Page 137 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 123 21.9.2005 11:41pm

Mechanization of Cognition 123

other types of knowledge (e.g., sensory, event, or factual knowledge), only the action symbol
replay knowledge is rehearsed and solidified at night. The action symbol to action command
associations can only be learned and refined via awake rehearsal. This accounts for the fact that
anyone learning a new skill will frequently find themselves (either through vague memories of
dreams upon waking, or via reports from their sleep partner) carrying out ‘‘silent practice’’ of those
skills in their sleep. These do not involve launching the involved actions (a function that is normally
suppressed during sleep), but simply running through the involved action symbol sequences. Such
activities can help solidify the symbol sequences and this often yields improved skill performance
the next day.
Quite a bit of experience has been gained with learning and recalling action symbol sequences in
one of my UCSD graduate courses. For example, a checker-playing system that learns by expert-
guided rehearsal has been demonstrated. However, issues surrounding the replay of action sequence
hierarchies are complicated and not within the scope of an introductory chapter (e.g., provisions for
automatic real-time, moment-by-moment modification of an ongoing lower-level action sequence
replay in response to the exact current state of the world; with no modification at the higher level —
a process called instantiation — must be introduced). So action symbol sequence learning and
recall are not discussed in this chapter.
In summary, the theory proposes that the unidirectional symbol pair links used in confabulation
are the only knowledge learned and stored in cortex that is used in cognition. However, as described
in this section, there is a second kind of knowledge learned and stored in cortex: the associations
between each symbol and the action commands that its expression should launch. This knowledge
is not really part of cognition. It is the mapping from decisive cognitive conclusions (single active
symbols resulting from confabulations) to behaviors. Thus, the ultimate end product of cognition is
the origination of action commands; some of which are unconditionally executed immediately and
others, termed suggested actions, must be approved (vetted) by the basal ganglia before they can be
executed.

3.A.7 Discussion

The theory’s hypothesized cortical implementation of knowledge links has some important
universal properties. First, the locally random wiring of the cortical axons can be established during
development and then frozen, essentially for life (although there may be a very slow replenishment
of some types of neurons throughout life that helps keep the brain functional as neurons slowly die;
but this has not been established — the vast majority of neurons probably live a very long time,
perhaps for the full life span of the individual). Knowledge links, by means of a parallel, two-step
synfire chain communication process through the random cortical signaling network, can be
immediately formed, as appropriate, between almost any two symbols in any two lexicons that
genetics have provided connection possibilities for. A link can be temporarily established instantly
(via the short-term memory mechanism) and then, if it is warranted, the link can be progressively
transformed into permanent knowledge during the subsequent few sleep periods.
The price of this ability to instantly learn almost anything without need for rewiring (to carry out
such wiring by growing new axons would take days and would require the involved axons to have
unbelievable navigation abilities) is probably a vast over-wiring of cortex. A prediction of the
theory is that only roughly 1% of cortical synapses are actually used to store knowledge (i.e., have
been strengthened). The rest are there to provide the capacity for instant arbitrary learning. Thus,
the old saw that ‘‘we only use 10% of our brain’’ is probably wrong on the high side. Ninety nine
percent of unstrengthened synapses are hypothesized to simply be sitting around waiting to be
needed. This may seem wasteful; but unstrengthened cortical knowledge synapses and axon
collaterals are small, and humans have about 10 14 to 10 15 of them (Mountcastle, 1998; Nicholls
et al., 2001; Nolte, 1999; Steward, 2000). Clearly, the survival value of instant arbitrary learning
vastly outweighs whatever inefficiency is incurred. This hypothesis helps explain one of the most