Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Benchmarking Quantum Convolutional Neural Networks for Classification and Data Compression Tasks

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A 17-parameter QCNN matches a 113-parameter ansatz at about one-tenth the training time per sample.

desk verdict Plausible benchmark showing a parameter-light RY-QCNN matches larger HEAs, but the headline speed/accuracy comparison lacks error bars and matched budgets. read the letter →

arxiv 2411.13468 v1 pith:GLLBOY3S submitted 2024-11-20 quant-ph

Jun Yong Khoo , Chee Kwan Gan , Wenjun Ding , Stefano Carrazza , Jun Ye , Jian Feng Kong This is my paper · ORCID
classification quant-ph
keywords quantumconvolutionalneuralnetworkshardware-efficientansatzphaseclassificationautoencodertransversefieldIsingmodelXXZmachinelearningtrainability
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 benchmarks quantum convolutional neural networks (QCNNs) against hardware-efficient ansatze (HEAs) on two tasks: classifying ground-state phases of the transverse-field Ising and XXZ models, and compressing Ising ground states with quantum autoencoders. The central result is that a QCNN restricted to RY rotation gates, with only 17 trainable parameters, reaches essentially the same test accuracy as the best HEA (0.931 versus 0.938 at 16 qubits) while costing roughly one tenth of the training time per sample. More expressive QCNN variants perform worse, which the authors attribute to trainability problems rather than expressive power. For equal parameter counts, the RY-QCNN is the most efficient architecture on these benchmarks, and the same parameter-count advantage explains faster convergence in the compression task.

What carries the argument

The central object is the QCNN circuit, whose convolutional and pooling layers halve the qubit count at each step, giving logarithmic depth and a structured reduction in trainable parameters; the RY-restricted variant keeps only real-valued rotation gates, cutting parameters from 85 (complex QCNN) to 17. The HEA comparator is a layered ansatz of single-qubit RY rotations and entangling gates whose parameter count grows with layers. The argument runs through parameter count: fewer parameters shrink the optimization landscape and per-sample cost, and the logarithmic-depth QCNN structure is the reason trainability is preserved. For compression, the autoencoder uses the encoder circuit's inverse as the decoder, with a cost function that drives the discarded qubits to $|0\rangle^{\otimes n_d}$.

What would settle it

Recompute the 16-qubit classification using exact ground states (or high-fidelity tensor-network states) of the TFI and XXZ models at the same field values; if the RY-QCNN's accuracy falls below the HEA's once the labels are exact, the paper's flagship comparison fails.

Watch

Extended reading notes

Core claim

The paper claims that for phase classification of TFI and XXZ ground states on 4, 8, and 16 qubits, a QCNN built from RY gates matches the best-performing HEA in accuracy while needing far fewer parameters and much less training time. Its flagship comparison at 16 qubits is a 17-parameter RY-QCNN at 0.931 test accuracy and 18.7 seconds per training sample versus a 113-parameter three-layer HEA at 0.938 accuracy and 188.7 seconds per sample. The authors also claim that increasing QCNN expressibility beyond this real-valued form degrades performance, while HEAs improve with more layers at linearly increasing cost. In the autoencoder task, both architectures reconstruct TFI ground states with high fidelity, but QCNNs converge faster because of their smaller parameter counts.

Load-bearing premise

The 16-qubit dataset labels come from optimized variational circuits bundled with a quantum machine-learning software repository, not from exact ground states; if those circuits deviate from the true TFI and XXZ ground states, the 16-qubit accuracies compare classifiers on the wrong labels.

Editorial extensions

If this is right

  • If the benchmark is representative, RY-QCNN achieves near-identical accuracy to the best HEA with a roughly tenfold reduction in per-sample training time at 16 qubits.
  • For equal trainable-parameter budgets, the RY-QCNN is the most efficient architecture for classifying these ground states.
  • The more expressive QCNN variants' poorer accuracy is a trainability effect, not a capacity limit, so architecture choice should weigh optimizability, not just expressibility.
  • On the compression task, QCNN-based autoencoders reconstruct TFI ground states with high fidelity while converging faster than HEA-based autoencoders.

Reading between the lines

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

  • If the training-time advantage survives on real hardware, the RY-QCNN's low parameter count makes it a practical default for near-term devices; the paper only reports simulation and lists hardware comparison as future work.
  • The 16-qubit result rests on variational proxy ground states rather than exact diagonalization, so the 16-qubit ranking should be read as provisional until exact-state labels confirm it.
  • The real-valued RY restriction is motivated by the fact that these ground states are real; on Hamiltonians whose ground states require complex phases, the RY-QCNN advantage may shrink or disappear.
  • A natural testable extension is to run the same benchmark on disordered or long-range models, where the QCNN's locality assumption may fail.
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

5 major / 6 minor

Summary. The paper benchmarks quantum convolutional neural networks (QCNNs) against hardware-efficient ansätze (HEAs) on two tasks: phase classification of transverse-field Ising (TFI) and XXZ ground states at 4, 8, and 16 qubits, and quantum autoencoder compression of TFI ground states. The authors propose real-valued QCNN variants, including an RY-gate version, and report that the 17-parameter QCNN(RY) reaches test accuracy 0.931 at 16 qubits for 60 training samples, close to the 113-parameter HEA(3L) accuracy of 0.938, while taking 18.7 s per training sample versus 188.7 s. They conclude that QCNNs with RY gates are the most parameter-efficient architecture and also converge faster in the autoencoder task. The manuscript reports no error bars or repeated-seed statistics, gives limited experimental detail for the training-time comparison, and includes only one quantitative figure/table in the visible text.

Significance. If the performance comparison were statistically and experimentally validated, the result would be practically useful: it would identify a low-parameter QCNN architecture that matches a higher-parameter HEA on phase classification while reducing training time, and it would provide evidence for QCNN advantages in quantum data compression. Strengths of the paper include the use of established simulation tools (Qibo), explicit parameter counts for all models, comparison against a standard HEA baseline, and clear statements of the cost functions. The paper does not claim a theoretical advance; its contribution is empirical. The main barrier to significance is the lack of statistical grounding and the unverified 16-qubit input states.

major comments (5)
  1. [Sec. III, Fig. 2 and Table I] The claim that QCNN(RY) matches HEA(3L) performance rests on a single unseeded comparison (0.931 vs 0.938 at 60 training samples). Because quantum circuit training is stochastic, this 0.7 percentage point gap is plausibly within run-to-run variance; without error bars, multiple random initializations, or a significance test, 'matching performance' is not established. This is load-bearing for the paper's central conclusion.
  2. [Sec. III, Fig. 2b and Table I] The reported 10x training-time advantage is not grounded because the training time per sample is presented without specifying the optimizer settings, convergence criteria, or number of epochs, and without evidence that the QCNN and HEA runs were given equal optimization budgets. If the models were trained for different durations or stopped by different rules, the time comparison conflates parameter count with training budget. The authors should report wall-clock time to reach a fixed target accuracy, or per-epoch time under matched budgets, across multiple seeds.
  3. [Sec. II.C] The 16-qubit datasets are described as 'optimized variational circuits provided in the tensorflow-quantum repository' rather than exact ground states of the TFI and XXZ models. The paper does not verify how closely these circuits approximate the true ground states, so the 16-qubit accuracies in Fig. 2 and Table I may not correspond to the claimed phase-classification benchmark. This is a concern for external validity even if the internal QCNN-versus-HEA comparison remains meaningful.
  4. [Sec. II.B and Sec. III] The data-compression section contains no quantitative results: no reconstruction fidelities, learning curves, parameter counts, or training times are reported for the autoencoder experiments. The claims that QCNNs 'demonstrated faster training convergence' and exhibit 'minimal tradeoff in compression capability' are therefore unsupported in the manuscript as written.
  5. [Sec. I and Sec. III] The paper states that system sizes of 4, 8, and 16 qubits were examined, but Fig. 2 reports results only for 16 qubits; the 4- and 8-qubit classification results and the autoencoder results are not shown. To support the claimed scaling behavior and the comparison across system sizes, these results should be presented or the scope statement should be revised.
minor comments (6)
  1. [Throughout] There are formatting issues: Eqs. (1) and (2) render with broken spacing, the inner product in the label prediction is missing angle brackets, and 'nd = N (1 − 1/2l)' should read 'nd = N(1 − 1/2^l)'.
  2. [Table I] The header says 'best QCNN and VQC models'; 'VQC' should likely be 'HEA' to match the text.
  3. [Abstract] The phrase 'through simulation were examined' is grammatically incomplete and should be revised.
  4. [Sec. II.C and Fig. 2b] The paper reports 'training time per training sample (s)' but does not define how this is measured; please clarify the measurement procedure, hardware, and software versions.
  5. [Sec. IV] The statement that 'Powell is one of the most reliable optimizers' is not supported by any reported optimizer comparison; either add the comparison data or soften the claim.
  6. [Sec. II.A] The exact architecture details of QCNN, QCNN(real), QCNN(RY), and HEA are not fully specified; a circuit diagram or explicit gate-layer list for each variant would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the QCNN-vs-HEA comparison is an external empirical benchmark with architectural parameter counts, not a derived prediction that reduces to fitted inputs or to a load-bearing self-citation.

full rationale

The paper's central claims are empirical: QCNN with RY gates matches HEA accuracy while training faster due to fewer trainable parameters. These claims are supported by simulation experiments on external benchmarks (TFI and XXZ ground states), not by a derivation whose output is equivalent to its input. The number of trainable parameters is an architectural choice, not a fitted constant, and the accuracy comparison is an independent measurement against a baseline ansatz. The only self-citations are to software libraries Qibo and Qibolab, co-authored by one of the present authors; these are computational tools used to run the experiments, not evidence invoked to justify the scientific conclusion, so they are not load-bearing. The use of optimized variational circuits from the tensorflow-quantum repository for the 16-qubit dataset raises a possible external-validity concern about whether those states are true ground states, but it does not make the argument circular: the dataset construction does not presuppose the relative performance of QCNN versus HEA. No equation, fitted parameter, or cited result is shown to be identical by construction to the claimed outcome, so there is no significant circularity.

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

No free parameters or invented entities appear in this benchmark paper; the only tunable numbers are trained circuit weights and the chosen model parameter counts, which are architectural design choices rather than fitted constants in a derivation. The key assumptions are about dataset fidelity and labeling.

assumptions (3)
  • domain assumption The TFI and XXZ ground states are correctly labeled by phase across the selected ranges of h.
    Section II.B defines the classification task using the phase boundaries; the labels are taken as ground truth for training and testing.
  • standard math Exact diagonalization yields the correct 4- and 8-qubit ground states.
    Section II.C states exact diagonalization is used for these sizes; this is a standard numerical method but the fidelity is not quantified.
  • domain assumption The TFQ variational circuits used for 16 qubits faithfully approximate true ground states.
    Section II.C substitutes TFQ optimized variational circuits for exact ground states at 16 qubits without reporting fidelity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking Quantum Convolutional Neural Networks for Classification and Data Compression Tasks." pith.science (2026). https://pith.science/paper/GLLBOY3S

@misc{pith2026241113468,
  author       = {Pith},
  title        = {Pith review of: Benchmarking Quantum Convolutional Neural Networks for Classification and Data Compression Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GLLBOY3S}},
  note         = {Machine review of arXiv:2411.13468}
}
read the original abstract

Quantum Convolutional Neural Networks (QCNNs) have emerged as promising models for quantum machine learning tasks, including classification and data compression. This paper investigates the performance of QCNNs in comparison to the hardware-efficient ansatz (HEA) for classifying the phases of quantum ground states of the transverse field Ising model and the XXZ model. Various system sizes, including 4, 8, and 16 qubits, through simulation were examined. Additionally, QCNN and HEA-based autoencoders were implemented to assess their capabilities in compressing quantum states. The results show that QCNN with RY gates can be trained faster due to fewer trainable parameters while matching the performance of HEAs.

Figures

Figures reproduced from arXiv: 2411.13468 by the authors.

Figure 1
Figure 1. FIG. 1: Framework of quantum phase classification (left) and data compression (right) for 4 qubits. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2: The results of (a) test accuracy for 16 qubits as a function of number of training samples, with their [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Inherent interpretability provides inherent value in quantum machine learning

    quant-ph 2026-07 conditional novelty 4.0 of 10

    Quantum machine learning models can be valuable for their inherent mathematical interpretability, and quantum Fourier models offer a bottom-up route to Gaussian-process kernel design that random Fourier features do no...

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [1]

    Quan- tum convolutional neural networks

    Iris Cong, Soonwon Choi, and Mikhail D Lukin. Quan- tum convolutional neural networks. Nature Physics , 15(12):1273–1278, December 2019

    work page 2019

  2. [2]

    Quantum con- volutional neural network for classical data classification

    Tak Hur, Leeseok Kim, and Daniel K Park. Quantum con- volutional neural network for classical data classification. Quantum Machine Intelligence , 4(1):3, February 2022

    work page 2022

  3. [3]

    Cerezo, Samson Wang, Tyler Volkoff, Andrew T

    Arthur Pesah, M. Cerezo, Samson Wang, Tyler Volkoff, Andrew T. Sornborger, and Patrick J. Coles. Absence of barren plateaus in quantum convolutional neural net- works. Phys. Rev. X , 11:041011, Oct 2021

    work page 2021

  4. [4]

    Quantum autoencoders for efficient compression of quantum data

    Jonathan Romero, Jonathan P Olson, and Alan Aspuru- Guzik. Quantum autoencoders for efficient compression of quantum data. Quantum Science and Technology , 2(4):045001, aug 2017

    work page 2017

  5. [5]

    Quantum autoencoders with en- hanced data encoding

    Carlos Bravo-Prieto. Quantum autoencoders with en- hanced data encoding. Machine Learning: Science and Technology, 2(3):035028, jul 2021

    work page 2021

  6. [6]

    Qibo: a framework for quantum simulation with hardware acceleration

    Stavros Efthymiou, Sergi Ramos-Calderer, Carlos Bravo- Prieto, Adrián Pérez-Salinas, Diego García-Martín, Ar- tur Garcia-Saez, José Ignacio Latorre, and Stefano Car- razza. Qibo: a framework for quantum simulation with hardware acceleration. Quantum Science and Technology, 7(1):015018, dec 2021

    work page 2021

  7. [7]

    Martinez, Jae Hyeon Yoo, Sergei V

    Michael Broughton, Guillaume Verdon, Trevor McCourt, Antonio J. Martinez, Jae Hyeon Yoo, Sergei V. Isakov, Philip Massey, Ramin Halavati, Murphy Yuezhen Niu, Alexander Zlokapa, Evan Peters, Owen Lockwood, An- drea Skolik, Sofiene Jerbi, Vedran Dunjko, Martin Leib, Michael Streif, David Von Dollen, Hongxiang Chen, Shux- iang Cao, Roeland Wiersema, Hsin-Yua...

    work page 2021

  8. [8]

    Qibolab: an open-source hybrid quantum operating system.Quantum, 8:1247, February 2024

    Stavros Efthymiou, Alvaro Orgaz-Fuertes, Rodolfo Carobene, Juan Cereijo, Andrea Pasquale, Sergi Ramos- Calderer, Simone Bordoni, David Fuentes-Ruiz, Alessan- dro Candido, Edoardo Pedicillo, Matteo Robbiati, Yuanzheng Paul Tan, Jadwiga Wilkens, Ingo Roth, José Ignacio Latorre, and Stefano Carrazza. Qibolab: an open-source hybrid quantum operating system.Qu...

    work page 2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.