Lorentz-equivariant transformers for the LHC
by
E/3rd floor-E.349 - Seminar room (E.349)
Marc de Hemptinne (chemin du Cyclotron, 2, Louvain-la-Neuve)
Lorentz-equivariant neural networks are becoming the leading architectures for high-energy physics. We introduce Lorentz Local Canonicalization (LLoCa) and the Lorentz-equivariant Geometric Algebra Transformer (L-GATr), two complementary frameworks to construct Lorentz-equivariant transformers. L-GATr employs specialized layers operating on Clifford Algebra representations, similar to the Dirac Algebra representations used in QFT. LLoCa is a recipe to make any neural network Lorentz-equivariant using equivariantly predicted local reference frames, inspired by the tetrad formalism in GR. We compare L-GATr and LLoCa with other architectures for QFT amplitude regression and jet tagging, and use them to construct generative networks.
Zoom: https://cern.zoom.us/j/62591938215?pwd=qSqjURfb693pfZR7aFvYR2vJw3YnuD.1