Definition
A probabilistic generative model that learns a distribution over latent representations and reconstructs or generates data by sampling from that latent space.
Distinguish it from nearby terms
A conventional autoencoder maps inputs to latent codes and back. A VAE learns a structured probability distribution over the latent space, which enables sampling but adds a variational training objective.
Check your understanding
A VAE can generate new examples by sampling latent points, not only reconstruct inputs it has already encoded.