Speaker
Description
Quantum machine learning (QML) models are often designed mainly to maximize accuracy, but accuracy alone is not enough for deployment on near-term quantum hardware. In practice, quantum circuits must be transpiled to match backend constraints such as qubit connectivity and native gate sets. This can increase circuit depth, two-qubit gate count, SWAP overhead, and noise accumulation. In addition, QML models should remain reliable under noisy, adversarial, and distribution-shifted inputs.
This work proposes a deployability-aware neural architecture search (NAS) framework for automatically identifying QML circuits that balance accuracy, adversarial robustness, and hardware resource usage after transpilation. The framework searches over feature maps, ansatz structures, entanglement patterns, circuit depth, qubit layout, and measurement strategies. Each candidate architecture is evaluated using a multi-objective function that includes clean accuracy, adversarial risk, noisy-simulation performance under backend-calibrated error models, transpiled depth, two-qubit gate count, SWAP overhead, estimated hardware error, and shot efficiency. Tools such as Qiskit, Qiskit Aer, PennyLane, Optuna, and adversarial attack methods can support implementation and evaluation.
The outcome is a set of compact and reliable QML architectures that maintain competitive accuracy while reducing hardware overhead and sensitivity to noise. These models should also demonstrate stronger robustness against adversarial perturbations compared with accuracy-only designs.
Overall, this work reframes QML circuit design as a deployability-aware optimization problem. By jointly considering accuracy, hardware constraints, noise resilience, and adversarial robustness, the proposed framework helps bridge the gap between QML research and practical quantum-ready applications.
| Topic | Quantum-enhanced robustness in machine learning models |
|---|---|
| Additional Topic | Quantum variational circuits and their applications |