Extracting Interpretable Physical Parameters from Spatiotemporal Systems using Unsupervised Learning
ORAL
Abstract
Real-world data from spatiotemporal systems is often difficult to analyze and interpret due to complex dynamics as well as uncontrolled experimental variables. We demonstrate an unsupervised machine learning technique for extracting interpretable physical parameters from noisy spatiotemporal data and for building a transferable predictive model of the system. This is accomplished without prior knowledge of the underlying dynamics or the governing partial differential equation (PDE). Numerical experiments using simulated data governed by PDEs show that our method accurately identifies and extracts relevant parameters that characterize independent variations in the system dynamics. Our method for discovering interpretable latent parameters in spatiotemporal systems will allow us to better understand real-world phenomena by analyzing datasets with varying dynamical behaviors that are difficult to disentangle.
*This work is supported in part by the U.S. Department of Defense through the National Defense Science & Engineering Graduate Fellowship (NDSEG) Program, the MIT-SenseTime Alliance on Artificial Intelligence, the Army Research Office under Cooperative Agreement Number W911NF-18-2-0048, and the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR00111890042.
–
Presenters
-
Peter Lu
- Physics, Massachusetts Institute of Technology
- Department of Physics, Massachusetts Institute of Technology