Department seminar by Neil Spencer

Fast approximate BayesBag model selection via Taylor expansions​​​​​​​

In recent years, BayesBag has emerged as an effective remedy for the brittleness of traditional Bayesian model selection under model misspecification. However, computing BayesBag can be prohibitively expensive for large datasets. In this talk, I propose a fast approximation of BayesBag model selection---based on Taylor approximations of the log marginal likelihood---which can achieve results comparable to BayesBag in a fraction of the computation time. I provide concrete bounds on the approximation error and establish that it converges to zero asymptotically as the dataset grows. The utility of this approach is demonstrated using simulations, as well as model selection problems arising in business, neuroscience, and forensics.