·
Self-Supervised Jet Representation Learning
Contrastive learning for jets where the two views come from the simulator. They share the collision and the parton shower, and the later random stages are run twice, so the invariances come from the data-generating process.
Key result
Re-simulation
as the augmentation
Positive pairs come from the simulation chain: the same collision and parton shower, with hadronization and detector response run twice. Tested on a Geant4-based detector simulation, with the backbone used for quark/gluon classification, hadronic tau identification and dark-jet anomaly detection.
Resources