PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.39172v2quant-phcs.LG

TL;DR

Prior quantum SSL work leaves open whether benefits extend beyond a single contrastive objective and whether they can be attributed to the quantum circuit. The paper compares width-matched QuFeX hybrids and classical models across three SSL frameworks for fingerprint recognition, finding higher single-run Top-5 scores across settings but noting substantial uncertainty and compute costs.

  • Problem

    Prior quantum SSL studies focus on a single contrastive objective, leaving its cross-objective benefits and attribution to the circuit unresolved.

    Source: §1 · p. 2

  • Method

    The study inserts QuFeX into SimCLR, MoCo v2, and BYOL and compares width-matched hybrids with classical counterparts using KNN on encoder features.

    Source: §1 · p. 2, §3.1 · p. 5

  • Results

    Top-5 rose from 48.43% to 57.06% for fingerprint SimCLR, from 28.22% to 40.30% for fingerprint MoCo v2, and from 23.70% to 25.55% for fingerprint BYOL.

    Source: §5.1 · p. 8, §7 · p. 18

  • Takeaways & Limitations

    The observed QuFeX Top-5 gains span contrastive and non-contrastive objectives, but the evidence supports a directional conclusion rather than a claim of statistical significance or circuit-only benefit.

    Source: §6.3 · p. 17, §7 · p. 18

    Single-seed runs provide no across-seed variance, and width matching does not match module capacity or isolate the quantum circuit's contribution.

    Source: §6.3 · p. 17

Abstract

from arXiv · show

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.

1 Introduction

The paper asks whether quantum-enhanced self-supervised learning improves fingerprint recognition across different objectives when representation width is matched, and whether any gains come from the circuit itself. It compares QuFeX hybrids with classical counterparts across three SSL frameworks, with QNet as a secondary comparison.

1 Introduction

  • Prior quantum SSL studies used a single contrastive objective, leaving unclear whether reported benefits generalize across objectives or are attributable to the quantum circuit.The introduction also identifies single or mixed accuracy metrics [2], single runs with large variance [20], unmatched representation width and qubit count [24], and the absence of prior quantum SSL on fingerprints.Source: §1 · p. 2
  • The study inserts QuFeX into SimCLR, MoCo v2, and BYOL, comparing each hybrid with its classical counterpart at matched width of 8 features and 8 qubits.It evaluates encoder features with KNN Top-1 and Top-5 on SOCOFing, with CIFAR-10 as a control.Source: §1 · p. 2
  • The paper frames its contribution as a unified hybrid framework and a width-matched comparison spanning contrastive and non-contrastive SSL objectives.The introduction presents these as contributions rather than evidence that the circuit alone causes the observed gains.Source: §1 · p. 2
  • QuFeX hybrids score higher than classical counterparts across all three objectives and on CIFAR-10, but the margins are modest, from single runs, and at higher compute cost.The secondary QNet comparison does not show the same advantage at the tested 15-epoch budget.Source: §1 · p. 2

2 Related Works

The related work covers contrastive and non-contrastive SSL methods, hybrid quantum-classical models, and fingerprint preprocessing. The study adapts these methods and preprocessing choices to compare classical and hybrid recognition models.

2.1 Classic Contrastive Classification Methods:

  • SimCLR learns from two augmented views without labels, while MoCo v2 [5] uses query-key matching and a queue of encoded keys; BYOL predicts an EMA target without negative pairs.The paper uses ResNet-18 backbones and defines the hybrid version by injecting a quantum-circuit layer into the representation network.Source: BYOL · pp. 3–4, §2.1 · p. 3

2.2 Hybrid Classic-Quantum Models

  • QSSL combines SimCLR with a parameterized quantum circuit, while QuFeX [15] combines QCNN-style convolution and pooling with parallel processing of feature maps and preserved qubit outputs.The study applies both QuFeX and QNet designs to the selected classical models for comparison on fingerprint recognition.Source: §2.2 · p. 4

2.3 Fingerprint Image Preprocessing

  • Gabor filtering extracts local spatial-frequency texture features, and Canny edge detection extracts fingerprint contours while reducing noise.The paper uses the described methods for fingerprint preprocessing and enhancement.Source: §2.3 · p. 4

3 Method

The method evaluates whether replacing the classical representation network with a same-width quantum circuit improves identification across contrastive and non-contrastive SSL. It uses a shared ResNet-18 pipeline and compares QuFeX and QNet circuit designs, while noting limits in capacity matching and quantum attribution.

3.1 Problem Formulation

  • The protocol pretrains on unlabeled fingerprint images from 600 identities, then evaluates frozen encoder representations with labeled-gallery KNN Top-1 and Top-5 accuracy.Within each classical-hybrid pair, the backbone, width, data split, and training hyperparameters are held fixed; circuit width equals qubit count.Source: §3.1 · p. 5

3.2 Encoder and hybrid representation network

  • All models use a ResNet-18 backbone and width-8 representation; KNN uses encoder features before the projection head, while the hybrid representation network uses QuFeX or QNet.The figure depicts two augmented views passing through a shared encoder and then a projection head; BYOL adds an EMA target branch and an online predictor.Source: §3.2 · p. 5, Figure 1 · p. 6

3.3 Comparison of Backbone Circuits: Expressibility, Entanglement, and Residual Design

  • QuFeX is QCNN-inspired, whereas QNet is a generic, highly expressive circuit; QNet has trainable entangling strength, while QuFeX uses fixed entangling gates and trainable single-qubit rotations.QNet adds mid-circuit projective measurement for nonlinearity; QuFeX remains unitary until final measurement, with practical nonlinearity from encoding and the classical residual.Source: §3.3 · p. 6
  • QuFeX's residual connection provides a trainable fallback to the identity map, but its advantage cannot be separated from that connection without a residual-matched classical control.The study does not test that control; QNet lacks a fallback path, so downstream representation is sent through its quantum channel.Source: §3.3 · p. 7

4 Experimental Settings

The experiments compare matched-width hybrid and classical SSL models on SOCOFing, with CIFAR-10 as a control, using encoder-feature KNN recognition. The section specifies the training setup, fingerprint preprocessing, and evaluation protocol.

4.1 Setup

  • Each configuration is a single run using ResNet-18, batch size 128, width 8, and Adam; BYOL trains for 200 epochs and MoCo v2 for 50.The SimCLR figures come from a separately trained run under the same backbone and width.Source: §4.1 · p. 7

4.2 Dataset and preprocessing

  • SOCOFing provides 6,000 grayscale images from 600 subjects; after augmentation and filtering, 29,514 images are preprocessed and split by identity into train, validation, and test sets.Preprocessing applies a Gabor filter followed by Canny edge detection.Source: §4.2 · p. 7

4.3 Integration into SSL frameworks and evaluation

  • SimCLR and MoCo v2 use contrastive pairs, with MoCo v2 adding a momentum encoder and negative queue, while BYOL uses no negatives; encoder-feature KNN reports Top-1 and Top-5.Hybrid/classical pairs use width 8, equal to 8 qubits, with the backbone, data split, image size, and batch size fixed.Source: §4.3 · p. 8

5 Results

QuFeX hybrids score higher than their classical counterparts in the reported single-run comparisons, with the contrastive-model advantage persisting across epoch budgets. Three-seed BYOL results qualify the single-run finding: its advantage varies by budget and is reliable only at 200 epochs.

5.1 Classical and Hybrid Self-Supervised Models Comparison

  • SimCLR, MoCo v2, and BYOL QuFeX hybrids all improve fingerprint test KNN Top-5 over their classical counterparts in single runs: 48.43%→57.06%, 28.22%→40.30%, and 23.70%→25.55%, respectively.On CIFAR-10, BYOL Top-5 rises from 97.03% to 97.52%; hybrid models cost 1.5× to 6× more per epoch.Source: §5.1 · p. 8, Table 1 · p. 9
  • For MoCo v2, the hybrid separates from the classical model within ten epochs, stays ahead on validation Top-5, and peaks near epoch 20 while the classical model is still improving.Its InfoNCE training loss is also consistently lower; the curves are single-run evidence, not across-seed estimates.Source: §5.1 · p. 8, Figure 2 · p. 10, Figure 3 · p. 11

5.2 Classical and Hybrid SimCLR and MoCo v2 with different epoch lengths

  • The SimCLR-QuFeX model reaches 57% Top-5, while QNet scores below classical SimCLR at its only tested budget of 15 epochs.The 200-epoch classical SimCLR and QNet runs were not completed, so their converged performance is unknown.Source: §5.2 · p. 10, Table 2 · p. 10
  • At 15, 50, 100, and 200 epochs, QuFeX SimCLR and MoCo v2 score higher than their classical counterparts on both test Top-1 and Top-5.The gap appears early and persists through each full run; QuFeX also has the highest accuracy and lowest loss at the reported training checkpoints.Source: §5.2 · pp. 10, 12, 13, Figure 9 · p. 15

5.3 BYOL across multiple seeds

  • Across three BYOL seeds, QuFeX's mean paired Top-5 gap is −1.51 at 50 epochs, +2.75 at 15, +0.66 at 100, and +2.59 at 200.Only at 200 epochs does QuFeX exceed the baseline across all three seeds with non-overlapping ranges; the paired test gives p ≈ 0.07 at n = 3.Source: §5.3 · p. 16, Table 8 · p. 16

6 Discussion

QuFeX improves KNN recognition across contrastive and non-contrastive settings, but the modest gains do not establish statistical significance or isolate the quantum circuit’s contribution.

6.1 A consistent but modest quantum benefit

  • QuFeX improves KNN recognition across SimCLR, MoCo v2, BYOL, and the CIFAR-10 control, while QNet does not show the same pattern at the tested single-layer depth.The paper characterizes the margins as modest and does not claim a large or state-of-the-art improvement.Source: §6.1 · p. 17

6.2 Why absolute accuracy is modest, and QNet versus QuFeX

  • A 600-class task, small grayscale images, an 8-dimensional bottleneck, and limited quantum-model training budgets bound absolute accuracy for both model types.QuFeX’s residual composition offers a trainable fallback absent from QNet; completing compute-limited QNet runs and comparing the circuits empirically remain future work.Source: §6.2 · p. 17

6.3 Limitations

  • Single runs leave across-seed variance unknown; width matching does not match module capacity, and the simulated hybrids cost 1.5× to 6× more per epoch.Epoch budgets differ across objectives, and the circuits were not run on real hardware; a capacity-matched classical module is an untested control.Source: §6.3 · pp. 17–18

7 Conclusion

The paper concludes that width-matched QuFeX hybrids score higher on KNN Top-5 across three SSL objectives and a CIFAR-10 control. The result is limited to single runs, higher compute cost, and one-layer injection.

7 Conclusion

  • The unified QuFeX framework scores higher than its classical counterpart on KNN Top-5 across SimCLR, MoCo v2, BYOL, and a CIFAR-10 control.The reported margins are modest and come from single runs at higher compute cost.Source: §7 · p. 18
  • The authors identify multiple seeds, error bars, capacity-matched ablations, and completed QNet runs as priorities for future work.These would help confirm the margins and isolate the quantum contribution.Source: §7 · p. 18
  • The reported results cover injection into one base-model layer; evaluating injection into more than one layer is left for future work..Source: §7 · p. 18

A.1 Upstream SimCLR testing at different depths

The appendix describes upstream SimCLR training and a depth comparison for classical and QNet-based models.

A.1 Upstream SimCLR testing at different depths

  • Upstream SimCLR training uses augmented photos as ground truth to produce fingerprint feature representations, then evaluates representation accuracy with model-predicted photos.Source: §A.1 · p. 18
  • Table 10 compares average training Top-5 accuracy for classical SimCLR and hybrid SimCLR-QNet across three representation layers.Source: Table 10 · p. 18

A.2 SimCLR Downstream Verification Accuracy

The appendix distinguishes linear evaluation from KNN monitoring for downstream testing of upstream training results.

A.2 SimCLR Downstream Verification Accuracy

  • Linear evaluation discards the projection head and adds a linear layer after the encoder.Source: §A.2 · pp. 18–19
  • Table 11 compares testing Top-5 accuracy for classical and hybrid SimCLR-QNet using two evaluation methods.Source: Table 11 · p. 19

Related papers

Loading 2609.39172v2…