Distributed Variational Bayes Based on Consensus of Probability Densities

This paper discusses a distributed design for the variational Bayes problem in a time-varying multi-agent network. In this circumstance, all the data are distributed stored and unavailable to all agents and each agent has to update based on its own computation and communication with neighbors. By virtue of distributed subgradient methods and consensus of probability densities idea, a distributed framework for variational Bayesian over exponential family is introduced, which can be implemented only based on local datasets. Simulations on synthetic datasets are presented to illustrate the effectiveness of the algorithm empirically, even though the communication network topology is time-varying.