Yeah, one shot learning in humans is built on top of enormous model representations of the world built up over many years, so I wouldn’t be surprised if we’re a bit premature in trying to pull it off in ML other than in very simple cases.