Lorentz Boost Networks: autonomous physics-inspired feature engineering

Abstract

A: We present a two-stage neural network architecture that enables a fully autonomous and comprehensive characterization of collision events by exclusively exploiting the four momenta of final-state particles. We refer to the first stage of the architecture as Lorentz Boost Network (LBN). The LBN allows the creation of particle combinations representing rest frames. The LBN also enables the formation of further composite particles, which are then transformed into said rest frames by Lorentz transformation. The properties of the composite, transformed particles are compiled in the form of characteristic variables that serve as input for a subsequent network. This second network has to be configured for a specific analysis task such as the separation of signal and background events. Using the example of the classification of ttH and tt+bb events, we compare the separation power of the LBN approach with that of domain-unspecific deep neural networks (DNN). We observe leading performance with the LBN, even though we provide the DNNs with extensive additional input information beyond the particle four momenta. Furthermore, we demonstrate that the LBN forms physically meaningful particle combinations and autonomously generates suitable characteristic variables.

Jet tagging via particle clouds

“…The prevailing choice is a sequence, where particles are sorted in a specific way (e.g., with decreasing transverse momentum) and organized into a 1D list. Using particle sequences as inputs, jet tagging tasks have been tackled with recurrent neural networks (RNNs) [36][37][38][39]44], 1D CNNs [40][41][42][43] and physics-oriented neural networks [45][46][47]. Another interesting choice is a binary tree, which is well motivated from the QCD theory perspective.…”

Invisible Higgs search through vector boson fusion: a deep learning approach

“…This is evident from the fact that convolutional neural networks perform best with data structures that have an underly-ing Euclidean structure, while recurrent networks work best with sequential data structures. In the context of classifying boosted heavy particles like W , Higgs, top quark or heavy scalars decaying to large-radius jets from QCD background, a lot of efforts [24,25,[27][28][29]111] went into representing the data like an image in the (η, φ) plane to use convolutional layers for feature extraction, while some others [112,113], use physics-motivated architectures. Convolutional architectures work in these cases because the differences between the signal jet and the background (QCD) follows a Euclidean structure.…”

Report on 1902.09914v2

“…The Lorentz Boost Network (LBN) is designed to autonomously extract a comprehensive set of physics-motivated features given only low-level variables in the form of particle fourvectors [19]. These engineered features can be utilized in a subsequent neural network to solve a specific physics task.…”