Direct optimisation of the discovery significance when training neural networks to search for new physics in particle colliders

Adam Elwood, Dirk Krücker

We introduce two new loss functions designed to directly optimise the statistical significance of the expected number of signal events when training neural networks to classify eve…