'Less Than One'-Shot Learning: Learning N Classes from M

There was an interesting discussion on this paper recently:

https://news.ycombinator.com/item?id=24820298