Symmetries, Safety, and Self-Supervision

Aug 15, 2022, 4:20 PM
Auditorium VMP8 (University of Hamburg)

Auditorium VMP8

University of Hamburg

Von-Melle-Park 8 20146 Hamburg Germany
Presentation ML


Lorenz Vogel (Heidelberg University)


Collider searches face the challenge of defining a representation of high-dimensional data such that (i) physical symmetries are manifest, (ii) the discriminating features are retained, and (iii) the choice of representation is data-driven and new-physics agnostic. We introduce JetCLR (Contrastive Learning of Jet Representations) to solve the mapping from low-level jet constituent data to optimized observables through self-supervised contrastive learning. Using a permutation-invariant transformer-encoder network, physical symmetries such as rotations and translations are encoded as augmentations in a contrastive learning framework. As an example, we construct a data representation for top and QCD jets and visualize its symmetry properties. We benchmark the JetCLR representation against other widely-used jet representations, such as jet images and energy flow polynomials (EFPs).

Primary authors

Barry Dillon (Heidelberg University) Gregor Kasieczka (University of Hamburg) Hans Olischläger (Heidelberg University) Tilman Plehn (Heidelberg University) Peter Sorrenson (Heidelberg University) Lorenz Vogel (Heidelberg University)

Presentation materials