My academic research has been centered around the development and enhancement of diffusion models in deep learning, especially in the context of computer vision and non-convex optimization. My main work focuses on Score-Based Generative Models, a recent class of deep generative models with state-of-the-art performance in many applications including images and audio generations. In this model, we add noise to the original data via a forward process determined by a stochastic differential equation (SDE), then reverse this process by denoising the results to produce synthetic data. The idea was inspired by the existing score-based models using critically-damped Langevin Dynamics.
Ongoing project
I am currently working on a higher order version of score-based model, where I apply a third order SDE system in both the forward and backward process. In this project, I and my co-authors have a theoretical analysis on the optimal choice for the coefficients of the SDE, and the implementation is underway to demonstrate the theory.
Some of my potential future projects on score-based models include:
- Denoising Diffusion Bridge Models
- Flow Matching for Generative Modeling
- Text-to-image using score-based models
Wasserstein Convergence Guarantees for a General Class of Score-Based Generative Models (published)
arxiv githubDue to the existence of many families and many orders of SDE that can serve the forward process, I and my co-authors have explored these potential options to further improve the models. This results in my publication of “Wasserstein Convergence Guarantees for a General Class of Score-Based Generative Models”, which shows our joint theoretical analysis of the models and my implementation of new SDEs.