Variational Autoencoders

In a Variational Autoencoder, there’s both a Prior, P, and a posterior, Q. Q is the approximate variational posterior.

The prior might be a mix of prediction and goal state. The posterior is the activity activated by the current inputs.