Pith. sign in

REVIEW 4 major objections 5 minor 30 references

Image Classification Using CNN-QNN Hybrid Model with Optimized Correlated Features

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

Pith's one-line read Inducing intermediate feature correlation—around 0.5—before quantum encoding improves CNN-QNN image classification accuracy and stability.

desk verdict Fresh idea and a neat closed-form derivation, but the Cor≈0.5 design rule is not fully supported because the simulation uses untrained VQC parameters and the baseline handling is post-hoc. read the letter →

arxiv 2608.04379 v1 pith:RNGBRU3X submitted 2026-08-05 cs.CV cs.AI

classification cs.CVcs.AI
keywords hybridquantum-classicalneuralnetworkfeaturecorrelationvariationalquantumcircuitimageclassificationentanglementregularizationCIFAR-10radarmicro-Doppler
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 argues that the usual preprocessing step of decorrelating or orthogonalizing CNN features is the wrong choice when those features are fed into a quantum neural network. It claims that deliberately imposing an intermediate average correlation—around $\mathrm{Cor} \approx 0.5$—between the eight classical features before quantum encoding improves binary classification accuracy and reduces run-to-run variance. The argument is supported by a closed-form expression for the variational quantum circuit's output and by Monte Carlo simulations showing an accuracy peak at $\mathrm{Cor}=0.5$ across feature dimensions. Experiments on Fashion-MNIST, CIFAR-10, and radar micro-Doppler signatures confirm that a correlation-regularization loss, which drives off-diagonal entries of the feature correlation matrix toward a target, yields the best mean accuracy near $\mathrm{Cor}\approx0.5$ and the lowest variance in the range $0.4$–$0.6$. If right, the paper supplies a one-hyperparameter recipe for making shallow, qubit-limited quantum classifiers more accurate and stable without changing the quantum circuit.

What carries the argument

The load-bearing object is the correlation-regularization loss added to the CNN's latent features, $L_{\mathrm{corr}}=\frac{1}{d^2}\sum_{i,j}(C_{ij}-T_{ij})^2$, where $C$ is the batch correlation matrix of the $d$ features and $T$ is a target matrix with diagonal $1$ and off-diagonal entries set to a constant $\mathrm{Cor}$. This single scalar hyperparameter steers the feature correlation before quantum encoding. The quantum side is a $d$-qubit variational circuit whose phase-encoding feature map applies $R_z(2x_j)$ rotations, followed by an initial $R_y(\alpha_j)$ layer and a reverse entangling chain that applies controlled-NOT gates from higher-index qubits down to the first, so that a single $Z_1$ measurement captures the multi-qubit correlations. The mathematical carrier of the argument is the derivation that conjugates this circuit's observable through the gates, reducing the output to the product-of-cosines expression in Eq. (24), which the Monte Carlo analysis then evaluates at different correlation levels.

What would settle it

Recompute the accuracy-versus-correlation curve using the closed-form expression at the trained rotation angles from the main experiments, or rerun the Monte Carlo simulation with angles drawn from a trained circuit's distribution; if the peak moves away from $\mathrm{Cor}\approx0.5$, then the random-parameter assumption, not the entanglement-alignment mechanism, is producing the design rule.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the output of its proposed variational quantum classifier—built from a phase-encoding feature map and a reverse entangling chain of controlled-NOT (CNOT) gates that compresses multi-qubit correlations into the first qubit—has a closed form, $f_\theta(x)=A_1\prod_{j=1}^{d}\cos(2x_j)+\sum_{r=1}^{d-1}A_2\prod_{j=1}^{r}\cos(2x_j)$, where the coefficients $A_1,A_2$ are products of sines and cosines of the trainable angles. Using this expression on Gaussian data, the paper finds that classification accuracy peaks when the input features carry an average pairwise correlation near $0.5$. The proposed mechanism is that moderate correlation matches the entanglement structure of the circuit: too little correlation leaves the entangling gates with nothing to couple, while too much collapses the effective dimensionality and adds redundancy. In the experiments, sweeping the target correlation $\mathrm{Cor}$ over $\{0,0.2,\ldots,0.99\}$ on three binary tasks shows the best mean accuracy near $\mathrm{Cor}\approx0.5$, with the lowest standard deviation in $\mathrm{Cor}\in[0.4,0.6]$, and the correlation-regularized training prevents the training collapse observed in the unregularized baseline.

Load-bearing premise

The load-bearing premise is that the optimum correlation of about 0.5, found in simulation with random untrained quantum-circuit rotation angles, also holds when those angles are trained on real data; if trained circuits respond to input correlation differently than random ones do, the predicted optimum does not transfer.

Editorial extensions

If this is right

  • The paper establishes a practical design rule: before quantum encoding, induce intermediate feature correlation, defaulting to $\mathrm{Cor}\approx0.5$, and fine-tune within $0.4$–$0.6$; no change to the quantum circuit is required.
  • The result directly challenges the classical practice of feeding decorrelated or whitened features into quantum classifiers, showing on three datasets that such inputs can be suboptimal and even unstable.
  • With a shallow CNN backbone, the correlation-controlled quantum head outperforms a classical multilayer-perceptron head in both accuracy and variance, suggesting quantum classifiers can be competitive without deep feature extractors.
  • Because the benefit appears across natural images, fashion items, and radar micro-Doppler spectrograms, the correlation-alignment rule is not specific to one data modality.

Reading between the lines

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

  • Beyond the paper: the optimal correlation level likely depends on the entangling architecture; a circuit with a different CNOT layout or more entangling layers may peak at a different $\mathrm{Cor}$, so the value $0.5$ should be read as a starting point rather than a universal constant.
  • Beyond the paper: the closed-form output being a product of cosines suggests the correlation optimum may be interpretable as a bias-variance trade-off in the cosine-transformed feature space, which could be tested by computing the Fisher information of the transformed features as a function of $\mathrm{Cor}$.
  • Beyond the paper: the variance reduction and collapse prevention may be the more robust benefit than the small accuracy gain; a natural extension is to test whether the same regularization stabilizes multiclass quantum classifiers or runs on noisy hardware.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a correlation-regularization method for hybrid CNN-QNN classifiers. A CNN extracts latent features, a correlation loss drives the off-diagonal entries of the feature correlation matrix toward a target value Cor, and the features are encoded into an 8-qubit variational circuit with a reverse entangling chain. Section 5 derives a closed-form expression for the QNN output f_theta(x) (Eq. (24)) and, using Monte Carlo simulations with random VQC parameters, reports that classification accuracy peaks at Cor=0.5. Section 6 validates this design rule on three binary tasks, Fashion-MNIST, CIFAR-10, and radar micro-Doppler signatures, using shallow CNN and ResNet-18 backbones. The paper concludes that inducing moderate feature correlation improves accuracy and reduces variance in hybrid CNN-QNN models.

Significance. If substantiated, the proposed design rule is simple and practically relevant: it requires no circuit modification and only one scalar hyperparameter. The closed-form derivation in Sec. 5 is a useful technical contribution, and the three-dataset, two-backbone experimental protocol is a reasonable attempt at validation. However, the central empirical claim is not yet established. The simulation-to-experiment transfer is not justified, and at least one baseline is reported under a post-hoc retraining protocol that makes the comparison invalid. The significance of the paper therefore rests on a claim that currently exceeds the evidence.

major comments (4)
  1. [Sec. 5, Step 5; Sec. 6.1] The design rule Cor≈0.5 is derived from Monte Carlo simulations in which α and β are drawn once from N(0,1) and only the classification threshold is fit. In Sec. 6.1, however, the VQC parameters are trained end-to-end with Adam. Since Eq. (24) depends on α and β only through the global coefficients A1 and A2, training can shift or flatten the location of the accuracy peak in Cor. The paper provides no evidence that the Cor–accuracy landscape for random parameters is representative of the landscape for trained parameters. Without such evidence, the Sec. 6 results cannot be read as confirmation of the Sec. 5 prediction; please repeat the simulation with trained or optimized VQC parameters, or provide a theoretical argument for transferability.
  2. [Table 1, Sec. 6.2] The unregularized Fashion-MNIST baseline is not a valid comparison. It reached 86.36% only after 2 of 5 collapsed runs were discarded and the model was retrained until success. This post-hoc selection makes the baseline incomparable to the regularized columns. Moreover, the abstract's claim that intermediate correlation 'consistently improved accuracy compared to low, high, or unregulated correlations' is contradicted by the same table: all regularized entries, including Cor=0.5 (86.18%), are below the retrained baseline (86.36%), and Cor=0.0 (86.26%) is higher than Cor=0.5. Please report all seeds without retraining, or use a fixed protocol that accounts for unstable runs.
  3. [Tables 1–2, Sec. 6.2] The accuracy differences supporting the Cor≈0.5 optimum are often within one standard deviation of the 5-fold cross-validation. For example, ResNet-18 on CIFAR-10 gives Cor=0.5 at 91.33%±0.54 pp and Cor=0.6 at 92.13%±1.05 pp; Fashion-MNIST gives Cor=0.4 at 92.16%±2.19 pp and Cor=0.5 at 92.96%±0.99 pp. No significance tests, confidence intervals, or paired fold-wise comparisons are provided. The claim that moderate correlation consistently improves accuracy needs a statistical basis, particularly where the best regularized Cor is 0.6 or 1.0 rather than 0.5 (Table 2, CIFAR-10).
  4. [Sec. 4.2, Eq. (9)] The total loss is written as L = L_classification + L_correlation with no relative weight or schedule. If L_correlation is small compared with L_classification, the target Cor may not be effectively enforced during training. The paper should state whether the two terms are used with equal weight, report the actual λ if a weight is used, or provide an ablation showing that the target correlation is reached at the reported accuracy values.
minor comments (5)
  1. [Eq. (11)] Equation (11) writes Ry(θ_j)† Z_i Ry(θ_j) = Z_i cos(θ_j) + X_i sin(θ_j) and then defines it as Z_i c_j − X_i s_j; if c_j=cos(θ_j) and s_j=sin(θ_j), the two expressions are inconsistent. Please fix the notation and verify that the signs in Eqs. (17)–(24) follow consistently.
  2. [Eq. (22)] In Eq. (22), the coefficient A2 depends on the summation index r but is written without an index outside the sum. Define A2^{(r)} for clarity.
  3. [Fig. 4 and Sec. 6.1] The MLP comparison is a useful sanity check, but the MLP head is a single linear layer; it is not matched to the VQC in parameter count or optimization budget. Please state this limitation explicitly when interpreting the comparison.
  4. [Sec. 7] The concluding statement that QNN 'can surpass' classical classifiers as more qubits become available is speculative and not supported by the 8-qubit experiments. Please soften this to a conjecture or remove it.
  5. [General] The paper does not mention code or data availability. Since the experiments are simulator-based, releasing code would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Cor≈0.5 rule is an output of an explicit QNN-output simulation and is then tested on real data; no load-bearing argument reduces to its own inputs.

full rationale

The paper's derivation chain is self-contained. Section 5 derives fθ(x) in closed form (Eq. 24) from the specified feature map and VQC using standard Pauli identities, then runs a Monte Carlo scan on synthetic Gaussian data with prescribed correlation matrices; the observed optimum at Cor≈0.5 is an emergent result of that simulation, not an input assumption. The correlation regularizer (Eq. 9) is defined independently and the target Cor is swept in Sec. 6, so the experimental confirmation is a genuine hyperparameter comparison rather than a fitted parameter renamed as a prediction. No equation in the paper reduces to its own inputs by construction, and there are no load-bearing self-citations: the only author self-citation ([16]) supports radar data acquisition, not the theoretical claim. The principal weaknesses are external-validity and interpretation concerns, not circularity: Sec. 5 draws α and β from N(0,1) without training while Sec. 6 trains the VQC end-to-end, and the unregularized Fashion-MNIST baseline required excluding collapsed runs, which conflates a generic regularization benefit with the proposed quantum-specific mechanism. These issues bear on whether the 0.5 rule is fully supported, but they do not make the derivation circular.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a single swept hyperparameter Cor, a hand-chosen loss weight, and several assumptions about the transfer from random-parameter simulations to trained circuits and from Gaussian synthetic data to real CNN features.

free parameters (4)
  • Correlation target Cor = 0.5 (swept in {0,0.2,0.4,0.5,0.6,0.8,0.99})
    The central hyperparameter; the paper sweeps it and finds 0.5 optimal in simulation and most real tasks. It is chosen by search, not derived.
  • Correlation loss weight lambda = 1 (implicit)
    The total loss is L_classification + L_correlation with no scaling. This weight is hand-chosen and could affect the balance between task loss and correlation enforcement.
  • Latent feature dimension d = 8
    The CNN outputs 8 features for 8 qubits; chosen for the qubit budget, not derived.
  • Simulation decision threshold = Exhaustive search on training scores
    In the Monte Carlo simulation, the threshold is fitted by exhaustive search; this is part of the accuracy measure, not an independent prediction.
assumptions (5)
  • standard math Standard quantum gate identities (Eqs. 11-15) and the Heisenberg picture
    Used in Sec. 5 to derive f_theta(x); from Nielsen and Chuang [21].
  • domain assumption CNN feature outputs are approximately normally distributed
    Invoked in Sec. 5 Step 5 to justify synthetic Gaussian data, citing Garriga-Alonso et al. [9].
  • ad hoc to paper Randomly initialized VQC parameters in simulation represent trained VQC behavior
    The simulation in Sec. 5 uses alpha, beta ~ N(0,1) without training, but Sec. 6 trains the VQC; the transfer is assumed without evidence.
  • ad hoc to paper A uniform off-diagonal correlation target is a sufficient summary of feature correlation structure
    The regularizer forces all off-diagonal entries to a single constant Cor; the paper does not test whether non-uniform targets are better.
  • domain assumption The chosen VQC architecture (reverse entangling chain) is representative of QNNs generally
    The derivation and design rule are specific to this circuit; the conclusion generalizes to QNNs without evidence of generality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Image Classification Using CNN-QNN Hybrid Model with Optimized Correlated Features." pith.science (2026). https://pith.science/paper/RNGBRU3X

@misc{pith2026260804379,
  author       = {Pith},
  title        = {Pith review of: Image Classification Using CNN-QNN Hybrid Model with Optimized Correlated Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RNGBRU3X}},
  note         = {Machine review of arXiv:2608.04379}
}
read the original abstract

We propose a method to optimize the correlation among convolutional neural network (CNN) features that are used as inputs to quantum neural network (QNN) to enhance image classification accuracy. Unlike prior approaches that employ orthogonal decomposition as preprocessing, we intentionally introduce correlated features that are more physically compatible with QNN. This design leverages the QNN's inherent ability to exploit quantum entanglement for representing correlated states-an advantage unavailable to classical neural networks. We hypothesize that aligning feature correlations with the entanglement structure of QNN improves binary classification performance. Based on a mathematical derivation of QNN outputs, Monte Carlo simulations indicate that an average correlation between features of 0.5 yields optimal classification accuracy. To validate this finding, we evaluate a quantum-classical hybrid model on three tasks: CIFAR-10 (automobile vs. truck), Fashion-MNIST (shirt vs. coat), and radar micro-Doppler signatures (robotic dogs vs. non-robots). To regulate feature correlations, we introduce a correlation-regularization term on the outputs of the CNN, driving the off-diagonal entries of the feature correlation matrix toward a target constant. Across all datasets, inducing intermediate correlation consistently improved accuracy compared to low, high, or unregulated correlations, while also reducing classification accuracy variance. These results demonstrate that imposing moderate feature correlations-without modifying the quantum circuit-enhances classification accuracy and stability by aligning feature statistics with the QNN's entanglement structure. This study highlights the potential of QNN to surpass the performance of classical classifiers as more qubits become available.

Figures

Figures reproduced from arXiv: 2608.04379 by the authors.

Figure 1
Figure 1. Overview of the proposed model. An input image is [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Proposed framework. Feature extractor extracts features from an input image, then the features are encoded to quantum states using a quantum feature map, followed by a VQC and an estimator to predict the label of the input image. Beyond a standard hybrid CNN–QNN, we add a correlation regularizer to the CNN output. The regularizer steers the batch correlation matrix of the latent features toward a predefined target. … view at source ↗
Figure 4
Figure 4. Examples of feature correlation matrices targeting (a) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (2 more)
Figure 7
Figure 7. Figure 7: Measurement setup and examples for micro-Doppler [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 6
Figure 6. Figure 6: Examples for Fashion-MNIST and CIFAR-10. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 24 canonical work pages

  1. [1]

    Atp: Adaptive threshold pruning for efficient data encoding in quantum neural networks

    Mohamed Afane, Gabrielle Ebbrecht, Ying Wang, Juntao Chen, and Junaid Farooq. Atp: Adaptive threshold pruning for efficient data encoding in quantum neural networks. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20427–20436, 2025. 2

  2. [2]

    Quantum–classical image processing for scene classi- fication.IEEE Sensors Letters, 6(6):1–4, 2022

    Avinash Chalumuri, Raghavendra Kune, S Kannan, and BS Manoj. Quantum–classical image processing for scene classi- fication.IEEE Sensors Letters, 6(6):1–4, 2022. 1, 2

  3. [3]

    Micro- doppler effect in radar: phenomenon, model, and simulation study.IEEE Transactions on Aerospace and electronic sys- tems, 42(1):2–21, 2006

    Victor C Chen, Fayin Li, S-S Ho, and Harry Wechsler. Micro- doppler effect in radar: phenomenon, model, and simulation study.IEEE Transactions on Aerospace and electronic sys- tems, 42(1):2–21, 2006. 7

  4. [4]

    Simulating noisy quantum circuits with matrix product density operators

    Song Cheng, Chenfeng Cao, Chao Zhang, Yongxiang Liu, Shi-Yao Hou, Pengxiang Xu, and Bei Zeng. Simulating noisy quantum circuits with matrix product density operators. Physical review research, 3(2):023005, 2021. 1

  5. [5]

    Hqnn-fsp: A hybrid classical- quantum neural network for regression-based financial stock market prediction.arXiv preprint arXiv:2503.15403, 2025

    Prashant Kumar Choudhary, Nouhaila Innan, Muhammad Shafique, and Rajeev Singh. Hqnn-fsp: A hybrid classical- quantum neural network for regression-based financial stock market prediction.arXiv preprint arXiv:2503.15403, 2025. 2

  6. [6]

    Reducing overfitting in deep networks by decorrelating representations.arXiv preprint arXiv:1511.06068, 2015

    Michael Cogswell, Faruk Ahmed, Ross Girshick, Larry Zitnick, and Dhruv Batra. Reducing overfitting in deep networks by decorrelating representations.arXiv preprint arXiv:1511.06068, 2015. 2

  7. [7]

    Improv- ing stdp-based visual feature learning with whitening

    Pierre Falez, Pierre Tirilly, and Ioan Marius Bilasco. Improv- ing stdp-based visual feature learning with whitening. In2020 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2020. 2

  8. [8]

    Hybrid quantum-classical convolutional neural network model for image classification.IEEE transactions on neural networks and learning systems, 2023

    Fan Fan, Yilei Shi, Tobias Guggemos, and Xiao Xiang Zhu. Hybrid quantum-classical convolutional neural network model for image classification.IEEE transactions on neural networks and learning systems, 2023. 2

Show all 30 references
  1. [9]

    Deep convolutional networks as shallow gaussian processes.arXiv preprint arXiv:1808.05587, 2018

    Adri`a Garriga-Alonso, Carl Edward Rasmussen, and Lau- rence Aitchison. Deep convolutional networks as shallow gaussian processes.arXiv preprint arXiv:1808.05587, 2018. 6

  2. [10]

    A hybrid quantum-classical cnn architec- ture for semantic segmentation of radar sounder data

    Raktim Ghosh, Amer Delilbasic, Gabriele Cavallaro, and Francesca Bovolo. A hybrid quantum-classical cnn architec- ture for semantic segmentation of radar sounder data. In2024 IEEE Mediterranean and Middle-East Geoscience and Re- mote Sensing Symposium (M2GARSS), pages 366–370. IEEE,

  3. [11]

    Quantum convolutional neural network based on varia- tional quantum circuits.Optics Communications, 550:129993,

    Li-Hua Gong, Jun-Jie Pei, Tian-Feng Zhang, and Nan-Run Zhou. Quantum convolutional neural network based on varia- tional quantum circuits.Optics Communications, 550:129993,

  4. [12]

    H-qnn: A hybrid quantum–classical neural network for im- proved binary image classification.AI, 5(3):1462–1481, 2024

    Muhammad Asfand Hafeez, Arslan Munir, and Hayat Ullah. H-qnn: A hybrid quantum–classical neural network for im- proved binary image classification.AI, 5(3):1462–1481, 2024. 1, 2

  5. [13]

    Supervised learning with quantum-enhanced fea- ture spaces.Nature, 567(7747):209–212, 2019

    V ojtˇech Havl ´ıˇcek, Antonio D C ´orcoles, Kristan Temme, Aram W Harrow, Abhinav Kandala, Jerry M Chow, and Jay M Gambetta. Supervised learning with quantum-enhanced fea- ture spaces.Nature, 567(7747):209–212, 2019. 4

  6. [14]

    Quantum convo- lutional neural network for classical data classification.Quan- tum Machine Intelligence, 4(1):3, 2022

    Tak Hur, Leeseok Kim, and Daniel K Park. Quantum convo- lutional neural network for classical data classification.Quan- tum Machine Intelligence, 4(1):3, 2022. 2

  7. [15]

    Quan- tum machine learning beyond kernel methods.Nature Com- munications, 14(1):517, 2023

    Sofiene Jerbi, Lukas J Fiderer, Hendrik Poulsen Nautrup, Jonas M K¨ubler, Hans J Briegel, and Vedran Dunjko. Quan- tum machine learning beyond kernel methods.Nature Com- munications, 14(1):517, 2023. 2

  8. [16]

    Human detection and activity classification based on micro-doppler signatures using deep convolutional neural networks.IEEE geoscience and remote sensing letters, 13(1):8–12, 2015

    Youngwook Kim and Taesup Moon. Human detection and activity classification based on micro-doppler signatures using deep convolutional neural networks.IEEE geoscience and remote sensing letters, 13(1):8–12, 2015. 7

  9. [17]

    Human detection by neural networks using a low-cost short-range doppler radar sensor

    Jihoon Kwon and Nojun Kwak. Human detection by neural networks using a low-cost short-range doppler radar sensor. In2017 IEEE Radar Conference (RadarConf), pages 0755–

  10. [18]

    A flexible representation of quantum images for polynomial preparation, image compression, and processing operations.Quantum Information Processing, 10(1):63–84, 2011

    Phuc Q Le, Fangyan Dong, and Kaoru Hirota. A flexible representation of quantum images for polynomial preparation, image compression, and processing operations.Quantum Information Processing, 10(1):63–84, 2011. 2

  11. [19]

    Radar hrrp target recognition based on hybrid quantum neural networks.IEEE Transactions on Aerospace and Electronic Systems, 2025

    Xin Liu, Daiying Zhou, and Qiuyan Huang. Radar hrrp target recognition based on hybrid quantum neural networks.IEEE Transactions on Aerospace and Electronic Systems, 2025. 2

  12. [20]

    Barren plateaus in quan- tum neural network training landscapes.Nature communica- tions, 9(1):4812, 2018

    Jarrod R McClean, Sergio Boixo, Vadim N Smelyanskiy, Ryan Babbush, and Hartmut Neven. Barren plateaus in quan- tum neural network training landscapes.Nature communica- tions, 9(1):4812, 2018. 1

  13. [21]

    Cambridge university press,

    Michael A Nielsen and Isaac L Chuang.Quantum computa- tion and quantum information. Cambridge university press,

  14. [22]

    Switchable whitening for deep representation learning

    Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaoou Tang, and Ping Luo. Switchable whitening for deep representation learning. InProceedings of the IEEE/CVF international conference on computer vision, pages 1863–1871, 2019. 2

  15. [23]

    Quantum computing in the nisq era and beyond

    John Preskill. Quantum computing in the nisq era and beyond. Quantum, 2:79, 2018. 1

  16. [24]

    Hybrid quantum-classical graph neural networks for tumor classification in digital pathology

    Anupama Ray, Dhiraj Madan, Srushti Patil, Pushpak Pati, Marianna Rapsomaniki, Aviwe Kohlakala, Thembelihle Rose Dlamini, Stephanie Julia Muller, Kahn Rhrissorrakrai, Filippo Utro, et al. Hybrid quantum-classical graph neural networks for tumor classification in digital patholo...

  17. [25]

    Regularizing cnns with locally constrained decorrelations.arXiv preprint arXiv:1611.01967,

    Pau Rodr´ıguez, Jordi Gonzalez, Guillem Cucurull, Josep M Gonfaus, and Xavier Roca. Regularizing cnns with locally constrained decorrelations.arXiv preprint arXiv:1611.01967,

  18. [26]

    Evaluating analytic gradients on quan- tum hardware.Physical Review A, 99(3):032331, 2019

    Maria Schuld, Ville Bergholm, Christian Gogolin, Josh Izaac, and Nathan Killoran. Evaluating analytic gradients on quan- tum hardware.Physical Review A, 99(3):032331, 2019. 3

  19. [27]

    Ex- pressibility and entangling capability of parameterized quan- tum circuits for hybrid quantum-classical algorithms.Ad- vanced Quantum Technologies, 2(12):1900070, 2019

    Sukin Sim, Peter D Johnson, and Al ´an Aspuru-Guzik. Ex- pressibility and entangling capability of parameterized quan- tum circuits for hybrid quantum-classical algorithms.Ad- vanced Quantum Technologies, 2(12):1900070, 2019. 1

  20. [28]

    Transition role of entangled data in quantum machine learning.Nature Communications, 15(1): 3716, 2024

    Xinbiao Wang, Yuxuan Du, Zhuozhuo Tu, Yong Luo, Xiao Yuan, and Dacheng Tao. Transition role of entangled data in quantum machine learning.Nature Communications, 15(1): 3716, 2024. 2

  21. [29]

    Barlow twins: Self-supervised learning via redundancy reduction

    Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and St´ephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. InInternational conference on machine learning, pages 12310–12320. PMLR, 2021. 2

  22. [30]

    The extraction of micro-doppler sig- nal with emd algorithm for radar-based small uavs’ detection

    Yichao Zhao and Yi Su. The extraction of micro-doppler sig- nal with emd algorithm for radar-based small uavs’ detection. IEEE Transactions on Instrumentation and Measurement, 69 (3):929–940, 2019. 7

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.