Curious as to what folks think of this wired.com/story/googles-
11
14
38
Conversation
1) It's a pretty natural thing to do with RL. Most progress in architecture design is pretty much try-things-and-see-if-they-work, so it's
2
2
2) a pretty grunt work-like process. Even if the "try things" part may be motivated by some theory. Ultimately if we can automate this
1
3) search process it should perform better than humans doing the search and design. Will hell the researchers in dl focus on developing new
1
4) modules as well as the actually theoretical work (for instance why do some networks generalize so well...etc).