variational autoencoders are powerful models for unsupervised learning.
However deep models with several layers of dependent stochastic variables are
difficult to train which limits the improvements obtained using these highly
expressive models. We propose a new inference model, the La