Ghost CEO John Hayes: A New Way to Drive

Imitation learning vs. hand coding:

An idea I had from reading this post and also inspired by George Hotz’s talk with Lex Fridman):

Why not train a neural network to drive end-to-end and then use neural network intepretability/explanability tools to see what visual representations it learns? If there’s something new there, turn it into a hand-crafted representation and teach it to an NN in a fully supervised setting.