They went on to make AlphaZero, a generalised version that could learn chess, shogi or any similar game. The chess version beat a leading conventional chess program 28 wins, 0 losses, and 72 draws.
That seemed impressive to me.
Also they used loads of compute during the training but not so much during play.(5000 TPUs, 4TPUs).
Also it got better than humans in those games from scratch in about 4 hours whereas humans have had 2000 years to study them so you can forgive it some resource usage.
What was impressive was the way Stockfish9 was beaten. AlphaZero played like a human player, making sacrifices for position that stockfish thought were detrimental. When it played as white, the fact that is mostly started with the Queen pawn (despite that the King pawn is "best by test") and the way AlphaZero used Stockfish pawnstructure and tempo to basicaly remove a bishop from the game was magical.
Yes, since its a game, it's "useless", but it allowed me (and i'm not the only one) to be a bit better at chess. It's not world hunger, not climate change, it's just a bit of distraction for some people.
PS: I was part of the people thinking that Genetic algorithm+deep learning was not enough to emulate human logical capacities, AlphaZero vs Stockfish games made me admit i was wrong (even if i still think it only works inside well-defined environments)
Just because Fischer preferred 1. e4, it doesn't make it better than other openings. https://en.chessbase.com/post/1-e4-best-by-test-part-1
Playing like a human for me also means making human mistakes. A chess-playing computer playing like a 4000 rated "human" is useless, one that can be configured to play at different ELOs is more interesting, although most can do that and there's no ML needed, nor huge amounts of computing power.
Without its opening database and without its endgame tablebase?
Frankly, the Stockfish vs AlphaZero match was the beginning of the AI Winter in my mind. The fact that they disabled Stockfish's primary databases was incredibly fishy IMO and is a major detriment to their paper.
Stockfish's engine is designed to only work in the midgame of Chess. Remove the opening database and remove the endgame database, and you're not really playing against Stockfish anymore.
The fact that Stockfish's opening was severely gimped is not a surprise to anybody in the Chess community. Stockfish didn't have its opening database enabled... for some reason.
Most games are also closed systems, and conveniently grokkable systems, with enumerable search spaces. Which gives us easily produceable measures of the contraptions' abilities.
Whether this is the most effective path to understanding deeper questions about intelligence is an open question.
But I don't think it's fair to say that deeper questions and problems are being foregone simply to play games.
I think most 'games researchers' are pursuing these paths because they themselves and no one else has put forth any other suggestion that makes them think, "hmm, that's a really good idea, that seems like it might be viable and there is probably something interesting we could learn from it."
Do you have any suggestions?
And comparing Alpha Go Zero against those "other chess programs that existed for 30 years" is exactly missing the point also. Those programs were not constructed with zero-knowledge. They were carefully crafted by human players to achieve the result. Are we also going to count in all the brain processing power and the time spent by those researchers to learn to play chess? Alpha Go Zero did not need any of that, besides the knowledge about the basic rules of the game. Who compare compute requirements for 2 programs that have fundamentally different goals and achievements? One is carefully crafted by human intervention. The other one learns a new game without prior knowledge...
Sounds more like religion and less like science to me.
I guess we could argue until the end of the world that no intelligence will emerge from more and more clever ways of brute-forcing your way out of problems in a finite space with perfect information. But that's what I think.
On the topic of the different algorithmic approaches, I find it so fascinating how different these two approaches actually end up looking when analyzed by a professional commentator. When you watch the new style with a chess commentator, it feels a lot like listening to the analysis of a human game. The algorithm has very clearly captured strategic concepts in its neural network. Meanwhile, with older chess engines there is a tendency to get to positions where the computer clearly doesn't know what its doing. The game reaches a strategic point and the things its supposed to do are beyond the horizon of moves it can computer by brute force. So it plays stupid. These are the positions that, even now, human players can beat better than human old style chess engines at.
But attacking not-well-constrained problems is what's needed to show real progress in AI these days, right?
This. Learning to play a game is one thing. Learning how to teach computers to learn a game is another thing. Yes chess programs have been good before, but that's missing the point a little bit. The novel bit is not that it can beat another computer, but how it learned how to do so.
That's a pretty major shift for humanity.
But it's a mistake to think that a system learning by playing against itself is something new. Arthur Samuel's draughts (chequers) program did that in 1959.
It's not that it's new, it's that they've achieved it. Chess was orders of magnitude harder than draughts. The solution for draughts didn't scale to chess but Alpha Go zero showed that chess was ridiculously easy for it once it had learned Go.
Big Blue is fine - it's referring to the company and not the machine. From Wikipedia "Big Blue is a nickname for IBM"
Consider it as the perfect lab.
Seems like a lab so simplified that I'm unconvinced of its general applicability. Perfect knowledge of the situation and a very limited set of valid moves at any one time.
an awful lot of graph and optimization problems. See for instance some examples in https://en.wikipedia.org/wiki/A*_search_algorithm
Did they manage to extend it to games with hidden and imperfect information?
(Say, chess with fog of war also known as Dark Chess. Phantom Go. Pathfinding equivalent would be an incremental search.)
Edit: I see they are working on it, predictive state memory paper (MERLIN) is promising but not there yet.
(You said problems, not games...)
The thing is, an algorithm that can work with fewer samples and robustly tolerating mistakes in datasets (also known as imperfect information) will be vastly cheaper and easier to operate. Less tedious sample data collection and labelling.
Working with lacking and erroneous information (without known error value) is necessarily a crucial step towards AGI; as is extracting structure from such data.
This is the difference between an engineering problem and research problem.
I completely agree about the importance of imperfect information problems. In practice, many techniques handle some label noise, but not optimally. Even MNIST is much easier to solve if you remove the one incorrectly-labeled training example. (one! Which is barely noise. Though as a reassuring example from the classification domain, JFT is noisy and still results in better real world performance than just training on imagenet.)
I guess in the same way as lab chemistry isn't interesting anymore ? (Since it often happens in unrealistically clean equipment :-)
I think there is nothing preventing lab research from going on at the same time as industrialization of yesterday's results. Quite on the contrary: in the long run they often depend on each other.
The real challenge is to devise a general algorithm that will learn to be a good poker player in thousands of games, strategically, from just a bunch of games played. DeepStack AI required 10 million simulated games. Good human players outperform it at intermediate training stages.
And then the other part is figuring out actual rules of a harder game...
A good example of a game of imperfect information is poker, because players have a private hand which is known only to them. Whereas all possible future states of a chess game can be narrowed down according to the current game state, the fundamental uncertainty of poker means there is a combinatorial explosion involved in predicting future states. There's also the element of chance in poker, which further muddies the waters.
Board games are often (but not always) games of perfect and complete information. Card games are typically games of imperfect and complete information. This latter term, "complete information", means that even if not all of the game state is public, the intrinsic rules and structure of the game are public. Both chess and poker are complete, because we know the rules, win conditions and incentives for all players.
This is all to say that games of perfect information are relatively easy for a computer to win, while games of imperfect information are harder. And of course, games of incomplete information can be much more difficult :)
If this were true, there would be a vast demand for grandmasters in commerce, government, the military... and there just isn’t. Poker players suffer from similar delusions about how their game can be generalised to other domains.
Oh that's so true
Poker players in the real life would give up more often than not, whenever they didn't know enough about a situation or they didn't have enough resources for a win with a high probability.
And people can call your bluff even if you fold.
I suspect that chess as a metagame is just so far developed that being "good at chess" means your general ability is really overtrained for chess.
Can AI make the world better? It can, but it won't since we are humans, and humans will weaponize technology every chance it gets. Of course some positive uses will come, but the negative ones will be incredibly destructive.
The practical uses of these technologies don't always make national news.
I'm sure you would also have scoffed at the "pointless impractical, wasteful use of our brightest minds" to make the the Flyer hang in the air for 30 yards at Kitty Hawk.
We solved nothing.
IBM Deep Blue doesn't exactly think like humans do.
Most of our algorithms really are 'better brute force'.
https://www.theatlantic.com/magazine/archive/2013/11/the-man...
Side observers are taking joy in the risker plays that it did -- reminded them of certain grand-masters I suppose -- but that still doesn't mean AGZ is close to any form of intelligence at all. Those "riskier moves" are probably just a way to more quickly reduce the problem space anyway.
It seriously reminds me more and more of religion, the AI area these days.
Most humans don't live 2000 years. And realistically don't spend that much of their time or computing power on studying chess. Surely a computer can be more focused at this and the 4h are impressive. But this comparison seems flawed to me.
Remember people reach peak play in ~15 years, but they don't nessisarily keep up with advances.
PS: You see this across a huge range of fields from running, figure skating, to music people simply spend more time and resources getting better.
So, in this sense, it's kind of like taking a human, teaching them the exact rules of the game and showing them how to run calculations, and then telling them to sit in a room playing games against themselves. In my experience from chess, you'd be at a huge disadvantage if you started with this zero-knowledge handicap.
One problem is that we can't play millions of games against ourselves in a few hours. We can play a few games, grow tired, and then need to go do something else. Come back the next day, repeat. It's a very slow process, and we have to worry about other things in life. How much of one's time and focus can be used on learning a game? You could spend 12 hours a day, if you had no other responsibilities, I guess. That might be counter productive, though. We just don't have the same capacity.
If you artificially limited AlphaGo to human capacity, then my money would be on the human being a superior player.
In a not equal fight, and the results are still not published. I'm not claiming that AlphaZero wouldn't win, but that test was pure garbage.
I agree AlphaZero had fancier hardware and so it wasn't really a fair fight.
Few would care. Your examiner doesn't give you extra marks on a given problem for finishing your homework quickly.
https://deepmind.com/blog/deepmind-ai-reduces-google-data-ce...
Just because alpha zero doesn't solve the problem you want it to doesn't mean that advancements aren't being made that matter to someone else. To ignore that seems disingenuous.
If you want an idea of where machine learning is in the scheme of things, the best thing to do is listen to the experts. _None_ of them have promised wild general intelligence any time soon. All of them have said "this is just the beginning, it's a long process." Science is incremental and machine learning is no different in that regard.
You'll continue to see incremental progress in the field, with occasional demonstrations and applications that make you go "wow". But most of the advances will be of interest to academics, not the general public. That in no way makes them less valuable.
The field of ML/AI produces useful technologies with many real applications. Funding for this basic science isn't going away. The media will eventually tire of the AI hype once the "wow" factor of these new technologies wears off. Maybe the goal posts will move again and suddenly all the current technology won't be called "AI" anymore, but it will still be funded and the science will still advance.
It's not the exciting prediction you were looking for I'm sure, but a boring realistic one.
What make this 3rd/4th boom in AI different?
The other AI winter, the funding for these science went from well funded to little funding.
I'm skeptical, with respect of course, on your statement because it doesn't have anything to back that up other than it produce useful technologies. Wouldn't this statement imply that the other previous AI which experience AI Winter (expert system, and whatever else) didn't produce useful enough technologies to have funding?
I'm currently on the camp of there is going to be an AI Winter III coming.
> None_ of them have promised wild general intelligence any time soon.
The post talk about Andrew Ng wild expectation on other things such as radiologist tweet. While it's not wild general intelligence. What I think the main article and also I am thinking is the outrageous speculation. Another one is the tesla self driving, it doesn't seem to be there yet and perhaps we're hitting the point of over promise like we did in the past and then AI winter happen because we've found the limit.
The current difference is that the technologies are actually useful right now. It's not about promised or expected technologies of tomorrow, but about what we have already researched, about known capabilities that need implementation, adoption, and lots of development work to apply it in lots and lots of particular use cases. If the core research hits a dead end tomorrow and stops producing any meaningful progress for the next 10 or 20 years, the obvious applications of neural-networks-as-we're-teaching-them-in-2018 work sufficiently well and are useful enough to deploy them in all kinds of industrial applications, and the demand is sufficient to employ every current ML practitioner and student even in absence of basic research funding, so a slump is not plausible.
Sure, keep moving timelines. It's what makes you money in the area. I am sure when around mid-2019 hits, it will suddenly be "most experts agree that the first feasible self-driving cars will arrive circa 2025".
You guys are hilarious.
Training is expensive but inference is cheap enough for Alpha Zero inspired bots to beat human professionals while running on consumer hardware. DeepMind could have released thousands of pro-level games if they wanted to and others have: http://zero.sjeng.org/
I am 100% in agreement with the author on the thesis: deep learning is overhyped and people project too much.
But the content of the post is in itself not enough to advocate for this position. It is guilty of the same sins: projection and following social noises.
The point about increasing compute power however, I found rather strong. New advances came at a high compute cost. Although it could be said that research often advances like that: new methods are found and then made efficient and (more) economical.
A much stronger rebuttal of the hype would have been based on the technical limitations of deep learning.
I'm not even sure how you'd go about doing that. You could use information theory to debunk some of the more ludicrous claims, especially ones that involve creating "missing" information.
One of the things that disappoints me somewhat with the field, which I've arguably only scratched the surface of, is just how much of it is driven by headline results which fail to develop understanding. A lot of the theory seems to be retrofitted to explain the relatively narrow result improvement and seems only to develop the art of technical bullshitting.
There are obvious exceptions to this and they tend to be the papers that do advance the field. With a relatively shallow resnet it's possible to achieve 99.7% on MNIST and 93% on CIFAR10 on a last-gen mid-range GPU with almost no understanding of what is actually happening.
There's also low-hanging fruit that seems to have been left on the tree. Take OpenAI's paper on parametrization of weights, so that you have a normalized direction vector and a scalar. This makes intuitive sense for anybody familiar with high-dimensional spaces since nearly all of the volume of a hypersphere lies around the surface. That this works in practice is great news, but leaves many questions unanswered.
I'm not even sure how many practitioners are thinking in high dimensional spaces or aware of their properties. It feels like we get to the universal approximation theorem and just accept that as evidence that they'll work well anywhere and then just follow whatever the currently recognised state of the art model is and adapt that to our purposes.
Who's to say we won't improve this though? Right now, nets add a bunch of numbers and apply arbitrarily-picked limiting functions and arbitrarily-picked structures. Is it impossible that we find a way to train that is orders of magnitude more effective?
Currently, people are projecting and saying that we are going to see huge AI advances soon. On which basis are these claims made? Showing fundamental limitations of deep learning is showing we have no idea how to get there. How to get there yet, indeed, just we have no idea how to do time travel yet.
The end result of this advancement to our world is earth shattering.
On the high compute cost. There is an aspect of that being true but we have also seen advancement in silicon to support. We look at WaveNet using 16k cycles through a DNN and offering at scale and competitive price kind of proves the point.
Current AIs have limitations but, at the tasks they are suited for, they can equal or exceed humans with years of experience. Computing power is not the key limit since it will be made cheaper over time. More importantly, new advances are still being made regularly by DeepMind, OpenAI, and other teams.
https://www.quora.com/Roughly-what-processing-power-does-the...
Unsupervised Predictive Memory in a Goal-Directed Agent
https://en.wikipedia.org/wiki/Moravec%27s_paradox
Moravec's paradox is the discovery by artificial intelligence and robotics researchers that, contrary to traditional assumptions, high-level reasoning requires very little computation, but low-level sensorimotor skills require enormous computational resources.
What do you think of recent papers and demos by teams from Google Brain, OpenAI, and Pieter Abbeel's group on using simulations to help train physical robots? Recent advances are quite an improvement over those from the past.
Now using models for RL is the obvious choice, since trying to teach a robot a basic behavior with RL is just absurdly impractical. But the problem here, is that when somebody build that model (a 3d simulations) they put in a bunch of stuff they think is relevant to represent the reality. And that is the same trap as labeling a dataset. We only put in the stuff which is symbolically relevant to us, omitting a bunch of low level things we never even perceive.
This is a longer subject, and a HN is not enough to cover it, but there is also something about the complexity. Reality is not just more complicated than simulation, it is complex with all the consequences of that. Every attempt to put a human filtered input between AI and the world will inherently loose that complexity and ultimately the AI will not be able to immunize itself to it.
This is not an easy subject and if you read my entire blog you may get the gist of it, but I have not yet succeeded in verbalizing it concisely to my satisfaction.
[1] https://mobile.nytimes.com/1992/06/05/business/fifth-generat...
Why not petaflops of bytes then?
You mean Monte Carlo Tree Search, which is not at all like Ma(r)kov chains. You're probably mixing it up with Markov decision processes though.
Before criticising something it's a good idea to have a solid understanding of it.
The biggest minds everywhere are working on AI solutions, and there's also a lot in medical/science going on to map brains and if we can merge neuroscience with computer science we might have more luck with AI in the future...
So we could have a draught for a year or two, but there will be more research, and more breakthroughs. This won't be like the AI winters of the past where it lay dormant for 10+ years, I don't think.
The first generation TPUs used 65536 very simple cores.
In the end you have so many transistors you can fit and there are options on how to arrange and use.
You might support very complex instructions and data types and then four cores. Or you might only support 8 bit ints, very, very simple instructions and use 65536 cores.
In the end what matters is the joules to get something done.
We can clearly see that we have big improvements by using new processor architectures.
The author is clearly informed and takes a strong, historical view of the situation. Looking at what the really smart people who brought us this innovation have said and done lately is a good start imo (just one datum of course, but there are others in this interesting survey).
Deepmind hasn't shown anything breathtaking since their Alpha Go zero.
Another thing to consider about Alpha Go and Alpha Go Zero is the vast, vast amount of computing firepower that this application mobilized. While it was often repeated that ordinary Go program weren't making progress, this wasn't true - the best, amateur programs had gotten to about 2 Dan amateur using Makov Tree Search. Alpha Go added CNNs for it's weighting function and petabytes of power for it's process and got effectiveness up to best in the world, 9 Dan professional, (maybe 11 Dan amateur for pure comparison). [1]
Alpha Go Zero was supposedly even more powerful, learned without human intervention. BUT it cost petabytes and petabytes of flops, expensive enough that they released a total of ten or twenty Alpha Go Zero game to the world, labeled "A great gift".
The author convenniently reproduces the chart of power versus results. Look at it, consider it. Consider the chart in the context of Moore's Law retreating. The problems of Alpha Zero generalizes as described in the article.
The author could also have dived into the troubling question as of "AI as ordinary computer application" (what does testing, debugging, interface design, etc mean when the app is automatically generated in an ad-hoc fashion) or "explainability". But when you can paint a troubling picture without these gnawing problems appearing, you've done well.
[1] https://en.wikipedia.org/wiki/Go_ranks_and_ratings