Quantum Kernel Design Improves Accuracy on Near-Term Hardware.

Quantum machine learning (QML) offers potential advantages over classical algorithms for certain computational tasks, yet realising this potential requires careful consideration of the limitations imposed by current quantum hardware. A key component of QML is the quantum kernel, a mathematical function that maps data into a quantum state allowing for pattern recognition. Designing effective kernels that account for both the specific problem and the characteristics of noisy intermediate-scale quantum (NISQ) devices presents a significant challenge. Researchers at Southeast University, Nanjing Tech University, and Purple Mountain Lab, including Yuxiang Liu, Fanxu Meng, Lu Wang, Yi Hu, Sixuan Li, Zaichen Zhang, and Xutao Yu, address this issue in their work, “Hardware-Aware Quantum Kernel Design Based on Graph Neural Networks”. They propose a framework, HaQGNN, which integrates device topology, noise characteristics, and graph neural networks to efficiently evaluate and select quantum circuits suitable for kernel-based machine learning tasks, demonstrating improved classification accuracy on benchmark datasets.

The development of practical quantum machine learning faces a substantial obstacle in designing effective quantum kernels for near-term, noisy intermediate-scale quantum (NISQ) devices. These devices, while offering potential computational advantages, are susceptible to errors and possess a limited number of qubits, demanding innovative approaches to kernel design. A new framework, HaQGNN, directly addresses this challenge by integrating device topology, noise characteristics, and graph neural networks (GNNs) to evaluate and select quantum circuits appropriate for a specific task, offering a pathway toward more robust and efficient quantum algorithms. HaQGNN predicts surrogate metrics, specifically fidelity—a measure of how accurately a quantum computation performs—and kernel performance, enabling efficient screening of circuits at scale, a crucial step for practical implementation and resource optimisation.

HaQGNN distinguishes itself through its innovative use of GNNs to model the complex relationships between circuit structure, hardware constraints, and resulting kernel performance. GNNs, a type of artificial intelligence adept at learning patterns from graph-structured data, allow HaQGNN to capture the intricate interplay between these factors, enabling it to identify circuits that maximise performance despite the limitations of current quantum hardware. By learning from these relationships, the framework effectively navigates the trade-offs between circuit complexity, noise resilience, and computational accuracy, leading to more practical and reliable quantum machine learning solutions. Feature selection further enhances the framework’s capabilities, improving compatibility with systems possessing a limited number of qubits and mitigating potential kernel degradation caused by noise, ensuring optimal performance even on resource-constrained devices.

Extensive experimentation across three benchmark datasets—Credit Card fraud detection, the handwritten digit dataset MNIST-5, and the fashion item dataset FMNIST-4—demonstrates HaQGNN’s superior performance and validates its effectiveness in optimising kernel design for practical applications. The framework consistently outperforms existing kernel baselines in terms of classification accuracy, showcasing its ability to extract meaningful patterns from data and make accurate predictions even in the presence of noise and limited resources. These results highlight the potential of hardware-aware strategies for advancing the field of quantum machine learning, moving beyond purely theoretical considerations and paving the way for real-world applications.

Future work will focus on extending HaQGNN’s capabilities to more complex datasets and quantum hardware architectures, pushing the boundaries of what is possible with quantum machine learning. Investigating the use of reinforcement learning techniques to further refine the circuit selection process represents a promising avenue for exploration, potentially leading to even more efficient and robust quantum algorithms. Additionally, exploring the framework’s applicability to other quantum machine learning algorithms beyond kernel methods warrants further investigation, broadening its impact and potential applications. A crucial next step involves validating HaQGNN’s performance on actual quantum hardware, moving beyond simulations to demonstrate its real-world viability and unlock its full potential.