Pith. sign in

REVIEW 4 major objections 5 minor 32 references

Benchmarking a Tunable Quantum Neural Network on Trapped-Ion and Superconducting Hardware

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

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

Pith's one-line read This paper demonstrates a tunable quantum neural network whose intermediate quantum uncertainty improves handwritten-digit classification over its classical limit on trapped-ion and superconducting hardware, with physical noise sometimes…

desk verdict Hardware confirmation of the intermediate-a effect is genuine, but the headline 'noise helps' claim rests on a single 10-shot image and a within-error-bar excess. read the letter →

arxiv 2507.21222 v2 pith:FAQJ6JM6 submitted 2025-07-28 quant-ph cond-mat.dis-nncs.LG

Djamil Lakhdar-Hamina , Xingxin Liu , Richard Barney , Sarah H. Miller , Alaina M. Green , Norbert M. Linke , Victor Galitski This is my paper · ORCID
classification quant-phcond-mat.dis-nncs.LG PACS 03.67.Lx
keywords quantumneuralnetworksbinarynetworkMNISTclassificationtrapped-ionprocessorsuperconductingqubitsactivationfunctionnoise-assistedinferencemid-circuitmeasurement
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 reports a quantum generalization of a binary neural network, called the BQNN, and runs its inference on trapped-ion and superconducting processors to classify handwritten-digit images. The network has a knob, the interpolation parameter $a$, that tunes it continuously between a classical deterministic network and a quantum network with measurement uncertainty. The core result is that at moderate $a$, quantum uncertainty in the measurement-based activations improves classification accuracy over the classical limit, and that on an intermediate set of ambiguous images the improvement appears even at $a=0$ because physical hardware noise pushes outputs toward the correct label. The paper therefore argues that certain physical noise, normally viewed as harmful, can be beneficial for some quantum machine learning tasks. Their experiments also show that injecting additional gate noise eventually degrades performance toward random guessing.

What carries the argument

The load-bearing object is the BQNN circuit: a three-layer feedforward network whose neurons are qubits, whose activations are projective measurement outcomes, and whose single-qubit rotation angles are functions of the previous layer's measurements. The rotation angle for neuron $i$ in layer $k$ is $\theta_i^k = \frac{\pi}{2}(1-\phi_a((W^1 I)_i))$ for the first layer and $\theta_i^k = \frac{\pi}{2}(d^{k-1}_i - \phi_a((W^k d^{(k-1)})_i))$ for later layers, with $\phi_a(x)=\operatorname{htanh}(x/a)$ reducing to $\operatorname{sgn}(x)$ as $a\to0$. At $a=0$ the rotations are multiples of $\pi$ and the feedforward is deterministic and classically equivalent, while at nonzero $a$ the rotations produce superposition states and hence random measurement outcomes. Training uses a clipped straight-through gradient estimator in simulation, and inference uses repeated runs with majority voting. The noise-injection procedure, inserting $n$ pairs of native gates $U$ and $U^\dagger$ after each layer, is the instrument used to benchmark how physical gate noise affects classification.

What would settle it

Run the same NY-image inference at $a=0$ on a device while independently logging gate fidelities and readout error rates; if the classification boost persists after error mitigation removes the incoherent component, or if it tracks a known systematic rotation or readout bias rather than random errors, the paper's attribution of the boost to physical noise is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a network whose feedforward pass consists of qubit rotations followed by projective measurements, with rotation angles set by previous measurement outcomes, displays a 'quantumness' window in which real-device inference outperforms the same network's classical $a=0$ limit. The boost comes from borderline images that the classical network deterministically misclassifies; at $a>0$, measurement randomness lets the output fluctuate between two nearby minima of the network's effective energy landscape, and on hardware the same fluctuation is caused partly by physical noise at $a=0$. For clearly classified images the outcome is stable, so the noise sensitivity is specific to ambiguous inputs. The paper verifies this picture across superconducting and two trapped-ion implementations, and shows that injecting extra single-qubit or two-qubit gate pairs eventually destroys performance, but that small or moderate noise can improve an NY image's validation rate rather than hurt it.

Load-bearing premise

The paper's noise-benefit claim assumes that hardware deviations from ideal simulation are random errors rather than a fixed bias that happens to favor the correct label, and it does not directly measure the noise.

Editorial extensions

If this is right

  • At moderate $a$, the BQNN on real superconducting and trapped-ion devices matches or exceeds the classical network's fraction of correctly classified images on the tested subset, with the gain concentrated on ambiguous borderline images.
  • NY images, misclassified classically but recovered quantumly, respond to hardware noise at $a=0$ in a way that clear YY images do not, suggesting that noisy quantum inference can flag ambiguous entries in a dataset.
  • Inserted single-qubit $UU^\dagger$ pairs can raise the NY-image validation rate on the trapped-ion device before enough noise drives all classifications toward random guessing.
  • Two-qubit gate-pair injection produces a sharp initial improvement on the trapped-ion NY image followed by decay to the random-guess rate, while the superconducting platform shows more resilience.
  • The qualitative agreement between simulation and three hardware implementations supports the BQNN as a portable benchmark for comparing device noise and for studying quantum-versus-classical behavior in neural networks.

Reading between the lines

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

  • If the NY-image sensitivity to noise is generic rather than peculiar to this digit dataset, quantum inference could serve as a data-quality probe: entries whose classifications are unstable under small hardware perturbations are exactly the ambiguous ones a classical network hides.
  • A natural testable extension, not performed in the paper, is to train the BQNN under small injected physical noise rather than only in clean simulation; the paper expects a systematic advantage from such noise-aware training, and that expectation could be checked by comparing validation rates.
  • Because the demonstrated circuit is separable and uses only mid-circuit measurements followed by classical feedback, scaling it to entangled partial-mid-circuit-measurement architectures would connect it to measurement-induced phase transitions; the authors flag this direction but do not demonstrate it, so the quantum-advantage claim remains prospective.
  • If the noise-injection metric, the number of $UU^\dagger$ pairs needed to reach random-guess performance, is reliable, it offers a neural-network-specific alternative to circuit-level error rates for comparing quantum platforms.
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 implements a quantum neural network (BQNN) on IBM superconducting and trapped-ion hardware, with a tunable classical-to-quantum interpolation parameter a. The network is trained classically but inference is run on hardware. The central claims are: (i) at moderate a, hardware validation rates exceed those of the classical a=0 network; (ii) for an ambiguous MNIST image (index 6929) that fails classically, hardware at a=0 sometimes classifies correctly due to physical noise; and (iii) injected single- and two-qubit gate pairs can, in some regimes, improve rather than degrade performance. The paper also presents a noise-injection benchmarking method. The experimental results are compared with QASM simulation and with the prior simulations of Ref. [5].

Significance. If the claims are established, this would be one of the few hardware demonstrations of a quantum neural network outperforming its classical counterpart on a standard image classification task, and it would provide a concrete example of noise-assisted quantum machine learning. The paper is also useful as a multi-platform benchmark (superconducting versus two trapped-ion gate implementations) with publicly available code. However, the statistical basis is thin: the main aggregate scan uses only 55 images with 10 shots each, the detailed noise-benefit analysis relies on a single image with 10 shots, and the paper provides no direct noise characterization. The result is therefore a promising proof-of-concept rather than a definitive demonstration, and the strongest claims in the abstract go beyond what the data currently support.

major comments (4)
  1. [Fig. 2a and following paragraph] The claim that the experimental validation rate 'slightly exceeds' QASM simulation for a ≤ 0.5 and that 'its presence can only be attributed to physical noise' is internally inconsistent with the same sentence's acknowledgement that 'the difference is within error bars'. A within-error-bar excess cannot establish the presence of a noise benefit, and it certainly cannot support the abstract's statement that physical noise 'is shown to improve' BQNN performance. Please provide the number of shots per point, confidence intervals, and either a direct noise measurement or a statistical test that quantifies the evidence for a hardware advantage.
  2. [Fig. 2b and 'Perhaps the most interesting observation' paragraph] The a=0 hardware successes on NY image 6929 (1/10 for the trapped-ion device and 5/10 for the IBM device) are each based on only 10 shots. The binomial confidence intervals for 1/10 and 5/10 overlap zero and overlap each other, so these data are statistically indistinguishable from sampling fluctuation. No calibration data (readout error rates, gate fidelities, or qubit coherence times) are reported for the specific runs, so the attribution to 'physical noise' cannot be distinguished from deterministic calibration bias or readout bias. The 'two nearby minima' picture is plausible but currently unfalsified by the reported data.
  3. [Fig. 2c and Fig. 3] The noise-injection scans that support the 'noise can be beneficial' conclusion are performed on a single image (NY 6929 and one YY image) and are plotted without error bars, shot counts, or replicates. The non-monotonic behavior in Fig. 3 (e.g., the trapped-ion jump after one two-qubit gate pair and the IBM improvement with additional pairs) is presented as evidence for a mechanism, but with no statistical uncertainty it is impossible to assess whether these features are reproducible or are fluctuations. These data need error bars, repeated runs, or a per-image statistical analysis before the conclusion can be accepted.
  4. [First paragraph after Fig. 2b] The statement that 'The NY images are the source of the boost of performance upon quantization' is a causal claim about the aggregate improvement in Fig. 2a, but Fig. 2a shows only the aggregate validation rate over 55 images. No per-image breakdown is provided to demonstrate that the improvement is concentrated in NY images rather than spread across the full test set. Please provide a per-image analysis or at least a comparison of accuracy on NY vs. non-NY images at different a values.
minor comments (5)
  1. [Equation (4)] Equation (4) appears to have a formatting error: the k=1 case reads 'W 1I' which should likely be 'W^1 I', and the parenthesis structure is unclear. Please correct the equation and ensure the argument of the activation function is written unambiguously.
  2. [After Eq. (5)] The phrase 'arbitrary angels of rotation' is a typo; it should be 'arbitrary angles of rotation'.
  3. [Second paragraph of Section after Fig. 2b] The sentence 'This predictably hinders classification performance and we observe validation rates dropping with with a' contains a duplicated 'with'.
  4. [Fig. 2c caption] The x-axis label '# U U†' does not define what U is or how it relates to the parameter n described in the text; please clarify in the caption that U denotes a native single-qubit or two-qubit gate and that n pairs are inserted after each layer.
  5. [Abstract and Introduction] The abstract states that 'Increasing [a] ... is shown to improve network performance', but the demonstration relies on a small sample (55 images, 10 shots each) with error bars that are not shown in Fig. 2a. Please add error bars or confidence intervals to the aggregate validation-rate plots, or soften the wording to reflect the statistical strength of the evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the hardware measurements provide independent external evidence for the prior same-group prediction, and no fitted parameter is renamed as a prediction.

full rationale

The paper's central claim is that the BQNN, implemented on trapped-ion and superconducting hardware, outperforms its classical counterpart at moderate values of the interpolation parameter a, and that certain physical noise can be beneficial. The architecture and the intermediate-a advantage originate in Ref. [5], which is co-authored by the same group, but this paper does not merely restate that prior prediction: it reports new experimental classification data from multiple hardware platforms, and the parameter a is scanned rather than fitted. No equation in the paper reduces a measured quantity to a fitted parameter, and no uniqueness theorem or machine-checked result is invoked to forbid alternatives. The paper's own admission that the hardware-QASM difference at a <= 0.5 is 'within error bars' undermines the strength of the 'physical noise helps' attribution, but that is a statistical and inferential weakness, not a circularity: the claim is not true by construction, and the authors do not define physical noise in terms of the classification outcomes they use it to explain. The selection of an NY image is definitional (an image that fails classically and succeeds quantumly), but the hardware behavior on that image is measured, not imposed by the definition. Because the load-bearing experimental facts are external to the prior self-citation, the derivation chain is not circular and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper mostly borrows its theory and trained weights from the same group's prior work, adds scanned parameters a and n, and introduces an interpretive energy-landscape explanation post hoc. No new physical entity is postulated.

free parameters (4)
  • Quantumness parameter a = scanned; optimal around 0.5
    Controls the classical-to-quantum interpolation. The claim of improved performance at moderate a depends on choosing a in this range, with the optimum read off from data rather than derived.
  • Network weights W^k = trained on MNIST, values not given in this paper
    The neural network parameters are trained via simulation in Ref. [5] and imported here. They are fitted to the MNIST training set and are not independently derived in this work.
  • Classification samples per image = 10
    The majority-vote classification uses 10 measurement shots per image. This choice affects the variance of the reported validation rates and is set by hand.
  • Injected noise gate pairs n = scanned, e.g., 0, 1, 5, 10, 15, 20
    The number of inserted U U-dagger pairs is scanned in the noise experiments. Observations of improvement at specific n values, such as 20 pairs on the trapped-ion device, are post hoc.
assumptions (4)
  • domain assumption At a=0, ideal noiseless circuits reproduce the classical sign-activation network exactly
    Equations (4) and (5) imply this for ideal rotations and projective measurements, assuming gates and measurements are perfect and hardware qubits behave as ideal two-level systems.
  • domain assumption Any experimental deviation from ideal simulation at fixed a is due to incoherent physical noise
    The paper attributes improved a=0 accuracy on the NY image to physical noise without directly characterizing gate fidelities, readout errors, or calibration biases.
  • ad hoc to paper The two-minima energy landscape picture for NY images is valid
    The explanation of noise-induced switching between nearby minima is invoked in the discussion of Fig. 2b to interpret the data. No measurement of the landscape is provided and the picture is chosen post hoc.
  • domain assumption Majority vote over 10 samples gives a reliable estimate of classification probability
    The reported validation rates use 10 shots per image, and the sampling error is not quantified in the figures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking a Tunable Quantum Neural Network on Trapped-Ion and Superconducting Hardware." pith.science (2026). https://pith.science/paper/FAQJ6JM6

@misc{pith2026250721222,
  author       = {Pith},
  title        = {Pith review of: Benchmarking a Tunable Quantum Neural Network on Trapped-Ion and Superconducting Hardware},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FAQJ6JM6}},
  note         = {Machine review of arXiv:2507.21222}
}
abstract

We implement a quantum generalization of a neural network on trapped-ion and IBM superconducting quantum computers to classify MNIST images, a common benchmark in computer vision. The network feedforward involves qubit rotations whose angles depend on the results of measurements in the previous layer. The network is trained via simulation, but inference is performed experimentally on quantum hardware. The classical-to-quantum correspondence is controlled by an interpolation parameter, $a$, which is zero in the classical limit. Increasing $a$ introduces quantum uncertainty into the measurements, which is shown to improve network performance at moderate values of the interpolation parameter. We then focus on particular images that fail to be classified by a classical neural network but are detected correctly in the quantum network. For such borderline cases, we observe strong deviations from the simulated behavior. We attribute this to physical noise, which causes the output to fluctuate between nearby minima of the classification energy landscape. Such strong sensitivity to physical noise is absent for clear images. We further benchmark physical noise by inserting additional single-qubit and two-qubit gate pairs into the neural network circuits. Our work provides a springboard toward more complex quantum neural networks on current devices: while the approach is rooted in standard classical machine learning, scaling up such networks may prove classically non-simulable and could offer a route to near-term quantum advantage.

Figures

Figures reproduced from arXiv: 2507.21222 by the authors.

Figure 1
Figure 1. FIG. 1. (a) The mapping from the classical binarized multi-layer perceptron network to the quantum circuit implementing [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. (a) Validation rates on 55 randomly selected test images obtained from superconducting hardware and simulation for [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. For an NY image at [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 21 canonical work pages

  1. [5]

    Natural Quantization of Neural Networks

    R. Barney, D. Lakhdar-Hamina, and V. Galitski, “Natural quantization of neural networks,” (2025), arXiv:2503.15482 [quant-ph]

    work page Pith review arXiv 2025

  2. [1]

    J. J. Hopfield, Proceedings of the National Academy of Sciences 79, 2554 (1982)

    work page 1982

  3. [2]

    D. J. Amit, H. Gutfreund, and H. Sompolinsky, Phys. Rev. A 32, 1007 (1985)

    work page 1985

  4. [3]

    Dotsenko, An introduction to the theory of spin glasses and neural networks, World Scientific lecture notes in physics, Vol

    V. Dotsenko, An introduction to the theory of spin glasses and neural networks, World Scientific lecture notes in physics, Vol. 55 (World Scientific, Singapore, 1994)

    work page 1994

  5. [4]

    Neural networks as spin models: From glass to hidden order through train- ing,

    R. Barney, M. Winer, and V. Galitski, “Neural networks as spin models: From glass to hidden order through train- ing,” (2024), arXiv:2408.06421 [cond-mat.dis-nn]

    arXiv 2024

  6. [6]

    K. L. Pudenz and D. A. Lidar, Quantum Information Processing 12, 2027 (2013)

    work page 2013

  7. [7]

    Quantum al- gorithms for supervised and unsupervised machine learn- ing,

    S. Lloyd, M. Mohseni, and P. Rebentrost, “Quantum al- gorithms for supervised and unsupervised machine learn- ing,” (2013), arXiv:1307.0411 [quant-ph]

    arXiv 2013

  8. [8]

    Z. Li, X. Liu, N. Xu, and J. Du, Phys. Rev. Lett. 114, 140504 (2015)

    work page 2015

Show all 32 references
  1. [9]

    Ciliberto, M

    C. Ciliberto, M. Herbster, A. D. Ialongo, M. Pontil, A. Rocchetto, S. Severini, and L. Wossnig, Proc. Math. Phys. Eng. Sci. 474, 20170551 (2018)

    2018

  2. [10]

    Schuld, I

    M. Schuld, I. Sinayskiy, and F. Petruccione, Contempo- rary Physics 56, 172 (2015)

    2015

  3. [11]

    Biamonte, P

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

    2017

  4. [12]

    Wang and J

    Y. Wang and J. Liu, Reports on Progress in Physics 87, 116402 (2024)

    2024

  5. [13]

    Quantum embeddings for machine learning,

    S. Lloyd, M. Schuld, A. Ijaz, J. Izaac, and N. Killoran, “Quantum embeddings for machine learning,” (2020)

    2020

  6. [14]

    Preskill, Quantum 2, 79 (2018)

    J. Preskill, Quantum 2, 79 (2018)

    2018

  7. [15]

    Abbas, D

    A. Abbas, D. Sutter, C. Zoufal, A. Lucchi, A. Figalli, and S. Woerner, Nature Computational Science 1, 403 (2021)

    2021

  8. [16]

    Huang, Y

    H.-L. Huang, Y. Du, M. Gong, Y. Zhao, Y. Wu, C. Wang, S. Li, F. Liang, J. Lin, Y. Xu, R. Yang, T. Liu, M.-H. Hsieh, H. Deng, H. Rong, C.-Z. Peng, C.-Y. Lu, Y.-A. Chen, D. Tao, X. Zhu, and J.-W. Pan, Phys. Rev. Appl. 16, 024051 (2021)

    2021

  9. [17]

    Quantum-enhanced llm efficient fine tuning,

    X. Kong, L. Li, Z. Chen, C. Xue, X. Xu, H. Liu, Y. Wu, Y. Fang, H. Fang, K. Chen, Y. Yang, M. Dou, and G. Guo, “Quantum-enhanced llm efficient fine tuning,” (2025), arXiv:2503.12790 [quant-ph]

    2025 arXiv

  10. [18]

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

    2024

  11. [19]

    Quantum convolutional neural networks are (effectively) classically simulable,

    P. Bermejo, P. Braccia, M. S. Rudolph, Z. Holmes, L. Cincio, and M. Cerezo, “Quantum convolutional neural networks are (effectively) classically simulable,” (2024), arXiv:2408.12739 [quant-ph]

    2024 arXiv

  12. [20]

    Gil-Fuster, C

    E. Gil-Fuster, C. Gyurik, A. Perez-Salinas, and V. Dun- jko, in The Thirteenth International Conference on Learning Representations(2025)

    2025

  13. [21]

    Y. Li, X. Chen, and M. P. A. Fisher, Phys. Rev. B 100, 134306 (2019)

    2019

  14. [22]

    Gradient- based learning,

    I. Goodfellow, Y. Bengio, and A. Courville, “Gradient- based learning,” in Deep Learning (MIT Press, Cambridge, MA, 2016) pp. 177–191, http://www. deeplearningbook.org

    2016

  15. [23]

    Estimat- ing or propagating gradients through stochastic neurons for conditional computation,

    Y. Bengio, N. L´ eonard, and A. Courville, “Estimat- ing or propagating gradients through stochastic neurons for conditional computation,” (2013), arXiv:1308.3432 [cs.LG]

    2013 arXiv

  16. [24]

    Courbariaux, Y

    M. Courbariaux, Y. Bengio, and J.-P. David, in Advances in Neural Information Processing Systems , Vol. 28, edited by C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett (Curran Associates, Inc., 2015)

    2015

  17. [25]

    NIST special database 19: Handprinted forms and charac- ters database,

    National Institute of Standards and Technology, “NIST special database 19: Handprinted forms and charac- ters database,” Available at https://www.nist.gov/ srd/nist-special-database-19 (1990), accessed: 2025- 06-12

    1990

  18. [26]

    Blais, R.-S

    A. Blais, R.-S. Huang, A. Wallraff, S. M. Girvin, and R. J. Schoelkopf, Phys. Rev. A 69, 062320 (2004)

    2004

  19. [27]

    J. Koch, T. M. Yu, J. Gambetta, A. A. Houck, D. I. Schuster, J. Majer, A. Blais, M. H. Devoret, S. M. Girvin, and R. J. Schoelkopf, Phys. Rev. A 76, 042319 (2007)

    2007

  20. [28]

    Natural quantization of neural net- works,

    R. Barney, “Natural quantization of neural net- works,” https://github.com/Galitski-theory-group/ quantum-nn (2025)

    2025

  21. [29]

    Natural-Quantization: A hybrid classical–quantum neural network frame- work,

    D. Lakhdar-Hamina, “Natural-Quantization: A hybrid classical–quantum neural network frame- work,” https://github.com/Galitski-theory-group/ natural-quantization (2025)

    2025

  22. [30]

    Debnath, N

    S. Debnath, N. M. Linke, C. Figgatt, K. A. Landsman, K. Wright, and C. Monroe, Nature 536, 63 (2016)

    2016

  23. [31]

    C. M. Bishop, Neural Computation 7, 108 (1995)

    1995

  24. [32]

    Tomesh, P

    T. Tomesh, P. Gokhale, V. Omole, G. S. Ravi, K. N. Smith, J. Viszlai, X.-C. Wu, N. Hardavellas, M. R. Martonosi, and F. T. Chong, in 2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA) (2022) pp. 587–603

    2022

Pith tools

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