Pith. sign in

REVIEW 4 major objections 4 minor 77 references

Assessing the Advantages and Limitations of Quantum Neural Networks in Regression Tasks

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper shows that a continuous-variable QNN's regression advantage is task-specific and shrinks when classical networks are compared on matched parameter counts.

desk verdict Abstract oversells the 'seven orders' result that the paper itself later calls an illusion; the body is a fair, honest benchmark with a smaller and more defensible parameter-matched advantage. read the letter →

arxiv 2509.00854 v1 pith:WFVKIMLD submitted 2025-08-31 quant-ph physics.comp-ph

Gubio G. de Limaa , Tiago de S. Farias , Alexandre C. Ricardo , Celso Jorge Villa Boas This is my paper · ORCID
classification quant-phphysics.comp-ph MSC 81P6868Q1268T07 PACS 03.67.-a07.05.Mh
keywords quantumneuralnetworkscontinuous-variablecomputingregressionbenchmarkingclassicalvsmachinelearningNoFreeLunchtheoremvariationalcircuitsKerrnonlinearity
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

Quantum neural networks have been claimed to outperform classical networks, but the conditions for any such advantage remain unclear. This paper takes a continuous-variable quantum neural network and compares it to small feedforward classical networks on two regression tasks: a smooth sine and a discontinuous Heaviside step. It shows that when the comparison is made layer-for-layer with one neuron per layer, the quantum network reaches a sine mean-squared error around 10^-8 while the classical networks stagnate near 10^-1; however, the paper itself identifies this as an artifact of the artificially narrow classical architecture. Under a parameter-matched comparison, the quantum network still achieves lower average sine error than the tested classical configurations, but classical networks with tanh or sigmoid activations fit the sine well once given enough width, and they clearly outperform the quantum network on the step function. The paper's conclusion is that quantum advantage in regression is task-specific, not universal, and that the basis of comparison - layers versus parameters - can determine which model looks superior.

What carries the argument

The engine of the comparison is the continuous-variable quantum neuron: a single-mode circuit layer consisting of a rotation R(theta1), squeezing S(xi), a second rotation R(theta2), displacement D(alpha), and an approximate Kerr interaction K(chi), with five trainable parameters per layer. Data is encoded by displacement of the vacuum, the output is the expectation value of the X quadrature, and the state is simulated in a Fock basis truncated at dimension 30. The Kerr gate supplies the nonlinearity that lets the circuit approximate smooth functions; the paper attributes the QNN's sine advantage to its ability to represent such functions in its state space, and its Heaviside failure to the s

What would settle it

Run the sine-regression task with matched parameter counts (5, 10, 15, 20, 25) but with classical networks allowed a broader architecture search: tanh and sigmoid networks with two to three hidden layers, varied widths, learning-rate annealing, and more than 10^4 epochs. If any classical configuration consistently attains average MSE at or below the QNN's reported roughly 10^-5 under the same 100-seed protocol, then the residual quantum advantage on sin(pi*x) collapses. A second check: add a smooth-but-steep function, such as tanh(100x), and see whether the QNN's error remains low; if not, the

Watch

Extended reading notes

Core claim

The central result is a conditional one: for the single-mode continuous-variable QNN studied here, approximating sin(pi*x) is a task where the quantum model makes efficient use of its parameter budget, reaching average MSE three orders of magnitude lower than some classical activation functions under matched parameter counts (and a best case near 10^-8 under the layer comparison); for the Heaviside step, the same QNN saturates at roughly constant error regardless of depth, while classical tanh/sigmoid networks improve with depth. The paper argues this pattern reflects the QNN's inductive bias: the displacement-rotation-squeezing-Kerr circuit naturally represents smooth, oscillatory structure

Load-bearing premise

The load-bearing premise is that the handful of classical architectures and hyperparameter choices tested (one neuron per layer in strategy one, the listed width/depth combinations in strategy two) fairly represent what a classical feedforward network can do at the same parameter count; if a wider or better-tuned classical network reaches equal or lower sine error, the claimed advantage is an artifact of the benchmark.

Editorial extensions

If this is right

  • If correct, QNN-versus-classical benchmark claims should be expressed as 'advantage under a specified comparison basis,' because switching from layers to parameters reverses the qualitative picture for sine.
  • Existing reported QNN advantages may be inflated when classical baselines are architecturally constrained (one neuron per layer), so re-benchmarking with matched parameter budgets and width/depth variation is needed.
  • The single-mode continuous-variable QNN is a viable model for smooth, low-dimensional regression with very few parameters (as low as 5), but not for functions with discontinuities.
  • For discontinuous targets, classical networks with sigmoidal activations remain the safer choice, and quantum circuits would need extra mechanisms such as hybrid classical post-processing to compete.
  • The practical path to quantum advantage in machine learning is problem-specific alignment of function smoothness with the variational circuit's inductive bias, not general supremacy.

Reading between the lines

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

  • A natural testable extension: compare the same continuous-variable QNN to classical networks of equal parameter count but with optimized width/depth and learning-rate schedules; if a classical tanh network then reaches MSE at or below the QNN's average on sine, the residual advantage disappears entirely.
  • The paper's QNN advantage is measured in expressivity only; because the QNN simulation took about a week and classical training is comparatively fast, the headline advantage does not account for computational cost, and a fairer metric would include time-to-solution or wall-clock error.
  • The smooth-versus-discontinuous dichotomy suggests a predictive hypothesis: continuous-variable QNNs with Kerr nonlinearity should excel on band-limited or Fourier-sparse functions, and lose on functions with sharp edges; this can be checked with a family of functions of varying smoothness, such as tanh(kx) for increasing k.
  • The parameter-count comparison still hides an asymmetry: classical parameters can be distributed across width and depth, while the QNN's parameters are all depth-like; future work could compare against classical networks with a deliberately sequential, chain-like architecture to isolate this effect.
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.

Referee Report

4 major / 4 minor

Summary. The paper compares continuous-variable quantum neural networks (QNNs) with classical feedforward neural networks (CNNs) on two single-variable regression tasks: smoothing a sinusoidal function and approximating a Heaviside step function. Two comparison strategies are used: matching the number of layers (with one neuron/layer for CNNs) and matching the number of trainable parameters. The authors report a large QNN advantage on the sine task in the layer-based comparison, a smaller average advantage in the parameter-matched comparison, and a classical advantage on the Heaviside task. They discuss methodological difficulties in fair QNN/CNN benchmarking and conclude with a No-Free-Lunch-style caveat.

Significance. If the claims are properly scoped, the paper makes a useful contribution to the QML benchmarking literature by explicitly demonstrating how the choice of comparison basis (layer count vs. parameter count) can reverse qualitative conclusions about quantum advantage. The authors should be credited for transparently acknowledging the 'illusion of quantum advantage' in their own layer-based comparison. The study is exploratory and limited to two functions, a single-mode QNN, and a small classical architecture set, but the methodological message is valuable. However, the headline quantitative claim in the abstract is not supported by the paper's own analysis, and several technical details (statistical significance, Fock cutoff validation, classical hyperparameter coverage) need attention before the conclusion can be accepted as stated.

major comments (4)
  1. [Abstract and Section 5.4] The abstract states that QNNs achieved errors 'up to seven orders of magnitude lower than their classical counterparts' on the sine function. This number comes from the layer-count comparison in Section 5.2 (Figure 2a), where each classical layer has only 2 parameters versus the QNN's 5. Section 5.4 (second paragraph) explicitly says that the constrained classical architecture 'led to its failure on the sine task, creating the illusion of quantum advantage.' The abstract therefore presents as a headline result a comparison the paper itself labels an artifact. This is an internal inconsistency that must be fixed by either removing the seven-orders claim or qualifying it as a structural-comparison artifact and reporting the parameter-matched result instead.
  2. [Section 5.3 and Figure 4] The parameter-matched comparison is the main evidence for a residual quantum advantage, yet no statistical significance tests are reported. The 100 initialization runs are averaged, but the figure does not show error bars and the text does not provide confidence intervals, distributions, or tests (e.g., paired bootstrap or Mann-Whitney) comparing the QNN against each classical activation function individually. The phrase 'three orders of magnitude smaller than the average of some classical activation functions' is ambiguous: it is not stated which activation functions are included in the average, and the high variance visible in Figure 4a for tanh and sigmoid makes the average alone misleading. Please provide per-condition distributions and formal comparisons.
  3. [Section 3, Fock cutoff] The simulation truncates the Fock basis at a cutoff of 30, with the statement that this 'ensures simulation fidelity.' No convergence check is provided. For displacement encoding of inputs x in [-1,1] and with trainable displacement, squeezing, and Kerr parameters, the photon-number distribution can extend well beyond n=30, especially for large displacement or squeezing values. If the cutoff truncates non-negligible amplitudes, the QNN errors could be contaminated by simulation artifacts. Please report a cutoff-convergence study (e.g., MSE vs. cutoff for representative trained parameters) and the maximum photon number reached across the trained models.
  4. [Section 5.3 and Appendix A.1] The parameter-matched comparison relies on a very small set of classical architectures: Table A.1 lists only 3-4 configurations per parameter count, and the Adam optimizer settings (learning rate 0.01, 10^4 epochs) are fixed for both models without classical hyperparameter tuning. The authors concede in Section 5.4 that the first strategy's classical failure is an artifact of the constrained architecture. To support the claim that the QNN 'makes more effective use' of its parameter budget, the classical side should be given a broader hyperparameter search (e.g., more width/depth combinations, different learning rates, or early stopping) or the conclusion should be explicitly restricted to the specific configurations tested. As written, the residual advantage in Figure 4a could still be a consequence of the limited classical search.
minor comments (4)
  1. [Abstract] Typo: 'sPuperior' should be 'superior'.
  2. [Section 3] Several typos: 'continuous-variable QNN based on the variational circuit model introduced in Ref. [27]' is missing a period; 'which user parameterized quantum circuits' should be 'which uses parameterized'; 'In this case, we user a validated approximation' should be 'we use'.
  3. [Section 5.1] The epoch count is written as '10 4 epochs' in the text; please format as 10^4. Also clarify whether the 200 test points are the same for all runs and whether training points are sampled uniformly over [-1,1] or include the discontinuity exactly for the Heaviside case.
  4. [Section 3, Eq. (2)] The quantum neuron notation is compressed; please define all gates (Kerr, displacement, rotation, squeezing) and the roles of parameters explicitly, and note the Kerr gate approximation from Ref. [57] is a key ingredient. The reference is an arXiv preprint; consider adding a more detailed description in the text.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline 'seven orders' advantage is an acknowledged artifact of the unequal layer-for-layer parameter budget; the fairer parameter-matched benchmark supports only a qualified QNN advantage.

  1. self definitional [Abstract; Section 5.2 (Comparison in terms of the Number of Layers, Figure 2a); Section 5.4 (Discussions)]
    "Abstract: 'QNNs excelled at approximating the sinusoidal function, achieving errors up to seven orders of magnitude lower than their classical counterparts.' Section 5.4: 'the constrained architecture of the classical network led to its failure on the sine task, creating the illusion of quantum advantage.'"

    The advertised seven-orders gap is the MSE gap of Figure 2a, obtained by comparing one QNN layer (Eq. 2: five trainable parameters, K(χ)D(α)R(θ2)S(ξ)R(θ1)) against one single-neuron classical layer (Eq. 1: two trainable parameters). The paper explicitly attributes the observed advantage to this parameter-count disparity ('the parameter count scales as 2 L, whereas in QNNs it scales as 5 L') and later states that the constrained classical architecture 'led to its failure on the sine task, creating the illusion of quantum advantage.' Thus the headline quantitative claim is generated by the unequal parameter budget of the layer-for-layer comparison, i.e., by the construction of the comparison itself, and is disowned by the same paper when the comparison is made fair.

  2. self citation load bearing [Section 3 (Quantum Neural Networks), 'Quantum Neuron' definition; Ref. [57] is arXiv:2412.15025 by Alexandre C. Ricardo et al.]
    "A critical component for expressive power is a nonlinear transformation. In this case, we use a validated approximation of Kerr gate ˆK(χ) tailored for trapped-ion systems [57]."

    The QNN's expressive advantage on the sine task rests on this nonlinear gate. The only support cited for the gate's validity is Ref. [57], which shares an author (A. C. Ricardo) with the present paper. The present paper does not re-derive, code-verify, or externally benchmark this gate; it imports the premise from the authors' own prior work. Because the sine-task advantage is produced by this nonlinear element, the central result is load-bearing on a self-citation rather than on an independently verified input.

full rationale

The central circularity is the abstract's 'seven orders of magnitude' claim. It derives from Figure 2a, where one QNN layer has five parameters and a one-neuron classical layer has two. Section 5.4 explicitly calls this an 'illusion of quantum advantage.' The paper's own parameter-matched comparison (Section 5.3, Figure 4a) is fair and independent, and it supports only a qualified, smaller advantage for this particular QNN on sine regression; that part is not circular. A second concern is the Kerr gate: the 'validated approximation' is supported only by a co-authored preprint [57], and since this nonlinearity is what gives the QNN its expressive power, the numerical advantage is ultimately contingent on a self-citation that is not independently checked in this paper. However, the paper does disclose the layer-comparison illusion and provides the fairer benchmark, so the score is 6 rather than 8-10.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central results do not introduce new physical entities. The main dependencies are the hand-chosen simulation cutoff, the selected classical baselines, and the self-cited Kerr approximation, all of which could affect the magnitude of the reported QNN advantage.

free parameters (3)
  • Fock basis cutoff = 30
    Chosen by hand to guarantee simulation fidelity; no convergence study is reported, yet all QNN results depend on the truncation being valid for every state visited during training.
  • Classical network architecture set = Configurations in Table A.1 (e.g., (3) for 10 params; (1,4), (4,1), (1,2,2), (2,2,1) for 15 params, etc.)
    The CNNs tested in the parameter-matched comparison are a small hand-picked set intended to equal QNN parameter counts; the observed QNN advantage depends on this set being representative of classical capability.
  • Training hyperparameters = Adam with learning rate 0.01 for 10^4 epochs
    Fixed across models based on 'prior experience and preliminary tests'; no per-model tuning is reported, so part of the performance gap may reflect suboptimal CNN optimization.
assumptions (3)
  • domain assumption The Kerr gate approximation of ref [57] correctly represents the nonlinear operation needed for QNN expressivity.
    The QNN's nonlinearity, essential to go beyond linear Gaussian operations, is imported from a co-authored reference and is not independently validated in this paper.
  • domain assumption Fock-basis truncation at dimension 30 is a valid approximation for all quantum states encountered during training.
    The paper asserts the amplitudes near cutoff are negligible but provides no measurement of occupation numbers or truncation error.
  • standard math With enough width and depth, the classical networks with tanh, sigmoid, or ReLU can represent the target functions (universal approximation).
    The paper invokes the universal approximation theorem, but the theorem is non-constructive and the specific architectures in Table A.1 may not satisfy its requirements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Assessing the Advantages and Limitations of Quantum Neural Networks in Regression Tasks." pith.science (2026). https://pith.science/paper/WFVKIMLD

@misc{pith2026250900854,
  author       = {Pith},
  title        = {Pith review of: Assessing the Advantages and Limitations of Quantum Neural Networks in Regression Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFVKIMLD}},
  note         = {Machine review of arXiv:2509.00854}
}
read the original abstract

The development of quantum neural networks (QNNs) has attracted considerable attention due to their potential to surpass classical models in certain machine learning tasks. Nonetheless, it remains unclear under which conditions QNNs provide concrete benefits over classical neural networks (CNNs). This study addresses this question by performing both qualitative and quantitative analyses of classical and quantum models applied to regression problems, using two target functions with contrasting properties. Additionally, the work explores the methodological difficulties inherent in making fair comparisons between QNNs and CNNs. The findings reveal a distinct advantage of QNNs in a specific quantum machine learning context. In particular, QNNs excelled at approximating the sinusoidal function, achieving errors up to seven orders of magnitude lower than their classical counterparts. However, their performance was limited in other cases, emphasizing that QNNs are highly effective for certain tasks but not universally sPuperior. These results reinforce the principles of the ``No Free Lunch'' theorem, highlighting that no single model outperforms all others across every problem domain.

Figures

Figures reproduced from arXiv: 2509.00854 by the authors.

Figure 1
Figure 1. (a) Algorithm versus problems addressed by quantum and classical ML; cases [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Prediction error with MSE in logarithmic scale of QNN and CNN, as a function [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Best results obtained after training the algorithms applied to the (a) sine and [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Prediction error results for each algorithm as a function of the number of param [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]
Figure 5
Figure 5. Figure 5: Best results obtained after training the QNN and CNN in the comparison based [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 51 canonical work pages

  1. [1]

    Schalkoff, Artificial Neural Networks , McGraw-Hill, 1997

    R.J. Schalkoff, Artificial Neural Networks , McGraw-Hill, 1997

    work page 1997

  2. [2]

    Pedrycz e S.M

    W. Pedrycz e S.M. Chen, Deep Learning: Algorithms and Applications , Springer International Publishing, 2019

    work page 2019

  3. [3]

    Pramila P. Shinde e Seema Shah, A Review of Machine Learning and Deep Learning Applications , in 2018 Fourth International Conference on Computing Communication Control and Automation (ICCUBEA) , 2018, pp. 1–6. https://doi.org/10.1109/ICCUBEA.2018.8697857 16

    work page arXiv 2018

  4. [4]

    Lei Zhou et al., Application of Deep Learning in Food: A Review , Com- prehensive Reviews in Food Science and Food Safety , vol. 18, no. 6, pp. 1793–1811, 2019. https://doi.org/10.1111/1541-4337.12492

    work page arXiv 2019

  5. [5]

    Samir Khan e Takehisa Yairi, A review on the application of deep learn- ing in system health management , Mechanical Systems and Signal Pro- cessing, vol. 107, pp. 241–265, 2018. https://doi.org/10.1016/j. ymssp.2017.11.024

    doi:10.1016/j 2018

  6. [6]

    Mihalj Bakator e Dragica Radosav, Deep Learning and Medical Diagno- sis: A Review of Literature , Multimodal Technologies and Interaction , vol. 2, no. 3, 2018. https://doi.org/10.3390/mti2030047

    work page doi:10.3390/mti2030047 2018

  7. [7]

    Daniel Hoang e Kevin Wiegratz, Machine learning methods in finance: Recent applications and prospects, European Financial Management, vol. 29, no. 5, pp. 1657–1701, 2023.https://doi.org/10.1111/eufm.12408

    work page doi:10.1111/eufm.12408 2023

  8. [8]

    Alberto Pliego Marug´ an et al., A survey of artificial neural network in wind energy systems , Applied Energy , vol. 228, pp. 1822–1836, 2018. https://doi.org10.1016/j.apenergy.2018.07.084

    work page 2018

Show all 77 references
  1. [9]

    Shiyu Liang e Rayadurgam Srikant, Why Deep Neural Networks for Function Approximation?, in International Conference on Learning Rep- resentations, 2016

    2016

  2. [10]

    Ben Adcock e Nick Dexter, The Gap between Theory and Practice in Function Approximation with Deep Neural Networks , SIAM Journal on Mathematics of Data Science , vol. 3, no. 2, pp. 624–655, 2021. https: //doi.org/10.1137/20M131309X

    2021 doi

  3. [11]

    Dennis Elbr¨ achter et al.,Deep Neural Network Approximation Theory , IEEE Transactions on Information Theory , vol. 67, no. 5, pp. 2581– 2623, 2021. https://doi.org10.1109/TIT.2021.3062161

    2021

  4. [12]

    Gudise e G.K

    V.G. Gudise e G.K. Venayagamoorthy, Comparison of particle swarm optimization and backpropagation as training algorithms for neural net- works, in Proceedings of the 2003 IEEE Swarm Intelligence Symposium. SIS’03 (Cat. No.03EX706) , 2003, pp. 110–117. https://doi.org/10. 1109/...

    2003 arXiv

  5. [13]

    2021, no

    Junxi Zhang e Shiru Qu, Optimization of Backpropagation Neural Net- work under the Adaptive Genetic Algorithm , Complexity, vol. 2021, no. 1, pp. 1718234, 2021. https://doi.org/10.1155/2021/1718234

    2021 doi

  6. [14]

    Leonard e M.A

    J. Leonard e M.A. Kramer, Improvement of the backpropagation al- gorithm for training neural networks , Computers Chemical Engineer- ing, vol. 14, no. 3, pp. 337–341, 1990. https://doi.org/10.1016/ 0098-1354(90)87070-6

    1990

  7. [15]

    3561–3562

    Abhinav Jain et al., Overview and Importance of Data Quality for Ma- chine Learning Tasks, in Proceedings of the 26th ACM SIGKDD Inter- national Conference on Knowledge Discovery Data Mining , 2020, pp. 3561–3562. https://doi.org/10.1145/3394486.3406477

    2020

  8. [16]

    4040–4041

    Nitin Gupta et al., Data Quality for Machine Learning Tasks , in Pro- ceedings of the 27th ACM SIGKDD Conference on Knowledge Discov- ery & Data Mining , 2021, pp. 4040–4041. https://doi.org/10.1145/ 3447548.3470817

    2021

  9. [17]

    Rahul Mishra et al., A survey on deep neural network compression: Challenges, overview, and solutions , arXiv preprint arXiv:2010.03954 , 2020

    2010 arXiv

  10. [18]

    Laith Alzubaidi et al., Review of deep learning: concepts, CNN architectures, challenges, applications, future directions , Journal of Big Data , vol. 8, no. 1, pp. 53, 2021. https://doi.org/10.1186/ s40537-021-00444-8

    2021

  11. [19]

    Amira Abbas et al., The power of quantum neural networks , Nature Computational Science, vol. 1, no. 6, pp. 403–409, 2021. https://doi. org/10.1038/s43588-021-00084-1

    2021 doi

  12. [20]

    Yunseok Kwak et al., Quantum Neural Networks: Concepts, Appli- cations, and Challenges , in 2021 Twelfth International Conference on Ubiquitous and Future Networks (ICUFN) , 2021, pp. 413–416. https: //doi.org/10.1109/ICUFN49451.2021.9528698

    2021

  13. [21]

    Renxin Zhao e Shi Wang, A review of quantum neural networks: meth- ods, models, dilemma , arXiv preprint arXiv:2109.01840 , 2021. 18

    2021 arXiv

  14. [22]

    Marco Cerezo et al., Challenges and opportunities in quantum machine learning, Nature computational science, vol. 2, no. 9, pp. 567–576, 2022. https://doi.org/10.1038/s43588-022-00311-3

    2022 doi

  15. [23]

    Maria Schuld et al., The quest for a quantum neural network , Quantum Information Processing, vol. 13, pp. 2567–2586, 2014. https://doi. org/10.1007/s11128-014-0809-8

    2014 doi

  16. [24]

    Altaisky et al., Quantum neural networks: Current status and prospects for development , Physics of Particles and Nuclei , vol

    Mikhail V. Altaisky et al., Quantum neural networks: Current status and prospects for development , Physics of Particles and Nuclei , vol. 45, pp. 1013–1032, 2014. https://doi.org/10.1134/S1063779614060033

    2014 doi

  17. [25]

    Han Qi et al., The barren plateaus of quantum neural networks: review, taxonomy and trends , Quantum Information Processing, vol. 22, no. 12, pp. 435, 2023. https://doi.org/10.1007/s11128-023-04188-7

    2023 doi

  18. [26]

    Yang Qian et al., The Dilemma of Quantum Neural Networks , IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 4, pp. 5603–5615, 2024. https://doi.org/10.1109/TNNLS.2022. 3208313

    2024 doi

  19. [27]

    Nathan Killoran et al., Continuous-variable quantum neural networks , Phys. Rev. Res., vol. 1, no. 3, pp. 033063, 2019. https://doi.org/10. 1103/PhysRevResearch.1.033063

    2019

  20. [28]

    Jinjing Shi et al., An approach to cryptography based on continuous- variable quantum neural network , Scientific reports, vol. 10, no. 1, pp. 2107, 2020. https://doi.org/10.1038/s41598-020-58928-1

    2020 doi

  21. [29]

    Sophie Choe e Marek Perkowski, Continuous variable quantum mnist classifiers—classical-quantum hybrid quantum neural networks , Journal of Quantum Information Science, vol. 12, no. 2, pp. 37–51, 2022. https: //doi.org/10.4236/jqis.2022.122005

    2022

  22. [30]

    Shikha Bangar et al., Experimentally realizable continuous-variable quantum neural networks , Phys. Rev. A , vol. 108, no. 4, pp. 042414,

  23. [31]

    Shikha Bangar et al., Continuous-variable quantum Boltzmann machine, Quantum Machine Intelligence , vol. 7, no. 1, pp. 1–15, 2025. https: //doi.org/10.1007/s42484-025-00259-3 19

    2025 doi

  24. [32]

    Giorgio Panichi et al., Quantum physics informed neural net- works for multi-variable partial differential equations , arXiv preprint arXiv:2503.12244, 2025

    2025

  25. [33]

    Stefano Markidis, On physics-informed neural networks for quantum computers, Frontiers in Applied Mathematics and Statistics , vol. 8, pp. 1036711, 2022. https://doi.org/10.3389/fams.2022.1036711

    2022

  26. [34]

    Kamila Zaman et al., A Comparative Analysis of Hybrid-Quantum Clas- sical Neural Networks , in Grid, Cloud, and Cluster Computing; Quan- tum Technologies; and Modeling, Simulation and Visualization Methods, Springer Nature Switzerland, 2025, pp. 102–115. https://doi.org/10. 10...

    2025

  27. [35]

    1447–1457

    Maximilian Wendlinger et al., A Comparative Analysis of Adversar- ial Robustness for Quantum and Classical Machine Learning Models , in 2024 IEEE International Conference on Quantum Computing and Engineering (QCE), 2024, pp. 1447–1457. https://doi.org/10.1109/ QCE60285.2024.00171

    2024

  28. [36]

    West et al., Benchmarking adversarially robust quantum machine learning at scale , Phys

    Maxwell T. West et al., Benchmarking adversarially robust quantum machine learning at scale , Phys. Rev. Res. , vol. 5, no. 2, pp. 023186,

  29. [37]

    00203–00208

    Taieba Tasnim et al., Quantum vs Classical: Performance Benchmark- ing of CNN and QCNN in Binary Image Classification , in 2025 IEEE 15th Annual Computing and Communication Workshop and Confer- ence (CCWC) , 2025, pp. 00203–00208. https://doi.org/10.1109/ CCWC62904.2025.10903816

    2025

  30. [38]

    https://doi.org/10.1103/PhysRevResearch.5.023186

    doi

  31. [39]

    classical: A comprehensive bench- mark study for predicting time series with variational quantum machine learning, arXiv preprint arXiv:2504.12416, 2025

    Tobias Fellner et al., Quantum vs. classical: A comprehensive bench- mark study for predicting time series with variational quantum machine learning, arXiv preprint arXiv:2504.12416, 2025. https://doi.org/10. 48550/arXiv.2504.12416 20

    2025 doi

  32. [40]

    Teppei Suzuki et al., Quantum support vector machines for classification and regression on a trapped-ion quantum computer , Quantum Machine Intelligence, vol. 6, no. 1, pp. 31, 2024. https://doi.org/10.1007/ s42484-024-00165-0

    2024

  33. [41]

    Matheus Cammarosano Hidalgo, Comparative Analysis of a Quantum SVM With an Optimized Kernel Versus Classical SVMs , IEEE Access, vol. 13, pp. 82378–82387, 2025. https://doi.org/10.1109/ACCESS. 2025.3567352

    2025

  34. [42]

    Kimleang Kea et al., A Hybrid Quantum-Classical Model for Stock Price Prediction Using Quantum-Enhanced Long Short-Term Memory , En- tropy, vol. 26, no. 11, pp. 954, 2024. https://doi.org10.3390/e26110954

    2024

  35. [43]

    Anujit Ghosh et al., Quantum vs. Classical: A Rigorous Comparative Study on Neural Networks for Advanced Satellite Image Classification , in 2024 International Conference on Trends in Quantum Computing and Emerging Business Technologies , 2024, pp. 1–5. https://doi.org/10. 110...

    2024

  36. [44]

    Md Nadim et al., Quantum Vs. Classical Machine Learning Algorithms for Software Defect Prediction: Challenges and Opportunities , in 2025 IEEE/ACM International Workshop on Quantum Software Engineer- ing (Q-SE) , 2025, pp. 35–42. https://doi.org/10.1109/Q-SE66736. 2025.00012

    2025

  37. [45]

    Sevdanur Gen¸ c,Performance Analysis of Quantum and Classical Ma- chine Learning Models for Feature Selection and Classification of the Diabetes Health Indicators Dataset , in 2024 8th International Artificial Intelligence and Data Processing Symposium (IDAP) , 2024, pp. 1–7. ...

    2024

  38. [46]

    Md Nadim et al., Comparative analysis of quantum and classical sup- port vector classifiers for software bug prediction: an exploratory study , Quantum Machine Intelligence , vol. 7, no. 1, pp. 1–17, 2025. https: //doi.org/10.1007/s42484-025-00236-w

    2025 doi

  39. [47]

    Adam et al., No Free Lunch Theorem: A Review , in Ap- proximation and Optimization : Algorithms, Complexity and Appli- cations, Springer International Publishing, 2019, pp

    Stavros P. Adam et al., No Free Lunch Theorem: A Review , in Ap- proximation and Optimization : Algorithms, Complexity and Appli- cations, Springer International Publishing, 2019, pp. 57-82. https: //doi.org/10.1007/978-3-030-12767-1_5 . 21

    2019 doi

  40. [48]

    Himanshu Gupta et al., Comparative performance analysis of quantum machine learning with deep learning for diabetes prediction , Complex & Intelligent Systems , vol. 8, no. 4, pp. 3073-3087, 2022. https://doi. org/10.1007/s40747-021-00398-7

    2022 doi

  41. [49]

    Wolpert, What Is Important About the No Free Lunch The- orems?, in Black Box Optimization, Machine Learning, and No-Free Lunch Theorems, Springer International Publishing, 2021, pp

    David H. Wolpert, What Is Important About the No Free Lunch The- orems?, in Black Box Optimization, Machine Learning, and No-Free Lunch Theorems, Springer International Publishing, 2021, pp. 373–388. https://doi.org/10.1007/978-3-030-66515-9_13

    2021 doi

  42. [50]

    David G´ omez e Alfonso Rojas, An Empirical Overview of the No Free Lunch Theorem and Its Effect on Real-World Machine Learning Clas- sification, Neural Computation , vol. 28, no. 1, pp. 216–228, 2016. https://doi.org/10.1162/NECO_a_00793

    2016 doi

  43. [51]

    ROBERT HECHT-NIELSEN, III.3 - Theory of the Backpropagation Neural Network , in Neural Networks for Perception , Academic Press, 1992, pp. 65–93. https://doi.org/10.1016/B978-0-12-741252-8. 50010-8

    1992 doi

  44. [52]

    Sazlı, A brief review of feed-forward neural networks, Commu- nications Faculty of Sciences University of Ankara Series A2-A3 Phys- ical Sciences and Engineering, vol

    Murat H. Sazlı, A brief review of feed-forward neural networks, Commu- nications Faculty of Sciences University of Ankara Series A2-A3 Phys- ical Sciences and Engineering, vol. 50, no. 01, 2006. https://doi.org/ 10.1501/commua1-2_0000000026

    2006 doi

  45. [53]

    Atilim Gunes Baydin et al., Automatic differentiation in machine learn- ing: a survey , Journal of machine learning research , vol. 18, no. 153, pp. 1–43, 2018

    2018

  46. [54]

    Nikhil Ketkar e Jojo Moolayil, Automatic Differentiation in Deep Learn- ing, in Deep Learning with Python: Learn Best Practices of Deep Learning Models with PyTorch , Apress, 2021, pp. 133–145. https: //doi.org/10.1007/978-1-4842-5364-9_4

    2021 doi

  47. [55]

    N. Fatima, Enhancing Performance of a Deep Neural Network: A Com- parative Analysis of Optimization Algorithms , ADCAIJ: Advances in Distributed Computing and Artificial Intelligence Journal , vol. 9, no. 2, pp. 79–90, 2020. https://doi.org/10.14201/ADCAIJ2020927990

    2020 doi

  48. [56]

    Derya Soydaner, A comparison of optimization algorithms for deep learning, International Journal of Pattern Recognition and Artificial In- telligence, vol. 34, no. 13, pp. 2052013, 2020. https://doi.org/10. 1142/S0218001420520138

    2020

  49. [57]

    Ricardo et al., Quantum computation over the vibrational modes of a single trapped ion , arXiv preprint arXiv:2412.15025 , 2024

    Alexandre C. Ricardo et al., Quantum computation over the vibrational modes of a single trapped ion , arXiv preprint arXiv:2412.15025 , 2024

    2024 arXiv

  50. [58]

    Augustine, A survey on universal approximation theorems , arXiv preprint arXiv:2407.12895 , 2024

    Midhun T. Augustine, A survey on universal approximation theorems , arXiv preprint arXiv:2407.12895 , 2024. 22

    2024 arXiv

  51. [59]

    Deepak Ranga et al., Quantum Machine Learning: Exploring the Role of Data Encoding Techniques, Challenges, and Future Directions, Math- ematics, vol. 12, no. 21, pp. 3318, 2024. https://doi.org/10.3390/ math12213318

    2024

  52. [60]

    Christian Weedbrook et al., Gaussian quantum information , Rev. Mod. Phys., vol. 84, no. 2, pp. 621–669, 2012. https://doi.org/10.1103/ RevModPhys.84.621

    2012

  53. [61]

    Minati Rath e Hema Date, Quantum data encoding: A comparative anal- ysis of classical-to-quantum mapping techniques and their impact on ma- chine learning accuracy , EPJ Quantum Technology , vol. 11, no. 1, pp. 72, 2024. https://doi.org/10.1140/epjqt/s40507-024-00285-3

    2024 doi

  54. [62]

    Mandaar B. Pande, A Comprehensive Review of Data Encoding Tech- niques for Quantum Machine Learning Problems , in 2024 Second Inter- national Conference on Emerging Trends in Information Technology and Engineering (ICETITE) , 2024, pp. 1–7. https://doi.org/10.1109/ ic-ETITE582...

    2024

  55. [63]

    Laurent Gajny et al., Best L 1 approximation of Heaviside-type functions from Chebyshev and weak-Chebyshev spaces, Numerical Algorithms, vol. 75, pp. 827–843, 2017. https://doi.org10.1007/s11075-016-0222-8

    2017

  56. [64]

    Freek Stulp e Olivier Sigaud, Many regression algorithms, one unified model: A review , Neural Networks , vol. 69, pp. 60–79, 2015. https: //doi.org/10.1016/j.neunet.2015.05.005

    2015 doi

  57. [65]

    Kingma e Jimmy Ba, Adam: A method for stochastic opti- mization, arXiv preprint arXiv:1412.6980 , 2014

    Diederik P. Kingma e Jimmy Ba, Adam: A method for stochastic opti- mization, arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  58. [66]

    Fern´ andez-Delgado et al., An extensive experimental survey of re- gression methods, Neural Networks , vol

    M. Fern´ andez-Delgado et al., An extensive experimental survey of re- gression methods, Neural Networks , vol. 111, pp. 11–34, 2019. https: //doi.org/10.1016/j.neunet.2018.12.010

    2019 doi

  59. [67]

    10.48550/arXiv.2008.05981

    Kanav Anand et al., Black Magic in Deep Learning: How Human Skill Impacts Network Training, arXiv, 2020. 10.48550/arXiv.2008.05981

    2020 doi arXiv

  60. [68]

    Ilemobayo et al., Hyperparameter Tuning in Machine Learn- ing: A Comprehensive Review , Journal of Engineering Research and 23 Reports, vol

    Justus A. Ilemobayo et al., Hyperparameter Tuning in Machine Learn- ing: A Comprehensive Review , Journal of Engineering Research and 23 Reports, vol. 26, no. 6, pp. 388–395, 2024. https://doi.org/10.9734/ jerr/2024/v26i61188

    2024

  61. [69]

    Steve Abel et al., Completely Quantum Neural Networks , Physical Re- view A , vol. 106, no. 2, pp. 022601, 2022. https://doi.org/10.1103/ PhysRevA.106.022601

    2022

  62. [70]

    Ezhov, On Quantum Neural Networks , arXiv preprint arXiv:2104.07106, 2021

    Alexandr A. Ezhov, On Quantum Neural Networks , arXiv preprint arXiv:2104.07106, 2021. https://doi.org/10.48550/arXiv.2104. 07106

    2021 doi arXiv

  63. [71]

    Olivier, Gaussian Process based Model Predictive Control to address uncertain milling circuit dynamics , IF AC-PapersOnLine, vol

    Laurentz E. Olivier, Gaussian Process based Model Predictive Control to address uncertain milling circuit dynamics , IF AC-PapersOnLine, vol. 54, no. 21, pp. 1–6, 2021. https://doi.org/10.1016/j.ifacol.2021. 12.001

    2021 doi

  64. [72]

    Min-Gang Zhou et al., Quantum Neural Network for Quantum Neural Computing, Research, vol. 6, 2023

    2023

  65. [73]

    Chia-Nan Ko e Cheng-Ming Lee, Short-term load forecasting using SVR (support vector regression)-based radial basis function neural net- work with dual extended Kalman filter , Energy, vol. 49, pp. 413–422, 2013.https://doi.org/10.1016/j.energy.2012.11.015

    2013 doi

  66. [74]

    Masini et al., Machine learning advances for time series forecasting, Journal of Economic Surveys , vol

    Ricardo P. Masini et al., Machine learning advances for time series forecasting, Journal of Economic Surveys , vol. 37, no. 1, pp. 76–111, 2023.https://doi.org/10.1111/joes.12429

    2023 doi

  67. [75]

    Matthias Rupp et al., Fast and Accurate Modeling of Molecular At- omization Energies with Machine Learning , Phys. Rev. Lett. , vol. 108, no. 5, pp. 058301, 2012.https://doi.org/10.1103/PhysRevLett.108. 058301. 24

    2012 doi

  68. [76]

    Shihao Gu et al., Empirical Asset Pricing via Machine Learning , The Review of Financial Studies , vol. 33, no. 5, pp. 2223–2273, 2020. https: //doi.org/10.1093/rfs/hhaa009

    2020 doi

  69. [2023]

    https://doi.org/10.1103/PhysRevA.108.042414

    doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.