Transfer learning with LundNet

In this work we explore the use of transfer learning in the context of the study of the structure of jets in hadron collisions. We adopt the tool LundNet, a graph neural network for jet tagging, and we investigate the performance of the tagger in different transfer learning scenarios in comparison to that of fully trained models. We find that the transfer of weights from early layers of a different pretrained tagger allows one to train new taggers with significantly reduced computational resources and time. We study in detail the performance of the resulting tagger, which nearly reproduces that of a model trained from scratch. This offers a promising avenue to facilitate the use of such tools in the context of collider physics experiments.