An interesting situation in electronics regarding machine learning: Anybody that has ever laid out a pc board can attest to what an intriguing problem is presented. I know one guy that excels at it. He can complete a layout with two layers that others might only be able to complete with eight layers, he's that good. Programmers have been trying to write software to lay out pc boards for decades, I've heard of some as early as the 1960's. Nothing exists to this very day that can do that task as well as even the average layout specialist, let alone the best. Most of that software isn't even very useful in assisting in the layout. My take on the current machine learning offerings is that they probably offer a lot more than they will deliver even in the upcoming decades. We are underestimating the fantastic CPU's that reside between our ears. That is why a country that allows for freedom and the pursuit of happiness does so well: It frees up those amazing CPU's between our ears to be productive.
*True* machine learning depends on stable data -- data where Xs have no hold-over value from T={t} to T={t+1}. Zero "serial correlation." Financial data is rife with it. But if you try to mention this, or how to address it, you *mostly* get the rolling eyes of a 12 year old being warned of reality by a parent. "Son, you don't want to put a rear wing on a front-wheel-drive 'race car' -- despite how 'cool' it may look. The faster you go, the worse things will get." Daddddddddddd!