Domain-aware Quantum Circuit Achieves Efficient Machine Learning on NISQ Devices

A significant challenge in quantum machine learning involves designing circuits that effectively process information on current, limited-capacity quantum computers, while also achieving high accuracy, and researchers are now addressing this problem with a novel approach to circuit design. Gurinder Singh from the Center for Computational Life Sciences, along with Thaddeus Pellegrini from IBM Quantum, and Kenneth M. Merz, Jr from the Lerner Research Institute, present a Domain-Aware Circuit (DAQC) that incorporates knowledge about the structure of images to improve performance. This new circuit design prioritises local connections between quantum bits, mirroring the relationships between neighbouring pixels in an image, and this strategy allows the circuit to efficiently process information without requiring excessive depth or complexity. The team demonstrates that DAQC achieves competitive results compared to established classical machine learning models on standard image datasets, and importantly, currently delivers the best reported performance for quantum machine learning on real quantum hardware, representing a substantial step forward in the field.

Quantum Extreme Learning for Image Recognition

This research details the development and evaluation of quantum machine learning models for image classification, specifically Quantum Extreme Learning Machines (QELMs). The study addresses limitations of classical machine learning and the challenges of training deep quantum neural networks, such as barren plateaus. Scientists implemented Extreme Learning Machines using quantum circuits, allowing for potentially faster computation and the ability to capture complex relationships in data. The quantum circuits function as feature maps, transforming images into quantum state representations and utilizing kernel methods for efficient output weight computation.

The team designed specific quantum circuits for feature maps, explored various data encoding strategies, and developed efficient methods for calculating the kernel matrix crucial for training. Error mitigation techniques, including zero-noise extrapolation and readout error mitigation, were incorporated to improve accuracy on noisy quantum hardware. Benchmarking against classical models and other quantum algorithms on datasets like MNIST, Fashion-MNIST, and MedMNIST demonstrated competitive performance. The research pioneered a method that integrates image-domain priors, specifically correlations between neighboring pixels, with the constraints of NISQ hardware. Researchers employed a non-overlapping, DCT-style zigzag scan to sequentially encode spatially neighboring pixels onto adjacent qubits, establishing a direct correspondence between image structure and quantum circuit layout. The circuit operates through interleaved cycles of feature encoding, local entanglement, and trainable one-qubit rotations, preventing long sequences of data or parameter-only layers to improve gradient flow.

Experiments on MNIST, FashionMNIST, and PneumoniaMNIST demonstrated performance competitive with strong classical baselines like ResNet-18/50, DenseNet-121, and EfficientNet-B0, and substantially outperformed other quantum circuit search frameworks. The study utilized a pure quantum circuit with a linear classical readout, enabling a clear attribution of quantum contributions and establishing a robust quantum baseline. Barren plateau analysis demonstrated improved performance compared to standard approaches, validating the effectiveness of the domain-aware design in mitigating common quantum training challenges.

Image Encoding with Domain-Aware Quantum Circuits

Scientists developed a domain-aware quantum circuit (DAQC) designed to improve machine learning performance on noisy intermediate-scale quantum (NISQ) hardware, achieving results competitive with strong classical baselines. The research concentrates on leveraging image structure, specifically correlations between neighboring pixels, to guide the encoding process and enhance optimization stability. The DAQC employs a non-overlapping, DCT-style zigzag scan to sequentially encode spatially adjacent pixels onto adjacent qubits, aligning with hardware connectivity and minimizing long-range interactions. Experiments involved partitioning input images into patches and traversing them with the zigzag scan, creating a feature vector representing the image data, which was then mapped to quantum states using angle encoding. Entanglement is achieved using hardware-friendly two-qubit gates applied to qubits hosting neighboring pixels, reducing two-qubit error exposure. The team demonstrates that DAQC achieves competitive performance on image classification tasks, using significantly fewer parameters and reduced input resolution compared to strong classical baselines. Specifically, DAQC maintains high accuracy and AUC scores on datasets including MNIST, FashionMNIST, and PneumoniaMNIST, while operating with only 16 logical qubits and a few hundred trainable parameters. This is accomplished through a design that prioritizes locality-preserving information flow, limiting two-qubit gate counts and depth, and mitigating the effects of barren plateaus. Against recent quantum circuit search baselines, DAQC delivers substantially higher accuracy, F1-score, and more balanced sensitivity-specificity, demonstrating the value of domain-aware and hardware-aligned circuit design.