Thank you so much Demis to share your results, knowledge and algorithms. 10:51 "So in humans is the dopamine system in the brain that implements a form of reinforcement learning. So we know both from mathematics and from biology that reinforcement learning is a sufficient solution to the intelligence question." 30:25 A clip in which a top professional Go player comments an AlphaGo move, so surprised. 47:28 Conclusion 48:33 Ethics 48:55 "Personal dream: to really accelerate the pace of scientific and medicine innovation in the world." Yes. Do that. Go ahead.

very interesting talk. Thank you for sharing CSAR.

VERY INTERESTING TALK. THANK YOU!!

2:50 talk starts

super cool topic to talk about.

Holy shit he is smart. I realize how dumb I am while trying to follow him, looking up words along the way. =(

@thatguyaknow5557

6 жыл бұрын

Picking up something this advanced from just a video is definitely difficult, but anyone could understand it with enough research.

Fascinating talk.

DeepMind's learning mechanism is based only on pixel input, or vision input. -- Gerald R. Tamayo

really nice speach!

Great talk!

Interesting how they share the codes as open source...

No 9 man endgame tablebase exists till date. Most are 7 but they are not used directly with computers. 6 piece tablebases are commonly used with chess engines.

Perhaps he will make a General AI.

25:16 "So we played him, and he's also [...] for being incredibly creative." What does he say there?

@GoSuNiguy

7 жыл бұрын

"So we played him and he's also *fated* for being incredibly creative."

@guepardiez

7 жыл бұрын

Thanks. I considered that, but it doesn't make much sense to me. As far as I know, you can say "fated to do something" (predestined to do something), but I can't see what "fated for being something" means.

@GoSuNiguy

7 жыл бұрын

What Demis Hassabis is suggesting is that Lee Sedol, in the context of Roger Federer, has played interesting and creative games in the past when he won various titles. He's saying that Lee Sedol is bound to come up with interesting and unconventional moves to test AlphaGo's limits which is exactly what happened in Game 4. In this context, that's what Demis Hassabis meant by "[Lee Sedol is] fated for being incredibly creative."

@evilteddy11

7 жыл бұрын

feted not fated. It means celebrated.

@guepardiez

7 жыл бұрын

Thanks a lot, evilteddy11. I knew "fated for" wasn't possible, but I would never have guessed that a word that is pronounced "feited" could be written "feted". O English orthography!

This is "guided brute force." Or, what we in the real world call, "parenting." ;-)

I had the idea to make an adversarial network for AlphaGo who maximise expected winning margin. That would be interesting in itself. Because AlphaGo is the angel that maximise his winning probability. The other would be a devil who don't slow down when he is ahead and does not conceal his full strength. This devil would be a hell of an agressive player.

@st2rseeker

7 жыл бұрын

This is a really interesting idea - I have wondered many times how strong could AlphaGo play were it "allowed" to compromise winning chance (within given bounds) in order to maximize winning margin. It's pretty documented how it plays "slack" moves when it's certain of victory.

@Silly.Old.Sisyphus

7 жыл бұрын

how is your devil going to calculate the score? in any case, a move which promises victory is much stronger than one which expects to win by a bigger margin but at a reduced likelihood. Alfa is smarter and stronger than your devil.

@danodet

7 жыл бұрын

Suppose Alphago and the devil play a 10 games tournament and these are the results: 3.5, 3.5, 3.5, -8.5, 1.5, 0.5, -10.5, 7.5, -8.5, 6.5 (minus sign means Alphago lost) Who won the tournament? In a way it's Alphago because he won 7 games. In another way it's the Devil since points add up to -1. Here AlphaGo would be like a president who won the electoral college and lost the popular vote... Go is a game of balance : territory vs influence, controling territory vs capturing stones, risk vs safety, yin vs yang, black vs white. To get better it's best to play someone who don't think like you do. For that reason I think this devil would be a good adversarial opponent to train AlphaGo.

@Silly.Old.Sisyphus

7 жыл бұрын

proof, as if any were needed, that the popular vote is meaningless, not least because the whole thing is rigged anyway by the Deep State of Mind Control

Can Deep Learning solve tactical parts of the game of Go? (I believe no.)

@Silly.Old.Sisyphus

7 жыл бұрын

Kato Masao would never ask such a question, for he more than anyone would know that Alfadog is bloody good at tactics, and only tactics, which is more than enough to knock humans off their pedestal

@HIDEKI53

7 жыл бұрын

Who confirmed if AlphaGo is good at tactics? No one other than the developers can do that. Solving complex L&D problems is a big weak point in MCTS so far. Can AlphaGo be an exception? If so, how? #Looks good is never "is" good.

@timothybolshaw

7 жыл бұрын

The program is good enough at the tactical parts to beat the top pros 60 to 0. Historically, many pros will initiate fights as soon as they judge they will win. Conclusion: AlphaGo can recognize positions where it would be at a disadvantage tactically and avoid them. This shows a tactical appreciation exceeding that needed to make the correct moves once a fight is initiated.

@HIDEKI53

7 жыл бұрын

The question is "Can solve". Your statements are not enough to prove this but only saying AlphaGo is strong enough to beat human players, who are obviously far from being able to solve tactical parts of the game of Go.

@Silly.Old.Sisyphus

7 жыл бұрын

i confirm she is good at tactics, because she is good at reading, even if she doesn't know what she's looking at until the end of the branch. the empirical justification is in her game record - she has won every tactical fight in every public game except one last year.

really nice speach!