Zhu Jun, Professor of Computer Science Department in Tsinghua University, Assistant Director of Brain and Intelligence Laboratory in Tsinghua University, Assistant Director and Dean of Basic Theory Center of Tsinghua University Institute of Artificial Intelligence, former Deputy Director of State Key Laboratory of Intelligent Technology and Systems and Adjunct Professor of Carnegie Mellon University. Mainly engaged in machine learning research, published more than one academic paper 100 in important international journals and conferences. He is the deputy editor-in-chief and member of the editorial board of the international periodical IEEETPAMI, the regional co-chair of the international conference ICML20 14, and the domain chairman of international conferences such as ICML and NIPS. The research work focuses on the basic theory, efficient algorithm and application of machine learning, and pays attention to the combination of theory and practical problems. Aiming at the problem of * * * in the research and utilization of complex data hiding structure, some key problems in structure learning and structure-based statistical learning are studied, and PAC-Bayes theory and method of 1 and maximum entropy discriminant learning are proposed. 2. Regularized Bayesian inference and regularized nonparametric Bayesian inference theory. 3. Maximum interval learning theory and efficient algorithm of Bayesian model. (4) Abacus depth probability programming library, etc. Aiming at many typical application scenarios, such as Internet data mining, social network analysis, multimodal data fusion, network recommendation and so on, this paper combines basic theory with practical problems and puts forward effective calculation models and algorithms. Including the application of regularized Bayesian inference to solve large-scale text classification, social network analysis, matrix low-rank decomposition, multimodal data fusion and other problems, and put forward an efficient learning algorithm. Structured maximum entropy discriminant learning is used to solve the problems of information extraction, entity relationship extraction, multimodal data fusion and retrieval under the network environment. A structure-based network data extraction framework and several statistical models including StatSnowball were established, and 17 invention patents were applied, including 3 American patents. The research results have been applied to several search engines of Microsoft, including people-centered search engine and academic search engine.