A1129
Title: Multiscale perspectives on computational statistical methods
Authors: Deniz Akyildiz - Imperial College London (United Kingdom) [presenting]
Abstract: Multiscale perspectives are presented on a number of long-standing problems in computational statistics. First, we show how multiscale and stochastic averaging techniques can be utilized to understand certain practical implementations of statistical inference methods, such as expectation-maximization. These insights are then used to understand a popular training method for generative models, namely contrastive divergence. Finally, if time permits, a novel application of multiscale ideas to multimodal sampling problems is introduced. A high-level introduction is provided to multiscale processes based on stochastic differential equations and how these can be connected to a number of computational statistical ideas and beyond.
