Kleanthes sent me this link to a recent lecture by Gary Marcus (GM) on the status of current AI research. It is a somewhat jaundiced review concluding that, once again, the results have been strongly oversold. This should not be surprising. The rewards to those that deliver strong AI (“the kind of AI that would be as smart as, say a Star Trek computer” (3)) will be without limit, both tangibly (lots and lots of money) and spiritually (lots and lots of fame, immortal kinda fame). And given hyperbole never cripples its purveyors (“AI boys will be AI boys” (and yes, they are all boys)), it is no surprise that, as GM notes, we have been 20 years out from solving strong AI for the last 65 years or so. This is a bit like the many economists who predicted 15 of the last 6 recessions but worse. Why worse? Because there have been 6 recessions but there has been pitifully small progress on strong AI, at least if GM is to be believed (and I think he is).
Why despite the hype (necessary to drain dollars from “smart” VC money) has this problem been so tough to crack? GM mentions a few reasons.
First, we really have no idea how open ended competence works. Let me put this backwards. As GM notes, AI has been successful precisely in “predefined domains” (6). In other words, where we can limit the set of objects being considered for identification or the topics up for discussion or the hypotheses to be tested we can get things to run relatively smoothly. This has been true since Winograd and his block worlds. Constrain the domain and all goes okishly. Open the domain up so that intelligence can wander across topics freely and all hell breaks loose. The problem of AI has always been scaling up, and it is still a problem. Why? Because we have no idea how intelligence manages to (i) identify relevant information for any given domain and (ii) use that information in relevant ways for that domain. In other words, how we in general figure out what counts and how we figure out how much it counts once we have figured it out is a complete and utter mystery. And I mean ‘mystery’ in the sense that Chomsky has identified (i.e. as opposed to ‘problem’).
Nor is this a problem limited to AI. As FoL has discussed before, linguistic creativity has two sides. The part that has to do with specifying the kind of unbounded hierarchical recursion we find in human Gs has been shown to be tractable. Linguists have been able to say interesting things about the kinds of Gs we find in human natural languages and the kinds of UG principles that FL plausibly contains. One of the glories (IMO, the glory) of modern GG lies in its having turned once mysterious questions into scientific problems. We may not have solved all the problems of linguistic structure but we have managed to render them scientifically tractable.
This is in stark contrast to the other side linguistic creativity: the fact that humans are able to use their linguistic competence in so many different ways for thought and self-expression. This is what the Cartesians found so remarkable (see here for some discussion) and that we have not made an iota of progress understanding. As Chomsky put it in Language & Mind (and is still a fair summary of where we stand today):
Honesty forces us to admit that we are as far today as Descartes was three centuries ago from understanding just what enables a human to speak in a way that is innovative, free from stimulus control, and also appropriate and coherent. (12-13)
All-things-considered judgments, those that we deploy effortlessly in every day conversation, elude insight. That we do this is apparent. But how we do this remains mysterious. This is the nut that strong AI needs to crack given its ambitions. To date, the record of failure speaks for itself and there is no reason to think that more modern methods will help out much.
It is precisely this roadblock that limiting the domain of interest removes. Bound the domain and the problem of open-endedness disappears.
This should sound familiar. It is the message in Fodor’s Modularity of Mind. Fodor observes that modularity makes for tractability. When we move away from modular systems, we flat on our faces precisely because we have no idea how minds identify what is relevant in any given situation and how it weights what is relevant in a given situation and how it then deploys this information appropriately. We do it all right. We just don’t know how.
The modern hype supposes that we can get around this problem with big data. GM has a few choice remarks about this. Here’s how he sees things (my emphasis):
I opened this talk with a prediction from Andrew Ng: “If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future.” So, here’s my version of it, which I think is more honest and definitely less pithy: If a typical person can do a mental task with less than one second of thought and we can gather an enormous amount of directly relevant data, we have a fighting chance, so long as the test data aren’t too terribly different from the training data and the domain doesn’t change too much over time. Unfortunately, for real-world problems, that’s rarely the case. (8)
So, if we massage the data so that we get that which is “directly relevant” and we test our inductive learner on data that is not “too terribly different” and we make sure that the “domain doesn’t change much” then big data will deliver “statistical approximations” (5). However, “statistics is not the same thing as knowledge” (9). Big data can give us better and better “correlations” if fed with “large amounts of [relevant!, NH] statistical data”. However, even when these correlational models work, “we don’t necessarily understand what’s underlying them” (9).
And one more thing: when things work it’s because the domain is well behaved. Here’s GM on AlphaGo (my emphasis):
Lately, AlphaGo is probably the most impressive demonstration of AI. It’s the AI program that plays the board game Go, and extremely well, but it works because the rules never change, you can gather an infinite amount of data, and you just play it over and over again. It’s not open-ended. You don’t have to worry about the world changing. But when you move things into the real world, say driving a vehicle where there’s always a new situation, these techniques just don’t work as well. (7)
So, if the rules don’t change, you have unbounded data and time to massage it and the relevant world doesn’t change, then we can get something that approximately fits what we observe. But fitting is not explaining and the world required for even this much “success” is not the world we live in, the world in which our cognitive powers are exercised. So what does AI’s being able to do this in artificial worlds tell us about what we do in ours? Absolutely nothing.
Moreover, as GM notes, the problems of interest to human cognition have exactly the opposite profile. In Big Data scenarios we have boundless data, endless trials with huge numbers of failures (corrections). The problems we are interested in are characterized by having a small amount of data and a very small amount of error. What will Big Data techniques tell us about problems with the latter profile? The obvious answer is “not very much” and the obvious answer, to date, has proven to be quite adequate.
Again, this should sound familiar. We do not know how to model the everyday creativity that goes into common judgments that humans routinely make and that directly affects how we navigate our open-ended world. Where we cannot successfully idealize to a modular system (one that is relatively informationally encapsulated) we are at sea. And no amount of big data or stats will help.
What GM says has been said repeatedly over the last 65 years. AI hype will always be with us. The problem is that it must crack a long lived mystery to get anywhere. It must crack the problem of judgment and try to “mechanize” it. Descartes doubted that we would be able to do this (indeed this was his main argument for a second substance). The problem with so much work in AI is not that it has failed to crack this problem, but that it fails to see that it is a problem at all. What GM observes is that, in this regard, nothing has really changed and I predict that we will be in more or less the same place in 20 years.
Since penning(?) the above I ran across a review of a book on machine intelligence by Gary Kasparov (here). The review is interesting (I have not read the book) and is a nice companion to the Marcus remarks. I particularly liked the history on Shannon’s early thoughts on chess playing computers and his distinction on how the problem could be solved:
At the dawn of the computer age, in 1950, the influential Bell Labs engineer Claude Shannon published a paper in Philosophical Magazine called “Programming a Computer for Playing Chess.” The creation of a “tolerably good” computerized chess player, he argued, was not only possible but would also have metaphysical consequences. It would force the human race “either to admit the possibility of a mechanized thinking or to further restrict [its] concept of ‘thinking.’” He went on to offer an insight that would prove essential both to the development of chess software and to the pursuit of artificial intelligence in general. A chess program, he wrote, would need to incorporate a search function able to identify possible moves and rank them according to how they influenced the course of the game. He laid out two very different approaches to programming the function. “Type A” would rely on brute force, calculating the relative value of all possible moves as far ahead in the game as the speed of the computer allowed. “Type B” would use intelligence rather than raw power, imbuing the computer with an understanding of the game that would allow it to focus on a small number of attractive moves while ignoring the rest. In essence, a Type B computer would demonstrate the intuition of an experienced human player.
As the review goes on to note, Shannon’s mistake was to think that Type A computers were not going to materialize. They did, with the result that the promise of AI (that it would tell us something about intelligence) fizzled as the “artificial” way that machines became “intelligent” simply abstracted away from intelligence. Or, to put it as Kasparov is quoted as putting it: “Deep Blue [the machine that beat Kasparov, NH] was intelligent the way your programmable alarm clock is intelligent.”
So, the hope that AI would illuminate human cognition rested on the belief that technology and brute calculation would not be able to substitute for “intelligence.” This proved wrong, with machine learning being the latest twist in the same saga, per the review and Kasparov.
All this fits with GM’s remarks above. What both do not emphasize enough, IMO, is something that many did not anticipate; namely that we would revamp our views of intelligence rather than question whether our programs had it. Part of the resurgence of Empiricism is tied to the rise of the technologically successful machine. The hope was that trying to get limited machines to act like we do might tell us something about how we do things. The limitations of the machine would require intelligent design to get it to work thereby possibly illuminating our kind of intelligence. What happened is that getting computationally miraculous machines to do things in ways that we had earlier recognized as dumb and brute force (and so telling us nothing at all) has transformed into the hypothesis that there is no such things as real intelligence at all and everything is “really” just brute force. Thus, the brain is just a data cruncher, just like Deep Blue is. And this shift in attitude is supported by an Empiricist conception of mind and explanation. There is no structure to the mind beyond the capacity to mine the inputs for surfacy generalizations. There is no structure to the world beyond statistical regularities. On this Eish viw, AI has not failed, rather the right conclusion is that there is less to thinking than we thought. This invigorated Empiricism is quite wrong. But it will have staying power. Nobody should underestimate the power that a successful (money making) tech device can have on the intellectual spirit of the age.
All-things-considered judgments, those that we deploy effortlessly in every day conversation, elude insight. That we do this is apparent. But how we do this remains mysterious. This is the nut that strong AI needs to crack given its ambitions. To date, the record of failure speaks for itself and there is no reason to think that more modern methods will help out much.ReplyDelete
This point about the challenges of everyday conversation was echoed recently by Microsoft's Satya Nadella in an interview with Quartz: "Before we even generate language, let us understand turn by turn dialogue. (…) Whenever you have ambiguity and errors, you need to think about how you put the human in the loop. That to me is the art form of an AI product."
While you connect 'error' with big data etc., Nadella's point is more subtle. Dealing with errors in interaction is something AI hasn't cracked yet, and it is where human competences shines. No other species can recover so gracefully from troubles of speaking, hearing and understanding. No other species, as far as we know, has a communication system that allows enough self-referentiality to carry out conversational repair as smoothly and frequently as we do it.
The second-most cited paper published in Language is Chomsky's review of Verbal Behavior (Chomsky 1959). It makes some interesting points about conversation, though these are often overlooked. For instance, there is the implicit realisation that even though this is clearly not behaviour under the control of some stimulus as Skinner envisaged, there are nonetheless rules of relevance, coherence and sequence that structure contributions to conversation.
The most cited paper published in Language (with twice the number of citations of Chomsky's review) is a study of the turn-taking system of informal conversation — using raw performance data to discover rules of conversational competence (Sacks et al. 1974). The field of conversation analysis has gone on to explore rules and regularities of sequence organization, repair, and so on — discovering and explaining some of the mechanisms by which people make meaning together in open-ended yet orderly ways. So some headway has surely been made.
Maybe the failure of AI isn't so much it's obsession with big data, but that it's forgot the problem that big data is supposed to solve.ReplyDelete
Go back far enough and you discover that before the early logicists were making particular claims about which particular kinds of logics can do the job of generating protocol sentences best, you have the idea of a logic as presenting us with a kind of world with its own individual set of implications which we can tease out and examine: logic was about exploring the world revealed by our most basic statements.
Big data doesn't imagine a world. Its concepts are flat points on a field which just don't have the semantic depth of a world. Sure, this another way of making Shannon's point about Type A vs Type B intelligence (and for that matter your own point about human intuition), but it still seems to me like a more optimistic way of putting it: we can at least imagine a computer possessing a kind of internal world.