92
3 Jet Substructure at the LHC
even though no b tagging information is included. Further variants of the DeepAK8
and ImageTop taggers are mass-decorrelated (MD) taggers. In the case of DeepAK8,
an adversarial training is used to achieve the decorrelation, whereas for ImageTop
the decorrelated version is obtained by training on samples with similar shaped mass
distributions. The MD taggers have the advantage of an unchanged jet mass distribution when selecting jets based on the algorithm’s output. This can be crucial in
analyses where the non-resonant background contribution is predicted from sideband regions. While the MD procedure results in a decrease in performance for the
DeepAK8 algorithm, the performance of ImageTop is very similar between the nominal and MD versions. Efficiency and misidentification rate measurements have been
carried out in data [525, 526], verifying the ML tagger performances from studies
in simulation. The derived correction factors can be used to adjust the efficiencies
in simulation. The corrections are typically smaller than 10% with uncertainties of
the same size. These measurements allow for the use of these ML taggers in future
analyses, promising a large gain in sensitivity relative to traditional approaches.
3 Jet Substructure at the LHC
even though no b tagging information is included. Further variants of the DeepAK8
and ImageTop taggers are mass-decorrelated (MD) taggers. In the case of DeepAK8,
an adversarial training is used to achieve the decorrelation, whereas for ImageTop
the decorrelated version is obtained by training on samples with similar shaped mass
distributions. The MD taggers have the advantage of an unchanged jet mass distribution when selecting jets based on the algorithm’s output. This can be crucial in
analyses where the non-resonant background contribution is predicted from sideband regions. While the MD procedure results in a decrease in performance for the
DeepAK8 algorithm, the performance of ImageTop is very similar between the nominal and MD versions. Efficiency and misidentification rate measurements have been
carried out in data [525, 526], verifying the ML tagger performances from studies
in simulation. The derived correction factors can be used to adjust the efficiencies
in simulation. The corrections are typically smaller than 10% with uncertainties of
the same size. These measurements allow for the use of these ML taggers in future
analyses, promising a large gain in sensitivity relative to traditional approaches.