Disentangling encoding, preprocessing, training budget and estimator artefacts in pre-training diagnostics for hybrid quantum-classical machine learning
A preregistered follow-up study testing six hypotheses that disentangle confounds in an earlier exploratory campaign on pre-training diagnostics for hybrid quantum-classical machine learning models. The study isolates the effects of data preprocessing, training budget, learning rate, matched circuit resources, and estimator sample size on the trainability failures previously observed at eight qubits.
A preregistered follow-up study testing six hypotheses that disentangle confounds in an earlier exploratory campaign on pre-training diagnostics for hybrid quantum-classical machine learning models. The study isolates the effects of data preprocessing, training budget, learning rate, matched circuit resources, and estimator sample size on the trainability failures previously observed at eight qubits.
A recurring claim in applied quantum machine learning [1],[2],[3]. The second is data scarcity relative to model capacity, which makes over-parameterized classical networks prone to memorizing scanner-specific artifacts rather than tumor morphology [4],[5],[6]. Convolutional and transfer-learning pipelines (ResNet, EfficientNet, VGG) remain the strongest classical baselines when training data is abundant [7], [8]. Quantum machine learning has been proposed as a partial answer to the capacity-versus-data problem [9]. In a variational quantum classifier, classical features are mapped onto qubit rotations or phases, and a parameterized circuit learns a decision boundary in the resulting Hilbert space [10], [11]. Because an n-qubit circuit spans a 2ⁿ-dimensional state space using only O(n) physical qubits and, for fixed depth, O(n) trainable parameters, a compact quantum classifier is argued to generalize more efficiently from small samples than a classical network with far more weights [12],[13]. Data re-uploading strengthens this claim: repeating the encoding block between variational layers turns even a single- qubit circuit into a universal approximator over the frequency spectrum the encoding induces [14],[15]. Whether this holds in practice is separate from whether it is theoretically available [18]. Published quantum medical-imaging classifiers typically report a single circuit, once, at a single qubit count, against classical baselines that are rarely varied over more than one architectural axis [16], [17]. It is therefore difficult to attribute a reported accuracy figure to quantum processing rather than to encoding, classifier family, qubit budget, or a favorable training run [19],[20]. This paper instead treats the quantum stage as an experimental platform, varying qubit count (8 vs. 16), encoding family (angle, phase, IQP/ZZ), entanglement topology (CZ-ring, CNOT-ring), and classifier family (QNN, QSVM, quantum ensemble, QCNN) one at a time against a fixed preprocessing and evaluation protocol [21],[22]. This design attributes an observed accuracy change to the pipeline component that produced it, and it lets us report honestly that a deeper, angle- encoded variational circuit does not converge reliably at the qubit
Quantum machine learning is a promising paradigm for learning from limited data, a central bottleneck in domains such as medical imaging, clinical trials, and rare diseases. Quantum convolutional neural networks (QCNNs) are particularly attractive in this setting, combining a hierarchical architecture with strong inductive bias and a parameter count that grows only logarithmically with system size. Their appeal rests on the generalization bounds of Caro et al. (2022), which show that the generalization error of a quantum model scales with the number of trainable parameters rather than with the Hilbert-space dimension, placing QCNNs in a potentially sample-efficient regime. We develop a hardware-compatible QCNN with mid-circuit measurement and classical feed-forward, and show on a binary handwritten-digit task that strong test performance is achievable from as few as 10 training samples, with the generalization error decreasing as the training set grows. At a matched 45-parameter budget the QCNN learns where an equally small classical convolutional network stays at chance, although an unconstrained classical baseline with roughly 25,000 parameters remains strongest when data are plentiful. Transpiling amplitude and angle encoded circuits across image resolutions from 2x2 to 512x512 pixels then exposes the dominant scaling bottleneck: amplitude encoding stays qubit-efficient but grows extremely deep, whereas angle encoding stays shallow but becomes qubit-prohibitive. On the medically motivated BreastMNIST benchmark the QCNN does not surpass the unconstrained classical network, yet it learns consistently above chance using orders of magnitude fewer parameters. Our results indicate that for QCNNs, learning from few samples is attainable in practice, whereas scaling to realistic image data is constrained less by optimization than by data encoding and hardware execution.
This work addresses the challenge of enabling practitioners without quantum expertise to transition from classical to hybrid quantum-classical machine learning workflows. We propose a three-stage framework: starting with a classical self-training model, then introducing a minimal hybrid quantum variant, and finally applying diagnostic feedback via QMetric to refine the hybrid architecture. In experiments on the Iris dataset, the refined hybrid model improved accuracy from 0.31 in the classical approach to 0.87 in the quantum approach. These results suggest that even modest quantum components, when guided by proper diagnostics, can enhance class separation and representation capacity in hybrid learning, offering a practical pathway for classical machine learning practitioners to leverage quantum-enhanced methods.
Abstract Hybrid quantum machine learning combines classical minimization with parameterized quantum circuits, enabling powerful learning models even on near-term quantum hardware. This work presents an evaluation of hybrid quantum machine learning classifiers using the data re-uploading framework across nine binary and multi-class classification tasks with varying geometric complexity. We compare three quantum circuit architectures, single-qubit, two-qubit, and two-qubit entangled, trained with three classical minimizers: COBYLA, Nelder-Mead, and SLSQP. Our experiments span training sizes from 1 to 200 samples and reveal that two-qubit classifiers, when paired with the SLSQP minimizer, achieve the highest overall test accuracy of 99.775% on structured tasks and strong performance (up to 89.78%) on complex multi-class pattern. Among minimizers, SLSQP consistently provides the best generalization, while entanglement improves accuracy in non-linear and fragmented tasks. By analyzing accuracy trends, generalization gaps, and confusion matrices, this study offers practical guidance for designing quantum classifiers that scale effectively and generalize well, providing a foundation for real-world Hybrid quantum machine learning applications.
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.
A growing body of work has questioned whether quantum machine learning (QML) offers any genuine advantage over classical methods on realistic data, with recent results showing that many apparent quantum kernel advantages are reproducible by classical surrogates once sample size, hyperparameters, and estimator variance are properly controlled. We conduct a pre-registered, falsification-first investigation into one specific, previously unexamined variant of this question: does the causal sparsity of a data-generating process as opposed to generic statistical structure modulate quantum kernel advantage or variational quantum circuit (VQC) trainability? Using a controllable structural-causal-model (SCM) data generator, a moment-matched dense-null control, and a battery of eight experiments spanning tabular regression, PDE solving, and gradient-landscape diagnostics, we find: (i) an apparent quantum-kernel advantage at small sample sizes N>=200 that consistently vanishes and reverses sign by N(approx):300, uniformly across all tested sparsity levels; (ii) no causal link between data sparsity and this advantage, once tested against a properly matched non-causal control (factorial ANOVA interaction p=0.194); (iii) no genuinely-quantum causal contribution to predictive performance distinguishable from estimator noise at any tested condition, including the resolution of one apparent positive result as a single-seed statistical artifact; (iv) a dissociation between classical-shadow kernel-matrix reconstruction (which fails) and task-level dequantization (which succeeds), indicating that matrix fidelity is an unreliable proxy for classical simulability; (v) a previously undocumented, reproducible failure mode in the PHASE domain-decomposition PINN architecture's boundary-continuity constraint, whose removal yields a solver that is simultaneously faster and more accurate across the full tested regime; and (vi) a robust, reproducible, but ultimately unexplained resonance phenomenon in VQC gradient variance at specific integer causal-edge counts, for which we construct, pre-register, and subsequently falsify a candidate mechanistic explanation via out-of-sample prediction. We argue that the recurring pattern of apparent effects collapsing under targeted multi-seed and out-of-sample scrutiny is itself a methodologically important finding, and we recommend such validation be treated as a mandatory component of QML advantage claims rather than an optional check applied only to surprising results.