From Variational to Deterministic Autoencoders

They show that simple regularized autoencoders can also learn a smooth, meaningful latent space without having to force the bottleneck code to conform to some arbitrarily chosen prior (ie Gaussian for VAEs).

For more details, visit the source.