Seminar: Graduate Seminar
Diffusion Transportation Cost for Optimal Transport in Domain Adaptation
Date:
January,08,2025
Start Time:
13:00 - 14:00
Location:
1061, Meyer Building
Add to:
Lecturer:
Gal Maman
Research Areas:
| Optimal Transport (OT) is a versatile framework for aligning distributions, with applications spanning computer vision, biology, and natural language processing. In recent years, there has been considerable interest in utilizing the OT problem for domain adaptation, a strategy shown to be highly effective. However, most approaches rely on the pairwise squared Euclidean distances for the transportation cost, potentially overlooking important intra-domain geometries.In this talk, I will introduce Diffusion-OT, a novel transport cost designed for Unsupervised Domain Adaptation. By leveraging diffusion geometry, Diffusion-OT employs a three-step diffusion process to capture both intra-domain structures and inter-domain relationships, with source label information incorporated to guide the diffusion. I will present theoretical results for the operator underlying the cost, analyze its asymptotic behavior to demonstrate that the cost function is defined by an anisotropic diffusion process between the domains, and illustrate its properties with a visual example. Finally, I will show how integrating the proposed cost into deep learning frameworks yields superior performance on benchmark datasets. Additionally, I will demonstrate its state-of-the-art results in Brain-Computer Interface (BCI) tasks, achieved by leveraging Riemannian geometry.M.Sc. student under the supervision of Prof. Ronen Talmon.
|