Pith. sign in

REVIEW 4 major objections 4 minor 51 references

TabularQGAN: A Quantum Generative Model for Tabular Data

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a quantum GAN with a 10–15 qubit generator and 37–104 parameters produces tabular data whose SDMetrics similarity beats the classical CTGAN and CopulaGAN baselines by an average of 8.5%.

desk verdict Nice encoding idea, but the headline claim of beating classical GANs is not supported by the experiments as run. read the letter →

arxiv 2505.22533 v1 pith:I6UVO37E submitted 2025-05-28 cs.LG cs.AIquant-ph

Pallavi Bhardwaj , Caitlin Jones , Lasse Dierich , Aleksandar Vučković This is my paper
classification cs.LGcs.AIquant-ph MSC 68Q1281P68
keywords quantumgenerativeadversarialnetworkstabulardatasynthesisvariationalcircuitsGivensrotationsone-hotencodingsyntheticparameterefficiencySDMetrics
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

The paper tries to establish that a quantum generative adversarial network — with a variational quantum circuit as the generator and a classical neural network as the discriminator — can synthesize realistic tabular data of mixed numerical and categorical features. On four small benchmarks drawn from the MIMIC-III healthcare and Adult Census datasets, it reports that the quantum model's SDMetrics overall similarity score beats CTGAN and CopulaGAN by an average of 8.5% while using roughly 0.072% of their parameters (37–104 versus up to 262,144). If true, this would make tabular data synthesis a plausible early application for quantum computers, relevant to privacy-preserving data sharing and dataset augmentation in healthcare, finance, and enterprise software. The central design bet is that encoding each categorical feature as a one-hot register preserved by particle-conserving Givens rotations keeps the circuit small without forcing the generator to rediscover categorical structure.

What carries the argument

The load-bearing object is the variational quantum circuit generator, whose categorical registers are built from Givens rotations — two-level unitary gates, each acting as a rotation in the subspace spanned by two one-hot bitstrings and conserving the total number of $|1\rangle$ excitations. Because the circuit is confined to the fixed-excitation subspace $\mathcal{H}_k$ of dimension $\binom{n}{k}$, a one-hot pattern such as $|100\rangle$ for a three-category feature stays one-hot under every gate, so the generator explores the categorical distribution without producing invalid bitstrings. Numerical registers use RY rotations, Ising-YY entanglers, and controlled RY gates; cross-register controlled single-excitation gates couple the registers so that pairwise feature correlations can be learned. The paper derives a total gate count of $O(N)$ in the number of qubits, and training differentiates through the circuit with the parameter-shift rule.

What would settle it

Rerun the four benchmarks under a single evaluation protocol — give CTGAN and CopulaGAN the same 3000 epochs and select each model's best checkpoint by the overall metric, or equivalently score TabularQGAN only at its final epoch — and recompute the average margin; if the roughly 8.5% gap shrinks to within the run-to-run spread across the five seeds, the headline claim is falsified. A second check is to record the Hamming weight of each categorical register before and after every cross-register controlled gate during training; if any gate moves an excitation out of a one-hot register, the encoding guarantee that the ansatz natively models categorical structure fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a shallow variational circuit can learn the joint distribution of a heterogeneous tabular dataset when the encoding matches the data's structure. Numerical features are discretized into bins and basis-encoded onto a qubit register; each categorical feature occupies a dedicated register whose one-hot pattern is preserved by Givens-rotation gates that conserve the number of excited qubits, and controlled single-excitation gates across registers are used to learn feature correlations. Benchmarking on Adults Census 10/15 and MIMIC 10/15, the paper reports overall SDMetrics scores — averages of column-wise and pairwise distribution-similarity measures — of 0.949, 0.930, 0.983, and 0.964 for TabularQGAN, against 0.855/0.845, 0.848/0.836, 0.888/0.887, and 0.938/0.924 for CTGAN/CopulaGAN, with the quantum generator using 37–104 parameters versus up to 262,144. The paper also reports downstream scores — the absolute difference in XGBoost accuracy or R² when a model is trained on synthetic versus real data — that are lower than or comparable to the classical baselines on three of four tasks, which it reads as evidence that the generated samples carry usable predictive signal. It presents this as one of the first demonstrations of a quantum generative model handling heterogeneous tabular data without an autoencoding or feature-reduction preprocessing step.

Load-bearing premise

The benchmark comparison is assumed to be fair even though the quantum models were scored at their best checkpoint out of 3000 epochs while the classical models were scored only at their final epoch of 1500; if that asymmetry inflated the quantum scores, the headline 8.5% improvement is not established as stated.

Editorial extensions

If this is right

  • If the benchmark results hold, tabular data synthesis becomes a candidate use case where a small quantum circuit matches or beats purpose-built classical tabular GANs while using three to four orders of magnitude fewer parameters.
  • The paper's own Unique-Row-Index experiment — replacing all categorical registers with a single numerical register — drops the overall metric substantially on the harder dataset, supporting the claim that the one-hot-preserving register structure, not just circuit expressivity, carries the performance.
  • The gate count scales linearly with qubit count, so for fixed register count the circuit cost grows only linearly as features and categories are added.
  • The reported downstream scores (0.006–0.133, lower than the classical baselines on three of four tasks) support the paper's contention that quantum-generated samples can stand in for real data in a supervised learning task without losing much predictive signal.
  • If the claims replicate, heterogeneous tabular data — the dominant data format in enterprise settings — becomes a target for quantum generative modeling alongside the homogeneous image and text data studied previously.

Reading between the lines

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

  • The paper's own evaluation protocol — best epoch of 3000 for the quantum model versus final epoch of 1500 for the classical models — makes the reported 8.5% margin an upper bound; a symmetric protocol could shrink or erase it. This is an inference from the paper's text, not a claim the paper makes.
  • Because the alternative encodings the paper tried (binary and Unique-Row-Index) change the qubit count and circuit topology at the same time as the encoding, the comparison does not isolate whether the Givens subspace restriction is the true cause of the performance; a controlled ablation holding qubit count fixed would settle the mechanism.
  • The generalization metrics (overlap fraction and downstream score) are computed only on a subsample of configurations — the best and worst ten per model — so the generalization claims rest on thinner evidence than the similarity headline; a complete sweep would be needed to confirm them.
  • If the one-hot wiring guarantee and the benchmark comparison both survive scrutiny, the natural next test is noisy hardware at larger scale; the paper's own discussion of barren plateaus and classical simulability (its references 48–51) already cautions that the parameter-efficiency edge may not persist at the qubit counts needed for enterprise-sized datasets.
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 introduces TabularQGAN, a hybrid quantum-classical generative adversarial network in which the generator is a variational quantum circuit with two proposed topologies (Boolean and non-Boolean) for encoding numerical and categorical features. Categorical features are one-hot encoded and manipulated with Givens rotations that preserve the one-hot subspace, while numerical features are binned and binary encoded. The model is benchmarked on four small configurations drawn from MIMIC-III and Adult Census against CTGAN and CopulaGAN, using the SDMetrics overall similarity score, an overlap fraction, and a downstream predictive score. The central claim is that TabularQGAN outperforms the classical baselines by an average of 8.5% on the SDMetrics overall metric while using about 0.072% of the classical parameter count.

Significance. If the central claim were established, this would be a noteworthy result for quantum generative modeling: a 10--15 qubit generator with tens of parameters would beat specialized classical tabular GANs with tens of thousands of parameters on standard SDMetrics similarity measures. The architecture idea of using particle-conserving Givens rotations to respect one-hot encodings is interesting and the gate-count analysis showing linear scaling is a strength. The use of third-party baselines (CTGAN, CopulaGAN) and the SDMetrics library is appropriate, and the authors are transparent about several limitations. However, the headline comparison is compromised by a documented evaluation asymmetry, and there are also reporting inconsistencies in the parameter counts. The current evidence does not support the 8.5% claim as stated, although the issue is fixable with a matched evaluation protocol.

major comments (4)
  1. [Section 3.2, Table 3] The evaluation is not apples-to-apples. The paper states: "In addition to the quantum model, the best epoch of the 3000 epochs was selected (this was not possible for the classical models, as only the parameters for the final epoch were accessible)." This means TabularQGAN is scored at its best checkpoint over 3000 epochs, while CTGAN and CopulaGAN are scored at the final epoch of 1500 epochs. Two simultaneous asymmetries favor the quantum model: checkpoint selection (best-over-time versus terminal) and training budget (3000 versus 1500 epochs). Table 2 also shows different learning-rate ranges for quantum (0.05--0.2) and classical (0.001--0.05) models. Since Table 3 reports single best-configuration values without variance or seed-level detail, the average 8.5% improvement in the abstract is not supported as stated. Please rerun the comparison with a matched protocol, e.g., same number of epochs and the same checkpoint-selection rule for all models, or report learning curves and best-epoch values for the classical models as well.
  2. [Section 2.3, Figure 1] The exact wiring of the cross-register controlled single-excitation gates is not specified. The text says that "entanglement between different registers is established by controlled single-excitation gates" and that this preserves one-hot structure, but it does not define which qubit is the control, which register is the target, or how the gates are ordered. Since the one-hot-preservation argument and the correctness of the proposed architecture depend on this wiring, please provide a formal gate-level definition (including controls and targets) or a clearly labeled circuit diagram. As written, the description is too underspecified to reproduce the circuit unambiguously.
  3. [Table 3, Adults Census 15 row] The parameter count reported for CopulaGAN on Adults Census 15 is 60, which is inconsistent with the other CopulaGAN entries (65,536 and 131,072) and with CTGAN's 131,072 for the same dataset. This entry also affects the abstract's "0.072% of the parameters" claim. Please clarify whether this is a typo and recompute the parameter-ratio claim with the corrected value.
  4. [Section 2.6] The overlap fraction formula is written as "1− (UR −U S/UR)", which is dimensionally inconsistent and does not match the prose definition. The text defines UR and US as counts of unique rows in the training and synthetic data, respectively, but an overlap fraction requires the intersection of the two sets. As written, the formula cannot be evaluated from the stated quantities. Please restate the definition unambiguously; this affects a reported generalization metric, though not the headline similarity claim.
minor comments (4)
  1. [Appendix A.4, Table 4] The layer width for CopulaGAN on Adults Census 15 is listed as "Data Width", but this quantity is not defined anywhere in the paper. Please define it or replace it with the numeric value.
  2. [Throughout] The dataset name is inconsistently written as "Adult Census", "Adults Census", and "Adult census" across the abstract, tables, and text. Please standardize the naming.
  3. [Reference [22]] Reference [22] is cited with an arXiv identifier but without a venue or year in the bibliography entry style used elsewhere. Please complete the reference.
  4. [Section 3.2] The hyperparameter search is said to repeat each configuration five times with different random seeds, but Table 3 and the figures report only the best value without variance or seed-level information. Reporting distributions (e.g., mean and standard deviation over seeds) would strengthen the empirical claims.

Circularity Check

1 steps flagged · score 6.0 of 10

TabularQGAN's headline 8.5% advantage is partly constructed by scoring the quantum model at its best epoch over 3,000 epochs while CTGAN/CopulaGAN are scored at the final epoch of 1,500; the reported comparison is a max-vs-terminal selection, not an independent prediction.

  1. fitted input called prediction [Section 3.2 (Hyperparameter Optimization); results in Table 3 and Abstract]
    "For each model, the best hyperparameter settings were selected with respect to the overall metric defined in section 2.6, and can be found in Appendix A.4. In addition to the quantum model, the best epoch of the 3000 epochs was selected (this was not possible for the classical models, as only the parameters for the final epoch were accessible)."

    The reported TabularQGAN overall metric in Table 3 is not the score produced by a fixed trained model; it is the maximum of the same SDMetrics overall metric over 3000 epoch evaluations. The headline 'outperforms classical models by an average of 8.5%' is then computed by comparing this selected maximum against CTGAN/CopulaGAN scores taken at the final epoch of a 1500-epoch run. Because a maximum over epochs is by construction at least as large as any single-epoch value, part of the claimed advantage is guaranteed by the selection protocol rather than by the quantum generator's learned distribution.

full rationale

The actual generative model derivation is largely self-contained: the Givens-rotation decomposition (Definition 1, Eqs. 4-5) is attributed to external work [39], the parameter-shift rule to external references [41,42], and the evaluation metrics to the external SDMetrics library [45]. The classical baselines CTGAN and CopulaGAN are third-party implementations [33,34], so the benchmark is anchored outside this paper's own definitions. The self-citation [22] is used only as architectural lineage ('an adaptation of the model presented in [22]'); it does not import a uniqueness theorem or forbid alternative encodings, so it is not load-bearing circularity. The one genuinely constructed step is the evaluation protocol in Section 3.2: the quantum model's score is selected as the best epoch over 3000 epochs on the very metric that is then reported, while the classical models are scored at their final epoch of 1500. This makes the headline 8.5% advantage partially a property of the selection rule (a maximum versus a terminal value), not purely an independent empirical finding. Because the central claim is not fully determined by the model's equations but is partly manufactured by checkpoint selection, the appropriate score is 6 rather than 0-2. A matched evaluation (same epoch budget, same checkpoint rule) would likely remove this circular component.

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

No new physical entities, particles, forces, or conserved quantities are introduced; the 'novel ansatz' is a circuit architecture, not a new entity. The free parameters are searchable hyperparameters and a hand-chosen discretization budget, none of which are fitted physical constants. The axioms are standard GAN/quantum-circuit assumptions plus the unverified wiring assumption about cross-register gates.

free parameters (5)
  • Circuit depth d = 1 to 4; best values: 4 for Adults 10, Adults 15, MIMIC 10; 1 for MIMIC 15
    Grid-searched hyperparameter chosen on the test metric (SDMetrics overall similarity), coupling hyperparameter selection to the evaluation metric.
  • Generator and discriminator learning rates = 0.05 to 0.2 for quantum; 0.001 to 0.05 for classical
    Different ranges per model type, chosen after 'initial experiments' (Section 3.2), which weakens comparability of the best configurations.
  • Batch size fraction = 10% or 20%
    Grid-searched; same range for all model types.
  • Numerical binning budget (qubits per feature) = 5 qubits per numerical feature in the reported runs (32 equal-width bins)
    Hand-chosen discretization; bin edges are computed from training-data min and max (Section 2.2), so the encoding uses training-set summary statistics.
  • Number of training epochs = 3000 for quantum, 1500 for classical
    The quantum model was given twice the epochs and best-epoch checkpoint selection; the classical models were limited to final-epoch weights, a protocol asymmetry that favors the quantum claim.
assumptions (5)
  • standard math Givens rotations form a universal gate set for preparing any state in the fixed-excitation (one-hot) subspace.
    Invoked in Section 2.1 and Definition 1 to justify that categorical registers can represent arbitrary one-hot distributions via single-excitation gates; the result is cited from [39].
  • domain assumption The discriminator loss provides a usable training signal for the quantum generator via the parameter-shift rule.
    Equations 13 to 15 assume standard GAN loss convergence and that single-shot measurement gradients remain informative; shot noise is not analyzed.
  • domain assumption The SDMetrics overall similarity score is a valid, comparable proxy for synthetic tabular data quality.
    Used as the selection metric and the headline claim (Section 2.6, Table 3); its two components, univariate marginals and pairwise correlations, ignore higher-order dependencies.
  • domain assumption Uniform equal-width binning of numerical features preserves enough information for the benchmark task.
    Section 2.2 and the limitations in Section 4 acknowledge discretization; all models were evaluated on the discretized data, so the comparison operates on a simplified task.
  • domain assumption The cross-register controlled single-excitation gates preserve each categorical register's one-hot subspace.
    Section 2.3 states that 'the entanglement between different registers is established by controlled single-excitation gates'; the exact qubit wiring is only shown in Figure 1, which is not legible in the text, so the one-hot preservation is asserted rather than demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TabularQGAN: A Quantum Generative Model for Tabular Data." pith.science (2026). https://pith.science/paper/I6UVO37E

@misc{pith2026250522533,
  author       = {Pith},
  title        = {Pith review of: TabularQGAN: A Quantum Generative Model for Tabular Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I6UVO37E}},
  note         = {Machine review of arXiv:2505.22533}
}
read the original abstract

In this paper, we introduce a novel quantum generative model for synthesizing tabular data. Synthetic data is valuable in scenarios where real-world data is scarce or private, it can be used to augment or replace existing datasets. Real-world enterprise data is predominantly tabular and heterogeneous, often comprising a mixture of categorical and numerical features, making it highly relevant across various industries such as healthcare, finance, and software. We propose a quantum generative adversarial network architecture with flexible data encoding and a novel quantum circuit ansatz to effectively model tabular data. The proposed approach is tested on the MIMIC III healthcare and Adult Census datasets, with extensive benchmarking against leading classical models, CTGAN, and CopulaGAN. Experimental results demonstrate that our quantum model outperforms classical models by an average of 8.5% with respect to an overall similarity score from SDMetrics, while using only 0.072% of the parameters of the classical models. Additionally, we evaluate the generalization capabilities of the models using two custom-designed metrics that demonstrate the ability of the proposed quantum model to generate useful and novel samples. To our knowledge, this is one of the first demonstrations of a successful quantum generative model for handling tabular data, indicating that this task could be well-suited to quantum computers.

Figures

Figures reproduced from arXiv: 2505.22533 by the authors.

Figure 1
Figure 1. Non-Boolean circuit design (a) and Boolean circuit design (b) for a [n5,c3,c2] register. The model [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Schematic diagram of TabularQGAN training. In Step 1, either a batch of training data or a batch [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Plot of the overall metric for each hyperparameter configuration for each dataset. The spread of the [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of the overall metric for each hyperparameter configuration on the Adults Census [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Plot showing the distribution of the overall metric value for each data set with the two different [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Plot showing the overlap fraction metric for different models and data types. Only a selected subset [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: (a) Effect of circuit depth on performance for quantum models. (b) Effect of the number of layers [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: (a) Downstream Score against Overall Metric for predicting categorical Variables (b) Downstream [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 34 canonical work pages

  1. [1]

    Quantum error correction below the surface code threshold.Nature, 2024

    Rajeev Acharya, Dmitry A Abanin, Laleh Aghababaie-Beni, Igor Aleiner, Trond I Andersen, Markus Ansmann, Frank Arute, Kunal Arya, Abraham Asfaw, Nikita Astrakhantsev, et al. Quantum error correction below the surface code threshold.Nature, 2024

    work page 2024

  2. [2]

    Evered, Alexandra A

    Dolev Bluvstein, Simon J. Evered, Alexandra A. Geim, Sophie H. Li, Hengyun Zhou, Tom Manovitz, Sepehr Ebadi, Madelyn Cain, Marcin Kalinowski, Dominik Hangleiter, J. Pablo Bonilla Ataides, Nishad Maskara, Iris Cong, Xun Gao, Pedro Sales Rodriguez, Thomas Karolyshyn, Giulia Semeghini, Michael J. Gullans, Markus Greiner, Vladan Vuletić, and Mikhail D. Lukin....

    work page 2023

  3. [3]

    Evidence for the utility of quantum computing before fault tolerance.Nature, 618(7965):500–505, June 2023

    Youngseok Kim, Andrew Eddins, Sajant Anand, Ken Xuan Wei, Ewout van den Berg, Sami Rosenblatt, Hasan Nayfeh, Yantao Wu, Michael Zaletel, Kristan Temme, and Abhinav Kandala. Evidence for the utility of quantum computing before fault tolerance.Nature, 618(7965):500–505, June 2023. Number: 7965 Publisher: Nature Publishing Group

    work page 2023

  4. [4]

    Optimization by decoded quantum interferometry.arXiv preprint arXiv:2408.08292, 2024

    Stephen P Jordan, Noah Shutty, Mary Wootters, Adam Zalcman, Alexander Schmidhuber, Robbie King, Sergei V Isakov, and Ryan Babbush. Optimization by decoded quantum interferometry.arXiv preprint arXiv:2408.08292, 2024

    arXiv 2024

  5. [5]

    Peter W. Shor. Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer.SIAM Journal on Computing, 26(5):1484–1509, 1997

    1997

  6. [6]

    Lov K. Grover. A fast quantum mechanical algorithm for database search. InProceedings of the Twenty-Eighth Annual ACM Symposium on Theory of Computing, STOC ’96, page 212–219, New York, NY, USA, 1996. Association for Computing Machinery

    work page 1996

  7. [7]

    Quantum measurements and the abelian stabilizer problem.arXiv preprint quant-ph/9511026, 1995

    A Yu Kitaev. Quantum measurements and the abelian stabilizer problem.arXiv preprint quant-ph/9511026, 1995

    arXiv 1995

  8. [8]

    Springer, 2018

    Maria Schuld and Francesco Petruccione.Supervised learning with quantum computers, volume 17. Springer, 2018

    2018

Show all 51 references
  1. [9]

    Quantum observables for continuous control of the quantum approximate optimization algorithm via reinforcement learning.arXiv preprint arXiv:1911.09682, 2019

    Artur Garcia-Saez and Jordi Riu. Quantum observables for continuous control of the quantum approximate optimization algorithm via reinforcement learning.arXiv preprint arXiv:1911.09682, 2019

    1911 arXiv

  2. [10]

    Quantum vision transformers.arXiv preprint arXiv:2209.08167, 2022

    El Amine Cherrat, Iordanis Kerenidis, Natansh Mathur, Jonas Landman, Martin Strahm, and Yun Yvonna Li. Quantum vision transformers.arXiv preprint arXiv:2209.08167, 2022

    2022 arXiv

  3. [11]

    Fast quantum algorithm for attention computation.arXiv preprint arXiv:2307.08045, 2023

    Yeqi Gao, Zhao Song, Xin Yang, and Ruizhe Zhang. Fast quantum algorithm for attention computation.arXiv preprint arXiv:2307.08045, 2023

    2023

  4. [12]

    Quantum machine learning for image classification.Machine Learning: Science and Technology, 5(1):015040, 2024

    Arsenii Senokosov, Alexandr Sedykh, Asel Sagingalieva, Basil Kyriacou, and Alexey Melnikov. Quantum machine learning for image classification.Machine Learning: Science and Technology, 5(1):015040, 2024

    2024

  5. [13]

    On the Quantum versus Classical Learnability of Discrete Distributions.Quantum, 5:417, March 2021

    Ryan Sweke, Jean-Pierre Seifert, Dominik Hangleiter, and Jens Eisert. On the Quantum versus Classical Learnability of Discrete Distributions.Quantum, 5:417, March 2021

    2021

  6. [14]

    A comprehensive review of quantum machine learning: from nisq to fault tolerance

    Yunfei Wang and Junyu Liu. A comprehensive review of quantum machine learning: from nisq to fault tolerance. Reports on Progress in Physics, 2024. 11

    2024

  7. [15]

    Better than classical? the subtle art of benchmarking quantum machine learning models.arXiv preprint arXiv:2403.07059, 2024

    Joseph Bowles, Shahnawaz Ahmed, and Maria Schuld. Better than classical? the subtle art of benchmarking quantum machine learning models.arXiv preprint arXiv:2403.07059, 2024

    2024 arXiv

  8. [16]

    The power of quantum neural networks.Nature Computational Science, 1(6):403–409, 2021

    Amira Abbas, David Sutter, Christa Zoufal, Aurélien Lucchi, Alessio Figalli, and Stefan Woerner. The power of quantum neural networks.Nature Computational Science, 1(6):403–409, 2021

    2021

  9. [17]

    Information perspective to probabilistic modeling: Boltzmann machines versus born machines.Entropy, 20(8), 2018

    Song Cheng, Jing Chen, and Lei Wang. Information perspective to probabilistic modeling: Boltzmann machines versus born machines.Entropy, 20(8), 2018

    2018

  10. [18]

    A generative modeling approach for benchmarking and training shallow quantum circuits.npj Quantum Information, 5(1):45, 2019

    Marcello Benedetti, Delfina Garcia-Pintos, Oscar Perdomo, Vicente Leyton-Ortega, Yunseong Nam, and Alejandro Perdomo-Ortiz. A generative modeling approach for benchmarking and training shallow quantum circuits.npj Quantum Information, 5(1):45, 2019

    2019

  11. [19]

    Quantum generative adversarial networks.Phys

    Pierre-Luc Dallaire-Demers and Nathan Killoran. Quantum generative adversarial networks.Phys. Rev. A, 98:012324, Jul 2018

    2018

  12. [20]

    Quantum generative adversarial learning.Physical Review Letters, 121(4), July 2018

    Seth Lloyd and Christian Weedbrook. Quantum generative adversarial learning.Physical Review Letters, 121(4), July 2018. arXiv:1804.09139 [quant-ph]

    2018 arXiv

  13. [21]

    D. Zhu, N. M. Linke, M. Benedetti, K. A. Landsman, N. H. Nguyen, C. H. Alderete, A. Perdomo-Ortiz, N. Korda, A. Garfoot, C. Brecque, L. Egan, O. Perdomo, and C. Monroe. Training of quantum circuits on a hybrid quantum computer.Science Advances, 5(10):eaaw9918, 2019

    2019

  14. [22]

    A performance characterization of quantum generative models

    Carlos A Riofrío, Oliver Mitevski, Caitlin Jones, Florian Krellner, Aleksandar Vučković, Joseph Doetsch, Johannes Klepsch, Thomas Ehmer, and Andre Luckow. A performance characterization of quantum generative models. arXiv e-prints, pages arXiv–2301, 2023

    2023

  15. [23]

    Gao, Z.-Y

    X. Gao, Z.-Y. Zhang, and L.-M. Duan. A quantum machine learning algorithm based on generative models. Science Advances, 4(12):eaat9004, 2018

    2018

  16. [24]

    Generative adversarial networks-based synthetic pmu data creation for improved event classification.IEEE Open Access Journal of Power and Energy, 8:68–76, 2021

    Xiangtian Zheng, Bin Wang, Dileep Kalathil, and Le Xie. Generative adversarial networks-based synthetic pmu data creation for improved event classification.IEEE Open Access Journal of Power and Energy, 8:68–76, 2021

    2021

  17. [25]

    Deep generative molecular design reshapes drug discovery.Cell Reports Medicine, 3(12), 2022

    Xiangxiang Zeng, Fei Wang, Yuan Luo, Seung-gu Kang, Jian Tang, Felice C Lightstone, Evandro F Fang, Wendy Cornell, Ruth Nussinov, and Feixiong Cheng. Deep generative molecular design reshapes drug discovery.Cell Reports Medicine, 3(12), 2022

    2022

  18. [26]

    Security and privacy of electronic health records: Concerns and challenges

    Ismail Keshta and Ammar Odeh. Security and privacy of electronic health records: Concerns and challenges. Egyptian Informatics Journal, 22(2):177–183, 2021

    2021

  19. [27]

    Privacy-preserving anomaly detection using synthetic data

    Rudolf Mayer, Markus Hittmeir, and Andreas Ekelhart. Privacy-preserving anomaly detection using synthetic data. InIFIP Annual Conference on Data and Applications Security and Privacy, pages 195–207. Springer, 2020

    2020

  20. [28]

    Cao Xiao, Edward Choi, and Jimeng Sun. Opportunities and challenges in developing deep learning models using electronic health records data: a systematic review.Journal of the American Medical Informatics Association : JAMIA, 25(10):1419–1428, June 2018

    2018

  21. [29]

    A review of Generative Adversarial Networks for Electronic Health Records: applications, evaluation measures and data sources, December 2022

    Ghadeer Ghosheh, Jin Li, and Tingting Zhu. A review of Generative Adversarial Networks for Electronic Health Records: applications, evaluation measures and data sources, December 2022. arXiv:2203.07018 [cs]

    2022 arXiv

  22. [30]

    Flöther, George Seegan, Nathan D

    Zoran Krunic, Frederik F. Flöther, George Seegan, Nathan D. Earnest-Noble, and Omar Shehab. Quantum Kernels for Real-World Predictions Based on Electronic Health Records.IEEE Transactions on Quantum Engineering, 3:1–11, 2022. Conference Name: IEEE Transactions on Quantum Engineering

    2022

  23. [31]

    Stewart, and Jimeng Sun

    Edward Choi, Siddharth Biswal, Bradley Malin, Jon Duke, Walter F. Stewart, and Jimeng Sun. Generating Multi-label Discrete Patient Records using Generative Adversarial Networks. InProceedings of the 2nd Machine Learning for Healthcare Conference, pages 286–305. PMLR, November ...

    2017

  24. [32]

    Cairns, Jingsong Li, and Tingting Zhu

    Jin Li, Benjamin J. Cairns, Jingsong Li, and Tingting Zhu. Generating synthetic mixed-type longitudinal electronic health records for artificial intelligent applications.npj Digital Medicine, 6(1):1–18, May 2023. Publisher: Nature Publishing Group

    2023

  25. [33]

    Modeling tabular data using conditional gan.Advances in neural information processing systems, 32, 2019

    Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Modeling tabular data using conditional gan.Advances in neural information processing systems, 32, 2019

    2019

  26. [34]

    The synthetic data vault

    Neha Patki, Roy Wedge, and Kalyan Veeramachaneni. The synthetic data vault. InIEEE International Conference on Data Science and Advanced Analytics (DSAA), pages 399–410, Oct 2016. Business Source License 1.1

    2016

  27. [35]

    Mimic-iii, a freely accessible critical care database.Scientific data, 3(1):1–9, 2016

    Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G Mark. Mimic-iii, a freely accessible critical care database.Scientific data, 3(1):1–9, 2016

    2016

  28. [36]

    & Mark R

    T. & Mark R. Johnson A., Pollard. Mimic-iii clinical database, 2016. 12

    2016

  29. [37]

    Barry Becker and Ronny Kohavi. Adult. UCI Machine Learning Repository, 1996. DOI: https://doi.org/10.24432/C5XW20

    1996 doi

  30. [38]

    Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C

    M. Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C. Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R. McClean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio, and Patrick J. Coles. Variational quantum algorithms. Nature Reviews Physics, 3(9):625–644, 2021

    2021

  31. [39]

    Universal quantum circuits for quantum chemistry.Quantum, 6:742, June 2022

    Juan Miguel Arrazola, Olivia Di Matteo, Nicolás Quesada, Soran Jahangiri, Alain Delgado, and Nathan Killoran. Universal quantum circuits for quantum chemistry.Quantum, 6:742, June 2022

    2022

  32. [40]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks, 2014

    2014

  33. [41]

    Gavin E. Crooks. Gradients of parameterized quantum gates using the parameter-shift rule and gate decomposition, 2019

    2019

  34. [42]

    Evaluating analytic gradients on quantum hardware.Physical Review A, 99(3), March 2019

    Maria Schuld, Ville Bergholm, Christian Gogolin, Josh Izaac, and Nathan Killoran. Evaluating analytic gradients on quantum hardware.Physical Review A, 99(3), March 2019

    2019

  35. [43]

    Parallel evaluation of quantum algorithms for computational fluid dynamics

    René Steijl and George N Barakos. Parallel evaluation of quantum algorithms for computational fluid dynamics. Computers & Fluids, 173:22–28, 2018

    2018

  36. [44]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K.Q. Weinberger, editors,Advances in Neural Information Proce...

    2014

  37. [45]

    DataCebo

    Inc. DataCebo. Synthetic data metrics, 10 2023. Version 0.12.0, MIT License

    2023

  38. [46]

    Xgboost: A scalable tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, page 785–794. ACM, August 2016

    2016

  39. [47]

    Alistair E. W. Johnson, Tom J. Pollard, Lu Shen, Li-wei H. Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G. Mark. MIMIC-III, a freely accessible critical care database.Scientific Data, 3(1):160035, May 2016. Number: 1 Pu...

    2016

  40. [48]

    Barren plateaus in quantum neural network training landscapes.Nature communications, 9(1):4812, 2018

    Jarrod R McClean, Sergio Boixo, Vadim N Smelyanskiy, Ryan Babbush, and Hartmut Neven. Barren plateaus in quantum neural network training landscapes.Nature communications, 9(1):4812, 2018

    2018

  41. [49]

    Theoretical guarantees for permutation-equivariant quantum neural networks.npj Quantum Information, 10(1):12, 2024

    Louis Schatzki, Martin Larocca, Quynh T Nguyen, Frederic Sauvage, and Marco Cerezo. Theoretical guarantees for permutation-equivariant quantum neural networks.npj Quantum Information, 10(1):12, 2024

    2024

  42. [50]

    Quantum convolutional neural networks are (effectively) classically simulable.arXiv preprint arXiv:2408.12739, 2024

    Pablo Bermejo, Paolo Braccia, Manuel S Rudolph, Zoë Holmes, Lukasz Cincio, and M Cerezo. Quantum convolutional neural networks are (effectively) classically simulable.arXiv preprint arXiv:2408.12739, 2024

    2024 arXiv

  43. [51]

    <=50K" and

    Marco Cerezo, Martín Larocca, Diego Garc’ia-Mart’in, N. L. Diaz, Paolo Braccia, Enrico Fontana, Manuel S. Rudolph, Pablo Bermejo, Aroosa Ijaz, Supanut Thanasilp, Eric R Anschuetz, and Zoe Holmes. Does provable absence of barren plateaus imply classical simulability? or, why we...

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.