COMMENT ⊗   VALID 00003 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002		BARWIS[W80,JMC] contains comments on Barwise's lecture and
C00004 00003	.require "memo.pub[let,jmc]" source
C00018 ENDMK
C⊗;
	BARWIS[W80,JMC] contains comments on Barwise's lecture and
his notes.  Let me suggest that others attending the lectures might
want to make similar files for exchanging comments.  It seems to me,
and others have complained, that we have slowed Barwise too much
by our comments in class and have made it difficult for the students
to follow his train of thought.  We give him printouts of the comment
files, and this may be of use to him.

	Incidentally, the E command αβXpo is useful for reading files
referred to in other files.  The effect of the command is to search
from the present location of the pointer for text that looks like
aaa[bbb,ccc] or aaa.bbb[ccc,ddd], interpret that text as a file
name and switch to that file.
.require "memo.pub[let,jmc]" source
.once center
Notes on Barwise's course - draft of {date}

[NB. The prudent won't look at drafts dated before the course ends].

	At the first meeting, B said that scenes would not be
individuals in his theory but part of its model theory, and that
the theory couldn't be done in first order logic.
It can be done in first order logic, e.g. by making scenes individuals.
Moravcsik called that bloating the ontology, but it can be regarded
as enriching it.  Barwise and I later agreed that the issue is
whether there are useful statements whose expression requires functions
from objects to scenes.

	In the jan 23 lecture in Carolyn's understanding, B said that
events would be entities in model theory but not in the language.  Useful
event valued functions or functions comparing events seem even easier.

Notes on Barwise's lecture notes.

p.3 - He is unfair to naive realism or at least to "Dick sees Jane".
There is no need to suppose
that "Dick sees Jane" is the only thing that can be said in order
to justify treating it as an example of the simplest thing that
can be said and from which the others can sometimes be deduced. 

The simplest deduction is to deduce from "John saw Mary" and "Mary
was running" to "John saw Mary running".  As Barwise points out,
this deduction isn't generally valid.  I think we will have to
want to say "If John saw Mary and Mary was running, then John saw
Mary running and knew that Mary was running unless something prevented
it".  We will appeal to non-monotonic reasoning.  The reason for
doubting the appropriateness of taking "John saw Mary running" as
basic is that it is only the first step down the slippery slope of
infinite qualification.  However, the non-monotonic reasoning hasn't
been developed in the intensional contexts that will be required to
use it to solve the problem.

	Maybe scenes will solve the problem, but I doubt it.

Remarks on the second installment of the notes

	1. It should be pointed out that the associationist "theory"
wasn't a theory, because it didn't present a mechanism but merely a
metaphor.  The cited arguments against it are also somewhat metaphoric for
that reason.

	2. Helmholtz had a better idea (as usual) than most of his
successors in his "unconscious inference", but it seems likely that the
pre-conscious processing of visual information is localized, and that the
results of the early stages are not available throughout the brain.
Therefore, while we might say that the first visual cortex knows
something, we won't be able to say that the person knows it.  This can be
discussed along the lines of Dennett's discussion of the localization of
various aspects of feeling pain.  Barwise says the same thing at the top
of p. 14.

	3. I haven't seen Bruner's paper yet, but if it is like the book
%2A Study of Thinking%1 by Bruner, Goodnow and Austin also written in the
fifties, there is a major error.  I can't decide whether Barwise makes
this mistake or not.  His examples suggest that he does.

	The trouble with the 1950s view of pattern recognition is that it
is based on the idea of categorization into one of a preassigned finite
set of categories and that a category is regarded as a boolean combination
of elementary categories.  At least this is the view taken by Bruner, et.
al and in Earl Hunts %2Concept Learning%1.

	This is the error that Russell points out pervades philosophy and
logic before Frege.  It amounts to using only the monadic predicate
calculus or even just the propositional calculus.  It doesn't see that
concepts require relations for their definitions.  Let me define a two
dimensional pattern as an example: One of them consists of four parts.
Three of the parts are horizontal line segments of equal length and the
fourth is a vertical line segment a bit longer than the horizontals.  The
left ends of the three horizontals co-incide with the bottom, middle and
top of the vertical.  Given this description, the pattern can be drawn.
The way I have described it, functions from segments to points and
functions from segments to numbers are used and the equality relation for
points and numbers are used.  Of course, the description could be done
entirely with relations, but it could not be given only monadic
predicates.

	The early artificial intelligence work on vision was motivated by
this disagreement with the "pattern recognizers" all of whom took the
boolean view.  Minsky and I pointed out that this approach would be
hopeless for controlling a robot that could be told to pick up the doll.
It cannot merely categorize the scene as %2doll-containing%1; it must
locate the doll and the shapes and attitudes of its parts.  The slogan was
"description not just discrimination".  In the 1970s, Azriel Rosenfeld and
others have developed picture grammars and have tried to make pattern
matchers like grammatical parsers.  Their success has been less than that
of the linguists, because visual patterns are more complicated, and the
articles I have seen don't seem to recognize that we are really interested
in recognizing a scene or an object as a three dimensional thing that
gives rise to two dimensional images by projection and that these images
are partially obscured by occlusion.

	4. I should mention that Marr's primal sketch is quite a lot like
the drawings generated at the Stanford AI Lab in the 1960s by Manfred
Hueckel.  Do Marr et. al. refer to Hueckel?  Binford says no and mentions
similar work by himself and Horn.

	5. I should like to see a definition of categorization to see if
its present use by Johnson-Laird is subject to the above criticism.

	6. On page 16 Barwise remarks that he doesn't "want to build any
accidental features of human or animal perception into the logic of
perception".  I agree, but I am surprised at his eagerness to build what
seem to be accidental features of English treatment of perception verbs
into his logic.

	7. Another question about the logic of perception.  One may say "I
see an Englishman" or "I hear an Englishman", and the giant said "I smell
an Englishman".  At this level all the modalities of perception are
treated the same.  Is it Barwise's intention that the whole logic of
perception should be independent of the modality?  Will he isolate the
part that can be treated this way?  I haven't thought about it much, but I
suppose that I have always thought that each modality requires its own
treatment in order to treat the relation between appearance and reality.
This treatment for vision involves the fact that objects have a three
dimensional structure and that we input a sequence of two dimensional
projections of parts of an object.  Also objects hide one another.

	8. Barwise characterizes the task of a vision system as that of
describing the scene or answering questions about it.  As I mentioned
before, we originally thought about AI perception of scenes in terms of
getting the information required to manipulate the scene.  The logic of
perception in the service of manipulation may be different and perhaps
more fundamental than that of perception in the service of conversation.

	9. This correct description of the analogy between the description
of a scene (noticing that the scene is taken as three dimensional) is not
really consistent with the earlier quotations from Bruner.

	10. Marr is looking for the "universals of human vision" as
Chomsky looks for the universals of human language.  Neither distinguishes
very clearly what constraints are peculiar to humans, and what are imposed
by the nature of light.  A perception system that used ultra sound that
penetrates solid objects and is partially reflected by internal surfaces
would be quite different.  Its images might be three dimensional from the
start since relative range information might be available.

	11. I repeat that I am unimpressed by the linguistic peculiarities
of perception verbs in English.  Of course, it might be taken as a clue or
surface manifestation of some more fundamental property of the logic of
perception, but these properties will have to stand on their own feet.  I
also doubt that anaphora is related to the logic of perception in any
fundamental way.

	All the above are haggles with Barwise's treatment of
preliminaries to his semantics.  The proof of the pudding for me will be
the expressiveness of his formalism and the reasoning it permits.  The
idea that perception can be usefully formalized by taking scenes as
fundamental seems quite plausible to me.