Quantum Machine Learning Bypasses Complex Coding with New 12-Gate Circuits

Researchers at Lahore University of Management, led by Adil Mubashir Chaudhry, have detailed a novel methodology for designing quantum feature maps specifically tailored for implementation on IBM quantum processors.

The team’s approach circumvents the need for complex circuit modifications, employing a hardware-aware Neural Architecture Search (NAS) to automatically generate circuits utilising exclusively the native gate set of IBM’s Torino processor. This resulted in a 12-gate circuit, leveraging 10 qubits, achieving 91.23% accuracy on the UCI Breast Cancer Wisconsin dataset. These findings signify a substantial improvement over conventionally designed quantum feature maps and represent a crucial step towards realising practical quantum kernel methods on currently available near-term quantum hardware, effectively eliminating errors inherent in universal gate compilation procedures. Automated quantum circuit design exceeds hand-crafted feature maps on NISQ hardware A newly developed hardware-aware Neural Architecture Search (NAS) attained 91.23% accuracy, demonstrably exceeding the performance of hand-crafted quantum feature maps by a margin of 27 percentage points. Prior to this advancement, such hand-crafted maps were constrained to a maximum accuracy of 64%. This progress facilitates the effective operation of quantum kernel methods on noisy intermediate-scale quantum (NISQ) devices, addressing limitations imposed by error-prone circuit translation, commonly known as transpilation. Transpilation is the process of converting a quantum algorithm designed for an idealised quantum computer into a sequence of operations executable on specific hardware, and it inevitably introduces errors due to the limitations of physical qubits and gate implementations. The automated design process successfully discovered a 12-gate circuit utilising only IBM’s native gates, specifically, ECR, RZ, SX, and X, demonstrating performance comparable to unconstrained searches while crucially avoiding decomposition errors. Prior to this advancement, hand-crafted maps were constrained to a maximum accuracy of 64%. This progress facilitates the effective operation of quantum kernel methods on noisy intermediate-scale quantum (NISQ) devices, addressing limitations imposed by error-prone circuit translation, commonly known as transpilation.

The team designed circuits that actively avoid decomposition errors. Decomposition errors arise when complex quantum gates are broken down into a series of native gates, increasing circuit depth and susceptibility to noise. Relaxing a single architectural constraint within the NAS, namely the fixed placement of RZ rotations, further improved accuracy to 94.73%. The utilisation of native gates guarantees hardware compatibility and allows for immediate execution without the need for modification, thereby streamlining the development process and paving the way for practical quantum machine learning applications. When benchmarked against several alternative methods, including classical Support Vector Machines achieving 93% accuracy, hand-crafted quantum feature maps limited to 64%, and unconstrained all-gates NAS approaches, the hardware-aware Neural Architecture Search (NAS) consistently outperformed manual designs. However, it is important to note that evaluation was conducted on a reduced six-qubit representation of the dataset, and scaling this approach to larger, more complex problems remains a considerable challenge requiring further investigation into resource allocation and algorithmic optimisation. The genetic algorithm employed within the NAS framework operates by iteratively evolving a population of quantum circuit architectures. Each circuit is evaluated based on its performance on the target dataset, and the fittest circuits select for reproduction and mutation, creating new generations of circuits with potentially improved performance. The fitness function incorporates both accuracy and circuit complexity, encouraging the discovery of compact and efficient designs. The constraint of using only native gates enforces throughout the evolutionary process, ensuring hardware compatibility. This differs significantly from traditional NAS approaches which often explore a vast search space of arbitrary quantum circuits, requiring extensive transpilation and potentially leading to suboptimal performance on real hardware. Direct hardware compatibility streamlines quantum machine learning for cancer diagnosis Automated circuit design offers a promising pathway towards unlocking the full potential of near-term quantum computers, effectively sidestepping the limitations inherent in manually crafted algorithms. This work successfully constructs quantum circuits directly compatible with IBM’s hardware, representing a significant advancement towards practical application in areas such as medical diagnostics. However, the current evaluation remains confined to the UCI Breast Cancer Wisconsin dataset, necessitating further research to assess the approach’s generalisability and effectiveness when applied to other, more complex and diverse problems. The UCI dataset, while valuable for initial validation, represents a relatively well-structured problem, and performance on more challenging datasets may differ. Initial tests focused on identifying cancerous cells within the dataset, but this specific application should not detract from the broader significance of the advance in quantum circuit design. The hardware-aware approach successfully evolved circuits achieving high accuracy on a breast cancer dataset, utilising only native gates and thereby minimising the accumulation of errors. The resulting 12-gate circuit establishes a new benchmark for performance, avoiding reliance on error-prone universal gate compilation and circumventing the need for transpilation, a process that introduces errors when adapting algorithms for use on quantum processors. Quantum kernel methods, which underpin this approach, map data into a high-dimensional quantum feature space, allowing for the identification of complex patterns and relationships. By optimising the quantum feature map, the researchers have effectively enhanced the ability of the quantum computer to discriminate between different classes of data. The implications of this research extend beyond cancer diagnosis. The ability to automatically design hardware-aware quantum circuits could accelerate the development of quantum machine learning algorithms for a wide range of applications, including materials discovery, financial modelling, and drug design. Future work will focus on scaling the NAS approach to larger datasets and more complex quantum processors, as well as exploring the use of different evolutionary algorithms and fitness functions to further optimise circuit performance. Investigating the robustness of these circuits to variations in hardware parameters and noise levels will also be crucial for ensuring their reliability in real-world applications. The researchers successfully designed a 12-gate quantum circuit for a Support Vector Machine, achieving 91.23% accuracy on the UCI Breast Cancer Wisconsin dataset. This is significant because the circuit used only the native gates of IBM quantum processors, eliminating errors from the usual process of adapting algorithms for specific hardware. This hardware-aware approach outperformed hand-crafted quantum feature maps, reaching accuracy levels approaching a classical baseline. The authors intend to scale this method to larger datasets and more complex quantum processors to further refine performance. 👉 More information 🗞 Hardware-Aware Quantum Support Vector Machines 🧠 ArXiv: https://arxiv.org/abs/2604.07856 Tags: