UCPH Statistics Seminar: Wenkai Xu

Title: Statistical Machine Learning with Kernels and Stein's Method

Speaker: Wenkai Xu from University of Tübingen

Abstract: The core of statistical machine learning relies on comparing distributions. The choice of statistical measures of distribution comparison is crucial in practical machine learning tasks. In this talk, we will discuss recent development of discrepancy measures based on reproducing kernels and Stein operators, where their corresponding integral probability metric (IPM) has shown both statistical and computational advantages dealing with unnormalised models, heavily censored data, and data with network structures. We give examples on various machine learning applications including hypothesis testing, optimisation, and generative modelling. We draw connections of these IPMs with other family of statistical measures such as f-divergence and conclude the session by discussing open problems and future research.