报告题目：Monte Carlo’s View of AI Developments
报告时间：2025-12-13（周六） 15:00-16:30
报 告 人 ：刘军 讲席教授 清华大学
报告地点：雷军科技楼六楼644报告厅
报告摘要：
Monte Carlo methods first appeared in early days (1945-55) of electronic computing. The technique was named after the famed gambling resort because its procedures incorporate the element of chance. Initially, statistical physicists introduced a Markov Chain-based dynamic Monte Carlo method for the simulation of simple fluids. This method was later named as “Markov chain Monte Carlo (MCMC)” and extended to cover more and more complex physical systems. At almost the same time, a sequential (recursive) construction was proposed to simulate long chain polymers, which can be seen as the ancestor of the popular “particle filters” (aka sequential Monte Carlo). Nowadays, Monte Carlo has been widely used as a powerful computational tool for optimization and integration in diverse fields, especially for various AI tasks. We will first review of Monte Carlo’s history, and then discuss a few recent directions and developments, e.g. Monte Carlo tree search, reparameterization, diffusion sampling, resampling and optimal transport, and particle flow via variational approximation.
专家简介：
刘军现为美国科学院院士、清华大学兴华卓越讲席教授、统计与数据科学系主任。他1985年于北京大学获得数学学士学位，1991年于美国芝加哥大学获统计学博士学位。1991-2025年间，他曾任美国哈佛大学和斯坦福大学统计系助理教授、终身教授；还曾任美国统计协会会刊(JASA)联席主编及多个国际一流统计杂志副主编等职。刘军从事贝叶斯理论、蒙特卡洛方法、统计机器学习、生物信息学等领域的研究，对机器学习、生物医药和人工智能领域有深远影响。他曾获美国统计领域最高荣誉考普斯会长奖（2002）、华人数学家大会晨兴应用数学金奖（2010）、泛华统计协会许宝騄奖（2016）；于2004、2005和2022年分别成为美国数理统计学会、美国统计学会、和国际计算生物学会会士(Fellow)。截至2025年10月，他在各类国际顶尖学术杂志及书刊上发表论文300余篇和一本专著，被引用9.5万余次(Google Scholar)， 已经指导了40多位博士生、30多位博士后。
