·
BumpNet — Architecture for Bump Hunting
A network that reads an invariant-mass histogram and returns the statistical significance of a bump in every bin. Each part of the architecture answers one feature of the bump-hunt problem.
Key result
4.5σ
on the Higgs discovery histogram
Applied to the diphoton mass spectrum from the ATLAS Higgs discovery paper, BumpNet predicts a maximum significance of 4.5σ at the right mass, against 4.2σ from the likelihood-ratio test, with no retraining. Its per-bin values are unbiased, so they can be read as significances.
Resources