I started watching the stream too late this time, and the commentary team was already saying AlphaGo was going to lose. That meant there wasn't the drama of the previous matches for me, and I didn't see the moment where they could see the tide turn or Lee get ahead. I hope to read some analysts later to see what was really going on.
I'm glad we got to see what it looks like when AlphaGo loses. We all know it chooses moves with the highest probability of winning, but it still doesn't have a solution for 100% chance, if there is such a thing. If it takes a pet of 90% victory, you jus tha even to find one of the 10% of sequences it's hoping you don't pick. And it's clearly betting you make a move with a high probability of winning for you.
What was most interesting was before it resigned it made some apparently dumb moves. Some were so dumb I could even see how bad they were. But really they weren't bad. They gave AlphaGo the best chance to win. Of course, they only had a chance if the opponent was so dumb they didn't play the obvious counter. Only after attempting, and failing, at all of those moves did it resign.
Sorry I'm late to the party. I picked up the subject after the first game and have been following since (a day behind or so). This is so awesome. In my opinion, the most incredible time to pay attention to a perfect information game is at this window. I love that Lee won a game. I am, admittedly, far too poorly informed on the subject. I am very dependent on the English language commentator to understand what is going on. I do, often, run into people saying slightly inaccurate things about AI and computation day-to-day regarding this series of events... and I try to correct them... but either way... this is so fun to watch.
I do not understand enough about high level Go play to be able to interpret much of anything it's doing without the commentary ;^)
The same is true for me, except for the extremely bad moves it made at the very end of game four. I played Go on an iPad for a few days way back. That's all you need to see how bad some of those moves are. They were like, onside kicks, but kicked so poorly such that only if the entire other team were unconscious could you recover them for a touchdown.
Oh, I also want to mention. The commentary guy who ISN'T the Go Master, how did he get that job? I mean, he knows more than me, but not enough to offer anything. And his commentary and general broadcasting skills, regardless of knowledge, aren't a problem. Sportscasters who specialize in major pro sports do just fine when they have to give flavor commentary for swimming or some shit during the Olympics. Sure, there are no other North American Go masters besides this one guy. But is there no one else who can even discuss the game who speaks English? There isn't a Chinese or Japanese master who is partially fluent in English? There isn't a broadcaster who can at least sound good and add flavor despite knowing nothing about Go? I feel like some people I knew from the RIT Go Club could easily do a better job. They could at least get someone who knows a lot about computers to balance the guy who knows a lot about Go. Come on!
I wish we could get the Korean commentary with subtitles...
Given that Go is unknown to many people, and it's such a deep well of knowledge, it actually seems like an excellent idea to have a less-informed and a more-informed commentator. That way, the less informed commentator can ask the questions that people have (because he also has them) and the high-level commentator can answer them. If they were both pros, they might not comment on things that are obvious to them that uninformed viewers should know.
That's a difficult problem in general. What if the informed viewers get annoyed at this chump asking simple questions all the time? How do you effectively communicate to an audience with potentially vastly different levels of understanding?
At least for sports announcing, the solution is simple, but almost nobody ever does it: multiple announcing teams. I really want to be able to turn on the super-pro announcers for football games. Actually talk about formations, coverages, etc. I want all the inside baseballs.
At least for sports announcing, the solution is simple, but almost nobody ever does it: multiple announcing teams. I really want to be able to turn on the super-pro announcers for football games. Actually talk about formations, coverages, etc. I want all the inside baseballs.
I think the NFL actually bans the tv networks from discussing such deep strategy in the broadcast contracts. They consider that to be trade secret/IP that is sold separately, if at all. They control who can watch the film and who can't.
I stopped watching the official feed because the non-expert presenter was so bad. It's really rough when the expert commentator is also the best host, is more comfortable in front of the camera, has better talking skills and is generally just a more relatable person.
I was super surprised when the non-expert wasn't replaced after the first match. Literally any random person would have a chance of being better at that job than he is.
Also, why don't they have basic things like a buzzer going off when a player makes a move? The two of them are always so wrapped up in their own board and musings that a play in the real game always always always passes them by, then they have to catch up with real state of the board again.
I would have been pretty happy if they went with an expert in computer science/artificial intelligence as one of the presenters. At least then there could be some back and forth discussion on different information.
I do not understand enough about high level Go play to be able to interpret much of anything it's doing without the commentary ;^)
The same is true for me, except for the extremely bad moves it made at the very end of game four. I played Go on an iPad for a few days way back. That's all you need to see how bad some of those moves are. They were like, onside kicks, but kicked so poorly such that only if the entire other team were unconscious could you recover them for a touchdown.
Yeah, but it's pretty easy to understand why the AI makes those kicks. There is only the game, not an opponent. It doesn't realize it's playing against one of the best players in the world, that "shouldn't ever" make a basic mistake. So it's kicking onside because that's the only path to victory at that point. "If the other guy fucks up monumentally here, I'm still in the game." The game was lost before that point, and it should have resigned, but it still had an onside kick chance. And it did resign soon after. I'm actually very curious how the AI evaluates resigning. Does it just go on raw odds? Or does it calculate that out to the end? In my outsider opinion, it resigned later than a player would. Probably because it still had that onside kick chance... which would never actually happen with a pro player that was of sound mind.
The AGA has a stream where they are commenting on the game at the same time as the official games, the commentators are Andrew Jackson (5d amateur, he runs a mid/high level study class at the Seattle Go Center and he has also been commentating on other high level games in the past few months) and Kim Myungwan (9p korean, he's really strong, it's hard to compare but I believe he's stronger than Redmond by a fair amount, not nearly as strong as Lee Sedol)
There should also be commentary by Younggil An (8p korean, does lectures in Australia iirc) in english floating around as well.
The AGA has a stream where they are commenting on the game at the same time as the official games, the commentators are Andrew Jackson (5d amateur, he runs a mid/high level study class at the Seattle Go Center and he has also been commentating on other high level games in the past few months) and Kim Myungwan (9p korean, he's really strong, it's hard to compare but I believe he's stronger than Redmond by a fair amount, not nearly as strong as Lee Sedol)
There should also be commentary by Younggil An (8p korean, does lectures in Australia iirc) in english floating around as well.
That is the stream I've been watching for the last two matches. It's just waaaay better. Andrew Jackson is much more likable than either guys on the official stream, and actually has something to contribute. He also works for Google, though makes it clear he has nothing to do with AlphaGo or Deepmind or anything like that. On the third match he had another co-host who was a lot more fun than Kim Myungwan, and easier to understand her English, but I don't recall her name right now.
Was it Lee Hjajin 3p (haylee on youtube)? She has a go channel on youtube where she plays games and talks about the games as she plays. I've been pretty busy this weekend so I haven't seen much of the 3rd and 4th game yet. I plan on going over them tonight though!
That article makes a silly argument that the AI is "less efficient" than human intelligence, because it can play so many more games so much faster.
Human intelligence is the product of civilization and thousands of years of refinement. We learned the same way, just over longer periods of time due to the slow nature of our meat.
It's not silly if the only way to develop AlphaGo requires Google levels of data to train your network. Humans are able to learn incredible amounts of information from very sparse input data. It allows us to be extremely adaptable in new situations.
Also, there is some very real contention in the AI world about whether reinforcement learning is actually that useful for anything in the real world.
It's not silly if the only way to develop AlphaGo requires Google levels of data to train your network. Humans are able to learn incredible amounts of information from very sparse input data. It allows us to be extremely adaptable in new situations.
Also, there is some very real contention in the AI world about whether reinforcement learning is actually that useful for anything in the real world.
Exactly. The article explicitly says "data efficiency", and AlphaGo is not very good in that regard.
One could reasonably argue that human Go players cheat in terms of data efficiency by learning from the old masters, but at a fundamental level humans have a capacity for abstract reasoning that AlphaGo can compete with only due to masses and masses of highly domain-specific data.
Well, if my understanding is correct, from having done like a handful of machine learning tutorials, AlphaGo doesn't actually keep a huge amount of data. There is a tremendous amount of input data in order to train it, but that data isn't kept around in some database. It is just used to populate and refine a matrix of values. The matrix is all that is needed to make decisions. The game data that was used as input can be discarded. In that way, it's really quite similar to human learning.
It's a lot like writing a spam filter. You can teach the spam filter how to recognize spam. Then it doesn't really need to keep this huge database of spam e-mails around.
What's holding us back now is that systems like AlphaGo take very smart people to make, and lots of expensive computing resources to run. If we could get over those two things, such systems could replace most human intellectual labor.
That's significantly more true now than it was when Deep Blue defeated Kasparov 20 years ago, but there's a lot we're still missing apart from engineering expertise and hardware.
AlphaGo, like Deep Blue before it, is a very special-purpose AI. Yes, the underlying techniques of deep neural networks, reinforcement learning, and Monte-Carlo Tree Search are much more general than minimax tree search with alpha-beta pruning was, but they're still very, very far from the flexibility of human intelligence.
What's holding us back is that all these very smart people don't really, truly understand deep, multiplayer convolution networks. Sure, we are really good at constructing the scaffolding for this machine that can perform very specific tasks given enough (and the right) training data. However, we don't know anything about the underlying structure of this machine once it's been trained. It's essentially a black box.
Comments
What was most interesting was before it resigned it made some apparently dumb moves. Some were so dumb I could even see how bad they were. But really they weren't bad. They gave AlphaGo the best chance to win. Of course, they only had a chance if the opponent was so dumb they didn't play the obvious counter. Only after attempting, and failing, at all of those moves did it resign.
I wish we could get the Korean commentary with subtitles...
At least for sports announcing, the solution is simple, but almost nobody ever does it: multiple announcing teams. I really want to be able to turn on the super-pro announcers for football games. Actually talk about formations, coverages, etc. I want all the inside baseballs.
Do you have any source for that or baseless speculation? :O
I was super surprised when the non-expert wasn't replaced after the first match. Literally any random person would have a chance of being better at that job than he is.
Also, why don't they have basic things like a buzzer going off when a player makes a move? The two of them are always so wrapped up in their own board and musings that a play in the real game always always always passes them by, then they have to catch up with real state of the board again.
The AGA has a stream where they are commenting on the game at the same time as the official games, the commentators are Andrew Jackson (5d amateur, he runs a mid/high level study class at the Seattle Go Center and he has also been commentating on other high level games in the past few months) and Kim Myungwan (9p korean, he's really strong, it's hard to compare but I believe he's stronger than Redmond by a fair amount, not nearly as strong as Lee Sedol)
There should also be commentary by Younggil An (8p korean, does lectures in Australia iirc) in english floating around as well.
http://www.theguardian.com/media-network/2016/jan/28/google-ai-go-grandmaster-real-winner-deepmind
Human intelligence is the product of civilization and thousands of years of refinement. We learned the same way, just over longer periods of time due to the slow nature of our meat.
Also, there is some very real contention in the AI world about whether reinforcement learning is actually that useful for anything in the real world.
One could reasonably argue that human Go players cheat in terms of data efficiency by learning from the old masters, but at a fundamental level humans have a capacity for abstract reasoning that AlphaGo can compete with only due to masses and masses of highly domain-specific data.
It's a lot like writing a spam filter. You can teach the spam filter how to recognize spam. Then it doesn't really need to keep this huge database of spam e-mails around.
AlphaGo, like Deep Blue before it, is a very special-purpose AI. Yes, the underlying techniques of deep neural networks, reinforcement learning, and Monte-Carlo Tree Search are much more general than minimax tree search with alpha-beta pruning was, but they're still very, very far from the flexibility of human intelligence.