Thinking aloud

[From Oded Maler (931208) noon]

* (Bill Powers (931206.1515 MST))

···

*
* Oded Maler (931206) --
*
* >... (so they tell me) there's a lot of room for fancy
* >perceptual functions, that can give you various distorted and
* >noisy images of boss reality. Partial, indirect and unreliable
* >information is dealt with by control engineers and
* >mathematicians.

(What I wanted to say is that there are tools to express a perceptual
signal p, as a probability distribution depending on one ore
more CEVs. It is not the question of accuracy which is crucial but
the question of objective (however noisy) definability.)

* The problem isn't quite that easy to solve. Just consider this
* question: if an artificial system's perceptions are distorted in
* relation to the Boss Reality, how would you go about determining
* that fact? Just to make the problem more concrete, suppose your
* artificial system had to perceive the distance from it to some
* object. What procedure would you follow to find out whether its
* perception of distance was accurate?

First of all, the "procedure follower" is me, that is, a limited
observer of boss reality equipped with some measurement devices.
Suppose I have an artificial distance sensor, I put it in front of a
wall ("objects" and their boundaries pose additional problems) and
read its measurement. In order to see that this perception is
accurate I need another (provably accurate) device and compare
its measurement with the tested one. This is of course very limited
theoretically but it works in practical situations. The question
is whether the "real" distance has an objective existence as function
of the world - I would say yes because sometimes you bump into
the wall and sometimes you don't.

* Then, instead of determining how well the artificial system's
* perceptions represent the actual distance, suppose you wanted to
* check the accuracy of your own perception of the distance of the
* object from you. Now what procedure would you follow? To see the
* nature of the problem, you have to realize that whatever method
* you describe, I will ask you how to determine whether that method
* gives an accurate picture of the distance. Please give this a try
* and see what you come up with.

About myself, one problem is that I am inflicted with consciousness,
and I have first to separate my perception into implicit and explicit
parts. When I look toward the wall my perception of the distance is
not in terms of objective coordinates, only when I contemplate about
it I can say "it's 50cm" or "if I put a distance sensor it will probably
show 50cm". This is not what you mean. The answer you are expecting
is the following: in order to verify my perception I'll try to reach
the object with my hand (or run away from it if it is perceived as
bad) and I'll do it independently of boss reality. The objective distance
does not exist within myself (when I get rid of higher levels and
consciousness), yet my perception is related to it somehow.

The question is whether something is going on objectively in such
situations, regardless of our ability to penetrate into and express
this "something". This is what Science was supposed to be - try to
find an external description (= mathematical model) that approximates
phenomena. Not having an "external description" theory in the case of living
systems, is either an indication that a new math is needed, or that
the kind of scientific explanation for the living world is not the
same as we were told in kindergarten.

--Oded

p.s.

Concerning the enthropy discussion, I remember once browsing at some
thesis from the Hebrew University were the author said that there
were many failed attempts to find physical meaning to informational
entropy (i.e., to connect thermodynamics to information). Being
ignorant in both I don't have anything to add on this, but I might locate
the reference in my forthcoming visit to the holyland.

--

Oded Maler, VERIMAG, Miniparc ZIRST, 38330 Montbonnot, France
Phone: 76909635 Fax: 76413620 e-mail: Oded.Maler@imag.fr

[From Bill Powers (931208.1140 MST)]

Oded Maler (931208) --

Sorry to keep mixing your name up with Hans Blom's. If this keeps
up I may have to seek professional help (a plumber to stop the
leak).

(What I wanted to say is that there are tools to express a
perceptual signal p, as a probability distribution depending on
one or more CEVs. It is not the question of accuracy which is
crucial but the question of objective (however noisy)
definability.)

The conventional way to think about such probabilities, as I
understand it, would be to speak of "the probability that this
object is an X." The way the PCT model is constructed, with one
degree of freedom per perceptual variable, there would be a
slight change: "The probability that the degree of perceived X-
ness is y." The perceptual signal in one control system can only
represent one variable, so if there is any signal at all, it
represents the X that the system is organized to perceive.

The first way treats X as a category, with membership of the
object in category X being uncertain but with no intermediate
values possible. The second treats X as a continuous perceptual
variable. In the second case, y is the mean magnitude of the
perceptual signal standing for AMOUNT OF X-ness, and the
probability is associated with a random noise distribution
superimposed on the steady value y.

I think these two cases are sometimes confused in net
discussions. For example, the second case is sometimes spoken of
as if all signals were near the threshold of detection. That
makes the two cases equivalent, because the second case becomes
"the probability that the value of X is nonzero." That reduces
the continuous case to a binary one, zero or nonzero, and makes
it look like the categorical case.

If a perceptual signal is well above threshold, then in the
control model what matters is its magnitude. The lower-level
perceptual signals, under ordinary conditions, are many standard
deviations of the noise level above threshold (just look for
random noise in your perceptions of the room you are in: there
isn't any). This noise level determines the dynamic range of
control, but it has no relevance to the probability that a given
signal is present (or, the probability that it is present is 1
with negligible uncertainty).

So the continuous case does not normally present a problem of
identification in any probabilistic sense. The categorical case
does present such a probabilistic problem, but only for normal
signals from lower levels that define a perception on the
boundary between two categories. Even that may not necessarily be
a statistical problem; both category-perceivers could produce
unequivocal signals. The problem is then a logical one: which
signal to treat as relevant in the current context.

Neither of these ways of treating a perceptual signal is relevant
to the epistemological problem. That problem is not one of
probability or magnitude , but of physical impossibility. It is
impossible for a system that knows the environment only through
perceptual signals to test the correspondence between its
perceptions and hypothetical entities outside to which those
signals are thought to correspond.

To bring probabilities into this subject is to beg a question. If
you can't say whether there is an external X' corresponding to a
perceptual signal X, then asking the probability that X
represents X' simply assumes that X' does exist. The underlying
question is given an answer by the very way the question is
asked.

When I look toward the wall my perception of the distance is
not in terms of objective coordinates, only when I contemplate
about it I can say "it's 50cm" or "if I put a distance sensor
it will probably show 50cm". This is not what you mean. The
answer you are expecting is the following: in order to verify
my perception I'll try to reach the object with my hand (or run
away from it if it is perceived as bad) and I'll do it
independently of boss reality. The objective distance does not
exist within myself (when I get rid of higher levels and
consciousness), yet my perception is related to it somehow.

Almost, but not quite. I'll let you calibrate your perception by
using a meter stick. You see on that meter stick, "50 cm." The
very fact that you see it says that this is a perception -- both
the meter stick itself and the number on it. Now, as I promised,
I must ask "how do you determine that there is actually a meter
stick there, and that there is something on it corresponding to
your perception of '50'?"

You might answer that you could feel the meter stick by touching
it. But touching simply yields another perception. We could carry
this through as many steps as you like, but eventually we would
agree, I think, that all you can do is compare one perception
against another. None of these perceptions reveals the "objective
distance" -- that is, something that exists independently of
perception.

The question is whether something is going on objectively in
such situations, regardless of our ability to penetrate into
and express this "something". This is what Science was supposed
to be - try to find an external description (= mathematical
model) that approximates phenomena. Not having an "external
description" theory in the case of living systems, is either an
indication that a new math is needed, or that the kind of
scientific explanation for the living world is not the same as
we were told in kindergarten.

Think about scientific explanations. They are all based
ultimately on statements about perceptions, aren't they? It
doesn't matter whether we are explaining living or nonliving
systems. Instrument readings are known to us only as the
appearance of needles on scales or displays of digital numbers.
It's only theory that puts labels on the meters stating what the
perceived reading means, and that constructs mathematical
relationships between the numbers that we read. But theories
reside in our heads, in the same place where perceptions live.

You said "The question is whether something is going on
objectively in such situations... ". That is indeed the question.
Before control theory, it was impossible to answer. All the facts
about ALL situations seem to point us toward solipsism. Of course
solipsism is ridiculous and nasty and self-defeating, as well as
conducive to madness, so it's important to do away with it
whether we know how to do that or not. Philosophers have gone
through all sorts of verbal contortions to get rid of solipsism,
but the facts all still seem to indicate that that is the answer.

PCT provides the way out. Certain of our perceptions are called
"actions." By producing actions, we can cause changes in other
perceptions. We can often see that there are no visible links
between the actions and the changes in perceptions. From this we
can deduce that there are hidden connections between action and
perception. This is evidence about an unperceived world. Also, we
can see that some perceptions will change spontaneously, even
when we don't act. That is more evidence.

By acting systematically, we can build up a model of the unseen
world. The model proposes a structure of some kind, containing
points where our actions can affect it, and other points where
the hypothetical reality affects our perceptions. Once the model
fits the relationship we can observe between an action and other
perceptions, we can use the model to predict a new effect on
perception resulting from some new action. If the prediction is
wrong, we try to modify the model to cover not only the first
case, but the new one as well. In this way models gradually
become more general, fail less often.

After three centuries or so of modifying models describing the
relationship between actions and perceptions, we have built up
some pretty serviceable models: physics and chemistry, for
example. Even small failures of these models is material for
front-page stories. But all of these models were constructed
without the realization that human perceptions played a
fundamental part in their very definitions. Most physical
scientists thought they were studying reality itself,
objectively. In fact they were studying their own perceptions.

Now we are building a model that is consistent with physics and
chemistry, but in which human perception plays an explicit part.
We have found cases in which it seems clear that there are no
external entities that correspond to some perceptions, like the
taste of lemonade. In other cases, it seems that there must be
external counterparts; for example we can experience a phenomenon
called "light" with our eyes, but we can also build devices that
report the presence of something under the same conditions that
we see light, and the existing models of physics and chemistry
account for both modes of perception in considerable detail.

However, between perception and that which is perceived there are
still input functions with forms that are an integral part of
whatever we perceive. If we want to move a step closer to
grasping what the external world is like, we must learn how these
perceptual functions are organized, and apply their inverses to
the world of perception. This will be necessary whether we are
speaking of lemonade or control systems, whether the theories we
are discussing are those of psychology or those of physics.
Between the external reality and all of our models there is still
an unknown transformation. When we learn what that transformation
is, all of our models will look different.

This is why I say that in the final analysis, PCT is about
perception and about the fact that we can act to control our
perceptions. It is not about objective models of control systems;
those models are merely attempts to explain what must be going on
behind the scenes for us to be able to act in a way that controls
perception. If we were all struck by amnesia right now and forgot
all we know about control systems, physics, and all the rest, it
would still be possible to observe that by acting we can control
our perceptions. That is the bed rock on which we can build a
theory of life.

Theoreticians need to put their theories aside periodically, and
simply experience the world. Theories exist to explain
experience. If you don't pay sufficient attention to experience,
with the theorizing module turned off, theory can perform a
random walk into universes that never were.

···

--------------------------------------------------------------
Best,

Bill P.