That's not how deep learning works as I understand. Training and testing a NN on CBR would give you, at best, a NN with weights that was really good at recognizing parts of the CBR. Would you be inspecting the arrangement of weights and netowork? Would you look at other data trying to find hidden instances of parts of CBR? In pop music or old paintings?!
Or, maybe you could expand on what you mean by learn and understand in this instance.