S'abonner à l'agenda

Mercredi 20 décembre 2023

Salle de conférences IRMA

This is a colloquium in Physics and Mathematics, organised biannually in Strasbourg by a group of scientists from IRMA, ISIS and IPCMS. This session is devoted to "AI and Machine learning".

Organizers : N. Anantharaman, R. Côte, S. Klevtsov, Y. Le Floch, C. Tauber, M. Vogel, X. Zeng (IRMA) — D. Hagenmuller, G. Pupillo, J. Schachenmayer (ISIS) — P.-A. Hervieux, R. Jalabert, G. Manfredi, G. Weick, D. Weinmann (IPCMS).

Speakers :

  • Florian Marquardt
  • Laurent Navoret

Location : Salle de conferences IRMA.

For practical and other questions please contact the organizers..

  • Mercredi 20 décembre 2023

  • 09:30

    Florian Marquardt, Universität Erlangen-Nürnberg

    Training learning machines by purely physical dynamics

    In the field of neuromorphic computing, we consider physical systems as information-processing devices that can be trained. The hope is for these new 'learning machines' to eventually replace digital artificial neural networks, with the advantages of higher energy efficiency and parallel data processing. One outstanding challenge in this field is to use physics not only for the information processing but also during the training process. In this talk I will introduce the approach of Hamiltonian Echo Backpropagation. This represents a new general learning procedure that accomplishes in an entirely physics-based way both the evaluation of gradients and the update of trainable parameters. It works for any time-reversal-invariant Hamiltonian dynamical system and does not require any access to the internal parameters of the system. Reference: Self-Learning Machines Based on Hamiltonian Echo Backpropagation Víctor López-Pastor and Florian Marquardt Phys. Rev. X 13, 031020
  • 10:30

  • 11:00

    Laurent Navoret, Université de Strasbourg

    Reduced modeling using neural networks

    Numerical simulations of complex physical models can require significant computing resources. Therefore, reduced modeling has been developed to construct smaller representations of the unknowns, whose dynamics are faster to simulate. In this talk, we will present two methods showing how neural networks can be used to obtain such reduced models. The first method tackles the question of extending the validity of fluid equations to weakly collisional regimes: there, neural networks have been used to learn non-linear dependencies of the heat flux and the pressure tensor on the fluid quantities (density, momentum and energy). The learning is based on numerical simulations of the underlying kinetic dynamics. The second method consists in directly determining a reduced representation using neural networks and then learning the reduced dynamics. In both cases, a specific attention has to be paid regarding the stability of the obtained models. These works have been carried out in collaboration with Léo Bois, Raphaël Côte, Emmanuel Franck, Guillaume Steimer and Vincent Vigon.
  • 12:00