Metalearning and data augmentation for mass-generalized jet taggers

Abstract

Deep neural networks trained for jet tagging are typically specific to a narrow range of transverse momenta or jet masses. Given the large phase space that the LHC is able to probe, the potential benefit of classifiers that are effective over a wide range of masses or transverse momenta is significant. In this work we benchmark the performance of a number of methods for achieving accurate classification at masses distant from those used in training, with a focus on algorithms that leverage metalearning. We study the discrimination of jets from boosted ${Z}^{\ensuremath{'}}$ bosons against a QCD background. We find that a simple data augmentation strategy that standardizes the angular scale of jets with different masses is sufficient to produce strong generalization. The metalearning algorithms provide only a small improvement in generalization when combined with this augmentation. We also comment on the relationship between mass generalization and mass decorrelation, demonstrating that those models which generalize better than the baseline also sculpt the background to a smaller degree.