S-R is alive and well

CSGnet Archive CSG1995
1

From Tom Bourbon [950730.1315]

"Nobody really believes in S-R anymore." So goes the mantram of the
contemporary behavioral-cognitive-neuro scientist. If that's so, why do I
keep finding things, like the passage I quote below, published in _the_
premier journals in those fields?

A few months ago, when I stopped being employed as a psychologist (or as
anything else) I resigned from the American Psychogical Association.
Apparently my membership was independent of my subscriptions to a couple of
the really big journals; I still receive a couple of them, including
_Psychological Review_, than which there is nothing holier in all of psych
land. I just recieved the latest issue and found the following:

James L. McClelland, Bruce L. McNaughton & Randall C. O'Reilly (1995). Why
are there complementary learning systems in the hippocampus and neocortex:
Insights from the successes and failures of connectionist models of learning
and memory. _Psychological Review_, _102_, 419-457. Those certainly look
like some well-known people, at some well-known universities, writing on
some well-known topics.

For now, I'll only briefly mention the fact that much of what the authors
say about the hippocampus, the neocortex, retrograde amnesia, and
"consolidation" of "memory" for discrete "events" looks suspiciously like
what I first encountered on those topics when I was an undergraduate
student, a few decades ago. What impresses me the most is this sample of
what they say about neuro-cognitive plans and how they are related to S-R
behavior:

"Throughout this article, we discuss the structure present in ensembles of
events. What we mean by the term _structure_ is any systematic relationship
that exists within or between the events that, if discovered, could then
serve as an efficient representation of novel events or for appropriate
responses to novel inputs. Marr (1970) noted that events almost never repeat
themselves exactly (sic), yet people do learn from past experience to
respond appropriately to new experiences. If there is no structure -- no
systematicity in the relationship between inputs and appropriate responses
-- then, of course, (TB: of course! :wink: there will be no basis for responding
appropriately to novel inputs. But if a systematic relationship does exist
between inputs and appropriate responses, and if the organism has discovered
that relationship, then appropriate responding may be possible (TB: may
be?)." page 436.

"Systematic relationships" exist between discrete "events" in the
environment. The person-organism "discovers" those objective environmental
relationships, then creates "representations" of them, then uses the
representations as the "basis" for "matching the proper responses to novel
stimuli." If that isn't a description of the plan-driven input-output (S-R)
model that Bill Powers and I used in "Models and their worlds," I'll eat the
floppy disk on which I stored our manuscript after its final rejection. And
to think, all of our our reviewers said nobody believed that sort of thing
anymore.

So this is the best that behavioral-cognitive-neuroscience has to offer
here, most of the way through "The Decade of the Brain."

Later,

Tom

2

[Martin Taylor 950731 14:00]

Tom Bourbon [950730.1315]

I have borrowed, but not yet read, one of the articles excoriated by Tom
in a recent series of postings. From what Tom quoted, I could not tell
whether the comments were justified, since the quotes could equally well
have applied to a pure PCT theory as to an S-R theory or, as Tom put it,
"the plan-driven input-output (S-R) model that Bill Powers and I used
in "Models and their worlds,"" (By the way, Tom and Bill, I haven't yet
received the version of that paper as you wish to see it in IJHCS).

Here's the passage Tom quoted, with my insertions added to his:

"Throughout this article, we discuss the structure present in ensembles of
events.

[MMT:In this, there is no evident requirement that "event" be discontinuous or
discrete, any more than there is in the "event level" of the PCT hierarchy.]

What we mean by the term _structure_ is any systematic relationship
that exists within or between the events that, if discovered, could then
serve as an efficient representation of novel events or for appropriate
responses to novel inputs.

[MMT: The word "discovered" seems to have tweaked Tom, but as the article
is about the effects that occur in distributed networks, it is quite clear
that this "discovery" has nothing to do with conscious perception. It has
more in common with the weighting-linking structural changes that constitute
"discovery" (aka "reorganization") in a learning PCT hierarchy.]

Marr (1970) noted that events almost never repeat
themselves exactly (sic), yet people do learn from past experience to
respond appropriately to new experiences. If there is no structure -- no
systematicity in the relationship between inputs and appropriate responses
-- then, of course, (TB: of course! :wink: there will be no basis for responding
appropriately to novel inputs.

[MMT: As I read this, it is equivalent to saying that if an increase in the
output of a control loop increases the perceptual signal sometimes and
decreases it at random other times, rapidly and erratically switching between
them, the loop will not control the perception.]

But if a systematic relationship does exist
between inputs and appropriate responses, and if the organism has discovered
that relationship, then appropriate responding may be possible (TB: may
be?)." page 436.

[MMT: Rephrase: "But if a systematic relationship does exist between change
of outputs and effects on inputs, then control may be possible." Is not
"appropriate responding" equivalent to "appropriate sign of the relation
between error and output"?]

Tom:

"Systematic relationships" exist between discrete "events" in the
environment.

Discrete or not, this is indeed a requirement for control.

The person-organism "discovers" those objective environmental
relationships, then creates "representations" of them,

Which are embodied in the changing signs and magnitudes of connection weights,
just as in a control hierarchy as it reorganizes...

then uses the
representations as the "basis" for "matching the proper responses to novel
stimuli."

Just as does the control hierarchy that is able to produce outputs of the
correct sign for patterns of perceptual error not previously encountered.

If that isn't a description of the plan-driven input-output (S-R)
model that Bill Powers and I used in "Models and their worlds," I'll eat the
floppy disk on which I stored our manuscript after its final rejection.

That COULD be a description of plan-driven input-output. It NEED not be.
I suggest a little salt might go well with the floppy.

But if you want a "final rejection" of your manuscript, send it along:-)

And
to think, all of our our reviewers said nobody believed that sort of thing
anymore.

I'm in no way trying to say that McClelland et al even know about control.
But as I skim the paper, they are applying the same techniques of modelling as
do the PCT people--making a hypothesis, putting it into a computer program,
and seeing if the result is akin to what people do under equivalent
conditions. Moreover, the statements about what is learned (in the abstract
of the paper) seem to be quite applicable to our usual view of reorganization.
Perhaps I'll change my opinion when I read the paper more fully, but this
interpretation seems to be supported by the abstract and the graphs.

Shooting mouth off is good way to get sore foot (old non-Chinese proverb).

Martin

3

From Tom Bourbon [950801.1029]

[Martin Taylor 950731 14:00]

Tom Bourbon [950730.1315]

I have borrowed, but not yet read, one of the articles excoriated by Tom
in a recent series of postings. From what Tom quoted, I could not tell
whether the comments were justified, since the quotes could equally well
have applied to a pure PCT theory as to an S-R theory or, as Tom put it,
"the plan-driven input-output (S-R) model that Bill Powers and I used
in "Models and their worlds,""

Why, of course, Martin. How silly of me not to notice that the authors of
the article on "complementary learning systems" understood that they were
talking about perceptual control systems. I should have picked up on the
fact that, where I saw them use the words "stimulus" and "response," they
were really talking about systems that make certain of their own perceptions
match their intended perceptions. And my interpretation was wrong, again,
where I saw them say:

Marr (1970) noted that events almost never repeat
themselves exactly (sic), yet people do learn from past experience to
respond appropriately to new experiences. If there is no structure -- no
systematicity in the relationship between inputs and appropriate responses
-- then, of course, (TB: of course! :wink: there will be no basis for responding

I should have seen, as you did, that they really meant organisms do not
learn to match their outputs to their inputs. They were really saying that
organisms vary their actions (outputs) and make some of their perceptions
match intended perceptions.

I'm glad you brought me back to my senses, Martin.

(By the way, Tom and Bill, I haven't yet
received the version of that paper as you wish to see it in IJHCS).

Perhaps with good reason. Bill and I have decided that the way we wish to
see "Models and their worlds" is the way it is. We decided against making
the changes you suggested. We understand why you think we should revise the
text, to address many of the objections people _might_ raise in defense of
their preferred versions of S-R and plan-driven theories. But there are too
many of them, and, in the past, just as in your reactions to my assessment
of one of the newest S-R/connectionist papers, you have thought we did not
treat them fairly. This is not unusual. Every time we have tried to
compare PCT with any particular S-R or cognitive/plan-driven/command-driven
models, defenders of those models accused us of misrepresenting their pets.

Now CSG has a home page on the World Wide Web. It links to a collection of
references about PCT, and some of those references link to text material.
We have decided to put "Models and their worlds" there, as is, where anyone
who wishes to read it can do so. We are delighted at the chance to put the
paper out for comments and reactions, of any kind, from the widest possible
audience. "Publication" on the web will allow us to reach that kind of
audience, with the text _we_ want to put there. If we have screwed up
royally, our readers will let us know.

If that isn't a description of the plan-driven input-output (S-R)
model that Bill Powers and I used in "Models and their worlds," I'll eat the
floppy disk on which I stored our manuscript after its final rejection.

That COULD be a description of plan-driven input-output. It NEED not be.
I suggest a little salt might go well with the floppy.

And the Queen of England COULD be a Martian. But, you are right, of course.
McClelland, McNaughton and O'Reilly _were_ talking about perceptual control
systems, even though their words were about stimuli and responses. You win.
I prefer my floppies cooked on a charcoal grill, with a spicy sauce. And
I'd like a cold beer with that. Maybe that should be two beers -- it's a
hard floppy.

But if you want a "final rejection" of your manuscript, send it along:-)

We already have it.

Shooting mouth off is good way to get sore foot (old non-Chinese proverb).

Thanks ever so much for the lesson, old boy!

Tom

4

[Martin Taylor 950801 13:30]

Tom Bourbon [950801.1029]

Thank you, Tom, for your deep understanding that when I said:

I'm in no way trying to say that McClelland et al even know about control.

I really meant:

ยทยทยท

that the authors of
the article on "complementary learning systems" understood that they were
talking about perceptual control systems.

---------------

(By the way, Tom and Bill, I haven't yet
received the version of that paper as you wish to see it in IJHCS).

Perhaps with good reason. Bill and I have decided that the way we wish to
see "Models and their worlds" is the way it is. We decided against making
the changes you suggested.

And as I told you before (Messages of Jan 11-12 95), if you wanted to keep
the paper the way it was, not making the changes I suggested, that was
acceptable to me as editor. As I said, I would regret your not making the
changes, but it would not bar the paper from appearing in the special issue.
If you remember, I went directly to Brian Gaines to ask him if Closed Loop
publication was considered "prior publication" that would bar the paper
from appearing in the special issue, and he said that there was no problem.

You did say that you wouldn't mind incorporating material on reference
variation as an annex, though. You also said:

Right now we plan to say a little
more early on in the paper about what we mean by "core concepts" in the
three alternatives. I might also add a few recent quotes and citations
representing the other two models. We don't see many other places to make
changes.

If you no longer feel either change to be a good idea, it's your call.

But if you want a "final rejection" of your manuscript, send it along:-)

We already have it.

Only by your own choice.
-------------------------

Do you read all your messages as carefully as these?

Shooting mouth off is good way to get sore foot (old non-Chinese proverb).

Well, I still don't know whether my foot is sore--perceptual signals from
mouth shoot-off haven't reached my dinosaur brain yet. (By your comment
that follows, I get the impression you thought I meant that you were
shooting your mouth off, which isn't the case. I meant me. If you got
a different impression, I apologize).

Martin