> The only thing is that nobody understand why they work so well.

This is exactly what I have ascertained from several different experts in this field. Interesting that a machine has been constructed that performs better than expected and/or is performing more advanced tasks than the inventors expected.

The linear regression model "ax + b" is the most simplest one and is still quite useful. It can be interesting to discover some phenomenon that fits the model, but that's not something people have control over. But imagine spending years (expensively) training stuff with millions of weight to ultimately discover it was as simple as "e = mc^2" (and c^2 is basically a constant, so the equation is technically linear)