Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Quantum Latent Diffusion Models

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A quantum latent diffusion model whose denoiser is made of three variational circuits on 10 qubits generates better images than a comparable classical model, with less data and fewer epochs.

desk verdict A concrete three-VQC latent diffusion architecture that is worth a serious look, but the quantum-advantage claim rests on a classical baseline missing the QLDM's skip connection and temporal branch. read the letter →

arxiv 2501.11174 v1 pith:BIHJSJJO submitted 2025-01-19 quant-ph

Francesca De Falco , Andrea Ceschini , Alessandro Sebastianelli , Bertrand Le Saux , Massimo Panella This is my paper · ORCID
classification quant-ph
keywords quantumlatentdiffusionvariationalcircuitsmodelsgenerativespaceimagegenerationFréchetInceptionDistancefew-shotlearning
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 proposes a quantum latent diffusion model (QLDM) that inserts variational quantum circuits into the denoising step of a diffusion model, after a classical autoencoder compresses images to 10-dimensional latent vectors. The central claim is that this quantum denoiser produces better images than a classical fully connected denoiser with a comparable number of parameters, as measured mainly by Fréchet Inception Distance, on MNIST, Fashion MNIST, and a two-class EuroSAT subset. The paper also claims the quantum version learns faster, with a training loss about an order of magnitude lower from the first iterations and better FID by the tenth epoch, and that it needs much less data, with 20–40% of the dataset enough to match or beat the classical model trained on everything. If these claims hold, quantum circuits would be a practical alternative for the parametric core of latent diffusion models rather than a separate theoretical idea.

What carries the argument

The load-bearing object is a three-circuit variational quantum denoiser operating on 10 qubits. After a classical autoencoder maps an image to a latent vector $\mathbf{x}$, the forward diffusion process adds Gaussian noise in the usual way; the reverse step then estimates the noise $\boldsymbol{\epsilon}_\theta(\mathbf{x}_t, t)$ with three parameterized circuits: one angle-encodes $\mathbf{x}$, one processes a sinusoidal positional encoding of the time step $t$, and the third takes the sum of their outputs. Expectation values measured from the third circuit are added to $\mathbf{x}$ through a skip connection, giving the residual-learning structure. The ansatz is built from single-qubit rotations around the $X$ and $Z$ axes interleaved with fixed entangling gates, at depth 3 or 4, with measurement in the $Z$ or $X$ basis; parameter counts are 120, 270, or 360, comparable to the 330-parameter classical baseline.

What would settle it

Give the classical baseline the same three-branch ResNet-style structure, with a separate branch for the time embedding, the same skip connection, and a comparable parameter budget; if its FID matches the QLDM's on MNIST and EuroSAT, the reported quantum advantage is explained by architecture rather than quantum mechanics.

Watch

Extended reading notes

Core claim

The paper's central claim is that a diffusion model can be made quantum without leaving image space: a standard autoencoder maps images into 10-dimensional latent vectors, and the denoiser that inverts the forward noising process is a stack of three variational quantum circuits (VQCs), one acting on the latent vector, one acting on a sinusoidal encoding of the diffusion time step, and one combining both, with a skip connection added to the output. With 10 qubits, angle encoding, ansätze built from rotations around the $X$ and $Z$ axes at depth 3 or 4, and measurements in the $Z$ or $X$ basis, the model estimates the noise $\boldsymbol{\epsilon}_\theta(\mathbf{x}_t, t)$ that the reverse diffusion step needs. The authors report that this quantum denoiser outperforms a classical fully connected denoiser with 330 parameters: FID drops from 44.36 to 40.40 on MNIST, from 90.37 to 84.89 on Fashion MNIST, and from 30.56 to 20.11 on a two-class EuroSAT subset. They also report that quantum models reach these values with 20–40% of the training data and stabilize their loss an order of magnitude earlier than the classical model.

Load-bearing premise

The comparison assumes the 330-parameter fully connected classical network is the fair counterpart to the QLDM, so the better quantum metrics are credited to the quantum circuits rather than to the three-branch ResNet-style structure, skip connection, and separate temporal processing that the QLDM also has.

Editorial extensions

If this is right

  • A quantum denoiser with 120–360 parameters can match or beat a 330-parameter classical denoiser on FID for standard image datasets, so parameter count alone is not the bottleneck for quantum generative models.
  • Because quantum models reach good FID with 20–40% of the training data, latent-space quantum diffusion is a candidate regime for few-shot image generation.
  • The FID advantage appears by epoch 10 and the training loss is roughly an order of magnitude lower from the first iterations, so quantum denoisers converge faster in epoch count.
  • Raising the number of qubits from 8 to 12 improves FID monotonically on MNIST, which the paper reads as evidence that a richer quantum latent representation improves generation quality.
  • The shallow circuit depth keeps the design within reach of noisy intermediate-scale hardware, provided noise can be mitigated.

Reading between the lines

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

  • Because the classical baseline is a simple fully connected network, the reported advantage may be partly attributable to the QLDM's three-branch ResNet-style architecture and separate temporal encoding; an immediate comparison would give the classical model the same three-branch structure with a matched parameter budget.
  • If the few-shot advantage survives that comparison, latent-space quantum diffusion becomes a plausible tool for domains with scarce imagery, such as satellite scenes, where only a handful of classes are available.
  • The paper simulates noiseless circuits; on real hardware, gate and measurement errors could erase the FID gain, so the decisive next experiment is a hardware run with error mitigation on the same 10-qubit circuits.
  • Since the latent space is fixed at dimension 10 to keep angle encoding cheap, scaling to higher-resolution datasets will require either more qubits, a larger latent dimension with more complex encoding, or a return to amplitude encoding with its exponential measurement cost.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a Quantum Latent Diffusion Model (QLDM) in which a classical autoencoder maps images into a 10-dimensional latent space and three variational quantum circuits (VQCs) implement the denoising function of a diffusion model: one processes the latent vector, one processes a temporal embedding, and one processes their sum, with a residual skip connection. The authors compare this quantum architecture against a classical fully-connected denoiser with a comparable parameter count, evaluating generated images on MNIST, Fashion MNIST, and a two-class EuroSAT subset using FID, KID, and IS. They report that several quantum variants outperform the classical baseline on FID, converge in fewer epochs, and achieve better FID when trained on small fractions of the data, concluding that the quantum version exhibits an advantage. The paper also includes comparisons with classical and quantum GANs, hyperparameter sensitivity analyses, and a discussion of noise mitigation.

Significance. If the central claim were established, a 10-qubit QLDM with 120–360 parameters outperforming a classical denoiser on FID would be a notable empirical data point for quantum generative modeling, especially with the few-shot and early-convergence observations. The paper's strengths include a concrete hybrid architecture, evaluation on three datasets, inclusion of GAN/QGAN baselines, explicit parameter counts, and a transparent acknowledgment in Sec. 4.7 that all experiments use noiseless simulation. However, the significance is currently bounded by two load-bearing gaps: the classical baseline is not architecture-matched (it omits the temporal branch and the residual connection), and FID differences are reported without repeated-seed error bars, so the claimed advantage cannot yet be attributed to quantum effects rather than classical architectural features.

major comments (4)
  1. [Sec. 3, classical baseline paragraph] The classical baseline is not architecture-matched to the QLDM. The text states that a classical network 'perfectly replicates the fully quantum architecture' but then describes it as implemented 'with simple fully connected layers with an input of 10 and an output of 10' (330 parameters). The QLDM denoiser consists of three VQCs (latent, temporal, and combined) plus a residual skip connection that adds x to the output. A fully connected layer cannot represent the temporal embedding branch or the residual connection, so the comparison conflates quantum effects with classical architectural components known to improve diffusion denoisers. The central claim in the abstract—that better metrics arise from the 'quantum version'—is therefore not supported by the current experimental design. Please add a classical control that mirrors the three-branch structure and the skip connection, with a comparable parameter count.
  2. [Sec. 4.2, Tables 1–3 and Figs. 7, 9, 11] FID values are reported as single point estimates without repeated-seed error bars, while KID and IS are reported with standard deviations. The claimed quantum advantages (e.g., MNIST FID 44.3568 vs 40.4031; EuroSAT FID 30.5643 vs 20.1073) need seed-to-seed variance to assess statistical significance. A 4–5 point FID gap on MNIST may be within the noise of a single training run, especially given the small latent dimension. Please report mean ± std over at least three to five independent seeds for all metrics, and, if possible, a paired significance test.
  3. [Sec. 4.2, BasicQ exclusion] The 'Basic' architecture with measurement on observable X is excluded from the results because it 'showed extremely worse metric values compared to the other quantum architectures considered.' This is a post hoc exclusion based on the outcome, which biases the comparison in favor of the quantum models. Please report the excluded variant's results (or at least its final FID/KID/IS) and justify the exclusion with a criterion stated before training, or treat the full set of architectures as the comparison.
  4. [Sec. 4.3, Table 2 and abstract] The abstract claims that the quantum version achieves 'better metrics' than the classical version, but on Fashion MNIST the classical model has the best KID (0.0711 vs 0.0734–0.0773) and the best IS (3.6283 vs 3.33–3.52); only FID favors the quantum models. This discrepancy is acknowledged in the text ('the best IS is that of the classical model' and 'it is the classical model that has the lowest KID'), but it directly contradicts the abstract's sweeping claim. The conclusion should be narrowed to FID, or the abstract should be revised to state that quantum models improve FID while KID/IS are comparable or slightly better classically.
minor comments (5)
  1. [Sec. 3, ansatz description] The text describes the more expressive ansatz as 'a rotation on RX, followed by RZ, and finally RX' and later refers to 'rotations on RX-RZ-RZ', which is inconsistent; Fig. 6 shows RX-RZ-RX. Please correct the wording.
  2. [Table 2] In Table 2 several KID entries contain a spurious space between the leading zero and the decimal point (e.g., '0 .0773±0.0024'); please fix the formatting.
  3. [Sec. 3, parameter counts] The description of the classical baseline is too vague: 'simple fully connected layers with an input of 10 and an output of 10' does not explain how 330 parameters arise (e.g., number of layers, hidden widths, biases). Please specify the exact architecture so the parameter comparison is reproducible.
  4. [Sec. 4.4, EuroSAT] The EuroSAT experiment uses only two classes (Forest and Herbaceous Vegetation) and images resized to 28×28; the paper acknowledges this limitation, but the conclusion 'paves the way for using our model not only for toy problems but also for real and more complex problems' overstates the evidence. Please temper this sentence.
  5. [Sec. 4.7, Impact of noise] The paper explicitly states that all experiments assume a noiseless quantum environment (Sec. 4.7). This is a proper limitation, but it should also be stated in the abstract or introduction so that the practical reach of the claim is clear to readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the quantum-advantage claim is empirical and the cited self-work is not load-bearing.

full rationale

The paper's central claim is an empirical performance comparison between a QLDM and a classical latent diffusion denoiser, evaluated by FID/KID/IS on MNIST, Fashion MNIST, and a two-class EuroSAT subset. There is no derivation chain in which an output is constructed from its own input: no fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in by citation. The self-citations (refs. [14], [20], [21]) are used as background, training-method context, or as an extra QGAN baseline; the QLDM comparison does not depend on their validity. The main methodological weakness is the unmatched classical baseline: Sect. 3 states that the classical network 'perfectly replicates the fully quantum architecture' but then describes it as 'simple fully connected layers with an input of 10 and an output of 10', so the residual connection and the temporal VQC branch are not structurally replicated. That is a threat to the fairness of the comparison and to the quantum-advantage conclusion, but it is an experimental design issue rather than a circular reduction: the QLDM metrics are not equal to any input by construction. Likewise, the grid-search selection of hyperparameters is a selection procedure, not a fitted parameter presented as a prediction. Under the stated circularity criteria, no circular step can be identified.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The QLDM's claims rest primarily on the expressive-power assumption for the three VQCs, the fairness of the 330-parameter MLP baseline, the validity of Inception-based metrics on small grayscale images, and the representativeness of noiseless simulation. None of these are established independently in the paper. No new physical entities are introduced.

free parameters (6)
  • Number of qubits = 10 (also tested 8 and 12)
    Chosen as a trade-off between expressivity and simulation cost; Section 4.5 shows performance improves with more qubits (Table 4).
  • Ansatz depth = 3 layers (basic) or 4 layers (expressive)
    Depth is hand-chosen; Section 4.5 reports marginal gains from depth 3 to 4, and the total parameter count changes (120 to 360).
  • Measurement observable = Z or X; X best on MNIST FID, Z best on Fashion MNIST and EuroSAT
    Measurement basis is a design choice; BasicX was excluded after seeing poor performance (Section 4.2).
  • Learning rate = 1e-3
    Selected by grid search; Section 4.5 shows higher and lower rates give worse FID (Table 5).
  • Latent dimension = 10
    Latent vectors from the autoencoder have dimension 10, which sets the number of qubits and enables angle encoding.
  • Classical baseline parameter count = 330
    The classical MLP has 330 parameters, chosen to be comparable to the QLDM's 120 to 360, but it is not a ResNet-style architecture.
assumptions (5)
  • standard math The reverse diffusion conditional q(x_{t-1}|x_t) is approximately Gaussian when beta_t is small, so a parametric denoiser trained with the simplified loss in Eq. (8) is valid.
    Standard DDPM assumption from Ho et al. [2], used in Section 2.1.
  • ad hoc to paper The expectation values measured from the three VQCs form a function class expressive enough to approximate the noise estimator epsilon_theta(x_t,t).
    No proof is given; the paper cites expressibility results [27, 28] as motivation in Section 3. This is the core architectural assumption.
  • ad hoc to paper The classical MLP with 330 parameters is a fair control, so metric differences can be attributed to the quantum denoiser rather than to other architectural features.
    The QLDM uses three processing streams and a skip connection (ResNet-like), while the baseline uses simple fully connected layers (Section 3).
  • domain assumption FID, KID, and IS computed with InceptionV3 features on grayscale images replicated across three channels are valid quality measures for these datasets.
    Common practice in generative model evaluation, but Inception features are not designed for MNIST-style images; the KID citation is incorrect.
  • domain assumption Noiseless simulation of the variational circuits is representative of anticipated performance on real quantum hardware.
    The paper explicitly assumes a noiseless environment in Section 4.7 and provides no empirical noise study.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum Latent Diffusion Models." pith.science (2026). https://pith.science/paper/BIHJSJJO

@misc{pith2026250111174,
  author       = {Pith},
  title        = {Pith review of: Quantum Latent Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIHJSJJO}},
  note         = {Machine review of arXiv:2501.11174}
}
read the original abstract

The introduction of quantum concepts is increasingly making its way into generative machine learning models. However, while there are various implementations of quantum Generative Adversarial Networks, the integration of quantum elements into diffusion models remains an open and challenging task. In this work, we propose a potential version of a quantum diffusion model that leverages the established idea of classical latent diffusion models. This involves using a traditional autoencoder to reduce images, followed by operations with variational circuits in the latent space. To effectively assess the benefits brought by quantum computing, the images generated by the quantum latent diffusion model have been compared to those generated by a classical model with a similar number of parameters, evaluated in terms of quantitative metrics. The results demonstrate an advantage in using a quantum version, as evidenced by obtaining better metrics for the images generated by the quantum version compared to those obtained by the classical version. Furthermore, quantum models continue to outperform even when considering small percentages of the dataset for training, demonstrating the quantum's ability to extract features more effectively even in a few shot learning scenario.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 34 canonical work pages

  1. [1]

    Pro- ceedings of Machine Learning Research, vol

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., Ganguli, S.: Deep unsuper- visedlearningusingnonequilibriumthermodynamics.In:Bach,F.,Blei,D.(eds.) Proceedings of the 32nd International Conference on Machine Learning. Pro- ceedings of Machine Learning Research, vol. 37, pp. 2256–2265. PMLR, Lille, France (2015).https://proceedings.mlr.press/v37/sohl-di...

    work page 2015

  2. [2]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H. (eds.) Advances in Neural Information Processing Systems, vol. 33, pp. 6840–6851. Curran Associates, Inc., NY, USA (2020)

    work page 2020

  3. [3]

    2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10674–10685 (2021)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10674–10685 (2021)

    work page 2021

  4. [4]

    In: Ghahramani, Z., Welling,M.,Cortes,C.,Lawrence,N.,Weinberger,K.Q.(eds.)AdvancesinNeu- ral Information Processing Systems, vol

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. In: Ghahramani, Z., Welling,M.,Cortes,C.,Lawrence,N.,Weinberger,K.Q.(eds.)AdvancesinNeu- ral Information Processing Systems, vol. 27. Curran Associates, Inc., NY, USA (2014)

    work page 2014

  5. [5]

    Farhi, E., Neven, H.: Classification with Quantum Neural Networks on Near Term Processors (2018)

    2018

  6. [6]

    npj Quantum Information (2021)

    Johri, S., Debnath, S., Mocherla, A., Singh, A., Prakash, A., Kim, J., Kerenidis, I.: Nearest Centroid Classification on a Trapped Ion Quantum Computer. npj Quantum Information (2021)

    work page 2021

  7. [7]

    EPL (Europhysics Letters)119(6), 60002 (2017) https://doi.org/10.1209/0295-5075/119/60002 35

    Schuld,M.,Fingerhuth,M.,Petruccione,F.:Implementingadistance-basedclas- sifier with a quantum interference circuit. EPL (Europhysics Letters)119(6), 60002 (2017) https://doi.org/10.1209/0295-5075/119/60002 35

    doi:10.1209/0295-5075/119/60002 2017

  8. [8]

    Machine Learning 90(2), 261–287 (2013) https://doi.org/10.1007/ s10994-012-5316-5

    Aïmeur, E., Brassard, G., Gambs, S.: Quantum speed-up for unsupervised learning. Machine Learning 90(2), 261–287 (2013) https://doi.org/10.1007/ s10994-012-5316-5

    work page 2013

Show all 46 references
  1. [9]

    Quantum Science and Technology4(4), 043001 (2019) https://doi.org/10.1088/2058-9565/ab4eb5

    Benedetti,M.,Lloyd,E.,Sack,S.,Fiorentini,M.:Parameterizedquantumcircuits as machine learning models. Quantum Science and Technology4(4), 043001 (2019) https://doi.org/10.1088/2058-9565/ab4eb5

    2019 doi

  2. [10]

    Science362(6412), 308–311 (2018) https://doi.org/10.1126/science.aar3106

    Bravyi, S., Gosset, D., König, R.: Quantum advantage with shallow circuits. Science362(6412), 308–311 (2018) https://doi.org/10.1126/science.aar3106

    2018 doi

  3. [11]

    Bravyi, S., Gosset, D., König, R., Tomamichel, M.: Quantum advantage with noisyshallowcircuits.NaturePhysics 16(10),1040–1045(2020)https://doi.org/ 10.1038/s41567-020-0948-z

    2020 doi

  4. [12]

    Nature Computational Science1(6), 403–409 (2021) https://doi.org/10.1038/s43588-021-00084-1

    Abbas,A.,Sutter,D.,Zoufal,C.,Lucchi,A.,Figalli,A.,Woerner,S.:Thepowerof quantum neural networks. Nature Computational Science1(6), 403–409 (2021) https://doi.org/10.1038/s43588-021-00084-1

    2021 doi

  5. [13]

    Huang, H.-L., Du, Y., Gong, M., Zhao, Y., Wu, Y., Wang, C., Li, S., Liang, F., Lin, J., Xu, Y., Yang, R., Liu, T., Hsieh, M.-H., Deng, H., Rong, H., Peng, C.-Z., Lu, C.-Y.,Chen,Y.-A.,Tao,D.,Zhu,X.,Pan,J.-W.:Experimentalquantumgenerative adversarialnetworksforimagegeneration.Ph...

    2021 doi

  6. [14]

    https://arxiv

    Chang, S.Y., Thanasilp, S., Saux, B.L., Vallecorsa, S., Grossi, M.: Latent Style- based Quantum GAN for high-quality Image Generation (2024). https://arxiv. org/abs/2406.02668

    2024 arXiv

  7. [15]

    IEEE Transactions on Quantum Engineering4, 1–19 (2022)

    Tsang, S.L., West, M.T., Erfani, S.M., Usman, M.: Hybrid quantum–classical generative adversarial network for high-resolution image generation. IEEE Transactions on Quantum Engineering4, 1–19 (2022)

    2022

  8. [16]

    (eds.) Advances in Neural Information Processing Systems, vol

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs 36 trainedbyatwotime-scaleupdateruleconvergetoalocalnashequilibrium.In: Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R. (eds.) Advances in Neural Informati...

    2017

  9. [17]

    Cacioppo,A.,Colantonio,L.,Bordoni,S.,Giagu,S.:QuantumDiffusionModels (2023)

    2023

  10. [18]

    Parigi, M., Martina, S., Caruso, F.: Quantum-Noise-driven Generative Diffusion Models (2023)

    2023

  11. [19]

    Nature Reviews Physics3(9), 625–644 (2021) https://doi.org/10

    Cerezo, M., Arrasmith, A., Babbush, R., Benjamin, S.C., Endo, S., Fujii, K., McClean, J.R., Mitarai, K., Yuan, X., Cincio, L., Coles, P.J.: Variational quantum algorithms. Nature Reviews Physics3(9), 625–644 (2021) https://doi.org/10. 1038/s42254-021-00348-9

    2021

  12. [20]

    early access

    Scala,F.,Ceschini,A.,Panella,M.,Gerace,D.:Ageneralapproachtodropoutin quantumneuralnetworks.AdvancedQuantumTechnologies n/a(n/a),2300220 (2023) https://doi.org/10.1002/qute.202300220 . early access

    2023 doi

  13. [21]

    Quantum Machine Intelligence5(2) (2023) https://doi.org/10.1007/ s42484-023-00126-z

    Incudini, M., Grossi, M., Ceschini, A., Mandarino, A., Panella, M., Vallecorsa, S.,Windridge,D.:Resourcesavingviaensembletechniquesforquantumneural networks. Quantum Machine Intelligence5(2) (2023) https://doi.org/10.1007/ s42484-023-00126-z

    2023

  14. [22]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recogni- tion. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    2016

  15. [23]

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L., Polosukhin, I.: Attention Is All You Need (2023)

    2023

  16. [24]

    Nature physics10(9), 631–633 (2014) 37

    Lloyd, S., Mohseni, M., Rebentrost, P.: Quantum principal component analysis. Nature physics10(9), 631–633 (2014) 37

    2014

  17. [25]

    Physical review letters109(5), 050505 (2012)

    Wiebe, N., Braun, D., Lloyd, S.: Quantum algorithm for data fitting. Physical review letters109(5), 050505 (2012)

    2012

  18. [26]

    Physical Review X7(3), 031041 (2017)

    Yao, X.-W., Wang, H., Liao, Z., Chen, M.-C., Pan, J., Li, J., Zhang, K., Lin, X., Wang, Z., Luo, Z.,et al.: Quantum image processing and its application to edge detection: theory and experiment. Physical Review X7(3), 031041 (2017)

    2017

  19. [27]

    Physical Review Research2(3), 033125 (2020)

    Du,Y.,Hsieh,M.-H.,Liu,T.,Tao,D.:Expressivepowerofparametrizedquantum circuits. Physical Review Research2(3), 033125 (2020)

    2020

  20. [28]

    arXiv preprint arXiv:2310.07528 (2023)

    Yu, Z., Chen, Q., Jiao, Y., Li, Y., Lu, X., Wang, X., Yang, J.Z.: Provable advantage of parameterized quantum circuit in function approximation. arXiv preprint arXiv:2310.07528 (2023)

    2023 arXiv

  21. [29]

    Zaman, K., Ahmed, T., Kashif, M., Hanif, M.A., Marchisio, A., Shafique, M.: Studying the Impact of Quantum-Specific Hyperparameters on Hybrid Quantum-Classical Neural Networks (2024)

    2024

  22. [30]

    Bergholm, V., Izaac, J., al.: PennyLane: Automatic differentiation of hybrid quantum-classical computations (2022)

    2022

  23. [31]

    com/google/flax

    Heek, J., Levskaya, A., Oliver, A., Ritter, M., Rondepierre, B., Steiner, A., Zee, M.:Flax:AneuralnetworklibraryandecosystemforJAX(2023).http://github. com/google/flax

    2023

  24. [32]

    ATT Labs [Online]

    LeCun, Y., Cortes, C., Burges, C.: Mnist handwritten digit database. ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist2 (2010)

    2010

  25. [33]

    CoRR abs/1708.07747 (2017) arXiv:1708.07747

    Xiao, H., Rasul, K., Vollgraf, R.: Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. CoRR abs/1708.07747 (2017) arXiv:1708.07747

    2017 arXiv

  26. [34]

    Helber,P.,Bischke,B.,Dengel,A.,Borth,D.:EuroSAT:ANovelDatasetandDeep Learning Benchmark for Land Use and Land Cover Classification (2017) 38

    2017

  27. [35]

    Choi,J.,Lee,J.,Shin,C.,Kim,S.,Kim,H.,Yoon,S.:PerceptionPrioritizedTraining of Diffusion Models (2022)

    2022

  28. [36]

    Kingma,D.P.,Ba,J.:Adam:Amethodforstochasticoptimization.arXivpreprint arXiv:1412.6980 (2014)

    2014 arXiv

  29. [37]

    In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R

    Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., Chen, X., Chen, X.: Improved techniques for training GANs. In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. (eds.) Advances in Neural Information Processing Systems, vol. 29. Curran Associates, I...

    2016

  30. [38]

    Betzalel, E., Penso, C., Navon, A., Fetaya, E.: A Study on the Evaluation of Generative Models (2022)

    2022

  31. [39]

    21105/joss.04101

    Detlefsen,N.S.,Borovec,J.,Schock,J.,Jha,A.H.,Koker,T.,Liello,L.D.,Stancl,D., Quan,C.,Grechkin,M.,Falcon,W.:Torchmetrics-measuringreproducibilityin pytorch.JournalofOpenSourceSoftware 7(70),4101(2022)https://doi.org/10. 21105/joss.04101

    2022

  32. [40]

    ACM Computing Surveys (CSUR)54(7), 1–35 (2021)

    Resch, S., Karpuzcu, U.R.: Benchmarking quantum computers and the impact of quantum noise. ACM Computing Surveys (CSUR)54(7), 1–35 (2021)

    2021

  33. [41]

    PRX Quantum 3(4), 040313 (2022)

    Guo, Y., Yang, S.: Quantum error mitigation via matrix product operators. PRX Quantum 3(4), 040313 (2022)

    2022

  34. [42]

    Physical Review Applied21(6), 067001 (2024)

    Gujju, Y., Matsuo, A., Raymond, R.: Quantum machine learning on near-term quantum devices: Current state of supervised and unsupervised techniques for real-world applications. Physical Review Applied21(6), 067001 (2024)

    2024

  35. [43]

    Physical Review A86(3), 032324 (2012)

    Fowler, A.G., Mariantoni, M., Martinis, J.M., Cleland, A.N.: Surface codes: Towards practical large-scale quantum computation. Physical Review A86(3), 032324 (2012)

    2012

  36. [44]

    Nature614(7949), 676–681 (2023)

    Acharya, R., Aleiner, I.,et al.: Suppressing quantum errors by scaling a surface 39 code logical qubit. Nature614(7949), 676–681 (2023)

    2023

  37. [45]

    Wang, S., Czarnik, P., Arrasmith, A., Cerezo, M., Cincio, L., Coles, P.J.: Can error mitigation improve trainability of noisy variational quantum algorithms? Quantum8, 1287 (2024)

    2024

  38. [46]

    Reviews of Modern Physics95(4), 045005 (2023) 40

    Cai, Z., Babbush, R., Benjamin, S.C., Endo, S., Huggins, W.J., Li, Y., McClean, J.R., O’Brien, T.E.: Quantum error mitigation. Reviews of Modern Physics95(4), 045005 (2023) 40

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.