[Martin Taylor 920701 0240] Today is Canada Day, the 125th and possibly last.
The following posting has been gestating for about 4 weeks. I hope it is
intelligible.
Here is a very long statement (over 450 lines) of something that I think
is an important consequence of Perceptual Control Theory, in a direction that
has not been discussed, so far as I remember, in the last 18 months of CSG-L,
or perhaps ever.
It deals with parallel and serial processing, similarity and difference
detector systems, alerting systems and situation awareness, and finishes with
a speculation about hyperlexia and autism. If you want to read it at all, you
may prefer to print it out rather than to read it off the screen.
Warning: the latter part of this note is hand-waving speculation. But the
earlier part represents what I think is necessarily true, given the basic
premises of PCT.
Martin
The following material is copyright M.M.Taylor, 1992.
Permission is granted for quoting within the mailing list CSG-L, and for
use in Closed Loop and other publications of CSG, provided credit is given.
ยทยทยท
----------------
Summary
The sensory systems provide many orders of magnitude more degrees of
freedom for input than the skeletal system permits for output. Some reduction
of the input degrees of freedom can be achieved by exploiting the natural
redundancy of the world, but there remain many times more possible degrees of
freedom for perception than there are for action. Hence, not all percepts
that can be controlled are being controlled at any given moment. This
discrepancy leads to the conclusion that there exist Elementary Control
Systems (ECSs) whose perceptual input functions determine percepts that are
not being controlled. These ECSs may be passive (only passing on the result
of their perceptual input function) or monitoring (controlling, but through
the imagination loop rather than through the real world).
Maintenance of percepts in desired stated is accomplished through the
operation of active ECSs, but a monitoring ECS can determine when its percept
is departing too far from its reference levels. If it is to achieve control,
restoring its error to a tolerable level, it must wrest control from some
active ECS, relegating that ECS to a passive or monitoring role.
Alternatively, it can emit an alerting signal that causes a sibling ECS with
less tolerance for error to take control[1]. A monitoring ECS contributes to
"situation awareness," a previously elusive concept that makes sense in the
context of PCT.
The shift of status of ECSs among monitoring, passive, and active states
demands some kind of switching, either within an ECS (changing its gain
function, for example) or in some separate system that can move control around
within the hierarchy.
A monitoring ECS requires tolerance for amounts of error that would cause
an active ECS to emit a substantial output signal. Likewise, a passive
template-based alerting detector may exist, which brings some part of the
external environment under control (shifting attention is one way of putting
it). Such a detector would also be expected to have tolerance for error.
Both seem to perform much as the human "similarity detection" system--
parallel, fast, and tolerant of error, in contrast to the "difference
detection" system which is slow, accurate, and seems to correspond with what
one would expect of an active ECS.
The differences between monitoring and active ECSs seem to provide a
natural reason for the two "tracks" of processing that I postulated in 1983 to
account for the results of experiments in reading. In the process, they
plausibly account for the relation between a pathology called "hyperlexia",
and autism. Studies of overt control seem unlikely to be able to penetrate
far into the functions of perception, since at any one time, most of
perception is not under control.
---------------
Introduction
ln 1983 I published something I called the BLC (Bilateral Cooperative)
theory of reading [2]. At the time, it seemed justified by a host of
surrounding data, but lacked an evolutionary rationale. I now think that PCT
gives it one, and in the process relates many concepts, including alerting
functions, situation awareness, autism, and the role of attentional focus.
The argument follows from the degrees-of-freedom (DOF) factors that I have
mentioned on CSG-L from time to time over the last few months.
Background on perception in reading:
Logically, the concepts of similarity and of difference seem related, in
that one is the polar opposite of the other. If two things are more similar,
they are less different. Psychologically, that seems not to be so. In many
respects, judgments of similarity act differently from judgments of
difference, to the extent that they seem to produced by quite independent
processes, and indeed to be preferentially processed by different brain
hemispheres (I'm not up on the recent literature in this area, but this was
the way it seemed in 1983, and I doubt that new data would much change this
general statement).
Similarity processes seem to work fast, in parallel, and possibly
unconsciously, whereas difference processes tend to be slower, serial, and
attentive. Similarity processes tend to work on whole entities, whereas
difference processes analyze the components of the entities. Where exactness
of identification is required, the difference processes dominate ("Is this
exactly an X?"), whereas if plausibility is more useful, the similarity
processes work faster and may be used without the slower difference processes
("Might this be an X?"). Similarity processes can give multiple answers
("This is like an X, like a Y, and like a V"), whereas the difference
processes can give only one answer in most situations ("This is not an X or a
V, but is a Y").
The BLC model of reading postulates a layered set of levels of abstraction
for the perception of elements of text, ranging from the visual features of
letters up to the concepts inherent in the material. At every level of
abstraction, parallel similarity processes and serial difference processes are
both available for the construction of the next level of abstraction. The
degree to which each is used at any level depends on several factors,
including the skill of the reader, the familiarity of the material, the
importance of exactness, and so forth. A highly skilled reader skimming
familiar material would tend to use largely the similarity processes at all
levels, to form very quick impressions of the conceptual content of the
material, whereas a poor reader might work through the same material syllable
by syllable using the exacting difference processes to avoid error, but doing
it very slowly. The skilled reader might also use the difference processes if
reading for copy-editing, or as an editorial critic looking for logical flaws
or rhetorical imprecision.
There is an interesting case of a French speed-reader who, after a stroke,
lost the ability to read slowly, but could still speed-read [3]. At a very
minimum, this case shows that the ability to analyze words slowly is not a
prerequisite for reading them fast.
PCT background--degrees of freedom:
By my very rough count, the joints of the human body, together with such
shape-changing functions such as facial expression and voice, admit around 125
degrees of freedom. There are probably far fewer, inasmuch as it is hard, for
example, to flex the top finger joint while keeping the others straight, but I
want to make this number as high as is reasonable. This number, 125, is the
largest number of degrees of freedom that can be used in total by ANY set of
ECSs within a layer of the hierarchy to control their actual perceptions from
the environment. More ECSs in a layer may be involved in control, but then
their perceptions will not be mutually independent[4]. Of course, the degrees
of freedom controlled by the ECSs in one layer depend completely on those
controlled by ECSs in the layer before, so the number of controlled degrees of
freedom cannot increase (but may decrease) as we go up the layers, and the
total number of independently controlled degrees of freedom overall cannot
exceed about 125.
Again by a very rough count, the degrees of freedom for sensory input can
be estimated as follows: about a million optic nerve fibres, about 60,000
auditory nerve fibres, many (I have no idea how many) touch and pain sensors
on the skin, and quite a few taste and smell sensors. The actual numbers are
not important; the point is that there are far more than there are degrees of
freedom for affecting the external environment, by a factor of many thousand.
Each incoming sensory nerve fibre IN PRINCIPLE (though not in practice)
provides a degree of freedom for sensation. Let's take a low number for how
many, and say 2.5 million, or about 20,000 times as many as there are
controllable degrees of freedom. How can ECSs with 125 degrees of freedom for
output control perceptions with 2.5 million degrees of freedom? The answer is
that they can't, not all at the same time; but over time, the system could
change which perceptual degrees of freedom are being controlled, so that any
of them is potentially the subject of control.
To compound the problem, the intensity values of the sensory degrees of
freedom can vary faster, usually much faster, than can the angles of the
joints. It would probably be fair to put a limit of 1 to 2 Hz (cycles per
second) for the average rate at which the joints can be oscillated, although
some can go faster. The visual, touch, and auditory fibres can all vary their
outputs much faster, though (to pre-empt myself) some of their speed comes
from their working in a coordinated fashion. It would probably be
conservative to allow the sensory fibres an average bandwidth of 10 Hz. Given
these numbers, and assuming that the control of the joints can be no more
precise than that of the sensory inputs, it is not unreasonable to think that
the incoming information flow is in principle capable of reaching a rate about
5 orders of magnitude greater than can be controlled.
In practice, the world is not so unkind as to provide us with information
that changes character in every part of the visual field 10 times per second.
There is a great deal of redundancy. A chair remains a chair; neighbouring
points tend to have much the same brighness and colour as each other and as
they themselves did a second or two previously. Only at edges and when new
objects come into view do the intensities change rapidly in space or time.
But it is unreasonable to think that only one part in 100,000 of the incoming
information is useful and new. Even if the redundancy is 99.9%, there is
still 100 times as much information coming in as can be controlled for. And
99.9% seems very high. Regardless of the numbers, it is clear from immediate
experience that there is a lot of sensory input for which we are not at any
particular moment controlling. The point of the numbers is to suggest how
very much uncontrolled perception there may be.
For two reasons, we can argue that there is no sensory input for which we
inherently cannot control. The first is that it would violate the efficiency
we expect of an evolved organism. Just as behaviour is the control of
perception, so perception is the mechanism whereby we can act so as to survive
long enough to propagate our genes. These actions control the relevant
perception. Any perceptual capability that does not support this objective
(cannot be controlled) will be unhelpful baggage, in an evolutionary sense,
and will be selected against if it entails any cost.
The second argument comes from the theoretical arguments and experiments
of J.G.Taylor, who showed at least how much more readily perceptual abilities
are developed if they are the subject of control than if the corresponding
sensory input is passively observed and identified by a teacher. His theory
proposes that there is NO perceptual capability unless it can form part of a
control loop, and this agrees with the evolutionary argument.
If there is no perceptual function that cannot be the subject of control,
and if at any moment only a tiny proportion of the sensory degrees of freedom
are under control, then (1) there must be some way of changing which
perceptual degrees of freedom are under control from moment to moment, and (2)
there must be some way that the living organism can detect which perceptual
degrees of freedom should profitably be controlled at any moment. These two
requirements are fundamental to the theoretical argument, and all of the
foregoing has been devoted to showing that they almost necessarily are real
requirements on a living system of the complexity of a human, not options.
ECSs that are not at a particular moment controlling their perceptions
through the environment might nevertheless be controlling them through
imagination, so they are not necessarily inactive as controllers. We can call
them "impotent" or "monitoring" ECSs, since their operation does not affect
any CEV in the real world at that moment. Other ECSs may not be controlling
at all, having their output gain set to zero, while nevertheless continuing
their perceptual function of abstracting from the incoming sensory data and
passing the abstraction on to higher level ECSs. We can call these ECSs
"passive." (Still others might be turned off completely, but we can ignore
them in this discussion; all of the ECSs of interest have their perceptual
input functions in operating normally).
Monitoring or passive ECSs can be supplied with reference signals, just as
if they were actively controlling, and therefore can develop error signals
that lead to outputs that provide references for lower ECSs. The only thing
they cannot do is to have their outputs actively affect a CEV outside the
person. If they control anything, it is through the imagination loops of ECSs
in their part of the hierarchy. Through the imagination loops, they could
control their prediction, or plan, for what they may perceive, but they cannot
control what they actually perceive from sensory data. Most ECSs must be
monitoring or passive at any given moment, but there is no reason to believe
that their perceptual functions are switched off.
To recapitulate, there are, at any moment, three kinds of interesting
condition in which an ECS might find itself: (1) "active," controlling its
perception through a loop that uses muscular function; (2) "impotent" or
"monitoring," controlling its prediction for incoming perception through loops
which are completed only through imagination; (3) "passive," not controlling
for anything, but nevertheless performing its perceptual function within the
hierarchy.
Situation Awareness
"Situation awareness" is a nebulous concept that has often been associated
with workload assessment. A naive reading of the words suggests that there is
some kind of conscious awareness of the state of a complex environment. In
PCT terms, consciousness has no explicit place, and, I believe, it should not
have a place in assessing situation awareness. In everyday life, one is
constantly acting within a complex and changing environment without being
aware of acting appropriately in respect of disturbances in that environment.
Operationally, one is "aware" of the disturbances without being conscious of
them.
From the degrees of freedom argument, only a small proportion of the
percepts based on the environment are being controlled, but it is possible
that many more are being monitored through imagination-based control. Within
HPCT, imagination-based control is like real control, except that at some
level of the hierarchy the output is connected back to the perceptual input by
an "imagination loop" rather than by the action of subordinate ECSs that
eventually act through the real world. Hence, imagination works much like
real control, with two important exceptions: (1) lower-level conflicts may not
occur, permitting the simultaneous satisfaction of references that could not
be simultaneously satisfied by control of the real world [5], and (2)
imagination-based control can work very fast, not being constrained by the
dynamics of objects in the real world. Imagination therefore can perform a
predictive function, projecting what percepts would be obtained if the control
were effective.
It is tautological that active ECSs (those really controlling percepts
from the environment) are aware of the situation in respect of the percept
that they control (if one removes the concept of consciousness from the
connotations of "aware"). It also seems reasonable to suppose that monitoring
ECSs are aware, in that they are shadowing the control of their percepts. The
output of a monitoring ECS has no effect on the percept it receives, but the
ECS can continually assess what effect it would have if it did take real
control. A monitoring ECS can become passive simply by reducing its output
gain to a negligible level, but it cannot become active without denying
control to some other ECS. Passive ECSs (not providing any output to lower
levels) do not need to make any use of their perceptual input such as to
present it to the comparator, and cannot necessarily be considered to be
situationally aware.
Of course, to identify situation awareness as being associated with
monitoring rather than passive ECSs is speculation, but it is speculation with
some justification. A monitoring ECS is controlling through imagination, and
thus is prepared to take active control if necessary, without the introduction
of an initialization transient. This ability seems to correspond with the
notion of situation awareness, according to which the individual can react to
the exigencies of a situation without the need for an initial period of
updating the perception of that situation. According to this view, situation
awareness is connected not so much with perception, but with the state of
perceptual control implicit in monitoring as opposed to passive operation.
Alerting and parallel/sequential function:
ECSs can change state. A monitoring ECS can become an active controller,
but if all the available output degrees of freedom had been used, some other
ECS must change from the active state to being either a monitor or passive
[but see again note 4]. Under what conditions should this happen?
One low-level example of an induced shift of control can be found in the
motion-detecting system of the visual periphery. The central part of the
visual field, imaged in the part of the retina called the fovea, is used for
detailed, coloured vision. The periphery, on the other hand, has relatively
poor resolution for detail, and the further out one goes, the less colour
sensitivity exists. What does exist is a motion detection system that tends
to lead the eyes to fixate on any location at which an unexpected movement has
occurred. The movement is not an uncontrolled S-R event, since it can be
suppressed, but in the absence of such suppression, it tends to be an
automatic response as if it were a simple S-R event. The effect is to allow
higher-level ECSs access to detailed information about the moving object that
would allow them to determine whether there exists a conflict with any
relevant reference value. The perpiheral motion detectors say "there might be
something important here" and the detailed examination permits other ECSs to
determine whether there is.
The peripheral motion detector system can be said to provide an alerting
signal that (usually) leads to a change in the perceptual degrees of freedom
that are under control, by changing both the portion of the environment from
which detailed visual sensation is received and the variables in the
environment that are the subject of attention. They induce a shift of
attentional focus.
Two almost opposite conditions seem appropriate for the issuance of an
alerting signal: (1) the perception coming into a monitoring ECS deviates
sufficiently either from its imagination-driven prediction or from its
reference, or (2) the perception in some passive ECS (or equivalently in a
specialized pattern-matching system that is not an ECS) matches sufficiently
closely some predetermined template that signals the need for control. A
mother's awakening to the cry of her baby in a noisy environment might be an
example of the latter.
Both proposed conditions for alerting signals depend on two things: rapid
parallel operation of many possible alerting entities, and a loose tolerance
for the identity of a perceived state with some reference. In the case of a
monitoring ECS, the alerting state should occur only when the actual percept
departs too far from the predicted percept (or perhaps when either departs too
far from the reference signal in the ECS). In the case of the template-driven
alerting signal, the fact that the relevant percept is not being controlled
means that it will be highly variable and thus should be tested with a wide
tolerance for error. Each of these conditions employ the concept of
similarity rather than of identity.
I have been talking of "alerting signals" as if they were real signals
that are sent from the ECS that detects the problem to some entity that has
the responsibility of switching control among the myriad perceptual degrees of
freedom. But this is not really necessary, especially in the case of
monitoring ECSs. If an ECS has a gain function that stiffens with increasing
error, then a monitoring ECS with a sufficiently high error could wrest
control from some unknown other. If this generated high error in other ECSs,
they also would wrest control, probably in some other direction, and
eventually the system would stabilize in a condition in which all the ECSs
were working in a relatively low-error regime. Such a system could not,
however, maintain any of its percepts in a zero error condition, because it
requires that there be a tolerance zone for error around which the dynamic
gain is near zero. Without such a tolerance zone, all the monitoring and
active ECSs would be trying to control, which they could not do because of the
insufficiency of output degrees of freedom. There would necessarily be a
great deal of conflict among them. Although there might be stability, that
stability would not be achieved with zero error levels.
The conclusion seems inescapable (usually a warning sign!) that the system
must incorporate some kind of switching mechanism. Either (1) an ECS can
switch between high and low (zero?) gain mode, or (2) there are at least two
types of ECS in sibling relationships (tolerant ones with zero gain near zero
error and insistent ones with high gain near zero error) in which signals from
one sibling can cause the other to assert control, or (3) an ECS can change
its perceptual function from moment to moment, or (4) there are separate
subsystems for alerting and for control, the controlling subsystem being able
to reconfigure itself in response to signals from the alerting subsystem.
I am sure that there are other possibilities, but I see no escape from
some kind of switching mechanism. A massively parallel set of passive or
monitoring elements must in some way direct the operation of a set of active
elements that among themselves are capable of controlling sequentially all the
degrees of perceptual freedom monitored by the parallel elements.
Of the four listed possibilities, (2) and (4) are mutually compatible.
There could well be two subsytems, one consisting of tolerant ECSs that
normally act as a parallel monitoring system, the other consisting of
insistent sibling ECSs most of which are inactive at any particular moment.
This is very close to the BLC hypothesis, except that the BLC model was posed
in terms of a two-way flow of expectations and recognitions, in which the
RIGHT track consisted of a multilevel set of parallel, tolerant, passive
recognition units that provided goals for the LEFT track's sequential,
exacting, slow units, and the LEFT track selected for correctness, pruning the
RIGHT track's efflorescence. In both tracks, higher levels provided
expectations to the lower, while lower levels provided results to the higher.
Hyperlexia and autism
The BLC model predicted various kinds of dyslexia and failure of
understanding as depending on the failure of one or other of the tracks at
some level of abstraction. One book reviewer pooh-poohed the whole idea,
dismissing it with the comment that if the BLC model were true, one would
expect there to exist people with right-hemisphere damage who would fail to
see jokes, and would take the world literally. Such people do exist, with
exactly the symptoms that the reviewer thought were so silly as to scuttle the
model entirely.
Hyperlexia is one pathology that was not predicted by the BLC model, but
that is consistent with it. A hyperlectic child is usually more able than a
normal reader to extract words out of fragmentary or noisy representations,
but is usually very poor at connecting the words identified into coherent
structures, or to extract meaning from connected text. In [2], we speculated
that the hyperlectic child had devoted too much of his or her resources to the
RIGHT track, thus impoverishing the LEFT. We also noted that the literature
suggested a high probability that a hyperlectic child would be autistic. It
now seems that the PCT approach, using the arguments persented above, accounts
at least plausibly for this connection.
In a pre-PCT evolutionary argument supporting the BLC model, I suggested
that the need for a LEFT track derives from a requirement to choose between
conflicting overt actions, whereas the RIGHT track permitted a wide-ranging
appreciation of the state of the world, having only a slender connection with
current overt action. One of the jobs of the RIGHT track was to alert the
LEFT track to possibilities for action, among which the LEFT track could make
precise judgments and initiate the overt actions. These characteristics now
seem to be those of active ECSs in the LEFT track, and monitoring (or passive)
ECSs in the RIGHT.
If we were correct that the hyperlexic child devoted too many resources to
the RIGHT track at the expense of the LEFT, it would now seem to follow that
the child would have few active ECSs, and would not control many of its
percepts. It would passively observe the world much of the time, until jolted
by necessity (or perhaps by chance) into producing some overt action. It
would be situationally aware, but not visibly active.
Of course, a child that interacted very little with the world would have
little chance to develop effective ECSs, so the situation would feed upon
itself. The imbalance between active and monitoring ECSs would tend to grow.
The child has "learned" that active control does not work very well, and
devotes resources instead to monitoring, so that control can be exerted in
those finely determined conditions where it is really necessary. But there
would be very little development of subtle discriminations, the learning of
which would normally be based on continued interaction with the world,
interaction not experienced by the autistic child.
Comment
Once again, PCT ties together many threads that were disparate, though
possibly seeming to be connected for non-obvious reasons. I find that this
argument makes the BLC model natural and plausible, whereas before it was just
a description that seemed to fit the data of experiments on reading and
related studies. The concept of "situation awareness" becomes less vague,
though remaining hard to experiment with. Studies of overt control seem
unlikely to be able to penetrate far into the functions of perception, since
at any one time, most of perception is not under control.
------------
[1] A "sibling" ECS means any ECS or set of ECSs that can control much the
same percept as the one in question. The notion is meant to take into account
the idea that there may be separate but related ECSs for monitoring and for
controlling actively a particular percept.
[2] Taylor, I. and Taylor M. M. The Psychology of Reading. Academic
Press, 1983. (The BLC material is mainly concentrated in Chapter 11, with
related experimental data in Chapter 10)
[3] Andreewsky, E. DeLoche, G. and Kossanyi, P. Analogies between speed-
reading and deep dyslexia: Toward a procedural understanding of reading. In
M. Coltheart, K. E. Patterson and J. C. Marshall, (Eds.) Deep Dyslexia.
London: Routledge and Kegan Paul, 1980.
[4] It is simplistic to say that one degree of freedom is controlled by
one ECS, for various reasons. Nevertheless it is convenient in the following
discussion to maintain this fiction, which does not affect the argument.
[5] Which may explain why people sometimes prefer to "live in a fantasy."
Martin Taylor
DCIEM, Box 2000, North York, Ontario, Canada M3M 3B9
(416) 635 2048