Quantum-enhanced Learning Framework Improves Accuracy from 0.31 to 0.87, Enabling Practical Hybrid Workflows

The increasing potential of quantum computing presents a significant challenge for those without specialist knowledge to integrate these methods into existing workflows. Silvie Illésová from Gran Sasso Science Institute, Tomáš Bezděk, and colleagues address this problem by developing a practical framework that guides users from standard classical machine learning to hybrid quantum-enhanced approaches. Their method begins with a conventional self-training model, introduces a minimal quantum component, and crucially, employs diagnostic feedback to optimise the hybrid architecture. Experiments demonstrate that this approach dramatically improves performance, boosting accuracy from 0. 31 to 0. 87 on a standard dataset, and suggests that even small quantum enhancements, when properly guided, can significantly improve machine learning capabilities for practitioners without deep expertise in quantum computing.

The study begins with a classical self-training model, specifically Partial Least Squares Regression applied to the Iris dataset, initially assigning arbitrary labels to the 150 samples with four features. This model undergoes iterative refinement, updating labels based on cross-validation accuracy, with a maximum of 20 iterations, to establish a baseline for comparison. Next, the team engineered a minimal hybrid model, termed Quantum-FAST, mirroring the classical structure but incorporating a quantum component.

Input data undergoes Principal Component Analysis, reducing the four features to a compact representation suitable for quantum encoding. This reduced data then feeds into a two-qubit EstimatorQNN, a quantum neural network, followed by a small classical neural network that maps the quantum outputs to label predictions. The design prioritizes minimal quantum resource usage while still demonstrating the potential for improvement. To assess and refine this hybrid model, the study employs QMetric, a toolkit used to analyze training behavior, feature-space properties, and quantum circuit characteristics.

This analysis provides targeted feedback, guiding modifications to the minimal hybrid model to improve its training characteristics and address identified weaknesses. The researchers demonstrate this workflow using the Iris dataset, implemented in Qiskit, Qiskit Machine Learning, PyTorch, and scikit-learn, comparing the performance of the classical model, the initial minimal hybrid model, and the refined hybrid model. The resulting improvements in accuracy, increasing from 0. 31 in the classical approach to 0. 87 in the quantum approach, demonstrate the potential of this framework to enhance class separation and representation capacity in hybrid learning.

Hybrid Learning with Quantum Diagnostics Improves Accuracy

This research successfully demonstrates a practical pathway for individuals with classical machine learning expertise to incorporate quantum computational elements into their models, even without prior quantum knowledge. The team developed a three-stage framework, beginning with a classical self-training model, then introducing a minimal hybrid quantum-classical variant, and finally applying diagnostic feedback using a tool called QMetric to refine the hybrid architecture. Experiments on the Iris dataset reveal that this approach improved accuracy from 0. 31 to 0. 87, demonstrating that even small quantum contributions, when guided by appropriate diagnostics, can significantly enhance learning and improve the representation of data.

Further analysis using QMetric showed that the refined hybrid model not only achieved greater accuracy but also exhibited stronger decisiveness in label assignment and more global entanglement within the quantum circuit. Importantly, these improvements in expressiveness were achieved without compromising the stability of the training process or encountering issues like barren plateaus. The team proposes an iterative refinement loop, where diagnostic feedback guides the progressive development of a classical model towards an increasingly effective hybrid design, moving beyond trial-and-error approaches to model architecture. The authors acknowledge that the current work focuses on a specific dataset and model configuration, and future research will likely explore the generalizability of this framework to more complex datasets and machine learning tasks.