Key challenges in quantum machine learning include designing circuits that efficiently process information while achieving high accuracy on today's capacity-limited quantum computers, and researchers are now tackling this problem with new approaches to circuit design. Gurinder Singh of the Center for Computational Life Sciences, along with Thaddeus Pellegrini of IBM Quantum and Kenneth M. Merz, Jr. of the Lerner Research Institute, introduce domain-aware circuits (DAQC) that incorporate knowledge of the structure of an image to improve performance. This new circuit design prioritizes local connections between qubits and reflects the relationships between adjacent pixels in the image. This strategy allows the circuit to process information efficiently without requiring excessive depth or complexity. The team demonstrated that DAQC achieves competitive results compared to classical machine learning models established on standard image datasets and, importantly, currently provides the best performance reported for quantum machine learning on real quantum hardware, representing a significant advance in the field.
Quantum limit learning for image recognition
This study details the development and evaluation of quantum machine learning models for image classification, specifically quantum limit learning machines (QELMs). This research addresses the limitations of classical machine learning and the challenges of training deep quantum neural networks, such as on barren plains. Scientists have used quantum circuits to implement extreme learning machines, potentially enabling faster calculations and the ability to capture complex relationships in data. The quantum circuit acts as a feature map, converting the image into a quantum state representation, and utilizes a kernel method for efficient output weight computation.
The team designed a specific quantum circuit for the feature maps, considered different data encoding strategies, and developed an efficient method to compute the kernel matrix essential for training. Error mitigation techniques such as zero-noise extrapolation and readout error mitigation are incorporated to improve accuracy on noisy quantum hardware. Benchmarks against classical models and other quantum algorithms on datasets such as MNIST, Fashion-MNIST, and MedMNIST demonstrated competitive performance. This work pioneered a method to integrate image region priors, especially the correlations between neighboring pixels, with the constraints of the NISQ hardware. The researchers employed a non-overlapping DCT-style zigzag scan to sequentially encode spatially adjacent pixels into adjacent qubits, establishing a direct correspondence between the image structure and the quantum circuit layout. The circuit operates through interleaved cycles of feature encoding, local entanglement, and trainable one-qubit rotations to prevent long sequences of data- or parameter-only layers and improve gradient flow.
Experiments with MNIST, FashionMNIST, and PneumoniaMNIST demonstrate competitive performance with strong classical baselines such as ResNet-18/50, DenseNet-121, and EfficientNet-B0, and significantly outperform other quantum circuit search frameworks. In this work, we utilized a pure quantum circuit with a linear classical readout, allowing unambiguous attribution of quantum contributions and establishing a robust quantum baseline. Barren plateau analysis demonstrated improved performance compared to standard approaches, validating the effectiveness of domain-aware designs in mitigating common quantum training challenges.
Image encoding using domain-aware quantum circuits
Scientists have developed a domain-aware quantum circuit (DAQC) designed to improve machine learning performance on noisy intermediate-scale quantum (NISQ) hardware, achieving results comparable to strong classical baselines. Our research focuses on exploiting image structure, especially the correlation between neighboring pixels, to guide the encoding process and increase the stability of the optimization. DAQC employs nonoverlapping DCT-style zigzag scans to sequentially encode spatially adjacent pixels into adjacent qubits and adjusts hardware connectivity to minimize long-range interactions. In our experiments, we divided the input image into patches, scanned them with a zigzag scan, created feature vectors representing the image data, and mapped them to quantum states using angular encoding. Entanglement is achieved using a hardware-friendly two-qubit gate applied to qubits hosting neighboring pixels, reducing the exposure of two-qubit errors. The team demonstrated that DAQC achieves competitive performance on image classification tasks using significantly fewer parameters and lower input resolution compared to traditional strong baselines. Specifically, DAQC operates with just 16 logical qubits and hundreds of trainable parameters while maintaining high accuracy and AUC scores on datasets such as MNIST, FashionMNIST, and PneumoniaMNIST. This is achieved by a design that prioritizes locality-preserving information flow, limits the number and depth of two-qubit gates, and reduces the effects of sterile plateaus. Compared to recent quantum circuit search baselines, DAQC achieves significantly higher accuracy, F1 scores, and more balanced sensitivity-specificity, demonstrating the value of domain-aware and hardware-aware circuit design.
