Dear all /Bonjour à tous,

We are happy to invite you to the Predoc III evaluation of Léna Néhale Ezzine on December 17th at 2h30 pm (hybrid mode).

Vous êtes cordialement invité.e.s à l'évaluation du Predoc III de Léna Néhale Ezzine, le 17 décembre à 14h30 (mode hybride).

Jury

Abstract

Can AI help us advance our understanding of the physical world at the
atomic scale ? This question gave rise to numerous research directions in
the machine learning community, in particular geometric deep learning and
generative modeling. In this predoc, we focus on the latter. Two classes of
generative models, namely diffusion models and flow-matching, have become
state-of-the-art for generating energetically favorable 3D structures of
molecules and proteins. However, both rely on the availability of
experimentally generated data. In the absence of such data, we would only
have access to an oracle that indicates the goodness of a datapoint. Unlike
diffusion models and flow-matching, GFlowNets allow to sample useful and
diverse objects according to a reward coming from the orcale. We explore
different facets of GflowNets: first, we introduce GflowNets for
Maximum-Likelihood Estimation (MLE), an algorithm that allows to train
GflowNets on a dataset only, without the need for a reward function. Then,
we provide an overview of the theory of continuous GflowNets, a
generalization of discrete GflowNets to general state spaces, and a
required framework for sampling the structures of molecules and proteins in
continuous space. Finally, we introduce an ongoing project and preliminary
results of a diffusion model trained with a GflowNet reward-matching
objective to sample from the Boltzmann distribution of molecules.