To John Flach and the PCT modeling interest group [changed to CSGnet by the
time I got to the end]:
We're coming down to the nitty-gritty here, with a paper by John Flach et.
al. about collisions. If John wants to cc a copy to this reply-all list,
that would help, but I will quote enough to get the points across, I hope.
Let's start with the paper's opening quotation from Gibson, which I include
in its entirety:
Approach to a solid surface is specified by a centrifugal flow of the
texture of the optic array. Approach to an object is specified by a
magnification of the closed contour in the array corresponding to the
edges of the object. A uniform rate of approach is accompanied by an
accelerated rate of magnification. At the theoretical point where the eye
touches the object the latter will intercept a visual angle of 180�; the
magnification reaches an explosive rate in the last moments before
contact. The accelerated expansion in the field of view specifies imminent
collision, and it is unquestionably an effective stimulus for behavior in
animals with well developed visual systems�.the fact is that animals need
to make contact without collision with many solid objects of their
environment: food objects, sex objects, and the landing surfaces on which
insects and birds alight (not to mention helicopter pilots). Locomotor
action must be balanced between approach and aversion. The governing
stimulation must be a balance between flow and non-flow of the optic
array. The formula is as follows: contact without collision is achieved by
so moving as to cancel the centrifugal flow of the optic array at the
moment when the contour of the object or the texture surface reaches that
angular magnification at which contact is made. Gibson (1958/1982,
p. 155 156)
One point of interest here, made more obvious in other writings by Gibson,
is that he had the idea of controlling perceptions (or at least inputs)
quite clearly. Collision avoidance, he says, is a problem of controlling
the flow of the optic array, not of producing any specific behavioral
action. One problem is that he mixed this view with the old
stimulus-response view.
Consider the sentence above, "The accelerated expansion in the field of
view specifies imminent collision, and it is unquestionably an effective
stimulus for behavior in animals with well developed visual systems." This
is, of course, the view against which PCT is arrayed: the idea that there
are stimuli which cause behavior. Note the assumption that the observer's
interpretation is an objective fact. I don't dispute the interpretation,
but I do dispute saying that it is an unquestionable aspect of objective
reality rather than a perception constructed by the brain of the observer.
Of course I also question the implication that the accelerated expansion
of the visual field is enough to cause some sort of behavior in the manner
of stimulus and response, with no need for a perceptual function to convert
an expansion of a retinal image to a perception of approach, or a need for
a reference signal to specify the desired degree, or maximum desired
degree, of approach.
Consider also, "The governing stimulation must be a balance between flow
and non-flow of the optic array." I don't dispute the theory that there is
an optical image on the retina, and that both the human observer and the
subject of observation will perceive a flow (which at one end of the scale
includes non-flow) of the image. That merely says that all human beings are
built alike. But there is a claim that this stimulation somehow "governs,"
which is a theoretical assumption contrary to control theory, and contrary
even to Gibson's idea of behavioral output controlling sensory input.
It's very clear that Gibson accepted his own private view of the world as
if it were a correct and objective picture. "Approach to a solid surface"
assumes the existence of an objective solid surface and an objective
dynamic relationship called approach. Elsewhere he has talked about
"tangible" objects, as if the sense of touch were somehow more objective
than other senses. "Centrifugal flow" assumes movement of image elements
away from a center, "the texture of the optic array" assumes something
called "texture," not to mention an "array." "Approach to an object"
assumes a changing distance relationship, "magnification of the closed
contour in the array" assumes a size relationship between the optic array
and the object. and "corresponding to the edges of the object" implies some
way of knowing about the object without depending on the optic array. In
short, Gibson takes the same view taken by any engineer, in which the role
of human perception in producing the world of experience is simply ignored
-- the next thing to naive realism. This is a perfectly useful point of
view for engineering, but it leads in the long run to a model of human
organization that is incompatible with our other models of reality.
In PCT terms, the problem here is that of exempting certain perceptions
from the general rule that what the brain knows (that is, what we know)
consists of perceptual signals. Certain aspects of the world, such as
"solid surfaces", are simply taken for granted by Gibson as existing
independently of the observer, and do not themselves apparently require any
explanation. I repeat that the issue here is not the veridicality of
perceptions; that remains an open question with answers that vary with
circumstances. The issue is that of modeling an organism which has no view
of the world independent of its perceptions -- and that includes organisms
like you, and me, and Gibson, and Flach.
I have started out discussing Gibson in the hope of directing my criticisms
toward a target outside the present circle and thus perhaps postponing any
direct confrontations. But now let us get into John Flach's paper.
The article begins in a promising way:
2. The comparator Problem
In a control system, the comparator is a junction with reference and
feedback signals coming in and error signals coming out. In a simple
system, such as the servomechanism illustrated in Figure 1, the comparator
is analogous to the simple mathematical operation of subtraction. The
feedback signal (specifying the current state of the system) is subtracted
from the reference signal (specifying the desired state of the system) in
order to get an error signal (the deviation from the goal). The error
signal then drives action in the direction that will reduce the difference
(bring the system state closer to the goal state).
There is nothing here to raise an eyebrow among aficionados of PCT. The
eyebrows will go up, however, at the next statement:
In engineering control systems an important step (that is critical in
practice, but rarely explicitly acknowledged) is to convert the various
signals (reference, feedback, and error) into a common (comparable)
medium (e.g., electrical current). Once this is accomplished, the
operation of the comparator is directly analogous to the simple
mathematical operation of subtraction. Figure 1: How is it possible for
a biological system to compare perceptual feedback to intentions in order
to specify appropriate corrective actions?
I will venture a guess that most PCTers will recognize the stated problem
as a non-problem. But this is because PCTers do not automatically assume
that there is an objective optic array in which the "signals" mentioned
above exist _outside of the brain_. If those signals exist outside of the
brain, then there is surely a comparator problem, and the question about
making comparisons is quite appropriate. But if the state of the system is
represented as a perceptual signal inside the brain, and the intended state
is also a signal inside the brain, then there is no difficulty at all with
converting these signals into a common medium: they are already in a common
medium, and subtraction poses no difficulties.
We have seen this problem in many forms. When I was first learning about
control systems, I fell prey to it, too. The problem is that while we can
see how an external variable like temperature can be represented by an
electrical signal of varying magnitude, it is very hard to understand how
the temperature could be compared with a _desired_ temperature, which
exists in a different place, inside somebody's head. The temptation is
strong to look for something else in the environment that will indicate the
desired temperature -- for example, if you are too warm you may see or feel
sweat on your skin. That labor of looking for an external explanation of
the reference condition is, as PCTers will recognize, exactly the wrong
strategy.
The difficulty arises from taking the external observer's point of view
instead of that of the control system. To the external observer, it seems
that the variable to be controlled is over _here_ in the environment, while
the desired state of that variable is specified over _there_ inside the
controlling person. But as soon as you move your point of view to the
interior of the control system, it becomes clear that the state of the
variable to be controlled is first known to the controller in the form of a
signal, and comparison of that signal to another one specifying the desired
state (also inside the controller) becomes a trivial problem, a
non-problem. As soon as we recognize that to any control system, the state
of the external world IS the state of its perceptual signals, the
comparison problem disappears.
At this point a distracting red herring gets dragged across the path:
Classically, this has been thought to require translation into a
symbolic neural representation. The idea of direct perception suggests
that lawful relations in perceptual arrays may support an indexical
coding in the nervous system that can be fully described in terms of the
perceptual referents.
In a way this is a valid and illuminating observation, but in another way
it discards a baby with the bathwater. As we know, many perceptions are
continuously variable, the magnitude (frequency) of a neural signal
indicating the magnitude of some (supposed) external variable. Perhaps this
is what Flach means by "indexical," though I don't really know. For this
class of perceptions, treating them as "events" or "states" to be
represented by discrete symbols is simply a mistake. To explain behaviors
like tracking, we need signals that vary on a continuum.
On the other hand, there are obviously perceptions that are in fact
symbolic, and which vary in an either-or way. What I am doing is called
"tracking," or it's not called "tracking," with no states between. If I
tell you to bring me "a sandwich," you can translate this verbal symbol
into an appropriate perception of an object we would both agree to call a
"sandwich," and then you can set the appropriate reference signals that
will result in our both perceiving that you have brought me a sandwich. We
can't ignore this kind of discrete control process, because such control
processes are the origin of most of the analog reference signals at lower
levels of organization. However, this whole subject brings up the concept
of levels of control, and that runs counter to the Gibsoniam premise that
says the environment sets our reference signals.
Now the confusion begins in earnest:
However, for animals, the three signals associated with the comparator
rarely come nicely packaged in a common medium or currency that would allow
simple subtraction of one from the other to produce the third. For
example, the
reference or intention may be to get to a meeting across town as quickly as
possible without collision. The information may be patterns in an optical
flow field. What does it mean to subtract the patterns from the intention?
The difference from this subtraction would have to specify the actions of
muscles (perhaps on control devices - steering wheel, accelerator, and
brake pedal). The natural units for each of these "signals" converging at
the comparator are different - desire not to be late for an important
meeting and to avoid collisions, a transforming pattern of texture
transduced through a retina, and a force or motion of a limb perhaps
transduced through a vehicle. How does an animal translate from one medium
to another in order to behave appropriately that is, in order to behave
in a way so that errors from intentions are kept within acceptable limits.
The idea of a multiordinate control organization with many systems at each
level controlling in one degree of freedom apiece solves these problems
automatically. First, the three signals are always in the common currency
because they are all neural signals to begin with. Second, the reference
signals do not have to be found in the external world, the optic array,
because they are specified by higher levels of organization in the brain
and do not _ever_ enter as sensory signals. And third, a complex goal such
as "get to a meeting across town as quickly as possible without collision"
breaks down into a set of parallel control processes each concerned with
one, or only a very few, degrees of freedom. Get to the meeting. Move
rapidly. Avoid collisions. Each of these control processes can be carried
out with reference signals set independently of the others. Each entails a
simpler perception that can be comparied with a reference signal set by a
higher system, such as a system that has the goal of not arriving early and
does this by setting the reference-speed for the trip to a lower value than
normal. The systems that want to avoid collision and want to get to the
meeting work as before.
Now the red herring returns:
Psychology has conventionally assumed that the comparator problem was
solved "in the head." That is, the general notion was that the three
signals (intention, feedback/perception, and error/motor command) were
converted to some common symbolic neural code (reflected in terms like
program, schema, mental map, mental model, gestalt, etc.). Thus, the
neural symbols associated with perceptions could be "compared" with the
neural symbols associated with intentions in a way that would specify the
appropriate neural symbols to guide actions..
If we put aside the difference between discrete and analog "codes", this is
exactly the PCT view (though I doubt it is the prevalent or conventional
"psychological" view). Perceptual functions convert physical stimulations
into perceptual signals, level upon level, with the lower levels behaving
as analog signals and the higher ones as discrete symbolic signals. This
provides a straightforward physical explanation of how living control
systems work, without filling in the details of course, but presenting a
framework within which we can hope to fill in the details.
But now Flach et. al. offer an alternative:
Gibson, however, suggested an alternative position. The radical notion of
"direct perception" suggests that the comparator problem can be solved in
the light. Gibson (1958/1982) wrote:
To begin locomotion, therefore, is to contract the muscles as to make the
forward optic array flow outward. To stop locomotion is to make the flow
cease. To reverse locomotion is to make it flow inward. To speed up
locomotion is to make the rate of flow increase and to slow down is to
make it decrease. An animal who is behaving in these ways is optically
stimulated in the corresponding ways, or, equally, an animal who so acts
as to obtain these kinds of optical stimulation is behaving in the
corresponding ways (p. 155).
If this is an alternative to the view expressed just before (it could be
taken as a simple restatement of it), the implication is that the "optic
array" is something existing outside the nervous system, presumably on the
retina. But if this is the intention, we are faced with something existing
on the retina that is not part of our physical theories about the world. We
know that according to simple physics, what exists on the retina is a field
of variable light intensities with different wavelengths. This field varies
with time, but it is never anything but a field of intensities and
wavelengths. More specifically, it contains no patterns, no objects, no
motion or flow, no sizes, no color, no boundaries, no relationships, no
events. Those are all impressions that arise after the initial neural
signals pass inward to higher structures in the brain. In short, there is
nothing "in the light" that can do any of the things Gibson, and now Flach,
want to be done there. Ask any physicist.
The above statement by Gibson can be converted to an equivalen PCT
statement by making one very small -- but profound -- change. All we have
to do is recognize that the "optic array" consists of the outputs of neural
perceptual functions specialized to produce signals indicating such things
as intensity, color, shape, motion, and position (and more) -- the
dimensions of experience that we find familiar (of course). Optical flow is
a neural signal. Size is a neural signal. Increase and decrease are neural
signals. Animals that behave as Gibson says are acting to obtain particular
neural signals in their brains by acting on the world and their
relationship to it to alter the optical inputs to the hierarchy of neural
computations making up many visual input functions at each of many levels.
Under this view there is simply no "comparator problem." Experience occurs
from the very start in the "common currency" of neural signals. Reference
signals do not enter the organism from outside it; they are specified for
any one level by systems at a higher level of organization. The highest
levels of reference signals come from heredity or experiment or memory --
but no reference signals ever come from outside the nervous system.
Now the denoument: Flach at. al. say
The implication of Powers' description of learning to drive and his
choice of the term "perceptual" control theory suggests that the currency
exchange that allows feedback to inform action relative to intentions is
typically
negotiated in the medium of perception. An important point of the PCT
approach is that the variable that drives a control system (e.g., a
thermostat) is
not the "output" variable per se (e.g., the actual room temperature), but the
measured or perceived temperature. Ideally, in a designed control system
there
would be a close correspondence between the output variable and the measured
variable. However, for the control system, the measured temperature is the
only
temperature. Analogously, the animal knows nothing of the world except its
own perceptions. Of course, if those perceptions are not fairly well tuned
to the actual situations in the world, the animal is not likely to survive
for very long. Although Gibson and Powers both focus on perception,
Powers treats perception as an "image . . . in your head," where as Gibson
focused on the structure in an optical array outside the head.
This is an accurate and fair representation of PCT. In defense of his
position, Flach says
An important implication of a closed-loop coupling of perception and
action is that neither perception nor action is causally prior to the
other. Thus, action and perception are locked in a circular causality.
This opens up the possibility that in many cases (not all) that the
problem is solved by putting muscles and intentions into the light. In
Gibson's words from the earlier citation "an animal so acts as to obtain
these kinds of optical stimulation." Muscles can be put into the light by
moving to create an optical flow field. Analogously, the dynamics of a
vehicle can be put into the light by manipulating the controls
transforming the optical flow (e.g., jiggling the steering wheel or
tapping the brakes). Further, we believe that by attending to the
consequences of motion, animals can learn to associate those consequences
with patterns in the optical flow field. This creates the possibility for
intentions to be specified in optical terms.
The phrase "putting muscles and intentions into the light" has about as
much meaning to me as a citation to "1st Corinthians" would. Putting that
obvious piece of ideology aside, we could interpret this paragraph to mean
that it is always possible to express the elements of (visual) control
processes in terms of equivalent states of the visual field on the retina.
And that is true: we have something in PCT called "the test for the
controlled variable" which does exactly that, though variables in other
parts of the external world (as we humans experience it) are also included.
We use physical models to represent the physical world, neural models to
represent the brain, and so on, and we can translate from one model into
the other if we choose and if we have complete enough models.
However, the things that are said to exist in the optical array on the
retina can be known ONLY when they are viewed by a human brain -- the brain
of the person or animal doing the behaving, and the brain of the
observer/analyst who is speaking of information, value, and action. They do
not exist in the physical light distribution itself, as the appropriate
theories of physics would inform us. The optic array and all its attributes
exists only inside the brain that receives the optical intensity signals.
To finish up section 2 of this paper:
It is important to note that the optical states are expected to have
correlated structure in the neural medium, (e.g., weights in a neural
network). However, the critical point of "direct perception" is that the
neural medium does not introduce additional constraints, at least, when we
are dealing with supra- threshold phenomenon which is generally the case
for control of locomotion. In other words, the claim of direct perception
is that the "comparison" of intention with ongoing perception to specify
action can be described in terms of lawful relations (e.g., physical laws)
that exist independent of any symbolic neural or cognitive process.
From the PCT point of view, this puts the cart squarely in front of the
horse. What we know as human beings is not the optical states, but the
perceptual signals: the optical states are theoretical and hypothetical.
The primary information we have about the world is in the form of neural
signals that we experience directly. In fact, even the idea of neural
signals is theoretical in this context. We begin with direct experience,
and only after considerable effort have we managed to devise physical and
neurololgical models that tell us what it is and where it comes from.
As the radical constructivist Ernst von Glasersfeld once said, "The brain
is not the black box; the environment is." The brain, PCT proposes,
receives physical effects from the environment, convertiong them to neural
signals which represent all that the brain can know about its surroundings.
All else is constructed by the brain, in the attempt to find and take
advantage of regularities in the connection between our outputs and our
inputs. Those connections remain completely invisible; they are accessible
only through making and testing hypotheses about the external reality.
···
===================================================================
I think that's far enough. The rest of the paper depends critically on the
assumptions and counterarguments discussed so far, and once these issues
have been settled the rest of the development would follow logically. I
think the lines between PCT and the Gibsonian position are clearly drawn. I
will not venture to serve both as the defense lawyer for PCT and the judge
of the case, so "more, deponent sayeth not."
Come to think of it, I'm changing the cc from the modelers' group to the
CSGnet list. John is welcome to copy this to his colleagues as he sees fit.
The issues here seem clear and sharp to me, and we need as much discussion
as we can get. Somewhere down the line, one point of view must prevail,
since they are radically divergent on most matters other than control
theory itself.
Best,
Bill P.