All Your Base Are Belong to Us

First they came, for Checkers, but I was silent, because I played chess. Then they came for chess, so I might have been a bit less silent, but hey, I still played Go. I whined when Go went down, but now, goldarn it, Alpha Zero has bested the strongest computer programs in Chess, Shogi, and Go, starting from scratch in each case (just the rules) and training itself solely by self-play, making zzero use of all the information humans have assembled over thousands of years of play. Even more annoyingly, it took only a few days to do it. Via Steve Hsu.

It's somewhat analogous to giving a five year old a chess set, explaining the rules to him, and coming back a few days later to find he was far better than the best player in the world. Clearly the best neural networks can now learn far more rapidly, and understand more deeply, than any human, at least in narrow domains.

All these games are narrowly constrained by definite rules and very specific outcomes. In most human activities we either have no definite rules or don't know the rules, so the strategies of Alpha Zero are not directly applicable. One domain where the rules are pretty definite is math. There have been a few situations where computers have been able to prove specific theorems of interest to humans, but I wonder if anybody has given a neural network like Alpha Zero the very simple axioms of say, group theory, and asked it to prove all the interesting theorems. An even better trick would be to get it to discover the interesting theorems.


Popular posts from this blog

Left, Right and Indian