Variational Inference - Amortized VI And VAEs

Amortized VI And VAEs

Classical VI fits one $q$ per dataset. Amortized VI trains a neural network $q_{\varphi }(z|x)$ that maps any $x$ to its approximate posterior, so inference at test time is a single forward pass. This is the encoder of a Variational Autoencoder, and the ELBO with the reparameterization trick is its training objective.