Fluid Concepts And Creative Analogies: Computer Models Of The Fundamental Mechanisms Of Thought [Paperback]

(1) Low-level perception. The best example of this type of work comes obviously from computer vision systems. These systems, given a set of input images, usually extract some important information from this input, generating, well, other images (i.e. depth image, edge contours etc.). But this extracted information is usually on a still very low, meaningless, level, to be used by, for instance, a theorem-proving system. To make it clear to all readers what is meant by "meaning", consider the information-processing that must occur whenever an animal, given its massive sensorial information, perceives danger. Going from a set of images and sounds to a feeling of danger involves extracting meaning from the original input, and this is not what is done by current low-level perception projects. It is almost as if these perceptual processes "delegate" the extraction of meaning to another upcoming process. To get into the meaning of a situation, low-level perceptual processes are not enough; there is a clear need for further perceptual processing.

(2) GOFAI symbolic manipulation. This is the other side of the AI coin, dubbed by philosopher John Haugeland as GOFAI, for "good-old-fashioned artificial intelligence", where programs usually handle (syntactically) a representation that supposedly should have been formed by a perceptual process. These systems, such as theorem-proving systems, chess playing, and others, do perform some impressive feats, but they do not have a clue about the semantics of their symbol manipulation. As an example, consider the following predicate-calculus statement: (philosopher (Socrates)). We all fully understand what that means, but what about the machine that executes it? Does it have any meaning to the machine? It is obvious that the answer is no, for that is just a syntactic symbol, as meaningful to the computer as (XzE (GgGggGG)), which doesn't mean anything. But how can a system that only manipulates meaningless syntactic symbols posses any meaning on those symbols? This seems to be an intrinsic problem to GOFAI projects.

Both of these avenues of AI research seem to be based on an unspoken hypothesis of a "center of meaning" arising in the brain (maybe the mind's eye?). The low-level perceptual processes should operate on information that has yet to reach such place, and GOFAI systems in turn handle information that seems to have long reached it. The problem is, what happens at the point of crossing the line? Nobody really knows.

Maybe, then, there is no such line after all - as Hofstadter clearly considers as true, by presenting us with an original alternative. His main thesis is based on the idea that meaning comes from an emergent process that combines perception with analogy-making. He argues, following philosopher Immanuel Kant, that perceptual processes are inseparable from high-level cognitive processes, and, moreover, that (1) perception is guided by analogy-making, and (2) this analogy-making process is itself derived from perception. This thesis has profounds implications for AI.

In his systems, perceptual observations activate concepts, and these activated concepts in turn guide (probabilistically) further perceptual observations. Hofstadter and his group ressurect the HEARSAY II architecture and extend it to other pattern-analysis domains. There is a mixture of bottom-up and top-down processing that eventually leads to the understanding of a situation arising as a combination of "platonic" concepts. This iterative (perception/analogy mapping) process gradually develops a coherent view of the context of the problem it is working on, and that view constitutes, in a sense, on the extracted meaning of the problem. We can say that "understanding p" is, in a sense, "to know what p is like", and this "what p is like" information comes from such analogy-mapping.

Not surprisingly, his projects cannot be found on the symbolic versus connectionist menu. Hofstadter points out that GOFAI (symbolic) systems are too optimal, too rational to be psychologically realistic (he calls them "the Boolean dream"), and that, on the other hand, connectionist systems operate on a level "too low" to be relevant, at present, to a greater understanding of the cognitive issues. Obviously, all mental phenomena may be reducible to a connectionist-system level, but, then again, these same phenomena will be reducible to a quantum physics level. What we should strive for at the moment, he argues, is the right level on which to conduct research. And that level may just be the level of the HEARSAY II speech-understanding system.

Probably the most ambitious AI project under development today is the Letter Spirit project, described on the last chapter. Striving to develop a system that deserves credit for its own creations, with a sense for esthetics, with true creativity and true style - almost taboo issues in AI -, this project messes with many important topics that lack serious study. And, just in case a skeptical reader is wondering, "but, doesn't the project X mess with these exact issues?", then, well, I would recommend Hofstadter's own criticism of "related" projects, given on the epilogue "On Computers, Creativity, Credit, Brain Mechanisms, and the Turing Test".

In summary, this is not your average AI book. This is a full redefinition of artificial intelligence, on a class of its own, an excellent book that deals with deep issues largely ignored by the AI community. Like all the great AI books, this one shuffles between philosophy, methodology, and architecture. Some, maybe even most, highly established AI researchers will not comprehend it completely -- they'll never realize its full scope. However, it is highly recommended to Graduate Students on AI (though not as an introduction to the field). It also seems to be making its mark among philosophers, and I think that neural network researchers will appreciate it as well, for, by extending the HEARSAY II architecture to other domains, it presents an alternative (emergent) architecture that brings us much closer to understanding what understanding is all about.

Many researchers in the field of Artificial Intelligence take the approach of attempting to mimick the
behavior of people with computer programs. On the surface, this might seem a logical direction to take, and so
AI researchers have a tendency to go and dream up batteries of tests that aim to characterize some area
of human behavior, then the sum up all the results and come up with the range of responses that fits cozily
into their bell-shaped curves. Armed with what they've assured themselves is normal human response to
all their scenerios, the go off and attempt to write computer programs that react the same way as John or
Jane Doe did. Once they've gotten a program that generally responds like 'most of the human subjects' did,
they usually beef it up by programming in more and more details about the domain of the scenerio at hand.
A good example of this line of thought is
Deep Blue, IBM's massively parallel chess playing supercomputer.

What Douglas Hofstader's latest book points out is that this sort of thinking about artificial intelligence is
the brute force approach. What you end up with is a computer that knows a *lot* about a particular
domain (i.e. chess), but has no other redeeming features whatsoever. Deep Blue could probably whip 99.9%
of the human population at chess, but it can't even begin recognize the elegance of a particular
strategy (such as the sicilian defense) because it has no ability to make analogies to other domains.

The ongoing thread of Hofstadter's work has always been quite clear. He's interested in understanding
human thought, not mimicking it. In his latest work, Analogies, he and his FARGonaouts (students at his
Fluid Analogies Research Group - FARG) introduce us to several of their long term projects that uncover
some of the 'fundamental mechanisms of thought'.

His usual modus operandi is to examine the problem space of extremely simple microdomains - problem sets having
very few parameters, but that scale up well into higher domains with the analogies it evokes.

For instance, he describes
a very simple game called "TableTop" in which two players face each other across a table in a cafe. On both
sides of the table are arranged various objects of the TableTop domain - knives, spoons, cups, plates, salt
and pepper shakers, etc. The game begins when one player touches an object on their side of the table, saying "Do This", and the
other player then must touch a corresponding object on their side of the table which best mirrors the other person's
choice.

The goal in each exchange is to choose the most appropriate corresponding object. Simple, right?
Say I touch the coffee cup sitting in the middle of my placemat. You don't have a coffecup on your side. But you
do have a soup bowl there. You touch it. You've made an analogy. The soup bowl's physical arrangement on the
table was similar to the situation of the coffee cup, and the 'round container-ness' also made it a good match, even
though it was a totally different object. This simple microdomain affords us a lot of insight into the
process of analogy making. That is, the lessons learned in the TableTop domain can be used in other domains
with different details, but similar problem space.

For instance, the Battle-Op Domain, where, two geographical entities are pitted against each other:

(Excerpt from "Analogies")

A war breaks out between California and Indiana over the former's attempt to divert rain clouds from
soggy Indiana to the parched San Joaquin Valley. Unfortunatley, the conflict goes nuclear, and California
obliterates Bloomington. The war council in Indianapolis, wishing to be appropriately punitive but not
risk further escalation, must then decide what Californian entity to annihilate in retaliation. Thus -
what is the Bloomington of California?

Given the act of agression committed by California, it would be nonsense to blast Los Angeles, a city with
a population over 100 times that of Bloomington. Attacking San Diego would be precluded because of its
world-famous zoo. And detonating an H-Bomb in the Pacific so as to cause a tidal wave to destroy Carmel would
be ruled out because an attack mounted on that jewel of a city would likely enrage Californians to a too-risky
degree. After some consideration, then, the war council might reason that the Hoosier Armed Forces would best
achieve 'the same result' not by destroying a city, but by offering all the migrant workers of California one dollar
an hour more to come and work in Indiana.

(Excerpt from "Analogies" ends.)

As you can see, lessons learned about analogy making in one domain can be easily mapped on to other domains of
problems. This is one of the uniquely human attributes of thought - that we can see analogies to things
we've experienced, and use those analogies to help us tackle new problems faced in other domains of life.
When a computer program can be endowed with this ability, then we'll be on the road to artificial intelligence.

The research outlined in Analogies is very intriguing and bucks the status quo in the field of AI at every
turn by focusing tightly upon the goal of understanding rather than mimicking human thought. If you ever find
yourself thinking about thinking - how we think and why we think, then I highly recommend you
pick up a copy of Fluid Concepts and Creative Analogies: Computer Models of the Fundemental Mechanisms of
Thought and curl up by the fireplace with it soon!