Archive for March, 2010

Picture: Martin Heidegger. This paper by Dotov, Nie, and Chemero describes experiments which it says have pulled off the remarkable feat of providing empirical, experimental evidence for Heidegger’s phenomenology, or part of it; the paper has been taken by some as providing new backing for the Extended Mind theory, notably expounded by Andy Clark in his 2008 book (‘Supersizing the Mind’).

Relating the research so strongly to Heidegger puts it into a complex historical context. Some of Heidegger’s views, particularly those which suggest there can be no theory of everyday life, have been taken up by critics of artificial intelligence. Hubert Dreyfus in particular, has offered a vigorous critique drawing mainly from Heidegger an idea of the limits of computation, one which strongly resembles those which arise from the broadly-conceived frame problem, as discussed here recently. The authors of the paper claim this heritage, accepting the Dreyfusard view of Heidegger as an early proto-enemy of GOFAI .

For it is GOFAI (Good Old Fashioned Artificial Intelligence) we’re dealing with. The authors of the current paper point out that the Heideggerian/Dreyfusard critique applies only to AI based on straightforward symbol manipulation (though I think a casual reader of Dreyfus  could well be forgiven for going away with the impression that he was a sceptic about all forms of AI), and that it points toward the need to give proper regard to the consequences of embodiment.

Hence their two experiments. These are designed to show objective signs of a state described by Heidegger, known in English as ‘ready-to-hand’. This seems a misleading translation, though I can’t think of a perfect alternative. If a hammer is ‘ready to hand’, I think that implies it’s laid out on the bench ready for me to pick it up when I want it;  the state Heidegger was talking about is the one when you’re using the hammer confidently and skilfully without even having to think about it. If something goes wrong with the hammering, you may be forced to start thinking about the hammer again – about exactly how it’s going to hit the nail, perhaps about how you’re holding it. You can also stop using the hammer altogether and contemplate it as a simple object. But when the hammer is ready-to-hand in the required sense, you naturally speak of your knocking in a few nails as though you were using your bare hands, or more accurately, as if the hammer had become part of you.

Both experiments were based on subjects using a mouse to play a simple game.  The idea was that once the subjects had settled, the mouse would become ready-to-hand; then the relationship between mouse movement and cursor movement would be temporarily messed up; this should cause the mouse to become unready-to-hand for a while. Two different techniques were used to detect readiness-to-hand. In the first experiment the movements of the hand and mouse were analysed for signs of 1/f? noise. Apparently earlier research has established that the appearance of 1/f? noise is a sign of a smoothly integrated system.  The second experiment used a less sophisticated method; subjects were required to perform a simple counting task at the same time as using the mouse; when their performance at this second task faltered, it was taken as a sign that attention was being transferred to cope with the onset of unreadiness to hand. Both experiments yielded the expected results.  (Regrettably some subjects were lost because of an unexpected problem – they weren’t good enough at the simple mouse game to keep it going for the duration of the experiment. Future experimenters should note the need to set up a game which cannot come to a sudden halt.)

I think the first question which comes to mind is: why were the experiments were even necessary?  It is a common experience that tools or vehicles become extensions of our personality; in fact it has often been pointed out that even our senses get relocated. If you use a whisk to beat eggs, you sense the consistency of the egg not by monitoring the movement of the whisk against your fingers, but as though you were feeling the egg with the whisk, as though there was a limited kind of sensation transferred into the whisk. Now of course, for any phenomenological observation, there will be some diehards who deny having had any such experience; but my impression is that this sort of thing is widely accepted, enough to feature as a proposition in a discussion without further support.  Nevertheless, it’s true that it this remains subjective, so it’s a fair claim that empirical results are something new.

Second, though, do the results actually prove anything? Phenomenologically, it seems possible to me to think of alternative explanations which fit the bill without invoking readiness-to-hand. Does it seem to the subject that the mouse has become part of them, part of a smoothly-integrated entity – or does the mouse just drop out of consciousness altogether? Even if we accept that the presence of 1/f? noise shows that integration has occurred, that doesn’t give us readiness-to-hand (or if it does, it seems the result was already achieved by the earlier research).

In the second experiment we’ve certainly got a transfer of attention – but isn’t that only natural? If a task suddenly becomes inexplicably harder, it’s not surprising that more attention is devoted to it – surely we can explain that without invoking Heidegger? The authors acknowledge this objection, and if I understand correctly suggest that the two tasks involved were easy enough to rule out problems of excessive cognitive load so that, I suppose, no significant switch of attention would have been necessary if not for the breakdown of readiness-to-hand.  I’m not altogether convinced.

I do like the chutzpah involved in an experimental attempt to validate Heidegger, though, and I wouldn’t rule out the possibility that bold and ingenious experiments along these lines might tell us something interesting.

Picture: Google chatbot. Bitbucket I was interested to see this Wired piece recently; specifically the points about how Google picks up contextual clues. I’ve heard before about how Google’s translation facilities basically use the huge database of the web: instead of applying grammatical rules or anything like that, they just find equivalents in parallel texts, or alternatives that people use when searching, and this allows them to do a surprisingly good – not perfect – job of picking up those contextual issues that are the bane of most translation software. At least, that’s my understanding of how it works.  Somehow it hadn’t quite occurred to me before, but a similar approach lends itself to the construction of a pretty good kind of chatbot – one that could finally pass the Turing Test unambiguously.

Blandula Ah, the oft-promised passing of the Turing Test. Wake me up when it happens – we’ve been round this course so many times in the past.

Bitbucket Strangely enough, this does remind me of one of the things we used to argue about a lot in the past.  You’ve always wanted to argue that computers couldn’t match human performance in certain respects in principle. As a last resort, I tried to get you to admit that in principle we could get a computer to hold a conversation with human-level responses just by the brutest of brute force solutions.  You just can a perfect response for every possible sentence. When you get that sentence as input, you send the canned response as output. The longest sentence ever spoken is not infinitely long, and the number of sentences of any finite length is finite; so in principle we can do it.

Blandula I remember: what you could never grasp was that the meaning of a sentence depends on the context, so you can’t devise a perfect response for every sentence without knowing what conversation it was part of.  What would the canned response be to;  ‘What do you mean?’  – to take just one simple example.

Bitbucket What you could never grasp was that in principle we can build in the context, too. Instead of just taking one sentence, we can have a canned response to sets of the last ten sentences if we like – or the last hundred sentences, or whatever it takes. Of course the resources required get absurd, but we’re talking about the principle, so we can assume whatever resources we want.  The point I wanted to make is that by using the contents of the Internet and search enquiries, Google could implement a real-world brute-force solution of broadly this kind.

Blandula I don’t think the Internet actually contains every set of a hundred sentences ever spoken during the history of the Universe.

Bitbucket No, granted; but it’s pretty good, and it’s growing rapidly, and it’s skewed towards the kind of thing that people actually say. I grant you that in practice there will always be unusual contextual clues that the Google chatbot won’t pick up, or will mishandle. But don’t forget that human beings miss the point sometimes, too.  It seems to me a realistic aspiration that the level of errors could fairly quickly be pushed down to human levels based on Internet content.

Blandula It would of course tell us nothing whatever about consciousness or the human mind; it would just be a trick. And a damaging one.  If Google could fake human conversation, many people would ascribe consciousness to it, however unjustifiably. You know that quite poor, unsophisticated chatbots have been treated by naive users as serious conversational partners ever since Eliza, the grandmother of them all. The internet connection makes it worse, because a surprising number of people seem to think that the Internet itself might one day accidentally attain consciousness. A mad idea: so all those people working on AI get nowhere, but some piece of kit which is carefully designed to do something quite different just accidentally hits on the solution? It’s as though Jethro Tull had been working on his machine and concluded it would never be a practical seed-drill; but then realised he had inadvertently built a viable flying machine. Not going to happen. Thing is, believing some machine is a person when it isn’t is not a trivial matter, because you then naturally start to think of people as being no more than machines.  It starts to seem natural to close people down when they cease to be useful, and to work them like slaves while they’re operative. I’m well aware that a trend in this direction is already established, but a successful chatbot would make things much, much, worse.

Bitbucket Well, that’s a nice exposition of the paranoia which lies behind so many of your attitudes. Look, you can talk to automated answering services as it is: nobody gets het up about it, or starts to lose their concept of humanity.

Of course you’re right that a Google chatbot in itself is not conscious. But isn’t it a good step forward?  You know that in the brain there are several areas that deal with speech;  Broca’s area seems to put coherent sentences together while Wernicke’s area provides the right words and sense. People whose Wernicke’s area has been destroyed, but who still have a sound Broca’s area apparently talk fluently and sort of convincingly, but without ever really making sense in terms of the world around them. I would claim that a working Google chatbot is in essence a Broca’s area for a future conscious AI. That’s all I’ll claim, just for the moment.

Picture: global workspace. Global Workspace theories have been popular ever since Bernard Baars put forward the idea back in the eighties; in ‘Applying global workspace theory to the frame problem’*,  Murray Shanahan and Baars suggest that among its other virtues, the global workspace provides a convenient solution to that old bugbear, the frame problem.

What is the frame problem, anyway? Initially, it was a problem that arose when early AI programs were attempting simple tasks like moving blocks around. It became clear that when they  moved a block, they not only had to update their database to correct the position of the block, they had to update every other piece of information to say it had not been changed. This led to unexpected demands on memory and processing. In the AI world, this problem never seemed too overwhelming, but philosophers got hold of it and gave it a new twist. Fodor, and in a memorable exposition, Dennett, suggested that there was a fundamental problem here. Humans had the ability to pick out what was relevant and ignore everything else, but there didn’t seem to be any way of giving computers the same capacity. Dennett’s version featured three robots: the first happily pulled a trolley out of a room to save it from a bomb, without noticing that the bomb was on the trolley, and came too; the second attempted to work out all the implications of pulling the trolley out of the room; but there were so many logical implications that it was stuck working through them when the bomb went off. The third was designed to ignore irrelevant implications, but it was still working on the task of identifying all the many irrelevant implications when again the bomb exploded.

Shanahan and Baars explain this background and rightly point out that the original frame problem arose in systems which used formal logic as their only means of drawing conclusions about things, no longer an approach that many people would expect to succeed. They don’t really believe that the case for the insolubility of the problem has been convincingly made. What exactly is the nature of the problem, they ask: is it combinatorial explosion? Or is it just that the number of propositions the AI has to sort through to find the relevant one is very large (and by the way, aren’t there better ways of finding it than searching every item in order?). Neither of those is really all that frightening; we have techniques to deal with them.

I think Shanahan and Baars, understandably enough, under-rate the task a bit here. The set of sentences we’re asking the AI to sort through is not just very large; it’s infinite. One of the absurd deductions Dennett assigns to his robots is that the number of revolutions the wheels of trolley will perform in being pulled out of the room is less than the number of walls in the room. This is clearly just one member of a set of valid deductions which goes on forever; the number of revolutions is also less than the number of walls plus one; it’s less than the number of walls plus two… It may be obvious that these deductions are uninteresting; but what is the algorithm that tells us so? More fundamentally, the superficial problems are proxies for a deeper concern; that the real world isn’t reducible to a set of propositions at all, that, as Borges put it

“it is clear that there is no classification of the Universe that is not arbitrary and full of conjectures. The reason for this is very simple: we do not know what thing the universe is.”

There’s no encyclopaedia which can contain all possible facts about any situation. You may have good heuristics and terrific search algorithms, but when you’re up against an uncategorisable domain of infinite extent, you’re surely still going to have problems.

However, the solution proposed by Shanahan and Baars is interesting. Instead of the mind having to search through a large set of sentences, it has a global workspace where things are decided and a series of specialised modules which compete to feed in information (there’s an issue here about how radically different inputs from different modules manage to talk to each other: Shanahan and Baars mention a couple of options and then say rather loftily that the details don’t matter for their current purposes. It’s true that in context we don’t need to know exactly what the solution is – but we do need to be left believing that there is one).

Anyway, the idea is that while the global workspace is going about its business each module is looking out for just one thing. When eventually the bomb-is-coming-too module gets stimulated, it begins sending very vigorously and that information gets into the workspace. Instead of having to identify relevant developments, the workspace is automatically fed with them.

That looks good on the face of it; instead of spending time endlessly sorting through propositions, we’ll just be alerted when it’s necessary. Notice, however, that instead of requiring an indefinitely large amount of time, we now need an indefinitely large number of specialised modules. Moreover, if we really cover all the bases, many of those modules are going to be firing off all the time. So when the bomb-is-coming-too module begins to signal frantically, it will be competing with the number-of-rotations-is-less-than-the-number-of-walls module and all the others, and will be drowned out. If we only want to have relevant modules, or only listen to relevant signals, we’re back with the original problem of determining just what is relevant.

Still, let’s not dismiss the whole thing too glibly. It reminded me to some degree of Edelman’s analogy with the immune system, which in a way really does work like that. The immune system cannot know in advance what antibodies it will need to produce, so instead it produces lots of random variations; then when one gets triggered it is quickly reproduced in large numbers. Perhaps we can imagine that if the global workspace were served by modules which were not pre-defined, but arose randomly out of chance neural linkages, it might work something like that. However, the immune system has the advantage of knowing that it has to react against anything foreign, whereas we need relevant responses for relevant stimuli. I don’t think we have the answer yet.

*Thanks to Lloyd for the reference.