Abstract
Annealed Importance Sampling (AIS) is a popular algorithm used to estimatesthe intractable marginal likelihood of deep generative models. Although AIS isguaranteed to provide unbiased estimate for any set of hyperparameters, thecommon implementations rely on simple heuristics such as the geometric averagebridging distributions between initial and the target distribution which affectthe estimation performance when the computation budget is limited. Optimizationof fully parametric AIS remains challenging due to the use ofMetropolis-Hasting (MH) correction steps in Markov transitions. We present aparameteric AIS process with flexible intermediary distributions and optimizethe bridging distributions to use fewer number of steps for sampling. Areparameterization method that allows us to optimize the distribution sequenceand the parameters of Markov transitions is used which is applicable to a largeclass of Markov Kernels with MH correction. We assess the performance of ouroptimized AIS for marginal likelihood estimation of deep generative models andcompare it to other estimators.