Optimal Transport for Generative Modeling: From Training to Fine-Tuning
김선우
27동 220호
0
666
05.14 16:23
| 구분 | ACM |
|---|---|
| 일정 | 2026-06-15(월) 16:00~18:00 |
| 세미나실 | 27동 220호 |
| 강연자 | 최재웅 (성균관대학교) |
| 담당교수 | 이다빈 |
| 기타 |
Optimal transport (OT) theory provides a principled framework for modeling transformations between probability distributions. From this perspective, generative models can be viewed as mechanisms for transporting a simple prior distribution to a target data distribution.
In this talk, we present a unified perspective on generative modeling through OT. We discuss how OT-based frameworks provide principled tools for both training and fine-tuning generative models.
First, we introduce Unbalanced Optimal Transport (UOT)-based methods for robust training, long-tailed generation, and unlearning. Second, we discuss Wasserstein Gradient Flow (WGF)-based methods, where generative modeling and reward-guided fine-tuning are formulated as steepest-descent dynamics of functionals in Wasserstein space. Finally, we present Schrödinger Bridge and stochastic optimal control perspectives for efficient path-space generative modeling and reward alignment.
