In important ways, arguably most of the ways that matter, we do not understand how or why modern neural networks work, or accomplish the things that they do.This has come up a lot recently, so I'm going to try to define what the boundary is between what we do and do not understand.In short: What we do understand is the actual math for making them. If we didn't, we couldn't make them. Someone has to write out code for the math, and this code defines both what it is made of and how it learns thing...