This is also the original way variational methods pick a parameterization of a model of known architecture which best matches some distribution which generated data but is not otherwise compactly expressible.