Page 84 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 70 21.9.2005 11:40pm

70 Biomimetics: Biologically Inspired Technologies

two then do two things: on phrase region one, only the symbol for The remains (since The canoe,
nor any further extension of it, is not in the phrase lexicon — for brevity, the manner in which the
phrase lexicon itself, and the additional knowledge bases of the Figure 3.2 architecture, are derived
using word-level knowledge — this process too is totally confabulation-implemented and does not
use any linguistic knowledge — is not described here). This parsing process continues down the
phrase lexicons (each possessing 126,008 symbols), quickly yielding the parse (with the phrase
symbol numbers in parentheses): The(8) canoe(25085) trip(1509) {was going}(63957)
smoothly(9723) when(64) all(56) {of a sudden}(69902). Thus, phrase lexicons 1, 2, 3, 4, 6, 7,
8, and 9 have symbols active on them (each phrase is represented on the lexicon immediately above
its first word). All the other phrase regions have no symbols active on them. Note that if the last
word (sudden) of the assumed fact phrase were not present, that phrase lexicon 9 would not have a
single phrase active on it; but would have several (representing all of the phrases that begin with of
a: e.g., of a, of a kind, of a sudden, of a sort, etc.). I thank my colleague Robert W. Means for
implementing and providing the details of this example.
The above processing sounds like it would take a long time. But remember that thinking is just
like moving. When you throw a baseball many tens of muscles are being commanded in parallel in a
precisely timed and coordinated way. The above thought process (‘‘parse sentence’’) is stored,
recalled, and executed just like a motor action such as throwing a baseball. The initiation of each
involved knowledge base activation and confabulation happens in close succession in a ‘‘ripple’’ of
processing that rapidly moves from the left end of the architecture to the right; terminating at the
end of the assumed fact phrase. The entire parsing process is completed in just a small multiple
of one knowledge base transmission time. Like some movement actions (e.g., dribbling a basket-
ball); thought actions are often divided up into small ‘‘macro’’ segments which, depending upon
their outcome (i.e., which symbol wins the confabulation competition), trigger alternative next-
segments.
As discussed in the Appendix, a key concept of hierarchical architecture design is the prece-
dence principle. There, it was discussed in the context of the constitution of individual symbols
within a single lexicon. However, the same principle holds between lower and higher abstraction
level lexicons within a hierarchy (such as that of Figure 3.2). In this expanded form, what the
precedence principle says is that as soon as content that is represented at a lower level of a hierarchy
is re-represented at a higher level, the involved active lower-level lexicon symbols must be shut off.
This is implemented in human cerebral cortex by use of the conclusion–action principle (see
Section 3.A.6).
In the case of the precedence principle, the action which is triggered by the expression of a
phrase representation symbol is to shut off the lexicons which supplied words to the phrase that the
symbol represents. For example, if the phrase that emerges from the parse has three words, the word
lexicon directly underneath the phrase lexicon, as well as the next two lexicons to its right, are shut
off (they stop expressing their word symbols). If the phrase has only one word, a different action is
triggered: namely, only shutting off the lexicon directly beneath. And so on. Note that these action
commands are not issued until the choices have narrowed to a single symbol; since it is only then
that the conclusion–action principle operates. This is a concrete example of how thought is not
software. It is a series of sets of action commands; each set being immediately originated (issued to
action nuclei) when a firm confabulation conclusion is reached (i.e., each conclusion has its own set
of action commands that are permanently associated with it, and which are originated every time
that conclusion is expressed as the lone final result of a confabulation operation by its lexicon).
This example illustrates a thought process that can be launched immediately with no further
evaluation (e.g., by basal ganglia). It also illustrates how we will need to implement the action
command output portion of cognition from the very outset of research. A great deal more could
be said about action command generation and action symbol sequence learning and recall using
confabulation architectures. But this topic would take us beyond the introductory sketch being
attempted in this chapter. Suffice it to say that quite a lot is known about how action sequences can