Neural message passing for jet physics



Supervised learning has incredible potential for particle physics, and one application that has received a great deal of attention involves collimated sprays of particles called jets. Recent progress for jet physics has leveraged machine learning techniques based on computer vision and natural language processing. In this work, we consider message passing on a graph where the nodes are the particles in a jet. We design variants of a message-passing neural network (MPNN);(1) with a learnable adjacency matrix,(2) with a learnable symmetric adjacency matrix, and (3) with a set2set aggregated hidden state and MPNN with an identity adjacency matrix. We compare these against the previously proposed recursive neural network with a fixed tree structure and show that the MPNN with a learnable adjacency matrix and two message-passing iterations outperforms all the others.


Isaac Henrion
Johann Brehmer
Joan Bruna
Kyunghyun Cho
Kyle Cranmer
Gilles Louppe
G Louppe
Gaspar Rochette
G Rochette