JON UDELL
FLUID CONCEPTS AND CREATIVE ANALOGIES: COMPUTER MODELS OF THE FUNDAMENTAL MECHANISMS OF THOUGHT by Douglas Hofstadter and the Fluid Analogies Research Group, Basic Books, ISBN 0-465-05154-5, $30
Fifteen years ago, Gödel, Escher, Bach: An Eternal Golden Braid exploded on the literary scene, earning its author a Pulitzer prize and a monthly column in Scientific American. Douglas Hofstadter's exuberant synthesis of math, music, and art, and his inspired thought experiments with ``tangled hierarchy,'' recursion, pattern recognition, figure/ground reversal, and self-reference, delighted armchair philosophers and AI theorists. But in the end, many people believed that these intellectual games yielded no useful model of cognition on which to base future AI research.
Now Flu
id Concepts and Creative Analogies presents that model, along with the computer programs Hofstadter and his associates have designed to test it. These programs work in stripped-down yet surprisingly rich microdomains. Here's one example from the Copycat domain: ``Suppose the letter-string abc were changed to abd; how would you change the letter-string xyz in the same way?''
A shallow analogy emerges if you answer xyd. But that's unsatisfying, because it doesn't acknowledge features such as sequence, successorship, and the special roles of a and z as first and last. A more subtle interpretation yields the more satisfying answer wyz, in which the role of d as the successor to c in an ascending sequence anchored at the beginning of the alphabet mirrors that of w as the predecessor of x in a descending sequence anchored at the end of the alphabet.
The challenge for Copycat, the program built to solve this class of problem, was not only to be able to arrive at the answer wyz but to get there in the s
ame way humans do. Hofstadter reports that just as human subjects usually answer xyd but are more satisfied when they sometimes discover wyz, so it is with Copycat.
Concepts? Discovery? Satisfaction? These are, of course, dangerously loaded terms. Here's how Copycat actually works: It uses one network, called the Slipnet, to model concepts, both literal (e.g., the letter a) and abstract (e.g., same, opposite, and successor). Links encode distances between pairs of concepts. The distances vary dynamically under the influence of another network, called the Workspace. This is where software agents assemble and tear down structures on various levels--bonds between adjacent letters, groups of letters, and correspondences between groups. Many agents are always running. They're chosen randomly from a larger population of agents scheduled to run. That randomness, governing interplay between conceptual (Slipnet) and perceptual (Workspace) activities, is what enables Copycat to sometimes ``discover'' wyz. The qu
ality of structure and depth of concepts assembled in the Workspace are what measure Copycat's ``satisfaction'' with that answer.
Hofstadter boldly claims that Copycat captures fundamental processes of creative intelligence. That's radical enough, but what will make this book even more controversial is that he considers and violently rejects the models put forward by other cognitive scientists.
Hofstadter, like the competitors he denounces, must ultimately appeal to the performance of a computer program in some artificial-problem domain (``look, it finds wyz'') as evidence of success. Thus, the Hofstadter-uber-alles attitude can be justified only by compelling explanations of why a domain is meaningful and how a program's performance in a domain models real aspects of intelligence. I find the explanations compelling. Maybe you won't. But either way, the cards are mostly on the table.
Jon Udell is a BYTE senior technical editor at large. You can reach him on the Internet or BIX
at
judell@bix.com
.
A Model for Future AI Research
