GT Transport optimal - EDP - Machine learning
Mind the reference measure of entropic OT: a study on Gaussian measures
16
mai 2022
mai 2022
Intervenant : | Hicham Janati |
Institution : | Télécom Paris |
Heure : | 15h15 |
Lieu : | 3L15 |
Since the introduction of entropic regularization to optimal transport, several works investigating its bias as well
as statistical properties have been carried out (debiasing, statistical complexity ..) In this presentation, I will discuss
the various implications of the very definition of entropy: discrete, continuous, semi-discrete and debiased. The first part
of the talk will be dedicated to comparing these formulations (+barycenters) for Gaussian measures. In the second part I will
present generalizations of closed form formulae and some potential conjectures for unbalanced OT (with/and without entropy) as well
as multi-marginal OT.