B1952
Title: Learning and testing heavy-tail distribution via stereographic projection
Authors: Wenkai Xu - University of Tuebingen (Germany) [presenting]
Jun Yang - University of Copenhagen (Denmark)
Abstract: The focus is on the problem of learning and inference for heavy-tailed distributions, e.g. student-t distribution, which can be challenging or even prohibitive due to its computational pitfalls associated with the learning objectives, especially in high dimensions. The proposed framework utilises stereographic projection, a conformal transformation mapping the Euclidean space to hyperspheres, where useful techniques and properties developed for directional statistics apply. We present a series of examples, including variational inference, goodness-of-fit testing, and generative modelling. We also show the advantages of our framework and superior performance in simulations.
