BIMSA >
Seminar on Control Theory and Nonlinear Filtering
Diffusion generative models and moment filtering
Diffusion generative models and moment filtering
Organizer
Speaker
Time
Tuesday, June 20, 2023 2:30 PM - 3:00 PM
Venue
数学系理科楼A-203
Abstract
In this talk, we present two of our on-going research projects. The first project is concerned with a generative model based on stochastic differential equations. More specifically, we design an approximate SDE bridging method that can draw samples from a target distribution by training a model with samples from a coupling of it. We show that this method is a state-of-the-art solution for image restoration problems (e.g., super-resolution or deblurring). The second project is concerned with a new class of convergent stochastic filters by using moment representations and a quadrature method. The key idea is to project the filtering distribution onto a sequence of moments, and use a moment quadrature method to propagate such a sequence in time. We show that this filter converges to the true distribution, and that the filter outperforms (in terms of convergence speed and computational cost) sequential Monte Carlo for low-dimensional models.