I am interested in the differences between using large, deep learning networks vs Probabilistic graphical models (PGMs), like Random Field models, for structured learning (e.g. on images, or labels of arbitrary graphs on surfaces, etc...). For instance, what areas/fields/problems would one or the other be preferred? Are there theoretical guarantees for one vs another? How can they be used together (e.g. in papers like this)?
I have heard lately that deep learning (i.e. conv-networks) has replaced classic structured learning approaches like conditional random fields (CRFs), and was interested in the thoughts of professionals in the field as to the truth of this.
I would of course prefer literature and math to just anecdotes (so no one closes the question :) ).