CRII: RI: Improving The Robustness of Neural Networks via Bayesian Inference: from Classical to Quantum Models

The rapid evolution of artificial intelligence (AI) techniques has revolutionized a wide range of applications from classical models to quantum circuits nowadays. Despite their transformative impact, AI models may be exposed to vulnerabilities under several scenarios, such as adversarial attacks, open-set environments, and gradient issues in optimization. These issues can not only degrade model performance but also lead to unpredictable outcomes, undermining the model's reliability. To address the above issues, this project aims to improve the robustness of AI models, which are critical to national and social security. Furthermore, this award will support the cross-disciplinary education between AI and quantum information through undergraduate and graduate-level courses at Boise State University. To achieve the above goals, this project systematically explores three thrusts to enhance the robustness of AI models, particularly quantum neural networks (QNNs), a type of AI model that harnesses unique properties of quantum circuits to process data. The first thrust focuses on improving the adversarial robustness of QNNs using Bayesian label transition. The second thrust tackles the open-set problems under perturbation circumstances by proposing novel quantum models. The third thrust advances the training robustness of QNNs from the optimization perspective. These three thrusts will be evaluated on public datasets, bridging theoretical advancements with practical implementations to ensure robust quantum AI applications. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.