I can't answer the question directly, but I will say this: machine learning is a lot of applied math. Suppose you are setting up a convolutional network to rec

I can't answer the question directly, but I will say this: machine learning is a lot of applied math. Suppose you are setting up a convolutional network to recognize some special object for a company. You will need to understand that math to know what parameters to tweak. Is it the learning rate? Is it the way you randomized the weights? Is it the activation function? Although, in fairness, I don't think even a PhD level candidate works out what the reason is likely to be. More than likely they have a few heuristics in their head (oh, it stops learning too soon, let's just drop the learning rate. Oh, it never converges? that activation function can't propagate error and so on). The point is that you have to know the theory to be useful. It hasn't been worked out. It is very much a living science project. That's the fun of it though.