Introduced on day 46
Variational autoencoder (VAE)
An encoder-decoder model trained to maximise a lower bound on the data likelihood, giving a smooth latent space from which new samples can be decoded.
In detail
The encoder maps an input to a distribution over a latent code and the decoder maps codes back to data. The training objective (the evidence lower bound) rewards reconstruction and penalises codes that drift from a simple prior, so sampling from the prior and decoding yields new data. Latent codes can be interpolated and searched; image samples are often described as softer than GAN or diffusion output.