Additive Coupling Layer

• $g(z_{d+1:D};m(z_{1:d}))=z_{d+1:D}+m(z_{1:d})$
• So no matter what m is , maybe a neural network, it is still possible to invert it and compute whatever we need
• They have a unit jacobian determinant, so neither contract nor expand space
  • preserves volume
  • but we don’t really want this because we want the model to be able to freely scale the amount of latent space it uses
  • So we use a Scaling matrix for coupling layers