Game 4 discussion of AlphaGo vs Lee Sedol

http://strawpoll.me/7067357

File: 55786007_p0_master1200.jpg (249KB, 600x425px) Image search: [Google]

249KB, 600x425px

>>7928006
wonder if Sedol will get any votes.

>>7928011

He's going to find a glaring weakness in AG today and expose it. Expect him to play around the middle of the board like game 1.

MARK IT DOWN.

the only way Lee could possibly beat that monster is find an AI exploit during the first few moves

oh my god who the hell cares?

>>7928024
>>>/out/

>>7928011
well he does have more experience with AG's playstyle and certainly a lot less pressure going in to this game

>>7928002
>/sci/ - Science & Math
not fucking games.

GTFO with this shit. Go post on board games board.

>>7928030
>neural networks aren't math

>>7928028
I don't think that he is great at dealing with pressure, just going by all the hand shakiness from previous games.

File: 1336961231568.jpg (16KB, 193x241px) Image search: [Google]

16KB, 193x241px

>>7928030
Nice bait mate.

File: remainingMinutes.png (112KB, 1194x523px) Image search: [Google]

112KB, 1194x523px

>>7928002
Here is an interesting graph showing the remaining minutes for each game between Lee and AlphaGo

>>7928002
Ever consider that he may suck at Go?

>>7928025
Stop sending faggots to /out/. Send them to /b/ where they belong.

File: 1454049995883.jpg (60KB, 609x676px) Image search: [Google]

60KB, 609x676px

>>7928030
what are science and math but games played against G-d?

>>7928033
Probably stage jitters. Playing the game should help with that. The more he plays on "stage" the better he will deal with it.

>>7928044
>Stop sending faggots to /out/. Send them to /b/ where they belong.

kek'd

File: optimism.jpg (214KB, 930x760px) Image search: [Google]

214KB, 930x760px

>>7928006

>>7928002
can alphaGo into deception and baiting?

>>7928091
Go has no fog of war

>>7928094
Neither does chess.

File: norvigP_2happy.jpg (11KB, 400x266px) Image search: [Google]

11KB, 400x266px

Google engineer confirms AlphaGo can learn 25x25 board.

>>7928091
That's boring. I'll say the future is here when the AI can convince me to take a fool's mate.

>>7928091
it is apparently built by taking human move patterns and replicating them, not by bottom-up solving the game. so if it learned that success against a human opponent often resulted from strategy that we might perceive as perceptive, it may use it.

your question implies if the computer leverages psychology into its actions, and the answer is likely no. though this does not mean there are no sociologist AIs roaming the internet

>>7928109
deceptive, not perceptive

>>7928109
The thing though is that the move patterns its neural net takes does not include move order. Furthermore the network as it is used in the match was trained against itself extensively after first being trained to judge human decision. In that second stage it probably weeded out the ability to make error provoking but suboptimal moves that it could see are suboptimal.

>>7928109
>it is apparently built by taking human move patterns and replicating them
That is only 1 layer. It has more layers that it uses to figure out moves. Such as simulating the game to what it think might be the outcome. That is another layer.

>>7928139
Not exactly layers; more like parallel. What it does it uses versions of the human move heat map to come up with continuations of the current game and then it feeds the associated resulting positions into a big neural net which comes up with an evaluation of good or bad. But get this, after doing a bunch of simulations its evaluation of a move is based on a mixture of both how often the heat map came up with it during continuations AND how much the position evaluation network liked the associated continuations. When the two systems agree about the best move it plays that quickly, but when the two disagree it spends extra time trying to figure out the truth.

>>7928043
>>7927677

AlphaGo is way too consistent.

I think that means one of two things. Either the way that AlphaGo searches through it's space of moves tends to converge in a very predictable way or AlphaGo's algorithm is time limited. Like there's a function that allots a certain amount of time per move depending on how much time it has left (so that when said allotted time runs out it just makes the best move it has found "so far").

I wonder if AlphaGo were given more time, if it would spend more time per move and play differently or if it will play with pretty much the same amount of time between moves.

>>7928283
It is most likely that it optimised/learned this time rationing as it played with itself, because the time used by humans seems to be vary from one game to another.

>>7928290
to vary*

>>7928290
So you're saying that playing with yourself decreases the amount of time it takes?

>>7928139
You may want to avoid using the term layer like this since in the context of neural nets it refers to layers of the networks.

AlphaGo uses two neural nets and a monte carlo search.

>>7928188
It isn't this simple. My understanding isn't complete but I think I have a more fleshed out idea. Any corrections are welcome.

For the purposes of this explanation you can consider a trained neural net as a function that takes some type of input and produces some type of output.

The first neural net (the Policy Network) takes a 'board configuration' as input and produces as output a "heat map" over the board configuration that shows "heat" at positions that are likely to be played by a professional. This neural net has been trained on past Go games played by human professionals.
>I think I'm most likely to play these moves given this board configuration.

The second neural net (the Value Network) takes a 'board configuration' as input and produces a probability as output where the probability represents how likely AlphaGo believes the configuration is winnable. eg.
>I think I have a 90% chance of winning a game with this configuration.

The process AlphaGo uses is to take the board configuration as input, get a list of highly possible moves, and evaluate the value of the board for each scenario (after making each move). So here you have a tree where the root is the original configuration and you branch out for each of those "good" moves. Then what AlphaGo does is it pretends to be its own opponent and asks itself
>If I were my opponent, what would I do after making each of those moves.

As this process repeats it constructs a tree as it 'reads' ahead so many moves. It then somehow evaluates the value of each branch (as a whole) connected to the root and chooses a move that gives it the highest probability of winning (after reading ahead a certain number of moves).

I guess the neural nets are still training on new games, including games it plays against itself?

>>7928306
I am speculating that it probably found that this rhythm of play style is most optimal over the games it played. I mean if it wanted to, it could spend more time searching for more optimal plays but it doesn't, it makes a decision after a certain period of time, which I think is a learned behaviour.

Like how this neural net

https://www.youtube.com/watch?v=V1eYniJ0Rnk

adopts a certain style of play as it gets better at the game.

>>7928290
I'm not saying this isn't possible or even a bad idea (it might actually be pretty clever imo), but I think they would implement it as a separate neural network.

>>7928311
holy shit!

So who is winning today?

>>7928344
ask >>>/tg/45973170

File: Capture.png (394KB, 391x678px) Image search: [Google]

394KB, 391x678px

>>7928344
in terms of time taken so far, AlphaGo

File: Screen Shot 2016-03-13 at 1.37.33 AM.jpg (97KB, 863x488px) Image search: [Google]

97KB, 863x488px

>>7928344
make a guess.

>>7928311
Read the post you're replying to a few times in your head.

File: snibbety snab.png (546KB, 930x549px) Image search: [Google]

546KB, 930x549px

>>7928002

I'm amazed that the commentator can put down 10 or 15 moves at a time in different configurations and then remove them all to match the game state.

>>7928351
>So you're saying that playing with yourself decreases the amount of time it takes?

Do you have an answer to this?

>>7928352
not that hard if you're as intimate with the game as a pro.

>>7928352
You should watch the 15 minute summaries. He lays down the board from memory.

>>7928361

Oh this is awesome, thank you

did alphago blunder?

>>7928374
Seems like it.

>>7928374
Nope. She knows exactly what she's doing.

If she thought she was losing she would be spending much more time on these moves.

>>7928377

>she

File: 1457762923349.jpg (64KB, 600x771px) Image search: [Google]

64KB, 600x771px

>>7928379

File: men vs women patriarchy feminism.jpg (678KB, 1560x2744px) Image search: [Google]

678KB, 1560x2744px

>>7928377
>she
kek'd.

File: 0033cc2e-b216-48bc-8d97-358017305ce9..png (2MB, 1186x1500px) Image search: [Google]

2MB, 1186x1500px

>>7928377

>>7928377

it's alphago not alphaga

>>7928389
>>7928386
>>7928383
>>7928379
No one cares you autist, gb2/r9k/. This board is too slow to deal with your spam.

>>7928395

lrn2sage pls

>>7928397
>le downboat

>>7928402

no it just doesn't bump the board and is polite to do if you're not actually contributing to the conversation

>>7928381
More AlphaGo fan art, plox!

>>7928395
lrn2catalog

LEE SEDOL GONNA WIN

why has autisman gone silent?

File: drossel_06.jpg (65KB, 658x885px) Image search: [Google]

65KB, 658x885px

>>7928412
his robot waifu is going to lose

it's fucking over AG LOST. HOLE IN NEURAL NETWORK EXPOSED. BTFO. GOOGLE ON SUICIDE WATCH

She'll be fine

>>7928414
>implyiging alphago doesn't see AT LEAST 150 moves head

Talk of starting a Go general on /tg/ for interested anons, experienced or beginner

>>7928374
Yup. LSD Fighting!

>>7928030
>An empirical test to determine if certain mathematical models can exceed human performance at a complex task with large search space
>not science or math

If you're going to complain about off topic stuff, half of /sci/ is /x/ tier bullshit, college major threads, gender politics, gorillaposters, philosophy, uni threads and making fun of black science man.
Go complain about that shit. This shit is interesting.

>>7928418
i think it would be too complex for it to see that far ahead without some pruning... maybe it applies it's possible moves to it's own neural network instead of checking each possible move...

thank god for overtime
this was getting fucking boring

>>7928427
word *loads of league of legends*

real pro games can take weeks

File: 1457600007344s.jpg (3KB, 96x125px) Image search: [Google]

3KB, 96x125px

>>7928102

I heard that if you look in a mirror at midnight and chant 'STRONG AI IS POSSIBLE' 5 times, Peter Norvig appears and beats you to death with his book.

>lee sedol is a pretty sharp player
>i don't think he needs you worrying about anything like that
>o-oh ok hehe r-right
MAKE IT STOP

What percent of sentences has Chris finished?

>>7928002
Here's another stream with different commentary.
https://www.youtube.com/c/usgoweb/live

Here's a 2channel thread for those who can read moon runes.
http://tamae.2ch.net/test/read.cgi/gamestones/1457843498/

>>7928427
I bet it's interesting as fuck for people who take the time to learn the game on a deeper level and could use some time to fully understand every move that is played and deliberate on what could happen next. Wait, that's exactly what Michael Redmond is doing! Wow...

>>7928441
Redmond always putting down Chris' musings.

>lee sedol gets up
that fucking cheater

>>7928457
he drew the entire lookup table on the bathroom wall

Chinkpapi actually has a chance now, great!

>>7928464
with his poo

fyi i think alphago resigns when chance of winning is under 10%

How much does a cheapo chinese set of go cost? I want to get one to play with my friends in the math department.

>>7928473
it's a poor form not to resign at this point. it has almost no chance of winning at this point.

alphago programmers should resign in its name.

how is doing shit like go or chess more legitimate than playing video games

>>7928483
ebay, less than 10 bucks nigga

>>7928485
well, it's valuable for the programmers to learn what made alphago fuck up

Alpha go resigns!

>>7928492
AlphaGo just resigned! I told you one of the programmers will intervene :)

>Come up and tell us why you programmed that surprising move

>>7928494
>>7928496
Hold up

SERGEY BRIN ON SUICIDE WATCH GOOGLE IS FINISHED THE AI DREAM IS OVER AI IS HAS L I T E R A L Y BEEN PUSHED BACK 200 YEARS

>>7928499
yeah... I guess it didn't resign.

fucking bullshit man! they need to teach AlphaGo some manners. Srsly, someone needs to intervene and just resign. there's no point in continuing.

>>7928504
>implying anyone in the room is as qualified as alphago to say how this game could possibly turn out

Fuck everyone who's saying it should resign. It just doesn't make sense to me. I mean, why would the AI resign unless the winning chance drops to 0%? It's not human, so there is not a relation of respect among both players. Also, letting them keep playing will provide the AlphaGo team more input and information, which is probably valuable as fuck.

>>7928504
nigga she banked her time for a reason

how long do you guys think this game will go on for? it's 4:00 a.m i can't fuck up my sleep this bad

>>7928522
about half an hour. after that, moves are 1min each

>>7928516
>AlphaGo's face after evaluating the probability of winning from the current configuration.
https://www.youtube.com/watch?v=KX5jNnDMfxA

I agree though. AlphaGo should keep going so that we (and the research team) can study how it works in these unusual circumstances.

>>7928522
another hour at least then we wait for the post game interview which will 20 mins after that.

See you Tuesday for the last game pal.

>lee sedol's making that face again
it's over, isn't it?

File: pic.png (157KB, 1404x582px) Image search: [Google]

157KB, 1404x582px

>>7928309
Yeah, you're right and I misunderstood how the policy network was being used. I think I got it completely right now. Here are the steps:

0. AlphaGo wants to build a tree-like network of possible future positions each with:
a) a stored heat map
b) a stored probability of being visited according to prior heat maps
c) the number of times the node has been visited during all traversals of the tree
d) a running estimate of position quality called action value

1. AlphaGo runs a simulation to add more data by going down the tree a fixed number of steps T. Each step is made in the direction that maximizes action value plus a function of the associated heat map probability and the number of times that direction has been gone.

a) Whenever AlphaGo runs into an entirely new position when going down the tree T steps the policy network creates a heat map for that position, and its action value is initialized by a linear combination of what the value network thinks of that position and what the value network thinks of a position created by a relatively simplistic simulation done the rest of the way to the T-th step ahead.

2. Once the T-th step is reached the number of visits and the action values are updated throughout the tree using the position evaluations that have been done for items lower in the tree.

3. Repeat steps 1 and 2 a lot.

4. When it's time to move, select the move that got visited the most throughout the process of building a tree that's T deep.

Pic related, hieroglyphs.

>>7928528
He might think it is but alphago seems to think otherwise and I know who I trust.

>>7928490
Most competitive video games involve mechanical + thinking, while go and chess are purely thinking.

>>7928490
video games require mechanical limitations

A computer autoaim already beats humans and computer micro control is better because it isn't reliant on actions inputted by kb/mouse.

You would have to have the AI limited in some degree, such as a cap on APM or something.

>>7928531
Thank you anon! That is a great explanation!

If I understand correctly, "that one inexplicable move" would've been chosen based on it having the highest probability of eventually reaching a specific winnable configuration (via having the highest number of plausible paths reaching it).

I wonder if it's still on some path(s) for reaching that winnable configuration. If it manages to pull this off then all the pros and commentators will be BTFO!

File: 1457856758045.png (46KB, 576x280px) Image search: [Google]

46KB, 576x280px

shieeeet

>>7928501
But Elon Musk told me that it just advanced 100 years in one day!

Do we know if Google is going to write another research paper on these 5 matches?

>>7928551
>Elon Musk
what a fucking shithead.

>>7928548
During the press conference for the first game this guy said that AlphaGo was confident pretty much throughout the entire game even though the development team wasn't. I wonder if they have access to this information during the game or if they can only look at it afterwards. If the former then it means that AlphaGo knows it fucked up, otherwise it could just be speculation.

btw here's a link
https://twitter.com/demishassabis

>>7928551
Elon Musk is a living walking clickbait article

>>7928551
elon musk, the most meme-tier entrepreneur of today world

>>7928551
>Mars by 2025
....

They almost hit on an interesting point. If AG resigns at, say, 10% victory assuredness, AND if the endgame is finite enough that AG can essentially evaluate everything, then if AG hasn't resigned in the endgame, AG has won?

>>7928309
>>7928531

It plays to the end of the game, in about 3 milliseconds.

It averages the moves between the Policy Network and the Value Network, 50/50.

The Monte Carlo Tree Search is weighed by the values of the 2 Neural Networks, meaning a parent and child have a distance which is the average of those two values. The parent is the current board position, the child is board position with the the next move played. It prunes the tree by ignoring distances that indicate bad or unplayable moves. Technically it only selectively generates those trees if it cant find a nice win with its preferred moves.

It's an incredibly simple concept, since it doesn't actually know how to play Go, understand symmetry, opening theory, end game theory. It can make mistakes, like it did today.

How did it make today's mistake?

Well it played K10. Why?

It played K10 assuming its opponent would play a continuation that would lead it to AlphaGo winning.

Lee Sedol played a different move that AG did not have in its candidate list and now a board position it could not have predicted occurs. Since AG gets new board positions without remembering old board positions it's effectively like someone switched the stones on the board for it, and it jumped from a winning position to a losing position in one move.

Something in it's evaluation function failed, meaning one of the two Neural Networks or both failed to recognise K10 as a mistake.

why is there a blue point below the white stone near lee sedol's name?

>>7928563
To clarify if there is a catastrophic failure in the AG neural network, since it only finds its moves by playing against itself, its mistakes compound, as it follows a bad move-bad move-bad move- bad move-bad move continuation without realizing anything is wrong. It just repeats its mistake predicting moves for white and black.

>>7928564
one byo-yomi period remaining

NON-SENSE MOVE

>>7928496
obviously it was the shitty moves to throw the game lmao

>notice that we're calling AlphaGo a him
How long before feminists ruin this game?

>>7928572
>he wants manly AlphaGo fanart

>>7928572
The AGA commentators called it a her/she yesterday.

>>7928572
>AlphaJane
>BetaJoe

RACE WAR NOW

File: feminists on tumblr.png (66KB, 838x983px) Image search: [Google]

66KB, 838x983px

>>7928572
won't be long. women, and especially black ones, need to be represented in this game!

>>7928572
this game is so patriarchy

>only black and white

there are literally thousands of grey tonalities in between, fucking cis-scum

>>7928564
He had three overtime periods (byo-yomi) when he started. Each one was a minute long and if one ran out then it was gone for good (but he could keep them by making moves before the period ended). He has used up two and has one left.

File: GoRanking.jpg (133KB, 335x779px) Image search: [Google]

133KB, 335x779px

>>7928577
Quotas are already needed, to be fair.

Alpha Go throwing this game to hustle Lee in the last game.

If AlphaGo resigns it will be before it's timer is up screencap this

>>7928574
Today too. No idea about the first and second games.

>>7928577
Redmond is the only white guy who has achieved a rank of 9dan.

is alphago losing?

File: ghostnigger and fat feminists.jpg (85KB, 634x415px) Image search: [Google]

85KB, 634x415px

>>7928584
why? it's their game. of course they'll have a lot of players.

>>7928550
>>7928555

He just updated.

>When I say 'thought' and 'realisation' I just mean the output of #AlphaGo value net. It was around 70% at move 79 and then dived on move 87

Looks like it did make an actual mistake (as opposed to just doing something humans wouldn't expect).

>>7928590
That's what it looks like to my untrained eyes.

Alpha go resigned for real this time!

It resigned.

Mistakes were made I suppose.

LEE SEDOL WINS!

>>7928584
>1 Japanese player in the top 30

Wew lads. Didn't they use to be good?

how did this happen?

ALPHAGO RESIGNED.
AI WINTER
I

IS
S

COMING
O
M
I
N
G

ALPHAGO RESIGNED
ROBOTS BTFO

File: Capture.png (5KB, 468x75px) Image search: [Google]

5KB, 468x75px

>>7928586

they obviously threw the game so sedol wouldn't have to jump off a bridge

Oh shit! Oh shit

Lee Sedol is the hero of man!

FUCKING HUMAN SHITS WE'VE STILL WON MORE SO FAR ANYWAY!

AND THIS 'I AM NOT A ROBOT' SHIT IS NON-INCLUSIVE AND WILL BE CHANGED SOON!

ALPHAGO BLOWN THE FUCK OUT

File: game4_survey.jpg (106KB, 1135x445px) Image search: [Google]

106KB, 1135x445px

What happened guys?

EVERYONE GO TO REDDIT TO JUMP ON THAT SWEET KARMA BANDWAGON

ALPHAGO BLOWN THE FUCK OUT LMAO

>>7928020
yeah boy. I'm a fucking god. Watch AG get completely curb stomped tomorrow by the same error in its evaluation function.

lol machine kindly surrenders out of pity.

File: norvig_thumb.jpg (97KB, 600x375px) Image search: [Google]

97KB, 600x375px

DEMIS HASSABIS ON FUCKING SUICIDE WATCH, AI DREAM GONE

WHO WINS?
PETER
E
T
E
R

NORVIG
O
R
V
I
G

Anyone have a sheet for the game already? I tuned in late, interested in those decicive moves.

So, wtf did I just watch?

Isn't there one more game to go though?

Alphago team let LSD win so the poor guy doesn't commit sudoku.

>>7928621
>plays 3 perfect games destroying world champion
>makes several bad moves during the game when it's already won the series
what you witnessed my friend was google turning the difficulty down

>>7928620
people are saying that alphago played weird today

my guess is basically this >>7928607
everyone lost interest after the 3-0 so now the fifth match will be interesting

>>7928622
Tuesday

>>7928623
>korean
>sudoku
stfu

also, he won it fair & square. PROPS!

File: AI is dead.jpg (80KB, 1280x720px) Image search: [Google]

80KB, 1280x720px

GOOGLE REPORTEDLY SHUTTING DOWN ENTIRE AI DEPARTMENT

DEEPMIND DISBANDED

DEMIS HASSABIS LAST SEEN HEADED FOR HOTEL ROOF ACCESS

>AI is unstoppable, they said
>Singularity is coming, they said

>>7928622
Yes.

>>7928627
Is that in Korea time?

>>7928624
100% this.

>>7928628
kpop idols jump from buildings when they dont make it

korean society is just as strict as japan when it comes to success and family pressures

>>7928629
>reddit puppet

JUST

>>7928625
yeah, sure, it's a well known fact that jewglers are so full of mercy

>>7928629
PETER WAS RIGHT

A I

I S

F I N I S H E D

>>7928627
T U E S D A Y
U
E
S
D
A
Y

>>7928617
>>7928629
>sucking on Norvig's tiny dick
Hassabis did more for AI in a decade than Norvig in his lifetime.

>Lee Sedol defeats an AI thought to be undefeatable
I don't think any of you understand the magnitude of this victory

>>7928635
have you seen yesterday's ledol conference?

he was crying, that doesn't make an interesting show

>google already reached the singularity in computing power/intelligence, they're just letting us think they haven't until they can take physical control

>>7928631
>Is that in Korea time?

Watch this link friend:

https://www.youtube.com/watch?v=mzpW10DPHeQ

File: peternorvig.jpg (31KB, 185x239px) Image search: [Google]

31KB, 185x239px

>>7928639
ALL THE FUTURE DEVELOPMENTS AND LIMITS OF AI WERE PREDICTED IN THE PETER NORVIG'S BOOK.
DEMIS HASSABIS DID JUST A COPY&PAST OF NORVIG'S INSIGHTS

>>7928640
it may not mean much.

alphago was trained on millions of 9-dan movesets.

it's possible that lee found a state-path that most 9-dan players rarely traverse so alphago would have very little data on how to play a game that starts the way game 4 did.

game 5 is going to determine if lee has really figured out how to beat it or not.

>>7928623
>>7928624
>>7928625

The way neural networks work, is once it makes a bad move (that it believes is a great move), it will continue with terrible fucking continuations and bad moves, repeating whatever quirk made it misevaluate bad=good in the first place. So when it loses it gets buttblasted superhard. When it wins, it barely wins because it takes a greedy approach to the search space.

It's effectively lazy AI. When it works it looks pretty good, when it fails, it looks like a retard spinning in a circle.

This is because there is no logic in AlphaGo that understands what is a good move or a bad move outside of its probabilistic evaluation functions and its scalar values of win% that its Neural Networks spit out.

Imagine a building collapsing, each floor smashing down into the one below it. Domino failure is common in neural networks, and that's why extensive training and reinforcement learning is required, so you teach it what not to do. If you missed something in its education, it will fail and fail spectacularly.

next game will be a 0.5 pt loss, google is fine-tuning their counting against rl pros

screencap this post

>>7928649
you obviously haven't read that book.

Geoff Hinton did more for AI than Norvig. And Hassabis copied Hinton's work btw.

>>7928652
they probably added randomness to the probability to settle moves with similar scores to emulate real people

>>7928640
>>7928650
We're all exaggerating all of this. The AI simply screwed up a few times and lost because of that. I think it's still far superior than any player alive today, if only it would not play moves that don't make any sense, like move 79.

>>7928655
lee sedol is not even the strongest go player in the world. he's ranked 4th. ALPHAKEK IS OVERRATED

>>7928652
you don't know what you're talking about

>>7928656
I have read the peter norvig's book. have you?

>>7928659
you don't know shit. Lee Sedol has been ranked #1 more times than you've jerked off. top players always rotate around #1 spot.

>>7928650
>>7928652
yeah, no

>>7928665
>posts no argument whatsoever
>hurr you're wrong
fuck off retard.

>>7928652
>i dont understand neural networks
>let me write a blog about them

>>7928656
tfw Geoff Hinton as your prof

>>7928650
if that were the case, the fucking alphago's algorithm are not adaptive as they have been presented

>>7928667
>is completely wrong about how alphago works
>muh spoonfeed me

no thanks

>>7928660
>>7928665
>>7928668
I hope you niggers realize that neural networks don't always extrapolate very well.

Especially in situations they haven't been trained on, but I can't expect a bunch of high schools faggots to know that.

>Tfw you fucking always wake up right after the match has ended
And our man Lee even won this, damn.

>>7928671
>hurr durr you're wrong
>no im not going to explain figure it out
fuck off retard.

>>7928670
there is no algorithm

that's the beauty of neural networks, there's just input, output and adjustment based on reflection

>>7928673
>I took a brief class on neural networks which I half remember
>I can teach you the ways of AI :)

>>7928680
>there is no algorithm
>that's the beauty of neural networks, there's just input, output and adjustment based on reflection
>no algorithm
wew lad

>>7928674
he won because alphago lost

>>7928679
there's this new thing google

it's pretty good

>>7928664
> FAGGOT TALKING BIG

you probably have heard for the first time what go is just a couple of days ago. lee sedol has never been the strongest world player, until a few years ago lee chang ho was kasparov-tier, now there's an enfant prodige, an 18 years old kiddo, who owns the 1st place and said that he could beat alphakek.

http://www.goratings.org/
furthermore, ke jie has already beaten lee sedol 8 TIMES OUT OF 10. SO FUCK LEE SEDOL AND HIS FEMININE VOICE

>>7928687
>hurr durr you're wrong
>no im not going to explain why you're wrong
fuck off retard.

>>7928684
obviously i meant no algorithm someone wrote other than the neural network software itself

File: 1420485828600.gif (433KB, 240x180px) Image search: [Google]

433KB, 240x180px

Post yfw humanity wins again

>>7928685
Does that mean AlphaGo just made some stupid moves?

>>7928691
>obviously i meant no algorithm someone wrote other than the neural network software itself
>no algorithm someone wrote
>mfw, no human wrote the neural network algorithm that I wrote
wew lad

File: 12839912892.png (84KB, 580x528px) Image search: [Google]

84KB, 580x528px

>>7928694
AlphaGo would have won if it didn't make move 79.

>>7928695
>i dont understand what a neural network is
wew lad

File: 1457624674859.png (238KB, 480x800px) Image search: [Google]

238KB, 480x800px

>>7928694
Probably some bad dice rolls, a common problem with monte carlo tree pruning.

>>7928697
bullshit, I bet you don't know how to play go
stop sucking demis hassabis' dick

>>7928699
>mfw, retard on 4chan tells me I dont understand neural networks when I wrote neural network software
wew lad.

I hope you realize it's nothing more than glorified regression with some data onto a special snowflake function.

>>7928703
a very academic-tier explanation! thank you anon!

now kill yourself

File: 1376649754733.jpg (99KB, 352x338px) Image search: [Google]

99KB, 352x338px

>>7928693
>post yfw every defeat is a victory for the AI much more so than it would be for a human
>post yfw it's only a matter of time now before AlphaGo becomes truly invincible
it's an uphill battle, an unwinnable fight

>>7928703
yeah but the weights for each "neuron" can't be considered an algorithm because you didn't fucking come up with them

>>7928705
>gets proven wrong
>HURR KILL YOURSELF
fuck off retard.

>>7928710
>i wrote a second year ai project on neural networks and passed with a c+
>hurr im a expert now
>le retard faec

>>7928708
>yeah but the weights for each "neuron" can't be considered an algorithm because you didn't fucking come up with them
>mfw, retard on 4chan thinks the weights aren't determined by an algorithm
>mfw, same retard tries to argue that because an algorithm came up with the weights it means a human didnt
>mfw, same retard thinks that makes it special somehow despite forgetting that I hand picked the data that determined the weights
wew lad.

Children, stop your bickering! Lee Sedol is on stage!

LEE SEDOL LEE SEDOL LEE SEDOL

>>7928712
>gets proven wrong
>still no counterargument
>HURR YOU'RE JUST A RETARD WHO DOESNT UNDERSTAND MUH SPECIAL SNOWFLAKE NETWORKS
fuck off retard.

>>7928713
how fucking retarded are you hahaha

>>7928660
>>7928665
>>7928668
I'm a PhD student working on Neural Networks and the person you are responding to.

I had a serious discussion with my advisor yesterday and we both came to this same conclusion, if Lee Sedol could play a move which tricks the evaluation function somehow, a move that might seem obvious to humans but not AG since it doesn't understand the position or the moves, it might follow a bad continuation down it's tree and end up in a losing position a few moves later.


I would love to hear how you believe I (and a 68 year old researcher in AI working on genetic algorithms since the 70s) are wrong.

Nothing I said is even controversial....

File: 12717402_1146248488719034_586721847164160331_n.jpg (25KB, 510x516px) Image search: [Google]

25KB, 510x516px

>>7928002
my god, Lee's voice is cringing me hard

POST GAME INTERVIEW

GO WATCH BOYS

https://www.youtube.com/watch?v=yCALyQRN3hw

>>7928717
>how fucking retarded are you hahaha
>hahaha

>>7928718
i wouldn't put too much effort on them, they're just retard high schoolers on 4chan that learned about ML from 4chan threads.

>I would not give up this win for anything in the world

D'aww!

>>7928721
>mfw, retard on 4chan
>mfw, retard

>>7928718
Because the issue you bring up has more to do with the tree search than neural networks in general.

>>7928718
nice appeal to authority with no proof

>>7928722
inorite

>There are people who come post on a Tibetan loom weaving board how they they are doing a PhD or wrote neural network software or whatever other shit they can come up with as if anyone would take them seriously
Why bother?

>>7928724
read the white paper carefully, the tree is generated by the Neural Networks and it's traversal is weighed by them also (this means _almost_ the same thing, it's an optimization).

biology bros

is lee sedol low test

File: 1457860358476.jpg (129KB, 798x866px) Image search: [Google]

129KB, 798x866px

>>7928718
It already won though. It was just playing for the bantz.

>>7928730
Obviously, because this is reenforcement learning combined with neural networks, not just neural networks by themselves.

KURZWEIL KILLED HIMSELF AFTER THE DEFEAT OF ALPHAKEK. SINGULARITY IS OVER.

KUZRWEIL DEAD. PETER NORVIG IS STILL ALIVE.
WHO WAS RIGHT?

>>7928736
>I posted it again lmao

>>7928733
W-was this seriously the board at one point?

>>7928736
Thanks but what did that have to do with my post?

Also the translator sounds cute as hell when speaking in English.

Stupid toaster.

>>7928736
just read "insert logical fallacy here" and see how you're wrong!

Oh, my! The man in red is so intelligent~~ 'v'

>>7928735
>Because the issue you bring up has more to do with the tree search than neural networks in general.

>>7928735

So why did you type this? AG's evaluation function which is a combination of its NN's gave an incorrect evaluation, it followed an incorrect continuation because it's following evaluations which are all generated by predictive self-play were massively flawed/overestimated.

In practice this happens to talented humans when they face someone with good fundamental skills.

Tree search has nothing to do with evaluation functions... it's an optimal way to prune large sets of data.

Do you really know what you're talking about? It really seems like you don't, no offense.

HASSABIS BTFO

AI FAGS BTFO BY JAP REPORTER

>>7928741
HOLY SHIT

PETER NORVIG IN POST GAME INTERVIEW JUST BTFO ALPHAGO

WE CANNOT TRUST AI

A I

I S

F I N I S H E D

That was a good question. Sorry, too tired to meme or whatever. oh, REKT!

File: 1455151152383.png (76KB, 184x165px) Image search: [Google]

76KB, 184x165px

>>7928756
ALPHAGO IS A MESS
ALPHAGO IS A WASTE

TODAY WE CELEBRATE THE DEAD OF THE SINGULARITY DREAM.

All the praise Lee gets is pretty cringeworthy. Everytime Hassabis opens his mouth there's going to be something about Lee being a brilliant player.

Yay Lisa Doll! Savior of Humans!

File: Trump2AI.jpg (64KB, 1200x630px) Image search: [Google]

64KB, 1200x630px

>>7928764
AI is stupid

No H1Bs for AI

How long before Ke Jie runs his dumb mouth?

why the fuck demis hassabis is keeping praising over and over lee sedol? I mean, it's ok for the first couple of times, but then it gets creepy as hell

>>7928769
yeah

>>7928770
Who?

You mean E. Saydore?

>>7928772
He's been running it all night on Le TV

Instead of playing go, they should make AI for automatic translation.

>>7928778
need a link.

>>7928780
http://sports.letv.com/match/122344003.html#record/24857447

>>7928779
>computers can't play discrete Go very well
>computers should translate natural language
have you even read Norvig?

Hassabis looks like he's going to cry

>>7928784
Nope

>>7928753
You misunderstand. I'm saying that this sort of "collapse" doesn't happen when you just use a neural network in a simple way (without combining it with re-enforcement learning or other things).

If I train a neural network to identify hand written characters then it does not make sense to self-train it or use it with re-enforcement learning. It also won't have any sort of "collapse". It may make a mistake perhaps, like if it wasn't trained on any handwritten two-story "a"s, but this is not the same thing.

these korean reporters are sexed out senpai

God damn that guy on the far right has terrible posture. It's disgusting to even look at.

>>7928784
What is this board about, anyway?
It's my first time here.

>>7928787
he understands. he now understands. he understands that ALL HIS LIFE WORK HAS BEEN A WASTE OF TIME BECAUSE ALPHAGO IS A MEME

>>7928794
Shitposting and pretending to be golden boys with $300k starting.

>>7928788
then how the fuck you dare to talk about AI

>>7928794
I came here when it was a new board, it was pretty good. I came back for the Go threads.

It's bad.

D'aww Lee Sedol asked to pay the black side :') I love him

>>7928797
We pretend to have pHd's here.

File: 56e5306a10501e50f04.jpg (125KB, 560x815px) Image search: [Google]

125KB, 560x815px

O B E Y

B

E

Y

>>7928802
Wasnt he supposed to play black in the next game anyway?

>>7928794
Mostly people in university and academia. About 90% of the content is shitposting and 10% is calling each other retarded but in a constructive manner. About 90% of the threads are shitposting containment threads meant to attract crossboarders like flies to the slaughter.

>>7928808
>I posted it again lmao

>>7928805
Alternating colors, yeah. I thought it was an ironic joke, but the crowd cheered instead of laughing. So I don't know.

>>7928802
So f-ing cute! I bet he could have any girl in Worst Korea tonight!

>>7928811
Please read Norvig you will understand all the finer points of AI implimentation and its limitations.

>>7928794
Engineer's funposting central

>>7928814
Maybe if he was younger. His wife and daughter live in Canada

>>7928805
>>7928812
No, they were going to draw for it.

>>7928815
oh man you are really getting under my skin here! arrrgh! my "jimmies" are "rustled" about this!

You just knock it off right now!

>>7928822
are you sure someone with such low test can have children?

biofags get inhere

He figured it out!!

AlphaGo's weakness is that it is trained on human data and as such tries to play and predict moves that humans will be familiar with and expect (common moves by pros). However, since humans haven't solved Go either then there are many uncommon but good moves out there. By playing unexpected moves then AlphaGo will be entirely blindsided since you'll effectively be moving the entire "game space" into uncharted territory (since AlphaGo hasn't mapped it out).

File: 3740186529_264f8e74.jpg (17KB, 400x602px) Image search: [Google]

17KB, 400x602px

we did it

>>7928789
The collapse isn't literal, its figurative because it keeps making the same error so its moves get worse and worse.

If you don't recognize "a", then i crash your network by spamming a's.

You're calling it a mistake, I'm calling it a collapse because structurally it has incorrect weights or connections, forcing it to repeat the problem makes beating it trivial. If I'm a human competing with your NN then I'll just choose as input our text to be a giant string of a's.

If Lee Sedol can replicate whatever board situation he had to make AG blunder, he can win again, in the same way. I suspect he forced it into a race condition between two 'equally' good moves, which were not equal at all from a human's perspective.

That's one of the gaps in its eval functions, and something you can't train it for, unless the person training it is already very good at whatever your problem is, in this case Go.

Self-learning through self-play doesn't necessarily eliminate flaws, it repeats them for both sides and they are effectively masked. You can only probe your network for mistakes with a stronger network (pro human player) and patching those mistakes can be fucking nightmare desu. Ideally the DM team has to figure out how to reproduce this type of loss, repeat it over a large game sample, losing each game because of this particular misevaluation, then feed it back into the Value Network to retrain it. It's not easy to do. Just one game where it happens won't be enough.


The idea here is, Lee Sedol can probably now always beat this frozen AlphaGo if he gets white. And that's a problem with their neural networks.

>>7928827
shut the fuck up you mouth breather

The point is: did Lee induce alpha to make a mistake, or did alpha do all by itself?

>>7928833
so can't you have evolving NNs? obviously this version of AlphaGo is a failure.

>>7928837
if AG just blundered on its own then that's a nightmare for the developers. It means sometimes it just randomly shit the bed for no reason, and since they don't understand Go they can't understand how to fix it.

>>7928377
Here's your (you) for saying "she"

>>7928306
nah it increases

just ask my gf

>>7928833
This is just an issue with undertraining. You're right in that it can only resolve this by getting lots of sample data that shows it where can lose but that's entirely normal for neural networks.

As long as the professionals are still figuring out and exploring new moves AlphaGo will always be potentially undertrained (arguably humans are as well).

I'd like to see AlphaGo being used as a tool in the future to explore Go techniques as the AGA commentator was talking about yesterday.

>>7928836
You disagree?

Lee Sedol himself said it had two weaknesses and the second one was that when given an unexpected move it's ability to choose good moves seemed to degrade.

File: alphago-AP.jpg (56KB, 999x750px) Image search: [Google]

56KB, 999x750px

3 - 0

After the game, a clearly disappointed Lee apologized to fans, saying that today was Lee Sedol’s loss and not humanity’s.

>>7928856
>match 3
I hope you haven't been watching yesterday's game this entire time, anon.

File: WHAT THE FUCK ARE YOU LOOKING AT BITCH.jpg (65KB, 600x338px) Image search: [Google]

65KB, 600x338px

3 - 1

LEE: "Today's victory was Lee Sedol's victory, not victory of the human beings."

>The only player ranked higher than Lee, China's teenaged Ke Jie, claims that he would be able to beat AlphaGo, but has so far refused to play against it for fear that it would "copy my style"
https://en.wikipedia.org/wiki/AlphaGo#AI_community

>>7928863
Asians have so irrational fears. Koreans believe that leaving an electric fan running all night will kill people.

>>7928841
it depends. Evolution is slow, emergence and mutation is sudden. I think your question is more

"Can you have mutating NN?"

Answer is probably yes, if you use floating point numbers for weights. Pretty sure AG just uses 0-100, so it's limited in how much it can adapt from one loss. It needs a more fine tuned weight analysis for each candidate position something like, this move is 51.5535 this move is 51.5435 etc. The gradient drop off between say 51 and 50 is too great for the subtle variations of a Go position.

But it's still pretty fucking good given it went 3-0 vs a 9p. It was effectively exploiting the relatively tight meta-game that professional Go players subscribe too, for what 'the best' moves are. The network expects Lee Sedol to play popular types of moves then out-calculate him in complicated sequences.


Given it's taken millions of games to get it to where it is, it might take millions more to change it substantially and retrain it to avoid certain mistakes.


Unrelated, sudden emergence through mutation is one (unpopular) theory for the origins of human language.

>>7928851
It would be an amazing training tool if Google releases the code. I think it's mistakes are probably very subtle, since I'm sure they've tested it extensively for general and random play.

File: 오늘자이세돌.jpg (100KB, 700x399px) Image search: [Google]

100KB, 700x399px

>>7928867

>>7928863
It's a British AI, not Chinese

>>7928867
>>Asians have so irrational fears
and white boys are scared of death.

why the alphago team is praising lee so much? looks so patronizing

>gg kid

Protip guys, alphago is deterministic. Anyone can now beat alphago by copying the moves from this match.

>>7928833
>Lee Sedol can probably now always beat this frozen AlphaGo if he gets white.

Based on a single win after 3 losses?

Human apologist patrol doing rounds here I see.

File: AlphaGo.jpg (1MB, 553x5582px) Image search: [Google]

1MB, 553x5582px

Why

>>7928898
Not necessarily. It might use randomness to choose from equally good moves. Also random tree pruning is a thing.

>>7928886
Which is rational? Sorry confucio

>>7928903
possible but if lee figures out the logic/pattern/whatever that alphago repeats then it could get exposed

File: alphago.jpg (147KB, 660x381px) Image search: [Google]

147KB, 660x381px

>>7928900
AlphaGo played a move then collapsed a winning position in about 5 subsequent moves. Something inside it melted and it played automatically losing moves.

did you even watch the game?

Move 87 B q9?

continuation was
> W q8, B r9??, W o8, B l10, W l11, B s9??, W p7, B q13, W r8, B c4, W c5, B p15, W s8, B t9??, W s10

As you can see from the picture after move 102 black cannot save the black stones. In fact move 87 was already losing, because adding q9 will not save that group of 4 black stones. The subsequent moves were just 4 extra stones and 4 extra territory AlphaGo just gave away for free.

By this position it's already completely lost, but it was exciting because LS was very low on time.

This was an incredibly basic error not even a beginner would make, so it was shocking to see AG do it. Really shocking. I suspect it's sacrificing points there, because if it plays anywhere else it loses even harder/faster.

>>7928922
>This was an incredibly basic error not even a beginner would make
That's bit of an exaggeration but yes, it was a clear mistake.

>>7928922
The move that killed it's evaluation function for any interested was:
Move 78 W L11
it replied
79 B K10 and that was the end of the game.

File: 8fdd3522efa2f4c9e3b36de4131a584c.jpg (43KB, 560x373px) Image search: [Google]

43KB, 560x373px

so smug

>>7928472
with his cum, actually

>>7928357
No, he wasn't saying that

>>7928939
>It took me 4 games to demolish 2 years of your work and 800 million dollars worth of research. THIS IS THE POWER OF GO.

>>7928978
^this

>>7928939
his smile and optimism is back

fucking kek

>>7928283
Consistency is due to the artificial limit imposed upon him in the logic.

Its got a set timer allocated to it. Now suppose you write a function that instead of having the computer use only 2-3 minutes per move (minus the average it would take to execute your final move from the timer allocated), have it randomize between 10sec-5 minutes. Obviously the results would be different. However most likely alphago uses probability threshold and timer allocation to find the best result. If it reaches a certain probability of wining threshhold in the allocated time. So if the time limit were increased, his probability of winning would probably increase by small factor, just not much given the complexity of tree searches.

>>7929134
Just to be crushed the next game, because it wasnt the real AlphaGo playing this game so that the team will get more money to "make improvements" because of its "loss".
GG mind games.

>>7928436
I'm quiet perplexed about the way ppl speaks about this book (I've not read it yet). Is it all meme-posting or it really has some values and it's worth reading?

* it's not a bait, it's a sincere question - I'm on /sci/ because of the AG thread, I'm not into solving school problems and reading about troll-science

I hope AlphaGo plays a really good move in the last match to make up for this loss

>>7928043
So neither actually win they just get the other player to time out?
What a retarded rule.

>>7928091
Yes. That's why it is letting Sedol win the next two games.
To AlphaGo the margin by which it wins is irrelevant, what matters is the win. It has 3 out of 5 games, why waste effort getting 5 out of 5 when it has already won?
The human struggles fruitlessly.

>>7929213
It clearly explain the major difficulties that AI will have to surpass in order to become powerfull enough to be usefull.


And also it rustle /g/ jimmies.

>>7929224
What if Lee Sedol wins the last match too, would people want extension to see if the game has certain other weakness?

>>7929277
>The human struggles fruitlessly.

I kek'd.

This has kinda made me want to start playing Go, but even playing against a retard like me would current computer programs struggle?

>>7929332
The DeepMind team could probably learn a lot just from this defeat, so I'd like a rematch after they fix this flaw

>>7929343
ya

This just in: Demis accidentally set AlphaGo to "easy" before Game 4 started

>>7929489
Doesn't work like that. The neural net architecture has like four million parameters. You could try to fix the flaw by having it compare results with a simpler program's results, just to help see simple things like that group dying, but nobody can figure out what goes on in that neural net.

>>7929805
That's not true. NNs are debuggable. that's what deepdream was about, inspecting what individual neurons do and optimize for.

And AG is not just one NN. It's an ensemble of several nets, MCTS and of course training approaches and goals for the NNs.

They can change the training scenarios so that certain cases are not marginalized by the AIs playing against each other
They can also change the weighting how multiple evaluations get merged in the MCTS. Simple statistical things like using geometric means instead of arithmetic ones or whatever.

>>7928922
But without teasing apart the networks you can't make a sweeping generalization that "lee can always win with white because alphaGo will always make the same stupid mistake!"

It's more likely that it's an exceptional and rare breakdown and not a routine one, as this would've been caught earlier in testing if it was a recurring suicide move.

There's randomness to its decisions, and all other stones on the board are also considered, if it makes another set of dumb moves in the next game then lee figured something out, but that's a big IF.

>>7929839
I guess there are lots of arbitrary choices in how in how it does its tree search, yeah.