清华主页 EN
导航菜单

Global well posedness of Score-Based Generative model via Sharp Lipschitz estimate

来源: 06-04

时间:Wednesday, 14:00-15:00 June 5, 2024

地点:C548, Shuangqing Complex Building A 清华大学双清综合楼A座 C548

组织者:荆文甲

主讲人:王中剑 新加坡南洋理工大学

Abstract:

We establish global well-posedness and convergence of the score-based generative models (SGM) under minimal general assumptions of initial data for score estimation. For the smooth case, we start from a Lipschitz bound of the score function with optimal time length. The optimality is validated by an example whose Lipschitz constant of scores is bounded at initial but blows up in finite time. This necessitates the separation of time scales in conventional bounds for non-log-concave distributions. In contrast, our follow up analysis only relies on a local Lipschitz condition and is valid globally in time. This leads to the convergence of numerical scheme without time separation. For the non-smooth case, we show that the optimal Lipschitz bound is $O(1/t)$ in the point-wise sense for distributions supported on a compact, smooth and low-dimensional manifold with boundary.

返回顶部
相关文章
  • Deep Generative Models

    Description: This course introduces how to develop deep generative models (DGMs) by integrating probabilistic graphical models and deep learning to generate realistic data including images, texts, graphs, etc. Course contents include 1) basics of probabilistic graphical models, including Bayesian network and Markov random field; 2) posterior inference methods, including message passing, variat...

  • YMSC Topology Seminar | Horospheres, Lipschitz maps, and laminations

    Abstract:Every horocycle in a closed hyperbolic surface is dense, and this has been known since the 1940's. We study the behavior of horocycle orbit closures in Z-covers of closed surfaces, and obtain a fairly complete classification of their topology and geometry. The main tool is a solution of a surprisingly delicate geometric optimization problem: finding an optimal Lipschitz map to the cir...