Google open sources its Tensorflow for 'wide and deep' learning. I very much like the thought. In our own exploration of pattern matching with neural nets we often did what was essentially deep learning. We left the 'wide' to what were variants of non procedural code (aka Declarative code, e.g. Prolog) .
Beyond that was what we might have called 'thin' ... that is classic analytic methods or algorithms to solve numeric sub-problems. Combining them harmoniously is an interesting thought. Am I thinking this right? Considering it further.