Mechanization of Cognition - Biomimetics: Biologically Inspired Technologies - page 118

Since not all symbols of the answer lexicon of Figure 3.A.6 receive knowledge links from all
four assumed facts a, b, g, and d, what will be the input excitation sums on symbols that receive
fewer than four link inputs (total excitation level of the entire ensemble of neurons representing that
symbol in the answer module)? For example, consider an answer lexicon symbol u which only
receives links from assumed facts b and d. The total input excitation sum I(u) of the set of neurons
which represent u will be:

I(u) K [a þ log b ( p (b j u))] þ K [a þ log b (p(d j u))]
¼ 2K a þ K log b [p(b j u) p(d j u)]

(3A : 5)

Thus, given that each individual term in the first lines of Equations (3A.4) and (3A.5) lies between
K 10 and K 60, the value of I(u) (Equation (3A.5)) could, in extreme cases, be larger than that of
I(l) of Equation (3A.4) (although in most cases I(u) will be smaller and u will not be the winning
symbol). In any event, the symbol with the highest I value will win the confabulation.
Note that in cognitive functions which employ binary knowledge (every knowledge link
transponder neuron synapse is either unstrengthened or is ''strong''), I(l) is roughly proportional
to the number of links that symbol l receives. Thus, in these cortical areas, confabulation devolves
into simply choosing the symbol with the most knowledge link inputs. Although it is not discussed
in this chapter, this is exactly what such cognitive functions demand.
The seeming problem identified above of having symbols which are missing one or more
knowledge links win the confabulation competition is not actually a problem at all. Sometimes
(e.g., in early visual processing) this is exactly what we want, and at other times, when we want to
absolutely avoid this possibility, we can simply carry out multiple confabulations in succession
to form a sequence of expectations. Also, some portions of cortex probably have smaller dynamic
ranges (e.g., 40 to 60 instead of 10 to 60) for strengthened synapses, which also helps solve this
potential problem.
As discussed in Section 3.1 of the main chapter, in mechanizing cognition we explicitly address
this issue by appropriately defining a constant called the bandgap (related to quantity above).
In summary, the theory claims that the above-sketched biological implementation of confabu-
lation meets all information processing requirements of all aspects of cognition; yet, it is blazingly
fast and can be accurately and reliably carried out with relatively simple components (neurons and
synapses) which operate independently in parallel. Confabulation is my candidate for the greatest
evolutionary discovery of all time (with strong runners-up being DNA and photosynthesis).

3.A.6 Action Commands

At the end of a confabulation operation, there is often a single symbol active. For example, the
triangular red cortical neurons (belonging to Layers II, III, and IV) shown in Figure 3.A.2 represent
one particular symbol of the module which is now active following a confabulation. Of course, in
a real human thalamocortical module, such an active symbol would be represented by tens to
hundreds (depending on the location of the module in cortex) of ''red'' neurons, not the few shown
in the figure.
A key principle of the theory is that at the moment a single symbol of a module achieves the
active state at the end of a confabulation operation, a specific set of neurons in Layer V of the
cortical portion of that module (or of a nearby module — this possibility will be ignored here)
become highly excited. The outputs of these cortical Layer V neurons (shown in brown in Figure
3.A.2) leave cortex and proceed immediately to subcortical action nuclei (of which there are many,
with many different functions). This is the theory's conclusion-action principle. In effect, every
time cognition reaches a definitive single conclusion, a behavior is launched. This is what keeps us
moving, thinking, and doing, every moment we are awake.