Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Solving MNIST with a globally trained Mixture of Quantum Experts

T0 review · 4 major / 5 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 A purely quantum mixture-of-experts circuit, trained with all experts updated together, classifies the parity of full-resolution MNIST images with up to 97.5 percent test accuracy using 10 qubits.

desk verdict A genuinely new MoQE architecture with an honest quadratic baseline and an interesting compute-accuracy claim, but the empirical results lack error bars and the infinite-expert guarantee is asserted rather than proved. read the letter →

arxiv 2505.14789 v1 pith:VI3RTQEB submitted 2025-05-20 quant-ph

Paolo Alessandro Xavier Tognini , Leonardo Banchi , Giacomo De Palma This is my paper · ORCID
classification quant-ph
keywords quantummachinelearningmixtureofexpertsamplitudeencodingMNISTvariationalcircuitsGaussianprocesslimitparameter-shiftruletrainability
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 purely quantum neural network that classifies full-resolution MNIST images by digit parity (odd versus even) with up to 97.5 percent test accuracy, using 10 qubits and no classical pre- or post-processing. The model is a mixture of quantum experts: each expert runs the same three-layer parametric circuit on the amplitude-encoded image, and the output of the network is the sum of the experts' measurements. The central empirical claim is that training all experts together, as one global model, gives a better compute-accuracy tradeoff than training fewer experts or training each expert independently. The paper further argues that in the limit of infinitely many experts the training algorithm can perfectly fit the training data, drawing on Gaussian-process limit results for wide quantum neural networks. If the claims hold, a small quantum device operating within a quadratic function class can approach the accuracy ceiling of a classical quadratic classifier.

What carries the argument

The central object is the Mixture of Quantum Experts (MoQE): a sum of n independent parametric circuits, each a three-layer ladder of two-qubit gates with 252 parameters acting on the same amplitude-encoded 10-qubit input, whose outputs are added and normalized to give the prediction. The mechanism that carries the argument is amplitude encoding, which packs the full image into the wavefunction and makes each single-qubit rotation act like a convolution over all pixels, together with global gradient-descent training in which every parameter update depends on the parameters of all experts. The theoretical engine is the identification of the n-expert mixture with the infinite-width limit of a quantum neural network, where Gaussian-process results predict trainability and a perfect fit to the training data.

What would settle it

Train the model with 64 or 128 experts while measuring the training loss per epoch and the output covariance across random initializations, then compare the loss decay and covariance with the analytic Gaussian-process prediction for the ladder expert: if the loss does not decay exponentially or the covariance deviates substantially from the predicted kernel, the infinite-expert claim fails even if the finite-n accuracy of 97.5 percent stands.

Watch

Extended reading notes

Core claim

Using amplitude encoding, the paper embeds each 28 x 28 MNIST image, zero-padded to 32 x 32, in the amplitudes of a 10-qubit state, so that a single one-qubit gate acts as a convolution over the whole image. Each expert is three layers of two-qubit gates arranged on a ladder, with 252 parameters per expert, and the full network sums the normalized Pauli-Z expectations of all experts; the sign of the sum gives the parity prediction. The model is trained by gradient descent on square loss using the parameter-shift rule, with every expert's parameters updated jointly in each step rather than expert by expert. Simulations with 1, 2, 4, 8, 16, and 32 experts reach 97.54 percent test accuracy at 16 experts, and the authors report that for equal computational cost, more jointly trained experts give higher accuracy until saturation. In the limit of infinitely many experts, the paper claims that Gaussian-process theorems apply directly because the experts are disconnected, guaranteeing exponential decay of the training cost and a perfect fit to the training data. The paper also observes that the model's function class is quadratic in the pixel intensities, which it connects to the roughly 98 percent ceiling of a classical quadratic classifier.

Load-bearing premise

The paper's claim that infinitely many experts guarantee a perfect fit to the training data rests on assuming that known Gaussian-process limit theorems for wide quantum neural networks apply to this specific three-layer ladder circuit as the number of experts grows, which the paper asserts without verifying every technical hypothesis.

Editorial extensions

If this is right

  • A 10- to 20-qubit device could process images of hundreds or thousands of pixels through amplitude encoding, without downsampling or classical feature extraction.
  • Within the model's quadratic function class, adding experts and training them jointly improves accuracy per unit of compute up to roughly 16 to 32 experts, after which the gain saturates.
  • If the infinite-expert limit holds, fitting a finite training set becomes easier as the number of experts grows, with the cost decaying exponentially in training time and local-minima obstructions disappearing.
  • Because the model function is quadratic in the pixel intensities, the demonstrated accuracy is near the ceiling of its function class; reaching higher accuracy requires adding nonlinearity beyond amplitude encoding.
  • The architecture is purely quantum on both input and output sides, so any observed accuracy is attributable to the quantum model itself rather than to classical preprocessing or postprocessing.

Reading between the lines

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

  • An untested but direct consequence of the paper's structure is that the compute-accuracy comparison is a property of the joint training algorithm, not of raw parameter count: a single large circuit with 8064 parameters would not necessarily match the accuracy of 32 jointly trained experts.
  • The Gaussian-process assumption could be checked directly by measuring the covariance of the MoQE output across random initializations and comparing it with the analytic kernel of the ladder expert, separating the finite-n behavior from the infinite-n guarantee.
  • The same amplitude-encoding-plus-experts recipe should transfer to other binary image tasks with similar pixel counts, but its quadratic function class suggests it will degrade on strongly nonlinear decision boundaries unless data re-uploading or other nonlinearity is added.
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 / 5 minor

Summary. The paper proposes a Mixture of Quantum Experts (MoQE) architecture for binary classification of full-resolution MNIST images (parity task). Each expert is a 3-layer 10-qubit variational circuit with 252 parameters; the input is amplitude-encoded, and the model output is a normalized sum of the experts' local Z-expectation values. All experts are trained jointly by gradient descent with the parameter-shift rule. Simulations with n = 1, 2, 4, 8, 16, 32 experts report a maximum test accuracy of 97.5%, and Figure 3 argues that for a fixed compute budget, training more experts jointly reaches higher accuracy. The paper further claims, via Gaussian-process limit theorems of Refs. [34-36], that in the limit of infinitely many experts the training algorithm can perfectly fit the training data. Appendix E shows that a classical quadratic classifier achieves about 98% test accuracy, supporting the interpretation that the current model is essentially a quadratic classifier on the pixel intensities.

Significance. If the empirical claims hold, this is a notable result for quantum machine learning: a purely quantum model on 10 qubits processes full-resolution MNIST images and approaches the classical quadratic-classifier ceiling within its function class, with a compute-accuracy tradeoff improved by jointly training more experts. The model-class identification in Appendix E is a genuine strength: it provides a falsifiable and parameter-aware classical benchmark against which the quantum accuracy can be judged. The main advertised theoretical contribution, however, is the infinite-expert perfect-fit guarantee, which currently rests on an unverified transfer of Gaussian-process limit theorems, and the empirical claims are based on single runs without error bars or a test-blind stopping rule. These issues are fixable and do not invalidate the core architecture idea.

major comments (4)
  1. [II.B] The transfer of the Gaussian-process limit theorems of Refs. [34-36] to the MoQE architecture is asserted, not demonstrated. The initialization distribution of the 252 parameters per expert is never specified, so one cannot check the randomness or gate-design hypotheses required by those theorems; the output normalization in Section II (empirical variance over the MNIST images set to one) is a data-dependent scaling and not obviously the fixed output normalization assumed in the GP limit; and the small-light-cone condition is argued only at the level of disconnected experts, not for the global observable (1/sqrt n) * sum of local Z measurements. Until these hypotheses are verified, the abstract's claim that 'in the limit of infinitely many experts, our training algorithm can perfectly fit the training data' is unsupported, even though the finite-n numerical results could still stand independently.
  2. [Table I, Figure 3, Appendix D] The empirical central result rests on a single run per expert count. Table I reports one training/test accuracy per n, and Appendix D states that 'the number of epochs is chosen to saturate the accuracy,' i.e., the stopping rule uses the test accuracy. No error bars, multiple seeds, or standard deviations are reported. The compute-accuracy advantage of joint training shown in Figure 3 could therefore be affected by run-to-run variance and by the test-based stopping rule. Please report multiple random seeds with mean and standard deviation, and adopt a stopping rule that does not use the test labels, or show the conclusion is robust under a fixed epoch budget.
  3. [Abstract and Section II.A] The advertised advantage that joint training 'significantly improves trainability with respect to training each expert independently' is not directly tested. All runs in Table I and Figure 3 train the global output jointly; there is no baseline in which n experts are trained independently and then combined (e.g., by averaging or voting), which is the comparison the sentence promises. Please add such a baseline, or restrict the claim to what is measured: more jointly trained experts improve the compute-accuracy tradeoff.
  4. [Section II, paragraph after Fig. 2(c)] The statement 'The normalization is such that the empirical variance of the result on the MNIST images is equal to one' is ambiguous and potentially problematic. It is not specified whether the variance is computed on training images only or on the full dataset including the test images; if it includes test images, the reported test accuracies are influenced by information from the test set. Please specify the data used for this normalization, and if necessary recompute the results with the normalization fitted on the training split only.
minor comments (5)
  1. [Abstract and Section IV] The phrase 'non-coarse grained' should be hyphenated as 'non-coarse-grained,' and the claim that this is 'the first time that a purely quantum architecture is successfully applied to the non-coarse grained MNIST database' is a strong historical claim that should be supported by a more thorough literature search or softened.
  2. [Appendix A] The description of the data split is confusing: 'We used 50,000 images from the training set, selected with a different seed in each iteration, and we kept the remaining 10,000 images available for potential statistical controls that eventually proved unnecessary.' Please clarify what 'each iteration' means, whether the training subset changes across runs, and whether the reserved 10,000 images were used at all.
  3. [Reference [33]] Reference [33] is cited only as a webpage without a version or author; please provide a stable, dated citation for the parameter-shift rule.
  4. [Figure 6 and Table II] The 'equivalent number of experts' comparison divides the number of classical parameters by 252, which treats one classical weight and one rotation angle as equivalent; the comparison should be labeled explicitly as a parameter-count comparison, not an expressivity comparison.
  5. [General] The paper does not include a code or data availability statement; since all results are simulations, releasing the code would substantially aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the 97.5% accuracy is standard supervised fitting benchmarked against an external classical quadratic classifier, and the infinite-expert guarantee is an application of independent published GP-limit theorems, not a construction-level reduction.

full rationale

The claimed derivation chain is not circular. The model function f = (1/sqrt n) sum_i f_i and the loss C = (f - l)^2 are defined independently of the reported predictions, and the 97.5% test accuracy is obtained by training on a 50k-image training split and evaluating on the held-out 10k test split (Appendix A), which is standard empirical fitting rather than a prediction forced by construction. The paper's own check against a classical quadratic classifier (Appendix E) is an external benchmark: it confirms that the quantum model saturates the quadratic function class instead of defining that class as the target. The only passage that could raise a self-citation concern is Section II.B, where the infinite-expert perfect-fit guarantee is imported from Refs. [34-36]; two of these are co-authored by De Palma, and the paper asserts 'These results can be directly applied to our quantum neural network' without verifying all hypotheses (initialization distribution, output scaling, barren-plateau freedom). However, those cited results are parameter-free published theorems with stated assumptions that do not include the MoQE architecture, so they are independent evidence under the evaluation rules rather than a self-referential reduction. The missing hypothesis check is a rigor/correctness concern, not circularity: no equation in the paper is equal to its input by construction, and no fitted parameter is renamed as a prediction. The single-qubit-convolution observation in Appendix C is a direct algebraic property of amplitude encoding, not an imported ansatz. Accordingly, no circular step meets the evidentiary bar, and the score is 0.

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

The paper introduces no new physical entities. Its free parameters are the trained circuit weights plus hand-chosen hyperparameters such as layer count, expert count, and normalization scale. The main non-standard assumption is the transfer of the infinite-width Gaussian-process theorem to this specific architecture, which is asserted rather than verified. All other axioms are standard quantum mechanics, standard optimization rules, or dataset assumptions.

free parameters (5)
  • Circuit rotation parameters per expert = 252 per expert, trained via gradient descent
    All predictive capacity of the model; values are fit to the MNIST training set and determine the reported accuracies.
  • Number of layers = 3
    Architecture hyperparameter chosen by the authors without a reported search procedure; affects model capacity.
  • Number of experts n = 1, 2, 4, 8, 16, 32
    Swept hyperparameter; results depend strongly on it and no selection rule is reported.
  • Output normalization scale = chosen so empirical variance on MNIST images equals 1
    Global rescaling of circuit outputs; does not affect the sign of predictions but changes the loss scale and training dynamics.
  • Optimizer hyperparameters (learning rate, Adam settings) = not reported
    These affect training dynamics and final accuracy but are absent from the manuscript, limiting reproducibility.
assumptions (5)
  • standard math Amplitude encoding maps normalized pixel intensities to amplitudes of a 10-qubit state (Eq. 1), and the measured expectation value of each expert is a quadratic polynomial of those amplitudes.
    Used in Section III and Appendix E; relies on standard linear-algebra facts about expectation values and linear encoding.
  • standard math The parameter-shift rule gives exact gradients for the Pauli rotation gates used in the circuit.
    Standard result in variational quantum circuits, cited as [33] and used for all gradient computations.
  • ad hoc to paper The infinite-width Gaussian process result of Refs. [34-36] applies to this MoE architecture when the number of experts goes to infinity, including the small-light-cone and no-barren-plateau hypotheses.
    The authors assert this in Section II.B without proving each hypothesis for the specific 3-layer ladder circuit; one reference is co-authored by G. De Palma.
  • standard math A single 1-qubit gate acting on an amplitude-encoded image implements a convolution with a 1x2 kernel with stride 2.
    Appendix C shows the algebra for a 2x2 matrix acting on amplitude pairs; this is a direct property of amplitude encoding.
  • domain assumption The 50,000-image training subset, with a different seed per iteration, and the batch size of 4 are representative enough for the reported comparisons.
    Data-handling choices are not justified and are not repeated across runs; they affect the measured accuracies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Solving MNIST with a globally trained Mixture of Quantum Experts." pith.science (2026). https://pith.science/paper/VI3RTQEB

@misc{pith2026250514789,
  author       = {Pith},
  title        = {Pith review of: Solving MNIST with a globally trained Mixture of Quantum Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VI3RTQEB}},
  note         = {Machine review of arXiv:2505.14789}
}
read the original abstract

We propose a new quantum neural network for image classification, which is able to classify the parity of the MNIST dataset with full resolution with a test accuracy of up to 97.5% without any classical pre-processing or post-processing. Our architecture is based on a mixture of experts whose model function is the sum of the model functions of each expert. We encode the input with amplitude encoding, which allows us to encode full-resolution MNIST images with 10 qubits and to implement a convolution on the whole image with just a single one-qubit gate. Our training algorithm is based on training all the experts together, which significantly improves trainability with respect to training each expert independently. In fact, in the limit of infinitely many experts, our training algorithm can perfectly fit the training data. Our results demonstrate the potential of our quantum neural network to achieve high-accuracy image classification with minimal quantum resources, paving the way for more scalable and efficient quantum machine learning models.

Figures

Figures reproduced from arXiv: 2505.14789 by the authors.

Figure 1
Figure 1. FIG. 1. Scheme of our Mixture of Quantum Experts (MoQE) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. (a) Our elementary 2-qubit gate with 4 parameters. This will be the building block of the layers of our quantum [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Test accuracy in function of the computation, which [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: ): using a 3 × 3 kernel, a stride of 1 and a padding of 1, the output of the convolutional layer will have the same dimension as the input. If instead of a single kernel we have c kernels (or analogously if the kernel is a tensor of dimension (3, 3, c)), we will have a…
Figure 5
Figure 5. Figure 5: FIG. 5. General structure of our classical Convolutional Neural Network. The image undergoes a series of convolutional layers [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Comparison between the accuracy of classical and quantum neural networks in the classification of the parity of the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Running and Test Accuracy using circuits with different experts. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Histogram of circuit outputs for odd (left) and even (right) images. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Test accuracy in function of the epochs. Here we presented the complete data. [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Quantum feature-map learning with reduced resource overhead

    quant-ph 2025-10 conditional novelty 6.0 of 10

    By classically reconstructing quantum model outputs, Q-FLAIR selects gates, features, and weights with O(M) quantum evaluations per iteration, decoupling quantum cost from feature dimension and enabling >90% MNIST acc...

  2. Image classification via a quantum-inspired strategy involving a mixture of experts

    cs.LG 2026-07 conditional novelty 5.0 of 10

    Joint multi-expert unitary convolution plus stabilizer-syndrome features roughly halves MNIST/Fashion-MNIST failure rates versus classical diffusion pooling, at moderate GPU overhead.

  3. QMoE: A Quantum Mixture of Experts Framework for Scalable Quantum Neural Networks

    quant-ph 2025-07 conditional novelty 5.0 of 10

    QMoE, a quantum mixture-of-experts architecture with a learnable quantum router and multiple parameterized quantum expert circuits, reports consistent accuracy gains over standard quantum neural networks on 8x8 MNIST ...

Reference graph

Works this paper leans on

47 extracted references · 18 canonical work pages · cited by 3 Pith papers

  1. [1]

    the classification is done purely at the quantum level, without any classical neural network pre- processing or post-processing

    work page

  2. [2]

    we leverage amplitude encoding to process full- resolution images without downsampling

    work page

  3. [3]

    inspired by CNNs, the variational circuit is defined assuming correlations between neighboring pixels

    work page

  4. [4]

    understanding the LEarning process of QUantum Neural networks (LeQun)

    we use a mixture of experts, where each expert is a quantum neural network and, crucially, all experts are trained together. As a relevant problem, we focus on the classification of the parity of MNIST digits, which has been widely used in the Quantum Vision literature [23–26]. MNIST images have a resolution of 28 × 28 pixels. We encode each full-resoluti...

    work page 2022

  5. [5]

    Preskill, Quantum 2, 79 (2018)

    J. Preskill, Quantum 2, 79 (2018)

    2018

  6. [6]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Nature 549, 195–202 (2017)

    2017

  7. [7]

    Cerezo, G

    M. Cerezo, G. Verdon, H.-Y. Huang, L. Cincio, and P. J. Coles, Nature Computational Science 2, 567–576 (2022)

    2022

  8. [8]

    Aaronson, Nature Phys

    S. Aaronson, Nature Phys. 11, 291 (2015)

    work page 2015

Show all 47 references
  1. [9]

    Schuld and N

    M. Schuld and N. Killoran, Prx Quantum 3, 030101 (2022)

    2022

  2. [10]

    A. W. Harrow, A. Hassidim, and S. Lloyd, Physical Review Letters 103 (2009), 10.1103/phys- revlett.103.150502

    2009 doi

  3. [11]

    Gily´ en, Y

    A. Gily´ en, Y. Su, G. H. Low, and N. Wiebe, in Pro- ceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing, STOC ’19 (ACM, 2019) p. 193–204

    2019

  4. [12]

    Quantum computational complexity of ma- trix functions,

    S. Cifuentes, S. Wang, T. L. Silva, M. Berta, and L. Aolita, “Quantum computational complexity of ma- trix functions,” (2024), arXiv:2410.13937 [quant-ph]

    2024 arXiv

  5. [13]

    Cerezo, A

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, Nature Reviews Physics 3, 625–644 (2021)

    2021

  6. [14]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Bab- bush, and H. Neven, Nature Communications 9 (2018), 10.1038/s41467-018-07090-4

    2018 doi

  7. [15]

    Cerezo, A

    M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, Nature Communications 12 (2021), 10.1038/s41467-021-21728-w

    2021 doi

  8. [16]

    Quantifying the barren plateau phenomenon for a model of unstructured variational ans¨ atze,

    J. Napp, “Quantifying the barren plateau phenomenon for a model of unstructured variational ans¨ atze,” (2022), arXiv:2203.06174 [quant-ph]

    2022 arXiv

  9. [17]

    Holmes, K

    Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, PRX Quantum 3 (2022), 10.1103/prxquantum.3.010313

    2022 doi

  10. [18]

    E. R. Anschuetz and B. T. Kiani, Nature Communica- tions 13 (2022), 10.1038/s41467-022-35364-5

    2022 doi

  11. [19]

    Grant, L

    E. Grant, L. Wossnig, M. Ostaszewski, and M. Benedetti, Quantum 3, 214 (2019)

    2019

  12. [20]

    M. S. Rudolph, J. Miller, D. Motlagh, J. Chen, A. Acharya, and A. Perdomo-Ortiz, Nature Commu- nications 14 (2023), 10.1038/s41467-023-43908-6

    2023 doi

  13. [21]

    Liu, T.-P

    H.-Y. Liu, T.-P. Sun, Y.-C. Wu, Y.-J. Han, and G.-P. Guo, New Journal of Physics 25, 013039 (2023)

    2023

  14. [22]

    LeCun, Y

    Y. LeCun, Y. Bengio, and G. Hinton, nature 521, 436 (2015)

    2015

  15. [23]

    T. P. Lillicrap, A. Santoro, L. Marris, C. J. Akerman, and G. Hinton, Nature Reviews Neuroscience 21, 335 (2020)

    2020

  16. [24]

    Abbas, R

    A. Abbas, R. King, H.-Y. Huang, W. J. Huggins, R. Movassagh, D. Gilboa, and J. McClean, Advances in Neural Information Processing Systems 36, 44792 (2023)

    2023

  17. [25]

    Gily´ en, S

    A. Gily´ en, S. Arunachalam, and N. Wiebe, in Proceed- ings of the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms (SIAM, 2019) pp. 1425–1444

    2019

  18. [26]

    E. A. Cherrat, I. Kerenidis, N. Mathur, J. Landman, M. Strahm, and Y. Y. Li, Quantum 8, 1265 (2024)

    2024

  19. [27]

    Kerenidis and A

    I. Kerenidis and A. Luongo, Physical Review A 101 (2020), 10.1103/physreva.101.062327

    2020 doi

  20. [28]

    Quantum kitchen sinks: An algorithm for machine learning on near-term quantum computers,

    C. M. Wilson, J. S. Otterbach, N. Tezak, R. S. Smith, A. M. Polloreno, P. J. Karalekas, S. Heidel, M. S. Alam, G. E. Crooks, and M. P. da Silva, “Quantum kitchen sinks: An algorithm for machine learning on near-term quantum computers,” (2019), arXiv:1806.08321 [quant- ph]

    2019 arXiv

  21. [29]

    Post-variational quan- tum neural networks,

    P.-W. Huang and P. Rebentrost, “Post-variational quan- tum neural networks,” (2024), arXiv:2307.10560 [quant- ph]

    2024 arXiv

  22. [30]

    Classification with quantum neural networks on near term processors,

    E. Farhi and H. Neven, “Classification with quantum neural networks on near term processors,” (2018), arXiv:1802.06002 [quant-ph]

    2018 arXiv

  23. [31]

    I. Cong, S. Choi, and M. D. Lukin, Nature Physics 15, 1273–1278 (2019)

    2019

  24. [32]

    A tutorial on quan- tum convolutional neural networks (qcnn),

    S. Oh, J. Choi, and J. Kim, “A tutorial on quan- tum convolutional neural networks (qcnn),” (2020), arXiv:2009.09423 [quant-ph]

    2020 arXiv

  25. [33]

    J. I. Latorre, arXiv preprint quant-ph/0510031 (2005)

    2005 arXiv

  26. [34]

    Cramer, M

    M. Cramer, M. B. Plenio, S. T. Flammia, R. Somma, D. Gross, S. D. Bartlett, O. Landon-Cardinal, D. Poulin, and Y.-K. Liu, Nature communications 1, 149 (2010). 6

    2010

  27. [35]

    A survey on mixture of experts,

    W. Cai, J. Jiang, F. Wang, J. Tang, S. Kim, and J. Huang, “A survey on mixture of experts,” (2024), arXiv:2407.06204 [cs.LG]

    2024 arXiv

  28. [36]

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al., arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  29. [37]

    Parameter-shift Rule,

    Pennylane.ai, “Parameter-shift Rule,” (2025)

    2025

  30. [38]

    Girardi and G

    F. Girardi and G. De Palma, Communications in Mathe- matical Physics 406 (2025), 10.1007/s00220-025-05238-0

    2025 doi

  31. [39]

    Quantitative convergence of trained quan- tum neural networks to a gaussian process,

    A. M. Hernandez, F. Girardi, D. Pastorello, and G. D. Palma, “Quantitative convergence of trained quan- tum neural networks to a gaussian process,” (2024), arXiv:2412.03182 [quant-ph]

    2024

  32. [40]

    Abedi, S

    E. Abedi, S. Beigi, and L. Taghavi, Quantum 7, 989 (2023)

    2023

  33. [41]

    Pennylane: Automatic dif- ferentiation of hybrid quantum-classical computations,

    V. Bergholm, J. Izaac, M. Schuld, C. Gogolin, S. Ahmed, V. Ajith, M. S. Alam, G. Alonso-Linaje, B. Akash- Narayanan, A. Asadi, J. M. Arrazola, U. Azad, S. Ban- ning, C. Blank, T. R. Bromley, B. A. Cordier, J. Ceroni, A. Delgado, O. D. Matteo, A. Dusko, T. Garg, D. Guala, A. Ha...

    2022 arXiv

  34. [42]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” (2017), arXiv:1412.6980 [cs.LG]

    2017 arXiv

  35. [43]

    P´ erez-Salinas, A

    A. P´ erez-Salinas, A. Cervera-Lierta, E. Gil-Fuster, and J. I. Latorre, Quantum 4, 226 (2020)

    2020

  36. [44]

    Cire¸ san, U

    D. Cire¸ san, U. Meier, and J. Schmidhuber, CVPR 2012, p. 3642-3649 (2012), arXiv:1202.2745 [cs.CV]

    2012 arXiv

  37. [45]

    Deng, IEEE Signal Processing Magazine 29, 141 (2012)

    L. Deng, IEEE Signal Processing Magazine 29, 141 (2012)

    2012

  38. [46]

    An ensemble of simple convolutional neural network models for mnist digit recognition,

    S. An, M. Lee, S. Park, H. Yang, and J. So, “An ensemble of simple convolutional neural network models for mnist digit recognition,” (2020), arXiv:2008.10400 [cs.CV]

    2020 arXiv

  39. [47]

    Lets keep it simple, using simple archi- tectures to outperform deeper and more complex archi- tectures,

    S. H. Hasanpour, M. Rouhani, M. Fayyaz, and M. Sabokrou, “Lets keep it simple, using simple archi- tectures to outperform deeper and more complex archi- tectures,” (2023), arXiv:1608.06037 [cs.CV]. 7 1 1 -2 -1 1 1 4 2 2 2 2 22 2 2 1 1 1 1 1 1 11 1 00 0 0 0 0 0 0 0 0 0 0 0 0 0 ...

    2023 arXiv

Pith tools

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