This podcast talks about the cognitive architecture that is needed to have a perceptual system, learn the dynamics of the world, and then have a routine program learning system to learn concepts, including the back propagation and credit assignment techniques. The inference side is more amenable to being constrained since experiments are much easier to perform.