Interpretable Deep Learning for New Physics Discovery

Explore a 24-minute video lecture on interpretable deep learning for new physics discovery. Learn about a method for converting neural networks into analytic equations using specific inductive biases. Discover how sparsification of latent spaces in deep neural networks, combined with symbolic regression, can recover physical laws for various systems. Gain insights into applications such as discovering gravity and planetary masses from data, performing cosmology with cosmic voids and dark matter halos, and extracting the Euler equation from graph neural networks trained on turbulence data. Follow along as the speaker, Miles Cranmer, discusses symbolic regression, genetic algorithms, PySR, the combination of deep learning and symbolic regression, graph neural networks, and their applications in recovering physics. Examine results on unknown systems and key takeaways from this innovative approach to interpretable machine learning in physics.

Introduction
Symbolic Regression Intro
Genetic Algorithms for Symbolic Regression
PySR for Symbolic Regression
Combining Deep Learning and Symbolic Regression
Graph Neural Networks
Recovering Physics from a GNN
Results on Unknown Systems
Takeaways