受限玻尔兹曼机（RBM）是深度学习领域具备严格理论分析的经典算法，其高效性与可靠性分别获得了学术界与应用领域的广泛关注，而两者的联合研究还处于起步阶段。本项目将首先从算法与理论两个维度分别研究：基于纠缠熵理论的高效RBM，主要体现在纠缠熵与网络高效性的理论联系与设计指导；基于对抗攻击的可靠RBM，主要体现在RBM的对抗样本设计与对抗训练方法。在此基础上，项目将挖掘RBM高效性与可靠性之间的相互关联，探索同时具备高效可靠的RBM网络架构设计与训练方法。项目由RBM与张量重正化群中高效性与可靠性的不同特点出发，预期发展出新型的RBM研究体系，并在典型应用问题上达到最佳性能，最终推动深度学习算法高效性与可靠性联合研究，以及推动RBM解决包括强关联物理问题在内的具体应用问题能力的发展。

Restricted Boltzmann Machine (RBM) is a classical algorithm of deep learning with rigorous theoretical analysis. Its efficiency and robustness have attracted widely attention from both academia and application fields, while the joint research of the above two is still in its infancy. This project will firstly study by algorithm and theory: efficient RBM based on entanglement entropy theory, mainly embodied in the theoretical connection and design guidance of entanglement entropy and network efficiency; robust RBM based on adversarial attack, mainly embodied in RBM adversarial examples design and adversarial training methods. On this basis, the project will excavate the correlation between RBM efficiency and robustness, and explore the network architecture design and training methods with efficient and robustness RBM at the same time. The project starts from the different characteristics of efficiency and robustness in RBM and tensor renormalization groups, expects to develop a new RBM research system and achieve the best performance on typical application problems, which will ultimately promote the development of joint research on efficiency and robustness of deep learning algorithms, and improve its ability of solving specific application problems including strongly correlated phytsical problems.