Reactive Message Passing for Scalable Bayesian Inference

We introduce reactive message passing (RMP) as a framework for executing schedule-free, scalable, and, potentially, more robust passing-based inference in factor graph representation of probabilistic model. RMP is based on the programming style, which only describes how nodes react to changes connected nodes. recognize suitable abstraction methods that improve robustness, scalability, and execution time procedure are useful all future implementations methods. also present our own implementation ReactiveMP.jl, Julia package realizing through minimization constrained Bethe free energy. By user-defined specification local form factorization constraints variational posterior distribution, ReactiveMP.jl executes hybrid algorithms including belief propagation, passing, expectation maximization update rules. Experimental results demonstrate great performance compared other packages Bayesian across range models. In particular, we show capable performing large-scale state-space models with hundreds thousands random variables standard laptop computer.