Lorentz Boost Networks: autonomous physics-inspired feature engineering

Abstract

We present a two-stage neural network architecture that enables a fully\nautonomous and comprehensive characterization of collision events by\nexclusively exploiting the four momenta of final-state particles. We refer to\nthe first stage of the architecture as Lorentz Boost Network (LBN). The LBN\nallows the creation of particle combinations representing rest frames. The LBN\nalso enables the formation of further composite particles, which are then\ntransformed into said rest frames by Lorentz transformation. The properties of\nthe composite, transformed particles are compiled in the form of characteristic\nvariables that serve as input for a subsequent network. This second network has\nto be configured for a specific analysis task such as the separation of signal\nand background events. Using the example of the classification of ttH and ttbb\nevents, we compare the separation power of the LBN approach with that of\ndomain-unspecific deep neural networks (DNN). We observe leading performance\nwith the LBN, even though we provide the DNNs with extensive additional input\ninformation beyond the particle four momenta. Furthermore, we demonstrate that\nthe LBN forms physically meaningful particle combinations and autonomously\ngenerates suitable characteristic variables.\n