[From Bill Powers (920324.0300)]
Copy direct to Mark Olsen, plus CSGnet
My address is powers_d%flc@vaxf.colorado.edu. Between the powers and the d
is an underscore (shift hyphen) and between the d and the flc is a percent
sign.
Send Closed Loop orders to Mary Powers, $5 each copy.
CSG
73 Ridge Place, CR 510
Durango, CO 81301
Mark Olsen (920323) --
You ask about the functions relating one level of perception to another.
This is indeed the question that HPCT poses -- but doesn't answer. What
lies behind HPCT is not any proposal as to how each level of perception is
derived from the one below it, but a proposal as to what the levels of
perception are and how they are related. This is the phenomenon that any
model must in the end explain.
The "H" part of HPCT can be taken in two ways: first, as a general sketch
of a hierarchy of control in the abstract, with the communication between
levels consisting of a series of perceptual re-representations of reality
and a corresponding set of reference signals used to control lower levels;
second, as a series of proposed levels of perception (and control) based
directly on an analysis of experience with the hierarchical control concept
as a guide. This is a beginning model; there may well be other modes of
communication between levels, but the basic one is probably valid.
The definitions of levels define the modeling problem. We can see that the
sensation level is probably derived by weighted summations of intensity
signals, the weights defining a vector in a perceptual space having fewer
dimensions than there are different sources of intensity signals. But that
answer to the modeling problem comes after noticing that sensations seem to
depend on intensities in a particular way, a way that could be modeled as
weighted summation. The phenomenon to be modeled comes before the model.
And that's as far as I can go. I don't know how configurations are derived
from sensations -- how it is that we can get the sense of, say, a
particular person's face over a range of distances and orientations and
expressions. If signals standing for the dimensions of a face existed, then
it's possible to make a rough guess that transitions of the face from one
state to another would be sensed using time functions and partial
derivatives; that's a feeble start toward a functional model that you could
run on a computer. As to the rest of the levels, the kinds of computations
involved are mostly a mystery to me. The few guesses we have come up with
are strictly stabs in the dark. You can use words like "integration" to
describe how some kinds of perceptions are put together to create others,
but the word is just a noise. It doesn't tell us anything about the
processes involved.
Behind this exploration of perception lies a fundamental postulate; if you
don't internalize it, I don't think you can even get started on the problem
of modeling the brain's perceptual systems, or for that matter, in
understanding HPCT. The postulate, simply put, is this: it's all
perception.
By that I mean that no matter what you attend to in the world of
experience, whether you refer to inner or outer experiences, concrete or
abstract, verbal or nonverbal, the object of your attention is a
perception. You are looking at or otherwise experiencing the brain's
perceptual activities, not the objective world itself.
Vision is the most important sense to understand this way if you're
sighted; understand vision and the rest (touch, taste, sound, etc.) will
follow. The world you see begins as pixels (individual picture elements).
The pixels are so close together that you see no spaces between them,
although the sensory nerves do not overlap and in fact do not completely
fill the retina. There's a world between the pixels, but we don't see it
unless the view shifts slightly -- and then what we had been seeing
disappears into the cracks between the pixels. This is invisible to direct
experience; the world seems continuous over the whole visual field. We get
a sense of seeing the world at infinite resolution, and can't imagine what
the whole field would look like if we had, say, ten times as many retinal
receptors and the optical acuity and brain power to take advantage of them.
This would be like seeing the world through a magnifying lens, except that
the whole world would look that way, not just one little part of it (which
we still see at human resolution). The only way to imagine this is to go
the other way: view the world at a lower resolution, as in a halftone
photograph or a television screen seen close up, and imagine that the
result is the only world you can ever see. That's how our picture of the
world would look to a different organism with higher visual resolution. But
we experience it as having continuous detail right down to the level where
it appears smooth. I suppose the fly sees the world in the same way. But
its world is smoother than ours.
Building up definitions of the rest of the levels in the hierarchy is then
a matter of noticing persistent types of structure in this world of picture
elements. The first level above the pixels themselves is sensation, a type
of perception that can't be analyzed in any way except into variations of
intensity. Color is a sensation, as is shading.
Perhaps things like edges are sensations, derived in one step from the
pixel distributions. When analyzing perceptions, however, don't use any
data but your own experience. Theory and neural data will tell you that in
the visual field, in the retina itself, all edges are enhanced, so that
there is a strong outlining effect. But look at the edge of a sheet of
paper on a dark tabletop. There is no outline. The closer you look at the
edge, the more nearly it seems to be an infinitely sharp line separating
uniform white from uniform dark. The edge itself is there -- but you can't
see it as an object. It's just a sense of edgeness. Only under special
conditions, as in looking at a smooth gradient of illumination going over a
relatively short distance from white to black do you see edge effects like
the "Mach band", the only clear subjective evidence of edge enhancement.
However those neural signals enhanced at edges are processed, the result is
that step changes look like step changes, not outlines as in cartoons.
Whatever model we come up with for how the nervous system processes pixel
information, it must result in edges that look this way, without borders.
If it doesn't, the model is wrong.
The next step is to notice that the edges and corners and broad white areas
of the piece of paper add up to -- a piece of paper. If you've made this
transition properly, it will come as a surprise. Where did that piece of
paper, or piece-of-paperness, come from? It wasn't there in the edge, or
the corner, or the whiteness, or the darkness. It comes into being only
when all those elements are seen grouped into a thing, a configuration with
a familiar shape, orientation, distance, size, and so on. The Gestalt
psychologists of old spent a lot of time looking at things like these. They
should have kept going. Or perhaps they shouldn't have been cowed by the
behaviorists.
You have to go slowly and by the smallest steps you can devise. If you go
too fast you'll miss the smallest steps; if you miss the smallest steps
you'll lose the sense of examining perceptions, and start projecting the
visual field into an external world again. You'll jump to the more abstract
levels and lose the connection from one level to the next. This is, if you
like, a form of meditation on experience in which you distance yourself
from experience and look at it merely as a display. You're not trying to
see anything about the world, but only something about the display. You're
trying to see what features the person who constructed it thought of
putting into it, just as when you read a program you think to yourself "Now
he's setting up an array to hold the results" instead of just reading the
code, or when you read a novel as a literary critic you think "Now he's
introducing tension" instead of just getting tense. Who the "he" is is
immaterial -- the point is to see what is before you as a construction that
has inner organization, and try to see how it is put together.
The general principle is that when you have found a level, like sensation,
the next level is going to depend on it; also, the current level depends on
the one below it. If you analyze a perception to see what it is made of, at
first you see just more perceptions of the same level -- big configurations
are made of little configurations. But when you analyze in just the right
way, you suddenly realize that all configurations, of whatever size or
kind, are made of sensations, which are not configurations of any kind. And
you realize that if it weren't for the presence of those sensations, there
couldn't be any configuration to see: a field consisting of a single
sensation, such as white, can't lead to any sense of configuration. There's
a relationship between these levels of perception. That gives us a hint
about building models of perception, a hint about how the brain's
perceptual system is constructed.
Sometimes you will identify what seems to be a higher level of perception,
some characteristic common to all perceptions, unconnected to lower levels
you have previously seen. Then you can use this kind of analysis to try to
fill in the gap. What is this new perception made of, besides smaller
perceptions of the same kind? When the gap is large, the missing steps are
obvious. You can, for example, look at spatial relationships such as "on"
-- something being "on" something else. You can see the on-ness clearly,
it's right in front of you. But what is it made of? If you said
"sensations" you would clearly be making too large a jump, because on-ness
involves objects, things, configurations. Some kind of object is "on" some
other kind of object. If it weren't for the impressions of distinct
objects, there couldn't be any sense of the relationship between them. But
is that step small enough? I've had to put two levels between relationships
and configurations: transitions (which can be zero), and events (which can
be as simple as mere duration). Seeing something "on" something else
involves more than a brief contact; there must be duration.
Perhaps someone else could find smaller steps still, or would characterize
the intervening steps differently. There's still a lot of room for
improving the definitions of the phenomena we're hoping ultimately to
model.
I'm not talking here about the models themselves. I'm talking about the
attitude you take toward your own experiences when you're trying to notice
phenomena that need modeling. If you were a physicist you wouldn't be
taking this attitude. You'd treat the world of perception in the normal
unanalytical way as if it lay outside yourself where everyone could see it,
and you'd search for laws relating changes of one kind of perception to
other kinds of perceptions. You would call these "natural laws" or
"behavioral laws" and assume you were discovering truths about an objective
universe.
As a CT psychologist, however, you have a different objective: to grasp the
natural world as a manifestation of human perception (your own), and to
ferret out of it some regularities that tell us about perception rather
than about the world perceived. If you stumbled onto this attitude
accidentally, without understanding what you were doing, you might well
find yourself in a state with a clinical name: dissociation. I don't
recommend this attitude as one suitable for ordinary living. It's difficult
and uncomfortable, and it tends to strip the meaning from experience (until
you get past a certain point, after which you realize that meaning, too, is
perception, and let it back in). If you're afraid that understanding your
girl friend as a set of intensities, sensations, configurations,
transitions, events, relationships, categories, sequences, programs,
principles, and system concepts in your brain might strain your feeling
toward her (and hers toward you), don't do this with your girl friend. Do
it with somebody else's, or a laboratory rat. It doesn't matter who or what
you do it to, because you're really talking about your own perceptions.
This is a private experience valid only in one person's world. It can
become public only to the extent that different people independently arrive
at the same analysis. I've always hoped for that, but only a very few
people, to my knowledge, have tried this for themselves. Most people just
memorize my definitions, which unfortunately are in words. It's easier to
push words around than to shut up and examine direct experience.
You'll hear objections to this process alluding to introspectionism, which
failed to get anywhere a long time ago. But introspectionism didn't fail
because it looked at the kinds of things I'm talking about here. It failed
because it confused the subjective with the objective (and so did its
critics). The world that I'm speaking of examining here would be called, by
most conventional scientists, the objective world, not the subjective one.
I'm not recommnending shifting attention off the objective world and
plunging into the dim and uncertain world of inner phenomena -- or what we
imagine to be inner phenomena. I'm recommending a change of attitude toward
the world we normally consider to be the objective one, which includes the
world outside us and our bodies as we experience them. I'm saying that you
will learn something if you look on this world as directly experienced
evidence about the nature of your own perceptual system, and only in a
conjectural way about the world that is actually outside you.
Instead of treating relationships like on, beside, after, with, and into as
properties of the external world, look on them as perceptions constructed
on a base of lower-level perceptions. Instead of seeing categories as made
of things that are inherently alike, think of categories as ways of
perceiving that MAKE things appear to be alike -- things that are actually,
at lower levels of perception, different. Instead of seeing sequential
ordering as a fact of nature, see it as a way of putting ordering into an
otherwise continuous of miscellaneous flow. In short, take nothing about
experience for granted, as if some aspects of experience were really
outside and others were inner interpretations. Put the whole thing inside,
and see what you come up with when you understand that it's all perception.
All of it.
Final notes:
In HPCT diagrams, we show signals coming out of perceptual functions and
going into higher-level ones (as well as the local comparator, if the
signal is under control). I think of these lines as representing single
neural signals that vary in only one dimension: how much. This can be
confusing, because we don't experience single signals under normal
circumstances (when we do they cease to be meaningful). Instead we
experience all the signals within the scope of awareness, at every level in
the state we call conscious. To understand what the single-signal concept
means, you have to break this world of simultaneous perceptions into its
components, the individual and independent dimensions in which the totality
of perception can vary. You have truly identified one isolated perception
when it can vary only in the degree to which it's present, which we
experience as its state. If the perception varies without in the slightest
changing its identity, you have probably noticed a single signal.
This can be important when you talk about control. We talk loosely about
controlling "a dog," for example. But that way of talking is really lumping
many independently variable aspects of the dog together. You don't control
its species, or its eye color, or the length of its tail. You don't even
control its behavior. If it's behavior you're controlling, you always
control SOME PARTICULAR VARIABLE ASPECT OF THE DOG'S BEHAVIOR. You may
control the radius within which it can move, by putting it on a chain. You
may control its speed of walking by saying "stay" or "follow," and its path
by saying "heel." Whatever you control, it must come down to a single
variable or small sets of variables independently controlled. If you're
controlling in more than one dimension, you must sense more than one
variable, and have a control system operating independently for each one.
That's because independent dimensions can be independently disturbed; you
need independent control systems so that a disturbance in one dimension can
be corrected without necessarily causing an error in another dimension.
None of this answers your question as to how perceptual signals in a
diagram depend on perceptual signals lower in the diagram. The only general
answer I can give is that some computation lies between them. The input
data consists of lower-level perceptions; the output data, the higher-level
perceptual signal, represents the value of the function being computed over
and over or continuously. At each level, I presume (judging from the way
the context changes every time you consider a higher level), a new type of
computation is involved, not simply a repetition of the kind of computation
at the lower level. The process of deriving categories from sets of
relationships can't be carried out by the same kind of computation that
derives relationships from sets of events or lower perceptions. There is no
one kind of computation that could serve at all levels.
But as I say, I am, we all are, a very long way from grasping what these
kinds of computations are. Every time people come up with a new computer
program for recognizing objects, they try to establish this new computation
as the blueprint for the whole perceptual system. This is a waste of time.
The blueprint changes with every level. Weighted algebraic summation is
simply not going to suffice to model our capacity to recognize and execute
a program described in words: a rule. Even though such networks are
purported to recognize categories, I think that the categoryness is read
into the results by a human observer. I don't think that any category-
recognizing back-propagation model will actually create what human beings
experience as categories -- for example, the category "wife." Of the eleven
levels of perception in my model, I think we know how model two of them,
the first two. All the rest of our modeling presents to us what a human
being might recognize as a higher-level perception, but which the circuit
or program itself does not recognize -- or control.
In that I could be wrong, of course, because I speak the truth when I say I
don't know how the higher levels of perception work. That means I don't
know how they don't work, too. I'm just expressing a hunch.
It's late and I've posted this so I could get to sleep (some ideas just
have to leak out through the fingers before they'll let you alone). I'll
get to comments on other interesting mail tomorrow.
···
--------------------------------------------------------------------
Best,
Bill P.
