Title: "Towards Generalizable Deep Dynamics Learning"
Abstract: Deep learning holds great promise in accelerating the prediction of physical dynamics relative to numerical solvers. However, current deep learning models for dynamics forecasting struggle with generalization. They only work in a specific domain and fail when applied to systems with different parameters, external forces, or boundary conditions. In this talk, I will demonstrate our efforts in improving the generalization of deep learning for forecasting physical dynamics. I will introduce (1) Equivariant-Net: a model that is inherently equivariant to groups of symmetries and thus generalizes automatically across groups. (2) DyAd: a model-based meta-learning method which can generalize across heterogeneous domains with latent task inference. I will showcase the advantage of our approaches on forecasting Rayleigh Benard convection, real-world ocean currents, and temperatures.
Bio: Dr. Rose Yu is an assistant professor at the University of California San Diego, Department of Computer Science and Engineering. She was a Postdoctoral Fellow at the California Institute of Technology. Her research focuses on advancing machine learning techniques for large-scale spatiotemporal data analysis, with applications to sustainability, health, and physical sciences. A particular emphasis of her research is on physics-guided AI which aims to integrate first-principles with data-driven models. Among her awards, she has won Faculty Research Award from Facebook, Google, Amazon, and Adobe, Several Best Paper Awards, Best Dissertation Award in USC, and was nominated as one of the 'MIT Rising Stars in EECS'.
To obtain the Zoom link for this event, please see the event announcements from MLFL on the college email lists or contact Kalpesh Krishna.