Place: Online Seminar: Please sign up for our mailing list at www.physicsmeetsml.org for zoom link
Speaker: Miles Cranmer, Princeton University
Abstract: If we train a neural network on some dynamical system in some region of phase space, and it learns a way to execute the dynamics more efficiently than a handwritten code, how do we distill physical insight from the learned model? In this talk, I will argue that symbolic learning should play a major role in the process of interpreting a machine learning model for physical systems. I will discuss our generic method for converting a neural network that has been trained on a physical system into a symbolic model, via genetic algorithm-based symbolic regression. One of the problems with this process is working with the fact that neural networks have high-dimensional latent spaces, and genetic algorithms scale poorly with the number of features. To work around this issue, I’ll then introduce our “Disentangled Sparsity Network,” which encourages a neural network to learn an easy-to-interpret representation. I will then share several recent applications of our techniques to real physical systems, and the various insights we have discovered and rediscovered.