Variational inference with copula augmentation

We develop a general methodology for variational inference which preserves dependency among the latent variables. This is done by augmenting the families of distributions used in mean-field and structured approximation with copulas. Copulas allow one to separately model the dependency given a factorization of the variational distribution, and can guarantee us better approximations to the posterior as measured by KL divergence. We show that inference on the augmented distribution is highly scalable using stochastic optimization. Furthermore, the addition of a copula is generic and can be applied straightforwardly to any inference procedure using the original meanfield or structured approach. This reduces bias, sensitivity to local optima, sensitivity to hyperparameters, and significantly helps characterize and interpret the dependency among the latent variables.