Page 79 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 65 21.9.2005 11:40pm

Mechanization of Cognition 65

or action sequences are triggered instantly each time a lexicon confabulation operation yields a
single active symbol (i.e., a decisive conclusion). As discussed in Appendix, this is the conclusion–
action principle of the theory. The action(s) automatically triggered by the winning symbol can be
of many characters. They can be immediate postural goal outputs that are sent down the spinal cord
to motor nuclei and the cerebellum, they can be immediate lexicon operation commands, they can
be immediate knowledge base operation commands, or they can be candidate actions (cortically
proposed thoughts and movements) which must first be sent to the basal ganglia for evaluation and
approval before they are executed.
The main advantage of confabulation-based cognition over traditional programmed computing
(formal computer programs, rule-based systems, etc.) is a much greater capacity for handling novel
arrangements of individually familiar objects. Programmed computing must essentially have a
predefined plan for dealing with every situation that is to be handled. For example, a plan for
breaking up a complicated ensemble of problems into isolated, disconnected sub-problems, so that
each can be handled in a predefined way. Unfortunately, in most real-world situations, this
approach fails badly because complicated real-world situations inevitably have unanticipatable
interrelations between their elements that disallow pre-defined decompositions. By virtue of their
huge stores of general-purpose and low-level knowledge, confabulation-based systems are inher-
ently able to take novel external context into account as each individual conclusion (or ensemble of
conclusions — if mutual solution constraints are to be honored — see the discussion of consensus
building in Section 3.3) is addressed. Confabulation-based systems can also adapt existing action
plans (e.g., by replacing specific elements of a stored plan with similar substitutions which are
relevant to the current situation) to fit novel circumstances. They do not typically run out of things
to try and, instead, tend to press on and do the best they can, given what they know. If a particular
approach yields no conclusion, other approaches are typically immediately launched. Yet, because
actions are triggered each time a conclusion is reached, almost all behavioral sequences are
dramatically novel. Also, each new experience can (with occasional help from a human educator)
be added to the knowledge base to further enlarge the system’s future repertoire.
More could be said regarding the benefits of the confabulation approach. However, the remain-
ing sections of this chapter present more concrete examples of this. The nature of cognition is very
different from that of computing. So much depends upon designing clever architectures of lexicons
and knowledge bases and upon using clever, highly threaded, but very simple, thought processes to
control these architectures. Since the information processing control which must be exerted at each
stage of an action process is triggered by the current cognitive world state (the collection of all
decisive confabulation conclusions that are active, or accessible from working memory, at that
moment), cognition has no need for ‘‘computer programs’’ or ‘‘software.’’ In effect, the conclusion
of each ‘‘cognitive microprogram’’ (lowest level action sequence) is a GOTO statement. There is no
overall program flow defined. Just action sequences completing and then triggering subsequent
action sequences (in a pattern that almost never exactly repeats). Things happen as they happen,
with no master program controller involved or needed (although a number of subcortical brain
structures can execute ‘‘interrupts’’ when certain conditions occur). This brain operating system (or
lack thereof, depending on your point of view) seems like an invitation to disaster. However,
beyond possible conflicting commands to the same action (movement or thought) resource (which
are impossible by design! — see Appendix Section 3.A.2), very little can go wrong.

3.3 LANGUAGE COGNITION

This section discusses the use of confabulation for representing and generating language. This
application arena is the most developed, and yet is transparently crude and primitive. An enormous
amount of work needs to be done in language. The hope of this section is to illustrate how promising
this research direction is.