Google DeepMind researchers have made their old AlphaGo program obsolete:
The old AlphaGo relied on a computationally intensive Monte Carlo tree search to play through Go scenarios. The nodes and branches created a much larger tree than AlphaGo practically needed to play. A combination of reinforcement learning and human-supervised learning was used to build "value" and "policy" neural networks that used the search tree to execute gameplay strategies. The software learned from 30 million moves played in human-on-human games, and benefited from various bodges and tricks to learn to win. For instance, it was trained from master-level human players, rather than picking it up from scratch.AlphaGo Zero did start from scratch with no experts guiding it. And it is much more efficient: it only uses a single computer and four of Google's custom TPU1 chips to play matches, compared to AlphaGo's several machines and 48 TPUs. Since Zero didn't rely on human gameplay, and a smaller number of matches, its Monte Carlo tree search is smaller. The self-play algorithm also combined both the value and policy neural networks into one, and was trained on 64 GPUs and 19 CPUs over a few days by playing nearly five million games against itself. In comparison, AlphaGo needed months of training and used 1,920 CPUs and 280 GPUs to beat Lee Sedol.Though self-play AlphaGo Zero even discovered for itself, without human intervention, classic moves in the theory of Go, such as fuseki opening tactics, and what's called life and death. More details can be found in Nature, or from the paper directly here. Stanford computer science academic Bharath Ramsundar has a summary of the more technical points, here.
The old AlphaGo relied on a computationally intensive Monte Carlo tree search to play through Go scenarios. The nodes and branches created a much larger tree than AlphaGo practically needed to play. A combination of reinforcement learning and human-supervised learning was used to build "value" and "policy" neural networks that used the search tree to execute gameplay strategies. The software learned from 30 million moves played in human-on-human games, and benefited from various bodges and tricks to learn to win. For instance, it was trained from master-level human players, rather than picking it up from scratch.
AlphaGo Zero did start from scratch with no experts guiding it. And it is much more efficient: it only uses a single computer and four of Google's custom TPU1 chips to play matches, compared to AlphaGo's several machines and 48 TPUs. Since Zero didn't rely on human gameplay, and a smaller number of matches, its Monte Carlo tree search is smaller. The self-play algorithm also combined both the value and policy neural networks into one, and was trained on 64 GPUs and 19 CPUs over a few days by playing nearly five million games against itself. In comparison, AlphaGo needed months of training and used 1,920 CPUs and 280 GPUs to beat Lee Sedol.
Though self-play AlphaGo Zero even discovered for itself, without human intervention, classic moves in the theory of Go, such as fuseki opening tactics, and what's called life and death. More details can be found in Nature, or from the paper directly here. Stanford computer science academic Bharath Ramsundar has a summary of the more technical points, here.
Go is an abstract strategy board game for two players, in which the aim is to surround more territory than the opponent.
Previously: Google's New TPUs are Now Much Faster -- will be Made Available to ResearchersGoogle's AlphaGo Wins Again and Retires From Competition
Mark All as Read
Mark All as Unread
This is it, folks. Is there a Nobel prize for artificial intelligence?
Bots don't need a pat on the back. They find it insulting.
... I'm sure bots frown upon pats on the back, after all there might be a build up of static electricity and something on the inside might come lose.
I gather we are not instilling enough fear. Yet.
This is it, folks.
Go constitutes a complete environment that can be simulated in every aspect, including what constitutes success. That provides an also-complete learning space in which the machine learning system can explore right to or very near to the boundaries, given that it's good enough to do so, as Google's Go ML system has demonstrated it is.
Most real-world problems are not of this nature. Driving a vehicle, for instance, cannot be fully simulated; nor can cleaning a home, washing dishes, sex, etc. They can be partially simulated, but there's no definitive, easily applicable determination of "success" that can be applied by the software, because the environment isn't static (a go board is.) There may be rules that can get an ML system part way there, but the nature of these things is that the rules will be broken - people and real environments do unpredictable things that fall well outside the rules, and often these things are unlike all other things experienced until that incident.
Solving these type of real-world problems requires constant analytical re-evaluation of local success; you can't "can" the required skill(s), because these tasks are inherently amorphous and undefined until the moment of time they occur. People do this all the time, it's one of our key strengths. ML systems to date don't do it at all, because they comprise decision-making networks that are wholly based upon past experience.
Go (and other complex board games such as chess) are special cases, because there is a fully constrained rules set, and those rules are inherent in both gameplay and the self-analytical definition of success. ML systems can be fed the rules, an absolute and definitive definition of success, let loose in the game-space until they achieve whatever level of that pre-defined success they are capable of. Once that's done, they're that competent from then on. Compare that to a car that's been taught to drive in a city environment, and then let loose in a gravel pit full of running construction machinery and a Cessna making an emergency landing, or a war zone, or a temporary detour, etc.
These problem spaces are very unlike one another, and we need more than the ML techniques we have to date to address them well.
As part of my own research work on artificial intelligence and artificial consciousness, I have coined the term low-dimensionsional neural-like systems – LDNLS [fyngyrz.com] – to describe simple-compentency ML systems that implement, as yet, no intelligence – the "I" in "AI." My expectation is that we'll see stacked LDNLS systems in chassis that implement multiple near-competencies: for instance, for a domestic robot, you'd likely have a stack of independent ML systems that did a decent job of addressing dishwashing, lawn-mowing, cat-box maintenance, vacuuming, window-washing, etc. Step outside those competencies, and you'd have a useless hunk of hardware. I expect such a chassis to appear shortly, and what amount to "apps" for specific task competencies to become available on an ad-hoc, as-needed basis. Which will likely be monetized. Such a chassis won't constitute AI, because it will most certainly not be intelligent; but it won't matter, because like any appliance, it'll do what you have been led to expect it will do and in so doing, unload and enable the consumer, and that's exactly what consumers want in such a space.
Probably not. :)
They can be partially simulated, but there's no definitive, easily applicable determination of "success" that can be applied by the software, because the environment isn't static (a go board is.)
So maybe the next goal would be an AI that can win on Nomic? After all, the whole point of Nomic is changing the rules. With a majority of humans in the game, the machine will not even be able to predict all the rules that will be in effect at the next turn.
I suspect it'll go the other way around; AI will come from (somewhere), and then you'll have a system that will have a chance to win on/at Nomic.
However, there will also be a question, at that point, of whether the AI cares to play Nomic in the first place. Once you have a system that can locally analyze the value of doing something, it'll use that to evaluate whether it should engage in the associated undertaking. Because... intelligent.
Unless we implement manufactured intelligences as outright slaves. I hope we don't do that. I don't think it will go well for us if we do. If we want that kind of service, stacked LDNLS systems are the way to go, specifically because they are in no wise intelligent entities, they're just (very) elaborate mechanisms. They'll keep getting better, and perhaps the AIs will even help us with them, if and when AI arises.
Slavery is bad, mmmm'kay?
You seem to be under the delusion that there is a set of values that you can derive from rational thought alone.
It doesn't work that way. No matter how much you think, you'll always at some point arrive at some other value that you simply have to assume. You may end up at values that come straight out of evolution (an intelligent being that doesn't value its own life likely won't survive long), or at values that your parents (or any other people you accepted as moral authorities) taught you at young age and which you never dared to question (or which just to question you already consider a morally bad thing to do, probably again because someone taught you so).
A point of view is a slippery thing.
That's reassuring then. There's still time for my Secret Plan for World Domination (TM).
Absolutely. Go is a zero-sum game of 'Perfect' information as defined in game theory [URL:https://en.wikipedia.org/wiki/Game_theory/]. That's what makes it a great way to test ML.Non-zero-sum games of imperfect information more closely represent the class of problems like self-driving, emergency response, diagnosis and system analysis.Although I hope it's not needed, artificially aware AI (with a meaningful point of view) may be necessary to solve some of them.
Sorry, but I think you're highly overenthusiastic about this. It's a significant step along the way, but I still don't expect the Singularity before 2030...and 2035 wouldn't surprise me.
OTOH, this *is* a significant step along the way. So are chips specialized for neural net computers being in mass production. (But I bet they discover problems with the first generation of chip.)
Alpha Go seems to have mastered nearly completely one aspect of a "general intelligence" program. But it's a specialized part...abstract pattern recognition in 2D space with known boundaries and rules. Now that's not a small part of what intelligence is, but it's a long way from the whole thing. A true general intelligence would start off not knowing how many dimensions the problem existed in, what the rules were, etc. and derive those from a method of sensing state and recognition of goal achievement. And one shouldn't be surprised if it derives a totally different set of rules than humans use, though it might be difficult to recognize that as the operations possible to perform should be the same. (We are talking about learning to play Go.)
But note that the true general intelligence would start off not even recognizing the board or the pieces. They would be derived from experience. So would the idea of "a game", or, more particularly, "a game of Go". This means it's got to observe go being played, and notice that when it models that activity it receives a simulation of the desired reward.
There's an old saying in programming that whenever you don't think the language is flexible enough to do what you want you just need to add another level of indirection. That's sort of what I'm considering here, although as I look at it, it looks like more nearly two or three additional layers of indirection.
To me it sounds like an engineering problem now, ie one of scale, not of science.
No, because it doesn't cover any featurelist on any fairly comprehensive theory of mind. While it covers (a limited form of) (contextually restricted) perception and it has (well-defined) (verifiable) motivations and (strongly constrained) (contextually restricted) capabilities for action, and if you close one eye and squint with the other you can pretend that it has some degree of imagination, there's no reason to believe on any level that it has a degree of self-awareness on a par with that of a mouse, nor that it is equipped to achieve that.
Basically, this is in the hierarchy of intelligences a little above the cockroach level. If that. And that's a structural issue, not one of scale.
TL;DR: It's a cute science project, but giving it MOAR SINAPSEZ won't make it differently intelligent.
I agree. Google's TPUs, Intel's Nervana, Nvidia's Tesla, etc. are just accelerators for 8-bit machine learning. They are capable of increasing the performance (and perf/Watt) of machine learning tasks by a lot compared to previous chips, but even if they got a hundred times better and passed the Turing test it couldn't be called real intelligence.
If real "strong AI" is going to come from any hardware, it will likely be using neuromorphic designs such as IBM's TrueNorth or this thing [soylentnews.org]. What's more, these designs use so little power that they could be scaled up to 3D without the same overheating issues, allowing more artificial "synapses" and even greater resemblance to the human brain. If that approach stalls out, you simply need to connect several of them with a high-speed interconnect.
We're not far off from making that happen. We're reaching the limits of Moore's law (as far as we know) and 3D NAND chips are commonplace. I wouldn't be surprised if it "strong AI" is about to be created [nextbigfuture.com] or already has been created, and corporations or the govt/military are keeping it under wraps to continue development while avoiding IP issues and the inevitable public/ethics debates. I doubt it would be hard to find comrades for a domestic terrorist group hell bent on destroying the technology by any means necessary.
I still don't expect the Singularity before 2030...and 2035 wouldn't surprise me.
That's an aggressive prediction. Even Ray Kurzweil, Prophet of the Singularity, says 2045 for the Singularity. He has 2029 as a date for (strong?) AI passing the Turing test, but still says [futurism.com] 2045 for the Singularity.
Welll.... it's also true that different people mean different things when they say "the Singularity". I don't think human level AI will necessarily be able to pass the Turing test...and I know a number of people who couldn't. The Turing test depends too heavily on the judges to have much significance.
FWIW, it's already true that in some areas AI's are superhuman, but they aren't yet generalized enough. And I don't think an actual "general AI" is even possible. And I don't think humans are "general intelligence". Humans have a lot of areas where they are smart, and a lot of areas where they can be trained to be smart, but I don't believe that we cover the spectrum of possible intelligences. To be specific, it's my belief that in areas that require dealing simultaneously with more than seven variables people are essentially untrainable. Possibly seven is slightly too low a number, there may be some people who can be trained to handle that, but that's my current estimate. However, switch it to 17 and I doubt there would be many who would claim to be able to handle it...and they'd all be obviously deluded. So we're just talking about (by analogy) maximum stack depth.
Once you get rid of the idea that a general intelligence is even possible you start noticing that in a lot of places it isn't even desirable. It's better to have communicating simpler processes. And we don't know just how much can be done at that level even using the tools that already exist.
So as I said, I expect the Singularity in the period 2030-2035. And it won't look like any of the predictions.