报告题目：Higher-Order Accurate Approximations for Latent Variable Models
报告人：金少博 博士
报告时间：2020年11月25日(星期三)15:30-16:30
线下报告地点：藕舫楼802
线上参与方式：腾讯会议ID402 908 031或点击链接：https://meeting.tencent.com/s/BQhfaYi74whk
主持人：曹春正 教授
报告内容简介：
In statistics, various latent variable models consist of intractable integrals that need to be approximated. The Laplace approximation is an accurate and computationally feasible method to approximate an intractable integral. For a ratio of two integrals, the fully exponential Laplace approximation takes the ratio of two first-order Laplace approximations and is known to be second-order accurate under mild conditions. It has been used to approximate the gradient for the estimation of various latent variable models, especially the gradient in the EM algorithm, since it has been presumed that the improved error rate is attained also for these cases. In this talk, we will show that the fully exponential Laplace-approximated gradient is the same as the gradient of the Laplace-approximated observed log-likelihood function of the same order, if it is used to approximate the gradient in latent variable models. The implication is that the estimator using the fully exponential Laplace-approximated gradient does not have an improved error rate as previously thought. It implies that direct maximization, if possible, is already an appealing approach and that neither FELA nor the EM algorithm is not likely to offer any major improvement to direct maximization.
报告人简介：金少博博士，2015年毕业于瑞典乌普萨拉大学统计系，现为乌普萨拉大学统计系副教授、博士生导师。曾为伦敦政治经济学院统计系和首尔国立大学统计系访问学者。主要研究领域包括潜变量建模和推断，基于结构方程模型和项目反映理论的心理测量学，以及非贝叶斯方法的模型平均。曾于2015年被提名为国际心理测量协会年度最佳毕业论文，在统计学及交叉学科Top期刊Biometrika、British Journal of Mathematical & Statistical Psychology、Psychometrika、Structural Equation Modeling等发表学术论文10余篇。
欢迎广大师生踊跃参加！
数学与统计学院
江苏省统计科学研究基地
2020年11月24日