Three Variations on Variational Autoencoders