Pith. sign in

REVIEW 4 major objections 5 minor 30 references

This paper argues that automated architecture search over parameterized quantum circuits can find quantum classifiers that outperform manually designed ones and match or beat classical baselines on multisource classification and change-dete

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 21:53 UTC pith:ESNUEMN4

load-bearing objection The abstract overstates the ONERA result—0.752 is the classical MLP, not an AQML circuit—but the MNIST experiments genuinely show automated search beating a manual PQC baseline. the 4 major comments →

arxiv 2602.18642 v1 pith:ESNUEMN4 submitted 2026-02-20 quant-ph cs.CVcs.LG

Auto Quantum Machine Learning for Multisource Classification

Tomasz Rybotycki , Sebastian Dziura , Piotr Gawron This is my paper
classification quant-ph cs.CVcs.LG
keywords quantum machine learningautomated architecture searchquantum architecture searchdata fusionmultisource classificationchange detectionparameterized quantum circuitshyperparameter optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that automated quantum machine learning (AQML) — treating the ansatz architecture as a hyperparameter and searching it with classical hyperparameter optimization — can find parameterized quantum circuits (PQCs) that outperform manually designed quantum classifiers on multisource classification and change-detection problems. On a synthetic two-source digit-classification task, AQML-found circuits reached 0.939–0.944 accuracy versus 0.886 for a hand-picked circuit, approaching a classical multilayer perceptron at 0.960 while using over ten times fewer total parameters. On a real multispectral change-detection benchmark, the best AQML-found circuit had only 8 trainable parameters and reached 0.752 accuracy, exceeding a previously published QML model at 0.720 and matching a classical baseline. The overall message is that circuit architecture is a large and automatable part of QML performance.

Core claim

The central discovery is that a relatively simple automated search over the structure of a parameterized quantum circuit — specifically, over the sequence and type of QML blocks composed of data re-uploading and variational layers — yields hybrid quantum-classical classifiers that beat manually designed PQCs and are competitive with classical MLPs. In the synthetic experiment, the best AQML-discovered circuit achieves 0.944 accuracy compared to 0.886 for the manual design, and the AQML circuits need only 72–135 classifier parameters versus 18 for the manual circuit. On the real-world change-detection task, an 8-parameter circuit (amplitude encoding followed by a single entangling layer) reac

What carries the argument

The key object is the parameterized quantum circuit (PQC), built from QML blocks — each block pairs a data-encoding step (which may be an identity, fixed encoding, or re-uploading of the input) with a variational unitary containing trainable gates. The machinery is the automated search: the circuit architecture is treated as a hyperparameter, and a classical hyperparameter optimizer explores combinations of QML blocks, selecting the sequence that maximizes validation accuracy. This transforms quantum architecture search from a manual design task into an optimization problem, and the discovered circuits are the vehicles of the argument. The paper also shows that appending a final classical li

Load-bearing premise

The central claim assumes that the accuracy gains come from the automated search itself and not from the AQML circuits having more trainable parameters (72–135 vs. 18 for the manual PQC in the synthetic study) or from more extensive search effort.

What would settle it

Retrain the manually designed PQC on the synthetic two-source digit task with its classifier expanded to 72–135 parameters, matching the AQML classifiers' capacity, using the same 5-fold cross-validation and optimization protocol; if it then reaches or exceeds the AQML accuracy (0.939–0.944), the claimed architectural-search advantage is an artifact of capacity rather than search.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • AQML can replace manual ansatz design for multisource fusion tasks, producing circuits that match classical MLPs with far fewer parameters.
  • The simplest discovered circuits (e.g., 8 parameters and few gates) are easier to transpile and run on noisy quantum devices, reducing the impact of gate errors.
  • The stabilizing effect of a final classical linear layer suggests a simple robust design rule for hybrid quantum-classical classifiers: read out all qubits through a learned linear map.
  • On the change-detection benchmark, AQML exceeds a previously published QML result (0.752 vs 0.720), indicating that architecture search alone can recover performance improvements without changing preprocessing or training loss.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If architecture search is the main driver of the observed gains, the same AQML approach could transfer to other QML applications beyond data fusion, where manual ansatz selection is still common.
  • A testable extension would be to let the search also propose the data-encoding strategy (angle vs amplitude vs re-uploading) jointly with the variational layers, rather than fixing the encoding in advance.
  • The paper reports best-of-many runs; reporting the full distribution of found architectures would clarify whether the improvement is a robust property of the search or of a few lucky circuits.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes an automated quantum machine learning (AQML) approach, based on the authors' aqmlator library, for multisource classification and change detection. It reports two sets of experiments: a synthetic multisource MNIST task and the real-world ONERA change-detection dataset. On MNIST, AQML-found PQCs achieve mean accuracies of 0.939–0.944, versus 0.886 for a manually designed PQC and 0.960 for a classical MLP, with far fewer classifier parameters than the MLP. On ONERA, the paper claims that AQML improves over a previously reported QML accuracy of 0.720, noting a best PQC of 0.743 with only 8 trainable parameters, and that the average AQML PQC_Linear accuracy outperforms the average MLP accuracy. The authors also observe that adding a final linear layer stabilizes training. The quantitative support for the headline ONERA claim is internally inconsistent, particularly concerning which number is the AQML/quantum result and which is the classical baseline.

Significance. If the MNIST results are reproducible, they provide a useful empirical demonstration that architecture search can discover compact PQCs that outperform a hand-picked ansatz and approach classical MLP accuracy; the associated Zenodo code/data deposit is a strength. The ONERA result would be significant for QML applications in remote sensing if the numbers were internally consistent and statistically grounded. As it stands, the central ONERA comparison is unreliable because the reported 0.752 appears to be the classical MLP, and the best PQC value is a best-of-many draw compared with a single prior point without statistical support.

major comments (4)
  1. [§4.2, Results; Abstract] The headline improvement over [24] is not supported by the numbers as written. The text states that 'the accuracy of the best classical model outperformed the best PQC model — 0.752 vs. 0.743', and then reports 'AQML approach led to the results that were better than originally reported (with accuracy 0.752 vs. 0.720)'. Thus 0.752 is the classical MLP, not an AQML-found circuit. The best AQML PQC is 0.743, and the abstract claims improved accuracy over previously reported QML-based change detection. The comparison should be restated as the AQML PQC result (0.743) versus 0.720, or the claim should be explicitly limited to the classical pipeline. Additionally, Eq. (3) gives average PQC_Linear accuracy 0.738 while the text gives 0.730; these conflict and must be reconciled.
  2. [§4.2, Results] The ONERA improvement claim is based on a best-of-many selection without statistical support. The best PQC accuracy 0.743 is the maximum over 36 PQC_Solo and 41 PQC_Linear runs, compared with a single published 0.720 value. No standard deviations, confidence intervals, seed-matched baselines, or multiple-testing correction are reported. Without these, the apparent improvement could be a selection artifact. Please report the distribution of accuracies (e.g., mean ± std for each family), run a paired or at least a simple significance test where possible, or explicitly label the 0.743 as a best-case observation rather than a demonstrated improvement.
  3. [§4.1, Table 3] The comparison between manual and AQML PQCs is confounded by model capacity and search effort. The manual PQC classifier has 18 trainable parameters and no data re-uploading, while the AQML-found classifiers have 72 and 135 parameters and use re-uploading blocks (SimplifiedTwoDesign, BellmanLayer). The manual PQC was also selected with 100 HPO trials versus 50/35 trials for the two AQML variants. The conclusion that AQML 'outperforms' manual design is therefore not a head-to-head comparison of equal-capacity models. Please discuss this confound or weaken the causal claim; at minimum, report whether a manual PQC with comparable parameter count and re-uploading would close the gap.
  4. [§4.2, Results; Eq. (4)] The statement that the 16-parameter MLP scores (max 0.731, avg 0.667, min 0.525) are 'all below AQML-found PQCs' is contradicted by Eq. (4), which reports min PQC_Solo = 0.505 < 0.525. The statement holds only for the PQC_Linear family (min 0.710), not for the full AQML set. Please qualify this claim to avoid an unsupported comparison.
minor comments (5)
  1. [Abstract/line 1] Typo: 'It this paper' should be 'In this paper'.
  2. [§4.1, Eqs. (1)–(2) and Table 3] Eqs. (1)–(2) report accuracies averaged over models from a group, while Table 3 reports fold-averaged accuracies for the single best HPO-selected architecture. The two quantities are not immediately comparable; please clarify the definitions and report standard deviations for the group averages.
  3. [§4.1, Results] The phrase 'over 10 times less parameters' is not supported by the total parameter counts in Table 3 (75,995 vs. 10,899 is roughly 7×); the claim is accurate only for the classifier component (37,623 vs. 72/135). Please make the comparison explicit.
  4. [Notation] The names PQC_Solo_AQML and PQC_Linear_AQML are abbreviated inconsistently; define the notation once and use it consistently throughout the tables and equations.
  5. [§4.2, first paragraph] The sentence 'We ran 26, 36 and 41 experiments...' could be clarified: it lists MLP, PQC_Solo_AQML, and PQC_Linear_AQML, but the order is not obvious from the names. A table or explicit enumeration would help.

Circularity Check

0 steps flagged

No significant circularity: the AQML comparison is an empirical benchmark; internal inconsistencies and self-citations do not reduce the claims to their inputs.

full rationale

The manuscript contains no formal derivation chain whose output equals its input. Its central claims are empirical: AQML (from the authors' aqmlator, ref [23]) searches over ansatze, and the reported MNIST and ONERA numbers are measured accuracies from repeated runs. The comparison with the authors' earlier ONERA work [24] is a baseline comparison, not a definitional reduction; the reported best PQC accuracy (0.743) and the average PQC_Linear accuracy (0.730 in text, 0.738 in Eq. 3) are obtained from experiments rather than constructed from the baseline value. The paper does contain internal inconsistencies (0.752 is first attributed to the best classical model and later used in the AQML comparison; Eq. (3) states 0.738 while the text says 0.730), and the best-of-77 comparison to a single prior point estimate lacks significance testing, but these are correctness/comparability concerns, not circularity. No prediction is a fitted parameter renamed, no ansatz is imposed by a self-citation, and no uniqueness theorem is imported. Under hard rule 1, circularity requires a quotable reduction to the paper's own inputs or a fitted parameter renamed as prediction; none is present here.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No derivation is attempted; the ledger captures experimenter degrees of freedom that the central claims depend on. The main degrees of freedom are HPO-selected hyperparameters, the fixed 8-qubit/PCA setup, the limited QML block search range, and arbitrary MNIST preprocessing choices. The paper introduces no new physical entities.

free parameters (6)
  • MNIST HPO architecture hyperparameters (learning rate, batch size, feature-extractor hidden/output sizes, classifier hid = 1e-3, 83, 90, 196, 95
    Selected by Optuna on validation data (Table 1); the MLP/PQC comparisons depend on these choices.
  • MNIST PQC HPO hyperparameters (batch sizes, hidden/output sizes, QML block counts) = varies: 32/76/121, 128/163/97, 6/8/8, 3/3/5
    Selected by Optuna for manual and AQML PQCs (Table 2); AQML circuits have 4-7.5x more classifier parameters than the manual PQC.
  • ONERA PCA components per source = 4
    Chosen to reduce 13-band input to keep simulation tractable (Section 3.2); affects all models but fixes the information available to the quantum classifier.
  • ONERA number of qubits = 8
    Fixed as in [24] (Section 4.2); constrains the AQML search space.
  • ONERA QML blocks search range = [1,5]
    Search-space bound for aqmlator; a wider range might yield different conclusions.
  • MNIST preprocessing recipe (digit subset {5,6,7}, split, binarization, downsampling, vertical mean to 14 features) = N/A
    Arbitrary design choices (Section 4.1); the reported accuracy may not generalize to other splits or feature encodings.
axioms (4)
  • domain assumption Noiseless circuit simulation faithfully represents the performance of the QML models
    All PQC results come from simulators; Section 3.2/4 do not include hardware runs, yet conclusions about deployability rely on simulator accuracy.
  • ad hoc to paper The aqmlator block library (data re-uploading and variational layers, incl. SimplifiedTwoDesign and BellmanLayer) spans the relevant circuit space for these tasks
    Section 3.3 defines the search space; the claim that AQML 'finds' better models is only meaningful within this space.
  • domain assumption The ONERA setup from [24] is correctly reproduced (PCA, 8 qubits, fusion scheme, accuracy metric)
    Section 4.2 states the setup remains the same; any divergence invalidates the 0.743 vs 0.720 comparison.
  • standard math Standard supervised-learning assumptions (iid samples, cross-entropy loss, Adam optimization, validation-based HPO)
    Section 4.1 experimental setup; the paper does not test robustness to violations.

pith-pipeline@v1.3.0-alltime-deepseek · 10284 in / 17186 out tokens · 132503 ms · 2026-08-02T21:53:11.715268+00:00 · methodology

0 comments
read the original abstract

With fault-tolerant quantum computing on the horizon, there is growing interest in applying quantum computational methods to data-intensive scientific fields like remote sensing. Quantum machine learning (QML) has already demonstrated potential for such demanding tasks. One area of particular focus is quantum data fusion -- a complex data analysis problem that has attracted significant recent attention. In this work, we introduce an automated QML (AQML) approach for addressing data fusion challenges. We evaluate how AQML-generated quantum circuits perform compared to classical multilayer perceptrons (MLPs) and manually designed QML models when processing multisource inputs. Furthermore, we apply our method to change detection using the multispectral ONERA dataset, achieving improved accuracy over previously reported QML-based change detection results.

Figures

Figures reproduced from arXiv: 2602.18642 by Piotr Gawron, Sebastian Dziura, Tomasz Rybotycki.

Figure 1
Figure 1. Figure 1: A schematic representation of all the models used in this work: yellow — feature extractor, blue — classifier (here the fusion takes place), green — optional output logits for quantum classifiers. XOR on the schematic means that only one type of classifier is used at the time. Top classifier represents MLP approach, and the bottom one parametrized quantum circuit approach. In this architecture, the inputs … view at source ↗
Figure 2
Figure 2. Figure 2: A general schematic of a parameterized quantum circuit. It consists of K QML blocks, each containing a data (re-)uploading operation and a variational operation. Notice that both Uload and Uvar could, in principle, be an identity. On the schematic, x denotes the input, and θ denotes the full set of PQC parameters. The PQC is concluded with a measurement. enhanced this architecture with an additional classi… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the synthetic multisource preprocessing pipeline. (a) Original input image. (b) Horizontal splitting into top and bottom halves. (c) Binarization of each half. (d) Spatial downsampling. (e) Computation of column-wise mean pixel intensities, resulting in vector representations that constitute the synthetic multisource inputs xtop and xbottom. Experimental setup The xtop and xbottom are used dire… view at source ↗
Figure 4
Figure 4. Figure 4: A pair of images from the ONERA dataset. They show Saclay, a city in France. The images were taken on 15 III 2016 (Left) and 29 VIII 2017 (Right). even more visible than in the case of the MNIST dataset, with: accuracy avg PQCSolo AQML = 0.676 < 0.738 = accuracyavg PQCLinear AQML , (3) accuracymin PQCSolo AQML = 0.505 < 0.710 = accuracymin PQCLinear AQML . (4) This observation positions training stability … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 5 canonical work pages

  1. [1]

    Nature Computational Science1(6), 403–409 (Jun 2021)

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

    doi:10.1038/s43588-021-00084-1 2021

  2. [2]

    Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M.: Optuna: A next-generation hyperparameter optimization framework (2019), https://arxiv.org/abs/1907.10902

    Pith/arXiv arXiv 2019

  3. [3]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 41(2), 423–443 (Feb 2019)

    Baltrušaitis,T.,Ahuja,C.,Morency,L.P.:MultimodalMachineLearning:ASurvey and Taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(2), 423–443 (Feb 2019). https://doi.org/10.1109/TPAMI.2018.2798607

    arXiv 2019

  4. [4]

    Nature549(7671), 195–202 (Sep 2017)

    Biamonte, J., Wittek, P., Pancotti, N., Rebentrost, P., Wiebe, N., Lloyd, S.: Quantum machine learning. Nature549(7671), 195–202 (Sep 2017). https://doi.org/10.1038/nature23474

    doi:10.1038/nature23474 2017

  5. [5]

    In: 2018 25th IEEE International Conference on Image Process- ing (ICIP)

    Caye Daudt, R., Le Saux, B., Boulch, A.: Fully convolutional siamese networks for change detection. In: 2018 25th IEEE International Conference on Image Process- ing (ICIP). pp. 4063–4067 (2018). https://doi.org/10.1109/ICIP.2018.8451652

    arXiv 2018

  6. [6]

    https://doi.org/10.1109/TAES.1986.310699

    Chair, Z., Varshney, P.: Optimal data fusion in multiple sensor detection sys- tems.IEEETransactionsonAerospaceandElectronicSystemsAES-22(1),98–101 (1986). https://doi.org/10.1109/TAES.1986.310699

    arXiv 1986

  7. [7]

    Cheng,G.,Huang,Y.,Li,X.,Lyu,S.,Xu,Z.,Zhao,Q.,Xiang,S.:Changedetection methods for remote sensing in the last decade: A comprehensive review (2023), https://arxiv.org/abs/2305.05813

    Pith/arXiv arXiv 2023

  8. [8]

    Csurka, G., Volpi, R., Chidlovskii, B.: Semantic image segmentation: Two decades of research (2023), https://arxiv.org/abs/2302.06378

    Pith/arXiv arXiv 2023

  9. [9]

    Proceedings of the IEEE103(9), 1585–1601 (Sep 2015)

    Dalla Mura, M., Prasad, S., Pacifici, F., Gamba, P., Chanussot, J., Benediktsson, J.A.: Challenges and Opportunities of Multimodality and Data Fusion in Remote Sensing. Proceedings of the IEEE103(9), 1585–1601 (Sep 2015). https://doi.org/10.1109/JPROC.2015.2462751, https://ieeexplore.ieee.org/abstract/document/7194740

    arXiv 2015

  10. [10]

    https://doi.org/10.5281/zenodo.18717347, https://doi.org/10.5281/zenodo.18717347

    Dziura, S., Rybotycki, T., Gawron, P.: Automated quantum ma- chine learning for multisource classification: Experimental data and the code (Feb 2026). https://doi.org/10.5281/zenodo.18717347, https://doi.org/10.5281/zenodo.18717347

    work page doi:10.5281/zenodo.18717347 2026

  11. [11]

    Information Fusion45, 150–152 (Jan 2019)

    Fortino, G., Ghasemzadeh, H., Gravina, R., Liu, P.X., Poon, C.C.Y., Wang, Z.: Advances in multi-sensor fusion for body sensor networks: Algorithms, architectures, and applications. Information Fusion45, 150–152 (Jan 2019). https://doi.org/10.1016/j.inffus.2018.01.012

    work page doi:10.1016/j.inffus.2018.01.012 2019

  12. [12]

    Bulletin of the Polish Academy of Sciences Technical Sciences73(5), e154279 (2025)

    Gupta, M.K., Romaszewski, M., Gawron, P.: Potential of quantum ma- chine learning for processing multispectral earth observation data. Bulletin of the Polish Academy of Sciences Technical Sciences73(5), e154279 (2025). https://doi.org/10.24425/bpasts.2025.154279

    arXiv 2025

  13. [13]

    Journal of Big Data7(1), 28

    Hancock,J.T.,Khoshgoftaar,T.M.:Surveyoncategoricaldataforneuralnetworks. Journal of Big Data7(1), 28. https://doi.org/10.1186/s40537-020-00305-w Auto Quantum Machine Learning for Multisource Classification 15

    doi:10.1186/s40537-020-00305-w

  14. [14]

    Na- ture567(7747), 209–212 (Mar 2019)

    Havlíček, V., Córcoles, A.D., Temme, K., Harrow, A.W., Kandala, A., Chow, J.M., Gambetta, J.M.: Supervised learning with quantum-enhanced feature spaces. Na- ture567(7747), 209–212 (Mar 2019). https://doi.org/10.1038/s41586-019-0980-2

    doi:10.1038/s41586-019-0980-2 2019

  15. [15]

    Nature Communications 12(1), 2631 (May 2021)

    Huang, H.Y., Broughton, M., Mohseni, M., Babbush, R., Boixo, S., Neven, H., Mc- Clean, J.R.: Power of data in quantum machine learning. Nature Communications 12(1), 2631 (May 2021). https://doi.org/10.1038/s41467-021-22539-9

    doi:10.1038/s41467-021-22539-9 2021

  16. [16]

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization (2017), https://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2017

  17. [17]

    Koike-Akino, T., Wang, P., Wang, Y.: Autoqml: Automated quantum ma- chine learning for wi-fi integrated sensing and communications (2022), https://arxiv.org/abs/2205.09115

    Pith/arXiv arXiv 2022

  18. [18]

    LeCun, Y.: MNIST Handwritten Digits (Dec 2025), https://www.kaggle.com/datasets/hichamachahboun/mnist-handwritten-digits

    2025

  19. [19]

    IEEE Transactions on Neural Networks and Learning Systems 33(2), 747–761 (Feb 2022)

    Li,H.C.,Hu,W.S.,Li,W.,Li,J.,Du,Q.,Plaza,A.:A3CLNN:Spatial,Spectraland Multiscale Attention ConvLSTM Neural Network for Multisource Remote Sensing Data Classification. IEEE Transactions on Neural Networks and Learning Systems 33(2), 747–761 (Feb 2022). https://doi.org/10.1109/TNNLS.2020.3028945

    arXiv 2022

  20. [20]

    Chinese Journal of Aeronautics37(7), 24–58 (Jul 2024)

    Li, X., Dunkin, F., Dezert, J.: Multi-source information fusion: Progress and future. Chinese Journal of Aeronautics37(7), 24–58 (Jul 2024). https://doi.org/10.1016/j.cja.2023.12.009

    doi:10.1016/j.cja.2023.12.009 2024

  21. [21]

    In: 2024 IEEE International Conference on Quantum Com- puting and Engineering (QCE)

    Martyniuk, D., Jung, J., Paschke, A.: Quantum architecture search: A survey. In: 2024 IEEE International Conference on Quantum Com- puting and Engineering (QCE). p. 1695–1706. IEEE (Sep 2024). https://doi.org/10.1109/qce60285.2024.00198

    arXiv 2024

  22. [22]

    Prentice Hall, 3rd edn

    Russell, S., Norvig, P.: Artificial Intelligence: A Modern Approach. Prentice Hall, 3rd edn. (2010)

    2010

  23. [23]

    Computer Science26(SI) (Jul 2025)

    Rybotycki, T., Gawron, P.: AQMLator – an Quantum Ma- chine Learning e-Platform. Computer Science26(SI) (Jul 2025). https://doi.org/10.7494/csci.2025.26.si.7063

    work page doi:10.7494/csci.2025.26.si.7063 2025

  24. [24]

    Rybotycki, T., Gupta, M.K., Gawron, P.: Explainable quantum ma- chine learning for multispectral images segmentation: Case study (2025), https://arxiv.org/abs/2503.08962

    arXiv 2025

  25. [25]

    Physical Review Letters122(4), 040504 (Feb 2019)

    Schuld, M., Killoran, N.: Quantum machine learning in feature Hilbert spaces. Physical Review Letters122(4), 040504 (Feb 2019). https://doi.org/10.1103/PhysRevLett.122.040504

    doi:10.1103/physrevlett.122.040504 2019

  26. [26]

    https://doi.org/10.1103/PhysRevA.103.032430, arXiv:2008.08605 [quant-ph]

    Schuld, M., Sweke, R., Meyer, J.J.: The effect of data encoding on the ex- pressive power of variational quantum machine learning models (Mar 2021). https://doi.org/10.1103/PhysRevA.103.032430, arXiv:2008.08605 [quant-ph]

    Pith/arXiv arXiv 2021

  27. [27]

    Sebastianelli, A., Zaidenberg, D.A., Spiller, D., Saux, B.L., Ullo, S.L.: On circuit- based hybrid quantum neural networks for remote sensing imagery classification (2021), https://arxiv.org/abs/2109.09484

    Pith/arXiv arXiv 2021

  28. [28]

    IEEE Transactions on Aerospace and Electronic Systems59(2), 2002–2012 (2023)

    Stooβ, V., Ulmke, M., Govaers, F.: Quantum computing for applications in data fusion. IEEE Transactions on Aerospace and Electronic Systems59(2), 2002–2012 (2023). https://doi.org/10.1109/TAES.2022.3212026

    arXiv 2002

  29. [29]

    Information Fusion10(1), 25–50 (Jan 2009)

    Sycara, K., Glinton, R., Yu, B., Giampapa, J., Owens, S., Lewis, M., Grindle, L.C.: An integrated approach to high-level information fusion. Information Fusion10(1), 25–50 (Jan 2009). https://doi.org/10.1016/j.inffus.2007.04.001

    work page doi:10.1016/j.inffus.2007.04.001 2009

  30. [30]

    https://doi.org/10.48550/arXiv.2601.07856

    Wu, Y., Zhou, Q., Geng, J., Deng, X., Jiang, W.: Feature Entanglement- based Quantum Multimodal Fusion Neural Network (Jan 2026). https://doi.org/10.48550/arXiv.2601.07856

    work page doi:10.48550/arxiv.2601.07856 2026