DL-SP+PM

A crash course for interdisciplinary applications of modern DL in Signal Processing and Physical Modeling.

June 13, 2024 - Building 660 DIGITEO, ground level amphitheater ( location)

Main objectives


inliers


Thank you very warmly to the speakers and to all the 40 participants who attended bravely the crash course !

Many scientific fields become very rapidly more digitized and data-driven, following the significant advances of the last decade in statistical machine learning.

This series of two half-day courses is proposed in order to support PhD students from fields such as Electrical Engineering, Physics or Mechanics who do not necessarily have an advanced ML background, and who start using more complex DL techniques for their research.

Computer Science PhD students are welcome as well if the courses are (or may become) relevant for their activity.

In addition to UPSaclay PhD students from STIC doctoral school, we welcome other PhD students, and (within the limit of places available) Master interns. In this case, the registration procedure is different (see below), and you will be notified by email.

This is a pilot edition, and your participation and feedback about the content will be very valuable for a future event!

Prerequisites: Before taking this course, students should have studied prerequisite courses such as advanced mathematics and linear algebra. Some experience with the theoretical and experimental bases of statistical and deep learning will help.

Participation: on site

  • ED STIC PhD students will receive official credit points for participation (1.6 points).
  • Students from other doctoral schools who participate (e.g., EOBE, SMEMAG etc.): we will discuss with these EDs in order to register the course in their accepted course list as well, with the same number of points (1.6). UPDATE : EOBE, SMEMAG, EDPIF, Matisse and Interfaces Doctoral Schools confirmed that the course will count for credits (the registration is still done using the Google form).

    Crash course program



    • 8h30 - 8h45: Welcome message and introduction (Emanuel Aldea and Alain Denise, UPSaclay)
    • 8h45 - 12h30: A crash course on GNNs for Signal Processing applications (Wei Hu, Peking University and UPSaclay visiting researcher) :

      Graph neural networks (GNNs) extend traditional deep learning models to non-Euclidean data, such as brain neural systems, molecular structures, knowledge graphs, traffic networks, social networks, three-dimensional geometry and other graph data. The main goal of this course is to introduce the basic concepts, theories, and classical methods of graph neural networks, and their applications in representative cutting-edge fields. The course aims to deepen students' understanding of the key ideas and cutting-edge research of graph neural networks.

    • 14h00 - 18h15: Deep Learning and Physics (Guillaume Charpiat, INRIA & LISN UPSaclay) and some perspectives in fluid mechanics(Anne Sergent, LISN Sorbonne Université) and in numerical simulations(Matthieu Nastorg, INRIA & LISN UPSaclay) :

      We will cover the main approaches to apply neural networks to physics, in particular dynamical systems. The key question is the incorporation of physical knowledge in the modelling of the machine learning task.

      • Exact invariance or equivariance by architecture design (e.g., invariance to permutations, to rotations)
      • Solving known PDEs: PINNs (Physically-Informed Neural Networks)
      • Learning a Partial Differential Equation (from recurrent networks to Neural-ODE)
      • Learning and controlling dynamical systems (Koopman - von Neumann as a justification for auto-encoders)

    This part will end with discussions of perspectives in fluid mechanics and numerical simulations, with examples (turbulent convection and Poisson problems).

    Two mid-course coffee breaks are offered by the Computer Science (ISN) Graduate School. The lectures are in English.

    How to register


    Registration is free, but mandatory by June 12.
    • ED STIC PhD students: register using ADUM ; no confirmation is necessary. You are registered, and you may attend even if you did not receive an official confirmation from ADUM.
    • other PhD students, Master interns : register using this form https://forms.gle/QuEktFA5xpQJuYN77 ; you will be informed by email.

    Organizers


    • Emanuel Aldea: Associate Professor at Paris-Saclay University, SATIE Laboratory
    • Catherine Nizery : Management Assistant, ENS Paris-Saclay, SATIE Laboratory