Neural networks in LHC physics must be accurate, reliable, and controlled. We first show how activation functions can be systematically tested with KANs. For reliability and control, we learn an uncertainty together with the target amplitude over phase space. While systematic uncertainties can be described by a heteroscedastic loss, a comprehensive learned uncertainty requires Bayesian networks or repulsive ensembles. We compute pull distributions to show that the learned uncertainties are calibrated correctly.
Ankita Budhraja, Juraj Klaric, Johannes Michel, Maria Laura Piscopo