Quantum Networks: Accuracy & Resource Optimisation

Kooshan Maleki and colleagues at Amirkabir University of Technology present QNAS, a neural architecture search framework that simultaneously optimises for accuracy, efficiency, and the challenges posed by limited qubit availability. QNAS addresses a key need in the field by moving beyond accuracy-focused searches to incorporate practical considerations like evaluation time and circuit complexity, ultimately guiding the development of hybrid quantum-classical neural networks suitable for deployment on existing devices. Across multiple benchmark datasets, MNIST, Fashion-MNIST, and Iris, QNAS identified architectures achieving high accuracy with remarkably compact circuits, such as 97.16% accuracy on MNIST using only eight qubits and two layers.

Optimised circuit division enables high accuracy with minimal qubits

A quantum neural network attained 97.16% test accuracy on the MNIST dataset utilising just eight qubits and two layers, representing a significant improvement over prior methods. These previous methods were hampered by the practicalities of limited quantum hardware, meaning they struggled with the constraints of available resources. Building quantum networks exceeding this performance with such compact circuits was previously impossible due to the substantial computational cost of ‘circuit cutting’, a process of dividing complex quantum operations into smaller, manageable segments for near-term devices.

Kooshan Maleki and colleagues have developed a new framework, QNAS, directly addressing this challenge by simultaneously optimising for accuracy, runtime, and the number of necessary circuit divisions. This holistic approach unlocks designs previously considered too expensive to execute. On the more complex Fashion-MNIST benchmark, the network achieved 87.38% accuracy with five qubits and two layers, while a remarkably concise four-qubit, two-layer circuit delivered 100% validation accuracy on the Iris dataset. Analysis revealed that angle-y embedding and sparse entangling patterns proved most effective for image datasets, whereas amplitude embedding excelled with tabular data; these insights surfaced automatically during the search process.

Demonstrating balanced accuracy and resource efficiency across established image and classification

Practical quantum neural network design requires a delicate balance between accuracy and the limitations of current quantum hardware. QNAS offers a promising route to achieving this, automatically searching for designs that minimise the computational burden of ‘circuit cutting’, dividing complex operations into smaller segments when qubit numbers are insufficient. The work acknowledges a key constraint: performance was only demonstrated across three datasets, namely MNIST, Fashion-MNIST, and Iris.

Successful navigation of the trade-offs between performance and resource use on these diverse problems suggests broader applicability. A ‘SuperCircuit’, a configurable quantum neural network, is trained and optimised using NSGA-II, allowing for efficient design exploration. QNAS, a framework automating the design of hybrid quantum-classical neural networks, introduces a new approach, moving beyond solely prioritising accuracy to encompass practical hardware limitations. Simultaneously optimising validation error, runtime, and the computational cost of ‘circuit cutting’, QNAS identifies designs balancing performance with resource constraints. The system employs a ‘SuperCircuit’ and NSGA-II, a multi-objective optimisation technique, to efficiently explore architectural possibilities and reveal trade-offs between these competing factors, reducing average cutting overhead by 62.8% compared to methods that ignore this factor.

The research successfully demonstrated a method for designing quantum neural networks that balances accuracy with the practical limitations of current quantum hardware. QNAS, a new neural architecture search framework, optimises for validation error, runtime and the computational cost of dividing circuits when using a limited number of qubits. Across benchmarks including MNIST and Iris, the system identified network architectures achieving up to 97.16% test accuracy using just eight qubits and two layers. The authors state that this approach provides clear trade-offs between performance and efficiency, potentially enabling more deployable hybrid quantum-classical neural networks.