Optimization of Annealed Importance Sampling Hyperparameters

Abstract Annealed Importance Sampling (AIS) is a popular algorithm used to estimates the intractable marginal likelihood of deep generative models. Although AIS guaranteed provide unbiased estimate for any set hyperparameters, common implementations rely on simple heuristics such as geometric average bridging distributions between initial and target distribution which affect estimation performance when computation budget limited. In order reduce number sampling iterations, we present parameteric process with flexible intermediary defined by residual density respect mean path. Our method allows parameter sharing annealing distributions, use fix linear schedule discretization amortization hyperparameter selection in latent variable We assess Optimized-Path models compare it more computationally intensive AIS.