机器学习与数据科学博士生系列论坛（第七十八期）—— Recent Advances in Anytime-Valid Inference
发文时间：2024-10-31
Speaker(s)：谢楚焓（北京大学）
Time：2024-10-31 16:00-17:00
Venue：腾讯会议 568-7810-5726
摘要：
Traditional statistical inference for parameters is based on asymptotic coverage guarantees at fixed sample sizes. Such guarantees tend to become problematic in sequential experimental design due to the issue of "p-hacking". In recent years, there has been an emerging literature on safe anytime-valid inference that focuses on mitigating this problem and provide valid coverage guarantees for any stopping times. The main technique is called the supermartingale method, coupled with Ville's inequality.
In this talk, we give a detailed discuss on the original intuition, popular applications, and recent advances on anytime-valid inference. Particularly, we will focus on its usage in hypothesis testing of parametric examples, construction of confidence sequences for sample means, and modification to adapt to asymptotic theory. We will finally introduce an application to parameter inference in stochastic approximation.
论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。