[From Bill Powers (950502.0910 MDT)]
Martin Taylor (950501 17:10)--
If all functions in the loop are acting simultaneously,
then it follows that at the time any perception is varying, it is
subject to the effects of present-time actions that are going on.
Absolutely NOT so. At the very least there is the light-speed
limitation
That's really reaching for it, Martin. Anyway, you still don't get the
point. At a given instant, a perceptual signal is subject to EFFECTS
from actions at literally the same time it is subject to EFFECTS from
disturbances. There is no spatial gap there to allow you to slip in two
or three nanoseconds of delay (a delay which, considering the dispersion
in time of neural signals taking different paths around the loop, is of
no consequence whatsoever). And even where you can show that,
technically, there are delays, the effects of the delays are very often
totally negligible.
In any practical control system the delays are many, many orders of
magnitude longer. Eyes take time to deliver signals after photons
arrive (Parenthetically, have you ever tried playing table-tennis
by the light of the full moon? It's a very edifying experience in
that regard*). Actions don't move things instantaneously. No, the
effect of present time actions never, ever, form part of present
time perception.
But effects of actions at SOME previous time are ALWAYS present in
present-time perception. I think you actually agree with this.
The primary limitation on the speed of spinal reflexes is not the delay
in these reflex loops (10 msec or less) but the mass of the limb,
inertial interactions with other degrees of freedom, the strength of the
muscles, the internal viscosity of the muscles, and the spring constant
of the muscles. Even in a zero-delay system with infinite signal
bandwidth, there would be a maximum speed of error correction, and a
need for neural rate feedback to prevent instability. These limitations
occur at much lower bandwidths than suggested by the 9 millisecond
transport lag in a spinal control loop.
There is also the problem that we can't define "a" transport lag. A
spinal control loop actually consists of dozens to hundreds of motor
neurons with their associated input, comparison, and output functions.
all working in parallel. Typically, the individual input functions come
into play at different levels of stimulation and saturate at different
signal levels. Furthermore, path lengths are not all exactly the same.
What we see is the composite behavior of all these loops, which as it
happens approximates that of a linear system quite closely as far as the
composite neural signals are concerned.
Since the speed of motor actions is dictated primarily by inertial
effects and spring constants, this limitation applies to all higher
systems that use motor actions (that is, all higher systems). The
primary limitation at any level is the speed with which a change in
output signal can cause a significant change in the affected perceptual
signal; neural delays are relatively insignificant, while environmental
limitations on speed (both integral lags and transport lags) are almost
always the most significant. Higher-level systems deal with aspects of
the environment that have inherently slower rates of change.
As to playing Ping-Pong by moonlight, the primary limitation is not a
transport lag in the visual system, but signal-to-noise ratio and the
integration time-constant needed to smooth the signal. That effect would
remain even if the transport lags were zero. This is an "edifying
experience" only if one analyzes it correctly.
Yes, all functions in the loop are acting simultaneously. The
amalgamated effects of possibly a long period of past action older
than some minimum age is represented in present perception (if the
output is a pure integrator, the effect of ALL past action older
than the minimum age is part of present perception). But that
minimum age is intrinsic to the limitations on the control system.
Martin, it's time you stopped saying this; the mechanical-inertial
limitations are much more important than the signal-delay limitations. I
know that the signal-delay limitations are very important in the
theoretical scheme you are trying to develop, but they are not important
in real behavior. Certainly, it takes appreciable time to come to a
decision about buying or selling a stock, but it takes far more time for
the decision to be changed into a new ownership status. Even if the
decision could be reached instantaneously, there would still be a
transport lag due to the environment (particularly if the decision is
made on a Saturday), and buying-selling behavior would not be much
improved, if at all.
The question here is quantitative: how much difference do the perceptual
lags we know about make in behavior? We have a rough estimate from
modeling tracking behavior. An integral-output model with no delay
accounts for 95% of the behavior; adding the best delay factor to the
model brings us to 97% or so. The behavior of the system with delay is
almost identical to the behavior of the system without delay. The only
reasonable conclusion is that while neural delays do have a detectable
effect, the effect is at the lower limits of our ability to measure it,
and plays no major role in behavior.
You have been considering the effects of delays as if all other
processes in the loop occurred instantly. In applying Nyquist sampling
criteria, you are considering a hypothetical system in which events take
place at discrete and synchronized instants, and in which the ONLY
stability consideration is the gain at the upper sampling frequency. In
such systems, which can be built, everything you say about limitations
to performance holds true. But in real physical organisms there are
mechanical-inertial effects which completely dominate; furthermore,
there is neither any actual sampling nor any actual synchronization of
events.
We have been through this many times, and your response has always been
similar to this:
ME:
In real human control systems, lags are
relatively unimportant because inertial effects and integrations that
would be there anyway take care of most of their deleterious effects.
YOU:
Yes, we animals have evolved in an efficient way in many respects.
We don't waste many resources on actions or processing that would
ordinarily be useless. Evolution is often neat that way, to the
extent that there are still people who insist that its neatness is
evidence for the existence of a personal Designer.
Evolution is only "neat that way" if you assume the truth of the
preceding sentence: that we don't "waste resources." The only reason for
assuming that we don't waste resources is the assumption that evolution,
being so neat, would never let that happen. But I see no evidence that
human organization is optimized in any way: to assume that it is is just
an act of faith or a questionable technique for winning an argument. I
think we waste resources quite happily, and that doing so is inevitable.
"Waste" can be defined only relative to some better usage that we can
imagine, but what we imagine has no influence on how the real system
works.
You think I concentrate too much on loop delay or external
transport lag. I think you slough over it too much, and disregard
its critical nature in the dynamics of control.
Well, there you have our basic disagreement. I don't think I can be
accused of disregarding loop delay, since I was the one who introduced
it to the model and found (with Rick) how much delay there probably is
in a tracking task. I have never disagreed with the formal idea that lag
is critical in a system where it is the primary determinant of
stability. But you have never seemed to realize that in most control
systems, there are other and far more important determinants of
stability, which could not be removed even if lags could be reduced to
zero. The actual bandwidth of behavioral control systems is far less
than one would guess just from looking at transport lags.
As to _external_ lags, they can become important (especially in higher-
level systems), but control systems with lags anywhere in the loop can
often be stabilized by adding a time integration to the loop -- and if
they can't be stabilized, we can't control anyway.
···
---------------------------------------------------------------------
Avery Andrews (950502)--
I think there may be a very simple reason why people are so
gullible: in our evolutionary environment, the fact that Uncle Fred
was still alive was an excellent reason to believe that he knew
what he was talking about, over a broad range of topics - where
there is water to be found, where the antelope go in springtime,
etc. Presumably believing Uncle Fred's junk information was less
costly than being sceptical of the stuff he was actually on target
about, so our dumb genes bias us in favor of believing what we are
told, at least until the teller is clearly revealed as a liar
and/or fool....But in a modern environment, this isn't true at all
Very neatly put, which is normal for you. One new factor that the modern
environment may bring in is that we're more used to hearing about facts
that have scientific or engineering validity -- we expect them to be
100% true essentially all of the time. We expect factual reporting to be
true, too, since it is so much easier to check on what actually happened
(when we try). Back in Uncle Fredericus's day, however, most facts were
found empirically, and nobody expected them to be true 100% of the time.
If Uncle Fred, casting the bones, said that the Visigoths were coming,
and they failed to show up, nobody would be particularly surprised. By
the same token, if Uncle Fred said they were coming, people would
probably not have gone to any great inconveniences to prepare for the
attack unless they also got the news from a breathless messenger on a
dying horse.
It's a truism that people seem to believe things most when they're
printed, especially in books. Of course now we have TV and other visual
electronic media, which can tell the most convincing lies of all
(President Kennedy shaking hands with Forrest Gump). I think we have an
emergency need to teach skepticism in the schools.
-----------------------------------------------------------------------
John E. Anderson (950501.2245 EDT)--
Well, I never said that NSD is a control model. And I have to
confess that, not being a psychologist and so not having been
steeped in the S-R controversy, I'm not even sure if I CARE about
the S-R controversy.
Ah, but all this means is that you've accepted what the S-R theorists
have to say without knowing it.
What I DO care about is helping to understand how the brain works,
how it gives us our mental and physical behaviors.
The critical question is, what are those mental and physical behaviors
that we want to understand? Most of the neuroscience literature seems to
assume that the behaviors are outputs that are created by sensory
inputs, which is the legacy of 100 years of S-R theory and its
popularization. You're not to blame for that, nor are today's
neuroscientists. But that doesn't mean they're on the right track. The
cruel fact is that many people have spend many years of their lives
developing explanations of phenomena that don't actually happen.
NSD is important to me.
I presumed so. It is an unfortunate fact of human nature that we easily
fall in love with theories for their own sake, just because they have a
certain internal beauty. This is especially true when they are our very
own theories, our brain-children. The hardest task of a theorist is to
look the child over for birth defects, or to send the child out into the
mean streets of experimental testing.
However, I don't respect theories that are defended simply by saying
that one has spent a lot of time and effort on them. That doesn't make
them good theories, or right. It just makes them hard to give up when
dispassionate reason and evidence call for doing so.
I am thinking about writing a grant, in collaboration with a
neurologist who is interested in neuroscience theory, asking for
support to see how far a PCT-like hierarchy can be demonstrated in
nervous system anatomy as it is understood now. In fact, I am
going to meet with him tomorrow to talk about this. I think it is
the place to start developing a biology of PCT.
That, of course, gives me the shivers because you might find out that
there is no evidence for PCT in the nervous system. But I've had those
shivers many times in my career, and have always gone ahead to take the
chance. If the theory isn't exposed to potentially lethal challenges, we
will never know if it's any good. Have at it!
I hope, however, that in making the judgments you will prepare yourself
by learning as much as you can about the PCT model as it applies to
hierarchical control processes. Feel free to consult CSG-L as much as
you like to get obscure points explained. Also, feel free to read the
literature described in the monthly Intro to PCT.
I strongly suspect that this project of yours, which starts out by
looking for comparisons between theoretical PCT and
biological/neurological facts, will evolve rather quickly into basic
research that will greatly improve PCT as a theory of a real system. A
lot of PCT as it applies to higher brain functions is speculative and
tentative; we simply don't know what we should be modeling in these
areas. You may end up telling us about a lot of phenomena that we must
one day handle, and that alone would be a major contribution to PCT.
One specific thing you might be able to help me with now is to
explain to me the difference between PCT and other control system-
based models of brain function, for instance those of Scott Kelso,
whom my neurologist friend has mentioned.
All right, that's a very good way to start. I suggest that those who
have more ready access to the literature than I do start supplying some
quotes from Kelso (and others in the same line) which show what he and
they really say, so we can formulate specific statements about
differences from PCT. The models proposed by Kelso et. al. are very
definitely different from PCT, although in a few regards not hopelessly
so. Let's start a RE: Kelso thread, meaning Kelso's theories and others
like them.
-----------------------------------------------------------------------
Rick Marken (950501.2120)--
I think we have to be more specific about just why classical
conditioning data are likely to be incomplete from the PCT viewpoint.
You state the reason, but not clearly enough.
The problem is that in these classical experiments, nobody measured the
behavior of the controlled variable. Unless you can show that the sum of
the system's output and certain other environmental events turns out to
be a constant value of a controlled variable, you can't determine the
properties of the control system, if any, that was responsible. You
can't even prove that there was a control system.
When we do PCT experiments, one of the variables we record continuously
is the state of the controlled variable, or at least the variable we
think is being controlled. We also record the states of disturbing
variables and the variables we use to measure the action of the system.
So we have data for ALL the variables that an external observer might
see that are involved in the hypothetical control process. Even if we're
guessing about the nature of the controlled variable, we have enough
data so we can formally compute the state of the controlled variable
from the data that are available, independently of our assumptions about
the control system.
Once we have the right data, we can propose a control model with
appropriate parameters and reference signal, and fit it to the data. We
can try various models and pick the one that reproduces the behavior the
best, and that best predicts new behavior under changed conditions. Then
we can be comfortable with claiming that we have a control-system model
that explains the observations.
Unfortunately, when we try to make control models of classical
conditioning -- the salivation experiment, for example -- we find that
we have to make up some of the data. We can guess that the controlled
variable is the "liquidity" or the taste of the food-powder-saliva mix,
but we have no data from which we can independently compute the state of
the proposed controlled variable. All we can do is propose plausible
definitions of possible controlled variables, and then show that the
model works under the assumption that those definitions are right. But
we have no way, even indirectly, to MEASURE the state of the controlled
variable given only the existing data. The necessary data were simply
not taken. So our models become just self-fulfilling prophecies.
This points up one of the major problems in dealing with existing
experimental data taken by conventional scientists. When we look into
constructing a working model of the situation, we usually find that what
is going on is FAR more complex than the experimenters realized. In the
case of the salivation experiments, we find that to get the kind of data
that would be needed to arrive at ANY defensible explanation, we would
have to develop very sophisticated monitoring equipment and keep track
of variables that the original experimenters never dreamed would be
relevant. So in effect, the data that were taken are useless. We can't
just go back and fill in the missing data. The only way to come up with
a real analysis would be to do the experiment over again, this time
paying attention to everything that needs to be noticed.
Plausible explanations are easy to find. All you have to do is imagine
just the right missing data to make the explanation fit the data that
were taken. But plausible explanations aren't enough: we need to have
enough data so that NO OTHER EXPLANATION IS POSSIBLE. It's only when the
data rule out every explanation we can think of -- all but one -- that
we can claim to have a real scientific explanation, real meaning the
best that anyone can now offer.
I know that you really said all this, Rick, but these points are
critical and can't be emphasized too much.
----------------------------------------------------------------------
Best to all,
Bill P.
