Page 78 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 64 21.9.2005 11:40pm

64 Biomimetics: Biologically Inspired Technologies

conclusion symbols which strongly support the truth of the set of assumed facts being considered
(this is the duck test, as described in Hecht-Nielsen, 2005).
But an even more important aspect of cognition is the underlying assumption that the knowledge
we possess is exhaustive (see Hecht-Nielsen, 2005). In other words, once the available knowledge
has been used, we can be reasonably sure that no other possible conclusions exist. This effectively
causes the known knowledge to act as an implied constraint. In particular, those possible conclu-
sions identified as known to be supportive of the assumed facts are not just viable alternatives; they
are probably the only viable alternatives. Thus, in non-Aristotelian information environments, the
exhaustive knowledge assumption leads to answers which are ‘‘almost logical deductions.’’ Thus,
beyond just being an implementation of the duck test, confabulation might be termed a ‘‘strong’’
(although not logically rigorous) form of inductive reasoning.
Another factor that makes confabulation so powerful is its ability to support the construction and
use of lexicon hierarchies. In the simplest case, the symbols of a higher-level lexicon each represent
an ordered set of symbols that meaningfully co-occur on lower-level lexicons. But much more is
possible. For example, in vision (see Section 3.5), higher-level symbols each represent several
groups of lower-level symbols. These are symbol groups that are seen in successive ‘‘eyeball
snapshots’’ of the same object at the same fixation point (but at slightly varied object poses). In this
way, these higher-level symbols respond to the appearance of a localized portion of an object at a
number of different poses. They are pose-insensitive localized visual appearance descriptors.
In language hierarchies, knowledge can be used to discern symbols which are highly similar in
meaning and usage (in a particular given context) to a particular symbol. These are termed
semantically replaceable elements (SREs). Knowledge possessed about an SRE of a symbol can
sometimes be used to augment knowledge possessed about that base symbol. This can significantly
extend the ‘‘conceptual reach’’ of a system; without requiring training material covering all possible
combinations of all symbols. For example, what if a friend tells you about the food ‘‘guyap’’ that
they had for breakfast. They poured the flakes of guyap from its cardboard box into a bowl; added
milk and sweetener, and then ate it with a spoon. It was good. By now, you are fairly sure that
‘‘guyap’’ is a breakfast cereal of some kind, and at least in the ‘‘breakfast food’’ context, you can
apply your knowledge about breakfast cereal to ‘‘guyap.’’
Hierarchies can work backwards too. For example, if you say you are looking for a ruler on your
desk; then links from the word ruler (and perhaps some of its SREs) go to the visual system and
provide input to high-level visual attribute (‘‘holistic’’) representation symbols which, in the past,
have meaningfully co-occurred with visual sightings of rulers. This is accomplished via a CKF
effect; which leaves an expectation of all such symbols that have previously been significantly
linked to the word ruler. During perception, which takes place immediately after this expectation
symbol set has been generated, knowledge links from primary, and then secondary visual lexicons
arrive at this high-level visual lexicon and a W, CK, or WN is issued at the same time. Only
elements of the expectation can be activated by the visual input; and the net result is a set of
symbols that are consistent with both the word ruler and with the current visual input. As discussed
further in Section 3.5, the final step is a rapid bidirectional knowledge link interaction of the higher-
level expected symbols with those of the lower-level lexicons to shut off any symbols that are not
participating in ‘‘feeding’’ (i.e., are not consistent with) symbols of the high-level expectation. This,
in effect, causes low-level symbols representing portions of the visual input that are not part of the
ruler to be shut off. It is by this visual object segmentation mechanism that sensory objects are
almost instantly isolated so that they can be analyzed without interference from surrounding objects
(segmentation is also a key part of sound and somatosensory processing). Without an expectation,
sensory processing cannot proceed (this point, which seems to be widely unappreciated in the
biologically oriented neuroscience disciplines, is the subject of a wonderful book describing clever
experiments that well illustrate this point [Mack and Rock, 1998]).
Above all else, cognition works because of the huge hierarchical repertoire of learned and stored
action sequences (programs of thought and/or movement — see Section 3.6). Appropriate actions