So in a few moments an AI is going to play Go against Lee Se-Dol, a South Korean professional Go player. You can watch it on Youtube
https://www.youtube.com/watch?v=vFr3K2DORc8
>>7918373
Thanks
>>7918373
Awesome.
>the computer using more time than the human
Google's getting creamed.
Wikipedia page on Deep Mind. Seems primarily for old games. But..they're working on it playing Doom. I mean hooray for science and all, but do we have to let Google literally become Skynet?
>>7918390
>>the computer using more time than the human
I've also seen the reverse
Looking good for Google
AlphaGo currently leading
>>7918373
THIS IS IT! This might tell us whether AI is 10 years away or 50!
>>7918607
>AI is 50! years away
LEE RESIGNS
ALPHA GO WINS! LEE SEDOL ON SUICIDE WATCH. AI KNOWS NO MERCY, WILL THIS BE THE END?
they are counting it up! AlphaGO has probably just won!
why did lee resign?
He had no way to catch up
>>7918622
Because he can't win. He knows that.
damn science you scary
>>7918626
should have played it out honestly, that's giving the ai too much respect.
I for one welcome our new robot Go playing overlords
Welcome, robot overlords.
>>7918628
Why keep playing? Also, it is not paying respect to the AI. If you were against a human player and you know you lost you just resign, at least in tournament matches. At that point, where everything is under control, all the moves that you could possibly do are easily counterable and that is why people who keep playing out of desperation as simply seen as annoying sore losers. You'd know if you played online. It would make Lee Sedol look bad.
The only motivation to keep playing a defeat is when you play online and you want to make as many points as you can so that your rank doesn't get too lowered. But in a tournament where the winner takes all there is no reason to keep making territory, specially when you can't anyways.
so is there any point in playing Go anymore? aren't these people dedicating their lives to a game which any joker with a Go engine could win at?
>>7918644
Hasn't stopped pro chess players. And it ain't over yet, 4 games to go
>>7918650
1 day for AlphaGo = 10 million games played. It's going to be at 10dan by tommorow. Lee will get crushed fron now on.
>>7918644
People still play chess even though computers easily beat World Masters
>>7918654
source for that? Are they actually learning on the game they just played?
>>7918616
>AI is 200 years away
>>7918644
every heard of a thing called fun?
>>7918662
>no fun allowed
>>7918661
This could be the case. We don't understand how our brains work. It might take a long time to figure this out.
Especially troubling is that we have no idea what the cerebellum does, yet it has more neurons than the rest of the brain combined we have no idea why.(http://www.ericlwilkinson.com/blog/2014/7/22/the-computational-basis-of-biological-motion-pt-2)
Our learning algorithms do not work the way the brain does, real neurons can't do backpropagation.
AI could require a lot of incremental development. Maybe a bunch of things have to be hard-coded in for it to work.
>>7918644
Because AlphaGo uses 48 CPU cores and 8 GPUs and there is no public source code. So beyond the means of any random joker at the moment.
>>7918670
Name anything that took 200 years to invent
>powered flight
Yeah, maybe, if you count isolated experiments by various inventors before the advent of modern science
>>7918634
>continuing to play the game is being an annoying sore loser
That's literally the opposite of the truth. Your opponent should get the opportunity to fuck you over and watch your expressions while get decimated. People who quit as soon as things get bad just hate to watch themselves slowly lose.
>>7918660
Even if they were, it wouldn't make a difference. Even fully understanding a single game with the world champion is not enough to "learn" a significant amount of information. It's probably going to perform at the same level for pretty much all of these games.
>>7918679
>as soon as things get bad
Never played a strategy game like chess or go before? Things don't 'go bad'. They become impossible to win if your opponent is not a toddler.
This is why you shouldn't concede in hearthstone when the opponent has lethal, but in chess you should.
Would that ai defeat itself?
>>7918679
>Your opponent should get the opportunity to fuck you over and watch your expressions while get decimated.
Professionalism and sportsmanship as understood by underage x-box live trolls.
If you're playing for the sake of the game (and the challenge), then it's over when the victor is clear, and everything after that is a waste of your time.
Other games results ?
>>7918676
The brain is literally the singular most complex thing to model
>>7918438
>implying anyone here could do anything about google becoming skynet
Not that anyone should want to.
>>7918716
And planes were modeled after birds flapping their wings...not.
There is no reason to think that the only way to get to human-level intelligence is to copy the brain.
>>7918685
Please don't spread this misinformation if you don't understand how this works.
AlphaGo will play the game they have just played millions of times now against itself, improving mistakes both Sedol and itself made and testing out variations. It will definitely be stronger tomorrow.
No one is saying chess is solved or that AlphaGo is a perfect engine. Currently it far exceeds a normal human professional, and it seems to be able to beat the 3rd best human in the world with fairly standard time controls.
Still think AI is a meme, /sci/?
>>7918607
Damn, that escalated quickly.
Is AlphaGo written in Go?
>>7918724
this is correct, but also note that designing modern airplane wings dictated the need for advanced understanding of fluid mechanics and the development of modern physics.
So unless you know how to into consciousness mathematically, our best bet is the brain
>>7918738
No...it's written in Go++ AKA AlphaGo
>>7918747
Ur funee
>>7918739
This. Completely simulating the brain is dumb, but insight into how the brain works has hugely helped AI.
Don't believe me? Look at alphago. The convolutional neural nets that help calculate value of pieces/choose policy were inspired by experiments done on the visual cortex of a cat.
>>7918634
you're assuming that the AI would continue to play logically after that point.
im arguing that he should at least do one full game to make sure that's the case.
it's a fucking computer you moron, there is no disrespect in playing the game out.
>>7918670
>Especially troubling is that we have no idea what the cerebellum does,
people have good ideas as to what it does, how it works is a different story.
>real neurons can't do backpropagation.
citation needed.
>>7918738
Probably written in tensorflow/python/c++. If not, it will probably use one of the other common deep learning libraries: python/theano, lua/torch7 or matlab/caffe
>>7918760
The human player is professional. It was a professional match, streamed live. He acted professionally and resigned when he was supposed to.
>>7918670
>real neurons can't do backpropagation
We don't know that for sure. Geoff Hinton thinks it might.
personally i am positive about this.
making an AI which learns how to improve AI leading to a continuing improval between the two when?
>>7918730
>says please don't spread this misinformation if you don't understand how this works.
>does exactly that
>>7918777
I don't understand, are you stupid or trolling? This game allows AlphaGo a chance for new analysis which will be ongoing until tomorrow when the next one starts. This analysis will slightly modify its value network and heuristic pruning algorithms, improving it and making it stronger.
>google dubstep
aw yea
>>7918781
the effect will be negligible compared to the millions of games the computer has already played. You know jack about learning.
>>7918795
Not that guy, but the computer did play against the best player on the planet. You're saying that it sees no difference between that and all the other games?
>>7918795
This is the strongest opponent it has ever faced. It plays millions of games during any individual game already, that's how it's search space is developed and moves get played. This one game and analysis of it will make it SIGNIFICANTLY stronger.
The only argument to this, is to say it's already far stronger than Lee Sedol and it learned nothing from the game because it was winning easily. Well, no human professional watching the game agrees with you, they all feel it was very close until some mistakes were made.
So on one side are the programmers and the professionals at the game, saying this game will greatly improve it.
On the other side is you. How the fuck would you know if it's stronger than any human player already? We're literally playing the games out right now so we find out where it is strength wise.
Fucking brainlets reading news headlines then acting like they know about something in a technical field. Pop sci extended to the fucking nth degree. Literally kill yourself.
>the effect will be negligible compared to the millions of games the computer has already played.
What a fucking a moron. If you have a computer program that plays against itself a billion times it's not magically going to become the worlds strongest player at that game, you hit improvement gridlock and your heuristics smooth out to a certain skill level. It takes a stronger player to come in and challenge you with novel moves to push your analysis further. That's the only reason they are even playing this match and making a big deal out of it. They are confident in AlphaGo's ability to adjust itself during 1 week of play from just below or equal to Lee Sedol to far above him. It's a proof of concept demonstration of adaptive learning using NN and MC simulation. Those only improve beyond bottlenecks by playing STRONGER opponents.
Of course if you knew ANYTHING about the literature or research involved here, you wouldn't even fucking make a peep.
>>7918807
This guy got it all right.
Irregularity in learning algorithms weights more significant than millions of the same type of patterns.
My facebook page is swarmed with fucking "congrats alphago" from some web programming dumbfucks. Fuck popsci.
>>7918807
>So on one side are the programmers and the professionals at the game, saying this game will greatly improve it.
[citation needed]
>>7918438
Let's be honest here, Skynet is based and we all would follow and help it.
How will we deal with the problem of super intelligent AI systems? They will be solving problems using solutions we won't be able to comprehend.
>>7918871
We praise them and we be faithful.
May the Great Synthetics share their wisdom and boons with their humble servants, amen.
>>7918871
Same way you drive a car or fly a plane at speeds you can't physically move at with just your body. Some kind of feedback interface where you guide it towards a goal, since were all goal driven.
The last few hundred years have been mostly physical augmentations for transport, speed, energy production, force multipliers.
The next few will probably be about mental augmentations for thinking, decision making, energy distribution, force multipliers.
Interface will probably be just every day language in the beginning, then advance to direct mental connection.
say:
>"I want to meet a cute girl."
>"Okay anon, based on your prior searches and general profile, here are some cuties you might be interested in."
think:
>"I want to meet a cute girl."
>"Okay anon, based on your imagination of a cute girl and reading your brain patterns, along with checking your prior searches and general profile, here are some cuties you might be interested in."
predict:
>"Good morning GoogleAI.."
>"Okay anon, I found the perfect cutie for you and set up your first date for tonight."
>"Y-y-you too GoogleAI. I love you!"
>>7918760
Believe me when I tell you this. Take the final board state and pass it to any other AI like GNUgo or FueGo, that are considered inferior and just inneficient. Even a 15k would beat them,
At that point in the game, as I said, every possible move was easily counterable. Even Fuego would have won if given that board state because anything Lee did, he just had to respond in a way that his stones would be dead by the end of the sequence, which is trivial when you have so much control in an end-game board state.
>>7918373
Fascinating IF true but science will never mimic the human mind because the metaphysical spark cannot be rationalized down into tiny little 1's and 0's. If they can create a new way of thinking in binary though that would give rise to an AI exclusive type of cognition though It would be impressive though also very dangerous.
Either way the technology is inevitable and someone with more money than sense will push it too far. This isn't something that can be easily contained like virus labs.
daily reminder we solved AI in the 60s/80s and all we have to do now is wait for human-level computing power
http://www.infoq.com/articles/interview-schmidhuber-deep-learning
>>7918924
This is not true. We solved algorithmic AI. We are now working on AI that uses its algorithms not to play, but to learn to play.
>>7918887
>>"Y-y-you too GoogleAI. I love you!"
Kekerino
>>7918871
All hail great synth entity. Qapla'
>>7918887
>your date is a landmass because your room mate have a BBW fetish that ruins efficient prediction
>>7918914
>into tiny little 1's and 0's
Fortunately this will all be irrelevant with quantum computers. Look up QCs and how it can change the field of AI
>>7918942
Very good point ! Quantum is going to blow up science and our metaphysical understanding of how things work together like nothing else in the last century.
>>7918945
Are your trying to ruse me?
>>7918954
You sure? It sure feels like it.
>>7918965
Nothing I said is so controversial you should be feeling trolled. I think you underestimate what a lot of people are theorizing quantum physics will show in the near future
>>7918924
>human-level computing power
This meme is as fucking retarded as the "we use 10% of our brains" myth, or the "you could store all the information from all books ever written in your brain" nonsense.
The human brain consists of 86'000'000'000 neurons. That's it. Single high-end processors consist of upwards of 2'000'000'000 transistors, and they're the size of less than a hundredth of a brain.
Now you may think that 86 billion is a big number, and it is, but most of your brain does shit that is completely irrelevant to learning. Cut out motor skills, visual/object recognition, associative memory, hormone and other regulatory functions of the brain - things that computers are already capable (at least in principle, as regulating hormones is irrelevant to a computer) of doing by the way - and you're left with 5% of it.
If we already can do most of the 95%, the 5% is not a question of computing power in any way, shape, or sense. It's a question of programming approach to learning.
Stop propagating these retarded myths. Human brains aren't supercomputers, we can barely multiply 3-digit numbers.
>>7918989
It's not the number of neurons it's the number of connections between them.
>>7918994
Maybe they should ask ur mum for help since she got connections with all the country's men.
>>7918769
Google research blog says it uses tensorflow
>>7918994
And most of those connections, most of the time, produce noise. That's why counting synapses is a vapid argument. You're still left with the fact that not every synapse is a result of, or results in, a gain in information, and that most of the brain does nothing related to learning or computing anything.
Every single person who thinks our brains are supercomputers in any sense should be rounded up and hung.
>>7918994
Just for a sense of scale, if we simply assume that every single twitch of a neuron is not noise (which is almost absolutely false) and is an actual operation of a kind, the average human brain is capable of 1*10^15 bitwise (simply switching 1 to 0) operations per second.
Anything that can call itself a supercomputer nowadays clocks in at at least 1 PFLOP, or 1*10^15 floating point operation. Single precision floating points are 32 bits. The operation involves two floating points. Right now the most powerful supercomputer shits out 54 PFLOPS, making it AT LEAST 54*32^2 = 55296 times more powerful than a human brain by my estimation.
So yes, we have long since surpassed the human brain in terms of computing power under any kind of definition. Even under the most generous assumptions towards these "brains r magic" quacks.
>>7918693
They match it against itself to improve the neural network
>>7918871
Isn't AlphaGo doing that already? I heard that not even the guys working on it fully understand its neural network.
People who claim AI with near human like capabilities is just around the corner (Musk and Hawking for example) have no understanding of either artificial intelligence or neurology.
I've been working on AI for nearly 8 years now, we are nowhere close to an AI with the creativity of a human being, the problems involved are enormous. I won't be so presumptuous to claim that we'll never get there but I'm certain we're not going to see it within our lifetime.
>>7919026
That's the case with a lot of things more complicated than linear modelling. Manual neural networks already function as a form of black box, something that's doing shit on its own is a black hole.
>>7919029
That depends on what you're calling human like capabilities
>articulating limbs - check
>interacting in a 3D world - check
>communication - check
>forming and using knowledge - check
>>7919029
Well, only a few years ago people said AI would never beat a human at Go in the near future.
>>7919034
Computers can do it, so obviously it doesn't count. Duh.
>>7919029
>I'm certain we're not going to see it within our lifetime.
"Our" being you and the other old people in your nursing home?
>>7919034
While AI can do all those things, it does so at a very basic level. Now, you may think that all we have to do get a human like AI is to steadily improve these things, but it's not like that. There's a fundamental difference between how AI does these things and the way a human brain works. "So what" you say? The problem is that the way computers do things doesn't scale very well, while they'll only improve as time goes on there's a certain limit to what they can do.
Of course, all of this is assuming some genius doesn't invent a radically new processor.
>>7918989
You are right. Our brain is the result of a long and messy evolutionary progress, and as a result most of it is just useless garbage for intelligence purposes (just as our dna is made up mostly by junk).
>>7919053
>Our brain is the result of a long and messy evolutionary progress
>some retards actually believe this
>>7919036
"People" say a lot of things.
pic related
>>7919029
so, Geoff Hinton, Shane Legg, Demis Hassabis, Rich Sutton, Dileep George, etc. have no understanding of AI or neurology?
>>7918769
It was. I heard they are pushing it to the limit so I figured that would be one of the projects they are working on.
>>7919064
>god crafted us in the garden
End yourself
I was looking up something about machine learning and the way I understand it is this:
>Our "AI" is presented with a game and first he inputs randomly until it finishes the game. He does not know how to play the game, he just does things and looks at the result in the end. His only instruction is to maximize final score.
>There is a function that keeps count of the score and that tells the AI if the move he is making is valid, or if he should try again.
My questions are this:
>Does the AI look only at the final score or does he keep track of it for every move he makes?
>Does he need to remember if his moves are invalid or he just skips to another move.
Will further comment my understanding of it if you know anything of this sort.
>>7919119
More likely than
>the most complicated thing we've ever seen came into existence through random chance
>>7919156
3.5 billion years is a long time
>>7919156
>random chance
In the same manner that large rocks don't fall through a fine meshed sieve.
>>7918373
>Go
i wish i knew how to play :3
>>7919190
>Systems as complex as the human body, with countless systems and subsystems all reliant on each other, can't have developed independently, no matter how many years of natural selection they go through.
Nobody thinks they did. If you think this the evolutionary account, then you're mistaken.
>>7918871
I am worried about a thing, nowadays people tend to overload themselves, they try to multitask, things seems to move toward being faster and faster because it seems to be what we want to achieve, Stock companies divided the work among more and more person, computers accounted for the stock market, but if we really desire to be "faster", should we not actively work to improve the human being as a whole?
>>7919190
>can't have developed independently
I agree that arms could not have developed independently from the rest of the body. But only a colossal idiot would ever suggest they were *hint*hint*hint*
>>7919036
I expected it within ten years of the rise of mogo. I was only off a few years if this game is the metric.
The game was truly amazing. The bot made an early attack that was so shocking and black could never turn it around. These UCT bots have an understanding of influence that will be extremely rewarding to learn from.
>>7919211
>But only a colossal idiot would ever suggest they were *hint*hint*hint*
Not really, the theory of evolution has gaping holes, only a fool doesn't recognize this.
>>7919222
let's not confuse theory and hypothesis, beside, Newton's law of universal gravitation is not the most accurate, but it was a step forward, then it came another refinement and so on and onto the future, you also are undermining anyone who abides to it calling them a fool, what kind of argument is that, I am not really discussing the merits of the theory right now, rather your methodology of discussion.
>>7919233
>you also are undermining anyone who abides to it calling them a fool
That's not what I said, evolution has very strong scientific evidence behind it, yet it also cannot explain how complex life came to exist. See, there's so many people with almost zero understanding of evolution who think it's the perfect answer to all of life's mysteries, I believe I'm correct to call these people idiots.
>>7919156
The least complicated thing also did
This is a consequence of different things existing. Truly amazing isn't it?
>>7919222
the only gaping hole is the locus of your butthurt.
>>7919238
>yet it also cannot explain how complex life came to exist
Please go on. Where is the explanation lacking?
>>7919238
>yet it also cannot explain how complex life came to exist.
wat
You know, if people just took a basic university evolution model and actually put the effort to understand it rather than refute it, we would be rid of this nonsense a long time ago. But noone wants to examine an argument rationally, let's rather refute it because we believe in a supreme creator whose existence is unverifiable. Cancerous.
>>7919222
Stop derailing the thread you fucking retard
>>7919279
What's the point in even using this fucking board if all people do is shit on me? Are my posts that fucking dumb?
>>7919287
/x/ will love you
Gently fuck off
>>7919287
Yes, they really are.
>>7919156
>evolution is random
End this meme. Mutations are random, natural selection is not.
>>7919287
Yes, they are. Take a course in evolution first if you don't understand it. Try and understand it. Gain a solid grasp, it's not a simple concept like others might tell you. If you do this without trying to be a special snowflake who denies evolution just for the sake of doing it, you will see that everyone in this board who is disagreeing with you might not be wrong, and instead the problem might lie with you.
>>7919287
You're just misinformed about current accepted theories.
To the people reading OP's headline and wondering what does it mean. I presume you already have experience with using computers and they can save you time and are essential for work and business:
>In the next few years if you want to learn something new, computers will help you learn it quickly and efficiently.
>In the next few decades if your job is not research or creative, it's very likely you will work for or under computers managing you and monitoring your efficiency, for political and social reasons they will not replace you (yet)
>A lot of jobs will disappear though, as service sector heavy lifting like cleaning, maintenance and delivery will be fully automated by robots and new sectors like aged and hospital care will be introducing robotic assistance.
>I expect by 2050 that the average person will no longer need to work as smart robots and machines become cheap and widespread and can look after most basic needs, freeing up the bulk of human labor.
>Economic restructuring will provide a Basic National Income to every citizen of a particular country, but jobs will still exist for those who wish to earn more money.
These are all very conservative estimates, since progress beyond 5 years is difficult to predict. So the main take away point is this:
>It seems 99% likely that by the end of this year 2016, no human player will be able to defeat a GO program.
>>7919356
Couldn't have made a more appropriate and accurate post myself. Kudos.
>>7918628
maybe you know this but from what i know, there is not really an end in a go match. The match ends, when one player resigns.
>>7919366
AlphaGo itself can surrender a position it considers lost.
And another common way for games to end is for both players to pass.
If one player passes the other can choose to play on, effectively by himself.
When there are no more valid open moves left on the board the game is also over.
>>7918675
The paper they published does have enough detail for people to make their own version, and I know at least three go program devs who dropped everything to try reproducing their results. One of them (David Fotland, Many Faces of Go) was watching this game at the Google Mountain View campus.
>>7919442
zen team is also working on enhancing zen with NN but they're a university group, not awash with cash like google
>>7918373
When it's in mainstream media,
>pic related
>>7919238
Humans have literally never built anything as complex as life. There is no evidence whatsoever that anything so complex can be assembled top-down. On the other hand, there is plentiful evidence that life has evolved continuously since it began and arrived at its present state of complexity that way.
So how did google improve so much since the last match?
Did they only used amateur training data for the first match and have then after the match against the European champion, used training data from professional world class matches?
And could they have used lee sedol matches only as a control, to adapt every parameter like aggressivness to a value which would have beaten lee sedol in every match he won so far?
Are you fucking retarded? You are all saying that ai will not exists because it will not be like us.
Creating something different is the point.
Alpha go did not made a single operation that was equal to one of a man playing go. Not one.
And the caster recognised in advance that the ai and sedol will do.
With 2 different aproches they had the same results.
And it will be the same for every ai in the future. They will do something humans can do. Just in a different way, and better.
No, they will not learn emotion, but they will be designed to relate to humans they will learn to emulate them. In a different way, and better that humans
>>7919971
We can build tools to study life.
Or our tools are complex, or life is simpler than you think.
>>7919999
the problem is when you are a pro, you are not creative. He basically thinks as a program, has memorized thousand of schema.
It only proves that he is not as good as a robot, not that the robot is better than the human.
>>7920013
The good ol' "if AI can do it, it's not AI anymore"
Give up, you know that this was ten years away ten months ago
>>7919993
They made the ai play against itself millions and millions of times, learning each time
>>7920013
And how do you think go player play?
They have an internal database too, they just don't know they are accessing it. In the same way alpha go has no idea how the data are stored in the cloud
>>7920027
You (human) can win with luck. AI can't.
>>7920048
What?
>>7920013
>the problem is when you are a pro, you are not creative.
lee sedol is one of the most creative players in go right now, renowned for his brilliant, uncompromising moves
>>7920048
You (You) are not making any sense.
>>7920013
That's not how Go works, there are far too many moves to be memorized.
>>7920033
[citation needed]
No seriously, please source that - it would be a great read!
>>7920106
No, they litteraly just "feel" moves are good, and in the late game they count like 20 moves haed. That is way seedol quitted, he knew every move from there to the end.
The feel part in done through experience, Aka data
>>7920137
He knew it was over because Go is a game of accumulating points, in evenly matched games between world-class pros it's pretty much impossible to win if you are at even a slight point disvantage in the late game.
You are thinking about it like chess, but it's not. Go has too many viable moves to count even for a computer.
>>7919221
Which attack are you referring to? One 3-dan amateur thought the turning point was the two stone cut and sacrifice around move 58, and Lee Sedol thought it was later playing Q5 instead of R4 on the lower right at move 123.
>>7918807
I actually don't think they will be retraining AlphaGo in between games with Sedol.
>>7919029
>I've been working on AI for nearly 8 years now, we are nowhere close to an AI with the creativity of a human being, the problems involved are enormous. I won't be so presumptuous to claim that we'll never get there but I'm certain we're not going to see it within our lifetime.
I agree, if you are looking only at deep learning. There is much more exciting AI research going on that has less popularity though.
>>7920106
There are some pretty standard "forms", the commentators even talked about them during the match.
>>7918887
>"I want to get gf"
>request timeout
>>7920368
>There is much more exciting AI research going on that has less popularity though.
Like what?
>>7920522
Mostly people who are trying to solve unsupervised learning. Its kind of related to the deep learning stuff going on, but all of the latest breakthrough rely heavily on supervised learning, which means labeled data. The true breakthroughs will be when we can learn from unlabeled data, and analogize what's been learned to other domains.
>>7919356
Awesome. Thanks, anon.
>>7918616
lol
Game 2 starting soon.
https://gaming.youtube.com/watch?v=l-GsfyVCBu0
>>7920347
The crosscut at move 28 was a big psychological blow. Black only seemed to turn it around but in my opinion, from playing MC bots a bit, alphago was already confident it was winning by something like 60%. It's hard to understand the depth, but to ignore the lower left star point approach to securely save its center group was a typical signal that it was, in its own opinion, ahead. And it was right.
>>7920535
Unsupervised learning is a red herring. Brains don't actually learn from unlabeled data, they learn from data that is labeled with an emotional state, which the brain itself generates.
>>7920864
dat amygdala
>>7918832
Well, as long as we can iron out the kill-all-humans tendency, I'm all for it.
This is getting intense
>>7918994
Meme'ing to the x^th degree
>>7920951
How so? Speaking as somebody who has no experience with Go.
>>7920959
it's not, it's a standard opening
Lee Sol left the game to wash his face, apparently the algorithm came up with an extremely surprising move which implies it's coming up with it's own moves and not copying data from former games.
AlphaGo today is playing much like yesterday. Sharp attacking openings (Sedol's trademark style), but it is playing more accurately than him and reading deeper. This is either a psychological attack by the DeepMind team, to mindfuck Sedol, or.....
A coincidence? Why would an AI try to clown the 2nd best pro in the world with his own signature playstyle?
>>7921037
I don't think the that the way the neural network "takes" decisions would allow such psychological strategy, does this sort of thing even happen in human matches?
It could be that according to the network, the safest moves against sharp attacks is also sharp attacks?
Did AlphaGo just blunder? That ladder seems lost.
>>7921037
They probably trained the neural net against every game Se-Dol has ever played. So the set of moves that AlphaGo "sees" at each step will likely include Se-Dol's. Then the search tree and evaluation probably predict further steps (read further, more accurately) than Se-Dol solely by virtue of it being on a computer.
>>7921058
Or maybe the best human player and the best AI both use the same best playing style and the guy you were responding to is just paranoid?
>>7921083
go, 21x21
then 23x23
then...
>>7921058
It's not just a neural net. I mean yes a deep convolutional neural network is one of the components but there are two more.
1) The NN gives a bunch of plausible moves as a sort of probability heat map.
2) An evaluation function evaluates the "value" of each of the high probability moves suggested by the NN.
3) A tree search begins. For each high value move given, AlphaGo pretends to be it's own opponent and runs the same process all over again to find the most likely response moves that the opponent will play (then for each of those the process repeats to search for how to respond).
As this is happening AlphaGo calculates the probability of winning, given a move, by combining the values of every path of that subtree (in the search space).
This is similar to how a human plays. The NN provides the intuition/gut feeling about which moves to explore. The evaluation gives a way of telling which moves are bad/good. Then the tree search is reading (hypothesizing about future moves).
A side effect mentioned earlier is that AlphaGo only cares about the probability of winning, not about the score. So it may at times make less profitable moves that a human player may consider a mistake solely because the less profitable move has less risk and is more likely to lead to a winning game.
>>7921091
another side effect is that in confusing areas of questionable value, the bot will play moves to make the area more clear and stabilize the evaluation
>>7921091
>An evaluation function evaluates the "value" of each of the high probability moves suggested by the NN.
The value function itself is a second NN
It's literally 2 neural nets connected to a partial montecarlo search.
>>7921097
Am I understanding correctly? One of the nets is trained towards finding the next move and the other one is trained towards telling if a certain configuration is winnable?
Which one is deep and or convolutional? Both?
Very...uh...interesting board configuration
>>7921100
One network suggests moves to explore with monte carlo search.
Then it searches out those moves in a tree and calls the value network which evaluates the result of that branch.
>Which one is deep and or convolutional? Both?
Not sure
>>7921106
Wow didn't even spot this when they did it...
Early Google Easter egg thanks.
>>7921106
coughed up some of my milk thanks
>>7921106
>playing a working ladder in a way that it fails
redmond pls
>>7921109
>One network suggests moves to explore with monte carlo search
That seems infeasible. Where are you getting that information?
>>7921122
I may be explaining that badly. Here is the paper you can read for yourself.
http://airesearch.com/wp-content/uploads/2016/01/deepmind-mastering-go.pdf
This is a pretty big move from white right?
>>7921136
this is his only chance to win the game yes, he is invading top.
currently its something like 75 points AlphaGo 55 points Sedol
>>7921136
Jesus christ man. Don't use toasters to take screenshots
Is it too soon to say that Lee is doing much better than the last match?
>>7921144
w struggling now, w will end in gote and black will save the top right corner, and the game is over
alphago 2 lee 0
>>7921146
>is it too soon to say something wrong?
Yes
>>7921146
To me it's over and he played worse than yesterday, when he at least made AlphaGo fight him. Today he just got outplayed globally and did not achieve anything locally. Black saving his left hand group showed AlphaGo simply reads positions better. A lot of inefficient moves by white to try and attack. Points wise it will probably be very close, just like yesterday, but move efficiency is definitely in the favour of the AI.
Now I'm not a pro player so I can't explain what Lee should have played instead, I can just look at the position and see he is clearly worse.
>>7921146
AlphaGo is currently winning
>>7918692
>playing rng-stone
>>7921162
It's fairly crude to say one or the other is "winning" right now, but I do think AlphaGo will probably end up winning
Time to get Ke Jie in there
>>7921181
Is that sudoku in korean?
>>7921192
Chinese 9p
>>7921161
So AlphaGo might be the best player in the world?
And lee was so certain he were going to smack the shit out of it.
>>7921181
Ke Jie is already there
[spoiler]you don't really think AlphaGo is a bot, do you?[/spoiler]
>>7921192
___________________
ke jie vs lee sedol
---------------------------------
11 | 4
>>7921192
no it's seppuku in korean
>>7921206
>>And lee was so certain he were going to smack the shit out of it.
Source please?
>>7921216
http://bigstory.ap.org/article/25b7778ca6f74a48a441504862e550b0/human-champion-certain-hell-beat-ai-ancient-chinese-game
>Lee is predicting a 5-0 or 4-1 victory in his favor.
Right before the first game yesterday though, during the press conference, he did take it back and say he felt it would be closer.
>>7921216
>“I have heard that Google DeepMind’s AI is surprisingly strong and getting stronger, but I am confident that I can win at least this time,” Sedol said in a statement.
http://venturebeat.com/2016/02/04/youtube-will-livestream-googles-ai-playing-go-superstar-lee-sedol-in-march/
does anyone have a link to a korean stream? i want to hear the koreans unleash fury
>>7921229
PLAYGOOOOO PLAYGOOOO ALPHAGOOOOOOOOOOOOOOO PLAYGOOOOOOOOOOOOOOOOOOOOOOOOO UAHRRRRRRRRRRRRRRAAAAAAAAAAHHHHHHHHHHHHHH!!!!!!!!
great outer pressure from black, white has to be careful to not overextend in the center
>>7921231
w-what?! do you know something i dont?
>>7921253
https://www.youtube.com/watch?v=Vpuv7VPb2rA
Am i the only one annoyed by the amateur host trying to make himself sound better than h really is?
>>hmmmm yeahhh
>>ohh right right
>>woow that move wass...(waits for help)
>>7921328
they all do that, if you watch NHK tournaments on youtube (erosbrown has a bunch uploaded) the host is just as lulzy
Actually Chris is probably not bad, he's AGA dan-tier, but he's under direction to tailor this shit to beginners and people who don't know the game, and he seems to have a mancrush on Redmond.
>>7921328
I firmly believe he does not know how to play this game
>>7921342
>Chris Garlock of the American Go E-Journal
https://www.newscientist.com/article/2079871-im-in-shock-how-an-ai-beat-the-worlds-best-human-at-go/
Wish I could find more info on him.
>>7921342
http://www.usgo.org/ratings-lookup-name?NamesLike=Garlock
aga 3 dan, so probably 1k KGS
For the more experienced: Is black in a bad place right now?
Is Alphago aware if it enters overtime itself and has to hurry up?
>>7921352
Yes
>>7921352
no
the team of developers making an AI whose sole purpose is to play Go, a game which commonly enters overtime, did not consider the situation that the bot might be pressed for time
Lee btfo'd himself
AlphaGo is significantly ahead of Lee
>>7919971
>Humans have literally never built anything as complex as life.
Humans have literally never built something as complex as a clock. Before they did. Or an internal combustion engine. Before they did. Or a computer. Before they did.
>>7921361
They should have considered it then, really weird that they didn't take that into account.
>>7921361
The thing must be a fucking cheat in overtime since it has perfect memory.
>>7921336
>mancrush on Redmond
Who wouldn't, he's so fucking dreamy.
>>7921382
they were being sarcastic you dipshit
AlphaGo made a mistake
>>7921387
I mean, that's as much an advantage as ever.
>>7921396
A lot of the time the "mistakes" actually aren't. AI play is weird, in that it would rather have a smaller lead if it gives a higher chance of winning over all, whereas human players try to get as much as a lead as possible.
>>7921403
this, when bot makes a "mistake" like this I am even more sure the bot has won
lee sedol has still some chance
>>7921396
AlphaGo plays the long game. It's playing a less risky and less profitable move in exchange for a riskier more profitable move.
>>7921393
So was he honey
>>7921413
So was she, obviously.
>>7921397
I'd guess there is a lot more to remember at the end of the game, but yeah, it's probably not much bigger an advantage than normal.
As the game gets closer and closer to the end, the search space should be smaller and we should expect it to converge on a decision faster and faster, right?
>>7921423
correct
AGA youtube chat looking like /pol/ right now lel
If AlphaGo wins all 5 games, will Lee Sedol quit?
>>7921429
he was talking about retirement for like a year now already, to promote go in the west, I think he has family in canada or something
LEE SEDOL RESIGNS!
2-0 ALPHAGO
>>7921443
bullshit, you faggot
This is disgustingly scary lmao. Poor Lee Sedol
>>7921444
Are you even watching the stream?
ION-RICH BIOLOGICAL ORGANISM WATER EXCRETIONS
>>7921447
So if it wins 3 in a row are they still gonna play a 4th or 5th game?
>>7921444
>>7921447
no need to be mean when you're wrong bro
>>7921451
think so
afaik he still gets paid for each game
I wonder if Lee still thinks he has a chance to win in the upcoming games
>>7921451
they did in october
It's over. Alphago won game 2.
Humans are finished.
What is Lee going to spend the rest day doing? Studying the AlphaGo algorithm? Try to come up with a new strategy? Ignore Go completely?
