Symmetries, safety, and self-supervision

Collider searches face the challenge of defining a representation high-dimensional data such that physical symmetries are manifest, discriminating features retained, and choice is new-physics agnostic. We introduce JetCLR to solve mapping from low-level optimized observables through self-supervised contrastive learning. As an example, we construct for top QCD jets using permutation-invariant transformer-encoder network visualize its symmetry properties. compare with alternative representations linear classifier tests find it work quite well.