Seminar- Patricia Ning, Texas A&M

Patricia Ning, Texas A&M

Title: Beyond Prediction: World Models for Understanding, Planning, and Action

Abstract: AI systems can now generate language, recognize images, and solve increasingly complex problems. Yet they still struggle with abilities that humans and animals acquire naturally: understanding how the physical world works, anticipating the consequences of actions, and planning efficiently. A promising direction is to build AI systems that learn world models—internal representations that predict how the world will evolve. Joint-Embedding Predictive Architectures (JEPAs) offer a powerful framework for learning such representations. But accurate prediction alone is not enough: the learned representation must also make it easy to reason about possible actions. In this talk, I will discuss recent progress by my group toward making learned world models more suitable for planning and action. I will show how structuring representations around the effects of actions can make planning substantially more efficient, and discuss what this suggests about the representations needed for AI systems that can understand, predict, and act in the physical world.

Bio: Patricia Ning is an Assistant Professor in the Department of Statistics and Data Science at Texas A&M University. She received her Ph.D. in Statistics and Applied Probability from UC Santa Barbara and held postdoctoral positions at the University of Washington and the University of Michigan. Her research focuses on scalable machine learning and quantum algorithms for high-dimensional prediction, uncertainty quantification, data-driven inference, latent structure detection, and complex network modeling. She currently serves as an Associate Editor for Bayesian Analysis, Journal of Computational and Graphical Statistics, and Statistics and Computing.