In my opinion, this push towards "mathematical elegance " and "interpretability" is hurting us and is unscientific. About interpretability, I'll quote Yann LeC

In my opinion, this push towards "mathematical elegance " and "interpretability" is hurting us and is unscientific. About interpretability, I'll quote Yann LeCun here, you don't really care too much about whether or not your taxi driver is interpretible, he's a black box. Not only that, but there are a variety of new methods such as attention mechanisms, gradient masks and auxiliary explainer networks that give really good interpretability. And it's funny how this criticism never comes up for SVMs... I also think it's unscientific to expect mathematical elegance from everything. Math is ultimately a logical descriptor tool for AI, a means, not an end in itself. Besides, there's actually a host of involved mathematics underlying the seemingly simple SGD of deep nets. For example, tropical geometry has been used to analyze the loss surfaces of ReLU networks and random matrix theory has been employed to analyze the loss surfaces with respect to the quality of local minima. Tl;dr: deep nets are far more interpretible than they've been given credit for and they're also more mathematical than some (such as this author ) would have you believe.