The chair is organized along the following subjects

Theory: algorithmic dierentiation, structural results, robustness

  • A proper theory for automatic dierentiation: ReLU and nonsmooth Neural Networks
  • Landscapes: structure of Neural Network losses (linear networks, classication of critical points)
  • Neural Networks with quantized weights (for embarked NN)
  • Parameter identications for interpretability, \soft-certication »
  • Neural Networks for solving blind inverse problems
  • Sensitivity of Neural Networks: Lipschitz constants

Algorithms and applications: convergence theory and rates

  • Small steps methods for training networks: a convergence theory?
  • Large steps methods for non Euclidean methods
  • Algorithms for adversarial learning
  • Optimization algorithm with certication for semi-innite programmation on measure space
  • Improvement of imaging systems, in particular in magnetic resonance, for microscopes and telescopes

Programme : IA certifiable
Thèmes : IA et modèles physiques, Optimisation et théorie des jeux pour l’IA

Porteur :
Jérôme Bolte, Université Toulouse Capitole

Permanent researchers:
F. de Gournay (Insa)
F. Malgouyres (UT3),
E. Pauwels (UT3),
P. Weiss (CNRS),
Post-doctoral fellows
1 Year, L. Glaudin (JB and EP)
2 Years, R. Rios-Zertuche (JB and EP)
Master intership: E. Dadalto-Camara-Gomes (FM), R. Razine (PW), E. Pauwels (T. L.Ngoc)
3 PhD students, R. Dragomir (JB), J. Bona-Pellisier (FM), E. M. Achour (FM)
For the year 20-21: 10 people are working on the project, 3 internships
Regular collaborators out of ANITI:
A. d’Aspremont (ENS)
L. Miclo (CNRS & UT1)
S. Villeneuve (CNRS & UT1)
S. Sabach (The Technion)