Or it could be the beginning of a new kind of science where people form theories to explain machine findings....https://twitter.com/_SalvorHardin/status/1078053904577904641 …
-
-
Replying to @etzioni
Or where authors write wildly romantic things about mechanical procedures. Honestly, this article is way over the top! It also confusing an optimization (AlphaZero) with supervised learning (medical imaging). When solving optimization problems, computers can find new solutions 1/
1 reply 0 retweets 5 likes -
Replying to @tdietterich @etzioni
But when learning from human examples, it can rarely do much better than people, because it is just learning to imitate people (and sometimes be more consistent). Supervised learning does not create a new kind of intellect. end/
1 reply 0 retweets 5 likes -
Replying to @tdietterich @etzioni
P.S. I don't think most problems of interest in the physical/biological world can be formulated as optimizations.
3 replies 0 retweets 4 likes -
Replying to @tdietterich @etzioni
The optimization problem is model discovery. The model itself is data compression. For instance, a mathematical proof compresses statements to axioms. Discovering a proof is finding a computable function that performs that compression.
2 replies 1 retweet 3 likes -
I agree that finding proofs is like playing games, because we can evaluate any proof to determine if it is sound.
1 reply 0 retweets 0 likes -
The problem with proofs is you still have the open vs closed domain issue. You aren't selecting from a fixed, clear set of actions per well defined turn. In addition to difficulty in generating valid proofs, evaluation remains an issue because there can be many valid, long paths
1 reply 0 retweets 1 like -
I had in mind finding proofs in a formal proof system (e.g., SMT, resolution, etc.). Those are closed worlds once you have formulated the formal claim you wish to prove.
1 reply 0 retweets 1 like -
Ah, right. It still seems there'd be difficulty training this. Given a theorem, it must initiate a proof search that emits a proof, unlike games where things can have rounds or be sliced. RL also hard given undecidability of each unfolding tree. Self-play also hard to see
1 reply 0 retweets 0 likes -
I was thinking of pure optimization problems, such as Go, where the space of possible moves, the rules of the game, and the definition of winning are all known. In principle this is purely a computational problem. There is no need for machine learning. Proof systems are similar.
2 replies 0 retweets 1 like
It turns into an optimization problem when we can only afford to conquer a tiny part of the search space, and need to figure out where to look first.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.