Page 135 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 121 21.9.2005 11:41pm

Mechanization of Cognition 121

currently expressed symbol (these neurons are shown in red in Figure 3.A.8, and reside in Layer II,
III, or IV [I am not sure]) with the apical dendrites of the now-validated action command neurons
of Layer V (shown in brown in Figure 3.A.8) to be incrementally strengthened. Essentially every
neuron representing the expressed symbol gets its direct synaptic connections with the action
command neurons incrementally strengthened.
Notice how different the situation of Figure 3.A.8 is from that of knowledge links. In a
knowledge link, the source symbol must first amplify its signal by briefly recruiting thousands
of transponder neurons to retransmit it. Even then, when the knowledge link signals arrive at
the target lexicon module, only a relatively small fraction of each target symbol’s neurons receive
a sufficient number of inputs to complete the link. In Figure 3.A.8, we presume that almost all
of the expressed symbol’s representing neurons synapse directly with the apical dendrites of
each Layer V action command neuron. The reason that this is a sensible hypothesis is that Layer
I is well known to be fed extensively with axons from the neurons below it (i.e., neurons of the
module that represent symbols), and to be profusely supplied with dense apical dendrites from
Layer V neurons.
The synapses from symbol representing neurons to action command neurons are hypothesized
to be quite different from those used in knowledge links. In particular, these synapses can slowly
and gradually get stronger (if repeatedly strengthened over many trials over time), and can
slowly and gradually get weaker (if not strengthened very often, or not at all, over time). This is
why ‘‘skill knowledge’’ decays so fast (in comparison with cognitive knowledge, which lasts for
very long periods of time, even if not used). A major benefit of this dynamic synapse characteristic
is that occasionally erroneous strengthening of synapses (e.g., when a random action command set
includes some irrelevant commands along with some effective ones) will, in general, not cause
problems (as long as the vast majority of strengthenings are warranted). This is very different from
cognition, where correction of erroneous knowledge is often impossible (and then the only solution
is to specifically learn not to use the erroneous knowledge).
The universal truism that ‘‘practice makes perfect’’ is thus exactly correct when it comes to
behavior. And for a difficult skill (e.g., landing a jet fighter on an aircraft carrier at night) to be
usable; that practice must have been recent. The associations from symbols to action command sets
are constantly being reshaped during life. If we live in a highly stable information environment we
might not notice much change in our behavioral repertoire over many years. If we are subjected to a
frequently and radically changing information environment, our behavior patterns are constantly
changing. In some respects, people who undergo such changes are being constantly ‘‘behaviorally
remade.’’ The workings of the neuronal network of Figure 3.A.8 are now briefly discussed.
Clearly, the size of the set of specific Layer V action command neurons which need to be
triggered by the expression of a particular symbol is arbitrary. One symbol’s association might
involve activating a set of five specific Layer V neurons, another might involve activating 79, and
yet another might activate no Layer V neurons. Keep in mind that each individual neuron in the
population of tens to hundreds of neurons which together represent one particular symbol in a
lexicon also participates in many other such representations for other symbols. So, this association
must be between the population representing a symbol and a specific set of Layer V neurons.
This requirement suggests a unidirectional Willshaw-type associative network structure
wherein the ‘‘retrieval keys’’ all have almost exactly the same number of neurons (which is exactly
what the symbol representation neuron sets are like); but where the ‘‘output’’ neurons activated by
each key have an arbitrary number of neurons. This is exactly what a Willshaw structure can do —
the retrieval keys (‘‘stable states’’) x k MUST be random and MUST each have almost the same
number of neurons; but there can be as many or as few ‘‘output neurons’’ in the associated y k as
desired, with no restriction; and the individual neurons making up each x k population can appear in
many other such populations.
(Note: If you don’t see this, consider again the computer experiments you performed in
Section 3.A.3. You will see that it does not matter how many y k neurons there are for each x k ,as