[From Bill Powers (920603.1700)]
Ray Allis (920603.1100) --
RE: AI's hope of computer intelligence:
Maybe there are side benefits in better programming methods and the
handling of complexity, but 34 years of "AI" has not yet touched
"Intelligence".
Didn't Columbus think that if he sailed West, he would reach Cathay? I
think it has been conclusively proven that the initial proposition was
false -- but he found something else (which a lot of people who live near
here wish he hadn't found, but that's a different subject). I haven't been
impressed by the symbol-manipulations of AI, either, but I sure do admire
the techniques. And some day, the AI types are going to wake up to the fact
that what they've been DOING, themselves, is what they should have been
studying. If you want to understand the organization of human behavior, you
can't dismiss anyone's approach to it: the approach itself is evidence.
... confusion of "model" with "simulate" (I call this the
"Mathematicians' Mistake").
A simulation, being a construction of logical statements, is unaffected
by events in the physical world.
You, when you program a computer to produce a "little man" on its
screen, have constructed a simulation. This is a Good Thing, because
now you can work out all the implications of the system of logic.
Just remind yourself that you are only making explicit the implications
YOU PUT THERE IN THE FIRST PLACE.
I think all these comments relate to a particular kind of modeling or
simulation, the kind that is based on logical statements or empirical
generalizations. Your version of a tornado model seems to be in the same
vein. But there's a different sort of stimulation that doesn't use any
logical statements and isn't a generalization. A supercomputer tornado
model doesn't just say that a funnel will have a certain shape as a
function of temperature. It doesn't actually deal with tornadoes or funnels
at all. It deals with little packets of air that are subject to laws of
physics. Given a certain water content, density, velocity, and temperature
as a starting point for each packet, the computer simply applies the laws
to generate the next state of each packet, and the next, and so on, one
millisecond following the next. All the packets interact with each other
according to their nature and the laws of physics -- as you say, to the
extent that we understand their nature and the pertinent laws. The computer
program presents a picture of what all the packets are doing as time
progresses. It's up to a human observer to give the result a name, such as
"tornado" or "funnel."
If you start with one set of initial conditions, you get a warm summer
breeze. Start with different conditions and you get a tornado. The point of
this kind of simulation is to see what we can't observe directly: how
physical factors influence the results. You can't run a real tornado over
and over, varying the humidity a little each time to see its influence. But
if you have a good model, you can do this via simulation in a computer.
The Little Man works that way. The basic simulation doesn't use logical
statements. It is built from mathematical descriptions of how the parts of
the body behave or are guessed to behave. How much signal does the tendon
receptor generate under a given stress from a tensing muscle? How much
feedback signal is there when a muscle lengthens by a specific amount? How
much does the muscle itself stretch when subject to tension? How much
shortening is there in the contractile part when signals reach it from the
spinal neurones? How does the arm respond physically to couples -- torques
-- applied at its joints? The answers to these questions are not in the
model; the questions are posed in terms of adjustable parameters. These
parameters all have direct physical significance.
By running the simulation and adjusting the parameters to get the closest
possible match of the model's behavior with that of a real arm, we can
estimate the values of the physical parameters of the real arm. This is the
basic method of modeling that is behind all the physical sciences and
engineering. In this world, "simulation" simply means constructing a
quantitative analog of the system, organized according to a model, and
running it.
···
-------------------------------------------------------------------
Andy Papanicolaou and Tom Bourbon (920603) --
Take for example, the phenomenon of skill acquisition or habit >formation
whereby a sensori-motor event sequence, originally produced >haltingly,
awkwardly with great effort and requiring conscious >attention comes to be
performed effortlessly automatically and reliably >following a number of
trials or repetitions.
All right, I'll take it. This looks to me like a control system gradually
becoming organized. It's a mistake, by the way, to assume that a behavior
that's always performed the same way is always performed by the same
combinations of muscle tensions. The opposite is most likely to be true.
Most behaviors are named in terms of outcomes, not outputs.
I think the basic problem here is the distinction between "conscious" and
"automatic." Habits are carried out without thought. If one thinks of
control systems as always involving thought, then of course it seems that
habits can't be controlled processes. But nothing in control theory says
that control has to involve thought or even awareness. In fact there are at
least six levels of control in my model BELOW the levels we would associate
with thinking. Even spinal "reflexes" are control systems.
Thus the gradual acquisition of the ability to pronounce the French "tu"
can be viewed as the gradual acquisition of the ability to make the correct
sound appear in perception (or more realistically, what one assumes is the
correct sound). This is certainly not hard to explain in control-theory
terms: make the sound you're hearing match the reference sound. This does
NOT have to involve awareness, although it probably does during the
learning phase.
To produce that syllable a specific sequence of partially overlapping
articulatory gestures is required constituting a pattern with some
invariant features.
According to my linguistics friends on this net, that is not generally
true. There are only a few phonemes that are closely associated with
specific configurations of the articulators. We can hear the same phoneme
with the articulators in a variety of arrangements; even a single speaker
will use different articulations to produce the same heard sound.
This says that a given articulator configuration or behavior does not
necessarily pin down the sound that will be recognized. It is auditory
perception that defines the range of DIFFERENT input sounds that will be
heard as the SAME word, correctly pronounced. All that the articulators
have to do is produce one of the configurations that will result in a
perception within the acceptable range. I'm sure that a Frenchman can say
"tu" in a friendly way, seductively, sarcastically, or condescendingly --
and that each way of saying it involves a different configuration of
articulators. I'm not at all sure, by the way, that it's necessary to round
the lips to say "tu" correctly. I can say "oo" with my lips stretched in a
smile. I think I can say "tu" with recognizeable correctness in the same
way. Of course that's just how I perceive it.
I'm sure that some motor learning takes place as the new pronunciation is
mastered. But what does that mean? It certainly doesn't mean "training the
muscles," as some people say. All muscles can do is contract; you can't
train them to do anything else. They pull their ends together, and that's
about it. What you can train is a lower-level control system, a kinesthetic
control system. And since muscles always behave the same way, training a
low-level control system is largely a matter of training its perceptual
function, so it's controlling the appropriate function of kinesthetic
sensations.
So: it isn't generally true that uttering a particular sound is a matter of
standardizing an articulator behavior, so it isn't generally true that a
"neuromotor plan" (thought of as generation of a fixed output pattern) is
necessary to cause a recognizeable particular sound to occur. I think this
is finally settled by experiments with disturbances that required DIFFERENT
articulations in order to produce the SAME pronunciation. We went through
this on the net some time ago with our linguists, and I believe the
consensus was that people do not need any practice to say words correctly
even when there is mechanical interference with pronunciation. The
articulators are simply used in a different way -- just the way required so
that a human listener will hear the same, or nearly enough the same, sounds
(even if a sound spectrograph says they're different). This phenomenon
can't be explained by the concept of a "neuromotor plan." As far as I can
see, it can be explained only as control of perception.
The perception tu cannot be achieved unless the same muscles are always
used to always do almost the same thing - to produce the same pattern >of
articulatory gestures.
It seems very unlikely to me that some behaviors would be organized
according to one fundamental principle, and other behaviors -- of the same
kind -- according to a different one. Such a proposition says that in the
nervous system, there are neuromotor plan generators that are used for all
actions in which it happens that the same motor output always produces
essentially the same perceptual result, but that in others, a control
system with a perceptual function, comparator, and output function is used
-- a completely different architecture.
There is certainly no penalty for using a control system even when no
disturbances occur, or can occur. Control organizations are generally far
simpler than plan generators. Just look at what the motor program people
have to go through with their arm models. They have to ask the nervous
system to specify the acceleration, velocity, and position of the arm at
every instant during a movement, and to calculate continually the inverse
kinematics of the arm in order to turn those specifications into torque
commands that produce the right movement (after passage through the FORWARD
kinematics of the arm). By the time the late Leonard Bernstein figured out
how to give the orchestra the downbeat in this way, the players would have
finished the symphony. And a control-system model does the same thing, far
better, with only a tiny fraction of the computations.
Yet there are all kinds of empirical findings suggesting that no >phoneme
can be perceived as such unless some "motor plans" for its >production have
already been formed even though the learner may not be >able as yet to
bring his articulation to implement the plant (and >produce the phoneme):
The native speaker of Japanese cannot hear, >cannot experience the
difference between r and l when first exposed to >these sounds. Slowly,
she may come to experience them (once a motor >plan for their production is
presumably formed)
What possible evidence can there be of a motor plan that has been formed
but is not yet being implemented? I could easily believe that a person
gradually learns a perceptual distinction between "r" and "l" -- for one
thing, I could ask the person if they sound different. I could also believe
that this perceptual learning would go faster if the person was actually
trying to produce those sounds at the same time; part of the perception is
the perception of how it feels to say the sound. When you say "No, not
'another,' 'A MMMMMother'," you emphasize the closed mouth; "M" is more a
feel than a sound, when being compared with "N".
You can't control the sounds of "l" or "r" reliably if you can't perceive
the difference between them -- auditorily and kinesthetically/tactily. If
you have the wrong feel going with the right sound, you'll feel the wrong
articulation and hear its result as right.
... we can hardly forego the indications that before acquiring a skill
to pronounce new speech sounds we cannot take it for granted that the
learner possesses auditory images of the new sounds to use as reference
signals.
Those indications just indicate the theoretical bias of the experimenter --
i.e., the explanation that makes the most sense, given the kind of
explanation the experimenter prefers. That's just how it is, with control
theorists and everybody else.
Before I took any of those indications at face value, I'd have to see the
original experimental data for myself, to satisfy myself that this isn't
just another of those statistical experimental facts that's true of some of
the people in the experiment and false of the others. The problem with
standard psychology is that most of its facts are of that statistical
variety, but they're always cited as if they were true for every
individual. If a fact isn't true of everyone, how can we use it to justify
a model that's supposed to be true of everyone?
I could believe this: that you could teach a Japanese how it FEELS to say
'l' and 'r' by manipulation of the mouth and tongue and by demonstrations
using mirrors, so that presented with a printed 'l' or 'r' the person could
produce an appropriate articulation, by kinesthetic control. And I can
believe that while an English speaker might recognize the resulting sound
as correct, the Japanese could still not hear the difference, having
learned the feel but not the sound. Would that answer the question you're
raising?
You propose something similar, but using vision as the controlled variable:
... it may not be the auditory image of the sound but the visual image
of the speaker's articulators that serves as a reference signal.
I think there has to be another step here: seeing the teacher's
articulators, the learner has to imagine how it would feel to make one's
own articulators look like that (or better yet actually try it, with a
mirror). The difficulty with supposing that what is learned is visual
feedback is that normally a speaker can't see his or her own articulators.
It isn't the LOOK of the tip of the tongue against the hard palate that
creates a 't', but the way it feels inside your mouth, where you usually
can't see.
--------------------------------------------------------------
I think that "habit formation" carries a lot of old-fashioned freight. It
assumes uniformity of output where in most cases there is uniformity only
in outcomes. Control theory works with or without awareness, and whether or
not disturbances are present. I think "learning" or "reorganization"
handles the situation perfectly well, and that the simplest general
explanation is that what is learned or reorganized is a control system, not
an output.
-----------------------------------------------------------------
Best to all,
Bill P.