Source-linked AI summary
Decorrelated Jet Substructure Tagging using Adversarial Neural Networks
Chase Shimmin, Peter Sadowski, Pierre Baldi, Edison Weik, Daniel Whiteson, Edward Goul, Andreas Søgaard
TL;DR
Powerful multivariate jet taggers can sculpt the jet-mass background, complicating background modeling in resonance searches. The paper uses adversarial training to balance classification with mass decorrelation and finds improved discovery power at high signal efficiency, with optimal significance around 90% efficiency in the discussed setting.
Problem
Extending decorrelation to powerful multivariate classifiers is challenging because their nonlinear responses require increasingly complex corrections.
Source: §I · p. 2Method
The paper trains a jet classifier alongside an adversary that infers jet mass from its response, and extends the approach to classifiers parameterized by signal mass.
Source: §I · p. 2, §VIII · p. 10Results
The adversarial classifier improves discovery power at high signal efficiency, with optimal significance around 90% efficiency when background-rate uncertainty makes sideband constraints valuable.
Source: §VI · p. 7, §VIII · p. 9Takeaways & Limitations
Reducing background morphing can preserve sideband constraints that help constrain background rates in searches.
Source: §VIII · p. 9
Abstract
from arXiv · showhide
We describe a strategy for constructing a neural network jet substructure tagger which powerfully discriminates boosted decay signals while remaining largely uncorrelated with the jet mass. This reduces the impact of systematic uncertainties in background modeling while enhancing signal purity, resulting in improved discovery significance relative to existing taggers. The network is trained using an adversarial strategy, resulting in a tagger that learns to balance classification accuracy with decorrelation. As a benchmark scenario, we consider the case where large-radius jets originating from a boosted resonance decay are discriminated from a background of nonresonant quark and gluon jets. We show that in the presence of systematic uncertainties on the background rate, our adversarially-trained, decorrelated tagger considerably outperforms a conventionally trained neural network, despite having a slightly worse signal-background separation power. We generalize the adversarial training technique to include a parametric dependence on the signal hypothesis, training a single network that provides optimized, interpolatable decorrelated jet tagging across a continuous range of hypothetical resonance masses, after training on discrete choices of the signal mass.
I. INTRODUCTION
Jet tagging separates boosted hadronic decays from quark and gluon jets, but powerful taggers can sculpt the jet-mass background and complicate its modeling. The paper builds mass decorrelation into adversarial training and extends it to classifiers parameterized by the signal hypothesis.
I. INTRODUCTION
- Jet tagging distinguishes jets from single quarks or gluons from jets containing two or three hadronic decay products, a key component of new-physics searches [1] [2] [3] [4] [5].Boosted decay products can overlap in the calorimeter and form a single large jet.Source: §I · p. 1
- Powerful multivariate classifiers can outperform individual tagging variables, but extending DDT-style decorrelation to their nonlinear responses requires increasingly complex corrections.Mass-correlated tagging quantities can distort background shapes and introduce systematic uncertainties when analyses interpolate the background under a signal peak from sidebands [9].Source: §I · pp. 1–2
- The paper incorporates mass decorrelation into adversarial training: a classifier tags jets while an adversary tries to infer jet mass from the classifier response.The jointly trained networks balance classification accuracy against dependence of the classifier response on the selected input.Source: §I · p. 2
- A scan over resonance masses motivates parameterization because optimal classifiers can differ by hypothesis, while simulations are usually available for only a small number of hypotheses.Producing signal simulations is computationally expensive.Source: §I · p. 2
- Parameterized classifiers can interpolate across signal hypotheses not used in training, allowing sensitivity evaluation without generating simulations at those points [17] [18].The paper proposes parameterizing the adversarial classifier by signal mass and reports decorrelation across the trained range.Source: §I · p. 2
II. BENCHMARK DATA
The benchmark uses simulated boosted Z′ decays recoiling against a photon, with γ+jets as the primary background. Large-radius jets are selected and characterized with kinematic and substructure variables, alongside DDT-style τ21 comparators.
II. BENCHMARK DATA
- The benchmark signal is a hadronically decaying Z′ recoiling against an initial-state photon, compared with γ+jets background.The photon channel is chosen for its simpler event topology.Source: §II · p. 2, Figure 1 · p. 2
- Signal and background samples are simulated, with an average of 15 overlaid pile-up interactions per event.The signal uses MadGraph, Pythia, and Delphes; the primary γ+jets background is generated with Sherpa.Source: §II · p. 2
- The selection reconstructs a single large-radius jet with R = 1.0 and requires photon and jet transverse momenta above 150 GeV.When multiple large-radius jets are present, the highest-pT jet is selected.Source: §II · p. 3
- The neural networks use eleven variables, including jet pseudorapidity, azimuthal angle, transverse momentum and invariant mass, C2 and D2, plus photon energy, pseudorapidity, azimuthal angle and transverse momentum.Source: §II · p. 3
- The DDT-style τ21 correction uses a second-order adjustment because a linear correction did not remove the mass dependence in this setup.The resulting τ′′21 shows reasonable independence from jet mass.Source: §II · p. 3
III. NEURAL NETWORKS
The tagger pairs a jet classifier with an adversary trained to infer jet mass from the classifier response. Their competing losses are optimized together, with the adversary trained on mass bins and the classifier using 11 input features.
III. NEURAL NETWORKS
- The classifier and adversary are trained concurrently using a tagger loss that trades classification performance against adversary loss.The positive weight λ controls the tradeoff; λ = 100 was a good value for this task.Source: §III · p. 3
- The adversary predicts one of 10 jet-mass deciles from a single input, while the classifier uses 11 features, three 300-node hidden layers, and a logistic output.The classifier uses hyperbolic tangent activations and binomial cross-entropy; the adversary uses a softmax output and multi-class cross-entropy.Source: §III · p. 4
- The adversary receives 100 initial updates while the classifier is fixed and is trained with learning rate 1.0, versus 10^-3 for the tagger objective.Training, validation, and test splits are 80%, 10%, and 10%, respectively.Source: §III · p. 4
- The adversary loss assigns signal events zero weight, so the adversary learns from background events only.Training uses mini-batches of 100 examples from each class.Source: §III · p. 4
IV. PERFORMANCE
The section compares neural-network and τ21-based taggers on signal efficiency, background rejection, and their effect on jet mass. The adversarial network improves background rejection over τ21 while reducing—but not eliminating—mass dependence.
IV. PERFORMANCE
- The performance comparison measures signal efficiency and background rejection across thresholds for neural-network and τ21-based discriminants.The candidates include the conventional and adversarial networks, τ21, and two DDT-modified τ21 variables.Source: §IV · p. 4, Figure 4 · p. 5
- The adversarial network’s background response is largely independent of jet mass, but the dependence is diminished rather than eliminated.This residual dependence reflects the trade-off between classification power and decorrelation.Source: §IV · p. 4
- The adversarial network’s response shows small pT-dependent effects, while training against log(m/pT) produced a flatter log(m/pT) response but greater jet-mass distortion.Two-dimensional decorrelation is left for future study because the pT dependence is not severe in this setting.Source: §IV · pp. 4–5
- Both neural networks improve background rejection over τ21 at similar signal efficiencies, while the adversarial network gives up some discrimination relative to the conventional network.The comparison uses signal samples with mZ′ = 100 GeV; the mass dependence of DDT-variable performance also varies with signal mass.Source: §IV · p. 4, Figure 4 · p. 5
V. STATISTICAL INTERPRETATION
This section evaluates taggers not only by discrimination but also by the distortion they induce in the background jet-mass spectrum. It uses a simplified likelihood test with background-rate uncertainty to represent the effect of that distortion on discovery significance.
V. STATISTICAL INTERPRETATION
- Figure 6 compares background jet-mass distributions after progressively stricter selections, with signal efficiencies from 50% to 100%.The displayed strategies include conventional and adversarial neural networks, τ21, and τ′′21.Source: Figure 6 · p. 6
- Background estimates rely on assumed functional forms constrained by sidebands, and tagger-induced distortion can weaken control-region checks and make the model harder to constrain.The simplified test likely underestimates the impact of more realistic systematics, though it is used to illustrate their effect on sensitivity.Source: §V · pp. 5–6
- The test applies a discriminator threshold and performs a binned likelihood fit to reconstructed large-radius jet masses using simulated signal and background templates.Expected discovery significance is calculated with a profile likelihood ratio and the CLs technique, marginalizing over the unknown background rate.Source: §V · p. 6
- A background-rate uncertainty penalizes taggers whose selections distort the mass spectrum to resemble a signal, because sidebands then constrain the rate less effectively.The study uses relative background-rate uncertainties of 5% and 50% for its small- and large-uncertainty cases.Source: §V · p. 6
VI. RESULTS
The discovery-significance benefit of discriminator thresholds depends on background-rate uncertainty: mass distortion can undermine the strongest classifier, while the adversarial network retains useful discrimination at high signal efficiency.
VI. RESULTS
- With 5% relative background uncertainty, tighter discriminator thresholds improve significance despite reducing signal efficiency, through stronger background suppression.Significance still improves at 50% signal efficiency even when the background is sculpted to resemble the signal; with a well-known background shape, reduced sideband constraints are not detrimental.Source: §VI · pp. 6–7
- With large background-rate uncertainty, the traditional classifier weakens discovery significance through mass distortion, while the adversarial network improves discovery power around 90% signal efficiency.The qualitative pattern persists for other signal-to-background ratios.Source: §VI · p. 7
- Figures 7 and 8 show jet-mass distributions after selections at 90% and 50% signal efficiency, respectively, with 50% background uncertainty.They compare the NN classifier, adversarial network, τ21, and τ ′′ 21.Source: Figure 7 · p. 7, Figure 8 · p. 7
- Figure 9 compares significance versus signal efficiency for discriminator thresholds under small and large background uncertainties, with Nsig = 100 and Nbg = 1000.The displayed significance curves are the main comparison across uncertainty scenarios.Source: Figure 9 · p. 7
VII. PARAMETERIZED NEURAL NETWORKS
A single network parameterized by the hypothesized Z′ mass handles related classification tasks, with an adversary parameterized as well to decorrelate predictions from jet mass at each hypothesis. The adversarial network gives up some separation relative to a traditional network but improves discovery significance.
VII. PARAMETERIZED NEURAL NETWORKS
- A single classifier takes mZ′ as an input to solve classification tasks across hypotheses, rather than training a separate network for every mass.Parameterized networks can interpolate to untrained values when the function is smooth; Ref. [18] also demonstrates successful interpolation.Source: §VII · p. 8
- Figure 10 shows the classifier using eleven variables plus mZ′, while the adversary uses the classifier output and mZ′ to predict invariant mass.The classifier output is therefore a function of the hypothesized mass.Source: Figure 10 · p. 8
- Parameterizing the adversary by Z′ mass makes background classifier predictions mostly mass-independent conditional on each hypothesis; without it, only marginalized predictions are independent.Figure 11 profiles background response against jet mass for the adversarial and traditional classifiers.Source: §VII · p. 8
- At all tested signal-mass hypotheses, the resulting classifier outperforms each individual input feature (τ21, τ′21, or τ″21), while the traditional neural network achieves the best separation.The adversarially trained network’s lack of background distortion preserves the ability to distinguish background and signal mass distributions, leading to improved discovery significance. The significance test fits a binned likelihood on jet mass after thresholding the discriminator; the traditional NN’s improved separation does not improve significance.Source: §VII · p. 8
VIII. DISCUSSION
The adversarial strategy combines multivariable discrimination with reduced correlation to jet mass, helping avoid background shapes that resemble the signal. Its effectiveness is a tradeoff, and residual dependencies remain; the strategy is also extended to parameterized networks.
VIII. DISCUSSION
- Adversarial training combines information from multiple inputs while decorrelating the classifier from jet mass, reducing the tendency for background to become degenerate with the signal.Avoiding that degeneracy matters when the background cannot be predicted reliably and its rate must be constrained in sidebands.Source: §VIII · pp. 8–9
- At roughly 90% signal efficiency, the adversarial classifier reaches optimal significance with limited background morphing, although sufficiently tight cuts cause significant morphing.The comparison is with τ21-based discriminants.Source: §VIII · p. 9
- Decorrelation remains incomplete: residual mass structure still sculpts background, and residual pT dependence could be addressed with a more sophisticated multi-variable adversary.These improvements are left for future work.Source: §VIII · p. 10
- The strategy is extended to parameterized classifiers for tagging specific signal hypotheses across theoretical parameter space.The supplied passage states the extension but does not provide its detailed performance here.Source: §VIII · p. 10