Resolving extreme jet substructure

• DOI: 10.1007/jhep08(2022)046
• Publication Date: 2022-08-04T11:06:03Z
• Authors: Yadong Lu; Alexis Romero; Michael James Fenton; Daniel Whiteson; Pierre Baldi
• Keywords: Substructure; Regularization

A bstract We study the effectiveness of theoretically-motivated high-level jet observables in extreme context jets with a large number hard sub-jets (up to N = 8). Previous studies indicate that are powerful, interpretable tools probe substructure for ≤ 3 sub-jets, but deep neural networks trained on low-level constituents match or slightly exceed their performance. extend this work up 8 using particle-flow (PFNs) and Transformer based estimate loose upper bound classification fully-connected network operating standard set observables, 135 N-subjetiness mass, reach accuracy 86.90%, fall short PFN models, which accuracies 89.19% 91.27% respectively, suggesting constituent utilize information not captured by observables. then identify additional able narrow gap, LASSO regularization feature selection rank most relevant provide further insights into learning strategies used constituent-based networks. The final model contains only 31 is performance approximate within 2%.