Seminario: Optimización y Aprendizaje Automático (2025)

A lo largo de 2025-2 el Seminario consistirá de varias sesiones extraordinarias a cargo de investigadores distinguidos.


Miércoles, Octubre 22, 2025 (14.30-16.00 Salón: 727-Gris (7mo. piso), FING)

Expositor: Marco Cuturi [link]

Afiliación: Apple ML Research, Institut Polytechnique de Paris.

Título: Optimal Transport and Generative Modeling

Resumen: After a brief introduction to computational OT, I will present in this talk a few recent works highlighting the interest of OT solvers in generative problems. I will start with a recent result on a factorization of probability measures that builds upon the idea of moment measures introduced by Cordero-Erausquin and Klartag in 2015. I will show how this idea can be modified to suit better the generative paradigm to present conjugate moment measures. Next, I will show how simple OT maps on activations can help control --- using very few intervention examples --- the generation of LLMs and text-2-image models. Finally, I will investigate to what extent the use of large coupling solvers can improve the training of flow matching models.

Referencias:

Lunes, Noviembre 3, 2025

Expositor: Joan Bruna [link]

Afiliación: Courant Institute (NYU)

Título: TBA

Miercoles, Noviembre 5, 2025

Expositor: Soledad Villar [link]

Afiliación: Johns Hopkins University

Título: TBA

Ver charlas de 2025-1 aquí.


ORGANIZADORES: