To be understandable, ML solutions need to cleanly separate the "characteristic finding" parts with the "decision tree" parts, however the most efficient networks may well have optimised these things together, like a compiler might.
For example the first impresive ImageNet solvers clearly worked by coming up with a number of characteristics based mainly around various "textures" rather than "shapes", but this wasn't obvious when it was first published. It really seemed like it could "recognise a Panda" etc.
For example the first impresive ImageNet solvers clearly worked by coming up with a number of characteristics based mainly around various "textures" rather than "shapes", but this wasn't obvious when it was first published. It really seemed like it could "recognise a Panda" etc.