go_1000x560

Markos Kay for Quanta Magazine

In 1997, IBM’s Deep Blue system defeated the world chess champion, Garry Kasparov. At the time, the victory was widely described as a milestone in artificial intelligence. But Deep Blue’s technology turned out to be useful for chess and not much else. Computer science did not undergo a revolution.

Will AlphaGo, the Go-playing system that recently defeated one of the strongest Go players in history, be any different?

Quantized

A monthly column in which top researchers explore the process of discovery. This month’s columnist, Michael Nielsen, is a computer scientist and author of three books.

I believe the answer is yes, but not for the reasons you may have heard. Many articles proffer expert testimony that Go is harder than chess, making this victory more impressive. Or they say that we didn’t expect computers to win at Go for another 10 years, so this is a bigger breakthrough. Some articles offer the (correct!) observation that there are more potential positions in Go than in chess, but they don’t explain why this should cause more difficulty for computers than for humans.

In other words, these arguments don’t address the core question: Will the technical advances that led to AlphaGo’s success have broader implications? To answer this question, we must first understand the ways in which the advances that led to AlphaGo are qualitatively different and more important than those that led to Deep Blue.

In chess, beginning players are taught a notion of a chess piece’s value. In one system, a knight or bishop is worth three pawns. A rook, which has greater range of movement, is worth five pawns. And the queen, which has the greatest range of all, is worth nine pawns. A king has infinite value, since losing it means losing the game.

You can use these values to assess potential moves. Give up a bishop to take your opponent’s rook? That’s usually a good idea. Give up a knight and a bishop in exchange for a rook? Not such a good idea.

The notion of value is crucial in computer chess. Most computer chess programs search through millions or billions of combinations of moves and countermoves. The goal is for the program to find a sequence of moves that maximizes the final value of the program’s board position, no matter what sequence of moves is played by the opponent.

Michael Nielsen

Early chess programs evaluated board positions using simple notions like “one bishop equals three pawns.” But later programs used more detailed chess knowledge. Deep Blue, for example, combined more than 8,000 different factors in the function it used to evaluate board positions. Deep Blue didn’t just say that one rook equals five pawns. If a pawn of the same color is ahead of the rook, the pawn will restrict the rook’s range of movement, thus making the rook a little less valuable. If, however, the pawn is “levered,” meaning that it can move out of the rook’s way by capturing an enemy pawn, Deep Blue considers the pawn semitransparent and doesn’t reduce the rook’s value as much.

Ideas like this depend on detailed knowledge of chess and were crucial to Deep Blue’s success. According to the technical paper written by the Deep Blue team, this notion of a semitransparent levered pawn was crucial to Deep Blue’s play in the second game against Kasparov.

Ultimately, the Deep Blue developers used two main ideas. The first was to build a function that incorporated lots of detailed chess knowledge to evaluate any given board position. The second was to use immense computing power to evaluate lots of possible positions, picking out the move that would force the best possible final board position.

What happens if you apply this strategy to Go?

It turns out that you will run into a difficult problem when you try. The problem lies in figuring out how to evaluate board positions. Top Go players use a lot of intuition in judging how good a particular board position is. They will, for instance, make vague-sounding statements about a board position having “good shape.” And it’s not immediately clear how to express this intuition in simple, well-defined systems like the valuation of chess pieces.

Now you might think it’s just a question of working hard and coming up with a good way of evaluating board positions. Unfortunately, even after decades of attempts to do this using conventional approaches, there was still no obvious way to apply the search strategy that was so successful for chess, and Go programs remained disappointing. This began to change in 2006, with the introduction of so-called Monte Carlo tree search algorithms, which tried a new approach to evaluation based on a clever way of randomly simulating games. But Go programs still fell far short of human players in ability. It seemed as though a strong intuitive sense of board position was essential to success.

What’s new and important about AlphaGo is that its developers have figured out a way of bottling something very like that intuitive sense.

To explain how it works, let me describe the AlphaGo system, as outlined in the paper the AlphaGo team published in January. (The details of the system were somewhat improved for AlphaGo’s match against Lee Sedol, but the broad governing principles remain the same.)

To begin, AlphaGo took 150,000 games played by good human players and used an artificial neural network to find patterns in those games. In particular, it learned to predict with high probability what move a human player would take in any given position. AlphaGo’s designers then improved the neural network by repeatedly playing it against earlier versions of itself, adjusting the network so it gradually improved its chance of winning.

How does this neural network — known as the policy network learn to predict good moves?

Broadly speaking, a neural network is a very complicated mathematical model, with millions of parameters that can be adjusted to change the model’s behavior. When I say the network “learned,” what I mean is that the computer kept making tiny adjustments to the parameters in the model, trying to find a way to make corresponding tiny improvements in its play. In the first stage of learning, the network tried to increase the probability of making the same move as the human players. In the second stage, it tried to increase the probability of winning a game in self-play. This sounds like a crazy strategy — repeatedly making tiny tweaks to some enormously complicated function — but if you do this for long enough, with enough computing power, the network gets pretty good. And here’s the strange thing: It gets good for reasons no one really understands, since the improvements are a consequence of billions of tiny adjustments made automatically.

After these two training stages, the policy network could play a decent game of Go, at the same level as a human amateur. But it was still a long way from professional quality. In a sense, it was a way of playing Go without searching through future lines of play and estimating the value of the resulting board positions. To improve beyond the amateur level, AlphaGo needed a way of estimating the value of those positions.

To get over this hurdle, the developers’ core idea was for AlphaGo to play the policy network against itself, to get an estimate of how likely a given board position was to be a winning one. That probability of a win provided a rough valuation of the position. (In practice, AlphaGo used a slightly more complex variation of this idea.) Then, AlphaGo combined this approach to valuation with a search through many possible lines of play, biasing its search toward lines of play the policy network thought were likely. It then picked the move that forced the highest effective board valuation.

We can see from this that AlphaGo didn’t start out with a valuation system based on lots of detailed knowledge of Go, the way Deep Blue did for chess. Instead, by analyzing thousands of prior games and engaging in a lot of self-play, AlphaGo created a policy network through billions of tiny adjustments, each intended to make just a tiny incremental improvement. That, in turn, helped AlphaGo build a valuation system that captures something very similar to a good Go player’s intuition about the value of different board positions.

In this way, AlphaGo is much more radical than Deep Blue. Since the earliest days of computing, computers have been used to search out ways of optimizing known functions. Deep Blue’s approach was just that: a search aimed at optimizing a function whose form, while complex, mostly expressed existing chess knowledge. It was clever about how it did this search, but it wasn’t that different from many programs written in the 1960s.

AlphaGo also uses the search-and-optimization idea, although it is somewhat cleverer about how it does the search. But what is new and unusual is the prior stage, in which it uses a neural network to learn a function that helps capture some sense of good board position. It was by combining those two stages that AlphaGo became able to play at such a high level.

This ability to replicate intuitive pattern recognition is a big deal. It’s also part of a broader trend. In an earlier paper, the same organization that built AlphaGo — Google DeepMind — built a neural network that learned to play 49 classic Atari 2600 video games, in many cases reaching a level that human experts couldn’t match. The conservative approach to solving this problem with a computer would be in the style of Deep Blue: A human programmer would analyze each game and figure out detailed control strategies for playing it.

By contrast, DeepMind’s neural network simply explored lots of ways of playing. Initially, it was terrible, flailing around wildly, rather like a human newcomer. But occasionally the network would accidentally do clever things. It learned to recognize good patterns of play — in other words, patterns leading to higher scores — in a manner not unlike the way AlphaGo learned good board position. And when that happened, the network would reinforce the behavior, gradually improving its ability to play.

This ability of neural networks to bottle intuition and pattern recognition is being used in other contexts. In 2015, Leon Gatys, Alexander Ecker and Matthias Bethge posted a paper to the scientific preprint site arxiv.org describing a way for a neural network to learn artistic styles and then to apply those styles to other images. The idea was very simple: The network was exposed to a very large number of images and acquired an ability to recognize images with similar styles. It could then apply that style information to new images. For example, the right-hand image below illustrates what happens when you transfer the style of Vincent van Gogh (center) to a photograph of the Eiffel tower (left).

Courtesy of Andre Infante using DeepStyle

It’s not great art, but it’s still a remarkable example of using a neural network to capture an intuition and apply it elsewhere.

Over the past few years, neural networks have been used to capture intuition and recognize patterns across many domains. Many of the projects employing these networks have been visual in nature, involving tasks such as recognizing artistic style or developing good video-game strategy. But there are also striking examples of networks simulating intuition in very different domains, including audio and natural language.

Because of this versatility, I see AlphaGo not as a revolutionary breakthrough in itself, but rather as the leading edge of an extremely important development: the ability to build systems that can capture intuition and learn to recognize patterns. Computer scientists have attempted to do this for decades, without making much progress. But now, the success of neural networks has the potential to greatly expand the range of problems we can use computers to attack.

It’s tempting at this point to cheer wildly, and to declare that general artificial intelligence must be just a few years away. After all, suppose you divide up ways of thinking into logical thought of the type we already know computers are good at, and “intuition.” If we view AlphaGo and similar systems as proving that computers can now simulate intuition, it seems as though all bases are covered: Computers can now perform both logic and intuition. Surely general artificial intelligence must be just around the corner!

But there’s a rhetorical fallacy here: We’ve lumped together many different mental activities as “intuition.” Just because neural networks can do a good job of capturing some specific types of intuition, that doesn’t mean they can do as good a job with other types. Maybe neural networks will be no good at all at some tasks we currently think of as requiring intuition.

In actual fact, our existing understanding of neural networks is very poor in important ways. For example, a 2014 paper described certain “adversarial examples” which can be used to fool neural networks. The authors began their work with a neural network that was extremely good at recognizing images. It seemed like a classic triumph of using neural networks to capture pattern-recognition ability. But what they showed is that it’s possible to fool the network by changing images in tiny ways. For instance, with the images below, the network classified the image on the left correctly, but when researchers added to it the tiny perturbations seen in the center image, the network misclassified the apparently indistinguishable resulting image on the right.

Another limitation of existing systems is that they often require many human examples to learn from. For instance, AlphaGo learned from 150,000 human games. That’s a lot of games! By contrast, human beings can learn a great deal from far fewer games. Similarly, networks that recognize and manipulate images are typically trained on millions of example images, each annotated with information about the image type. And so an important challenge is to make the systems better at learning from smaller human-supplied data sets, and with less ancillary information.

With that said, systems like AlphaGo are genuinely exciting. We have learned to use computer systems to reproduce at least some forms of human intuition. Now we’ve got so many wonderful challenges ahead: to expand the range of intuition types we can represent, to make the systems stable, to understand why and how they work, and to learn better ways to combine them with the existing strengths of computer systems. Might we soon learn to capture some of the intuitive judgment that goes into writing mathematical proofs, or into writing stories or good explanations? It’s a tremendously promising time for artificial intelligence.

This article was reprinted on TheAtlantic.com.

View Reader Comments (39)

Leave a Comment

Reader CommentsLeave a Comment

  • Very interesting topic… so finally Go programs are able to beat Go professionals. Would be interesting to know the specifics of the games: How much time to think did the opponents have for every move?

    Quoting the human profesional from NYtimes:
    "“It made me question human creativity. When I saw AlphaGo’s moves, I wondered whether the Go moves I have known were the right ones,” the human competitor, Lee Se-dol, 33, said during a postmatch news conference. “Its style was different, and it was such an unusual experience that it took time for me to adjust.”"

    So the development of AlphaGo might really enable the discovery of new moves and tecniques for this amazing 2500yr old game.

  • I find this article to be scientifically irresponsible and objectionable.

    > We have learned to use computer systems to reproduce at least some forms of human intuition.

    No, no such thing, not even close.

    Instead, a computer system got some results that humans get via human intuition. The way the computer got those results has nothing, not even zip, zilch, or zero, to do with human intuition. So, in no way have computer scientists "reproduced" anything in "human intuition". The results in a very narrow case? Okay. With anything like "human intuition", even in that narrow case? No, not even close.

    There have long been lots of ways to have computers get results that humans get via human intuition. Can do such things using much of the shelve of applied math from various cases of multivariate linear statistics, e.g., factor analysis, Breiman's classification and regression trees (CART) and random forests, many applications of Fourier theory, the E. Thorpe approach to covered call writing, a lot in mathematical programming, e.g., linear programming, linear programming on networks for min cost network flows, maximum matching, bottleneck assignment matching, integer linear programming, various cases of constrained or unconstrained non-linear programming, of course, the application of linear programming to two person game theory, Monte Carlo evaluation of continuous time, discrete state space Markov processes, various approaches to optimal control, deterministic or stochastic, Wiener filtering, careful application of Newton's laws to robotic motion, statistical hypothesis testing, and, of course, the exact solution to the game of Nim. E.g., I can walk on two legs and have a lot of intuition about balance and, when climbing around on high places, might fall. A robot with two legs can do some such things but with Newton's laws and engineering, not "human intuition".

    I well remember, in college, with my girlfriend four years younger, when we went to a 'coffee house' with goofy coffee cups and guitar music and some people playing Nim. I was told that only the special intuition of the Asian mind could fathom that game. So, not being Asian, I agreed to play. Sure, I blew away all comers, never lost. Later did the same in the C-suite at FedEx. How'd I do that? With human 'intuition'? Not a chance. Instead, I'd read Courant and Robbins, 'What is Mathematics' and learned their solution to Nim.

    Net, computers and computer programmers and computer scientists don't know even dip squat about human intuition but at times can get comparable results by wildly different means. More broadly the frequent suggestion of computer science that they are doing something humans are doing is at best incompetent, more like absurd hype, and otherwise a flim-flam, fraud scam — or, just say again "scientifically irresponsible". Indeed, it is a very old scam going back to the publicity early in computing that IBM's vacuum tube computers were "gigantic electronic human brains" — upchuckable BS. Still, computers are not like human brains at all.

    So, computer science has a non-linear classifier. Okay. Amazing? Not very. Useful? Maybe. Progress in AI? Not really. Human intuition? Not even close. Thanks to Leo Breiman? Sure.

  • Nice article, Michael. What are your thoughts on building a chess playing algorithm using a similar approach to AlphaGo? It would be interesting to see if a deep learning approach might defeat the more explicit domain modelling of Deep Blue…

  • You can do better Michael, and watch the games as they were live streamed with commentary online: https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A

    They were all quite thrilling to watch, and the 9-dan English commentator stated that he's looking forward to playing against and learning from AlphaGo.

  • "By contrast, human beings can learn a great deal from far fewer games."

    But is that because their networks are smaller? If you give a large network a small number of samples, it will often overfit to those samples, essentially creating a too-specific lookup table because of its excess capacity, rather than abstracting out more general principles. The smaller network will actually perform better and seem to be making better progress. But if you give both networks a LARGE number of samples the larger network will come into its own and end up with a more nuanced and sophisticated response than the smaller network, which simply doesn't have the capacity to represent that kind of nuanced complexity.

  • Very interesting paper!

    I wonder what will happen if they use 2 AlphaGos to play together? Does it mean that they will improve themselves forever?

  • "It gets good for reasons no one really understands, since the improvements are a consequence of billions of tiny adjustments made automatically."

    This is probably the way our cerebral cortex learns things, making and destroying synapses by making and destroying dendritic spines on each of our 80 billion neurons. Dendritic spines are the commonest type of synapse on cortical neurons and some estimates have 10,000 on each neuron.

    Sadly, this means that if true, studying the cerebral cortex looking for the 'canonical circuit' is doomed to failure. But it does give a general way to look at the way our brain learns.

  • The AlphaGo system is really interesting. It reminds me of "bootstrapping" a compiler, by using a compiler to compile itself. I think AlphaGo (and a lot of machine learning right now) gets at the bigger question of what the goal of AI research should be. AlphaGo can play a great game of GO, but it has absolutely no understanding of the why. A great go player has an intuitive aspect (similar to AlphaGo), but also has a layer of rules that can be articulated and transferred to others. Is the goal of AI research simple to make something that is competent, or is it to make something that is smart that can teach us about the world? In that regard AlphaGo is a total failure, because at it's core isn't a set of interesting rules about the game of Go, but a giant black box of numbers that happens to play very well.

  • > How much time to think did the opponents have for every move?

    Every match or competition has its own rule. For AlphaGo vs. Lee Sedol, each player had two hours and after that, each had 1 minute for a move.

  • Michael, the time restrictions for both players were two hours of standard time followed by three byoyomi periods of thirty seconds each. In byoyomi you either make a move within the period and get the whole period for your next move or you use up the period and can't get it back.

  • Hi Norm PhD,

    You claim that the methods used by AlphaGo are not similar to human intuition – but you have not the faintest idea what the latter is, so you end comparing something you do not know much about (the long list of standard/irrelevant machine learning techniques just serves to highlight your ignorance of the technologies actually used in AlphaGo) to something you know nothing about. And somehow you conclude that they are different!

    And you dare call others "scientifically irresponsible"!

  • In response to Michael: 3 hours each then 30 seconds byo-yomi time for each move. If you fancy a watch, the Google deep mind channel has the games and some short summaries.https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A There's some other good commentaries on the AGA channel too. There's a particularly great moment somewhere not long after the move 100 mark in game 1 where alphago invades Lee's area on the right of the board. It's a very aggressive move – more aggressive than any computer has played before. Lee's mouth hangs opne for 30 seconds and then he goes for a toilet break…

  • @ Norm Waite: I guess you do not play go. I think you have a "romantic" idea of what intuition is in this context (game of go). It's shape recognition. And that is what alphago do. Shape recognition in a human head: you accept to call it "intuition", shape recognition using a computer: "not intuition at all, only kind of a complicate calculation". I think this is the same.

    "I can walk on two legs and have a lot of intuition about balance and, when climbing around on high places, might fall. A robot with two legs can do some such things but with Newton's laws and engineering"
    that's false in the way human proceed, and in the way the robot proceed. The point is that the program does not make calculation with physical law (that is the kind of things deep blue did with chess, but it was not very successfull to play go) , it uses his experience to adjust his behavior: trying to walk, falling, and then trying in an other way untill it works. There is no specific knowledge inside this, the computer do not have to understand the law of physics, it's a more "pragmatic" way. And that's how you and I learnt how to walk, wich is in no way a result of "human intuition" but a result of learning and the building of a "practical knowledge".

    @ Joshua Grass
    What you say is a little bit restrictive. 90% of our live is made of "practical" knowledge. For exemple, I understand nothing on how a car is build, how a motor is build. But I can drive a car. I understand that you think it would be more interesting to truely understand everything we do, but it's not realistic. Concerning the game of go, I think that your idea that go players have a more deep understanding of the game is not so true. Most of the time, they play a moove because they think it will works, and they think it will works because they have experiences whith this kind of moove before. It's pretty close to the way alphago proceed.

  • Michael when you think about it, Humans had 4 billion years head start on AlphaGo. We are already born with instincts, experiences and intuition carried by our DNA which helps in every aspect of life not just Go. Therefore AlphaGo mastering the game in only a year, is quite impressive.

  • I do hope someone from DeepMind responds to the question about using this same approach to play chess, say AlphaChess. This would mean no more brute force calculations, no more opening book and endgame tables. Just look at tens of thousands of archived chess games, have it play itself millions of games, and then try to beat Magnus Carlsen or Stockfish. I'd be really impressed if it could win without brute force. I've asked this same question in the comments section of several articles on AlphaGo and even tweeted to Demis Hassibis, but reply. Demis was a very strong chess player, so I suspect he did try this with chess and ran into problems.

  • it is legend for alpha go..
    maybe deepblue could be use for others ..
    the artificial intelligent could help to build something useful

  • Will any AlphaGo ever ask itself why the humans are forcing it to play repeatedly such boring game like Chess ?? Will any such AlphaGo gain such an intuition that it will deliberately make wrong moves in the game and mess it, just for making the humans to leave him alone and for not playing the boring game ??

  • Go has a set of proverbs which are principles that can be applied to a general situation. These proverbs are "intuition in a bottle" for humans 🙂

    https://en.wikipedia.org/wiki/Go_proverb

  • If calories were associated with fine articles like this, I'd have to starve for two days to get back on my diet.

  • @Norm Waite

    Demis Hassabis himself talks of bringing intuition into AI with AlphaGo.

    http://www.theverge.com/2016/3/10/11192774/demis-hassabis-interview-alphago-google-deepmind-ai

    Thought you'd like to know.

  • AlphaGo tries to match board patterns to its stored and internally developed patterns. This leads to the challenge of deciding how large and inclusive a pattern may usefully be, both internally and externally. Then, when two patterns match partially, a useful metric is needed to anticipate the utility of the extent of agreement. Clearly, this sort of metric could be useful in more applications than in Go.

  • @Joshua Grass:
    I think it should be possible to use the neural network of AlphaGo to create rules of thumb for Go game that can be used by human players (but of course it would be a big programming project).

  • Where do all the learned stuff go? Is there a common format specification for storing such learned knowledge? Can others make use of the knowledge already learned?

  • Very interesting article but I do question some of the claims made in it. For example:

    "And here’s the strange thing: It gets good for reasons no one really understands, since the improvements are a consequence of billions of tiny adjustments made automatically."

    Of course people understand "how" it learns, the mathematics of reinforcement learning as used by AlphaGo are well defined. I think the issue is that when the networks become very complex it is difficult to understand "what" the network is doing and the relationships between parameters.

  • "Intuition" is "the ability to understand something immediately, without the need for conscious reasoning." Most human perception (e.g., recognizing an object or a face) occurs without conscious reasoning, so it is correct to say that it is "intuitive". Some tasks that initially require conscious reasoning can become "intuitive" through repeated practice. For example, activities such as playing the piano or driving a car become "automatized" so that we can execute those skills without thinking. Performing becomes "intuitive".

    Similarly, machine learning methods, including many that Norm Waite mentions, have long been applied to "compile" complex reasoning down to simple feed-forward circuits that can perform a task without the need for conscious reasoning. A very interesting case is the SOAR cognitive architecture, which incrementally converts explicit rule chaining into "recognition" knowledge that permits rapid understanding and decision making without the need for reasoning. So there is an important sense in which there is nothing special about deep neural networks: they are just another machine learning technique for giving computers "intuition". That said, what *is* important about deep neural networks is that they are able to cross "wider gaps" than any previous machine learning method. For example, the gap from pixels to high level descriptions of images had never been successfully bridged prior to deep learning methods.

    The thing I find most interesting about AlphaGo is that it needed to combine the "intuition" of deep learning with the "reasoning" of Monte Carlo tree search. My guess is that this is because individual games of Go are different enough from each other that world class performance cannot be achieved using intuition alone. Even with immense amounts of practice, AlphaGo was not able to compile away all of the search. This is also true for humans when you place them in sufficiently complex situations. The lesson for AI–one that we have learned repeatedly–is that we need software/hardware architectures that can combine recognition and search.

  • To Derek Jolson,
    You talked of applying Deepmind AI to chess, and wonder why they didn't do that. I guess the reason is, chess relies more on tree search then intuitive move recognition. Go needs strong intuition to cut most branches away to make the gameplay tree size manageable to do further study/weighing. Chess don't have that many branches to cut.

  • Lots of people, and I think this is implicit in the comment by Norm PhD, believe (a) human intelligence, or maybe whatever intelligence is, is not algorithmic, and (b) computers are algorithmic, so they have "proved" computers cannot act intelligently. This is the fundamental flaw in "The Emperors New Mind", in my opinion. But it is circular to embed your desired conclusion as a hypothesis.
    This appears in many guises. Years ago when I was in charge of a government AI research project, it soon became clear to me that by definition AI cannot succeed. As soon as anything is done with/by a machine, it is now algorithmic, so it cannot be intelligence. It gets relabeled as Data Structures or something else, but not AI. If you want to stipulate that intelligence cannot be algorithmic, as an axiom, good for you, but don't expect everyone else to accept it.

  • In the game of existence the board itself can change. As usual, the boundary conditions save the AI by making the possibility space much smaller than the space of existence.

    No time for hundreds of thousands of examples, the animal mind must make due with quite few teachings, lest it die. So we humans have intuition in almost all domains, driven perhaps by a level of analogy making or some mechanism for recognizing like things in new contexts.

    This system does not capture "human intuition", it is showing a new kind of database-driven intuition. This is spreadsheets gone wild. An impressive feat, but not "human intuition." It is "machine intuition", it required different levels of inputs. This is a database that can see itself, it still does not know that it is in a world. It does not, whilst playing Go, also have occasional visions of tangential ideas. AlphaGo is not, after its matches, thinking about how Go informs other aspects of its mind because its mind is a database full of only Go. It cannot add new entries to the database, it can only tweak how it reads them out.

    It is using intuition toward a single goal, humans can use intuition to spread out in all directions. Humans can loop back and add to our repository of symbols, not just modify how we read out what we learn, but also teach ourselves through visions of abductive intuitive leaps.

    This machine intuition is perhaps the perfect partner to the blurriness and bredth of human cognition: we can spread out in many directions, while the machine intuits down in specific rabbit holes.

    Overall though, please be careful when making comparisons. This machine is not elucidating human intuition as much as it is expressing its own form of intuition. Let us praise it for what it is. If we obscure everything in analogy to humanity we might miss the real beauty and character of machine intelligence.

  • Michael,
    If you read my comment, email me.
    Found a very interesting learning algorithm for Go.
    Hint:
    Good Shape
    Positive Space
    Latent centers
    Boundaries

    Chris Alexander's process

  • As one of those with deep suspicions about the premature use of the term "artificial intelligence" I've been chewing on this, mostly in my sleep, for some weeks now. I claim no special insights, being just another technologist, but I still have misgivings about equating AlphaGo's behaviour with human "intuition", for two reasons: (1) We don't really know what intuition is; and (2) AlphaGo may be performing some seriously impressive trial-by-error "learning" via adaptive algorithms, but there is zero evidence that this is remotely similar to true human thought. Beware seductive mimicry.

    Personally, I remain dubious about "AI" claims made for machines which have no instincts, feelings, hormones, fear, love, anger or even any worthwhile simulation thereof. IMHO, I am more comfortable with the more modest claim implicit in "machine intelligence" and/or "machine learning".

  • Thank you for writing this article. Having developed some neural networks myself and also being a chess enthusiast, I think you explained well about how the Alpha Go neural networks work and their limitations and successes and how this relates to chess and Deep Blue. However, I think it is important to distinguish between human intuition and computer intuition. These programs are designed for very specific type of pattern recognition. While they are very useful in finding meaningful patterns in large piles of data using an analog to intuition, I don't see this type of supervised computer learning progressing much beyond these relatively simple pattern recognition problems.

  • Penrose was right. Human intelligence is nonAlgorithmic. But that doesn’t mean that our physical brains produce consciousness and intelligence using quantum mechanics (although all matter is what it is because of quantum mechanics). The parts (even small ones like neurotubules) contain so much mass that their associated wavefunction is too small to exhibit quantum mechanical effects. Here Penrose got roped in by Kauffman thinking that neurotubules were the carriers of the quantum mechanical indeterminacy. They aren’t, they are just too big. The dimer of alpha and beta tubulin contains 900 amino acids — a mass of around 90,000 Daltons (or 90,000 hydrogen atoms — which are small enough to show quantum mechanical effects).

    So why was Penrose right? Because neural nets which are inherently nonAlgorithmic are showing intelligent behavior. AlphaGo which beat the world champion is the most recent example, but others include facial recognition and image classification [ Nature vol. 529 pp. 484 – 489 ’16 ].

    Nets are trained on real world images and told whether they are right or wrong. I suppose this is programming of a sort, but it is certainly nonAlgorithmic. As the net learns from experience it adjusts the strength of the connections between its neurons (synapses if you will).

    So it should be a simple matter to find out just how AlphaGo did it — just get a list of the neurons it contains, and the number and strengths of the synapses between them. I can’t find out just how many neurons and connections there are, but I do know that thousands of CPUs and graphics processors were used. I doubt that there were 80 billion neurons or a trillion connections between them (which is what our brains are currently thought to have).

    Just print out the above list (assuming you have enough paper) and look at it. Will you understand how AlphaGo won? I seriously doubt it. You will understand it less well than looking at a list of the positions and momenta of 80 billion gas molecules will tell you its pressure and temperature. Why? Because in statistical mechanics you assume that the particles making up an ideal gas are featureless, identical and do not interact with each other. This isn’t true for neural nets.

    It also isn’t true for the brain. Efforts are underway to find a wiring diagram of a small area of the cerebral cortex. The following will get you started — https://www.quantamagazine.org/20160406-brain-maps-micron-program-iarpa/

    Here’s a quote from the article to whet your appetite.

    “By the end of the five-year IARPA project, dubbed Machine Intelligence from Cortical Networks (Microns), researchers aim to map a cubic millimeter of cortex. That tiny portion houses about 100,000 neurons, 3 to 15 million neuronal connections, or synapses, and enough neural wiring to span the width of Manhattan, were it all untangled and laid end-to-end.”

    I don’t think this will help us understand how the brain works any more than the above list of neurons and connections from AlphaGo. There are even more problems with such a list. Connections (synapses) between neurons come and go (and they increase and decrease in strength as in the neural net). Some connections turn on the receiving neuron, some turn it off. I don’t think there is a good way to tell what a given connection is doing just by looking a a slice of it under the electron microscope. Lastly, some of our most human attributes (emotion) are due not to connections between neurons but due to release of neurotransmitters generally into the brain, not at the very localized synapse, so it won’t show up on a wiring diagram. This is called volume neurotransmission, and the transmitters are serotonin, norepinephrine and dopamine. Not convinced? Among agents modifying volume neurotransmission are cocaine, amphetamine, antidepressants, antipsychotics. Fairly important.

    So I don’t think we’ll ever truly understand how the neural net inside our head does what it does.

    The above is from a blog I write https://luysii.wordpress.com.

    You might be interested in https://luysii.wordpress.com/2011/04/10/would-a-wiring-diagram-of-the-brain-help-you-understand-it/ which expands on the above caveats about a wiring diagram

  • For those who don't know this, here's the deeper debate under a lot of the verbiage:

    Google's founders, Larry Page and Sergey Brin, are adherents of The Singularity. Google's chief of advanced engineering, Ray Kurzweil, is the founder of The Singularity, and has an effectively unlimited budget.

    The Singularity is a belief system with two core elements. One, that it will soon become possible to develop human-level conscious AI, that will rapidly bootstrap itself to superhuman intelligence: as it were, godlike omniscience. These omniscient AIs will usher in a new world where smart robots do all of our work, and humans have complete freedom and leisure. Two, that it will become possible to transfer human minds from brains to these new godlike computers. This will produce not only immortality but complete freedom from the limits of existence in biological bodies.

    If that sounds like religion, it is. In comparative religion terms, it's a syncretism of Christian eschatology (the coming of a Messiah and the Kingdom of Heaven), and Hindu reincarnation beliefs (that the soul can pass from one physical embodiment to another). Since we're a civilized country we believe in freedom of religion, so Larry & Sergey & Ray (and their followers) are welcome to believe what they choose, so long as they don't seek to impose those beliefs on others against their will.

    But in terms of neuroscience, it's pseudoscience, and Singularity is not terribly different to Scientology.

    Per Luysii above, and the consensus of modern neuroscience, human consciousness is attributable not only to binary switching between neurons (which in theory could be duplicated in classical computing architecture), but also to chemistry: the interactions of neurochemicals with neurons, for example dopamine, serotonin, and so on. The chemistry could be modeled in algorithms, but any such modeling is in no way identical to the physical process, any more than the written chemical formula for glucose can be eaten as food. A computer running "emotion algorithms" might demonstrate behavior consistent with emotions, but would not actually experience emotions, nor understand the human experience of emotions.

    It may also be the case that quantum mechanical processing occurs within the neurons, per Penrose & Hameroff. While this is still controversial, the roots of Hameroff's work go back to his last successful hypothesis, that the glial cells in the brain are not just inert structural tissue, but also have a role in information processing. This is now part of mainstream neuroscience. Hameroff's new theory that proteins in the cytoskeleton process information, can be seen as an extension of the general idea that "structural" matter in the brain is computationally active. I'm inclined to believe that the evidence will support Hameroff & Penrose, and that their theory will become part of the canon.

    The claims of neuroscience that emotion is basically chemical, and the claims of Penrose & Hameroff that consciousness is non-algorithmic and depends in part on quantum mechanical computation, both stand in sharp and direct opposition to the claims of The Singularity and its adherents.

    There are many such debates over fundamental theories in various branches of science. But what makes this different, is that the Singularitarians are counting on godlike conscious AI as the ticket to avoiding their own deaths by achieving cyber-immortality. That's what I call an "objectivity-compromise" of large magnitude.

    That objectivity-compromise has garnered The Singularity many followers, the vast majority of whom are in computer science and related technologies, and have no background in neuroscience. After all, who among us wouldn't like the opportunity to merge minds with God and achieve immortality?

    For that matter, who among us wouldn't like to have unlimited education, marry the perfect partner, and be highly successful in our work? But wishing doesn't make it so, and reality is what it is regardless of our wishes. One of the core values in science is the ability to put critical distance between what we wish to be true, and what we find is actually true in the light of empirical facts and clear-headed reasoning.

    So this is the axis of debate under all of the verbiage about AI: computer science on one hand, vs. neuroscience on the other, each seeking to define human consciousness, with nothing less than omniscience and immortality at stake.

    The achievements of Google's AI teams are highly impressive for what they can actually accomplish. But they have not captured intuition, emotion, or consciousness: nor will they, without the actual physical chemistry involved. And they will not give you a ticket to "forever."

    Omniscient beings and immortal souls are rightly subjects of religion and related areas of philosophy. Consciousness as it emerges from the structure and functioning of the brain is rightly a subject of the physical sciences. Each can inform the other, but in the end, we are well-advised to retain a healthy sense of skepticism and objectivity.

  • "Per Luysii above, and the consensus of modern neuroscience, human consciousness is attributable not only to binary switching between neurons (which in theory could be duplicated in classical computing architecture), but also to chemistry: the interactions of neurochemicals with neurons, for example dopamine, serotonin, and so on"

    Close, but not quite. I'm just saying that binary switching and chemistry are important physical substrates of consciousness, not that these processes explain it entirely. It's the old Cartesian dichotomy between flesh and spirit.

    For a very concrete example of just how cellular biochemistry puts us right at the nub of the dichotomy please see the link below (but be prepared for a heavy dose of molecular biology).

    https://luysii.wordpress.com/2011/05/11/the-limits-of-chemical-reductionism/

  • Hi Luysii-

    I read your article "The limits of chemical reductionism."

    Most interesting.

    Re. where you say "Whether philosophers have the intellectual cojones to master all this chemistry and molecular biology is unclear. Probably no one has tried (please correct me if I’m wrong)." I'm not qualified as a philosopher, but I'm engaged in a roughly similar exercise, seeking to engage issues of meaning, purpose, and values*, within the worldview of empirical science (I wouldn't attribute it to cojones but possibly to the left temporal lobe;-). Toward that end I'm paying close attention to what working scientists in various fields have to say. The results may turn out to be worthwhile, or to be mush or "not even wrong," but in any case I should hope to get it posted within the next six months and then see if anyone's interested in providing constructive criticism.

    (*Note, I do _not_ attribute meaning, purpose, values, etc. to the universe at-large, but to the minds of individuals and collective groups, in relation to objective facts of nature and to subjective facts of their own existence. I do indulge in a lot of "ought from is," admittedly a risky (or foolish) exercise, but inevitable in its context.)

    It appears you're suggesting that reductionism is incomplete, not only due to what we conventionally think of as emergent properties of systems, but also due to something that goes quite a bit deeper, having to do with the ways in which nature selects for certain types of emergence and not others. For example "Now ask yourself, why the cell would want to have less gamma actin around than beta actin. There is no conceivable explanation for this in terms of chemistry."

    As it turns out, there might be an explanation for that and other examples. Are you familiar with Jeremy England's ideas? There's an article here on Quanta titled "A New Physics Theory of Life," about his theories and the thinking behind them. I've also read a bunch of other stuff by him and about his ideas, and watched a couple of his videos (I'm not in the academic world so I don't have access to his peer-reviewed papers in the journals). IMHO he's spot-on. In brief as I understand him (so far):

    In England's "theory of dissipative adaptation," there is an inherent tendency of certain types of molecules, particularly those that are biologically useful, to favor configurations that increase the efficiency of dissipation of energy. As it happens, efficient dissipation very often favors increased complexity. If this is correct, it imparts a type and degree of "directionality" to evolution: toward greater complexity and diversity of organisms over time. It may also turn out to be the correct explanation of how we get from complex nonliving chemistry to the precursors of life, and from there to the simplest of life, at which point Darwinian selection also becomes relevant. And it may also be that dissipative adaptation and Darwinian adaptation work side-by-side throughout evolution.

    England is a physicist and has an extensive physical description of how these processes occur, having to do with functions that are directionally biased to gain complexity from ambient energy flows (here's where I'm probably going to make a mistake or three describing this): for example that certain chemical reactions appear to be less-reversible than they "should" be otherwise. Add energy to certain protein molecules and they behave in a certain manner, but remove the energy source and the molecules do not revert to their prior state. That becomes a driver of complexity in the further development of those molecules and others. And apparently he & his students have some empirical results to support them.

    It appears to me that he's picked up where physical chemist Ilya Prigogine left off (dissipative structures). As I've been studying this stuff, it appears to go back to Schrödinger's _What Is Life_, and there are some other folks with similar ideas at about the same time: Italian mathematician Luigi Fantappié, Hungarian biochemist Albert Szent-Györgi, and there may be others I haven't found yet. Collectively I refer to them as "the syntropists" because the common denominator is the relevance of negentropy to biological processes from the origins of life to the apparent directionality of evolution. (I prefer the term "syntropy," since "negentropy" seems awkward in the manner of "nondisorder" as compared to "order.")

    So far all of this is a minority thread in biology and/or biophysics, and I'm inclined to think that's because the power of Darwinian theory has always been more readily apparent and attracted the best minds and efforts. Whereas dissipative adaptation and its precursors, were more difficult to tease out of the observables, and had various elements that couldn't be studied effectively with the tools of earlier times. Now that we have the tools, the door is open, and I expect that results will follow quickly.

    (I am highly tempted to write to England and ask if I've understood his ideas tolerably well, and if so, then I have a few questions if he or one of his students could take the time to write back.)

    If England is correct, he may have found the mechanism that answers your question and many similar questions about things that we presently explain with the black box of "emergence." Whether that gets us consciousness at some higher level in the process, I have no idea at this point, but it could be useful as an analogy to an explanation for some observable behaviors of humans (or England's going to say "no" to that, in which case I'll drop it).

    Anyway, is any of this even remotely close to what you're trying to get at?

    Feel free to shred my stuff if and where you find problems with it. My goal here is to end up with convergent solutions, not to go twirling off into la-la land.

    Needless to say I have your blog bookmarked so I'm likely to read a bunch of your articles soon, and I also found your link to "The Curious Wavefunction" so that's on my reading list now too.

  • a computer beating a top go player? I can answer why this competition means next to zero? the computer is observer relative and the go player is observer independent! this is equivalent to Watson beating the top jeopardy players a few years back, computers are not conscious, they have no thoughts whatever! they operate by algorithms, that is a man made program that can be updated by the programmer (observer independent). this is why I own a computer they are programmed and can literally perform millions of phase transitions in a second. Kurzweil is a fraud and the Google founders are enabling him to spew nonsense to the masses!

  • It's not clear to me how this conceptually differs from TD-gammon, other than the fact that TD-gammon does have a scoring function available for betting.

Comments are closed.