Applied Math Seminar-Supervised Generative Model for Density Estimation

Date and Time: Friday, October 4th 1:30pm-2:30pm

 

Place: Zoom Link: https://gwu-edu.zoom.us/j/4092408949

 

Speaker: Dr. Yanfang Liu, Oak Ridge Lab

 

title: Supervised Generative Model for Density Estimation

 

abstract: In this talk, we will present a supervised learning framework of training generative models for both unconditional and conditional density estimation. Our approach utilizes a score-based diffusion model to generate labeled data, enabling the training of the generative model in a supervised manner. Instead of relying on neural networks to approximate the score function, we employ a training-free score estimation technique using mini-batch Monte Carlo estimators. With score function estimation, labeled data are obtained by solving an ordinary differential equation (ODE) that corresponds to a reverse-time stochastic differential equation (SDE). The generative model is then learned using the labeled data with a simple mean squared error (MSE) approach. We demonstrate the effectiveness of this supervised generative model for both unconditional and conditional density estimation.