Learning and Optimization

Building on the machine learning (ML) and stochastic optimization expertise of the previous TAO project team, the TAU team aims to address the imprecise goals of Big Data. Assuming that (sufficiently) large data can, to some extent, compensate for the lack of knowledge, Big Data is supposed to fulfill all the commitments of artificial intelligence.

Research Themes

The A&O team – a joint Paris-Saclay, CNRS and INRIA Saclay project team – is interested in learning models from data, focusing on four fundamental axes.

  • The first one concerns adversarial learning based on the interaction of two or more learning agents, replacing the unknown objective function by a min-max approach (game theory); this axis is also crucial for the validation and certification of neural networks.
  • The second one deals with the selection and configuration of algorithms a priori, according to the available data, also called AutoML; this is not only a necessary condition for the democratization of AI, but also an open challenge for the last 40 years, related to the definition of the data order parameters.
  • The third one tackles the problems of learning complex models and deals with the identification of regularities allowing the well-founded augmentation of data in the many application domains where data are small compared to the complexity of the models sought.
  • Finally, far from replacing knowledge by models purely derived from data, one objective is to interact with the knowledge of the domain, for example expressed by partial differential equations; the challenge here is to bridge the gap between Learning and Numerical Engineering, in connection with the Department of Fluid Mechanics-Energy