Why does deep and cheap learning work so well?