BIMSA >
BIMSA Lecture
Generative Models based on Optimal Transport and their Application to Weather Map Super-Resolution
Generative Models based on Optimal Transport and their Application to Weather Map Super-Resolution
Organizer
Speaker
Milena Gazdieva
Time
Wednesday, May 28, 2025 5:00 PM - 6:00 PM
Venue
Online
Online
Zoom 204 323 0165
(BIMSA)
Abstract
Generative models based on Optimal Transport (OT) have attracted significant interest from the ML community in recent years due to their solid mathematical foundations, scalability, and applicability to unpaired data. Despite these advantages, methods relying on the classical OT formulation are sensitive to outliers and class imbalance in the given measures, and can not be used to perform the translation which maximally preserve the features of the input samples which limits their applicability in real-world tasks. These limitations can be addressed by considering the unbalanced and partial OT formulations, and developing models based on them. In this talk, I will provide an overview of my recent papers which propose these types of generative models, provide theoretical justifications of their performance and show their advantages over classical OT counterparts in a number of synthetic and high-dimensional experiments. Additionally, I will show the results of application of these models to real-world weather map super-resolution problem.
Speaker Intro
My research interests lie in the field of generative models rooted in the optimal transport theory. More specifically, I am building generative models on the unconventional formulations of the OT problem, i.e, partial and unbalanced OT formulations.