Xref: utzoo comp.ai:3393 sci.lang:4114 Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!ncar!boulder!sunybcs!rutgers!elbereth.rutgers.edu!harnad From: harnad@elbereth.rutgers.edu (Stevan Harnad) Newsgroups: comp.ai,sci.lang Subject: Re: Categorization Summary: On ad lib vs imposed categorization, Watanabe, the frame problem Message-ID: Date: 18 Feb 89 21:36:36 GMT References: <15585@cisunx.UUCP> <3200@uhccux.uhcc.hawaii.edu> <1435@client1.dciem.dnd.ca> Organization: Rutgers Univ., New Brunswick, N.J. Lines: 111 mmt@client1.dciem.dnd.ca (Martin Taylor) of D.C.I.E.M., Toronto, Canada writes about Watanabe's "Ugly Duckling Theorem," quoting: " "When we employ a concept, we usually understand that there is a " group of objects corresponding to this concept that any two " members of the group resemble each other more than a member and " a nonmember [but]... from the formal point of view there exists " no such thing as a class of similar objects in the world, " insofar as all predicates (of the same dimension) have the same " importance" (Watanabe 1969). " " [C]ategories cannot be logically derived as groups of objects sharing " features, but (as I interpret Harnad) as groups of objects toward which " actions have common consequences, and have been found to have common " consequences in past experience... [But] We have far more classes than " those composed of objects with which we have interacted. Most " categories are determined linguistically, by mutual agreement... Harnad " insists... someone [can have] "MIScategorized" something, as if the " category existed outside the linguistic agreement or the feedback from " experience. Earlier, he also used the term "natural kinds"... [B]oth " these usages assert a kind of... universe, in which a God has " prescribed some knowable structure; but we could not know such a " universe... We can know (and categorize) only what we can sense, " derive, and discuss. And in those categories there can only be grades " of usefulness, never error. Let me clarify some misunderstandings: (1) I have been writing AGAINST, not FOR the ontological (or "God's Eye") view of categories. Our internal representations are our provisional bases for sorting and labeling inputs based on what we've encountered so far, as guided by feedback from the consequences of mis-sorting and mis-labeling. Although this categorization mechanism may be converging on what things "really are," we have no way of knowing this. At best, our sorting and labeling is an approximation to reality. I AM a realist, though: There are things out there. I prefer to avoid ontological questions, however, because they are simply irrelevant to modeling human categorization. Our "errors" are determined relative to their pragmatic consequences, not an omniscient or ontological criterion. Forget about ontology; this discussion is only about whether or not our internal representations of categories are "classical," i.e., whether they are based on detectable features that provide conditions that are necessary and sufficient to guide our correct, all-or-none sorting/labeling performance. I'm claiming that they are indeed "classical," and that there's absolutely nothing wrong with the "classical" view that Rosch and Lakoff [and perhaps Wittgenstein] are widely interpreted as having invalidated in favor of "prototypes," family-resemblances, exemplars, or any other form of graded, hence "nonclassical" representation. (2) It doesn't matter what the source of the feedback about MIScategorization is, just as long as it comes from "out there." (Purely subjective "categories" would be susceptible, for example, to Wittgenstein's argument against "private language.") Feedback from a teacher or a parent or from the "linguistic community" to the effect that you have mis-labeled something as a "mushroom" is just as good as feedback from stomach cramps. In the end, unless everyone miraculously shares a purely subjective delusion, the "linguistic community's" own labeling of mushrooms will have to be guided by detectable features of mushrooms. (3) Watanabe's theorem is relevant to what I've called "ad lib" similarity judgments, where one sorts as one pleases, guided only by how similar things "look." He was right that in this case we are being guided by "weights" on a subset of an infinity of inherent similarities and dissimilarities (the weight being either arbitrary or governed by existing NON-ad-lib categories we already have). But I have stressed that models for categorization should not focus on ad lib similarity judgments but on IMPOSED categorization tasks, the ones with feedback from the consequences of miscategorizing. Apart from our arbitrary subjective "categories," all of our categories are of the imposed kind, and certainly most of the linguistic ones -- the ones that are labeled by the words in our vocabulary -- are imposed categories. It's precisely their imposed (i.e., constrained) as opposed to ad lib nature that makes these categories immune to Watanabe's theorem: For a "classical" basis for the correct sorting and labeling of the inputs must be SELECTED out of the Watanabean confusion matrix whenever the categorization successfully converges. Otherwise there's no way to explain our success! Watanabe's theorem is also related to (mostly irrelevant red herrings) associated with Hempel's Raven paradoxes (of theory confirmation: does a white swan confirm that all ravens are black?) as well as problems variously labeled as the "frame" problem (how to specify in general what remains invariant in a change of circumstances under a symbolic description?) and the "credit assignment" problem (when a set of formerly reliable features fails, i.e., miscategorizes, how is one to determine which features are to blame, or which features deserve the credit for restoring successful categorization?). These in turn are related to what I've called the "symbol grounding problem" (how are the meanings of symbols to be grounded in something other than just more meaningless symbols?), and ultimately to the problem of underdetermination: How do you pick out the "right" features to solve a categorization problem when there are so many confusable candidates? Which brings us right back to the problem of category learning and representation, guided by feedback from "error." So the take-home message is this: The ad lib similarity structure of a huge Laplacean feature matrix will not give you categorization, but a "classical" subset selected on the basis of feedback from miscategorization will. -- Stevan Harnad INTERNET: harnad@confidence.princeton.edu harnad@princeton.edu srh@flash.bellcore.com harnad@elbereth.rutgers.edu harnad@princeton.uucp BITNET: harnad@pucc.bitnet CSNET: harnad%princeton.edu@relay.cs.net (609)-921-7771