Pith. sign in

REVIEW 4 major objections 8 minor 57 references

Boosting Classification with Quantum-Inspired Augmentations

T0 review · 4 major / 8 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 The paper claims that applying small random Bloch-sphere rotations to amplitude-encoded images, then discarding the imaginary part, improves ImageNet classification beyond classical augmentation alone, and that the benefit comes from the…

desk verdict Quantum-inspired augmentation idea is interesting, but the accuracy claim is undercut by per-method hyperparameter tuning and a mismatch with the reported table. read the letter →

arxiv 2506.22241 v1 pith:DFV4UZEO submitted 2025-06-27 cs.CV cond-mat.dis-nncs.LGquant-ph

Matthias Tschöpe , Vitor Fortes Rey , Sogo Pierre Sanon , Paul Lukowicz , Nikolaos Palaiodimopoulos , Maximilian Kiefer-Emmanouilidis This is my paper · ORCID
classification cs.CVcond-mat.dis-nncs.LGquant-ph
keywords quantum-inspiredaugmentationdataBlochrotationsamplitudeencodingImageNetsingularvaluespectrumResNet-34differentialprivacy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that small random Bloch-sphere rotations—the kind of perturbation that appears as gate noise on quantum hardware—can be turned into a data augmentation that improves classical image classification. On ImageNet with a ResNet-34, applying a random Z-axis rotation to the amplitude-encoded image and keeping only the real part, after classical flip and perfect rotation, raises Top-1 accuracy to 63.40% and F1 to 12.51%, compared with 61.47% and 7.93% for the best classical augmentation. The authors argue the improvement is not due to the unitary rotation itself but to the non-unitary projection (real() or abs()), which changes the image's singular value spectrum; rotations that preserve singular values show no significant gain. They also show the augmentation can be computed in O(N log N) on a classical computer, and that it does not provide differential privacy.

What carries the argument

The central object is the amplitude-encoded image state |Image>_f, flattened into a 2^n-dimensional vector, acted on by tensor products of single-qubit Bloch rotations RX, RY, RZ with angles drawn from [0, Θ]. Because the rotations are separable SU(2) operations, they leave the singular values of the state invariant; the augmentation's effect is created by the subsequent element-wise real() or abs() projection, which breaks that invariance and reshapes the singular value spectrum. Computationally this is a sequence of n reshape-and-multiply steps, giving O(N log N) complexity, and for QRZ the action reduces to an element-wise multiplication by cosines of Gaussian-distributed phase sums.

What would settle it

Train the best classical method F(PR(x)) and the best combined method real(QRZ(F(PR(x)))) with identical hyperparameters across several random seeds; if the Top-1 gap shrinks to within the run-to-run variance, the central claim that the quantum-inspired step boosts classification would collapse.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a quantum-inspired augmentation built from single-qubit Bloch rotations on an amplitude-encoded image can outperform standard classical augmentations on ImageNet when the rotation is followed by a non-unitary projection. The best method, real(QRZ(F(PR(x)))), combines the classical perfect-rotation-plus-flip augmentation with a random Z-axis rotation whose real part is kept; it reaches 63.40% Top-1, 85.35% Top-5, and 12.51% F1. The authors isolate the mechanism by comparing with singular-value-preserving rotations such as real(QRY(x)) and QRXYZ(x), which stay near baseline, concluding that the benefit comes from the projection disturbing the singular value spectrum—an effect they describe as a positional, dropout-like masking—rather than from the unitary rotation itself.

Load-bearing premise

The reported accuracy gaps are assumed to reflect the augmentation itself, even though each method was tuned with its own hyperparameter search and no repeated runs or error bars are reported.

Editorial extensions

If this is right

  • If the result holds, quantum-inspired augmentation can be deployed on classical hardware at low overhead, since the full pipeline is O(N log N) and needs no actual quantum device.
  • The mechanism implies that non-unitary projections that disturb the singular value spectrum are the active ingredient; future augmentation design can focus on such projections rather than on the choice of unitary rotation.
  • real(QRZ(F(PR(x)))) nearly doubles F1 relative to baseline (6.41% to 12.51%), suggesting the method helps most in the long tail of classes where F1 is low.
  • The negative privacy result means strong unitary augmentations should not be used as a substitute for differential privacy mechanisms, despite their visually unrecognizable output.
  • Combining quantum-inspired rotations with classical augmentations works in nested order, so the method is a plug-in addition to existing classical augmentation pipelines.

Reading between the lines

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

  • A testable extension: because real(QRZ) acts as a register-position-dependent cosine mask, a purely classical mask built from the same Gaussian-distributed phase sums should reproduce the accuracy gain, which would show the quantum framing is not the operative part.
  • The paper's single-run comparisons leave open whether the 2-3 point Top-1 gap survives seed averaging; an immediate experiment is to fix one hyperparameter set and run both F(PR(x)) and real(QRZ(F(PR(x)))) across several seeds.
  • If the spectral distortion mechanism is the true cause, similar gains should appear on other large-scale datasets and architectures, and the augmentation could be transferred to contrastive learning pipelines where composite transformations are the standard tool.
  • The privacy proof applies to the augmentation as a standalone map; it does not rule out combining the augmentation with differentially private training, which is a separate and untested direction.
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 / 8 minor

Summary. The paper proposes quantum-inspired data augmentations based on random Bloch-sphere rotations applied to amplitude-encoded image vectors, optionally combined with classical augmentations, and evaluates them on ImageNet with ResNet-34. The best reported method, real(QRZ(F(PR(x)))), is claimed to improve Top-1, Top-5, and F1 over the best classical augmentation F(PR(x)). The authors attribute the gain to non-unitary projections (real() or abs()) rather than to the unitary rotation itself, support this with singular-value spectrum analyses, give an O(N log N) implementation argument, and include a differential-privacy negative result in the supplementary material.

Significance. If the empirical claim survives a controlled comparison, the paper offers a simple, cheap augmentation with a concrete mechanistic story (singular-value distortion via non-unitary projection) and a useful negative result on differential privacy. The O(N log N) implementation and the singular-value invariance analysis are genuine strengths. However, the central evidence is a single table without error bars, repeated seeds, or statistical tests, and the per-method hyperparameter search confounds the augmentation effect; the current support for the headline claim is therefore weak.

major comments (4)
  1. [Methods/Implementation and Results/Table 1] The genetic-algorithm hyperparameter search is performed separately for each augmentation method over batch size, learning rate, Adam betas, weight decay, and learning-rate decay schedule. Because the training configuration is not held constant across methods, the accuracy differences in Table 1 can be explained by tuning artifacts rather than by the augmentation itself. The paper needs a controlled comparison in which hyperparameters are fixed across methods (or at least matched by search budget and seeds), with repeated runs and error bars, before the headline claim is supported.
  2. [Augmentation Methods and Methods/Implementation] The rotation strength Theta is only described as being 'in the order of 10^-2', and the paper does not report the exact Theta values, the random seeds, or the final hyperparameter configurations used for each method. This makes the experiments unreproducible and leaves open the possibility that the reported gains depend critically on a particular untested choice of Theta. Please report the exact augmentation parameters and the hyperparameter configuration for every row of Table 1.
  3. [Results/Table 1 and Discussion] No seeds, repeated runs, or statistical tests are reported, and the phrase 'within the margin of error' in the Results section refers to a margin that is never defined. The differences among the top combined methods (63.40%, 63.36%, 63.52% Top-1) are small relative to the uncontrolled setup, so the ranking is not statistically grounded. At minimum, the authors should report mean and standard deviation over multiple seeds and a paired significance test for the main comparisons.
  4. [Abstract, Introduction, Discussion, and Table 1] The reported improvements are numerically inconsistent. Table 1 shows real(QRZ(F(PR(x)))) at 63.40% Top-1 versus 61.47% for F(PR(x)), a 1.93 percentage-point difference; Top-5 is 85.35% versus 84.12%, a 1.23 percentage-point difference; and F1 is 12.51% versus 7.93%, a 4.58 percentage-point difference. The abstract's '3%, 2.5%, and from 8% to 12%' are not percentage-point changes, and if they are intended as relative changes, the Top-5 value is about 1.5%, not 2.5%. The Introduction states '2%, 1%' instead. The comparison baseline and the metric convention must be stated precisely and used consistently.
minor comments (8)
  1. [Introduction] The sentence 'and its unclear if the same procedures adapt to significantly harder datasets' contains a typo; it should read 'it is unclear'.
  2. [Quantum Augmentation] The term 'Trait-Bryan angles' should be 'Tait-Bryan angles'.
  3. [Quantum Augmentation] In Eq. (3), the notation with the primed product over the ordered tuple (X,Y,Z) is confusing; please clarify explicitly that the order of the non-commuting rotation gates is the randomly drawn order.
  4. [Algorithm Description and Complexity] The sentence 'Transpose Image f and flatten the image back to the state |Image>_f' is not a complete algorithmic specification; please state the reshape/transpose operation precisely, including the roles of the two index sets.
  5. [Singular Value Spectrum Analysis] There is a missing closing parenthesis in 'F(PR(x), in both figures' and a doubled word 'of of' in the Discussion; these should be corrected.
  6. [Supplementary A] The differential-privacy proof uses the symbol R both for the rotation operation and for the output state, and the sentence about the singular-value decomposition of R(Theta)|Psi> is unnecessary for the argument; please rephrase the proof to state directly that a deterministic map cannot satisfy epsilon-DP for finite epsilon.
  7. [Supplementary B] The term 'unit8' appears throughout; it should be 'uint8'.
  8. [Data availability] The statement 'Data and Code will be made available on reasonable request' is too vague for reproducibility; please provide a repository or at least a detailed description of the experimental configuration, including exact Theta values and hyperparameter settings.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the augmentation claims are empirical comparisons against ImageNet and are not reduced by construction to their inputs.

full rationale

No circularity found. The paper's central claim is an empirical comparison of augmentation methods on ImageNet validation data; the classification accuracies are measured outcomes, not quantities assumed in the definitions of the augmentations. The singular-value invariance argument around Eq. (5) follows directly from the Schmidt decomposition and the fact that single-qubit rotations act locally on the two partitions; it does not presuppose any classification result. The real(QRZ) derivation in Eq. (9) is a straightforward consequence of RZ being diagonal and the Hadamard product structure; it explains the visual artifacts but is not used to force the reported accuracy gains. The non-differential-privacy theorem in Supplementary A is proven constructively for any rotation angle and is therefore not an imported uniqueness claim. The only self-citations that appear (refs. 17, 23, 48) are used for background, optical-implementation motivation, and the genetic-algorithm hyperparameter search, respectively; none of them supplies the load-bearing claim that real(QRZ(F(PR(x)))) improves classification. The methodological concern about per-method hyperparameter search and the absence of seeds or error bars is a statistical-validity issue, not a circularity issue: the numbers in Table 1 are not fitted parameters renamed as predictions, and the comparison is made against an external benchmark. Accordingly, the paper is self-contained with respect to the circularity patterns considered here.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No invented physical entities are introduced. The central claim rests on the amplitude-encoding and normalization pipeline, on the singular-value preservation theorem, and on the comparability of the per-method hyperparameter search. The free parameters are the rotation strength and the undisclosed training hyperparameters; neither is fitted to the target accuracy.

free parameters (2)
  • Rotation strength Theta = on the order of 10^-2 (exact value not reported)
    Angles are sampled uniformly from [0, Theta] for each qubit and axis; Theta controls augmentation strength and was chosen by the authors.
  • Per-method training hyperparameters = not reported
    Batch size, learning rate, Adam betas, weight decay, and learning-rate decay schedule were tuned per method by a genetic algorithm; without these values, the augmentation effect is confounded.
assumptions (4)
  • domain assumption Amplitude encoding with min-max normalization preserves class-discriminative information after Bloch rotations.
    The method treats each flattened image as a non-normalized quantum state (Eq. 1), applies rotations, then rescales per channel to the original range (Eqs. 7-8); the paper does not prove this pipeline retains the information needed for classification.
  • standard math Single-qubit rotations cannot change the singular values of the amplitude-encoded image because no entangling gates are used.
    Proved via Schmidt decomposition in Eq. 5; used to argue that unitary-only augmentations do not alter spectral structure.
  • domain assumption The Irwin-Hall distribution of summed rotation angles is well approximated by a Gaussian for n > 10.
    Used to describe real(QRZ) as a cos of Gaussian-distributed phase in the Results section; this is a standard approximation but not exact for finite n.
  • ad hoc to paper Per-method genetic-algorithm hyperparameter optimization yields comparable training conditions across augmentation methods.
    No hyperparameter values are reported; the central comparison implicitly assumes that tuning does not drive the reported gains.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Boosting Classification with Quantum-Inspired Augmentations." pith.science (2026). https://pith.science/paper/DFV4UZEO

@misc{pith2026250622241,
  author       = {Pith},
  title        = {Pith review of: Boosting Classification with Quantum-Inspired Augmentations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DFV4UZEO}},
  note         = {Machine review of arXiv:2506.22241}
}
abstract

Understanding the impact of small quantum gate perturbations, which are common in quantum digital devices but absent in classical computers, is crucial for identifying potential advantages in quantum machine learning. While these perturbations are typically seen as detrimental to quantum computation, they can actually enhance performance by serving as a natural source of data augmentation. Additionally, they can often be efficiently simulated on classical hardware, enabling quantum-inspired approaches to improve classical machine learning methods. In this paper, we investigate random Bloch sphere rotations, which are fundamental SU(2) transformations, as a simple yet effective quantum-inspired data augmentation technique. Unlike conventional augmentations such as flipping, rotating, or cropping, quantum transformations lack intuitive spatial interpretations, making their application to tasks like image classification less straightforward. While common quantum augmentation methods rely on applying quantum models or trainable quanvolutional layers to classical datasets, we focus on the direct application of small-angle Bloch rotations and their effect on classical data. Using the large-scale ImageNet dataset, we demonstrate that our quantum-inspired augmentation method improves image classification performance, increasing Top-1 accuracy by 3%, Top-5 accuracy by 2.5%, and the F$_1$ score from 8% to 12% compared to standard classical augmentation methods. Finally, we examine the use of stronger unitary augmentations. Although these transformations preserve information in principle, they result in visually unrecognizable images with potential applications for privacy computations. However, we show that our augmentation approach and simple SU(2) transformations do not enhance differential privacy and discuss the implications of this limitation.

Figures

Figures reproduced from arXiv: 2506.22241 by the authors.

Figure 1
Figure 1. An example of the proposed augmentation circuit for three qubits. For each qubit i and axis j ∈ (X,Y,Z), a rotation angle θi, j is independently sampled from the uniform distribution [0,Θ], where Θ represents the strength of the rotation. In addition, the j’s are randomly drawn from the ordered tuple (X,Y,Z) determining the order in which single-qubit rotations around the X, Y, and Z axes are applied. The quantum pa… view at source ↗
Figure 2
Figure 2. Visualizations of the previously mentioned augmentation methods. Methods Model Architecture In this work, we utilize the ResNet34 architecture without pre-training, consisting of 34 layers. The model is organized into stages, each featuring residual blocks of increasing complexity. Details and the visualization of the architecture of the model are presented in the work by Kaiming He et al.34. Despite being a relativ… view at source ↗
Figure 3
Figure 3. This figure visualizes the losses (left) and accuracies respectively Top-1 scores (right) of all augmentation methods. In [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Results of the singular value decomposition for all previously mentioned augmentation methods, including the baseline, applied to the camera man image, already shown in [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Results of the singular value decomposition averaged over 100 images from the ImageNet dataset using all augmentation methods, without the normalization and the unit8 transformation. The sub-figure in the top shows the full range of averaged singular values, while the …
Figure 6
Figure 6. Figure 6: Overview of differential privacy with real(QRXY Z(x)) augmentation, using various epsilon and strength values. DP is often defined using the concept of neighbouring datasets. Let D and D ′ be two datasets that differ by at most one data point, and A be a randomized alg…
Figure 7
Figure 7. Figure 7: Results of the singular value decomposition for all augmentation methods, including the baseline, applied to the cameraman image. Normalization and the datatype conversion are applied after the augmentation. The figure in the top shows the full range of singular number…
Figure 8
Figure 8. Figure 8: Results of the singular value decomposition averaged over 100 images from the ImageNet dataset using all augmentation methods. Normalization and the datatype conversion are applied after the augmentation.The sub-figure in the top shows the full range of averaged singul…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 36 canonical work pages

  1. [1]

    E., Loey, M

    Khalifa, N. E., Loey, M. & Mirjalili, S. A comprehensive survey of recent trends in deep learning for digital images augmentation. Artif. Intell. Rev. 55, 2351–2377, DOI: 10.1007/s10462-021-10066-4 (2022)

    work page doi:10.1007/s10462-021-10066-4 2022

  2. [2]

    & Özuysal, M

    Ba¸ stanlar, Y . & Özuysal, M. Introduction to machine learning.miRNomics: MicroRNA biology computational analysis 105–128 (2014)

    work page 2014

  3. [3]

    Van Dyk, D. A. & Meng, X.-L. The art of data augmentation. J. Comput. Graph. Stat. 10, 1–50 (2001)

    work page 2001

  4. [4]

    & Khoshgoftaar, T

    Shorten, C. & Khoshgoftaar, T. M. A survey on image data augmentation for deep learning. J. big data 6, 1–48 (2019)

    work page 2019

  5. [5]

    & Park, D

    Xu, M., Yoon, S., Fuentes, A. & Park, D. S. A comprehensive survey of image augmentation techniques for deep learning. Pattern Recognit. 137, 109347 (2023)

    work page 2023

  6. [6]

    & Grochowski, M

    Mikołajczyk, A. & Grochowski, M. Data augmentation for improving deep learning in image classification problem. In 2018 international interdisciplinary PhD workshop (IIPhDW) , 117–122 (IEEE, 2018)

    work page 2018

  7. [7]

    Biamonte, J. et al. Quantum machine learning. Nature 549, 195–202 (2017)

    2017

  8. [8]

    & Ferrie, C

    Perrier, E., Youssry, A. & Ferrie, C. Qdataset, quantum datasets for machine learning. Sci. Data 9, 582, DOI: 10.1038/ s41597-022-01639-1 (2022)

    work page 2022

Show all 57 references
  1. [9]

    & Manoj, B

    Chalumuri, A., Kune, R., Kannan, S. & Manoj, B. S. Quantum–classical image processing for scene classification. IEEE Sensors Lett. 6, 1–4, DOI: 10.1109/LSENS.2022.3173253 (2022). 11/18

    2022

  2. [10]

    & Feld, S

    Apak, B., Bandic, M., Sarkar, A. & Feld, S. Ketgpt – dataset augmentation of nbsp;quantum circuits using transformers. In Computational Science – ICCS 2024: 24th International Conference, Malaga, Spain, July 2–4, 2024, Proceedings, Part VI , 235–251, DOI: 10.1007/978-3-031-637...

    2024 doi

  3. [11]

    & Chen, Y

    Gao, H., Lin, K., Cui, Y . & Chen, Y . Quantum assimilation-based data augmentation for state of health prediction of lithium-ion batteries with peculiar degradation paths. Appl. Soft Comput. 129, 109515, DOI: https://doi.org/10.1016/j.asoc. 2022.109515 (2022)

    2022

  4. [12]

    Data augmentation experiments with style-based quantum generative adversarial networks on trapped-ion and superconducting-qubit technologies (2024)

    Baglio, J. Data augmentation experiments with style-based quantum generative adversarial networks on trapped-ion and superconducting-qubit technologies (2024). 2405.04401

    2024 arXiv

  5. [13]

    & Khalil, I

    Kottahachchi Kankanamge Don, A. & Khalil, I. Q-supcon: Quantum-enhanced supervised contrastive learning architecture within the representation learning framework. ACM Transactions on Quantum Comput. 6, DOI: 10.1145/3660647 (2025)

    2025 doi

  6. [14]

    & Cook, T

    Henderson, M., Shakya, S., Pradhan, S. & Cook, T. Quanvolutional neural networks: Powering image recognition with quantum circuits, DOI: 10.48550/ARXIV .1904.04767 (2019)

    2019 doi arXiv

  7. [15]

    R., Izaac, J., Schuld, M

    Mari, A., Bromley, T. R., Izaac, J., Schuld, M. & Killoran, N. Transfer learning in hybrid classical-quantum neural networks. Quantum 4, 340, DOI: 10.22331/q-2020-10-09-340 (2020)

    2020 doi

  8. [16]

    & Stirenko, S

    Gordienko, Y ., Trochun, Y ., Taran, V ., Khmelnytskyi, A. & Stirenko, S. Hnn-qc n: Hybrid neural network with multiple backbones and quantum transformation as data augmentation technique. AI 6, 36 (2025)

    2025

  9. [17]

    Palaiodimopoulos, N. et al. Quantum inspired image augmentation applicable to waveguides and optical image transfer via anderson localization. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 116–120 (IEEE, 2024)

    2024

  10. [18]

    Disqu: investigating the impact of disorder in quantum generative models.arXiv preprint arXiv:2409.16180 (2024)

    Werner, Y .et al. Disqu: investigating the impact of disorder in quantum generative models.arXiv preprint arXiv:2409.16180 (2024)

    2024

  11. [19]

    Quantum Computing in the NISQ era and beyond

    Preskill, J. Quantum Computing in the NISQ era and beyond. Quantum 2, 79, DOI: 10.22331/q-2018-08-06-79 (2018)

    2018 doi

  12. [20]

    & Wang, C

    Chia, N.-H., Lin, H.-H. & Wang, C. Quantum-inspired sublinear classical algorithms for solving low-rank linear systems, DOI: 10.48550/ARXIV .1811.04852 (2018)

    2018 doi arXiv

  13. [21]

    M., Delgado, A., Bardhan, B

    Arrazola, J. M., Delgado, A., Bardhan, B. R. & Lloyd, S. Quantum-inspired algorithms in practice. Quantum 4, 307, DOI: 10.22331/q-2020-08-13-307 (2020)

    2020 doi

  14. [22]

    & Macaluso, A

    Srinivasan, A., Geng, A. & Macaluso, A. e. a. Benefiting from quantum? a comparative study of q-seg, quantum-inspired techniques, and u-net for crack segmentation. arXiv (2024). 2410.10713

    2024 arXiv

  15. [23]

    & Moghiseh, A

    Srinivasan, A., Geng, A., Macaluso, A., Kiefer-Emmanouilidis, M. & Moghiseh, A. A comparative study of q-seg, quantum-inspired techniques, and u-net for crack image segmentation. tm-Technisches Messen (2025)

    2025

  16. [24]

    & Andronikos, T

    Papalitsas, C., Kastampolidou, K. & Andronikos, T. Nature and quantum-inspired procedures – a short literature review. In Vlamos, P. (ed.) GeNeDis 2020, 129–133 (Springer International Publishing, Cham, 2021)

    2020

  17. [25]

    Huynh, L. et al. Quantum-inspired machine learning: a survey. arXiv preprint arXiv:2308.11269 (2023)

    2023 arXiv

  18. [26]

    & Hinton, G

    Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning , 1597–1607 (PmLR, 2020)

    2020

  19. [27]

    Chen, C.-S., Tsai, A. H.-W. & Huang, S.-C. Quantum multimodal contrastive learning framework (2025). 2408.13919

    2025 arXiv

  20. [28]

    & Wang, X

    Ren, Y ., Wang, D., An, L., Mao, S. & Wang, X. Multi-positive sample quantum contrastive learning for human activity recognition. In GLOBECOM 2024 - 2024 IEEE Global Communications Conference , 3992–3997, DOI: 10.1109/ GLOBECOM52923.2024.10901296 (2024)

    2024

  21. [29]

    & Wang, X

    Ren, Y ., Wang, D., An, L., Mao, S. & Wang, X. Quantum contrastive learning for human activity recognition.Smart Heal. 36, 100574, DOI: https://doi.org/10.1016/j.smhl.2025.100574 (2025)

    2025

  22. [30]

    Jaderberg, B. et al. Quantum self-supervised learning. Quantum Sci. Technol. 7, 035005, DOI: 10.1088/2058-9565/ac6825 (2022)

    2022 doi

  23. [31]

    Kankeu, I. et al. Quantum-inspired embeddings projection and similarity metrics for representation learning (2025). 2501.04591

    2025 arXiv

  24. [32]

    & Yan, R

    Yu, Y ., Qiu, D. & Yan, R. A quantum-like tensor compression sentence representation based on constraint functions for semantics analysis. Int. J. Comput. Intell. Syst. 17, 1, DOI: 10.1007/s44196-023-00380-w (2024). 12/18

    2024 doi

  25. [33]

    & Schuld, M

    Bowles, J., Ahmed, S. & Schuld, M. Better than classical? the subtle art of benchmarking quantum machine learning models. arXiv preprint arXiv:2403.07059 (2024)

    2024 arXiv

  26. [34]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition , 770–778 (2016)

    2016

  27. [35]

    & Capmany, J

    Macho-Ortiz, A., Pérez-López, D. & Capmany, J. Optical implementation of 2 × 2 universal unitary matrix transformations. Laser & Photonics Rev. 15, 2000473, DOI: https://doi.org/10.1002/lpor.202000473 (2021). https://onlinelibrary.wiley.com/ doi/pdf/10.1002/lpor.202000473

    2021 doi

  28. [36]

    J., Bergholm, V

    Möttönen, M., Vartiainen, J. J., Bergholm, V . & Salomaa, M. M. Transformation of quantum states using uniformly controlled rotations. Quantum Info. Comput. 5, 467–473 (2005)

    2005

  29. [37]

    Quantum computing in the nisq era and beyond

    Preskill, J. Quantum computing in the nisq era and beyond. Quantum 2, 79 (2018)

    2018

  30. [38]

    F., Blank, C., Araújo, I

    Araujo, I. F., Blank, C., Araújo, I. C. & da Silva, A. J. Low-rank quantum state preparation. IEEE Transactions on Comput. Des. Integr. Circuits Syst. 43, 161–170 (2023)

    2023

  31. [39]

    & Protasov, S

    Lisnichenko, M. & Protasov, S. Quantum image representation: a review. Quantum Mach. Intell. 5, 2, DOI: 10.1007/ s42484-022-00089-7 (2022)

    2022

  32. [40]

    O., Endo, S

    Kubo, K., Nakagawa, Y . O., Endo, S. & Nagayama, S. Variational quantum simulations of stochastic differential equations. Phys. Rev. A 103, 052425 (2021)

    2021

  33. [41]

    & Sánchez-Soto, L

    de Guise, H., Di Matteo, O. & Sánchez-Soto, L. L. Simple factorization of unitary transformations. Phys. Rev. A 97, 022328 (2018)

    2018

  34. [42]

    Moll, N. et al. Quantum optimization using variational algorithms on near-term quantum devices. Quantum Sci. Technol. 3, 030503 (2018)

    2018

  35. [43]

    & Lukin, M

    Zhou, L., Wang, S.-T., Choi, S., Pichler, H. & Lukin, M. D. Quantum approximate optimization algorithm: Performance, mechanism, and implementation on near-term devices. Phys. Rev. X 10, 021067 (2020)

    2020

  36. [44]

    Bley, J. et al. Visualizing entanglement in multiqubit systems. Phys. Rev. Res. 6, 023077, DOI: 10.1103/PhysRevResearch. 6.023077 (2024)

    2024 doi

  37. [45]

    Https://people.math.sc.edu/Burkardt/data/tif/cameraman.tif [Online; accessed June 17, 2025]

    A man taking a picture (grayscale). Https://people.math.sc.edu/Burkardt/data/tif/cameraman.tif [Online; accessed June 17, 2025]

    2025

  38. [46]

    Fast r-cnn

    Girshick, R. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision , 1440–1448 (2015)

    2015

  39. [47]

    & Adam, H

    Chen, L.-C., Zhu, Y ., Papandreou, G., Schroff, F. & Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV) , 801–818 (2018)

    2018

  40. [48]

    & Lukowicz, P

    Tschöpe, M., Schneider, D., Suh, S. & Lukowicz, P. A novel guidance framework for nasal rapid antigen tests with improved swab keypoint detection. Smart Heal. 35, 100534 (2025)

    2025

  41. [49]

    The distribution of means for samples of size n drawn from a population in which the variate takes values between 0 and 1, all such values being equally probable

    Hall, P. The distribution of means for samples of size n drawn from a population in which the variate takes values between 0 and 1, all such values being equally probable. Biometrika 240–245 (1927)

    1927

  42. [50]

    Dwork, C., Roth, A. et al. The algorithmic foundations of differential privacy. F oundations Trends Theor. Comput. Sci.9, 211–407 (2014)

    2014

  43. [51]

    & Smith, A

    Dwork, C., McSherry, F., Nissim, K. & Smith, A. Calibrating noise to sensitivity in private data analysis. In Halevi, S. & Rabin, T. (eds.) Theory of Cryptography, 265–284 (Springer Berlin Heidelberg, Berlin, Heidelberg, 2006)

    2006

  44. [52]

    Seo, J., Choi, S., Ha, E., Kim, B. & Na, D. New benchmarks for asian facial recognition tasks: Face classification with large foundation models. arXiv preprint arXiv:2310.09756 (2023)

    2023 arXiv

  45. [53]

    Differential privacy

    Dwork, C. Differential privacy. In International colloquium on automata, languages, and programming , 1–12 (Springer, 2006)

    2006

  46. [54]

    & Naor, M

    Dwork, C., Kenthapadi, K., McSherry, F., Mironov, I. & Naor, M. Our data, ourselves: Privacy via distributed noise generation. In Advances in Cryptology-EUROCRYPT 2006: 24th Annual International Conference on the Theory and Applications of Cryptographic Techniques, St. Petersb...

    2006

  47. [55]

    Renyi differential privacy

    Mironov, I. Renyi differential privacy. In 2017 IEEE 30th computer security foundations symposium (CSF) , 263–275 (IEEE, 2017)

    2017

  48. [56]

    Dong, J., Roth, A. & Su, W. J. Gaussian differential privacy. J. Royal Stat. Soc. Ser. B: Stat. Methodol. 84, 3–37 (2022). 13/18

    2022

  49. [57]

    Ji, J. et al. Privacy-preserving face recognition with learnable privacy budgets in frequency domain. In European Conference on Computer Vision, 475–491 (Springer, 2022). Acknowledgements We gratefully acknowledge financial support from the Quantum Initiative Rhineland-Palatin...

    2022

Pith tools

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