Learning Symbolic Equations with Deep Learning

Explore the intersection of deep learning and symbolic equations in this insightful 59-minute conference talk by Shirley Ho, Acting Director of the Center for Computational Astrophysics. Delve into a novel approach for extracting symbolic representations from Graph Neural Networks (GNNs) through the use of strong inductive biases. Discover how this method successfully uncovers known physical equations, including force laws and Hamiltonians, and its application to a complex cosmology problem involving dark matter simulations. Learn about the potential of this technique for interpreting neural networks and uncovering new physical principles, as well as its superior generalization capabilities compared to traditional GNNs. Gain valuable insights from Ho's extensive experience in astrophysics, observational astronomy, and data science, and understand the broader implications of this research for fields ranging from fundamental cosmology to exoplanet statistics.

Introduction
Presentation
Linguistic Perspective
Machine Learning
Symbolic Regression
Questions
Outcomes Razor Rules
Interaction Network
Graph Network
Regularization
Dark Matter
Real Data
Input
QA