Deep transfer learning with convolutional backbones achieves high performance on multispectral remote sensing benchmarks when sufficient labeled data are available. However, model behavior under extreme label scarcity remains less explored. This paper investigates whether compact variational quantum classification heads exhibit favorable generalization behavior in low-label settings for Sentinel-2 land use and land cover classification on the EuroSAT dataset. A pretrained ResNet50 backbone is combined with three classification heads: a standard fully connected head, a parameter-matched compact classical head, and variational quantum circuit (VQC) heads with four and eight qubits. Experiments are conducted using $\mathbf{3 0 \%}, \mathbf{2 0 \%}, \mathbf{1 0 \%}$, and $\mathbf{5 \%}$ of labeled training data under identical validation and test splits. Across supervision levels, all models perform comparably under moderate data availability. However, as labeled data become scarce, quantum heads consistently exhibit higher test accuracy and smaller generalization gaps relative to classical baselines. At 10% and 5% supervision, the 8 -qubit VQC shows the largest improvements, achieving accuracy gains of up to 2.3 percentage points at 10% supervision and 4.3 percentage points at 5% supervision compared to the standard classical head across three random seeds. These observations are consistent with the hypothesis that structurally constrained quantum heads may provide an implicit regularization effect in data-scarce regimes. While no claim of quantum computational advantage is made, the results highlight the potential of compact hybrid architectures for low-label multispectral learning.
Deep transfer learning with convolutional backbones achieves high performance on multispectral remote sensing benchmarks when sufficient labeled data are available. However, model behavior under extreme label scarcity remains less explored. This paper investigates whether compact variational quantum classification heads exhibit favorable generalization behavior in low-label settings for Sentinel-2 land use and land cover classification on the EuroSAT dataset. A pretrained ResNet50 backbone is combined with three classification heads: a standard fully connected head, a parameter-matched compact classical head, and variational quantum circuit (VQC) heads with four and eight qubits. Experiments are conducted using 30%,20%,10%, and 5% of labeled training data under identical validation and test splits. Across supervision levels, all models perform comparably under moderate data availability. However, as labeled data become scarce, quantum heads consistently exhibit higher test accuracy and smaller generalization gaps relative to classical baselines. At 10% and 5% supervision, the 8 -qubit VQC shows the largest improvements, achieving accuracy gains of up to 2.3 percentage points at 10% supervision and 4.3 percentage points at 5% supervision compared to the standard classical head across three random seeds. These observations are consistent with the hypothesis that structurally constrained quantum heads may provide an implicit regularization effect in data-scarce regimes. While no claim of quantum computational advantage is made, the results highlight the potential of compact hybrid architectures for low-label multispectral learning.
We investigate variational quantum classifiers (VQCs) for land-cover classification from multispectral satellite imagery, adopting a feature-map perspective in which the quantum circuit defines a nonlinear data embedding while the readout determines how this representation is exploited. Using the EuroSAT-MS dataset, we perform a systematic one-vs-one evaluation across all class pairs under a controlled experimental protocol, comparing classical baselines (logistic regression, SVMs, neural networks) with VQCs employing both linear readout and quantum-kernel SVM strategies. Our results show that, while VQCs with linear readout do not outperform strong classical baselines such as RBF-SVM, the same trained quantum feature map can significantly improve performance when reused within a kernel-based decision framework. A qubit-count sweep further reveals saturation effects consistent with the mismatch between exponential Hilbert space dimension and linear parameter scaling. Overall, our findings highlight that the effectiveness of quantum models depends critically on the interplay between representation and readout, and that meaningful gains may arise from combining learned quantum feature maps with classical decision mechanisms rather than seeking direct replacement of classical models.
Abstract Metal–insulator transition (MIT) materials are promising for electronic, sensing, and neuromorphic technologies, but their data-driven classification remains challenging because experimentally verified examples are scarce, heterogeneous, and strongly imbalanced. We evaluate whether compact quantum classifiers provide useful predictive readouts under these conditions without assuming or claiming quantum advantage. In a protocol-controlled hybrid graph–quantum benchmark, fold-local crystal graph convolutional neural networks are trained on 316 structure-available compounds, and their frozen embeddings are supplied to linear, multilayer-perceptron, variational quantum circuit (VQC), and data re-uploading (REUP) heads. Grouped nested five-fold cross-validation, five random seeds, validation-only architecture selection, low-data training fractions, bottleneck-matched classical controls, calibration analysis, and dependence aware statistical tests are used to limit representation leakage and test-informedmodel selection. At full data, the validation-selected VQC achieves a precision–recall area under the curve (PR-AUC) of 0.870 [0.829, 0.910], comparable to the linear (0.869) and multilayer-perceptron (0.861) heads, whereas the selected REUP family reaches 0.811 [0.765, 0.858]. Classical bottleneck controls remove evidence of a quantum-specific gain, and REUP selects its single-upload limit in82.7% of fold–seed–fraction evaluations. In a complementary exploratory benchmark using six physical descriptors for 343 materials, training-fold resampling raises REUP F1 from 0.431 to at most 0.530, but a random forest remains stronger at 0.707 and provides better calibration. These results show that quantum circuits can serve as competitive compact readouts of learned materials representations, but their apparent benefits depend strongly on representation control, architecture selection, imbalance treatment, and calibration.
Label noise is prevalent in medical image datasets, arising from observer disagreement and ambiguous cases and deep networks memorize the mislabeled examples at the cost of generalization. Current solutions such as loss correction and sample selection add training complexity and generally assume the noise rate is known or estimable. This paper investigates whether robustness can be obtained architecturally, by routing a frozen backbone's features through a low-dimensional bounded bottleneck and classifying them with a head whose function class is intrinsically smooth. Variational quantum circuits with angle encoding and a linear cost function realise exactly such a class. Their use in medical imaging, however, has been limited by the small number of qubits available on noisy intermediate-scale devices, which typically restricts inputs to low-resolution grayscale images. We therefore couple a frozen ImageNet-pretrained ViT-B/16 backbone to a trainable linear bottleneck and an eight-qubit variational quantum circuit, allowing 224x224 RGB inputs to be processed. Under clean labels the model leads five of six MedMNISTv2 benchmarks against published hybrid quantum-classical models, with margins of 17.4 and 15.7 percentages on BloodMNIST and PathMNIST. The advantage emerges under corrupted labels. At symmetric noise rates of 30% and 50% the model outperforms many published quantum and noise-robust classical baselines on all six datasets without any explicit noise-handling component. Against a control that shares the backbone, bottleneck, parameter count and optimizer and replaces only the circuit with a ReLU layer, it retains 89.75% accuracy on PathMNIST at 50% noise where the control falls to 78.96%. Direct measurement of training dynamics shows that this advantage is not an inability to fit corrupted labels, which the circuit retains, but greater stability under prolonged training on them.
Quantum transfer learning combines pretrained classical deep learning models with quantum circuits to reuse expressive feature representations while limiting the number of trainable parameters. In this work, we introduce a family of compact quantum transfer learning architectures that attach variational quantum classifiers to frozen convolutional backbones for image classification. We instantiate and evaluate several classical-quantum hybrid models implemented in PennyLane and Qiskit, and systematically compare them with a classical transfer-learning baseline across heterogeneous image datasets. To ensure a realistic assessment, we evaluate all approaches under both ideal simulation and noisy emulation using noise models calibrated from IBM quantum hardware specifications, as well as on real IBM quantum hardware. Experimental results show that the proposed quantum transfer learning architectures achieve competitive and, in several cases, superior accuracy while consistently reducing training time and energy consumption relative to the classical baseline. Among the evaluated approaches, PennyLane-based implementations provide the most favorable trade-off between accuracy and computational efficiency, suggesting that hybrid quantum transfer learning can offer practical benefits in realistic NISQ era settings when feature extraction remains classical.
The practical value of quantum machine learning must be derived from efficiency rather than scale, as current NISQ processors support only small, shallow circuits. This paper evaluated whether variational quantum circuits (VQCs) can reduce the trainable parameter count in an image classification scenario while preserving accuracy. Experimentation compared quantum classifier heads to parameter-matched classical heads on an identical set of frozen features drawn from the Fashion-MNIST dataset by a CNN. 9 configurations of a quantum circuit crossed 4, 6, and 8 qubits with 1, 2, and 4 entangling layers, each matched to a multilayer perceptron with an approximately equal parameter count. The linear probes served as a floor baseline for all 105 runs across 5 random seed values. 7 out of 9 quantum-based configurations met the mean test accuracy of no more than 1 percentage point below its matched classical head, enforced by the pre-specified criterion. The 2 failures were single-layer circuits, with each falling roughly 8 points below its matched classical head and 12 points below the circuit-free linear probe; in total, 6 configs with 2+ layers passed, with the smallest at 74 trainable parameters. The general trend noticed was that accuracy increased progressively with depth. The 4-layer circuits exceeded the linear probe at every qubit count and produced the highest accuracy at 6 and 8 qubits. The simulation ran noiselessly; VQCs preserved accuracy at the matched parameter budget when depth was given, showcasing depth rather than qubit count gating performance.
Supplementary Data Description This supplementary dataset provides the complete experimental results underlying Figures 4–9 of the manuscript. Each CSV file corresponds to one figure and reports mean performance values (and standard deviations where applicable), computed over multiple independent runs with different random seeds. Fig4_Low_Data_Performance.csv Contains classification accuracy (mean ± standard deviation) for Linear, MLP, and VQL heads across varying labeled target data fractions (1%, 5%, 10%, 20%). This dataset supports the low-data performance analysis presented in Fig. 4. Fig5_Parameter_Matched.csv Reports accuracy statistics for parameter-matched adaptation heads, including Linear Probe, MLP Adapter, Bottleneck Adapter, and the proposed VQL Head. This file enables fair comparison of adaptation strategies under comparable trainable parameter budgets (Fig. 5). Fig6_Domain_Shift.csv Provides source-domain accuracy, target-domain accuracy, and the resulting accuracy drop for each model under domain shift conditions. These results quantify robustness to distributional changes, as shown in Fig. 6. Fig7_Ablation.csv Contains ablation results evaluating the effect of quantum circuit depth and number of qubits on classification accuracy. This dataset supports the architectural analysis in Fig. 7. Fig8_Noise_Aware.csv Includes classification accuracy values under increasing quantum noise probabilities for all adaptation heads, enabling noise robustness evaluation corresponding to Fig. 8. Fig9_Convergence.csv Reports training accuracy as a function of training epochs for Linear, MLP, and VQL heads, illustrating convergence behavior and optimization stability (Fig. 9). Reproducibility Note All values are reported as mean performance over multiple independent runs, ensuring statistical reliability and reproducibility of the experimental findings.