Tilman Plehn (Heidelberg)

Modern Neural Networks with Calibrated Learned Uncertainties

Abstract: Machine Learning is revolutionizing our lives and our research. However, ML-applications for instance in particle physics come with extra requirements on precision, uncertainties, and control. I will show how we can understand neural networks and their training from a statistical perspective, including learned uncertainties. A key question is if these learned uncertainties are calibrated with respect to statistical and systematic shortcomings. I will present a comprehensive study for loop-induced transition amplitudes in LHC physics and discuss the challenges for generative networks with learned uncertainties. Finally, I will show how transformers can be used to realize accurate and data-efficient Lorentz-covariant architectures and to extrapolate QCD patterns to phase spaces not seen during training.