Dear all / Bonjour à tous,

We are happy to invite you to the Predoc III Evaluation of Jarrid Rector-Brooks on August 20th at 2h30 pm (hybrid mode).

Vous êtes cordialement invité.e.s à l'évaluation du Predoc III de Jarrid Rector-Brooks , le 20 août à 14h30 (mode hybride).

Jury

Pierre-Luc Bacon

Yoshua Bengio

Gauthier Gidel

Abstract

Efficiently generating statistically independent samples from an
unnormalized probability distribution, such as equilibrium samples of
many-body systems, is a foundational problem in science. While the practice
of training generative models in the presence of datasets via maximum
likelihood is mature and able to scale to very high dimensional problems
the same is not true for sampling given only the unnormalized distribution.
Current approaches to sampling from unnormalized densities range from
Markov Chain Monte Carlo (MCMC) methods which strug- gle to move from mode
to mode in high dimensions to different neural samplers that,
unfortunately, have various drawbacks in their learning procedures holding
them back from scaling to very high dimensional problems. For example, some
require integrating an entire SDE trajectory to evaluate the loss, others
rely on con- sistency objectives where credit assignment is a struggle,
while others require the restrictive architecture choices preventing wider
adoption. In this thesis proposal we attempt to take inspiration from
approaches that work well when training via maximum likelihood, in
particular diffusion and flow matching models, to further the scalability
of neural sampling algorithms.