Empirical comparisons between hybrid quan-tum-classical machine learning (HQML) models and classical baselines are often confounded by mismatched model capacity and training budgets. We propose a fair, budgetaware benchmarking protocol for HQML evaluation on classical quantum simulators. Our proposed framework ensures controlled comparisons by matching the number of trainable parameters and optimization steps under a shared early-stopping policy, while explicitly reporting the wall-clock training cost. Variational HQML models with different circuit depths are evaluated on a binary FashionMNIST task $(N=12000)$ using PCA-reduced inputs mapped to $n=8$ qubits and are compared against parameter-matched classical multilayer perceptrons.The results obtained indicate that shallow HQML models achieve statistically comparable predictive performance under strictly matched computational budgets, albeit at substantially higher computational cost. These findings emphasize the necessity of budget-aware evaluation and provide a reproducible foundation for principled HQML benchmarking.
Empirical comparisons between hybrid quan-tum-classical machine learning (HQML) models and classical baselines are often confounded by mismatched model capacity and training budgets. We propose a fair, budgetaware benchmarking protocol for HQML evaluation on classical quantum simulators. Our proposed framework ensures controlled comparisons by matching the number of trainable parameters and optimization steps under a shared early-stopping policy, while explicitly reporting the wall-clock training cost. Variational HQML models with different circuit depths are evaluated on a binary FashionMNIST task (N=12000) using PCA-reduced inputs mapped to n=8 qubits and are compared against parameter-matched classical multilayer perceptrons.The results obtained indicate that shallow HQML models achieve statistically comparable predictive performance under strictly matched computational budgets, albeit at substantially higher computational cost. These findings emphasize the necessity of budget-aware evaluation and provide a reproducible foundation for principled HQML benchmarking.
Quantum machine learning has emerged as a promising approach to enhance computational efficiency by leveraging the principles of quantum computing. However, the limitations of current Noisy Intermediate-Scale Quantum (NISQ) devices, such as noise, limited qubit availability, and circuit depth constraints, restrict the implementation of fully quantum models. To address these challenges, this study focuses on the design, implementation, and performance evaluation of hybrid quantum-classical machine learning (HQML) models. The proposed approach integrates parameterized quantum circuits with classical optimization techniques to enable efficient learning within NISQ environments. The study employs standard benchmark datasets, including Iris, Breast Cancer, and MNIST, to evaluate the performance of the hybrid model. The results indicate that the HQML model achieves competitive accuracy on small and medium-sized datasets while maintaining balanced precision, recall, and F1-score. However, performance declines for complex datasets due to hardware limitations and noise effects. Additionally, the hybrid model demonstrates a lower number of parameters compared to classical deep learning models but requires higher training time due to iterative quantum-classical optimization. The findings highlight that hybrid quantum-classical models provide a practical and scalable approach for utilizing quantum computing in the current technological landscape. Although challenges related to noise, scalability, and computational overhead persist, advancements in quantum hardware and algorithm design are expected to improve performance. This study contributes to the growing field of quantum machine learning by providing a systematic framework for evaluating hybrid models on NISQ devices and identifying key areas for future research.
Hybrid Quantum-Classical Machine Learning (HQML) is an emerging technology that can be used for solving problems with higher computational intensity. This paradigm is designed to solve problems that remain challenging to classical machine learning approaches. Noisy Intermediate-Scale Quantum (NISQ) devices (which are today’s quantum computers) are prone to frequent errors and cannot perform fault-tolerant quantum computation. NISQ devices have a limited number of qubits, and as the number of qubits increases, circuit depth also increases, leading to increase in noise. Therefore, selecting the optimal number of qubits remains a challenge. The goal is to develop a framework that tests a Variational Quantum Classifier (VQC) with a different numbers of qubits. For each configuration, the model’s prediction accuracy and the hardware resources it consumes (gate count and circuit depth) are evaluated. Using this data, the best qubit configuration that provides the optimal balance between accuracy and resource utilization can be identified. Experiments were conducted using the Wisconsin Breast Cancer dataset, where benign and malignant tumors are classified. The dataset was preprocessed by applying Principal Component Analysis (PCA) to reduce the feature space to a maximum of six principal components. A ZZFeatureMap was used to encode the reduced features, which were then processed by a VQC with a RealAmplitudes ansatz. The COBYLA optimizer was used to train the VQC, and the experiment was repeated using different numbers of qubits from 2 to 6. Quantum circuits were transpiled into realistic IBM fake backends (FakeManilaV2 and FakeJakartaV2) for hardware-aware analysis, obtaining two key metrics: CNOT gate count and circuit depth. An AerSimulator configured with the FakeManilaV2 noise model was used to carry out the noise-aware simulation. The framework was evaluated using classification accuracy, training time, circuit depth, CNOT gate count, and noise-aware accuracy. Based on these criteria, the optimal qubit configuration that balances predictive performance with quantum hardware resource utilization was identified. The results identify 2 qubits as the optimal “sweet spot,” achieving 83.21% accuracy under ideal simulation and 81.75% under noise-aware simulation while maintaining lower resource requirements. These findings provide practical guidelines for resource-efficient hybrid quantum machine learning on NISQ hardware.
Quantum Transfer Learning (QTL) offers a promising approach for visual quantum machine learning under near-term constraints, where limited qubit counts, shallow circuit depths, and costly hybrid optimization restrict end-to-end quantum training. In this setting, pretrained classical backbones can extract high-level visual features, while compact quantum modules operate as trainable classification heads. However, existing QTL results are difficult to compare because they often differ in datasets, preprocessing, backbone settings, qubit budgets, circuit designs, optimization choices, and reporting protocols. This work presents a controlled benchmarking methodology for evaluating representative QTL methods under a unified transfer-learning pipeline. The benchmark compares DQN-QTL, QPIE-QTL, AE-CQTL, PVCQTL, and ED-QTL under shared preprocessing rules, frozen-backbone settings, training conditions, and reporting metrics. The evaluation focuses on Fashion-MNIST and Hymenoptera Ants vs. Bees as the two main datasets, while CIFAR-10 is used to provide additional configuration-level evidence on a harder natural-image task. Beyond predictive performance, the benchmark analyzes circuit size, trainable parameters, quantum parameters, training time, and architectural sensitivity to qubit count and circuit depth. The results show that no single QTL family dominates across all settings: performance depends on the dataset, encoding strategy, circuit design, and computational cost. These findings highlight the need for resource-aware QTL evaluation and provide guidance for selecting hybrid quantum-classical transfer models under near-term resource constraints.
Abstract Hybrid quantum machine learning (QML) relies on quantum variational circuits (QVCs) trained together with classical components. While gradient-based training is standard, it is costly on near-term hardware and vulnerable to barren plateaus, motivating gradient-free alternatives. However, there is limited systematic evidence comparing gradient-free methods to strong gradient baselines under matched quantum-evaluation budgets. A benchmark of 15 gradient-based, gradient-free, and hybrid optimizers is conducted on four binary classification datasets and two three-qubit QVC architectures using a unified experimental protocol under a fixed per-step quantum-evaluation budget. Across tasks, Adam is the strongest gradient-based method. Among the purely gradient-free schemes, Asexual Reproduction Optimization is the strongest optimizer; tournament-selection and elite-pool genetic algorithms (GA-Tourn, GA-Elite) also frequently match the best gradient-based configurations. Among the hybrid schemes, Gradient Lexicase Selection with Adam (GLS- Adam ) performs comparably. In contrast, the two gradient-free methods big bang–big crunch and a cross-entropy method variant often fail to improve beyond chance. Additional sensitivity analyses show that circuit depth can change the preferred optimizer, whereas finite-shot sampling budget and post-training backend-specific noise do not materially change the main conclusions. Overall, gradient-based methods remain robust defaults for hybrid QML, while a small subset of gradient-free and hybrid approaches provides viable alternatives under realistic quantum-evaluation constraints.
Image-based quantum machine learning usually follows a hybrid pipeline: a classical encoder compresses an image, and a quantum circuit then acts as a kernel, feature map, trainable classifier, or measurement-based readout. Yet it remains unclear which of these roles contributes useful predictive structure because prior studies often change the input representation, classical baseline, and measurement budget simultaneously. We present a controlled benchmark spanning five roles: quantum-kernel geometry, fixed product and entangled features, trainable data-reuploading representations, shot-aware probability readout, and task-adaptive Pauli measurement. Six quantum-derived methods, six matched classical controls, and two unmatched neural references are evaluated on MNIST, Fashion-MNIST, CIFAR-10, and BreastMNIST using a shared frozen ResNet-18 and train-only PCA interface. We assess exact and finite-shot inference, predictive performance, calibration, selective risk, resource requirements, and statistical reliability. PdrQC-Matched was the strongest quantum-derived method on all four datasets. C-RBF was better on Fashion-MNIST and CIFAR-10 by 2.19 and 3.50 accuracy points, while BreastMNIST showed no detectable difference. Despite lower exact accuracy, PdrQC-Matched achieved significantly lower Brier score and AURC than C-RBF on Fashion-MNIST and CIFAR-10. Finite-shot results showed no automatic benefit from entanglement or shot-aware training. Overall, image-QML conclusions depend on quantum role, measurement budget, classical control, and evaluation metric.
High-dimensional classification demands expressive feature representations, robust optimization, and manageable computational cost. This paper proposes a hybrid quantum– classical framework combining classical feature screening and compression, a shallow parameterized quantum circuit, measurement-derived quantum features, and a regularized classical classifier. The design explicitly addresses the mismatch between hundreds of original attributes and the small qubit counts available to near-term devices. Its training objective combines cross-entropy, complexity regularization, and a noise-sensitivity penalty, while alternating classical and quantum parameter updates. A reproducible evaluation protocol is specified using stratified data partitions, classical and quantum baselines, noise sweeps, and resource accounting. To demonstrate how the analysis should be reported, illustrative numerical results and plots are provided; they are not outcomes of executed quantum experiments and must be replaced by measured values before publication. The framework offers a testable route for evaluating whether compact quantum representations improve predictive performance or robustness relative to equally budgeted classical models, without presupposing quantum advantage.