Abstract
We compare the performance of a convolutional neural network (CNN) trained on jet images with dense neural networks (DNNs) trained on nn-subjettiness variables to study the distinguishing power of these two separate techniques applied to top quark decays. We find that they perform almost identically and are highly correlated once jet mass information is included, which suggests they are accessing the same underlying information which can be intuitively understood as being contained in 4-, 5-, 6-, and 8-body kinematic phase spaces depending on the sample. This suggests both of these methods are highly useful for heavy object tagging and provides a tentative answer to the question of what the image network is actually learning.
Highlights
New particles created above the electroweak scale are often expected to decay into the most massive members of the Standard Model family, namely the W and Z bosons, the (BroutEnglert-)Higgs boson h and the top quark
From each sample we extract from the leading pT jet both a jet image and a set of N subjettiness variables as defined in Sec 2.2 and Sec 2.4 respectively, which serve as the raw pseudodata to be fed to the classifiers after further preprocessing
In general the performance of the two methods is very comparable both with and without jet mass information included, which suggests the image networks are probing very similar information as the n-subjettiness one. This is a non-trivial test of whether or not our image network accesses information which can not be considered safe from a modeling perspective: since it saturates at 4-body kinematic phase space information, we can be fairly certain it is learning features of the hard splittings and first few splittings in the parton shower rather than low-energy features which should not be trusted
Summary
New particles created above the electroweak scale are often expected to decay into the most massive members of the Standard Model family, namely the W and Z bosons, the (BroutEnglert-)Higgs boson h and the top quark. It is natural to consider the possibility of converting information from LHC events into images on which we can train a machine learning algorithm on in order to teach it the mapping from hadronic activity to various classifications such as heavy particle decay and ‘QCD background’ or light quark and gluon jets [10,11,12]. There have been initial studies of the sensitivity of these algorithms to the modeling uncertainties involved in the use of Monte Carlo event generators which (just to mention one issue among many) rely entirely on phenomenological models to hadronise the final state after the (well-understood) parton shower has increased the coloured particle multiplicity considerably [16] Such questions of the extent to which we might be teaching the machine spurious modeling details rather real physics become more pertinent as more advanced algorithms such as jet images squeeze more information out of the radiation patterns.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
