A Width-Matched Comparison of Hybrid Quantum-Classical Self-Supervised Learning for Fingerprint Recognition

Maria S. Edwards; Kidwell Dlamini; Pin-An Lin; Wen-Hsien Hsu; Wen-Chieh Fang

quant-phSep 30, 2026

A Width-Matched Comparison of Hybrid Quantum-Classical Self-Supervised Learning for Fingerprint Recognition

Authors: Maria S. Edwards, Kidwell Dlamini, Pin-An Lin, Wen-Hsien Hsu, Wen-Chieh Fang

Organizations: Department of Computer Science and Information Engineering, National Dong Hwa University, Hualien, Taiwan · Department of Computer Science and Information Engineering, National Chiayi University, Chiayi City, Taiwan

Abstract

Fingerprint recognition is a widely deployed biometric, but supervised training requires large labeled enrollment sets. Self-supervised learning (SSL) removes this requirement, and hybrid quantum-classical models have been proposed to enrich the learned representations. Prior quantum SSL studies consider a single contrastive objective, so it is unclear whether reported benefits depend on the objective or can be attributed to the quantum circuit. We insert the QuFeX quantum feature-extraction module into three SSL frameworks, the contrastive SimCLR and MoCo v2 and the non-contrastive BYOL, and compare each hybrid with its classical counterpart at matched representation width (8 features, equal to 8 qubits) on the SOCOFing fingerprint dataset, with a CIFAR-10 control, using k-nearest-neighbor identification on encoder features. In single-run experiments the hybrid scores clearly higher for both contrastive objectives, whereas for BYOL a multi-seed analysis shows no reliable difference, suggesting that any benefit depends on the SSL objective. A hardware-efficient circuit (QNet) does not show the same gain. We examine whether the gains can be attributed to the quantum circuit, considering circuit architecture, trainable parameter count, nonlinearity, and the classical simulability of 8-qubit circuits.

Figures & tables

Panel of Figure 1 — Figure 1: Overall architecture. Each preprocessed fingerprint $\vec{x}_{i}$ (Gabor + Canny) is augmented into two views $\vec{x}^{1}_{i},\vec{x}^{2}_{i}$ and passed through a shared encoder: a CNN (ResNet-18) backbone followed by a representation network that is either a classical MLP or the quantum module (QuFeX / QNet). A projection head $g$ maps representations to $\vec{z}^{1}_{i},\vec{z}^{2}_{i}$ for the self supervised objective. For SimCLR and MoCo v2, the objective is contrastive with weight sharing; In BYOL, the target branch is an EMA target, the online branch adds a predictor $q$ , and a stop gradient ( $sg$ ) is applied to the EMA target. The KNN monitor evaluates the representation $\vec{y}^{1}_{i}$ before the projection head.

Method (variant)	Top-5
SOCOFing (600 cls.; rand. 0.83)
SimCLR, classical	48.43
SimCLR, hybrid	57.06
MoCo v2, classical	28.22
MoCo v2, hybrid	40.30
BYOL, classical	23.70

Table 1: Test KNN Top-5 accuracy (%) on encoder features (before the projection head). Hybrid uses the QuFeX quantum module; width

=8

(

=8

qubits) for the fingerprint runs.Training budgets: SimCLR 100 epochs, MoCo v2 50 epochs, BYOL 200 epochs; the CIFAR-10 control uses BYOL trained for 200 epochs. All values are from single runs, taken from the final-epoch model. Better value in each classical/hybrid pair in bold.

Figure 2: MoCo v2 on fingerprints (SOCOFing, 50 epochs), classical vs. hybrid (QuFeX) at width $8$ . The hybrid separates from the classical baseline within the first ten epochs and stays above it for the whole run on validation Top-5, so the advantage is not an artefact of the single best epoch. Markers are single-run values (seed 0); error bars show $\pm 1$ s.d. of the metric over a forward window of a few consecutive epochs, i.e. within-run epoch-to-epoch volatility, not confidence intervals, across-seed variance, or the significance of the hybrid–classical gap.

Panel of Figure 3 — Figure 3: MoCo v2 training loss (InfoNCE) on fingerprints, classical vs. hybrid (QuFeX). The hybrid loss is consistently lower, so loss and recognition (Fig. 2 ) agree. Markers are single-run values (seed 0); error bars show $\pm 1$ s.d. of the loss over a forward window of ${\approx}20$ consecutive batches, i.e. within-run batch-to-batch volatility inflated by the training-curve slope, not confidence intervals or across-seed variance.

Model	15 Epochs	50 Epochs	100 Epochs	200 Epochs
SimCLR-Classic	34.98	44.25	48.43	51.23
SimCLR-QuFex	45.49	55.25	57.06	56.13
SimCLR-QNet	32.27	—	—	—

Table 2: Highest Percentage of Top-5 Test Accuracy in SimCLR classical and hybrid models. In bold, the highest result obtained from all the runs.

Model	15 Epochs	50 Epochs	100 Epochs	200 Epochs
MoCo-Classic	21.28	28.22	41.91	47.77
MoCo-QuFex	30.52	40.30	45.03	51.79

Table 3: Highest Percentage of Top-5 Test Accuracy in MoCo v2 classical and hybrid models. In bold, the highest result obtained from all the runs.

15-Epochs
	SimCLR				MoCo
	Classical		Hybrid		Classical		Hybrid
Epoch N°	Loss	Acc	Loss	Acc	Loss	Acc	Loss	Acc
1 Epoch	3.51	23.47	3.27	24.32	6.39	11.28	6.33	14.53
10 Epoch	0.96	34.75	0.78	44.41	4.31	18.90	3.93	22.75
15 Epoch	0.82	34.55	0.64	45.24	4.02	21.27	3.68	30.52

Table 4: Comparison Table for epochs=15 denoting the testing values at specific epoch numbers

50-Epochs
	SimCLR				MoCo
	Classical		Hybrid		Classical		Hybrid
Epoch N°	Loss	Acc	Loss	Acc	Loss	Acc	Loss	Acc
1 Epoch	3.56	16.65	3.17	20.48	6.70	3.00	6.41	2.27
30 Epoch	0.59	42.89	0.50	51.61	3.26	23.65	2.78	34.29
50 Epoch	0.48	43.71	0.43	51.85	2.90	28.22	2.47	40.07

Table 5: Comparison Table for epochs=50 denoting the testing values at specific epoch numbers

100-Epochs
	SimCLR				MoCo
	Classical		Hybrid		Classical		Hybrid
Epoch N°	Loss	Acc	Loss	Acc	Loss	Acc	Loss	Acc
1 Epoch	3.69	16.11	3.23	25.33	7.43	2.71	6.35	7.59
30 Epoch	0.61	41.11	0.49	49.19	3.09	31.52	2.67	39.30
50 Epoch	0.49	45.00	0.41	53.40	2.66	41.23	2.47	43.12

Table 6: Comparison Table for epochs=100 denoting the testing values at specific epoch numbers

200-Epochs
	SimCLR				MoCo
	Classical		Hybrid		Classical		Hybrid
Epoch N°	Loss	Acc	Loss	Acc	Loss	Acc	Loss	Acc
1 Epoch	3.66	15.34	3.50	20.68	6.57	3.74	6.28	22.10
50 Epoch	0.56	40.72	0.45	52.96	2.54	38.29	2.35	47.25
100 Epoch	0.40	44.02	0.37	54.04	22.25	43.09	2.21	46.24

Table 7: Comparison Table for epochs=200 denoting the testing values at specific epoch numbers

Panel of Figure 5 — Figure 5: Hybrid Quantum-Classical accuracy versus loss for SimCLR models with one QuFex layer injected at different epoch sizes.

Panel of Figure 7 — Figure 7: Hybrid Quantum-Classical accuracy versus loss for MoCo v2 models with one QuFex layer injected at different epoch sizes.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Avg. Top-5 Accuracy ( $\%$ )	Classical SimCLR	Hybrid SimCLR-QNet
Repr. Layer with 3 layers	56.62	66.96
Repr. Layer with 2 layers	53.65	65.36
Repr. Layer with 1 layer	62.29	54.62

Appendix

Table 8: Average Training Top-5 Accuracy Comparison between Classical SimCLR and Hybrid SimCLR-QNet, tested with 3 different representation layers.

Top-5 Accuracy ( $\%$ )	Classical SimCLR	Hybrid SimCLR-QNet
Linear Evaluation	11	12
KNN Monitor	23	26

Appendix

Table 9: Testing Top-5 Accuracy Comparison between Classical SimCLR and Hybrid SimCLR-QNet, using two different evaluation methods

Explore similar work

Do We Really Need Quantum Machine Learning?: A Multidimensional Empirical Study
May 27, 2026Sudip Vhaduri, Ryan Gammon, Sayanton DibboQuantum Machine Learning Quantum Kernels
Department of Computer Science, University of Alabama, AL 35487 · Clemson University
Q-PhotoNAS: Hybrid Quantum Neural Architecture Search Framework on Photonic Devices
May 21, 2026Farah Elnakhal, Alberto Marchisio, Nouhaila Innan +2Neural Architecture Search Nanophotonic
Science Division, New York University Abu Dhabi, UAE · eBrain Lab, Division of Engineering, New York University Abu Dhabi, PO Box 129188, Abu Dhabi, UAE · Center for Quantum and Topological Systems, NYUAD Research Institute, New York University Abu Dhabi, UAE +1
On the Complementarity of Quantum and Classical Features: Adaptive Hybrid Quantum-Classical Feature Fusion for Breast Cancer Classification
Apr 24, 2026Yasmin Rodrigues Sobrinho, João Renato Ribeiro Manesco, João Paulo PapaHybrid Quantum-Classical Pipeline Direct Feature Fusion
Department of Computing, São Paulo State University, Ave. Engenheiro Luiz Edmundo Carrijo Coube, 14-01 - Vargem Limpa, Bauru, 17033-360, São Paulo, Brazil

May 27, 2026, cs.CVJ/K move · Enter open · S save

Do We Really Need Quantum Machine Learning?: A Multidimensional Empirical Study

Sudip Vhaduri, Ryan Gammon, Sayanton Dibbo

Department of Computer Science, University of Alabama, AL 35487 · Clemson University

Quantum Machine Learning Quantum Kernels Mnist Computational Cost

Open paper AskPDF arXiv

The rapid growth of computer vision and increasingly complex image recognition tasks has exposed fundamental computational limitations of classical machine learning models, motivating the exploration of quantum computing as an emerging new paradigm. This paper presents a comprehensive benchmarking study of classical and quantum machine learning models for image recognition on the MNIST handwritten digit dataset, evaluating both traditional models, a Classical Support Vector Machine (CSVM) and a Quantum Support Vector Machine (QSVM), and deep neural network models, a Classical Convolutional Neural Network (CCNN) and a Quantum Convolutional Neural Network (QCNN), across four performance dimensions: classification accuracy, computational runtime, parameter count, and memory requirements. Experiments are conducted as functions of both feature dimensionality and sample size, and across CPU and GPU execution environments, providing a controlled, multidimensional comparison to address gaps in prior work. For the SVM-based models, QSVM consistently outperforms CSVM in accuracy, reaching

\sim

0.90 versus

\sim

0.85 at 1,000 samples, with a higher computational cost. A feature count of 10 qubits and a sample size in the range of 200 -- 500 emerge as practical operating points that balance accuracy and runtime. For the neural network models, CCNN and QCNN achieve comparable classification accuracy, both exceeding 0.96 at 64 features and 60,000 samples, yet QCNN offers substantially superior parameter and memory efficiency, requiring

\sim

94% fewer parameters and

\sim

75% less memory than CCNN at higher feature counts, while incurring higher runtime. Across both model families, quantum models consistently outperform classical models by greater margins in accuracy as feature dimensionality or sample size increases.

Abstract

Explore similar work

Do We Really Need Quantum Machine Learning?: A Multidimensional Empirical Study

Q-PhotoNAS: Hybrid Quantum Neural Architecture Search Framework on Photonic Devices

On the Complementarity of Quantum and Classical Features: Adaptive Hybrid Quantum-Classical Feature Fusion for Breast Cancer Classification