dec 2021, Vol 374, Issue 6573, pp. 1385-1389.
• Machine learning surrogate models for prediction of point defect vibrational entropy, Physical Review Materials 4,063802, 2020.
• Solid harmonic wavelet scattering for predictionsof molecule properties, J. Chem. Phys. 148, 241732 (2018).

• Interaction avec des artistes (Interaction with artists)

• Diptychs of human and machine perceptions, NeurIPS 2020 workshop "Machine Learning for Creativity and Design".
• Interactive Neural Style Transfer with Artists, International Conference of Computational Creativity (ICCC) 2020, [photo pairs].
• Dialog on a canvas with a machine, NeurIPS 2019 workshop "Machine Learning for Creativity and Design".

• Développement logiciel (Software)

• Kymatio: Scattering Transforms in Python, Journal of Machine Learning Research, 21(60):1−6, 2020.

Implémentations open-source (open-source implementations)

• Python (numpy-scipy) implementation of Angular Fourier Series (AFS) descriptors.
• PyTorch implementation of 3D solid-harmonic Scattering Transform.
• PyTorch implementation of Style Transfer with Relaxed Optimal Transport and Self-Similarity (STROTSS), additional Sinkhorn Earth Mover's distance implemented.
• PyTorch implementation of Arakawa-lamb discretisation of multi-layer Quasi-geostrophic equations.
• PyTorch implementation of multi-layer Quasi-geostrophic equations new discretization.
• PyTorch implementation of projected multi-layer Quasi-geostrophic equations and multi-layer Shallow water equations.

Videos de simulations numériques de l'océan (Ocean numerical simulation)

• Video of mediterranedan see spin-up 3-layers QG simulation, code MQGeometry.
• Video of north-atlantic ocean spin-up 3-layers QG simulation with bottom topography, code MQGeometry.
• Double gyre with 3-layers Shallow-Water simulation starting from 3-layers QG initialisation (illustrates non-constant Coriolis force in SW) , code QGSW.

Présentations orales (Talks)

• Implicit eddy parameterization, workshop Ciao Andiamo, ICSD Sorbonne Universite, 8 sept. 2022, slides.
• Implicit eddy parameterization, séminaire Sandbox ERC STUOD project, INRIA Rennes qnd online , 23 mai 2022, slides.
• Efficiency of patch representations for image classification, séminaire Deep Learning des equipes INRIA Sirocco et Serpicon, INRIA Rennes, 21 mars 2022, slides.
• Soutenance de These (PhD defense), ENS Paris, 9 juillet 2021, slides.
• The unreasonable effectiveness of patches in convolutional kernel methods, séminaire de l'équipe MLIA du LIP6, laboratoire d’informatique de Sorbonne Université, 26 avril 2021, slides.
Efficiency of local methods in image classification and energy regression in physics, séminaire de statistiques de l'IRMAR, université Rennes 1, 19 mars 2021, slides.
• Image classification with patches neighborhood encoding, équipe PANAMA, INRIA Rennes, 12 novembre 2020, slides.
• Interactive neural style transfer with artists, ICCC 2020, virtual conference, 7-11 sept. 2020, slides, link.
• Décomposition des images par les algorithmes et conséquences sur leur circulation, Colloque IMAGO-ENS, La circulation des images, Paris (virtuel), 15-18 juin 2020, video (english), link.
• Deep Network Classification by Scattering and Homotopy Dictionary Learning, Séminaire Images Optimisation et Probabilités, Universite de Bordeaux, 30 jan. 2020, link, slides.
• Image classification with Scattering and Dictionary learning, INRIA Parietal Team Seminar, Saclay, France, 22 oct. 2019, slides.
• Solid harmonic scattering transform for atomization energy regression, CECAM discussion meeting “Coarse-graining with Machine Learning in molecular dynamics”, Tuesday, December 4 – Thursday, December 6, 2018 slides, link.
• Solid Harmonic Wavelet Scattering for Predictions of Molecule Properties, Cermics Lab, ENPC, Paris, 17 May 2018 slides, link.

E-mail:prenom.nom@ens.fr