Speaker: Dr. Dr. Hugh Chipman, Canada Research Chair in Mathematical Modelling, Acadia University
The last quarter-century has seen an explosion of flexible models invented by statisticians and machine learners. Increasing computing power and advances in learning algorithms have made it possible to fit such sophisticated models to large and complex data sets. At the same time, there have been breakthroughs in computational methods for Bayesian statistics, notably Markov chain Monte Carlo methods. In this talk I’ll outline some of the ways that Bayesian methods can be used to learn complicated models from data. Specific examples such as decision trees and ensemble models will be used to illustrate particular issues, including the extent to which statistical inference is possible with complex models, the role of prior information and its ability to regularize estimated models, and how a statistical approach can enrich what would otherwise just be “algorithms for learning from data”