Symmetry Aware Machine Learning.

Abstract

We investigate an equivariant neural network architecture that is equivariant with re- spect to operations of the Lorentz group. The basis of the architecture is the Equivariant Universal Approximation, which specifies constraints for any architecture so that it effec- tively simulates physical processes. We demonstrate that an equivariant architecture like this has fewer learnable parameters with its components being much more physically in- terpretable for classification tasks like top tagging in particle physics. The performance of the neural network is measured using the Top Quark Tagging Reference Dataset [1], for tagging hadronic top quark decays given the 4-momenta of jet constituents.