[From Bill Powers (2004.09.09.1904 MDT)]
Martin Taylor 2004.09.09.15.52--
The hormonal signals correspond more to broadcasts than to cable
distribution. They may affect the parameters of control systems at
several levels (at least I don't see a way of restricting chemical
distribution initiated at one level to having its effect only on
units at the level immediately above or below). They don't look like
signals of the same nature, to be combined with others, on the "cable
networks" of perceptual signals or of action signals.
Here's my stab at a discussion of broadcast signals.
As I understand it, the specificity of chemical signals is produced by the
target organ or cell systems that recognize signals by having receptor
molecules into which only certain signal molecules will fit. Only
variations in the concentration of those specific substances will be seen
as a variab le signal. Example: the thyroid-stimulating hormone affects
only the thyroid gland even though TSH circulates through the bloodstream
and reaches all possible target organs. High concentrations of TSH produce
a large output, low concentrations produce a low output (over a dynamic
range of around 10,000:1, yet no other organ even notices the TSH.
Thyroxine, the output of the thyroid gland, also reaches its targets
through the bloodstream, but here the picture is more complex. Various
target organs or cells chemically convert the thyroxine into a form that
they can sense -- one location into one form, another location into another
form. I don't know any more about it than that, but this certainly looks
like fanning out an output signal to multiple lower-order systems that do
different things.
Is my understanding of HPCT faulty in that it ignores the role of
broadcast signals, or is there a way in which these signals can and
do act like specific perceptual and/or action signals targets at
specific elementary control units in the hierarchy?
The originating system in the chemical hierarchy has nothing to say about
which subsystems recognize its output. As I understand it, that's
determined entirely by the receiving systems and the signals they're
prepared to recognize. As far as I know the higher-order system can't send
signals to one lower-order system at one time, and a different one at a
different time. I haven't heard of any biochemical control systems that can
alter which output molecule they produce, which would be necessary for
changing the target in a system where all signals reach all target systems.
We can, on the other hand, easily imagine neural output functions at higher
levels doing that by the simple process of gating. The myelin insulation of
neural pathways prevents most crosstalk, and different brain areas use
different neurotransmitter molecules which further reduces unwanted effects
of one system on its neighbors. And signals can be switched easily from one
wire to another.
In the neural hierarchy there is something like the broadcast effect in
that a single upgoing perceptual signal channel can branch into multiple
pathways, each carrying a copy of the original signal. The nature of neural
conduction assures that the same signal is present in all branches; the
signal isn't weakened by being split into multiple channels because the
impulsese are all-or-nothing in nature. Impulses are regenerated and
brought up to maximum amplitude by the mechanism of propagation. See B:CP
chapter 3.
Because perceptual signals can branch, they can reach multiple input
functions of higher level (the Byte articles, #3 I think, show how this
looks). When many signals branch and reach many higher input functions,
each higher input function can extract a different invariant by applying
different weightings (or other computations) to the same set of input
signals. So it's as if a general signal consisting of many neural signals
was being broadcast to a large set of receptors at the next level up, and
each system at the next level responded to the extent that the invariant it
was prepared to recognize was present, ignoring the others (once the PIFs
are orthogonalized bym reorganization).
Something similar happens between higher output functions and lower
reference inputs, as you noticed and embodied in your Reference Input
Function. A single higher output function may send copies of its output
signal to many RIFs at the next lower level; from the standpoint of each
RIF, the net reference signal is a function of multiple higher-order output
signals. I don't think there is any "recognition" in this downgoing path,
but who knows? I just haven's seen any need for it yet. Control systems are
very insensitive to the details of their output functiona, quite the
opposite of the case with their input functions. All the system needs,
basically, is to get the sign right in a majority of the connections to
RIFs, so the net feedback is negative for every loop. This, of course, is
closely related to the subject of orthogonality.
The connections in the neural systems are continually subject to
reorganization, in part by growth and atrophy of terminal arborizations of
axons and the synaptic connections they make. This means that on a slow
time-scale, the physical wiring of the brain can change. The only place in
the biochemical systems that I know of where similar reorganizations can
happen is in the immune system. Other than that, I don't think there is any
learning in the biochemical systems, but that could be very wrong. I know
too little about biochemistry to be spouting opinions like that.
We have to ask what the broadcast method accomplishes. It is a method for
sending signals to targets a long way off in space, and requires the
lock-and-key "packet" method to work for multiple signals. In the nervous
system, sending signals to distant targets is handled by a different
method: insulated wires, with a variety of neurotransmitters to handle the
problem of crosstalk in close quarters.
That's my understanding as of 040909.
Best,
Bill P.