Basics of Variational Autoencoders (VAEs)

An autoencoder learns to encode an input and reconstruct it from a latent representation. A variational autoencoder (VAE) adds a probabilistic model of that representation, so training supports both reconstruction and sampling new outputs. The difference involves the training objective as well as the encoder's output.

Compression and reconstruction

A conventional autoencoder has two parts. The encoder maps an input, such as an image, to a code. The decoder maps the code back to an output resembling the input. A compact code can constrain how much information passes through the network.