Hancock reply -- at long last

[From Rick Marken (930826.1300)]

Tom Hancock (930723.1000) --

I said:

What you are trying to figure out is something I don't understand at all,
which is "how does a person know that an imagined perception
corresponds to a previously experienced perception?"

Tom says:

I have thought that the way a person knows is by the error signal. If there
is a large error signal in the match of the imagined perception to the
reference then he will report not being certain. But if possible matches
have been accessed and the signals from the incorrect ones have been
completely suppressed and there remains a good match, then the subject
knows (or thinks he knows) that he has found the right one.

I think a diagram -- properly labelled -- would help me understand
your model. My problem is this: you present me an X and I am to
provide the appropriate Y associate. X exists as a perception for me; Y
is provided as an imagination. How do I know that the Y I imagine as
the associate of X 1) is an imagination based on previous exposure to
X and 2) that it is an imagination that corresponds to the Y that I
actually perceived previously? Maybe this is obvious but I can't think
of how to implement this as a working model.

I said:

The associations you are talking about as perceptions are perceptions --
the subject can perceive that after seeing X you say Y. p* would be the
reference for that perception; at first the subject may have no reference
for the association -- s/he doesn't care what comes after X. But you have
asked thats/he learn to have a reference for this association that is the
same what s/he experienced earlier. So after perceiving X-Y s/he is
supposed to store a reference value, p*, of X-Y . The next times/he sees X
s/he should produce Y to match my reference. So s/he can control the
perception of the association by producing Y (instead of Z,
etc) when you show him/her X.

Tom says:

Yes, that is the way I see it with one addition. The subject does not just
say Y automatically (at least at the learning stage). The subject mediates
between the associates--typically by building some verbal category
connections. So when X is perceived, S searches for the bridge or
mnemonic. When it is found and other similar mnemonics are rejected
(signals inhibited), then the S says Y.

What process decides that the appropriate mnemonic has been found?
What process rejects or inhibits mnemonics?

Your PCT schema if more fully developed than mine (you have many more
constraints guiding your imagination to PCT-correct interpretations), so
do you have any thoughts about what kind of data I would be looking for?

The main constraint guiding my imagination is that the model should
work like the system modelled -- ON ITS OWN. So the research that you
do should be aimed at discovering how the system does what it does
on it's own . Working "on its own" (autonomy) means that the model
(like the modellee) must be able to specify the inputs that are right
for itself and it must be able to act in order to to make those inputs
match the specifications. We get to control theory just by being
careful modellers.

I like your suggestion and I believe I will write a proposal partly along
that line for continuation funding from the Air Force. Maybe I could send
you a copy--if you are not getting overloaded by me.

Sure, I'd like to take a look at it.

I have looked at response
patterns for one subject who has been consistently correct with one item.
My problem has been interpreting this in terms of PCT

If you have the trial by trial data for individual subjects there may be
hints about how to model the process in there. Maybe you can post
some data for one subject?

Your hypothetical data appears similar to what the subjects actually do. In
addition I have been trying to include the response time, and post-
instructional information time which is associated with each trial. But so
far I am not sure what how it all fits with a precisely articulated model.

That's because we have no model. The next step must really be to
develop a model that can be tested in a simple experimental situation.
I don't know if I can really help with this modelling -- it seems too
complex for me. I used to do memory experiments and I know
there are some working models of memory but I don't know if they
are really autonomous;they might have a bit too much of the modeller
built into them. But maybe some existing models will help.

Rick, in light of our discussion to this point, do you have any additions (or
subtractions) on what the dynamics of the model should be like?

Nothing very clever. Why don't YOU develop the model. Don't worry
about being PCT orthodox. Just try to develop a model that works on
its own. Keep thinking "autonomy". The result will almost certainly
be a control system; don't worry.

That's the best I can do for now. I do hope you stick with the research;
if you do, your first step should be to develop a working model of a
person doing your confidence rating of association study. Make some
diagrams of the model and then maybe some of the smart people on
the net can help you converge to a working (autonomous) solution.

Best

Rick