Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Contrastive Normalizing Flows for Uncertainty-Aware Parameter Estimation

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read By encoding signal and background with contrastive normalizing flows, this paper turns learned densities into classifier inputs and reports calibrated frequentist confidence intervals for the Higgs signal strength under nuisance-parameter…

desk verdict A promising hybrid pipeline whose central claim about CNF robustness is not yet supported in the small-shift regime, but the work is worth refereeing. read the letter →

arxiv 2505.08709 v1 pith:GWAAWQSE submitted 2025-05-13 physics.data-an cs.LGhep-exhep-ph

Ibrahim Elsharkawy , Yonatan Kahn This is my paper · ORCID
classification physics.data-ancs.LGhep-exhep-ph
keywords contrastivenormalizingflowsuncertainty-awareparameterestimationsystematicuncertaintiesdomainshiftlikelihoodratioapproximationfrequentistconfidenceintervalsHiggsMLUncertaintyChallengesignalstrength
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Estimating the fraction of signal events in a particle-physics dataset is hard when detector miscalibration deforms the data. This paper proposes a pipeline that first trains contrastive normalizing flows (CNFs) on unperturbed signal and background samples, then feeds their learned densities, along with the raw kinematics, into a neural classifier, and finally converts classifier scores into frequentist confidence intervals via a binned likelihood and confidence-belt construction. The paper shows analytically that the contrastive loss drives the learned density toward a truncated likelihood-ratio contour, and argues this density view is more stable under nuisance-induced shifts than a bare decision boundary. On the HiggsML Uncertainty Challenge, the method reports a score of 0.823, an average 68% confidence interval length of 0.438 at coverage 0.672, roughly half the interval size of a likelihood-based comparison at comparable coverage and about ten times faster. If these numbers hold, the approach offers a practical way to do uncertainty-aware signal-strength fits without modeling a high-dimensional parameter grid.

What carries the argument

The machinery is the contrastive normalizing flow (CNF): a normalizing flow trained with the usual log-likelihood on signal plus a contrastive term $c\log p_\theta(x_b)$ on background. In the infinite-data limit the optimum is a truncated likelihood-ratio density, $\max\{p_s(x)-c p_b(x),0\}$ up to normalization, and the boundary where it vanishes is the Bayes-optimal classification contour. The paper builds two such flows per jet category ($c=2$ and $c=0.5$), forms bounded discrimination functions $\Phi_{s,b}=p_{\theta}^{(s,b)}/(1+p_{\theta}^{(s,b)})$, appends them to the 28 kinematic features, and trains a two-headed DNN classifier on 1000 sub-datasets with different nuisance settings. At inference, classifier-score histograms are fit with a binned likelihood in the signal strength $\mu$ and the two most influential nuisance parameters, and the intervals come from inverting a confidence belt built over many pseudo-experiments.

What would settle it

Train the same DNN on the 28 raw features but drop the CNF-derived $\Phi$ inputs, and evaluate both pipelines on the same 10 by 100 HiggsML pseudo-experiments; if coverage and interval width are statistically indistinguishable, the CNF features are not the load-bearing part of the claimed stability. A second check is to replace the CNF densities with ordinary non-contrastive flow densities and see whether the contrast term, rather than density estimation, is what preserves coverage.

Watch

Extended reading notes

Core claim

The paper claims that the ratio of signal to background densities, approximated through CNF-derived discrimination functions, remains informative under the small nuisance-parameter distortions of the benchmark, so a DNN trained on augmented features can approximate the full likelihood ratio and yield calibrated intervals for the signal strength. The analytic result is that the CNF's learned density minimizes the contrastive loss to $p^{(s)}_{\theta^*}(x) \propto \max\{p_s(x)-c p_b(x),0\}$, a truncated likelihood-ratio contour, which the authors interpret as capturing the whole class distribution rather than only a boundary. Empirically, on 10 by 100 pseudo-experiments, CNF+DNN gives score 0.823, mean 68% interval 0.438, coverage 0.672, RMSE 0.191, in 5 minutes, while the likelihood-based method gives score 0.001, interval 0.998, coverage 0.701, in 41 minutes. The paper states the method selects 100 to 3000 signal events out of $10^6$ events with a 1-$\sigma$ band of 150 to 500 events.

Load-bearing premise

The method assumes CNFs trained only on unperturbed signal and background remain discriminative enough when test data are shifted by the benchmark's small nuisance deformations; the paper's own Gaussian toy shows the plain DNN beats the CNF-based classifier precisely in that small-shift regime.

Editorial extensions

If this is right

  • If the reported benchmark numbers are correct, signal-strength fits in the HiggsML setting can be made with 68% intervals of length about 0.44, roughly half the width of the likelihood-based comparison at similar coverage.
  • The 5-minute runtime, versus 41 minutes for the likelihood method, makes large pseudo-experiment coverage studies practical on modest GPU allocations.
  • Because only two nuisance parameters enter the binned likelihood, the method avoids explicit profiling over all six systematic parameters at test time.
  • The analytic contrastive optimum ties the hyperparameter $c$ to the truncation contour of the likelihood ratio, giving an interpretable handle for trading signal-richness against stability.
  • A calibrated pipeline of this kind could be applied to any rare-signal mixture where a simulator can generate labeled signal and background events with nuisance settings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: its own toy study shows the plain DNN is more accurate than the CNF-featured DNN at small nuisance deformations, while the CNF features win at large deformations; since the challenge's nuisance widths are small, this leaves open whether the benchmark gain comes from the CNF densities or from the DNN learning raw-feature shifts.
  • Beyond the paper: the CNF's generative side suggests a testable extension, namely using the contrastive flow to sample signal-like events that are maximally unlike background and using them as data augmentation, which could improve sensitivity in even rarer-signal regimes.
  • Beyond the paper: the method's calibration relies on a simulated confidence belt; a natural stress test is to clip or remove the two nuisance-parameter templates and measure how much coverage degrades, which would isolate where the stability claim actually lives.
  • Beyond the paper: because the pipeline is built on classifier score templates rather than a parametric likelihood, it may transfer to other signal-fraction problems such as rare decays or astroparticle stacks whenever labeled simulations are available, but that transfer is not demonstrated here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a parameter-estimation pipeline for signal-fraction inference in the presence of nuisance parameters. A contrastive normalizing flow (CNF), trained with a loss that penalizes background probability under the signal flow, is used to define discrimination functions Phi_s and Phi_b; these are appended to 28 raw kinematic features and fed to a two-headed DNN classifier. A binned likelihood and Neyman construction are then used to produce coverage-calibrated confidence intervals for the signal strength mu. The authors provide an analytic characterization of the CNF optimum (Eq. 10), validate the approach on a 10D Gaussian toy (Sec. 5), and report results on the HiggsML Uncertainty Challenge (Sec. 6.3, Table 2), claiming a score of 0.823, an average 1-sigma interval of 0.438, coverage 0.672, and roughly 10x faster runtime than a likelihood-based method. The central claim is that CNF features make classification robust to distribution distortions induced by systematic uncertainties.

Significance. If the central claim is established, the paper offers a practical recipe for uncertainty-aware parameter estimation with classifiers in the presence of nuisance parameters, an important problem in HEP and related fields. The contrastive-loss modification is simple and the analytic minimization in Eq. (10) is clean and instructive. The toy study is clearly described, and the HiggsML results are concrete and falsifiable (score, interval length, coverage, runtime). The method's significance, however, rests on whether the CNF-derived features are actually responsible for the reported robustness; currently that attribution is not demonstrated, and the paper's own toy evidence points the other way in the small-shift regime that characterizes the HiggsML benchmark. If the missing isolation is supplied, the results would be a useful contribution; without it, the headline claim is not supported.

major comments (4)
  1. [Sec. 5 and Sec. 6.3, Fig. 3, Table 2] The central claim that CNF features drive the benchmark performance is not tested by an ablation. In the only controlled comparison, the plain DNN (i) is the best classifier at small values of the nuisance parameter (Fig. 3, center), while the DNN with CNF features (iv) surpasses other classifiers only for large deformations (Fig. 3, left and right). The HiggsML nuisance priors in Table 3 have widths from 0.001 to 0.25, i.e., the small-shift regime, and the two baselines in Table 2 differ from the proposed method in both feature set and estimation procedure (the 'Baseline MLE for mu only' does not profile nuisances and reports coverage 0.431). Consequently, the 0.823 score cannot currently be attributed to the CNF features. I ask for an ablation that keeps the entire pipeline (raw 28 features, the same DNN training set, the same Neyman construction, and the same profiling of alpha_jes and alpha_tes) fixed and compares performance with and without the CNF-derived Phi_s and Phi_b.
  2. [Sec. 4.2 and Eq. (10)] The theoretical argument does not establish robustness under nuisance shifts. Eq. (10) characterizes the density that minimizes the contrastive loss on unperturbed data, but no result is given for how this density, or the associated discrimination functions Phi_s and Phi_b, behaves when ps and pb are deformed by the nuisance parameters. The 'holistic view' argument in Sec. 4.2 is heuristic. Since the paper's central robustness claim is explicitly about data distribution distortions, I ask for either a formal stability statement (e.g., a Lipschitz or bounded-perturbation condition on the learned density as a function of the nuisance parameters) or an empirical demonstration at the small-shift scales actually listed in Table 3.
  3. [Sec. 6.3, Table 2] The abstract's claim of 'top performance on the HiggsML Uncertainty Challenge' is not supported by an external benchmark comparison. Table 2 compares only three internally constructed methods, and no challenge leaderboard scores, other participant results, or published baselines are reported. Without such a comparison, 'top performance' is unsubstantiated; the paper should either report the method's official challenge rank/score or compare against the publicly available leaderboard and at least one independent baseline using the same scoring function.
  4. [Sec. 6.2] Several design choices are made using the target data without a described validation protocol. The removal of 0-jet events, the selection of alpha_jes and alpha_tes as the only profiled nuisance parameters, and the choice of CNF contrast values c=2 and c=0.5 are all justified as 'empirical' in Sec. 6.2. These choices affect the reported score and therefore the claimed robustness. The manuscript should either present a pre-specified or cross-validated selection procedure or show that the headline metrics are stable under reasonable variations of these choices.
minor comments (5)
  1. [Sec. 3, Eq. (3)] The notation p_s(x | Phi_s[x;{nu_i}]) is not well-defined because Phi_s is itself a function of x; the intended likelihood ratio should be written as a function of the augmented feature vector (x, Phi_s(x), Phi_b(x)), not as a conditional probability on Phi_s.
  2. [Sec. 3] The statement that 1000 nuisance-parameter choices correspond to '1000^{1/6} approx 3 choices of each nuisance parameter value' is misleading, since the 1000 sub-datasets are random draws from a 6D distribution, not a grid with three values per parameter. Please rephrase to describe the sampling scheme accurately.
  3. [Appendix A.1 and Appendix E.2] There are minor typographical errors: in Eq. (14) the scale function appears with an extra parenthesis, 'exp(s(ub))', and in Appendix E.2 the word 'susbet' should be 'subset'.
  4. [Sec. 6.3, Table 2] The 'Runtime (min)' column is ambiguous as to whether it includes CNF and DNN training time or only the evaluation/Neyman-construction time; please specify exactly what the 5.0 minutes and 41 minutes cover.
  5. [General] The manuscript contains no data or code availability statement. For a benchmark-driven paper, making the training and evaluation pipeline available would substantially improve reproducibility and the usefulness of the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CNF loss, analytic minimizer, and Neyman construction are each self-contained, with the headline benchmark result coming from an external challenge rather than from fitting the reported parameter to its own inputs.

full rationale

The derivation chain is not circular. Eq. (5) defines a contrastive loss, and Eq. (10) is obtained by direct functional minimization of that loss with a Lagrange multiplier; Fig. 1 compares the trained flow to this same analytic minimizer, which is a consistency check, not an imported prediction. The discrimination functions Phi_{s,b} in Eq. (3) are architectural choices used as classifier inputs, and the DNN is trained on labeled signal/background events plus these features; no signal-strength parameter is fitted to the training labels. The Neyman construction in Sec. 6.2 builds confidence intervals from simulated pseudo-experiments, and the reported coverage 0.672 in Table 2 is a calibration property of that construction, not a quantity derived from the target estimate. References to CNFs [18], likelihood-ratio classification [21], and the HiggsML benchmark [1] are external; there are no load-bearing self-citations. The real weaknesses are evidentiary rather than circular: Sec. 5 (Fig. 3, center panel) explicitly states that at small nuisance shifts the plain DNN performs best while the CNF-feature DNN wins only at larger shifts, yet the HiggsML nuisance priors in Table 3 are small; and Sec. 6.3 attributes the HiggsML improvement to CNF features without a clean ablation against a raw-feature DNN. The choices of c, jet categories, and profiled nuisances are also empirically tuned on the benchmark. These are concerns about generalization and attribution, not reductions of the central claim to its inputs, so they do not raise the circularity score.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's central claim rests on the contrastive density estimator (with its hand-chosen c and architecture), the assumption that unperturbed CNF features stay informative under small nuisance shifts, and a Neyman construction that profiles only two of six nuisances. No new physical entities are introduced. The main free parameters are the contrast strengths, the loss clamp, the size of the nuisance grid, the choice of profiled nuisances, and the jet selection, most of which were set empirically on the target benchmark.

free parameters (5)
  • Contrastive strength c = c = 2 and c = 0.5 for HiggsML flows
    Controls how much background is suppressed in the CNF density; larger c emphasizes signal-rich regions but discards signal density. The two values were chosen to balance nuisance robustness and signal coverage, without a formal selection procedure.
  • Loss clamping threshold k = -50
    Bounds the contrastive loss to prevent mode collapse during training; value taken empirically and consistent with Ref. [18].
  • Number of nuisance sub-datasets = 1000
    The DNN training set is composed of 1000 sub-datasets, each with a different random 6D nuisance vector; this is a practical sampling choice (about 3 values per dimension).
  • Profiled nuisance parameters = alpha_jes and alpha_tes
    The binned likelihood used for final estimation depends only on these two of the six nuisance parameters, which the authors report as empirically the most impactful on signal/background discrimination; the other four are only varied in the Neyman construction.
  • Jet category selection = 0-jet events excluded
    The analysis uses only 1-jet and 2-jet events; 0-jet events were found empirically detrimental to accurate mu estimation.
assumptions (5)
  • domain assumption A binary classifier trained on CNF features approximates the likelihood ratio in Eq. (3) in finite-data settings.
    Section 3 invokes the infinite-data/expressivity result of Cranmer et al. [21] and assumes it carries over to their finite training set with Phi features.
  • domain assumption CNF densities trained on unperturbed data give well-defined probabilities on perturbed (out-of-distribution) data.
    The method evaluates fixed CNFs on test data with nuisance shifts, yet normalizing flows are known to be unreliable out-of-distribution (Ref. [16]); the paper provides no analysis of OOD behavior for these specific shifts.
  • domain assumption The analytic loss-minimizing contrastive distribution (Eq. 10) yields features that improve classifier robustness to domain shift.
    The vanishing contour of p_theta*(s) matches the Bayes decision boundary, but the claimed robustness of the full learned distribution to nuisance shifts is justified only heuristically in Sec. 4.2 and tested on a toy.
  • domain assumption The Neyman construction using simulated test sets with random nuisances gives correct coverage for the challenge evaluation.
    Standard frequentist calibration, but the guarantee depends on the simulator and nuisance priors matching the evaluation procedure.
  • ad hoc to paper The two nuisance parameters alpha_jes and alpha_tes are sufficient for the binned likelihood; the remaining four can be handled through random draws in the Neyman construction.
    Sec. 6.2 step 4 states these two are found to affect discrimination most, with no systematic check of the others' influence on the likelihood.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Normalizing Flows for Uncertainty-Aware Parameter Estimation." pith.science (2026). https://pith.science/paper/GWAAWQSE

@misc{pith2026250508709,
  author       = {Pith},
  title        = {Pith review of: Contrastive Normalizing Flows for Uncertainty-Aware Parameter Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GWAAWQSE}},
  note         = {Machine review of arXiv:2505.08709}
}
read the original abstract

Estimating physical parameters from data is a crucial application of machine learning (ML) in the physical sciences. However, systematic uncertainties, such as detector miscalibration, induce data distribution distortions that can erode statistical precision. In both high-energy physics (HEP) and broader ML contexts, achieving uncertainty-aware parameter estimation under these domain shifts remains an open problem. In this work, we address this challenge of uncertainty-aware parameter estimation for a broad set of tasks critical for HEP. We introduce a novel approach based on Contrastive Normalizing Flows (CNFs), which achieves top performance on the HiggsML Uncertainty Challenge dataset. Building on the insight that a binary classifier can approximate the model parameter likelihood ratio, we address the practical limitations of expressivity and the high cost of simulating high-dimensional parameter grids by embedding data and parameters in a learned CNF mapping. This mapping yields a tunable contrastive distribution that enables robust classification under shifted data distributions. Through a combination of theoretical analysis and empirical evaluations, we demonstrate that CNFs, when coupled with a classifier and established frequentist techniques, provide principled parameter estimation and uncertainty quantification through classification that is robust to data distribution distortions.

Figures

Figures reproduced from arXiv: 2505.08709 by the authors.

Figure 1
Figure 1. Randomly sampled signal (blue) and background (red) points with contours overlaid. CNF [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. A rotation of the data (black arrows) can move points across the DNN decision boundary [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Accuracy as a function of nuisance parameter [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Setup of the HiggsML Uncertainty Challenge. The left panel shows an example of the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Flow chart for our CNF-based method of estimating [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the CNF+DNN method for uncertainty-aware parameter estimation. The left [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Contours of the loss-minimizing p (s) θ (x) from Eq. (10) on the toy example of 2-dimensional Gaussians from Sec. 5, for various values of c. A.2 Modifications for contrastive normalizing flow Due to CNF’s contrastive objective, training a CNF in practice is complicate…
Figure 8
Figure 8. Figure 8: Samples from a NF trained on just MNIST zeros (left), compared to a CNF trained with [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: MNIST images correctly classified (left) and misclassified (right) with a CNF trained on [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: MNIST images correctly classified (left) and misclassified (right) with a CNF trained on [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Higgs Signal Strength Estimation with Machine Learning under Systematic Uncertainties

    hep-ph 2025-08 conditional novelty 6.0 of 10

    SAGE, a dual-branch GNN trained under nuisance fluctuations, estimates the Higgs signal strength with near-nominal coverage (0.662-0.683) but wider intervals than the top FAIR-HUC leaderboard methods.

Reference graph

Works this paper leans on

28 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [18]

    Positive Difference Distribution for Image Outlier Detection using Normalizing Flows and Contrastive Data

    R. Schmier, U. Köthe and C.-N. Straehle, Positive difference distribution for image outlier detection using normalizing flows and contrastive data, 2208.14024

    work page Pith review arXiv

  2. [21]

    Cranmer, J

    K. Cranmer, J. Pavez and G. Louppe, Approximating likelihood ratios with calibrated discriminative classifiers, 1506.02169

    arXiv

  3. [1]

    Bhimji, P

    W. Bhimji, P. Calafiura, R. Chakkappai, P.-W. Chang, Y .-T. Chou, S. Diefenbacher et al.,FAIR Universe HiggsML Uncertainty Challenge Competition, 2410.02867

    arXiv

  4. [2]

    Louppe, J

    G. Louppe, J. Hermans and K. Cranmer, Adversarial variational optimization of non-differentiable simulators, 1707.07113

    arXiv

  5. [3]

    Ovadia, E

    Y . Ovadia, E. Fertig, J. Ren, Z. Nado, D. Sculley, S. Nowozin et al.,Can You Trust Your Model’s Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift, 1906.02530

    arXiv 1906

  6. [4]

    P. W. Koh, S. Sagawa, H. Marklund, S. M. Xie, M. Zhang, A. Balsubramani et al., Wilds: A benchmark of in-the-wild distribution shifts, in Proceedings of the 38th International Conference on Machine Learning (M. Meila and T. Zhang, eds.), vol. 139 of Proceedings of Machine Learning Research, pp. 5637–5664, PMLR, 18–24 Jul, 2021

    work page 2021

  7. [5]

    Arjovsky, L

    M. Arjovsky, L. Bottou, I. Gulrajani and D. Lopez-Paz, Invariant risk minimization, 1907.02893

    arXiv 1907

  8. [6]

    Sagawa, P

    S. Sagawa, P. W. Koh, T. B. Hashimoto and P. Liang,Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization, 1911.08731

    arXiv 1911

Show all 28 references
  1. [7]

    Malinin and M

    A. Malinin and M. Gales, Predictive uncertainty estimation via prior networks, 1802.10501

    arXiv

  2. [8]

    Z. C. Lipton, The mythos of model interpretability, 1606.03490

    arXiv

  3. [9]

    Baldi, K

    P. Baldi, K. Cranmer, T. Faucett, P. Sadowski and D. Whiteson,Parameterized neural networks for high-energy physics, The European Physical Journal C 76 (Apr., 2016)

    2016

  4. [10]

    Louppe, M

    G. Louppe, M. Kagan and K. Cranmer, Learning to pivot with adversarial networks, 1611.01046

    arXiv

  5. [11]

    Ghosh, B

    A. Ghosh, B. Nachman and D. Whiteson, Uncertainty-aware machine learning for high energy physics, Physical Review D 104 (Sept., 2021)

    2021

  6. [12]

    Dorigo, A

    T. Dorigo, A. Giammanco, P. Vischia, M. Aehle, M. Bawaj, A. Boldyrev et al., Toward the end-to-end optimization of particle physics instruments with differentiable programming: a white paper, 2203.13818

    arXiv

  7. [13]

    L. Dinh, J. Sohl-Dickstein and S. Bengio, Density estimation using Real NVP, 1605.08803

    arXiv

  8. [14]

    D. P. Kingma and P. Dhariwal, Glow: Generative Flow with Invertible 1× 1 Convolutions, 1807.03039

    arXiv

  9. [15]

    Papamakarios, E

    G. Papamakarios, E. Nalisnick, D. J. Rezende, S. Mohamed and B. Lakshminarayanan, Normalizing flows for probabilistic modeling and inference, 1912.02762

    1912 arXiv

  10. [16]

    Nalisnick, A

    E. Nalisnick, A. Matsukawa, Y . W. Teh and B. Lakshminarayanan,Detecting out-of-distribution inputs to deep generative models using typicality, 1906.02994. 10

    1906 arXiv

  11. [17]

    J. Ren, P. J. Liu, E. Fertig, J. Snoek, R. Poplin, M. A. DePristo et al., Likelihood ratios for out-of-distribution detection, 1906.02845

    1906 arXiv

  12. [19]

    Cao and Z

    S. Cao and Z. Zhang, Deep hybrid models for out-of-distribution detection, in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4723–4733, 2022. DOI

    2022

  13. [20]

    Andreassen and B

    A. Andreassen and B. Nachman, Neural networks for full phase-space reweighting and parameter tuning, Physical Review D 101 (May, 2020)

    2020

  14. [22]

    Neyman, Outline of a theory of statistical estimation based on the classical theory of probability, Philosophical Transactions of the Royal Society of London

    J. Neyman, Outline of a theory of statistical estimation based on the classical theory of probability, Philosophical Transactions of the Royal Society of London. Series A, Mathematical and Physical Sciences 236 (1937) 333–380

    1937

  15. [23]

    Perlmutter system architecture

    National Energy Research Scientific Computing Center, “Perlmutter system architecture.” https://docs.nersc.gov/systems/perlmutter/architecture/

  16. [24]

    T. J. Boerner, S. Deems, T. R. Furlani, S. L. Knuth and J. Towns, ACCESS: Advancing Innovation: NSF’s Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support, in Practice and Experience in Advanced Research Computing 2023: Computing for the Common Good, PEARC ’...

    2023

  17. [25]

    6.2 as is the Neyman Construction once ˆµ is estimated

    In this procedure, the event selection and pre-processing are identical to the procedure described in Sec. 6.2 as is the Neyman Construction once ˆµ is estimated. The main differences reside in the training and evaluation portion of our method. 16 E.1 Training: DNN training wi...

  18. [26]

    Given a data pointx, letνtrue∈ Rkν andνfake∈ Rkν be two sets of different full 6D nuisance-parameter vectors, where onlyνtrue is the true set s.tx∼ps/b(x|νtrue)

  19. [27]

    During training, we append both copies to each training event, and let the ordering depend on the label: ˜x = x,νtrue,νfake if the event is signal, x,νfake,νtrue if the event is background (25)

  20. [28]

    Thus, the network implicitly learns how variations ofν shift the score distribution for each class

    The same two -headed DNN (one head for 1 -jet, one for 2 -jet events) with binary cross-entropy loss is trained on this augmented input. Thus, the network implicitly learns how variations ofν shift the score distribution for each class. E.2 Evaluation In this procedure, traini...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.