Keywords: normalizing flows, variational inference, approximate inference
TL;DR: We present a software framework for transforming distributions and demonstrate its flexibility on relaxing mean-field assumptions in variational inference with the use of coupling flows to replicate structure from the target generative model.
Abstract: Transforming one probability distribution to another is a powerful tool in Bayesian inference and machine learning. Some prominent examples are constrained-to-unconstrained transformations of distributions for use in Hamiltonian Monte-Carlo and constructing flexible and learnable densities such as normalizing flows. We present Bijectors.jl, a software package for transforming distributions implemented in Julia, available at github.com/TuringLang/Bijectors.jl. The package provides a flexible and composable way of implementing transformations of distributions without being tied to a computational framework. We demonstrate the use of Bijectors.jl on improving variational inference by encoding known statistical dependencies into the variational posterior using normalizing flows, providing a general approach to relaxing the mean-field assumption usually made in variational inference.