Kuan-Cheng Chen, Shang Yu, Chen-Yu Liu, Samuel Yen-Chi Chen, Huan-Hsin Tseng, Yen Jui Chang, Wei-Hao Huang, Felix Burt, Esperanza Cuenca Gomez, Zohim Chandani, William Clements, Ian Walmsley, Kin K. Leung
quant-phcs.ETcs.LGSubmitted Mar 16, 2026• Announced Mar 17, 2026• 0 citations
Photonic quantum processors naturally produce intrinsically stochastic measurement outcomes, offering a hardware-native source of structured randomness that can be exploited during machine-learning training. Here we introduce Photonic Quantum-Enhanced Knowledge Distillation (PQKD), a hybrid quantum photonic--classical framework in which a programmable photonic circuit generates a compact conditioning signal that constrains and guides a parameter-efficient student network during distillation from a high-capacity teacher. PQKD replaces fully trainable convolutional kernels with dictionary convolutions: each layer learns only a small set of shared spatial basis filters, while sample-dependent channel-mixing weights are derived from shot-limited photonic features and mapped through a fixed linear transform. Training alternates between standard gradient-based optimisation of the student and sampling-robust, gradient-free updates of photonic parameters, avoiding differentiation through photonic hardware. Across MNIST, Fashion-MNIST and CIFAR-10, PQKD traces a controllable compression--accuracy frontier, remaining close to teacher performance on simpler benchmarks under aggressive convolutional compression. Performance degrades predictably with finite sampling, consistent with shot-noise scaling, and exponential moving-average feature smoothing suppresses high-frequency shot-noise fluctuations, extending the practical operating regime at moderate shot budgets.
Photonic quantum processors naturally produce intrinsically stochastic measurement outcomes, offering a hardware-native source of structured randomness that can be exploited during machine-learning training. Here we introduce Photonic Quantum-Enhanced Knowledge Distillation (PQKD), a hybrid quantum photonic--classical framework in which a programmable photonic circuit generates a compact conditioning signal that constrains and guides a parameter-efficient student network during distillation from a high-capacity teacher. PQKD replaces fully trainable convolutional kernels with dictionary convolutions: each layer learns only a small set of shared spatial basis filters, while sample-dependent channel-mixing weights are derived from shot-limited photonic features and mapped through a fixed linear transform. Training alternates between standard gradient-based optimisation of the student and sampling-robust, gradient-free updates of photonic parameters, avoiding differentiation through photonic hardware. Across MNIST, Fashion-MNIST and CIFAR-10, PQKD traces a controllable compression--accuracy frontier, remaining close to teacher performance on simpler benchmarks under aggressive convolutional compression. Performance degrades predictably with finite sampling, consistent with shot-noise scaling, and exponential moving-average feature smoothing suppresses high-frequency shot-noise fluctuations, extending the practical operating regime at moderate shot budgets.
Abstract Quantum computing has emerged as a promising paradigm for addressing computational tasks intractable for classical systems, leveraging quantum mechanical principles such as superposition and entanglement to efficiently explore high-dimensional solution spaces. In recent years, hybrid quantum-classical approaches have gained increasing attention as a means to exploit the representational power of quantum systems while preserving the practicality of classical machine learning frameworks. This paper presents a study on quantum-enhanced model compression via feature-based knowledge distillation, in which a large classical neural network (the teacher) transfers knowledge to a substantially smaller hybrid student model. The proposed hybrid quantum-classical student architecture incorporates a parameterized quantum circuit (PQC) implemented in Qiskit, employing a linear $$\varvec{R_y}$$ angle-encoding feature map followed by an EfficientSU2 variational ansatz, with features bounded through a $$\varvec{\tanh (\cdot )\times \frac{\pi }{2}}$$ scaling to ensure stable gradient propagation. Gradient computation is performed via Reverse-Mode (adjoint) differentiation through the StatevectorEstimator primitive, which provides exact, shot-noise-free gradients at the cost of a single forward–backward pass per optimization step. The hybrid student is compared against a parameter-matched purely classical student under strict 1:1 parameter parity across a six-task binary benchmark spanning MNIST and Fashion-MNIST, as well as a four-class multi-class benchmark, totaling eleven classification experiments. A hardware validation experiment is additionally conducted on the IQM Garnet 20-qubit quantum processor using the Parameter-Shift Rule. The results indicate that the hybrid student outperforms its classical counterpart in five of six binary tasks (83%) and in the two hardest multi-class configurations, with advantages most pronounced in high-confusion regimes where compact classical classifiers approach the limits of their expressive capacity. On real quantum hardware, accuracy decreases from 99.04% in noiseless simulation to 80.50%, reflecting the impact of gate noise, readout errors, and transpilation overhead. While the hybrid model demonstrates improved parameter efficiency and expressive capacity, it incurs higher computational overhead and exhibits greater epoch-to-epoch training variability relative to classical baselines. We discuss practical limitations related to noise, scalability, and current hardware constraints, and identify directions for future work including hardware-efficient gradient estimation and advanced error mitigation protocols. Overall, our findings suggest that variational quantum circuits can serve as compact yet expressive components within neural compression schemes, with their representational advantage most effectively realized in binary and high-ambiguity classification scenarios.
Knowledge distillation (KD) is a widely adopted technique for compressing large models into smaller, more efficient student models that can be deployed on devices with limited computational resources. Among various KD methods, Relational Knowledge Distillation (RKD) improves student performance by aligning relational structures in the feature space, such as pairwise distances and angles. In this work, we propose Quantum Relational Knowledge Distillation (QRKD), which extends RKD by incorporating quantum relational information. Specifically, we map classical features into a Hilbert space, interpret them as quantum states, and compute quantum kernel values to capture richer inter-sample relationships. These quantum-informed relations are then used to guide the distillation process. We evaluate QRKD on both vision and language tasks, including CNNs on MNIST and CIFAR-10, and GPT-2 on WikiText-2, Penn Treebank, and IMDB. Across all benchmarks, QRKD consistently improves student model performance compared to classical RKD. Importantly, both teacher and student models remain classical and deployable on standard hardware, with quantum computation required only during training. This work presents the first demonstration of quantum-enhanced knowledge distillation in a fully classical deployment setting.
The rapid advancement of machine learning (ML) and the growing need for computational power have led to the exploration of quantum computing, which offers significant potential for faster complex calculations. However, Quantum Machine Learning (QML) faces challenges due to the limited number of qubits and noise of quantum circuits, particularly with Noisy Intermediate-Scale Quantum (NISQ) devices. These challenges severely limit the current capacity to train accurate and stable Quantum Machine Learning Models. In this paper, we propose a novel framework for QML that employs the knowledge distillation method to harness the power of well-trained classical machine learning (CML) models and enhance the training performance of QML models. In this framework, we utilize the well-trained CML as a teacher model to assist the training of the student QML model using the knowledge distillation method. By distilling knowledge from the robust CML model, our framework can potentially address the problem of the barren plateau which hinders effective model training. Knowledge distillation is well suited for this framework through the transfer of knowledge without parameter sharing. Through empirical tests, our framework has demonstrated not only an increase in the accuracy of QML models but also a notable improvement in training stability.
Quantum neural networks (QNNs), harnessing superposition and entanglement, have shown potential to surpass classical methods in complex learning tasks but remain limited by hardware constraints and noisy conditions. In this work, we present a novel framework for transferring knowledge from classical convolutional neural networks (CNNs) to QNNs via knowledge distillation, thereby reducing the need for resource intensive quantum training and error mitigation. We conduct extensive experiments using two parameterized quantum circuits (PQCs) with 4 and 8 qubits on MNIST, Fashion MNIST, and CIFAR10 datasets. The approach demonstrates consistent accuracy improvements attributed to distilled knowledge from larger classical networks. Through ablation studies, we systematically compare the effect of state of the art dimensionality reduction techniques fully connected layers, center cropping, principal component analysis, and pooling to compress high-dimensional image data prior to quantum encoding. Our findings reveal that fully connected layers retain the most salient features for QNN inference, thereby surpassing other down sampling approaches. Additionally, we examine state of the art data encoding methods (amplitude, angle, and qubit encoding) and identify amplitude encoding as the optimal strategy, yielding superior accuracy across all tested datasets and qubit configurations. Through computational analyses, we show that our distilled 4-qubit and 8-qubit QNNs achieve competitive performance while utilizing significantly fewer parameters than their classical counterparts. Our results establish a promising paradigm for bridging classical deep learning and emerging quantum computing, paving the way for more powerful, resource conscious models in quantum machine intelligence.
Quantum machine learning offers a practical opportunity to leverage near-term quantum devices without requiring quantum inference at deployment. We propose a quantum-classical hybrid knowledge distillation framework in which a variational quantum circuit, equipped with a Quantum Fourier Transform-inspired positional encoding, acts as a soft-label teacher for a compact classical student. Rather than serving as a standalone classifier, the quantum model is used offline to generate structured probability distributions that encode global spectral information and uncertainty. These quantum-derived soft labels are distilled into a lightweight classical network using a hybrid loss that combines hard and soft supervision. Experiments on MNIST demonstrate that students trained with quantum soft labels exhibit consistent and statistically meaningful robustness improvements to Gaussian noise and in-plane rotations compared to classical distillation baselines, while maintaining strong clean accuracy and calibration. These results highlight a distinct role for NISQ-era quantum models as supervisory signal generators, enabling quantum-informed learning within fully classical deployment pipelines.
While hybrid quantum–classical architectures and knowledge distillation have each been explored independently in medical imaging and model compression, no prior work has unified these domains within a single framework. This study addresses this gap by investigating the integration of quantum–classical neural networks with knowledge distillation for medical image analysis, with a focus on circuit behavior, parameter efficiency, and diagnostic performance. The framework (QuantumMedKD) employs classical CNN architectures (ResNet-50, EfficientNet-B0, Xception) as teacher models, transferring learned representations to parameter-efficient quantum student networks via parameterized quantum circuits across qubit configurations (3-8), demonstrated through pneumonia detection from pediatric chest radiographs. Experimental validation reveals remarkable parameter efficiency, requiring only 24-36 trainable parameters compared to millions for classical counterparts, achieving compression ratios exceeding 1 0 5 while maintaining competitive diagnostic performance. Optimal configurations achieve 84.00% accuracy (EfficientNet-B0, 6-qubit) and 73.33% (Xception, 4-qubit), with knowledge distillation providing statistically significant improvements ( p -vallues < 0.001, Cohen’s d > 2.0). Medical evaluation confirms diagnostic capability with sensitivity 81.4%, specificity 73.9%, and AUC 0.89, establishing quantum–classical knowledge transfer viability for resource-constrained healthcare deployment. QuantumMedKD reveals essential principles and proof-of-concept for quantum-enhanced healthcare AI systems within and beyond the noisy intermediate-scale quantum (NISQ) era for high efficiency and deployability, paving the way for future advancements in the field.