Pith. sign in

REVIEW 3 major objections 5 minor 50 references

PQFA applies shallow parallel quantum circuits after classical multimodal fusion and reports consistent gains over both an unaugmented backbone and a width-matched classical augmentation branch while using about one-tenth the augmentation p

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 05:06 UTC pith:N4GBLKOB

load-bearing objection PQFA is a carefully controlled study of post-fusion quantum augmentation, but the central claim of quantum-specific benefit is under-supported because no same-budget classical augmentation is tested. the 3 major comments →

arxiv 2607.13466 v1 pith:N4GBLKOB submitted 2026-07-15 cs.LG quant-ph

PQFA: Parallel Quantum Feature Augmentation of Fused Representations for Multimodal Classification

Mingzhu Wang , Yun Shang This is my paper
classification cs.LG quant-ph
keywords multimodal classificationquantum machine learningparameterized quantum circuitsfeature augmentationpost-fusion enhancementmissing-modality stabilitymultimodal fusionamplitude encoding
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 aims to establish that a small set of parallel variational quantum circuits, placed after the classical fusion stage of a multimodal classifier, can act as a parameter-efficient post-fusion augmentation module. On two image-text benchmarks, the proposed PQFA framework outperforms both the fusion backbone without augmentation and a width-matched MLP augmentation branch, while using roughly 2.2K augmentation parameters versus 24.0K for the MLP branch. Through controlled ablations, feature-space analyses, and missing-modality experiments, the paper argues that the gain does not come from feature-width expansion, random feature maps, untrained quantum circuits, or a larger parameter budget. The benefit is especially pronounced when the stronger textual modality is heavily degraded. If correct, this gives quantum circuits a practical near-term role as lightweight auxiliary feature extractors inside otherwise classical deep multimodal systems.

Core claim

The central discovery is that a post-fusion quantum readout branch can improve multimodal classification beyond what a width-matched classical augmentation branch achieves, at a fraction of the parameter cost. The fused classical vector—produced by frozen text and image encoders, bidirectional cross-attention, attentive pooling, and adaptive gated fusion—is amplitude-encoded into K parallel shallow variational circuits; their single-qubit measurement readouts are concatenated with the classical fused feature and sent to the classifier. On the movie-genre benchmark, PQFA raises Micro-F1 and Macro-F1 relative to the no-quantum backbone and the MLP augmentation baseline, and on the news benchma

What carries the argument

The load-bearing mechanism is the parallel quantum readout branch: K shallow brickwall parameterized quantum circuits that amplitude-encode the normalized fused representation and return single-qubit Pauli-Z expectation values, which are concatenated with the classical fused vector for the final prediction. Different branches can draw from a predefined 'brick pool' of two-qubit blocks, so the same fused input is transformed into structurally diverse nonlinear readouts. The decisive control is a width-matched MLP augmentation branch with the same output dimension but roughly 24.0K parameters, which isolates the effect of the quantum transformation from simple increases in feature dimensionali

Load-bearing premise

The argument depends on the width-matched 24K-parameter MLP augmentation being the correct classical control; since no classical augmentation with the same tiny ~2.2K parameter budget as the quantum branch was tested, the gain could be due to the quantum branch's smaller capacity rather than to the quantum transformation itself.

What would settle it

Train a classical augmentation branch with about 2.2K trainable parameters, matched to PQFA's branch, under the same frozen encoders, data splits, projection dimension, and output width. If that small classical branch matches or exceeds PQFA on the two benchmarks, the quantum-specific explanation is falsified. Running the trained circuits on physical hardware with finite-shot measurements would also test whether the simulated gains survive real device noise.

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

If this is right

  • If correct, PQFA offers a template for adding quantum modules to existing multimodal pipelines without changing the classical encoders or the fusion path.
  • The parameter efficiency (2.2K versus 24.0K) suggests quantum readouts can enrich fused representations in settings where trainable parameters are scarce.
  • Missing-modality gains, especially under severe text degradation, imply the augmentation can help when parts of the input are unavailable at inference time.
  • The failure of random or untrained quantum transformations and wider MLP baselines to reproduce the gain implies the benefit depends on task-driven optimization of the quantum circuits.
  • The controlled comparison methodology provides a template for attributing hybrid quantum-classical gains to the quantum component rather than to extra capacity.

Where Pith is reading between the lines

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

  • An untested control is a classical augmentation branch with the same ~2.2K parameter budget as PQFA; until that is run, the apparent quantum-specific advantage could stem from the quantum branch's smaller capacity or implicit regularization rather than from the quantum transformation itself.
  • All experiments use simulated quantum circuits; real-device tests with finite-shot noise, compilation overhead, and amplitude-encoding costs would be the decisive next check on the practical claims.
  • The framework points to a broader design principle: a cheap nonlinear readout after a strong fusion module can help even when fusion already performs well; comparing the quantum branch against other low-parameter classical nonlinearities would clarify what is unique to the quantum structure.
  • If the missing-modality result transfers, post-fusion augmentation could be viewed primarily as a resilience mechanism rather than solely as an accuracy booster.

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

3 major / 5 minor

Summary. The paper proposes PQFA, a hybrid quantum-classical post-fusion augmentation module for multimodal classification. Frozen RoBERTa and ViT encoders produce token features that are aligned by projections, refined by bidirectional cross-attention, pooled, and fused by an adaptive gate into a 128-dimensional vector. This vector is amplitude-encoded into n=7 qubits and processed by K parallel shallow brickwall parameterized quantum circuits; Pauli-Z readouts are refined and concatenated with the classical fused feature for final classification. Controlled comparisons on MM-IMDb and N24News are made against a no-quantum gated-fusion backbone (NoQ), a width-matched MLP augmentation baseline (MLP-Aug), and additional ablations (RFF-Aug, MLP-Aug-2x, PQFA-NoEnt, PQFA-FrozenQ). The paper reports consistent gains in Micro/Macro-F1 and accuracy, a much smaller augmentation-branch parameter count (2.2K vs 24.0K), improved missing-modality robustness, paired decision-transition statistics, PCA-based feature-space diagnostics, noisy-inference stability, entanglement analysis, and gate-weight analysis. The central claim is that a small trainable quantum readout branch provides a structured post-fusion transformation whose benefit is not attributable to feature expansion, random quantum maps, or additional classical width.

Significance. If the attribution holds, this is a useful contribution to hybrid quantum-classical multimodal learning. The paper's strengths are its controlled experimental protocol (same frozen encoders, fusion backbone, projection dimension, and output width), the inclusion of genuine negative controls (untrained quantum circuits, random features, wider classical branches), and the use of paired bootstrap/McNemar diagnostics on MM-IMDb. The feature-space and quantum-state analyses go beyond aggregate accuracy and are valuable. However, the significance is contingent on closing a control gap around the quantum-specific attribution and on reporting uncertainty for the primary aggregate metrics. The central claims are empirically plausible but not yet fully supported.

major comments (3)
  1. [§4.4, Table 4] The central attribution is under-determined by the absence of a same-parameter-budget classical nonlinear augmentation. MLP-Aug is matched only in output width (Km=56); it has 24.0K params vs PQFA's 2.2K. Table 9 ablations do not fill the gap: RFF-Aug and PQFA-FrozenQ are untrained, MLP-Aug-2x is wider (31.2K), and PQFA-NoEnt is still quantum. Thus PQFA > MLP-Aug is consistent with any compact trainable nonlinear map providing a regularization/inductive-bias benefit. This is load-bearing for §5's conclusion 'benefit ... from the learned transformation provided by the quantum readout branches' and the abstract's parameter-efficiency claim. Add a trainable classical branch with ~2.2K params (e.g., bottleneck MLP) and report its output width and paired result, or soften the quantum-attribution wording.
  2. [§4.1.2 and Tables 3, 5, 8] Primary metrics are reported only as five-seed point averages. The headline improvements are small (N24News: 84.70→85.35; MM-IMDb: 67.57→68.28 Micro-F1, 60.98→61.85 Macro-F1), so without seed-level variance, CIs, or significance tests, 'consistently outperforms' is not fully supported. The paired McNemar/bootstrap analysis in §4.6 mitigates this for MM-IMDb's Jaccard gain, but there is no equivalent for N24News, missing-modality, or K-sensitivity results. Report per-seed values or confidence intervals for all main controlled comparisons.
  3. [§4.1.1 and Table 1] MM-IMDb is evaluated on a filtered 'controlled test split' of 3,894 samples rather than the standard benchmark split. Table 1 therefore compares PQFA with prior published results obtained on a different test set, weakening the reference-comparison claim; the same filtering can affect the internal ablation if it changes the difficulty of the split. Please either report standard-split numbers, justify the filtering statistically, or explicitly present Table 1 only as a non-comparable positioning.
minor comments (5)
  1. [§3.6 / Table 4] The 'lightweight readout refinement layer' is counted in the 2.2K branch parameters but never defined. Specify its architecture, nonlinearity, and parameter count.
  2. [§4.8 / Table 9] PQFA-NoEnt is used as an ablation but is never defined in the text. State how entanglement is removed and confirm it is trained under the same protocol as PQFA.
  3. [§4.7] K=8 is selected for MM-IMDb because it gives the best Micro-F1, but K=1 gives a higher Macro-F1 (62.08 vs 61.85). Justify the selection criterion or discuss sensitivity to it.
  4. [§4.9.1] The text says 'We also report the probability drift and the label flip rate', but probability drift is never defined and does not appear in Table 10. Remove or define it.
  5. [Algorithm 1 / Table 5] Algorithm 1 line 4 has an apparent typo: 'Initialize the best validation scores ∗' should probably be 'score s∗'. Table 5 also has a minor spacing typo in the clean row ('61.22' adjacent to the next value).

Circularity Check

0 steps flagged

No significant circularity: PQFA is an empirical architecture/ablation study whose central comparison is not definitionally forced.

full rationale

The paper's central claim is empirical: a small trainable quantum readout branch improves classification over specified no-quantum and width-matched classical baselines. There is no derivation that takes its own conclusion as an input. The controlled ablation fixes encoders, fusion backbone, data splits, projection dimension, and augmentation output width, while varying only the augmentation branch; NoQ, MLP-Aug, MLP-Aug-2x, RFF-Aug, PQFA-FrozenQ, and PQFA-NoEnt are genuine alternative configurations rather than rescalings or restatements of PQFA. The parameter-efficiency contrast (2.2K vs 24.0K, Table 4) is a reported property of the fixed configurations, not a fitted quantity later renamed as a prediction. Branch number K is selected on validation and the sensitivity table reports all tested values, so reporting K=8 is model selection, not circular prediction. The paper itself is appropriately cautious in Section 4.3 ('should be interpreted as a reference comparison rather than as direct evidence') and reserves attribution for the controlled ablation. No load-bearing self-citation appears; no uniqueness theorem is imported; no ansatz is smuggled via citation. The manuscript's own limitation in Section 5 — that results are simulations and that 'comparisons with classical dropout and noise-injection baselines would also help distinguish hardware-induced robustness from robustness arising from the structure of the augmentation module itself' — is a missing-baseline/external-validity caveat, not circularity. Likewise, the absence of a ~2.2K-parameter classical nonlinear augmentation baseline (Table 4 shows only MLP-Aug at 24.0K and MLP-Aug-2x at 31.2K) is a soundness/under-determination concern, not a reduction of the result to its inputs. No equation in the paper equals another by construction, and no fitted parameter is fed back into the definition of the reported outcome. Therefore no circular steps are identified.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. Its load-bearing assumptions are engineering and simulation choices: exact amplitude encoding, statevector-based expectations, null-input missing-modality masks, and a nonstandard filtered test split. The most consequential free parameters are the validation-selected branch count and the unreported quantum-circuit depth/brick-pool configuration.

free parameters (6)
  • K (parallel quantum branches) = 8 (MM-IMDb); unreported for N24News
    Selected by validation performance; Table 8 shows non-monotonic sensitivity, so the choice affects the headline numbers.
  • L_Q (brickwall circuit depth) = not reported ('shallow')
    Controls the expressiveness of the quantum feature map; depth is never specified, which is critical for replication and for interpreting the 2.2K parameter count.
  • Brick pool composition and branch-wise refinement layers = not reported
    The structural diversity of the parallel branches and the post-measurement refinement are not fully specified, so the exact augmentation branch cannot be reconstructed.
  • MM-IMDb controlled test split = 3,894 test samples
    A filtered subset of the standard benchmark; filtering criteria are not given, so internal and reference comparisons may not align with standard MM-IMDb evaluation.
  • Learning rates = 1e-4 classical, 1e-5 quantum
    Separate learning rates are a hand-chosen hyperparameter that affects the optimization and could influence the comparison.
  • Validation-selected thresholds tau_c = selected on validation
    Needed for multi-label F1; a standard procedure, but the thresholding method is part of the pipeline.
axioms (5)
  • domain assumption Amplitude encoding exactly maps the normalized 128-dimensional fused vector into a 7-qubit state.
    Equations (21)-(22); the simulation assumes no state-preparation error or cost, which is optimistic for real hardware.
  • domain assumption The PennyLane statevector simulator with exact expectation values is a faithful proxy for the quantum readout branch.
    Section 3.9; no finite-shot noise or device errors are present during training, so the trained features may differ from hardware-realizable readouts.
  • domain assumption Replacing an image with a zero tensor and text with an empty string is a valid missing-modality perturbation.
    Section 4.5; frozen encoders may produce strong biased outputs for null inputs, which could confound the robustness comparison.
  • ad hoc to paper The filtered MM-IMDb test split is representative of the standard benchmark.
    Section 4.1.1; the 3,894-sample test subset is nonstandard and its filtering criteria are not disclosed, so its representativeness is unknown.
  • domain assumption McNemar's test on sample-label pairs treats the ~90K binary decisions as independent.
    Section 4.6; decisions are correlated within samples and labels, so the reported p-values may be anti-conservative.

pith-pipeline@v1.3.0-alltime-deepseek · 23165 in / 15026 out tokens · 148950 ms · 2026-08-02T05:06:10.392286+00:00 · methodology

0 comments
read the original abstract

Most multimodal learning methods improve how heterogeneous representations are aligned and fused, while post-fusion enhancement remains less explored. We propose Parallel Quantum Feature Augmentation (PQFA), a hybrid quantum-classical framework that applies multiple shallow variational quantum circuits to fused multimodal features. Text and image representations extracted by frozen RoBERTa and ViT encoders are processed through bidirectional cross-attention, attentive pooling, and adaptive gated fusion. The fused feature is then amplitude-encoded into parallel quantum circuits, whose measurement readouts are concatenated with the classical representation for prediction. We evaluate PQFA on MM-IMDb and N24News through controlled comparisons using the same encoders, fusion backbone, data splits, projection dimension, and augmentation output width. PQFA consistently outperforms both the fusion backbone without quantum augmentation and a width-matched MLP augmentation baseline, while using approximately 2.2K augmentation parameters compared with 24.0K for the MLP branch. Missing-modality experiments further show improved robustness when textual or visual inputs are incomplete, with particularly clear gains when the more informative textual modality is severely degraded. Controlled ablations and feature-space analyses indicate that the improvement cannot be reproduced by random feature mappings, increased classical width, or untrained quantum transformations. Quantum-state diagnostics additionally show stable predictive performance across the tested simulated noise levels and distinct branch-specific transformations of the encoded states. These results establish PQFA as an effective and parameter-efficient strategy for post-fusion augmentation in hybrid quantum-classical multimodal learning.

Figures

Figures reproduced from arXiv: 2607.13466 by Mingzhu Wang, Yun Shang.

Figure 1
Figure 1. Figure 1: Overall architecture of the proposed Parallel Quantum Feature Augmentation (PQFA) framework. Text and image inputs are first encoded [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic illustration of the Parallel Quantum Feature Augmentation module. (a) A shallow brickwall parameterized quantum circuit [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Fine-grained comparison between PQFA and MLP-Aug on MM-IMDb. (a) Per-label F1 di [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: PCA cumulative explained variance of augmentation features on MM-IMDb. Faster saturation indicates stronger variance concentration [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Post-training noisy inference analysis on MM-IMDb. Panel (a) reports Micro-F1 under di [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Entanglement entropy diagnostics of the trained PQFA quantum branches. Panel (a) shows layer-wise main-cut entropy averaged across [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Gate weight distributions on MM-IMDb. (a) Element-level distribution over all samples and dimensions. (b) Dimension-level mean gate [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] 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

50 extracted references · 2 linked inside Pith

  1. [1]

    Baltrušaitis, C

    T. Baltrušaitis, C. Ahuja, L.-P. Morency, Multimodal machine learning: A survey and taxonomy, IEEE transac- tions on pattern analysis and machine intelligence 41 (2) (2018) 423–443

    2018

  2. [2]

    P. Xu, X. Zhu, D. A. Clifton, Multimodal learning with transformers: A survey, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (10) (2023) 12113–12132

    2023

  3. [3]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, S. Lloyd, Quantum machine learning, Nature 549 (7671) (2017) 195–202

    2017

  4. [4]

    Schuld, N

    M. Schuld, N. Killoran, Quantum machine learning in feature hilbert spaces, Physical review letters 122 (4) (2019) 040504

    2019

  5. [5]

    Havlíˇcek, A

    V . Havlíˇcek, A. D. Córcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, J. M. Gambetta, Supervised learning with quantum-enhanced feature spaces, Nature 567 (7747) (2019) 209–212

    2019

  6. [6]

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

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

    2018

  7. [7]

    Bharti, A

    K. Bharti, A. Cervera-Lierta, T. H. Kyaw, T. Haug, S. Alperin-Lea, A. Anand, M. Degroote, H. Heimonen, J. S. Kottmann, T. Menke, et al., Noisy intermediate-scale quantum algorithms, Reviews of Modern Physics 94 (1) (2022) 015004

    2022

  8. [8]

    J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, H. Neven, Barren plateaus in quantum neural network training landscapes, Nature communications 9 (1) (2018) 4812

    2018

  9. [9]

    Z. Qu, Y . Meng, G. Muhammad, P. Tiwari, QMFND: A quantum multimodal fusion-based fake news detection model for social media, Information Fusion 104 (2024) 102172

    2024

  10. [10]

    Z. Qu, Y . Li, P. Tiwari, QNMF: A quantum neural network based multimodal fusion system for intelligent diagnosis, Information Fusion 100 (2023) 101913

    2023

  11. [11]

    Zheng, Q

    J. Zheng, Q. Gao, D. Dong, J. Lü, Y . Deng, A quantum multimodal neural network model for sentiment analysis on quantum circuits, IEEE Transactions on Artificial Intelligence 6 (5) (2024) 1128–1142

    2024

  12. [12]

    Y . Li, Y . Qu, R.-G. Zhou, J. Zhang, QMLSC: A quantum multimodal learning model for sentiment classification, Information Fusion 120 (2025) 103049

    2025

  13. [13]

    J. Gao, P. Li, Z. Chen, J. Zhang, A survey on deep learning for multimodal data fusion, Neural computation 32 (5) (2020) 829–864

    2020

  14. [14]

    T. Jiao, C. Guo, X. Feng, Y . Chen, J. Song, A comprehensive survey on deep learning multi-modal fusion: Methods, technologies and applications., Computers, Materials & Continua 80 (1) (2024). 26

    2024

  15. [15]

    Z. Liu, Y . Shen, V . B. Lakshminarasimhan, P. P. Liang, A. B. Zadeh, L.-P. Morency, Efficient low-rank multi- modal fusion with modality-specific factors, in: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), 2018, pp. 2247–2256

    2018

  16. [16]

    Z. Yu, J. Yu, J. Fan, D. Tao, Multi-modal factorized bilinear pooling with co-attention learning for visual question answering, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 1821–1830

    2017

  17. [17]

    Ben-Younes, R

    H. Ben-Younes, R. Cadene, M. Cord, N. Thome, Mutan: Multimodal tucker fusion for visual question answering, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 2612–2620

    2017

  18. [18]

    Z. Sun, P. Sarma, W. Sethares, Y . Liang, Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis, in: Proceedings of the AAAI conference on artificial intelligence, V ol. 34, 2020, pp. 8992–8999

    2020

  19. [19]

    X. Guo, A. Kong, H. Zhou, X. Wang, M. Wang, Unimodal and crossmodal refinement network for multimodal sequence fusion, in: Proceedings of the 2021 conference on empirical methods in natural language processing, 2021, pp. 9143–9153

    2021

  20. [20]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, Advances in neural information processing systems 30 (2017)

    2017

  21. [21]

    Nagrani, S

    A. Nagrani, S. Yang, A. Arnab, A. Jansen, C. Schmid, C. Sun, Attention bottlenecks for multimodal fusion, Advances in neural information processing systems 34 (2021) 14200–14213

    2021

  22. [22]

    Z. Xue, R. Marculescu, Dynamic multimodal fusion, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 2575–2584

    2023

  23. [23]

    W. Li, H. Zhou, J. Yu, Z. Song, W. Yang, Coupled mamba: Enhanced multimodal fusion with coupled state space model, Advances in Neural Information Processing Systems 37 (2024) 59808–59832

    2024

  24. [24]

    M. Ma, J. Ren, L. Zhao, D. Testuggine, X. Peng, Are multimodal transformers robust to missing modality?, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 18177–18186

    2022

  25. [25]

    Peruzzo, J

    A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, J. L. O’brien, A variational eigenvalue solver on a photonic quantum processor, Nature communications 5 (1) (2014) 4213

    2014

  26. [26]

    Zaman, T

    K. Zaman, T. Ahmed, M. A. Hanif, A. Marchisio, M. Shafique, A comparative analysis of hybrid-quantum clas- sical neural networks, in: World Congress in Computer Science, Computer Engineering & Applied Computing, Springer, 2024, pp. 102–115

    2024

  27. [27]

    Basilewitsch, J

    D. Basilewitsch, J. F. Bravo, C. Tutschku, F. Struckmeier, Quantum neural networks in practice: a comparative study with classical models from standard data sets to industrial images, Quantum Machine Intelligence 7 (2) (2025) 110

    2025

  28. [28]

    S. Wang, E. Fontana, M. Cerezo, K. Sharma, A. Sone, L. Cincio, P. J. Coles, Noise-induced barren plateaus in variational quantum algorithms, Nature communications 12 (1) (2021) 6961

    2021

  29. [29]

    Larocca, S

    M. Larocca, S. Thanasilp, S. Wang, K. Sharma, J. Biamonte, P. J. Coles, L. Cincio, J. R. McClean, Z. Holmes, M. Cerezo, Barren plateaus in variational quantum computing, Nature Reviews Physics 7 (4) (2025) 174–189

    2025

  30. [30]

    R.-X. Zhao, H. Siljak, S. Wang, Y . He, Y . Wang, HQCC: A hybrid quantum-classical classifier with adaptive structure, IEEE Signal Processing Letters (2025)

    2025

  31. [31]

    S. M. A. Rizvi, U. I. Paracha, U. Khalid, K. Lee, H. Shin, Quantum machine learning: Towards hybrid quantum- classical vision models, Mathematics 13 (16) (2025) 2645

    2025

  32. [32]

    Schnabel, M

    J. Schnabel, M. Roth, Quantum kernel methods under scrutiny: a benchmarking study, Quantum Machine Intel- ligence 7 (1) (2025) 58. 27

    2025

  33. [33]

    Nguyen, T

    T. Nguyen, T. N. Hoang, P. L. Nguyen, H. L. Vu, T. C. Thang, Expressive and scalable quantum fusion for multimodal learning, arXiv preprint arXiv:2510.06938 (2025)

    arXiv 2025

  34. [34]

    K. Yan, P. Lai, Y . Yang, Y . Ren, T. Badarch, Y . Chen, X. Zheng, Quantum-inspired multimodal fusion with lindblad master equation for sentiment analysis, Neurocomputing 648 (2025) 130710

    2025

  35. [35]

    Arevalo, T

    J. Arevalo, T. Solorio, M. Montes-y Gomez, F. A. González, Gated multimodal networks, Neural Computing and Applications 32 (14) (2020) 10209–10228

    2020

  36. [36]

    Z. Wang, X. Shan, X. Zhang, J. Yang, N24News: A new dataset for multimodal news classification, in: Proceed- ings of the thirteenth language resources and evaluation conference, 2022, pp. 6768–6775

    2022

  37. [37]

    Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, V . Stoyanov, RoBERTa: A robustly optimized BERT pretraining approach, arXiv preprint arXiv:1907.11692 (2019)

    Pith/arXiv arXiv 1907

  38. [38]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An image is worth 16x16 words: Transformers for image recog- nition at scale, in: International Conference on Learning Representations (ICLR), 2021

    2021

  39. [39]

    D. M. W. Powers, Evaluation: from precision, recall and f-measure to roc, informedness, markedness and corre- lation, Journal of Machine Learning Technologies 2 (1) (2011) 37–63

    2011

  40. [40]

    Y .-H. H. Tsai, P. P. Liang, A. Zadeh, L.-P. Morency, R. Salakhutdinov, Learning factorized multimodal repre- sentations, in: International Conference on Learning Representations (ICLR), 2019

    2019

  41. [41]

    S. M. Jayakumar, W. M. Czarnecki, J. Menick, J. Schwarz, J. Rae, S. Osindero, Y . W. Teh, T. Harley, R. Pascanu, Multiplicative interactions and where to find them, in: International conference on learning representations, 2020

    2020

  42. [42]

    I. Gat, I. Schwartz, A. Schwing, T. Hazan, Removing bias in multi-modal classifiers: Regularization by maxi- mizing functional entropies, Advances in Neural Information Processing Systems 33 (2020) 3197–3208

    2020

  43. [43]

    Sankaran, D

    S. Sankaran, D. Yang, S.-N. Lim, Multimodal fusion refiner networks, arXiv preprint arXiv:2104.03435 (2021)

    Pith/arXiv arXiv 2021

  44. [44]

    D. Li, Z. Weng, Z. Xuan, Z. Wang, ModiFedCat: A multi-modal distillation based federated catalytic framework, Information Fusion 124 (2025) 103378

    2025

  45. [45]

    L. Wen, Q. Dai, J. Liu, J. Zheng, Y . Dai, D. Wang, Z. Kang, J. Wang, Z. Xu, J. Duan, InfMasking: Unleash- ing synergistic information by contrastive multimodal interactions, Advances in Neural Information Processing Systems 38 (2026) 15529–15555

    2026

  46. [46]

    H. Zou, M. Shen, C. Chen, Y . Hu, D. Rajan, E. S. Chng, UniS-MMC: Multimodal classification via unimodality- supervised multimodal contrastive learning, in: Findings of the Association for Computational Linguistics: ACL 2023, 2023, pp. 659–672

    2023

  47. [47]

    X. Chen, Q. Shuai, F. Hu, Y . Cheng, SDDA: A progressive self-distillation with decoupled alignment for multi- modal image–text classification, Neurocomputing 614 (2025) 128794

    2025

  48. [48]

    Kumar, R

    R. Kumar, R. Singhal, P. Kulkarni, D. Mehta, K. Jadhav, M3CoL: Harnessing shared relations via multimodal mixup contrastive learning for multimodal classification, Transactions on Machine Learning Research 2025 (2025)

    2025

  49. [49]

    McNemar, Note on the sampling error of the difference between correlated proportions or percentages, Psy- chometrika 12 (2) (1947) 153–157

    Q. McNemar, Note on the sampling error of the difference between correlated proportions or percentages, Psy- chometrika 12 (2) (1947) 153–157

    1947

  50. [50]

    R. J. Tibshirani, B. Efron, An introduction to the bootstrap, Monographs on statistics and applied probability 57 (1) (1993) 1–436. 28

    1993