When reading up on deep learning from various sources, one is often presented with the notion that deep learning is essentially a black box: A set of recipes, heuristics and empirical best practices that have been found to often work, but where the mechanism by which good results are achieved is usually not well understood. See for example http://www.nature.com/news/can-we-open-the-black-box-of-ai-1.20731.
I am looking specifically for resources (e.g. specific papers or research groups working in the field) of how this is not the case, i.e. work aimed at understanding why deep learning techniques work, both in the backward and forward direction.
I know about work on optimizing the input to a network in such a way as to activate some unit maximally (see http://yosinski.com/deepvis) and this seems interesting, but that's also a rather descriptive than mechanistic approach.
I would greatly appreciate any pointers!