PaperlayerView exact revision on arXiv

Source-linked AI summary

Adversarially-trained autoencoders for robust unsupervised new physics searches

Andrew Blance, Michael Spannowsky, Philip Waite

arXiv:1905.10384v2hep-ph

TL;DR

Direct-data anomaly detection avoids theoretical uncertainties but still faces experimental smearing, while training on clean signal and background samples is difficult for rare or unknown processes. The paper combines an autoencoder with adversarial training to reduce sensitivity to smearing, and reports nearly identical ROC curves across smearing directions, including with signal-contaminated backgrounds.

  • Problem

    Data-driven anomaly detection avoids theoretical uncertainties but remains sensitive to experimental uncertainties in measured inputs, while pure training samples are difficult to obtain for rare or unknown signals.

    Source: §1 · p. 3

  • Method

    The paper adversarially trains an autoencoder on systematically smeared Monte Carlo events to reconstruct inputs without using smearing information.

    Source: §1 · p. 3, §4.1 · p. 8

  • Results

    The adversarial autoencoder's classification result is independent of smearing across the ROC curve and extends to backgrounds with a large signal admixture.

    Source: §4.1 · p. 11, §5 · p. 14

  • Takeaways & Limitations

    The method offers a robust unsupervised approach to resonance searches that could be applied analogously to experimental data using systematically smeared labelled datasets.

    Source: §5 · p. 14

    The study enlarges its smearing envelope by a factor of three as a conservative test of systematic uncertainty.

    Source: §2 · p. 5

Abstract

from arXiv · show

Machine learning techniques in particle physics are most powerful when they are trained directly on data, to avoid sensitivity to theoretical uncertainties or an underlying bias on the expected signal. To be able to train on data in searches for new physics, anomaly detection methods are imperative, which can be realised by an autoencoder acting as an unsupervised classifier. The last source of uncertainties affecting the classifier are then experimental uncertainties in the reconstruction of the final-state objects. To mitigate their effect on the classifier and to allow for a realistic assessment of the method, we propose to combine the autoencoder with an adversarial neural network to remove its sensitivity to the smearing of the final-state objects. We quantify its effect and show that one can achieve a robust anomaly detection in resonance-induced $t\bar{t}$ final states.

1 Introduction

Supervised searches can inherit theoretical uncertainties from simulated training samples, while training directly on data is difficult when pure signal and background samples are unavailable. The paper motivates an autoencoder-based anomaly detector and adversarial training to reduce sensitivity to experimental smearing.

1 Introduction

  • Supervised classifiers trained on simulated signal and background can inherit theoretical uncertainties, especially in exclusive phase-space regions and observables that may not be infrared-safe.The paper notes that such regions may be poorly understood perturbatively.Source: §1 · p. 2
  • Training on actual data avoids theoretical uncertainties, but pure, well-separated samples are difficult to obtain for rare or unknown signals.The proposed alternative is to learn background kinematics and flag events that do not share those features.Source: §1 · p. 3
  • Experimental and systematic uncertainties in measured inputs remain a problem for data-driven anomaly detection, particularly for complex top final states with many jets, leptons, and missing transverse energy.The paper says detector effects can make uncertainties on reconstructed observables large.Source: §1 · pp. 3–4
  • An autoencoder learns compressed background features through reconstruction, and adversarial training is proposed to reduce its sensitivity to systematic smearing.The study trains it on systematically smeared Monte Carlo pseudo-data so it can reconstruct events without using smearing information.Source: §1 · p. 3

2 Analysis setup and smearing procedure

The study uses semileptonic top-pair events and applies systematic smearing to background observables. The smearing can make signal-background discrimination either easier or harder, depending on its direction.

2 Analysis setup and smearing procedure

  • The benchmark uses 14 TeV semileptonic top-pair background and a 2 TeV Z′ signal, with a generator-level top-quark pT cut above 500 GeV.The Z′ width is 89.6 GeV.Source: §2 · p. 4
  • The analysis uses 21 observables from two b-jets, two light jets, an isolated lepton, and missing energy, smearing jets and missing energy in upward and downward directions.The smearing envelope is enlarged by a factor of three as a conservative test.Source: §2 · pp. 4–5
  • Smearing can make signal-background discrimination easier or harder depending on its direction.Figure 1 compares smeared-background distributions with unsmeared background and signal distributions.Source: Figure 1 · p. 5, §2 · p. 6
  • After selection, the study uses 100,000 events in each of three background samples and one unsmeared signal sample, retaining 20% for testing.The background samples are unsmeared, smeared upward, and smeared downward.Source: §2 · p. 5

3 Decorrelated jet smearing with supervised adversarial classifier

A supervised classifier's performance varies strongly with background smearing direction. Adding an adversarial network nearly removes that dependence by making the classifier's output less informative about the smearing class.

3 Decorrelated jet smearing with supervised adversarial classifier

  • The baseline classifier's performance depends strongly on smearing direction, consistent with greater signal-background overlap for upward-smeared background.The classifier is evaluated against the central signal sample for each background smearing set.Source: §3 · pp. 6–7
  • The adversary predicts which of three smearing classes produced a background event, while the classifier is penalized when the adversary performs better.This opposing objective aims to make classification predictions without using smearing information.Source: §3 · p. 7
  • The total-loss weight α must balance decorrelation against classification performance; a value that is too large severely harms classifier performance.The study reports α = 100 works well for its setup and finds alternating network updates more stable.Source: §3 · pp. 7–8
  • The adversarial classifier's ROC curves become very close across smearing directions as its outputs become nearly independent of background smearing.Figure 3 shows the adversarial classifier outputs and ROC curves.Source: Figure 3 · p. 8, §3 · p. 8

4 Extension to unsupervised autoencoder

The adversarial autoencoder uses reconstruction loss to make background-only anomaly detection insensitive to object smearing. Its performance remains robust to smaller smearing, signal contamination up to unrealistically large fractions, and tested resonance variations.

4.1 Adversarial autoencoder

  • The autoencoder uses three sigmoid hidden layers of 10, 3, and 10 nodes, followed by a linear output matching 21 observables, and minimizes mean squared input–output error.Its loss is supplied as an input to an adversary with two 20-node ReLU hidden layers and three softmax outputs; training alternates autoencoder and adversary updates.Source: §4.1 · pp. 8, 10
  • Without the adversary, upward-smeared backgrounds are more often mislabelled as signal, and the autoencoder performs worse overall than the dedicated supervised classifier.The autoencoder learns background kinematics only and is not optimised for strong signal-background classification.Source: §4.1 · p. 9, Figure 4 · p. 9
  • The adversarial autoencoder shapes background-loss distributions to be independent of smearing direction, making the three signal-versus-background ROC curves almost identical.The adversary receives autoencoder loss and learns to identify the smearing direction; the combined objective penalises losses that reveal that direction.Source: §4.1 · pp. 10–12, Figure 5 · p. 10
  • When tested with less smearing than used in training, the autoencoder remains insensitive to smearing and its AUC increases to a level similar to training and testing at that smaller smearing.The training used background data smeared by a much larger amount than the smaller-smearing test samples.Source: §4.1 · p. 11

4.2 Corrupted autoencoder and application to other new physics models

  • Training on backgrounds contaminated with signal has little effect on overall performance until the signal fraction becomes unrealistically large.The AUC band represents the difference between the upper and lower AUC scores, indicating how well the adversary desensitises the autoencoder to smearing.Source: §4.2 · p. 12, Figure 7 · p. 13
  • Cross-section limits are insensitive to resonance quantum numbers and comparable to ATLAS limits, while the adversary remains able to desensitise the autoencoder for the tested signals.The comparison is qualified: this study reports cross sections after event-selection and reconstruction cuts at 14 TeV, whereas ATLAS reports an inclusive cross section for a specific Z′ model at 13 TeV.Source: §4.2 · p. 13, Table 1 · p. 14

5 Conclusions

The paper proposes an adversarially trained autoencoder for unsupervised anomaly detection that is insensitive to smearing of reconstructed final-state objects. In semileptonic t̄t resonance searches, it reports robustness across the ROC curve and training with signal-contaminated backgrounds.

5 Conclusions

  • Even direct training on experimental data would leave residual uncertainties from imperfect reconstruction of final-state objects.The paper identifies these uncertainties as a remaining concern even when theoretical uncertainties and theoretical signal bias are avoided.Source: §5 · p. 14
  • The supervised classifier performs significantly better than the unsupervised approach, while the autoencoder still shows strong aptitude for separating signal from background.Reconstruction uncertainties substantially affect the evaluated performance of both classifiers.Source: §5 · p. 14
  • The adversarial autoencoder’s classification result is independent of reconstructed-object smearing across the ROC curve and extends to backgrounds with large signal admixtures.The authors propose that the procedure could be applied to experimental data using labelled datasets that have been systematically smeared.Source: §5 · p. 14

Related papers

Loading 1905.10384v2…