Generative and Invertible Networks for the LHCTilman Plehn, Heidelberg University, 12:00 EDT
Abstract: LHC physics is a unique field in the sense that we compare vast and highly complex data sets with precise first-principles predictions. These predictions usually rely on Monte Carlo simulations. I will show how generative neural networks can supplement these simulations and discuss conceptional advantages of this method. I will then explain how generative networks can invert event simulations. Flow-based invertible networks allow us to invert or unfold individual detector simulations of QCD parton showers in a mathemacially consistent manner. That means that they predict calibrated probability distributions in parton-level phase space for individual observed events. Finally, I will illustrate how the same networks can infer the structure of QCD splittings forming jets.