[nicolapede]

> But many of the problems that remain untouched by DL, like human cognition, are NOT driven by gradients.

This seems to me quite a deep insight. But how would you formally define a process that is gradient-based?