Bicker about Bickhard

[from Joel Judd]

STOPPED SHAKING YET, RICK?

Bill, Rick, Martin, Bruce, Tom:

Gary might not have provided enough of his Bickhard conversation to provoke
some discussion. Of course if you're controlling for not seeing more of the
conversation then this post won't help. But I think he IS on the
epistemological track, and that is something sorely needed. I am also
personally interested in getting feedback on a question regarding the
mechanics of an HPCT, so humor me for a bit.

On the epistemology, he is adamant that anyone dealing with knowledge
systems recognize that ahen one is at the point of modelling LEARNING, then
the system has to be able to develop interactions with its environment
without knowing what is to be represented in those interactions. As soon as
one assumes representation of knowledge in a system then one begins down
the road of infinite regress. In education, this fallacy leads to the
"transmission" notion of teaching: the teacher has some knowledge the
student doesn't, and must somehow transmit that knowledge to the student.
This is a paradox of learning that is yet to be widely admitted and studied
in educational theory. Currently, the field of study that is distracting
many from addressing this fallacy head-on is neural nets, because they
APPEAR to be learning, but in fact the nets are not developing their own
knowledge from scratch--they are deriving new combinations of prior
knowledge (provided by a programmer).

So while we may be a ways from modelling knowing systems from scratch, I
think that this basic epistemelogical fact needs always to be kept in
mind--especially in an educational setting. In this part of the
conversation Bickhard says:

"...in classic wax slate models errors are inevitable but the more you can
avoid them the better. Well try applying that to any skill learning. You
don't learn to drive, you don't learn to walk, you don't learn to play
golf, you don't learn any skill by practicing all of the subelements to
perfection and then combining them and pracitcing that to perfection. That
is absolutely the worst conceivable way to learn anything. And yet at the
cognitive level we assume that that's the optimal way to learn
anything...people learn things by progressive approximations just like they
learn any other skill and when they learn what this approximation needs to
do in order to become a better approximation, in effect what they have done
is to learn a new critical principle. They have learned a way that this
approximation still makes some sort of an error or another and by learning
in that manner they end up understanding what they've learned because they
have learned all of the error criteria that it is satisfying that make it
an acceptable or good or whatever proposal.."

At this point Gary mentions that just error doesn't give you the solution,
which Bickhard acknowledges by pointing out the role of environmental
influences and, in education, the classroom, and how good teachers are ones
that appreciate the process of "scaffolding" or working one's way through
problems. On a personal note, I don't feel I am as good a teacher as others
because I tend to FORGET how I learn something, whereas good teachers (I
think) tend to REMEMBER what it took to learn something and are adept at
identifying and appreciating other who are going through similar processes.

The point here is that Bickhard's epistemological arguments are basically
correct and there educational implications (as well as present or future
implications for AI, robotics, etc.) should be recognized and remembered,
even if the mechanics of his model don't interest PCTers.

And to the mechanics...

I think, I repeat I THINK, part of Bickhard's problem with reference levels
has to do with their origins. So I'm going to go ahead and ask this
question without consulting PCT scripture. If one understands a reference
level to be a "goal" in the sense that the organism is acting to perceive
perceptual inputs similar to the reference signal, then the reference
signal can only come from two sources: memory of past experience (i.e., the
recording of a past reference signal), or my imagination of a reference
signal. In either case, in order to EXPLAIN the system, one would like to
be able to explain where the reference signals come from. In the case of
LEARNING, memory signals would seem in a sense to be trivial, since they
result from prior learning (though they are anythign but trivial for, say,
"practice" or "fluency" or whatever). Imagination signals would seem to be
key in learning something new. But we're faced with the dilemma: How do I
know what I am trying to learn? In other words, How do I have a reference
signal for experience I have not had? How do I know when my reference
signal for 'ride a bike' or 'graduate from the university' have been
satisfied (I purposely chose these obvious examples)? More to my and some
other's interests: How do I know when my reference signal for 'learn
another language' is satisfied? I think in a considered answer to those
questions lies much of the variability one observes in others' attempts to
achieve their goals.

FInally (really) [is Lubin still out there?]: is there any
neurophysiological support for the following supposition (and here in
fairness Bickhard is just responding to Gary's query):

"...I argue that the nervous system (NS) should be looked at as--well, even
this is only a first approximation--as a complexly organized system of
oscillators that modulate each others' activity. Now that's only a first
approximation and I think a closer one is the NS should be looked at as a
complex topology of media for oscillations to move in. And those
oscillations move in the medium in accordance with whatever the local
properties of the medium are and in accordance with this topology and
thereby modulate each other. That's the general architecture I would argue
for. And one demonstration that something like that ought to be the case is
the fact that a very large proportion of our CNS neurons are silent, they
never fire...but in terms of modulatory effects they can be doing a lot."
(Such as affecting the ionic concentrations in the area which affect
dendritic computations, etc.)

Are topologies in use in neurophysiology? In engineering? Do they make
sense in CT terms? Can someone provide a laymen's explanation of why they
are or aren't to help my conceptualization of the brain?

Awaiting replies....

[Martin Taylor 920629 17:00]
(Joel Judd, undated (920629?))

I am also
personally interested in getting feedback on a question regarding the
mechanics of an HPCT, so humor me for a bit.

The questions that follow don't seem to need humoring.

[Bickhard says that] anyone dealing with knowledge
systems recognize that ahen one is at the point of modelling LEARNING, then
the system has to be able to develop interactions with its environment
without knowing what is to be represented in those interactions. As soon as
one assumes representation of knowledge in a system then one begins down
the road of infinite regress.

No, there's no infinite regress here, unless one's arguing style is singularly
obtuse. Knowledge can certainly be represented, but interactions do not
require the kind of representation that implies regress.

I think the fallacy here is the usual one perpetrated by philosophers, that of
using a word in two different senses, leading to an apparent contradiction.
Only from a cognitive-symbolic model does one come to the notion that there
is a problem in "interacti[ng] with its environment without knowing what is
to be represented in those interactions." "Knowing what is to be represented"
implies that the actor (interactor?) is viewing its own performance from
outside and conducting it by means of a set of rules based on some knowledge
representation. Interactions performed by a control system don't need to
work that way, and indeed will not work that way except at levels above
programme. But there is a representation, based in HPCT on the reorganization
that alters the sign of the gain function that turns an error signal into the
reference for a lower-level control system.

There is another representation, in the perceptual combining function for
each ECS. One may ask how that representation is learned, but it is not the
same question as "how does the system develop interactions with the environment
without knowing what is to be represented in those interactions." The system
as a whole does not "know" what is in these perceptual functions, but it
can learn them. Simple Hebbian learning may develop them, or genetic algorithms
(as we intend to attempt), or something akin to reorganization. The point
here is that if there is a complex environmental variable susceptible to
control in the environment, and an ECS controls it, even poorly, then that
ECS can adapt to control it better, by modification of its perceptual function.
No "knowledge" is required, unless one asserts that the perceptual function
that adapts is its own knowledge representation. There's certainly no
infinite regress implied by that.

In education, this fallacy leads to the
"transmission" notion of teaching: the teacher has some knowledge the
student doesn't, and must somehow transmit that knowledge to the student.
This is a paradox of learning that is yet to be widely admitted and studied
in educational theory.

There's no paradox that I can see. The teacher does indeed have knowledge
that the student doesn't, and must somehow transmit it. The fallacy is in
thinking that the teacher can do this if the student doesn't have the building
blocks for the new knowledge. To build a house needs bricks, pipes, wire,
paint, and myriads of other things, but just to pile them up doesn't make a
house. You have to organize them in previously unknown ways (new knowledge).
Neither can you make a house unless you have the bricks, etc.

In our BLC theory of reading (Taylor and Taylor, The Psychology of Reading,
Acvademic Press 1983) we discussed what we called a "three-phase" pattern
of learning: (1) acquisition of the perception of wholistic structures which
are subdividable in some way, (2) the perception of the subdivisions and the
relations that can and cannot occur among them, and (3) a new, more precise
perception of the whole structures, which now includes the perception of
the relationships among the units. Of these phases, only (2) can be "taught."
(1) and (3) are "learned." In reading, (2) corresponds to the teaching of
phonics, and it is the teacher's responsibility to provide the circumstances
in which the regularities that permit subdivision can be perceived. A teacher
can then inform the student about the ways the units can and cannot fit
together. Of course, if you already have the units, new ways of organizing
them can be described and "taught." So you can build both up and down from
any suitable percept or set of percepts.

There was an experiment done in the late 50s or early 60s by Wilson P (Spike)
Tanner at Michigan that illustrates the ability of people to develop brand new
perceptual functions. His claim was that through feedback, people could
learn to discriminate any pair of auditory signals that were physically
distinct. He designed a pair of signals that nobody could tell apart, but
that were physically quite different. Then he put people into a psychophysical
discrimination study using those signals, for a long time each day for many
days. Typically, the subject would get 50% correct (no discrimination) for
many days, and would insist that the signals were identical. Then, one day,
some distinction might be observed, and very quickly, perhaps within the
same day, the score would rise to near 100% (perfect discrimination). One
subject, as I recall, took 43 days before catching on.

Currently, the field of study that is distracting
many from addressing this fallacy head-on is neural nets, because they
APPEAR to be learning, but in fact the nets are not developing their own
knowledge from scratch--they are deriving new combinations of prior
knowledge (provided by a programmer).

And these combinations ARE new knowledge. No-one develops their knowledge
(if by that you mean performance ability) from scratch. We come from a few
billion years of evolution, and have many building blocks available at (and
before) birth. We develop new ones from the originals, and use them as
building blocks. Programmers seldom know what their neural networks are doing,
so if all the net's behaviour is knowledge provided by the programmer, it
is knowledge the programmer did not know she had. A strange kind of knowledge.

Does Bickhard claim that one knows only what one can teach?

Martin