Old Rules in a New Game: Mapping Uncertainty Quantification to Quantum Machine Learning

Kilian Tscharke; Maximilian Wendlinger; Pascal Debus

REVIEW 4 major objections 5 minor 39 references

Old Rules in a New Game: Mapping Uncertainty Quantification to Quantum Machine Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Classical uncertainty quantification can be mapped onto quantum machine learning; in this paper's comparative evaluation the Bayesian QML model is the best calibrated (ECE 0.0395), with Gaussian Monte Carlo dropout close behind (ECE…

desk verdict A useful pilot comparison of classical UQ methods in QML, but a sign error in the Bayesian objective undermines the headline ranking until fixed. read the letter →

arxiv 2507.14919 v1 pith:JLHB2YOU submitted 2025-07-20 cs.LG quant-ph

Maximilian Wendlinger , Kilian Tscharke , Pascal Debus This is my paper

classification cs.LGquant-ph

keywords uncertaintyquantificationquantummachinelearningBayesiancircuitsMonteCarlodropoutensemblesGaussianprocessesexpectedcalibrationerrorvariational

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

This paper argues that the main uncertainty quantification (UQ) techniques developed for classical deep learning can be transferred to quantum machine learning (QML), giving variational quantum circuits the ability to say how confident they are. The authors construct four families of quantum UQ models—Bayesian circuits, three Monte-Carlo dropout variants, circuit ensembles, and Gaussian processes with quantum kernels—and evaluate them on low-dimensional regression and classification tasks. They find the Bayesian QML model best calibrated, with expected calibration error 0.0395, and that adding Gaussian noise to circuit parameters (Gaussian dropout) yields well-calibrated uncertainty at 0.0818 with essentially no implementation overhead. If correct, this means QML practitioners do not need bespoke quantum uncertainty theories; classical tools, adapted to circuits, can open up the quantum black box.

What carries the argument

The machinery is a family of single-qubit variational circuits using data re-upload encoding, whose model functions are Fourier series $f(x;\theta)=\sum_{\omega\in\Omega} c_\omega e^{i\omega x}$; this representation is what lets the authors derive quantum kernels and transfer classical UQ constructions. Uncertainty is produced by repeating forward passes and aggregating empirical mean and variance, $\mu=\frac{1}{M}\sum_{m=1}^M \hat{y}_m$ and $\sigma=\sqrt{\frac{1}{M-1}\sum_{m=1}^M(\hat{y}_m-\mu)^2}$, applied to Bayesian posterior samples, dropout configurations, or ensemble members. For Gaussian processes, the kernel is the overlap $k(x,x')=|\langle 0|U(x')^\dagger U(x)|0\rangle|^2$, and for Bayesian circuits the variational posterior is a diagonal Gaussian over rotation angles using the reparameterization trick.

What would settle it

Train the Gaussian-dropout QML model on a regression task while steadily increasing the number of training points in one region and check whether the predictive variance in that region shrinks as the data grows. If the variance stays roughly constant, the dropout spread is not epistemic uncertainty and the Bayesian analogy fails; alternatively, compare the dropout predictive distribution against an exact posterior on a tiny circuit and measure the divergence.

Watch

Extended reading notes

Core claim

The central claim is that classical UQ methods—Bayesian variational inference, MC dropout, deep ensembles, and Gaussian processes—survive translation to parameterized quantum circuits and give genuinely useful uncertainty estimates there. Concretely, the paper reports a comparative evaluation in which the Bayesian QML framework shows the best calibration (ECE 0.0395), while a Gaussian-noise dropout variant is nearly as well calibrated (ECE 0.0818) and requires only repeated forward passes through the same circuit; bias-only dropout (0.1082), quantum ensembles (0.1091), and quantum-kernel Gaussian processes (0.1459) follow. The authors also observe that only the ensemble is overconfident, while dropout and GP models err toward underconfidence.

Load-bearing premise

The load-bearing premise is that randomly dropping gates or perturbing parameters in a quantum circuit, then averaging the outputs, gives a sample from a meaningful posterior over functions—the same assumption that justifies MC dropout in classical networks; the paper does not prove this for quantum circuits.

Editorial extensions

If this is right

Variational quantum circuits can be made uncertainty-aware without new quantum-specific UQ theory; each classical method has a direct circuit analogue.
Bayesian QML offers the best calibration but costs more than plain training; Gaussian MC dropout gives nearly as good calibration at near-zero overhead, making it a practical default for cheap uncertainty estimates.
Quantum ensembles are the only tested method that tends toward overconfidence, so they are a poor choice in safety-critical settings; dropout and GP models are mildly underconfident, which is safer.
Because circuit outputs are periodic, uncertainty estimates inherit that periodicity; reliable out-of-distribution detection requires matching the circuit architecture to the periodicity of the data.
The same aggregation recipe (empirical mean and variance over repeated forward passes) works for Bayesian sampling, dropout, and ensembles, so the methods compose naturally with each other.

Reading between the lines

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

If Gaussian dropout's calibration persists on real hardware, then physical gate noise—usually treated as an error—could be repurposed as a free source of epistemic uncertainty estimates.
The same mapping recipe could be applied to variational dropout and deterministic distance-aware UQ methods, which the paper names as future work but does not implement.
The Fourier/kernel connection suggests a testable scaling prediction: UQ quality should degrade as higher-frequency circuit components are added, since the kernel becomes more oscillatory; the paper's appendix on deep versus shallow GP kernels already hints at this.
Because all experiments use one- or two-qubit circuits and low-dimensional data, the claimed transferability to large QML models is an extrapolation; the empirical ranking could change with more qubits, layers, and entanglement.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a genuinely useful pilot comparison, and the three MC-dropout variants are the real contribution. But the Bayesian QML headline is compromised: Equation (14) is the negative of the standard Gaussian KL divergence, so the training objective as written is not BayesByBackprop. That needs to be resolved before the ECE ranking means anything.

What's new: nobody has mapped MC dropout to QML in these three flavors (bias-only, rotation, Gaussian noise), and nobody has put Bayesian QML, dropout, ensembles, and quantum-kernel GPs side by side with calibration curves. The writing is clear, the figures are informative, and the limitations appendix is honest. The qualitative result that Gaussian dropout gives cheap, reasonably calibrated uncertainty is plausible and worth replicating.

Soft spots, in order of severity. First, the sign error. The KL term for a diagonal Gaussian posterior against a unit Gaussian prior is 0.5*sum(σ²+μ²−1−log σ²). Equation (14) is the negative of that. So L = LKL + Lobj is actually −D_KL plus the data term. Minimizing that drives the posterior away from the prior; it is not the ELBO or BayesByBackprop. If the implementation used the printed equation, the reported ECE for Bayesian QML (0.0395) is an artifact of a mis-specified objective. If the code used the correct sign, the paper must say so. This is fixable, but it is load-bearing.

Second, the empirical rigor is pilot-level: no error bars or seeds reported for the ECE values, the regression target is a Fourier series the circuit can represent exactly, and the GP uses a different architecture (two qubits, two layers) than all other models (one qubit, six layers). Appendix A discloses this, which I appreciate, but it means the GP comparison is not on equal footing. Third, code is only available "upon reasonable request." That is not enough if the ECE numbers are to be trusted.

Fourth, the paper asserts MC dropout approximates the predictive posterior without argument. In classical ML that connection has theory; for random gate removal or adding noise in a quantum circuit, it is a heuristic. The paper should either supply a justification or label the uncertainty estimates as empirical scatter.

Who benefits: QML practitioners who want a menu of uncertainty tools for small circuits. The paper deserves a serious referee — it will need revision, but the direction is sound and the dropout variants are a real contribution.

Referee Report

4 major / 5 minor

Summary. The paper maps four classical uncertainty quantification (UQ) techniques to variational quantum circuits: Bayesian QML via BayesByBackprop, three variants of Monte-Carlo dropout (bias-only, rotation, Gaussian), quantum circuit ensembles, and Gaussian processes with quantum kernels. The methods are evaluated on a one-dimensional Fourier-series regression task with epistemic and aleatoric uncertainty gaps and on a two-moons classification task, using expected calibration error (ECE) as the main quantitative metric. The paper reports that Bayesian QML achieves the best calibration (ECE 0.0395), that Gaussian MC dropout offers well-calibrated uncertainty with low implementation overhead (ECE 0.0818), and that different methods exhibit different overconfidence or underconfidence behavior. The authors argue that classical UQ insights can be transferred to QML and should be integrated into QML design.

Significance. If the empirical claims hold, the paper would provide useful practical guidance for adding uncertainty estimates to variational quantum classifiers and regressors. The comparative evaluation of multiple UQ mechanisms on a common circuit architecture is a useful contribution, and the qualitative plots (Figures 3–6) give an intuitive picture of predictive uncertainty. However, the significance is currently limited by the sign error in the Bayesian objective and the lack of statistical rigor in the ECE comparison, which together undermine the headline ranking. The paper does not provide formal derivations for the quantum dropout posterior interpretation, so its central recommendation rests on an analogy rather than on a proven connection. The code is only 'available upon reasonable request', which hampers reproducibility.

major comments (4)

[IV-A, Eq. (14)] Equation (14) defines L_KL = 1/2 Σ_k (1 + log σ_k^2 − μ_k^2 − σ_k^2). For a diagonal Gaussian variational posterior q(θ)=N(μ, σ^2) and a unit Gaussian prior, the KL divergence is D_KL(q||p) = 1/2 Σ_k (σ_k^2 + μ_k^2 − 1 − log σ_k^2), so Eq. (14) is the negative KL divergence. Minimizing L = L_KL + L_obj therefore drives the variational distribution away from the prior rather than regularizing toward it, which is the opposite of the BayesByBackprop/ELBO objective introduced in Section III-B1. This is an internal inconsistency: the paper's headline result that Bayesian QML has the best calibration (Table I, ECE 0.0395) is attributed to a Bayesian procedure whose objective, as written, is not a Bayesian objective. If the implementation used the correct sign, the manuscript should state this explicitly; as written, the Bayesian claim is not established.
[VI-B, Table I] The ECE values in Table I are reported without error bars, multiple seeds, or details on the number of test points per bin. The differences between Bayesian QML (0.0395), Gaussian MC dropout (0.0818), and bias-only dropout (0.1082) may be within run-to-run variance, especially given the small one-dimensional regression set. The central comparative claim therefore lacks statistical support; please provide means and standard deviations over repeated training runs, or a statement of the evaluation protocol (e.g., number of seeds, test-set size, binning scheme).
[V-A and VI] The regression target is the degree-2 Fourier series from Eq. (21), which the authors explicitly note is a function the quantum circuit can represent exactly. This makes the regression task favorable to the model class and may inflate calibration performance; a single low-dimensional task, plus the two-moons classification in Appendix C, is a narrow basis for the general recommendation that classical UQ techniques map successfully to QML. The paper acknowledges the low-dimensional limitation in Appendix D, but the quantitative ranking in Table I should be framed as a proof-of-concept on a favorable task, not as a general empirical conclusion.
[IV-B] The three quantum MC dropout variants are justified solely by analogy to Gal and Ghahramani's classical result that MC dropout approximates Bayesian inference, but no argument is given that random gate removal or parameter perturbation in a quantum circuit produces samples from a meaningful posterior over functions, especially given the periodic Fourier structure of Eq. (11). The paper's recommendation that Gaussian MC dropout is a cheap and well-calibrated alternative depends on this analogy; without a derivation (or at least a clear statement that this is a heuristic), the uncertainty estimates from dropout are not formally connected to a posterior. Please either provide a formal connection or soften the Bayesian interpretation.

minor comments (5)

[III-B1, Eq. (1)] The integrand in Eq. (1) appears as p(y∗|x∗θ); it should be p(y∗|x∗, θ) for clarity.
[V-B, Figure 2] The circuit diagram is minimal; please specify the exact form of R(w_j ∘ x + b_j) and the measurement basis in the caption.
[Table I] The 'Cost' and 'Quality' columns are subjective; please define them operationally (e.g., wall-clock time, parameter count, or number of circuit evaluations).
[Appendix A] The spelling of 'BayesByBackprop' is inconsistent (BayesByBackprop vs. BayesByBackProp); please standardize.
[Appendix E] Stating that code is 'available upon reasonable request' is not sufficient for reproducibility; please provide a public repository or persistent archive.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the UQ mappings are imported from external classical and quantum results and benchmarked against held-out ECE; no fitted parameter is renamed as a prediction and no load-bearing self-citation appears.

full rationale

The paper's derivations are imports from independently established results: BayesByBackprop [21], MC dropout [17], deep ensembles [22], Gaussian-process regression [23], the Fourier-series characterization of re-upload encoding [27], and quantum feature maps [5]-[7]. None of these is a self-citation of the present authors, and none is used to forbid alternative constructions. The empirical ranking in Table I is obtained by training on the regression/classification losses and then computing ECE on fixed test sets; no parameter is fit to the ECE values, and no reported calibration number is a fitted input renamed as a prediction. The choice of a Fourier-series regression target in Section V-A deliberately matches the circuit family's expressive power, but the paper acknowledges this design choice; it is a selection bias, not a circular derivation. The MC-dropout construction in Section IV-B is an analogy to Gal and Ghahramani: the paper does not prove that dropping gates or adding Gaussian perturbation samples a meaningful posterior over functions, so those uncertainty estimates may be unjustified scatter. That is a correctness and validity risk, not a circular reduction. Similarly, the KL term in Equation (14) is the negative of D_KL(q||p) for q=N(mu,sigma^2) and p=N(0,1), so the stated BayesByBackprop objective is internally inconsistent; if implemented with that sign, the learned distribution is pushed away from the prior and the 'Bayesian' label may be unjustified. Again, this is an error in the stated objective rather than a case of the paper predicting, by construction, what was already fitted. No step in the claimed derivation chain reduces to its own inputs, so the honest circularity finding is score 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new free-floating entities; its burden is in the assumptions that classical UQ theory (especially the Bayesian interpretation of dropout) carries over unchanged to quantum circuits, and that the chosen toy architectures and data are representative. These assumptions are load-bearing for the empirical recommendations.

free parameters (6)

Dropout probability p = 0.1
Chosen by hand for bias-only, rotation, and Gaussian dropout variants; directly sets the spread of the MC samples and hence the uncertainty magnitude (Section VI-A).
Data noise standard deviation sigma_epsilon = 0.1
Used to generate aleatoric noise in the regression data; also the assumed noise level in the GP covariance, though the GP inference noise is not explicitly stated (Section V-A).
Fourier coefficients of regression target = c0=0.1, c1=c2=0.15-0.15i
Define the regression target as a degree-2 Fourier series, which a six-layer re-uploading circuit can represent exactly, so the data distribution favors the model family (Section V-A).
Circuit architecture = 6 layers, 1 qubit (base); 2 qubits, 2 layers (GP kernel)
The base architecture is chosen by hand; the GP used a different architecture after the original one proved unstable in training (Appendix A).
Ensemble size M = 8
Number of circuits aggregated in the quantum ensemble model (Section VI-A).
Early stopping thresholds = L_MSE < 0.005, L_CE < 0.3
Convergence criteria for all models; these thresholds influence the final fitted model and hence the calibration results (Section V-B).

assumptions (6)

standard math Bayes' theorem and variational inference with diagonal Gaussian variational distributions
Used in Section IV-A to define the Bayesian QML loss (Eq. 13-14).
domain assumption Classical MC dropout approximates Bayesian inference over model weights (Gal and Ghahramani, 2016)
Transferred to quantum circuits without proof in Section IV-B; the predictive mean and variance in Eqs. (16)-(17) assume this analogy.
standard math Parameterized quantum circuits with data re-uploading have a finite Fourier series representation
Invoked in Eq. (11) and for the quantum kernel in Eq. (20).
standard math Quantum kernels are valid positive-semidefinite covariance functions
Required for using the quantum kernel inside standard GP regression (Section IV-D).
domain assumption The six-layer, one-qubit circuit is expressive enough for the chosen tasks
The regression target is a degree-2 Fourier series, exactly the function class the circuit represents; the two-moons classification is only partially fitted (Appendix C), yet the model comparisons are drawn from this architecture.
domain assumption State-vector simulation faithfully represents the behavior of the quantum model
All results come from PennyLane simulators, so real-device noise, especially relevant for Gaussian dropout (which mimics noise), is not covered (Appendix A).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Old Rules in a New Game: Mapping Uncertainty Quantification to Quantum Machine Learning." pith.science (2026). https://pith.science/paper/JLHB2YOU

@misc{pith2026250714919,
  author       = {Pith},
  title        = {Pith review of: Old Rules in a New Game: Mapping Uncertainty Quantification to Quantum Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JLHB2YOU}},
  note         = {Machine review of arXiv:2507.14919}
}

read the original abstract

One of the key obstacles in traditional deep learning is the reduction in model transparency caused by increasingly intricate model functions, which can lead to problems such as overfitting and excessive confidence in predictions. With the advent of quantum machine learning offering possible advances in computational power and latent space complexity, we notice the same opaque behavior. Despite significant research in classical contexts, there has been little advancement in addressing the black-box nature of quantum machine learning. Consequently, we approach this gap by building upon existing work in classical uncertainty quantification and initial explorations in quantum Bayesian modeling to theoretically develop and empirically evaluate techniques to map classical uncertainty quantification methods to the quantum machine learning domain. Our findings emphasize the necessity of leveraging classical insights into uncertainty quantification to include uncertainty awareness in the process of designing new quantum machine learning models.

Figures

Figures reproduced from arXiv: 2507.14919 by the authors.

**Figure 2.** The basic architecture for all quantum models, on [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 4.** Regression results for all MC dropout QML models. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

Figures from the paper (7 more)

**Figure 6.** Figure 6: Gaussian Process regression results using a quantum [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Calibration curves for all models investigated in this [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Comparative overview over regression output functions [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: Classification results for deterministic QML model [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

**Figure 10.** Figure 10: Classification results for a Bayesian QML model. Left: [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

**Figure 12.** Figure 12: Classification results for an ensemble of eight quan [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]

**Figure 13.** Figure 13: Classification results for a Gaussian Process with a [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 22 canonical work pages

[1]

Jay Gambetta, The hardware and software for the era of quantum utility is here , Dec. 2023. (visited on 10/16/2024)

work page 2023
[2]

Algorithms for quantum computation: Discrete logarithms and factoring,

P. Shor, “Algorithms for quantum computation: Discrete logarithms and factoring,” in Proceedings 35th An- nual Symposium on Foundations of Computer Science , Santa Fe, NM, USA: IEEE Comput. Soc. Press, 1994, pp. 124–134, ISBN : 978-0-8186-6580-6. DOI: 10.1109/ SFCS.1994.365700. (visited on 10/18/2024)

arXiv 1994
[3]

Quantum algorithm for solving linear systems of equations,

A. W. Harrow, A. Hassidim, and S. Lloyd, “Quantum algorithm for solving linear systems of equations,” Physical Review Letters , vol. 103, no. 15, p. 150 502, Oct. 2009, ISSN : 0031-9007, 1079-7114. DOI: 10 . 1103 / PhysRevLett . 103 . 150502. arXiv: 0811 . 3171 [quant-ph]. (visited on 08/26/2024)

work page 2009
[4]

A rigorous and robust quantum speed-up in supervised machine learning,

Y . Liu, S. Arunachalam, and K. Temme, “A rigorous and robust quantum speed-up in supervised machine learning,” Nature Physics , vol. 17, no. 9, pp. 1013– 1017, Sep. 2021, ISSN : 1745-2473, 1745-2481. DOI: 10 . 1038 / s41567 - 021 - 01287 - z. arXiv: 2010 . 02174 [quant-ph]. (visited on 09/03/2024)

work page 2021
[5]

Schuld, Supervised quantum machine learning mod- els are kernel methods , Apr

M. Schuld, Supervised quantum machine learning mod- els are kernel methods , Apr. 2021. DOI: 10 . 48550 / arXiv.2101.11020. arXiv: 2101.11020 [quant-ph, stat]. (visited on 10/03/2024)

arXiv doi:10.48550/arxiv.2101.11020 2021
[6]

Quantum machine learning in feature Hilbert spaces,

M. Schuld and N. Killoran, “Quantum machine learning in feature Hilbert spaces,” Physical Review Letters , vol. 122, no. 4, p. 040 504, Feb. 2019, ISSN : 0031- 9007, 1079-7114. DOI: 10 . 1103 / PhysRevLett . 122 . 040504. arXiv: 1803.07128 [quant-ph]. (visited on 10/01/2024)

arXiv 2019
[7]

Quantum Gaussian Process Regression for Bayesian Optimization,

F. Rapp and M. Roth, “Quantum Gaussian Process Regression for Bayesian Optimization,” Quantum Ma- chine Intelligence, vol. 6, no. 1, p. 5, Jun. 2024, ISSN : 2524-4906, 2524-4914. DOI: 10 . 1007 / s42484 - 023 - 00138 - 9. arXiv: 2304 . 12923 [quant-ph]. (visited on 07/09/2024)

work page 2024
[8]

Bayesian Quantum Neural Networks,

N. Nguyen and K.-C. Chen, “Bayesian Quantum Neural Networks,” IEEE Access , vol. 10, pp. 54 110–54 122, 2022, ISSN : 2169-3536. DOI: 10.1109/ACCESS.2022. 3168675. (visited on 09/10/2024)

doi:10.1109/access.2022 2022

Show all 39 references

[9]

Schuld and F

M. Schuld and F. Petruccione, Quantum ensembles of quantum classifiers , Apr. 2017. DOI: 10 . 48550 / arXiv.1704.02146. arXiv: 1704.02146 [quant-ph, stat]. (visited on 09/18/2024)

2017 doi arXiv
[10]

Bayesian Deep Learning on a Quantum Com- puter,

Z. Zhao, A. Pozas-Kerstjens, P. Rebentrost, and P. Wittek, “Bayesian Deep Learning on a Quantum Com- puter,” Quantum Machine Intelligence , vol. 1, no. 1- 2, pp. 41–51, May 2019, ISSN : 2524-4906, 2524-4914. DOI: 10.1007/s42484-019-00004-7. arXiv: 1806.11463 [quant-ph, stat] ....

2019 arXiv
[11]

Uncertainty in Deep Learning,

Y . Gal, “Uncertainty in Deep Learning,” Ph.D. disser- tation, University of Cambridge, Sep. 2016

2016
[12]

W. He, Z. Jiang, T. Xiao, Z. Xu, and Y . Li, A Survey on Uncertainty Quantification Methods for Deep Learning, Jul. 2024. arXiv: 2302.13425 [cs, stat] . (visited on 10/08/2024)

2024
[13]

Uncertainty Quantification in Ma- chine Learning for Engineering Design and Health Prognostics: A Tutorial,

V . Nemani et al. , “Uncertainty Quantification in Ma- chine Learning for Engineering Design and Health Prognostics: A Tutorial,” Mechanical Systems and Sig- nal Processing, vol. 205, p. 110 796, Dec. 2023, ISSN : 08883270. DOI: 10.1016/j.ymssp.2023.110796. arXiv: 2305.04933 [...

2023
[14]

Park and O

S. Park and O. Simeone, Quantum Conformal Predic- tion for Reliable Uncertainty Quantification in Quan- tum Machine Learning , Oct. 2023. arXiv: 2304.03398 [quant-ph]. (visited on 07/09/2024)

2023 arXiv
[15]

Over- fitting in quantum machine learning and entangling dropout,

M. Kobayashi, K. Nakaji, and N. Yamamoto, “Over- fitting in quantum machine learning and entangling dropout,” Quantum Machine Intelligence , vol. 4, no. 2, p. 30, Dec. 2022, ISSN : 2524-4906, 2524-4914. DOI: 10 . 1007 / s42484 - 022 - 00087 - 9. arXiv: 2205 . 11446 [quant-ph]....

2022
[16]

A General Approach to Dropout in Quantum Neural Net- works,

F. Scala, A. Ceschini, M. Panella, and D. Gerace, “A General Approach to Dropout in Quantum Neural Net- works,” Advanced Quantum Technologies, p. 2 300 220, Dec. 2023, ISSN : 2511-9044, 2511-9044. DOI: 10.1002/ qute . 202300220. arXiv: 2310 . 04120 [quant-ph]. (visited on 08/28/2024)

2023
[17]

Dropout as a Bayesian Ap- proximation: Representing Model Uncertainty in Deep Learning,

Y . Gal and Z. Ghahramani, “Dropout as a Bayesian Ap- proximation: Representing Model Uncertainty in Deep Learning,” no. arXiv:1506.02142, Oct. 2016. arXiv: 1506.02142 [cs, stat] . (visited on 07/30/2024)

2016 arXiv
[18]

Polson, V

N. Polson, V . Sokolov, and J. Xu, Quantum Bayesian Computation, Mar. 2023. arXiv: 2208 . 08068 [cs, math, stat] . (visited on 07/09/2024)

2023
[19]

Quan- tum assisted Gaussian process regression,

Z. Zhao, J. K. Fitzsimons, and J. F. Fitzsimons, “Quan- tum assisted Gaussian process regression,” Physical Review A, vol. 99, no. 5, p. 052 331, May 2019, ISSN : 2469-9926, 2469-9934. DOI: 10 . 1103 / PhysRevA . 99 . 052331. arXiv: 1512 . 03929 [quant-ph, stat] . (visited on ...

2019
[20]

An introduction to variational methods for graph- ical models,

M. I. Jordan, Z. Ghahramani, T. S. Jaakkola, and L. K. Saul, “An introduction to variational methods for graph- ical models,” Machine learning, vol. 37, pp. 183–233, 1999

1999
[21]

Blundell, J

C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra, Weight Uncertainty in Neural Networks , May
[22]

Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles,

B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles,” no. arXiv:1612.01474, Nov

arXiv
[23]

Gaussian Processes in Ma- chine Learning,

Carl Edward Rasmussen, “Gaussian Processes in Ma- chine Learning,” in Advanced Lectures on Machine Learning, vol. 3176, 2004

2004
[24]

K. P. Murphy, Probabilistic Machine Learning: An In- troduction (Adaptive Computation and Machine Learn- ing Series). Cambridge, Massachusetts: The MIT Press, 2022, ISBN : 978-0-262-04682-4

2022
[25]

M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information , 10th anniversary ed. Cam- bridge ; New York: Cambridge University Press, 2010, ISBN : 978-1-107-00217-3

2010
[26]

P ´erez-Salinas, A

A. P ´erez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, Data re-uploading for a universal quantum classifier, https://arxiv.org/abs/1907.02085v3, Jul. 2019. DOI: 10 . 22331 / q - 2020 - 02 - 06 - 226. (visited on 10/03/2024)

1907 arXiv
[27]

The effect of data encoding on the expressive power of variational quantum machine learning models,

M. Schuld, R. Sweke, and J. J. Meyer, “The effect of data encoding on the expressive power of variational quantum machine learning models,” Physical Review A, vol. 103, no. 3, p. 032 430, Mar. 2021, ISSN : 2469- 9926, 2469-9934. DOI: 10.1103/PhysRevA.103.032430. arXiv: 2008.08...

2021 arXiv
[28]

Auto-Encoding Varia- tional Bayes,

D. P. Kingma and M. Welling, “Auto-Encoding Varia- tional Bayes,” Dec. 2013. (visited on 09/12/2024)

2013
[30]

Obtaining Well Calibrated Probabilities Using Bayesian Binning,

M. Pakdaman Naeini, G. Cooper, and M. Hauskrecht, “Obtaining Well Calibrated Probabilities Using Bayesian Binning,” Proceedings of the AAAI Conference on Artificial Intelligence , vol. 29, no. 1, Feb. 2015, ISSN : 2374-3468, 2159-5399. DOI: 10.1609/aaai.v29i1.9602. (visited on...

2015 doi
[31]

Accurate Un- certainties for Deep Learning Using Calibrated Regres- sion,

V . Kuleshov, N. Fenner, and S. Ermon, “Accurate Un- certainties for Deep Learning Using Calibrated Regres- sion,” no. arXiv:1807.00263, Jul. 2018. arXiv: 1807 . 00263 [cs]. (visited on 10/11/2024)

2018 arXiv
[32]

(visited on 10/05/2024)

arXiv: 1412.6980 [cs]. (visited on 10/05/2024)

2024 arXiv
[33]

Bergholm et al

V . Bergholm et al. , PennyLane: Automatic dif- ferentiation of hybrid quantum-classical computa- tions, Jul. 2022. arXiv: 1811 . 04968 [physics, physics:quant-ph]. (visited on 10/04/2024)

2022
[34]

J. Z. Liu et al. , Simple and Principled Uncertainty Estimation with Deterministic Deep Learning via Dis- tance Awareness, Oct. 2020. arXiv: 2006.10108 [cs, stat]. (visited on 07/09/2024)

2020 arXiv
[35]

Paszke et al

A. Paszke et al. , PyTorch: An Imperative Style, High- Performance Deep Learning Library, Dec. 2019. arXiv: 1912.01703 [cs, stat] . (visited on 10/05/2024)

2019 arXiv
[36]

Varia- tional Dropout and the Local Reparameterization Trick,

D. P. Kingma, T. Salimans, and M. Welling, “Varia- tional Dropout and the Local Reparameterization Trick,” no. arXiv:1506.02557, Dec. 2015. DOI: 10.48550/arXiv. 1506.02557. arXiv: 1506.02557 [cs, stat] . (visited on 09/28/2024). APPENDIX A MODEL OPTIMIZATION AND EVALUATION As ...

2015 doi arXiv
[38]

J. Z. Liu et al., A Simple Approach to Improve Single- Model Deep Uncertainty via Distance-Awareness , Dec
[2015]

arXiv: 1505

DOI: 10.48550/arXiv.1505.05424. arXiv: 1505. 05424 [cs, stat] . (visited on 09/10/2024)

2024 doi arXiv
[2017]

arXiv: 1612

DOI: 10.48550/arXiv.1612.01474. arXiv: 1612. 01474 [cs, stat] . (visited on 09/25/2024)

2024 doi arXiv
[2022]

00403 [cs, stat]

arXiv: 2205 . 00403 [cs, stat] . (visited on 07/12/2024)

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Jay Gambetta, The hardware and software for the era of quantum utility is here , Dec. 2023. (visited on 10/16/2024)

work page 2023

[2] [2]

Algorithms for quantum computation: Discrete logarithms and factoring,

P. Shor, “Algorithms for quantum computation: Discrete logarithms and factoring,” in Proceedings 35th An- nual Symposium on Foundations of Computer Science , Santa Fe, NM, USA: IEEE Comput. Soc. Press, 1994, pp. 124–134, ISBN : 978-0-8186-6580-6. DOI: 10.1109/ SFCS.1994.365700. (visited on 10/18/2024)

arXiv 1994

[3] [3]

Quantum algorithm for solving linear systems of equations,

A. W. Harrow, A. Hassidim, and S. Lloyd, “Quantum algorithm for solving linear systems of equations,” Physical Review Letters , vol. 103, no. 15, p. 150 502, Oct. 2009, ISSN : 0031-9007, 1079-7114. DOI: 10 . 1103 / PhysRevLett . 103 . 150502. arXiv: 0811 . 3171 [quant-ph]. (visited on 08/26/2024)

work page 2009

[4] [4]

A rigorous and robust quantum speed-up in supervised machine learning,

Y . Liu, S. Arunachalam, and K. Temme, “A rigorous and robust quantum speed-up in supervised machine learning,” Nature Physics , vol. 17, no. 9, pp. 1013– 1017, Sep. 2021, ISSN : 1745-2473, 1745-2481. DOI: 10 . 1038 / s41567 - 021 - 01287 - z. arXiv: 2010 . 02174 [quant-ph]. (visited on 09/03/2024)

work page 2021

[5] [5]

Schuld, Supervised quantum machine learning mod- els are kernel methods , Apr

M. Schuld, Supervised quantum machine learning mod- els are kernel methods , Apr. 2021. DOI: 10 . 48550 / arXiv.2101.11020. arXiv: 2101.11020 [quant-ph, stat]. (visited on 10/03/2024)

arXiv doi:10.48550/arxiv.2101.11020 2021

[6] [6]

Quantum machine learning in feature Hilbert spaces,

M. Schuld and N. Killoran, “Quantum machine learning in feature Hilbert spaces,” Physical Review Letters , vol. 122, no. 4, p. 040 504, Feb. 2019, ISSN : 0031- 9007, 1079-7114. DOI: 10 . 1103 / PhysRevLett . 122 . 040504. arXiv: 1803.07128 [quant-ph]. (visited on 10/01/2024)

arXiv 2019

[7] [7]

Quantum Gaussian Process Regression for Bayesian Optimization,

F. Rapp and M. Roth, “Quantum Gaussian Process Regression for Bayesian Optimization,” Quantum Ma- chine Intelligence, vol. 6, no. 1, p. 5, Jun. 2024, ISSN : 2524-4906, 2524-4914. DOI: 10 . 1007 / s42484 - 023 - 00138 - 9. arXiv: 2304 . 12923 [quant-ph]. (visited on 07/09/2024)

work page 2024

[8] [8]

Bayesian Quantum Neural Networks,

N. Nguyen and K.-C. Chen, “Bayesian Quantum Neural Networks,” IEEE Access , vol. 10, pp. 54 110–54 122, 2022, ISSN : 2169-3536. DOI: 10.1109/ACCESS.2022. 3168675. (visited on 09/10/2024)

doi:10.1109/access.2022 2022