Page 77 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 63 21.9.2005 11:40pm

Mechanization of Cognition 63

probably enriching. Perhaps the admonishment should be to take care not to soil or damage what
they handle. If you are punctilious in this regard, have your children wear disposable latex gloves
and force them to pay for any damage or breakage out of their allowance. But give them these
valuable experiences.
Evenmorethanbasicsensoryprocessing,learningtocarry outimportant behavioraltasksrequires
deliberate provision of examples and supervised rehearsal practice. This often includes feedback on
performance; something that will be ignored here since using such feedback requires noncognitive
functions, which as yet, we do not understand sufficiently to build. Because of this current lack of a
reinforcement learning adjunct to cognition (work is proceeding in this area — see Miyamoto et al.,
2004, for example), for the moment, education of confabulation-based cognitive systems will
probably be confined to strictly positive examples. In other words, examples, where learning should
definitely take place.
For example, consider a confabulation-based vision system viewing cars passing by on a busy
road. The visual portion of the system segments each car it fixates on (see Section 3.5) and then
rerepresents its visual form, color, and internal motion using high-level symbols that have invari-
ance properties (e.g., pose insensitivity). Thus, the final product of processing one such look is
activation of a set of high-level symbols, each describing one visual attribute of the object.
Imagine that a human educator sitting at a computer screen where each look (eyeball snapshot
image — see Section 3.5) to be processed by the confabulation-based vision system is being
displayed (each subsequent look is processed only after the previous look’s use for education has
been completed). The human examines the visual object upon which the center (fixation point) of
the eyeball image rests and describes it in terms of English phrases (spoken into a noise-canceling
microphone connected to an accurate speech transcriber — see Section 3.4). For example, if the
object is a green Toyota Tundra truck with a double cab; the educator might speak: ‘‘Toyota Tundra
truck,’’ ‘‘dark green,’’ ‘‘two rows of seats; in other words, a full-sized back seat,’’ ‘‘driving in the left
lane of traffic.’’ After accurate transcription, this text is represented by a set of active symbols in the
language module (see Section 3.3). Knowledge links are then established between the active visual
symbols representing the visual content of the look and the active language symbols representing
the education-supplied language content of the look.
Note that the language description is not exhaustive; it is just a sample of descriptive terms for
the visual object. For example, if a similar look of the same truck were presented on another
occasion, the educator might add: ‘‘oh, and there are four dogs in the bed of the truck.’’ This would
add further links.

3.2.3 Discussion

One of the most exasperating things about this theory is that it seems impossible that just forming
links between symbols and then using these links to approximately maximize cogency could ever
yield anything resembling human cognition. The theory appears to be nothing but a giant mountain
of wishful thinking!!! That such a simple construction can do all of cognition is indeed astounding.
Yet, that is precisely my claim. Some reasons why confabulation may well be able to completely
explain cognition are now discussed.
First is the fact that the number of links that get established (i.e., the number of individual
items of knowledge that are employed) is enormous. Even in the narrow domain of single proper
English text sentences (see Section 3.3), over a billion individual knowledge items are often
employed (contrast this with the world’s largest rule bases; which have about 2 million items
of knowledge). Slightly more elaborate proper English text confabulation systems (able to deal
with two successive sentences — see Section 3.3) often possess multiple billions of items of
knowledge.
The value of having such huge quantities of such a simple form of knowledge is best seen in
terms of how this knowledge is used in confabulation. The first use of knowledge is to excite those