Variational Autoencoder (VAE) (Kingma et al., 2013) is a new perspective in the autoencoding business. It views Autoencoder as a bayesian inference problem: modeling the underlying probability distribution of data.