Building a mind: a long unfinished tale with mathematical interludes
by
John Cleary
Dept of Computer Science, University of Waikato

Over the last 50 years it has slowly dawned on cognitive and computer scientists that many of the capabilities of the mind which we think of as hard are easy and things that we thought were easy (if we thought of them at all) are actually very hard. Hard in the sense that it has turned out to be difficult to construct computer systems to do them. Thus routine operations in the differential calculus are easy (despite the anguish of generations of undergraduate students). Picking up a pen from your desk has so far proved impossible (for a computer).

Picking up a pen involves many aspects of the human mind and body. Building a body which is as flexible as the human hand is as yet unsolved (and mass producing them with unskilled labor even harder). Also unbuilt are minds that can recognize a pen or direct a body to pick it up. This talk will concentrate on one small part of the race that we are in to catch up with evolution in the construction of bodies and minds. This is the area of adaptation or learning.

Human and animal minds have a ubiquitous ability to adapt to the circumstances they find themselves in. This may be contrasted with most modern software which has a rigid stimulus response behavior akin to that of, say, a flat worm and well short of a cockroach. Every hand and arm is different and the mind must adjust the reaching and grasping behavior to suit arms with a great variations in length and strength. Also our brains do not come loaded with software for recognizing pens or need to download a ``pen recognition'' plug-in.

The current view of adaptation is that it consists of a process of selecting between competing ``hypotheses''. Differing views about how to do this go back at least to Epicurus and William of Occam. Epicurus advocated treating all possible correct hypotheses as being equal while Occam's razor advocates selecting the simplest one. The mathematical reasoning required to resolve this conflict has been provided by the Reverend Bayes, Kant, Turing, Kolmogorov and others.

In the last ten years or so computers have become sufficiently powerful to successfully apply these theories. For example, the best text compression schemes all adapt their compression to the text that they see. After seeing as much text as an educated adult might see in their lifetime the compression algorithms are able to recognize text as well as the human mind. Because building eyes, hands and bodies is harder than building minds recent applications focus on areas where the data is easy for computers to access. This consists in the main of text held on the world wide web or in corporate databases. Tasks that are being attempted include: categorizing newspaper articles, for example, into sports or business articles; assessing whether articles about stocks are positive or negative; extracting names and addresses from within text; finding whether people who buy dog food also buy tennis balls; anticipating what you will type into your computer next; watching the decisions that experts make and building systems to mimic them ....

Each of these applications consists of an algorithm to select among a large number of competing hypotheses coupled to a large memory. Just as in the minds and brains constructed by evolution it pays to carefully tailor each application to the particular problem. Current capabilities probably exceed those of an ant and may soon approach the level of lizards.

Date received: November 7, 2000

Copyright © 2000 by the author(s). The author(s) of this document and the organizers of the conference have granted their consent to include this abstract in Atlas Conferences Inc. Document # caek-78.