Domain-Aware Quantum Circuit for QML

TL;DR

This work introduces the Domain-Aware Quantum Circuit (DAQC), a hardware- and domain-guided quantum architecture for image classification that encodes neighboring pixels with zigzag, locality-preserving patterns and interleaved encode–entangle–train cycles to balance expressivity and noise. Through 16-qubit experiments with 256 embeddings and 64 entanglers, the authors demonstrate Haar-like expressibility and strong trainability, while achieving competitive or superior performance to classical baselines on binary/unbalanced tasks and outperforming state-of-the-art quantum circuit search methods on MNIST-family datasets, including real hardware with error mitigation. Extensive ablations, barren-plateau analyses, and hardware error profiling on IBM Kingston support the design choices, showing that local cost functions mitigate gradient decay and enable stable optimization within NISQ budgets. The results suggest a practical, domain-aware path toward quantum advantage in QML, highlighting the value of structured encoding, hardware-aware entanglement, and robust mitigation for scalable quantum image classification. Future work points to scaling through tensor-network simulations and cautious depth/feature expansion as hardware capabilities improve, aiming to close the gap with large classical baselines while maintaining efficiency gains from quantum resources.

Abstract

Designing parameterized quantum circuits (PQCs) that are expressive, trainable, and robust to hardware noise is a central challenge for quantum machine learning (QML) on noisy intermediate-scale quantum (NISQ) devices. We present a Domain-Aware Quantum Circuit (DAQC) that leverages image priors to guide locality-preserving encoding and entanglement via non-overlapping DCT-style zigzag windows. The design employs interleaved encode-entangle-train cycles, where entanglement is applied among qubits hosting neighboring pixels, aligned to device connectivity. This staged, locality-preserving information flow expands the effective receptive field without deep global mixing, enabling efficient use of limited depth and qubits. The design concentrates representational capacity on short-range correlations, reduces long-range two-qubit operations, and encourages stable optimization, thereby mitigating depth-induced and globally entangled barren-plateau effects. We evaluate DAQC on MNIST, FashionMNIST, and PneumoniaMNIST datasets. On quantum hardware, DAQC achieves performance competitive with strong classical baselines (e.g., ResNet-18/50, DenseNet-121, EfficientNet-B0) and substantially outperforming Quantum Circuit Search (QCS) baselines. To the best of our knowledge, DAQC, which uses a quantum feature extractor with only a linear classical readout (no deep classical backbone), currently achieves the best reported performance on real quantum hardware for QML-based image classification tasks. Code and pretrained models are available at: https://github.com/gurinder-hub/DAQC.

Figures (7)

• Figure 1: Domain-aware quantum circuit for QML. Here $p\times q$ is the size of non-overlapping sliding window for the zigzag scan, $u\times v$ denotes the total number of non-overlapping sliding windows. In the trainable layers, $t=2v-1$, $t'=2v$, $t"=2(u\times v)-1$, and $t"'=2(u\times v)$. Encoding and trainable layers gates are selected based on uniform random sampling from $RXYZ$ design space.
• Figure 2: Expressibility and entangling capability of the 16-qubit DAQC. (a) Expressibility measured via $D_{KL}(P_{PQC}\Vert P_{Haar})$ (lower is better), (b) Mean Meyer--Wallach global entanglement measure $Q$ (higher is better), and (c) Fidelity histogram for the depth setting $(N_{embed}, N_{train}, N_{ECR}) \in (256,512,64)$ overlaid with the analytic Haar fidelity density.
• Figure 3: Specifications of DAQC before and after transpilation.
• Figure 4: Some of the example layouts selected by the Qiskit transpiler for the DAQC on the ibm_kingston hardware, red: physical qubits, blue: ancilla qubits.
• Figure 5: Hardware characterization data reported from the IBM Kingston quantum processor averaged across time on each error channel during the time-frame of September 01-23, 2025.