Variational Autoencoders (VAEs) differ from the standard autoencoders that we have discussed so far, in the sense that they describe an observation in latent space in a probabilistic, rather than deterministic, manner. As such, a VAE outputs a probability distribution for each latent attribute, rather than a single value.
Standard autoencoders are only really useful when you want to replicate the data that was input into it, which has somewhat limited applications in the real world. As VAEs are generative models, they can be applied to cases where you don't want to output data that is the same as the input data.
Considering this in a real-world context, when training an autoencoder model on a dataset of faces, one would hope that it would learn latent attributes such as whether the person is smiling, their skin tone, whether they are wearing glasses...
