Variational Autoencoder

A Variational Autoencoder – VAE – is a type of neural network used to generate data of various types, and especially images. VAEs use an encoder (a convolutional network) and a decoder (a deconvolutional network) in order to extract a latent vector from sample data. This latent vector can then be used with arbitrary parameters to generate new data.

Simplified diagram of a Variational Autoencoder
Simplified diagram of a Variational Autoencoder

For more information, read this presentation of VAEs and this complete tutorial on VAEs.

« Back to Glossary Index