Generative Artificial Intelligence | Yulin Du | TEDxMIT

I will talk about how we can use generative models beyond just creating fun art but for many practical and important real world problems. I’ll talk back how you can use them for perceiving the world as well as to both synthesize novel behaviors of agents as well as design new proteins. I am a fourth-year PhD student at MIT EECS, advised by Prof. Leslie Kaelbling, Prof. Tomas Lozano-Perez and Prof. Joshua B. Tenenbaum. Previously, I obtained my bachelor’s degree from MIT, was a research fellow at OpenAI, an intern at Deepmind and FAIR, and got a gold medal at the International Biology Olympiad.

I am interested in constructing machine learning tools that enable the development of autonomous embodied agents. In the embodied setting, the world is both highly uncertain and richly combinatorical in nature. To address these challenges, my recent research uses the tools of energy-based models to accurately generatively model the uncertainty in the world and as a tool to construct composable models which may be rapidly adapted to new experiences. My research further uses the underlying energy optimization procedure as an adjustable computational budget, enabling the use of longer computation times to adapt to novel out-of-distribution experiences. Furthermore, embodied learning is richly multimodal in nature, and we need models which universally capture structure across modalities such as vision, text, sound and touch. I am interested in leveraging neural fields as a generic way to discover and capture such rich structure in the world. Finally I’m interested in broader applications of these tools to other domains such as computational biology. This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at