A prior distribution is specified before data are observed or encoded. In a variational autoencoder, it provides a shared target distribution for each molecule’s latent representation. section 10.3
Using the standard normal prior N(0, I), training penalizes differences between each encoder-produced distribution and this shared target through KL divergence. section 10.3
Sampling a latent molecule
After training, draw a latent vector z from N(0, I) and pass it to the decoder to generate a molecule from a region the decoder understands. section 10.3











