(re)organization

[From Bill Powers (951027.0530 MDT)]

Peter Burke (951026.1335) --

     Consider for an ECU a number of input signals that get combined to
     form a perception. They are combined in a particular way, and of
     all of the things that might be combined, some are excluded. So my
     question really is how is it that the input signals that are
     included in the combination, and the particular form of the
     combination is the correct one for the particular reference signal
     of that ECU? I think that this may really be a question about how
     (re)organization takes place.

Ah, I see. Yes, it's a question as to how a new control system gets
organized. All I can tell you is a story; how I imagine it to come about
under my present conception of reorganization and neural function.

The first thing that has to appear is a perception. Reorganization, by
rearranging neural connections, creates a new signal that is a function
of existing lower-level perceptions or raw intensity inputs. Every
functional transformation of the input signals that is possible to make
out of the available neurons creates a perceptual signal that behaves in
some way as the inputs to the function vary. If there happens to be a
reasonably low amount of intrinsic error for a while, the perceptual
function will persist; otherwise reorganization will continue.

Once a perceptual function exists, the signal it generates, representing
some new aspect of the lower-level world, begins to be recorded. I don't
know how the recording process works, whether it works full time, or
whether there are special conditions under which recording takes place.
I also don't know how or to what extent memory association works -- how
recordings of perceptual signals become linked so one can evoke another.

Because memories are created in each elementary control system, the
played-back recording of a signal value is automatically associated with
the same system in which the perception originally appeared, and can
have only those values that have actually occured in that perceptual
channel. If, therefore, a memory is selected (by means as yet
unspecified, perhaps at random), the result will be a signal guaranteed
to represent a past value of a perceptual signal. The current value of
the perceptual signal can be compared with this replay of a selected
past value, and the difference can be computed (by subtraction, since
all signals are scalar). So, given a comparator function, we now have an
error signal: the difference between the present perception and some
value that perception has had in the past. If you liked that past value
(i.e., it was associated with low intrinsic error conditions), then you
might want to try to make the present perception match it.

All that remains now is to use the error signal as the basis for action.
Action consists of sending amplified copies of the error signal, through
some kind of output function, into the world that this system can
affect, which consists of the already-existing lower levels of control,
or in the very beginning, the muscles. Now the problem is to apply the
error signal to the reference inputs of lower-level systems (whose
actions alter the inputs to the new perceptual function) with the
correct sign (excitation or inhibition) to make the error signal
smaller. As those connections are formed by trial and error irrelevant
outputs are eliminated, incorrect signs are reversed, and the new
perceptual signal is brought closer and closer to the value of the
selected past value, the reference signal.

(I have demonstrated in simulation that it is possible to accomplish
this convergence to a working control system without any knowledge of
the external world and strictly on the basis of the error signal).

The final result is a new control system at a higher level. The system
has become able to act on the world outside it in such a way as to
reproduce in present time a past value of a perceptual signal.

This picture includes memory, but it is quite possible to construct a
control hierarchy without memory, as long as there is some source of
signals to serve as reference signals. In this model where all neural
signals are scalar quantities, all reference signals are physically
alike: they are just neural signals that have higher or lower
frequencies. There is nothing special about a signal that makes it a
reference signal as opposed to an error signal, a perceptual signal, or
an output signal. There is no physical difference between signals in one
control system and signals in a different one. All neural signals are
alike.

What makes the difference is the neural function performed on the
signals, and the connections between the output of the physical
implementation of that function and the inputs to other functions
(perhaps this is what the "connectionists" have discovered). So if there
is a comparator receiving a perceptual signal, ANY second signal
affecting the comparator with an opposite sign will function as a
reference signal. The origin of the second signal makes no difference --
if could come from an experimenter's stimulating probe. Its significance
is given by where in the system it enters. Thus postulating that a
reference signal comes from memory is optional; it doesn't matter, for
the control system in question, where it comes from.

Returning to memory for a moment. Recording scalar signals doesn't make
much sense, since all you get is a numerical value. However, if memory
is really a structure with lateral connections, then it makes a
difference whether the replayed signals in many parallel control systems
are replayed in _combinations_ that have occurred before. All behavior
is multidimensional, so to reproduce a past experience generally
requires setting multiple reference signals to some pattern of values.
Now it does make a difference if that pattern has never occurred before.
It may be impossible for some combination of reference signals to be
satisfied at the same time, even though they can be satisfied
separately. Consider perceptual signals pertaining to the spatial
positions of your hands. You never experience the two hands occupying
the same space at the same time, although you can place either hand
alone anywhere within the possible range. So you will have no memories
in which those two perceptual signals imply the same spatial location at
the same time. The time index, therefore, may be a way of selecting
combinations of reference signals that are in fact feasible, because
they have occurred before.

You would, of course, discover the same problem without the aid of
memory, because you can't actually put both hands in the same place;
there would be a conflict between the control systems, which presumably
would get resolved through reorganization. I don't know how necessary
these ideas about memory are. As I say, the hierarchy can be made to
work without memory, although without some form of memory, remembering
can't be accounted for.

···

-------------------------------------
You may have noticed something missing in this story: any mention of the
actual properties of the real world, except at the end. The point of
this story is to try to see the problem of becoming organized to control
experience as it would appear from inside the brain becoming organized.
In the beginning there are no organized perceptions; there is nothing to
control. Perceptions have to be created by connecting neurons and
ajusting their input weightings. There can be no question of perceptions
struggling to get from the real world into the brain, where they would
overload the machinery if the unwanted ones were't filtered out. Until a
perceptual function is formed, the perception it's going to generate
simply doesn't exist anywhere. The brain's problem isn't how to keep
perceptions out; it's how to make any coherent perceptions appear so
they can be experienced and controlled.

Even after perceptions have appeared, the brain must still discover what
outputs to generate to bring these perceptions under control. Since the
brain knows nothing of the external world, it has to do this
experimentally. There is no _a priori_ reason for producing any given
output; the only way the right outputs can be found is to try everything
that is possible and retain the connections that result in bringing the
perception closer to the reference level specified in the form of a
reference signal. The environment is a black box.

And why would any particular way of perceiving be retained, and why
would any perceptual signal be brought to some specific value? Only
because doing so makes the organism feel better. The driving force
behind reorganization is a set of intrinsic reference levels which, when
not satisfied, lead to reorganization. The final result is a hierarchy
of control systems controlling major aspects of a world of perceptual
signals created by neural functions which themselves are products of
reorganization. In this hierarchy we find all of human experience,
including theories about behavior, explanations of why doing some things
is good and doing others is bad, concepts of biology, psychology,
physics, and all the rest. In one corner of one brain, we find the story
I have just told.
---------------------------------------
In order to work this way, a brain would have to start with some degree
of organization. This might include some pre-wired (although not
permanent) ways of perceiving, even some built-in (but temporary)
control systems. The intrinsic reference levels that provide the
ultimate criteria of good and bad must be inherited, as well as the
reorganizing machinery. I have proposed that the brain is predisposed to
becoming organized in levels, with specialized neurons at each level
suited to perceiving different classes of variables derived from lower-
level signals (although no particular examples of them). Just how much
preorganization exists, and what kind it is, remains to be settled. One
of the goals of those who model such organizations, which crops up now
and then, is to find the principle of reorganization that would allow a
simulation to start with no organization at all, and build itself into
an organized system. I doubt that this is really possible; most actual
attempts to do this of which I know have had to install a considerable
degree of organization to make further self-organization possible.

Any permanent preorganization that might exist is a two-edged sword. It
makes further self-organization easier, but puts constraints on what
self-organization can accomplish. A horse is born knowing how to stand
and walk, how to produce certain gaits, how to jump. But when horses are
taught different gaits they are extremely limited and awkward, because
the built-in control systems are hardly modifiable. When people watch
the Lipizaners [sp?] going through their paces, they say "Oh, how
graceful, how beautiful, how skillful!" But they should add, "... for a
horse." A person who moved the same way would be considered clumsy.

Built-in control systems reduce skill because they can't be adapted to
all the details of the world into which an organism is born. They can't
be improved through further learning. They can even get in the way of
skillful control. In a human being, an instinctive fear of fire would
have been a great disadvantage, but in a moth it would have had
considerable survival value. What we see in the apparent progression of
species, along one dimension anyway, is a simultaneous decrease in
built-in organization and an increase in the capacity to reorganize
within a single lifetime. Of course all organisms without exception have
to inherit a working body with all its life-support systems functioning;
if the brain is not kept alive and well, nothing of much interest will
follow.
----------------------------
Addendum.

Suppose there is no memory involved in a new control system. If we're
talking about a control system at the highest existing level (at some
stage of development), there is no higher source of reference signals.
If we imagine that a perception has been formed, and the signal enters a
comparator, there can be no reference signal also entering the
comparator. This is entirely equivalent to the presence of a reference
signal with a value of zero. A control system with a reference signal
set to zero behaves to maintain the perceptual signal near zero: it
avoids experiencing any amount of that perceptual signal. So we would
predict that the first control processes involving higher levels of
perception would avoid perceiving the new variable! Could this explain
the fact that when babies develop the ability to recognize different
faces, they also develop a "fear of strangers?"

Just a shot in the dark.
-----------------------------------------------------------------------
Best,

Bill P.