[comp.cog-eng] The symbol grounding problem: "Fuzzy" categories?

harnad@mind.UUCP (Stevan Harnad) (06/30/87)

In comp.ai.digest: Laws@STRIPE.SRI.COM (Ken Laws) asks re. "Fuzzy Symbolism":

>	Is a mechanical rubber penguin a penguin?... dead...dismembered
>	genetically damaged or altered...? When does a penguin embryo become
>	a penguin?... I can't unambiguously define the class of penguins, so
>	how can I be 100% certain that every penguin is a bird?... and even
>	that could change if scientists someday discover incontrovertible
>	evidence that penguins are really fish. In short, every category is a
>	graded one except for those that we postulate to be exact as part of
>	their defining characteristics.

I think you're raising the right questions, but favoring the wrong
answers. My response to this argument for graded or "fuzzy" category
was that our representations are provisional and approximate. They
converge on the features that will reliably sort members from
nonmembers on the basis of the sample of confusable alternatives
encountered to date. Being always provisional and approximate, they
are always susceptible to revision should the context of confusable
alternatives be widened.

But look at the (not so hidden) essentialism in Ken's query: "how can I
be 100% certain that every penguin is a bird?". I never promised that!
We're not talking about ontological essences here, about the way things
"really are," from the God's Eye" or omniscient point of view! We're
just talking about how organisms and other devices can sort and label
APPEARANCES as accurately as they do, given the feedback and
experiential sample they get. And this sorting and labeling is
provisional, based on approximate representations that pick out
features that reliably handle the confusable alternatives sampled to
date. All science can do is tighten the approximation by widening the
alternatives (experimentally) or strengthening the features
(theoretically).

But provisionally, we do alright, and it's NOT because we sort things
as being what they are as a matter of degree. A penguin is 100% a bird
(on current evidence) -- no more or less a bird than a sparrow. If
tomorrow we find instances that make it better to sort and label them
as fish, then tomorrow's approximation will be better than today's,
but they'll then be 100% fish, and so on. 

Note that I'm not denying that there are graded categories; just that
these aren't them. Examples of graded categories are: big,
intelligent, beautiful, feminine, etc.

>	You are entitled to such an opinion, of course, but I do not
>	accept the position as proven...

(Why opinion, by the way, rather than hypothesis, on the evidence and
logical considerations available? Nor will this hypothesis be proven:
just supported by further evidence and analysis, or else supplanted by
a rival hypothesis that accounts for the evidence better; or the
hypothesis and its supporting arguments may be shown to be incoherent
or imparsimonious...)

>	...We do, of course, sort and categorize objects when forced to do so.
>	At the point of observable behavior, then, some kind of noninvertible
>	or symbolic categorization has taken place.  Such behavior, however,
>	is distinct from any of the internal representations that produce it.
>	I can carry fuzzy and even conflicting representations until -- and
>	often long after -- the behavior is initiated.  Even at the instant of
>	commitment, my representations need be unambiguous only in the
>	implicit sense that one interpretation is momentarily stronger than
>	the other -- if, indeed, the choice is not made at random.

I can't follow some of this. Categorization is the performance
capacity under discussion here. ("Force" has nothing to do with it!).
And however accurately and reliably people can actually categorize things,
THAT'S how accurately our models must be able to do it under the same
conditions. If there's successful all-or-none performance, the
representational model must be able to generate it. How can the
behavior be "distinct from" the representations that produce it?

This is not to say that representations will always be coherent, or
even that incoherent representations can't sometimes generate correct
categorization (up to a point). But I hardly think that the basis of
the bulk of our reliable all-or-none sorting and labeling will turn
out to be just a matter of momentary relative strengths -- or even
chance -- among graded representations. I think probabilistic mechanisms
are more likely to be involved in feature-finding in the training
phase (category learning) rather than in the steady state phase, when
a (provisional) performance asymptote has been reached.

>	It may also be true that I do reduce some representations to a single
>	neural firing or to some other unambiguous event -- e.g., when storing
>	a memory.  I find this unlikely as a general model.  Coarse coding,
>	graded or frequency encodings, and widespread activation seem better
>	models of what's going on.  Symbolic reasoning exists in pure form
>	only on the printed page; our mental manipulation even of abstract
>	symbols is carried out with fuzzy reasoning apparatus.

Some of this sounds like implementational considerations rather than
representational ones. The question was: Do all-or-none categories
(such as "bird") have "defining" features that can be used to sort
members from nonmembers at the level of accuracy (~100%) with which we
sort? However they are coded, I claim that those features MUST exist
in the inputs and must be detected and used by the categorizer. A
penguin is not a bird as a matter of degree, and the features that
reliably assign it to "bird" are not graded. Nor is "bird" a fuzzy
category such as "birdlike." And, yes, symbolic representations are
likely to be more apodictic (i.e., categorical) than nonsymbolic ones.
-- 

Stevan Harnad                                  (609) - 921 7771
{bellcore, psuvax1, seismo, rutgers, packard}  !princeton!mind!harnad
harnad%mind@princeton.csnet       harnad@mind.Princeton.EDU

franka@mmintl.UUCP (Frank Adams) (07/02/87)

In article <936@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes:
|The question was: Do all-or-none categories (such as "bird") have "defining"
|features that can be used to sort members from nonmembers at the level of
|accuracy (~100%) with which we sort? However they are coded, I claim that
|those features MUST exist in the inputs and must be detected and used by the
|categorizer. A penguin is not a bird as a matter of degree, and the features
|that reliably assign it to "bird" are not graded.

I don't see how this follows.  It is quite possible to make all-or-none
judgements based on graded features.  Thermostats, for example, do it all
the time.  People do, too.  The examples which come to mind as being
obviously in this category are all judgements of actions to take based on
such features, not of categorization.  But then, we don't understand how we
categorize.

But to take an example of categorizing based on a graded feature.  Consider
a typical, unadorned, wooden kitchen chair.  We have no problem categorizing
this as a "chair".  Consider the same object, with no back.  This is
clearly categorized as a "stool", and not a "chair".  Now vary the size of
the back.  With a one inch back, the object is clearly still a "stool"; with
a ten inch back, it is clearly a "chair"; somewhere in between is an
ambiguous point.

I would assert that we *do*, in fact, make "all-or-none" type distinctions
based precisely on graded distinctions.  We have arbitrary (though vague)
cut off points where we make the distinction; and those cut off points are
chosen in such a way that ambiguous cases are rare to non-existent in our
experience[1].

In short, I see nothing about "all-or-none" categories which is not
explainable by arbitrary cutoffs of graded sensory data.

---------------
[1] There are some categories where this strategy does not work.  Colors are
a good example of this -- they vary over all of their range, with no very
rare points in it.  In this case, we use instead the strategy of large
overlapping ranges -- two people may disagree on whether a color should be
described as "blue" or "green", but both will accept "blue-green" as a
description.  The same underlying strategy applies: avoid borderline
situations.
-- 

Frank Adams                           ihnp4!philabs!pwa-b!mmintl!franka
Ashton-Tate          52 Oakland Ave North         E. Hartford, CT 06108

harnad@mind.UUCP (Stevan Harnad) (07/03/87)

In Article 176-8 of comp.cog-eng: franka@mmintl.UUCP (Frank Adams)
of Multimate International, E. Hartford, CT.writes:

>	I don't believe there are any truly "all-or-none" categories. There are
>	always, at least potentially, ambiguous cases... no "100% accuracy
>	every time"...  how do you know that "graded" categories are less
>	fundamental than the other kind?

On the face of it, this sounds self-contradictory, since you state
that you don't believe "the other kind" exists. But let's talk common sense.
Most of our object categories are indeed all-or-none, not graded. A
penguin is not a bird as a matter of degree. It's a bird, period. And
if we're capable of making that judgment reliably and categorically,
then there must be something about our transactions with penguins that
allows us to do so. In the case of sensory categories, I'm claiming
that a sufficient set of sensory features is what allows as to make
reliable all-or-none judgments; and in the case of higher-order
categories, I claim they are grounded in the sensory ones (and their
features).

I don't deny that graded categories exist too (e.g., "big," "smart"), but
those are not the ones under consideration here. And, yes, I
hypothesize that all-or-none categories are more fundamental in the
problem of categorization and its underlying mechanisms than graded
categories. I also do not deny that regions of uncertainty (and even
arbitrariness) -- natural and contrived -- exist, but I do not think that
those regions are representative of the mechanisms underlying successful
categorization.

The book under discussion ("Categorical Perception: The Groundwork of
Cognition") is concerned with the problem of how graded sensory continua
become segmented into bounded all-or-none categories (e.g., colors,
semitones). This is accomplished by establishing upper and lower
thresholds for regions of the continuum. These thresholds, I must
point out, are FEATURES, and they are detected by feature-detectors.
The rest is a matter of grain: If you are speaking at the level of
resolution of our sensory acuity (the "jnd" or just-noticeable-difference),
then there is always a region of uncertainty at the border of a category,
dependent on the accuracy and sensitivity of the threshold-detector.

But discrimination grain is not the right level of analysis for
questions about higher-order sensory categories, and all-or-none
categorization in general. The case for the putative "gradedness" of
"penguin"'s membership in the category "bird" is surely not being
based on the limits of sensory acuity. If it is, I'll concede at once,
and add that that sort of gradedness is trivial; the categorization
problem is concerned with identification grain, not discrimination grain.
All categories will of course be fuzzy at the limits of our sensory
resolution capacity. My own grounding hypothesis BEGINS with
bounded sensory categories (modulo threshold uncertainty) and attempts
to ground the rest of our category hierarchy bottom-up on those.

Finally, as I've stressed in responses to others, there's one other
form of category uncertainty I'm quite prepared to concede, but that
likewise fails to imply that category membership is a matter of
degree: All categories -- true graded ones as well as all-or-none ones
-- are provisional and approximate, relative to the context of
interconfusable members and nonmembers that have been sampled to date. If
the sample ever turns out to have been nonrepresentative, the feature-set that
was sufficient to generate successful sorting in the old context must
be revised and updated to handle the new, wider context. Anomalies and
ambiguities that had never occurred before must now be handled. But what 
happens next (if all-or-none sorting performance can be successfully
re-attained at all) is just the same as with the initial category learning
in the old context: A set of features must be found that is sufficient to
subserve correct performance in the extended context. The approximation
must be tightened. This open-endedness of all of our categories, however, is
really just a symptom of inductive risk rather than of graded representations.

>	"Analog" means "invertible". The invertible properties of a
>	representation are those properties which it preserves...[This
>	sounds] tautologically true of *all* representations.

For the reply to this, see my response to Cugini, whose criticism you
cite. Sensory icons need only be invertible with the discriminable properties
of the sensory projection. There is no circularity in this. And in a dedicated
system invertibility at various stages may well be a matter of degree, but
this has nothing to do with the issue of graded/nongraded category membership,
which is much more concerned with selective NONinvertibility.

>	It is quite possible to make all-or-none judgements based on graded
>	features [e.g., thermostats]

Apart from (1) thresholds (which are features, and which I discussed
earlier), (2) probabilistic features so robust as to be effectively
all-or-none, and (3) gerrymandered examples (usually playing on the
finiteness of the cases sampled, and the underdetermination of the
winning feature set), can you give examples?

>	"chair"... with no back... [is a] "stool"... Now vary the size
>	of the back

The linguist Labov, with examples such as cup/bowl, specialized in
finding graded regions for seemingly all-or-none categories.
Categorization is always a context-dependent, "compared-to-what"
task . Features must reliably sort the members from the nonmembers
they can be confused with. Sometimes nature cooperates and gives us
natural discontinuities (horses could have graded continuously into
zebras). Where she does not, we have only one recourse left: an
all-or-none sensory threshold at some point in the continuum. One can
always generate a real or hypothetical continuum that would foil our
current feature-detectors and necessitate a threshold-detector. Such
cases are only interesting if they are representative of the actual
context of confusable alternatives that our category representation
must resolve. Otherwise they are not informative about our actual
current (provisional) feature-set.

>	I see nothing about "all-or-none" categories which is not explainable
>	by arbitrary cutoffs of graded sensory data... [and] avoid[ing]
>	borderline situations.

Neither do I. (Most feature-detection problems, by the way, do not
arise from the need to place thresholds along true continua, but from
the problem of underdetermination: there are so many features that it
is hard to find a set that will reliably sort the confusable
alternatives into their proper all-or-none categories.)
-- 

Stevan Harnad                                  (609) - 921 7771
{bellcore, psuvax1, seismo, rutgers, packard}  !princeton!mind!harnad
harnad%mind@princeton.csnet       harnad@mind.Princeton.EDU

sher@rochester.arpa (David Sher) (07/05/87)

In article <967@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes:
>
>Most of our object categories are indeed all-or-none, not graded. A
>penguin is not a bird as a matter of degree. It's a bird, period. 

Just for the record is this an off hand statement or are you speaking
as an expert when you say most of our categories are all or none.  Do
you have some psychology experiments that measure the size of human 
category spaces and using a metric on them shows that most categories
are of this form?  Can I quote you on this?  Personally I have trouble
imagining how to test such a claim but psychologists are clever
fellows.
-- 
-David Sher
sher@rochester
{ seismo , allegra }!al A: stwter s 

harnad@mind.UUCP (Stevan Harnad) (07/05/87)

In Article 185 of comp.cog-eng sher@rochester.arpa (David Sher) of U of
Rochester, CS Dept, Rochester, NY responded as follows to my claim that
"Most of our object categories are indeed all-or-none, not graded. A penguin
is not a bird as a matter of degree. It's a bird, period." --

>	Personally I have trouble imagining how to test such a claim...

Try sampling concrete nouns in a dictionary.
-- 

Stevan Harnad                                  (609) - 921 7771
{bellcore, psuvax1, seismo, rutgers, packard}  !princeton!mind!harnad
harnad%mind@princeton.csnet       harnad@mind.Princeton.EDU

dmark@sunybcs.uucp (David M. Mark) (07/08/87)

In article <974@mind.UUCP> harnad@mind.UUCP (Stevan Harnad) writes:
>
>
>In Article 185 of comp.cog-eng sher@rochester.arpa (David Sher) of U of
>Rochester, CS Dept, Rochester, NY responded as follows to my claim that
>"Most of our object categories are indeed all-or-none, not graded. A penguin
>is not a bird as a matter of degree. It's a bird, period." --
>
>>	Personally I have trouble imagining how to test such a claim...
>
>Try sampling concrete nouns in a dictionary.

Well, a dictionary may not always be a good authority fro this sort of
thing.  Last semester I led a graduate Geography seminar on the topic:
"What is a map?"   If you check out dictionaries, the definitions seem
unambiguous, non-fuzzy, concrete.  Even the question may seem foolish, since
"map" probably is a "basic-level" object/concept.  However, we conducted
a number of experiments and found many ambiguous stimuli near the boundary
of the concept "map".  Air photos and satellite images are an excellent
example: they fit the dictionary definition, and some people feel very
strongly that they *are* maps, others sharply reject that claim, etc.
Museum floor plans, topographic cross-profiles, digital cartographic
data files on tape, verbal driving directions for navigation, etc., are
just some examples of the ambiguous ("fuzzy"?) boundary of the concept
to which the English word "map" correctly applies.  I strongly suspect
that "map" is not unique in this regard!