An autoencoder whose code is a smooth probability space, so you can sample new data from it.
- Encode to a distributionOutput a mean and spread, not one point.
- SamplePick a point from that distribution (reparameterisation trick).
- DecodeRebuild data from the sampled point.
- Two lossesReconstruction plus a KL term keeping codes smooth.
Explain 1/5
Source: Kingma and Welling: Auto-Encoding Variational Bayes