[Martin Taylor 941221 14:10]
Rick Marken (941221.0900)
Dennis Delprato (941220)
I'm not sure that Rick's lucid description of the reorganizing system
actually answers Dennis's question:
Why is it necessary to posit reorganizing systems? In other
words, cannot the control systems reorganize without the need
for a separate set of reorganizing systems?
The reorganizing system is needed to explain why the result of learning
(changes in existing control systems or development of new one's) is
typically the maintenance of physiological variables in states required for
survival. ... [followed by a clear description of what the reorganizing
system is supposed to do].
Firstly, it is not clear that the result of learning IS typically the
maintenance of physiological variables in states required for survival.
This seems to me to be a view that depends on the view of reorganization
that it is intended to support, in other words, a circular proposition.
"Living things are alive" is what it seems to say, and what they have learned
is not incompatible with their staying alive. And one shouldn't forget
the maxim "A little learning is a dangerous thing." Sometimes, learning
does have the effect Rick mentions in his next sentence:
If control systems parameters were randomly changed, it is very
likely that many of these changes would quickly have fatal physiological side
effects.
If I understood Dennis's question, he was asking why the mechanism for
learning has to be a system separate from the perceptual control hierarchy,
not why learning is normally accomplished without causing the organism
to kill itself in the process. The question of "why a separate system"
is not answered by providing a description of what the separate system
does, no matter how clear and beautiful the description. It is necessary
to show that no integral system can do the job.
I have asked much the same question a few times, especially in the early
days of my acquaintance with PCT, and have not yet discovered a satisfactory
answer, either from CSG-L or from my own thinking. In fact, it has seemed
to me more than probable that there is no separate reorganization system.
It is clear that however reorganization occurs, it must have the property
that it USUALLY results in intrinsic variables being maintained near their
intrinsic reference levels--levels that may change over the lifetime of
an individual, but that are redetermined on an evolutionary time scale
rather than from the experience of the individual. But keeping
a signal near its reference level is the job of a control system, so
whatever the reorganization mechanism may be, it is a control system.
Rick's diagram shows such a control system, and it has to be correct in
principle:
intrinsic intrinsic
reference -->C<------ perception---------
> ^
intrinsic error |
v |
perceptual |
control physiological |
hierarchy---side effects-->physiological
> ^ variables
> >
v |
outside world
Now, this control system is drawn a little differently from the usual
way of drawing a control hierarchy, because the "physiological side effects"
are signals like any other, though they are probably chemical rather than
electrical. Normally such signals are drawn going upward, parallel to the
downward going reference signals that Rick has labelled "intrinsic error"
(I would have labelled that arrow "Output of the Reorganizing system.")
So, completing the part of this discussion that I hope is non-controversial,
here is a slightly modified version of Rick's diagram. It is modified
mainly in where the lines are placed, not in what they represent.
intrinsic reference
>
>
---intrinsic-----C---intrinsic error---
> perception |
> intrinsic
physiological output
variables function
> /
physiological /
effects /
\ /
---------- ------------------
> >
^ V
> >
perceptual
control
hierarchy
> ^
> >
v |
outside world
This is the same diagram as Rick's, except that I omitted the word "side"
from "physiological side effects" and added a stage "intrinsic output function."
Omitting "side" is a rather important modification to the diagram, because
it allows for the possibility of perceptual control that has the purpose
of causing a physiological effect, as well as allowing for side effects.
As I understand it, one reason for assuming that the upper part of the
diagram represents a separate "reorganization" system is that the states
of intrinsic physiological variables are not directly sensed, at least not
by the sensors that interact with the outside world, and thus the signals
corresponding to "intrinsic perception" cannot be derived from the
perceptual signals in the perceptual control hierarchy. This is clearly
true, since the outside world is not inside the body. But signals from
sensors inside the muscles are accepted as contributing to the perceptual
signals in the main hierarchy, and I see no evidence one way or the other
to suggest whether other signals from internal body states might enter
the main hierarchy. We may not be consciously aware of them, but that
does not mean that we don't use them.
Does the main hierarchy use ONLY neural axons and dendrites for its signal
paths? Nothing I know of in PCT theory makes any claim about the necessity
for such a limitation. Chemical states vary, and control systems within
cells control them in ways that seem to be understood by cell physiologists
(though not by me). It seems quite reasonable to suppose that, for example,
the sodium-potassium-calcium balance in the intracellular fluid might affect
the firing probability of some neurons under given conditions of input from
other neurons. If such a thing happened, it would act as a transducer that
brought physio-chemical states into the action of the perceptual hierarchy.
So, there seems to be no obvious reason why signals relating to physiological
states must be kept separate from signals derived from sensors of the outer
world. Signals from both internal and external sources might well
contribute to the perceptual signals that are controlled in the main
hierarchy.
Which brings us to the possible interpretations of Rick's diagram. The
two ways of drawing it suggest two different interpretations. In one,
the "intrinsic error" acts ON the perceptual control hierarchy, and in
the other, the intrinsic error acts IN the perceptual control hierarchy.
In the latter, the "intrinsic control systems" form the top-level ECUs
of the whole hierarchy, whereas in the former, they are the top-level
ECUs of a hierarchy for which the perceptual control system is the "outer
world."
Reorganization, as an effect, is the restructuring of the perceptual
control hierarchy, including the modification of output and perceptual
functions in individual ECUs. The usual result of reorganization is
that the perceptual control hierarchy is able to control, and it controls
in such a way that the intrinsic variables also remain under control.
This is true whether the intrinsic output works ON or IN the perceptual
control hierarchy.
If the intrinsic output works ON the perceptual control hierarchy, Rick's
description applies. If it works IN the hierarchy as a supplier of reference
signals for the rest of the hierarchy, then we need some explanation for
learning that does not demand the reconstruction of another kind of
reorganizing system outside the hierarchy.
In discussion the reorganizing system, one kind of "intrinsic variable"
is often implicated that is different in kind from the other, physiological,
variables. That interloper is related to the sustained error in the
perceptual control hierarchy. The exact nature of the relation is
not known, but is sometimes assumed to be the derivative of the square
of the summed error over the hierarchy or a region of the hierarchy.
It is quite easy to see how such a variable could have its value computed
by the neural system, and it is possible to see how it could be expressed
in chemical terms, to take its place among the physiological variables.
But nevertheless, it seems to be a variable unlike the other variables
that in themselves determine the health and survival of the organism.
The error of control, in itself, has no such effect.
The "orthodox" reorganizing system is a separate control system based on
intrinsic variables, which acts ON the perceptual control hierarchy, and
the error-based "intrinsic variable" is treated either with the others
or separately, depending on one's assumptions about the structure of the
reorganizing system. But if the physiological intrinsic variables are
taken to be part of a single control hierarchy, then the error-based
"intrinsic variable" is truly separate and different. It does not naturally
participate in the hierarchy, but is computable from the action of the
hierarchy, a hierarchy that DIRECTLY controls the physiological intrinsic
variables.
Learning, however it is done, usually has the result of reducing the error-
based intrinsic variable. An alternate plausible form of reorganization
places the locus of learning within each ECU or perhaps each group of
ECUs that interact closely. In this view, the only criterion for a
reorganization event is the local value of the error-based intrinsic
variable. To make this concrete, a single ECU might change its connections
and/or its perceptual or output function with higher probability the higher
the current value of the decayed integral of the derivative of the square
of the error signal. This means that in any short period of time, the
probability that the ECU undergoes a reorganization event is relatively
high when the error is large and increasing, or is increasing rapidly
whatever the magnitude of the error. The local environment of the ECU
determines its stability, not a separate reorganizing system.
An objection to this view of reorganization is that it requires new ECUs
to be inserted into an existing hierarchy between higher and lower levels,
rather than being added to the top of a growing hierarchy. This may indeed
prove to be a problem. It has been discussed on CSG-L, and may again be
discussed, but as yet there is no evidence one way or the other as to
whether the problem is real.
In a project I have mentioned before, we are developing a structure we
call the "Little Baby", in which some of these ideas of reorganization are
to be tested (in fact, the first tests are being started). I don't know
whether they will work in the environment we are trying (learning syntactic
structures without the use of symbols), but then, I don't know whether the
use of a separate reorganizing system would work, either.
To return to Dennis's question: I do not think that it has been shown that
control systems must have a separate reorganizing system in order to learn.
It seems quite plausible that if the control systems for intrinsic variables
are incorporated as part of the whole control hierarchy, and if failure
of an ECU to control increased the probability of reorganization in its
part of the hierarchy, then a control system could reorganize effectively
without having to posit a separate system for the purpose.
And to modify Rick's answer, the first sentence might be trivially reworded:
Rick:
The reorganizing system is needed to explain why the result of learning
(changes in existing control systems or development of new one's) is
typically the maintenance of physiological variables in states required for
survival.
Rewording:
The reorganizing system is able to explain why ...
Martin