Abstract
Probabilistic modelling is a general and elegant framework to capture theuncertainty, ambiguity and diversity of data. Probabilistic inference is thecore technique for developing training and simulation algorithms onprobabilistic models. However, the classic inference methods, like Markov chainMonte Carlo (MCMC) methods and mean-field variational inference (VI), are notcomputationally scalable for the recent developed probabilistic models withneural networks (NNs). This motivates many recent works on improving classicinference methods using NNs, especially, NN empowered VI. However, even withpowerful NNs, VI still suffers its fundamental limitations. In this work, wepropose a novel computational scalable general inference framework. With thetheoretical foundation in ergodic theory, the proposed methods are not onlycomputationally scalable like NN-based VI methods but also asymptoticallyaccurate like MCMC. We test our method on popular benchmark problems and theresults suggest that our methods can outperform NN-based VI and MCMC on deepgenerative models and Bayesian neural networks.