Score-based generative models with constraint

April 7, 2023
3:05pm ET
Cummings 265
Speaker: Yukun Li - Quals talk
Host: Liping Liu

Abstract

Quals talk:

Recently the score-based generative model has become popular. Rather than directly maximizing the log-likelihood of the density function, it models the score of the data to avoid explicitly calculating the normalizing constants. However, the current score-based model needs to fully consider the raw data property (e.g., molecule valency property ), which may generate sub-optimal samples. In this work, we propose a constrained score model by adding pre-designed sufficient statistics to the model so that it can guide the learning of the score model.