An important addition to this is priors: deep networks allow to express the prior that hierarchical representation, i.e. composing into multiple layers of abstr

An important addition to this is priors: deep networks allow to express the prior that hierarchical representation, i.e. composing into multiple layers of abstractions make sense (see e.g. conv nets).