Statistics Ph.D. Dissertation Defense - Zhendong Wang

This 2024 Dissertation Defense will be held on Friday November 15, from 8:00 a.m. to 10:00 a.m. with Zhendong Wang. This event will be hybrid. If you are able to attend in person, it will be held in WEL 5.204. If you need the Zoom link, please email stat.admin@austin.utexas.edu.

Title: Enhancing Efficiency and Controllability in Generative Models for Reinforcement Learning and Robotics 

Advisor: Dr. Mingyuan Zhou

Abstract: The rapid progress in generative models has enabled their application across diverse areas, from synthesizing realistic images to tackling complex decision-making tasks. My dissertation focuses on enhancing both the efficiency and controllability of generative models, addressing significant challenges within offline reinforcement learning and robotics.

To improve efficiency, I developed Diffusion-GAN, a new algorithm that uses diffusion techniques to stabilize GAN training, achieving faster and higher-quality image generation. Another key contribution is Patch Diffusion, which trains diffusion models on smaller, localized patches, significantly enhancing both training and data efficiency.

For greater controllability, I introduce Prompt Diffusion, the first framework exploring in-context learning within diffusion models for image-based applications, enabling more refined control over the generative process.

In addition, I leverage advanced diffusion models as expressive policy classes, boosting performance in offline reinforcement learning tasks. Continuingly, I explore using fast, distilled diffusion models to enable responsive control in real-world robotics.