SciPost Submission Page

Quark-Gluon Tagging: Machine Learning meets Reality

by Gregor Kasieczka, Nicholas Kiefer, Tilman Plehn, Jennifer M. Thompson

Submission summary

As Contributors: Jennifer Thompson
Arxiv Link:
Date submitted: 2019-03-04
Submitted by: Thompson, Jennifer
Submitted to: SciPost Physics
Domain(s): Theoretical
Subject area: High-Energy Physics - Phenomenology


Distinguishing quarks from gluons based on low-level detector output is one of the most challenging applications of multi-variate and machine learning techniques at the LHC. We first show the performance of different methods, including a new LoLa tagger, without and after considering detector effects. We then discuss two benchmark applications, mono-jet searches with a gluon-rich signal and di-jet resonances with a quark-rich signal. In both cases an immediate benefit compared to the standard event-level analysis exists.

Current status:
Editor-in-charge assigned

Submission & Refereeing History

Submission 1812.09223v1 on 4 March 2019

Reports on this Submission

Show/hide Reports view

Anonymous Report 1 on 2019-3-15 Invited Report


1-Thorough machine-learning analysis of the problem of quark/gluon discrimination applied to two search regions

2-Analysis of benchmark, standard observables which to compare with the machine learning results


1-Introduction and motivation of this work conflates and confuses many different aspects of theoretical studies, such as the importance and relevance of IRC safety

2-While the comparison to standard observables is good, the corresponding re-interpretation and combination to represent the "best" high-level observable tagger is misleading, especially when comparing to LoLa

3-I think the title of the article is misleading: "Quark-Gluon Tagging: Machine Learning meets Reality" would, to most people I think, imply that machine learning is being applied to experimental data, i.e., "reality". However all analyses in the paper are on Monte Carlo simulation.


The article "Quark-Gluon Tagging: Machine Learning meets Reality" applies machine learning methods, specifically the authors' LoLa algorithm, to the problem of quark versus gluon tagging in the context of two searches. I think the ultimate results of the paper are interesting and topical, but the introduction and comparison to high-level observables is a bit confusing and uses imprecise language. I would request the authors make minor changes to the article before I recommend it for publication.

Requested changes

1-First, the title should be changed. No analysis of experimental data is performed, therefore the authors are not working with "reality". Conflating simulation with data is ultimately detrimental to experimental science.

2-The authors cite their reference [26] several times throughout the paper when discussing theoretical issues of quark versus gluon discrimination. These citations include in the second and third paragraphs of the introduction and immediately before section 2.2. I am not sure why this reference is the catch-all for theory issues with quark/gluon tagging, as reference [26] consists of two papers published in 2018, and many issues were identified years or even decades earlier. I urge the authors to provide more representative references in place of [26]. For example, the first reference of [26] mentions the serious theoretical challenges of quark/gluon discrimination, but this was presented as a Les Houches report in their reference [27]. Further, the second reference of [26] mentions infrared and colliear safety, but this issue was known decades ago.

3-I'm somewhat confused by the setup in the first paragraph of the introduction regarding "kinematic observables". The authors write that there is a change from measuring kinematic observables on jets to measuring the particle four-vectors directly with machine learning. While I understand what the authors are trying to say here, a particle's momentum four-vector is the fundamental kinematic observable. The authors should clarify these statements and more precisely and distinguish between what they mean by "kinematic observable" and what they propose in this paper.

4-At the beginning of section 2, the authors make a couple of imprecise statements. In the first sentence of section 2, they write that "quarks and gluons are poorly defined in perturbative QCD". More precisely, quark and gluon flavor jets are ambiguous in perturbative QCD, and require some definition as there is no preferred definition. In the second paragraph of section 2, the authors write that "pile up could be dealt with by using standard techniques." I understand that this article is likely only to be read by other subject experts, but the authors could provide a few references to some "standard techniques" for context for researchers outside of the field. The authors should make these changes.

5-Finally, the authors should be careful with the interpretation and implications of their 6 variable BDT that is compared to LoLa. The authors demonstrate that each of the 6 observables are individually good quark versus gluon discriminants, but this in no way means that their combination is an "optimal" discriminant. It could be that the information in jets that they use for discrimination is identical, so when combined in a BDT would not improve performance. One is only guaranteed to have an optimal tagger if the observables feed to the machine form some complete basis on phase space. Individual particle four-vectors of course do this, which is why LoLa performs so well. However, the authors should add caveats in their construction and comparison to the 6 observable BDT. Locations in the draft where qualification should be added include the discussion on page 5 and page 8.

  • validity: good
  • significance: good
  • originality: good
  • clarity: ok
  • formatting: excellent
  • grammar: good

Login to report or comment