Quantum architecture search (QAS) has emerged as a prominent paradigm for designing quantum circuits in recent years. To accelerate the search for optimal circuit structures, circuit performance is often estimated without full training. Neural predictors have shown great potential in providing effective performance estimations by leveraging the power of deep learning. However, training such predictors requires a labeled dataset of circuit-performance pairs, where circuit performance must still be obtained via costly circuit training. A large training set undermines the original intent of using predictors to reduce the cost of obtaining circuit performance, while a small training set may hinder the predictor's generalization ability. To address this dilemma, we propose two data-efficient, predictor-based QAS algorithms that operate in a semi-supervised learning fashion to exploit the latent information of unlabeled circuits. The first algorithm adopts a teacher-student framework, where a teacher model guides the training of a student predictor using unlabeled circuits. The teacher produces smoother and more stable targets, encouraging the student to generate steady predictions under perturbations, thereby enhancing robustness. The second algorithm employs an uncertainty-aware circuit selection strategy that identifies unlabeled circuits with low-uncertainty predictions, which are then incorporated into the training set to effectively expand the available labeled data. Numerical results demonstrate that these two algorithms improve the predictor's generalization ability and facilitate the discovery of higher-performing quantum circuits, thereby enhancing the efficiency of predictor-based QAS in variational quantum algorithms.

Close

PrevNext

View figure in article

• Noise
• Optimization problems
• Quantum algorithms
• Quantum circuits
• Quantum computing models
• Artificial neural networks
• Machine learning