Feature Detection, Symbolic Rules and Connectionism

[Stevan Harnad]

I am redirecting to connectionists a segment of an ongoing discussion
of categorization on comp.ai that seems to have taken a connectionistic
turn. I think it will all be understandable from context. The issue
concerns whether category representations are "nonclassical" (i.e.,
with membership a matter of degree, and no features that provide
necessary and sufficient conditions for assigning membership) or
"classical" (i.e., with all-or-none membership, assigned on the basis
of features that do provide necessary and sufficient conditions).
I am arguing against the former and for the latter. Connectionism seems
to have slipped in as a way of having features yet not-having them too,
so to speak, and the discussion has touched base with the familiar
question of whether or not connectionist representations are really
representational or ruleful:

anwst at cisunx.UUCP (Anders N. Weinstein) of Univ. of Pittsburgh,
Comp & Info Sys wrote:

" I think Harnad errs... that reliable categorization *must* be
" interestingly describable as application of some (perhaps complex) rule
" in "featurese" (for some appropriate set of detectable features)...
" Limiting ourselves (as I think we must) to quick and automatic
" observational classification... If... the effects of context on such tasks
" are minimal... there must be within us some isolable module which can
" take sensory input and produce a one bit yes-or-no output for category
" membership...  But how does it follow that such a device must be
" describable as applying some *rule*? Any physical object in the world
" could be treated as a recognition device for something by interpreting
" some of its states as "inputs" and some as "yes-or-no responses." But
" intuitively, it looks like not every such machine is usefully described
" as applying a rule in this way. In particular, this certainly doesn't
" seem a natural way of describing connectionist pattern recognizers. So
" why couldn't it turn out that there is just no simpler description of
" the "rule" for certain category membership than: whatever a machine of
" a certain type recognizes?

For the points I have been trying to make it does not matter whether or
not the internal basis for a machine's feature-detecting and
categorizing success is described by us as a "rule" (though I suspect
it can always be described that way). It does not even matter whether
or not the internal basis consists of an explicit representation of a
symbolic rule that is actually "applied" (in fact, according to my
theory, such symbolic representations of categories would first have to
be grounded in prior nonsymbolic representations). A connectionist
feature-detector would be perfectly fine with me; I even suggest in my
book that that would be a natural (and circumscribed) role for a
connectionist module to play in a category representation system (if it
can actually deliver the goods).

To rehabilitate the "classical" view I've been trying to rescue from
well over a decade of red herrings and incoherent criticism all I need
to re-establish is that where there is reliable, correct, all-or-none
categorization performance, there must surely exist detectable features
in the input that are actually detected by the categorizing device as a
("necessary and sufficient") basis for its successful categorization
performance. I think this should be self-evident to anyone who is
mindful of the obvious facts about our categorization performance
capacity and is not in the grip of a California theory (and does not
believe in magic).

The so-called "classical" view is only that features must EXIST in the
inputs that we are manifestly able to sort and label, and that these
features are actually DETECTED and USED to generate our successful
performance. The classical view is not committed to internal
representations of rules symbolically describing the features in
"featurese" or operating on symbolic descriptions of features. That's
another issue. (According to my own theory, symbolic "featurese"
itself, like all abstract category labels in the "language of thought,"
must first be grounded in nonsymbolic, sensory categories and their
nonsymbolic, sensory features.)

[By the way, I don't think there's really a problem with sorting out
which devices are actually categorizing and which ones aren't. Do you,
really? That sounds like a philosopher's problem only. (If what you're
worried about is whether the categorizer really has a mind, then apply
my Total Turing Test -- require it to have ALL of our robotic and
linguistic capacities.) Nor does "whatever a machine of a certain type
recognizes" sound like a satisfactory answer to the "question of how in
fact our neural machinery functions to enable us to so classify
things." You have to say what features it detects, and HOW.] 

[Related to the last point, Greg Lee (lee at uhccux.uhcc.hawaii.edu),
University of Hawaii, had added, concerning connectionist
feature-detectors: "If you don't understand how the machine works, how
can you give a rule?" I agree that the actual workings of connectionist
black boxes need more analysis, but to a first approximation the answer
to the question of how they work (if and when they work) is: "they
learn features by sampling inputs, with feedback about
miscategorization, `using' back-prop and the delta rule." And
that's certainly a lot better than nothing. A fuller analysis would
require specifying what features they're detecting, and how they
arrived at them on the available data, as constrained by back-prop and
the delta rule. There's no need whatsoever for any rules to be
explicitly "represented" in order to account fully for their success,
however. -- In any case, connectionist black boxes apparently do not
settle the classical/nonclassical matter one way or the other, as
evidenced by the fact that there seems to be ample room for them in
both nonclassical approaches (e.g., Lakoff's) and classical ones
(e.g., mine).]

" [We must distinguish] the normative question of which things are
" *correctly* classified as birds or even numbers, and the descriptive
" question of how in fact our neural machinery functions to enable us to
" so classify things. I agree also with Harnad that psychology ought to
" keep its focus on the latter and not the former of these questions.

A kind of "correctness" factor does figure in the second question too:
To model how people categorize things we have to have data on what
inputs they categorize as members of what categories, according to what
constraints on MIScategorization. However, it's certainly not an
ontological correctness that's at issue, i.e., we're not concerned with
what the things people categorize really ARE "sub specie
aeternitatis":  We're just concerned with what people get provisionally
right and wrong, under the constraints of the sample they've
encountered so far and the feedback they've so far received from the
consequences of miscategorization.

I also see no reason to limit our discussion to "quick, automatic,
observational" categorization; it applies just as much to slow
perceptual pattern learning and, with proper grounding, to abstract,
nonperceptual categorization too (although here is where explicitly
represented symbolic rules [in "featurese"?] do play more of a role,
according to my grounding theory). And I think context effects are
rarely "minimal": All categorization is provisional and approximate,
dependent on the context of confusable alternatives so far sampled, and
the consequences (so far) of miscategorizing them.

Stevan Harnad    harnad at confidence.princeton.edu   harnad at pucc.bitnet