Linespace
Cycosaurus breaks free
Linespace

12 May 2005

home

Cycosaurus

A recent New Scientist piece announced that within the next few months, CYC would be unleashed on the Internet. The CYC project (pronounced 'psych', but derived from 'en cyclopedia') dates back to an earlier, more optimistic era of Artificial Intelligence; it was the most confident of those based on 'brute force' - the assumption that the problems could be conquered by huge databases and massive computing power, rather than requiring some revolutionary new insight. The survival of the project for more than twenty years, and its emergence now seems as surprising to me (and almost as impressive) as the sudden appearance of a Tyrannosaur on 21st century streets. But perhaps that just shows how wrong I've been.  

The problem confronted by CYC is the same, hydra-headed monster which in different ways has blocked the advance of most forms of AI - the problem of dealing with the chaotic complexity of real-life, real-world problems. Computers do well at solving simple problems and carrying out tasks in which every contingency can be nailed down in advance, but that is rarely possible in the wider world.

Linespace

A classic example often used in explaining flow-charts is the making of a cup of tea: we're asked to break down the job into stages (get the pot, fill the kettle), consider the correct order in which tasks should be carried out, and build in some elementary branching (is the kettle full?). This is fine so long as we're dealing with imaginary teapots in a grossly simplified world: but actual real world tea-making is a jungle of unforeseen problems and bear-traps which no computer could cope with. Just recognising the teapot is a formidable task: teapots don't have to be made of any particular material or be any particular size: they can be any shape and even their topological properties are not guaranteed - they don'thave to have a tubular spout or a handle. Is that a teapot we're heading for, or might it be a kettle, a coffee-pot, a jug, a picture or a mirror? Can we put down the paper we're holding, or will we lose track of it forever? When we move the kettle, should we note its new position, and must we also note that the cups, the walls, the table and the floor have not moved? Have the cups, as a matter of fact, fallen on the floor because we failed to anticipate that moving the kettle laterally would shove them off the table?

Linespace

Humans move through this infinite sea of troubles fairly effortlessly. Computers always seem to run into one or other of a set of related problems. Either the vast number of possible permutations in reality gives rise to a combinatorial explosion which exhausts the speed and capacity of even the most powerful computer, or the frame problem rears its ugly head, either in the classic form of being unable to keep track of what hasn't changed, or in the looser philosophical form, where we are unable to pick out the important contingencies from the infinite range of possible things we might think about.

In some expositions, these problems, especially the last one, seem so daunting that it begins to seem obvious that intelligence is actually impossible. But humans pull it off somehow. It seems that human beings have access to a vast body of knowledge about the world - a background, a common sense - which just tells them what is likely, or relevant, or reasonable, in any given context. One view is that this relies on a special faculty, a kind of intuition which we need to understand better before we can replicate it. There are plenty of theories about where this special faculty might come from, but they tend to be unfriendly to old-fashioned conceptions of AI. Another, pragmatic, approach has been to set up frames or scripts , which supply the computer with assumptions about particular limited contexts. This approach yielded some good results in the early days, but it suffers limitations almost by definition; we're still essentially talking about toy worlds. 

It might be that until you achieve a certain minimum quantity of factual 'common-sense' knowledge, a level well above anything yet realised by most projects, the thing won't really fly at all: if so, then CYC might prove to have been the only viable strategy. It certainly represents the most serious attempt to tackle the whole problem head on - to assemble a significant fraction of the entire corpus of human everyday common-sense knowledge, and use it to give a computer the same kind of background understanding that human beings automatically posess. Even CYC, of course, is not supposed to know everything immediately. A key part of the approach is an ability to generate new truths from its stock of information, using a system which is ultimately based on predicate calculus

Linespace

There's no denying, I think, that Doug Lenat, the founder and guiding spirit of the project, has pulled off a remarkable achievement in keeping such a large project on the road. When the original funding dried up, he set up Cycorp to carry the project forward: as progress was made various cut-down versions of the software, notably OpenCyc, have been made available to others to study and work on. You would expect that as the data acumulated, the software would gradually find it harder and harder to cope, but Lenat says that it actually gets easier for CYC to draw its own conclusions as time goes on, rather than being spoon-fed. This, it seems, is the rationale behind the creation of a direct internet-based interface with CYC; it will soon be so well-informed that it will be able to talk directly to people and gather its own information. I don't think this interaction is likely to take the form of chat-bot style conversations, but even a yes/no question facility would be interesting.  

Linespace

Can it be true, though? Are we really on the verge of breaking through to human-style comprehension, with all that that implies, and has the trick been pulled off by good old-fashioned AI and massive computing power after all? I'm afraid I think it is highly unlikely. Neither an encyclopedia nor predicate calculus seem to me to work in ways which are even remotely like the human brain.

Now that may seem rather sweeping. The brain does store a huge number of facts, and a well-informed person can indeed supply information, just as an encyclopedia does. Equally, the human brain does do predicate calculus - indeed, if it didn't, there wouldn't be any predicate calculus. But in both cases, these are very specialised forms of thought. An encyclopedia records facts in a fixed, explicit form: if you want to run off the same piece of text, well and good: if you want it even slightly reinterpreted, a formidable processing task awaits you. If I ask you, 'Do dogs have tails?', you answer without difficulty, and if it contains a suitable statement, an encyclopedia could answer just as readily. If I ask the question in a different form - 'Are tails usually part of a dog's anatomy?' - you answer just as easily. You don't have to ask yourself whether dogs have tails and then reason from there to the required conclusion. The encyclopedia, however, probably lacks the exact formulation required by the second question and a different process has to be invoked. If, moreover, I ask whether dogs wear beards, the question is easy for you (barring fancy breeds, cartoon dogs, and so on), but it will be a non-trivial task, perhaps an impossible one, to deduce the answer from the information in an encyclopedia 

Linespace

Deduction, after all, is a relatively weak and limited way of generating new conclusions. There is a strong impression about all of formal logic that you only get out what you put in, whereas normal human thought encompasses a range of creative, intuitive and insightful processes which far outrun pure logic. Human beings do predicate calculus the way they do tight-rope walking; by artificially restricting and then bringing to a new level of perfection, a special form of one of their more general abilities. To try to build a model of human cognition on the basis of predicate calculus is like trying to replicate human locomotion by stringing ropes everywhere.

Of course, CYC has been around long enough to address some of these problems, and gain a sophistication which earlier conceptions did not have. For one thing it seems that instead of tackling the whole world in one bite, CYC has a series of theories about particular fields or aspects of reality. It includes, apparently, a full statement of the Linnaean system to help it recognise and classify animals, for example. The fact that different theories are applied in different realms allows CYC to tolerate contradictory or inconsistent beliefs to some degree: lack of this ability is one of the leading weaknesses of AI systems based on classical logic, a problem which others have tried to address with non-monotonic logics or other systems.

Linespace

But this too raises an objection.  Common sense is supposed to be the opposite of theory, not an example of it. This mistake was exemplified by the "Naive Physics Manifesto" issued by Patrick Hayes back in the 70s: he saw the normal human understanding of physics as based, not on Einsteinian or even Newtonian theory, but a kind of secret, subtly mistaken theory which held that things have an inbuilt tendency to stop moving, and so on. The attempt to reconstruct this naive theory so that it could be built into artificial intelligences proved an interesting failure, and I believe the real reason was simply that normal understanding is not theoretical in nature at all.

Academics have a natural tendency to value theory rather highly, and they are prone to mistake the tools which consciousness uses to investigate reality (encyclopedias, formal logic, computers) with the mechanisms which sustain consciousness (and even reality) itself. Perhaps CYC is an example of this rather clerkish tendency. If so, the bad news comes with a possible silver lining. CYC might never be truly intelligent in the way human beings are, but it might still turn out to be a formidable tool.

I must admit that the idea of the old monster breaking out and rampaging triumphantly through the streets is one I am reluctant to abandon altogether.

Linespace

Earlier

home

later