VAE are not really autoencoders, except that the training algorithm resembles that of an autoencoder. Perhaps this section should be moved somewhere else.