Page 136 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 122 21.9.2005 11:41pm

122 Biomimetics: Biologically Inspired Technologies

long as we are not implementing the second, y field to x field, part of the cycle (this is not well
known, because for analytical simplicity, the original Willshaw model used the same number of
neurons in both the x k and x k vectors). Further, as long as the x k keys are random and have almost
exactly the same numbers of active neurons, the reliability of the y k neuron responses is extremely
high.)
However, as mentioned earlier, unlike the situation in knowledge links (where only a few of the
target symbol neurons receive connections from the transponder neurons of a source symbol), in
this case, all of the neurons of each active symbol must connect to each of the desired action
command neurons. Partial connectivity will not work here, since there is no feedback to implement
a ‘‘convergence’’ process. But the enormous local connectivity within Layer I is hypothesized to
make, achieving a sufficient level of this connectivity no problem.
By incorporating inhibitory neurons into its intrinsic design, such a one-way Willshaw network
(with inhibition added) will only respond with a y k when its input is a newly active SINGLE symbol
(multiple symbols will fail to yield any association output because they induce excessive inhibition;
which shuts down all of the Layer V neurons). This is hypothesized to be why action commands are
only issued when a confabulation produces a single winning symbol. Also, when considering what
action to take for a given x k input, only those Layer V neurons having a sufficient input excitation
will respond (much like in confabulation competitions). In other words, even near the beginning of
learning, when behavioral symbol to action associations are all weak, the Layer V response will be
based upon this ‘‘competitive’’ criterion, not a fixed threshold.
That a vast majority of cortex would be involved in issuing thought action commands, as
opposed to movement action commands, makes sense because there are many more feature
attractor modules (and knowledge bases) than muscles. (It is not discussed here, but each know-
ledge base may also need to receive an ‘‘operate’’ command in order to function — if this is true,
this function probably involves the large ‘‘higher order’’ [Sherman and Guillery, 2001] portion of
the thalamus that is not included in the thalamocortical modules.) So it probably requires a much
larger portion of cortex to producing such thought process control action commands (muscle action
commands come mostly from Layer V of lexicons located within the relatively small primary motor
area of cortex).
Most action commands represent ‘‘low-level housekeeping functions’’ that are executed reflex-
ively whenever a single symbol (often one of a large set of symbols that will elicit the same action
command set) is expressed on a lexicon. For example, if a confabulation in a lexicon that is
recalling a stored action sequence (such lexicons are typically located in frontal cortex) ends in the
expression of a single action symbol, then that lexicon must be immediately erased and prepared for
generating the next sequence symbol. This is an action command that is issued along with the
expression of the current action sequence symbol. Overriding such reflexive thought progressions is
possible; but generally involves shutting off tonic cortical arousal (one of multiple adjuncts to the
lexicon operation command input) in a general cortical area via action commands issued to
brainstem thought nuclei. The result is a momentary freezing of the halted function as a new
thought process stream is inaugurated. This is what happens when we see that we are about to step
on dog poop. It takes a only fraction of a second for us to recover from the suspension of the
ongoing action and activate an alternative. Further, since muscle tone and rhythmic actions such as
walking are nominally controlled by other brain nuclei (not cortex and thalamus), all the cortex
typically needs to do (once the prior action sequence has been suspended) in such instances is issue
a momentary set of corrective alteration action commands which are instantly executed as a
momentary perturbation to the ongoing (subcortically automated) process, which then typically
resumes.
It is important to note that the details of how sequences of ‘‘action’’ symbols — each represent-
ing (via its symbol to action command association) a particular specific set of action commands that
will be launched every time that symbol is the sole conclusion of a confabulation — are learned,
stored, and recalled are the same as with all other cognitive knowledge. However, unlike many