Evidence estimation using stochastic likelihood approximations
Thesis (MSc)--Stellenbosch University, 2020.
ENGLISH ABSTRACT: We consider the problem of estimating evidence for parametric Bayesian models in the large data regime. Many existing evidence estimation algorithms scale poorly due to their need to repeatedly calculate the exact likelihood, which requires iterating over the entire data set. This ineﬃciency can be circumvented with the use of stochastic likelihood estimates on small sub-samples of the data set. We therefore tackle this problem by introducing stochastic gradient Monte Carlo methods for evidence estimation, our main contribution being stochastic gradient annealed importance sampling. Our approach enables eﬃcient online evidence estimation for large data sets. SGAIS is considerably faster than previous approaches for single data sets, with improved order complexity for online estimation, without noticeable loss of accuracy.