Re: Categorisation and Prototypes

From: HARNAD Stevan (
Date: Tue Jun 04 1996 - 22:15:33 BST

> From: "Hooper, Molly" <>
> Date: Fri, 24 May 1996 16:05:58 GMT
> Categorisation, similarly classed as categorical perception,

No, categorisation and categorical perception are not the same thing.
Categorisation is sorting things by making one, response to one kind of
thing and another to another kind; categorical perception is the
increase in similarity between members of the same category and the
decrease in similarity between members of different categories that
sometimes occurs when categories are learned, or are inborn.

> as John
> R. Anderson in his book "Cognitive Psychology and its Implications" of
> 1995 states, refers to the perception of stimuli as belonging in
> distinct, discreet categories, and, the failure to perceive the
> graduations among stimuli present within a particular category. This
> classical view of Categorisation

This is not the classical view of categorisation. To know that, you
would have to come to lectures or do the skyreading; the glossary of the
Anderson book is not a good way to find these things out.

> can be explained further by
> exploring investigations conducted in 1969 by Allan Paivio who
> studied "Mental Representations" in conjunction with a dual coding
> approach and produced some remarkable insights. It appears that
> Categorisation occurs when the visual and auditary systems of an
> individual is organised by the mind into discreet and completely
> distinct categories whose members seem to resemble one another more
> so than they would resemble those members present in other
> categories. An example of this phenomenon can be found in Stevan
> Harnad's book of 1990 "Categorical Perception: The groundwork of
> Cognition" where he examines colour categories. Harnad argues that
> colours differ, physically , only in their wavelengths, which
> gradually become shorter across the spectrum of visible colours.
> However, the individual person will perceive qualitative changes from
> red to orange to yellow to green, and so on.

This would have been a fine answer if the question had been about
categorical perception, but it was not. It was about the classical view
of categorisation, which is that things are sorted into categories on
the basis of features shared by members of the same category, and not
shared with members of different categories. Together, such features
provide conditions that are NECESSARY if one category is to be
successfully told apart, and SUFFICIENT for doing so successfully.

> Such prompt
> superordinate classification suggests that abstract information is
> stored in organised categories, as it has been proved even in very young
> children.

This is again a non sequitur: What has it to do with what came before
and after, and with the question asked?

> Similarly, this process is true of appreciating musical
> pitches where slight increasing changes of frequency begin to be heard
> as categorical changes from C to C sharp to D to E flat. In each
> particular case, discreet specifics are quickly subordinated to
> broad categorical criteria , envelopes which subsume constant common
> attributes. This conception supposes that related events are grouped
> into concept ladders, and maintained with the aid of mnemonic summary
> codes. These "perceptual boundaries" have evolved along a physical
> continuum, separating it into discrete regions "with qualitative
> resemblances within each category and qualitative resemblances
> between them." Therefore, this classical view of Categorisation,
> incorporating these bounded categories, may indeed provide the
> groundwork for our higher-order cognition and language.

Unfortunately, the concepts that are put together here do not produce an
explanation that would make sense to kid-sib; nor do they answer the
question. Please read the other skywriting on this question, and the
skyreadings too. Anderson will not be sufficient. There IS a passage in
the Harnad chapter that answers this question; If you searched on the
word "classical" using "find" in netscape, as I had suggested, you would
have found it -- and been better off than using the index of the
Anderson text, which the pointers did not point to in the Archive.
Here is the relevant passage from :

    [footnote start] It seems to be a point of logic rather than one of
    theoretical preference that if a categorizer is able to perform
    error-free categorization then that performance must be based on
    detecting and using some set of features that is sufficient to
    serve as a basis for the successful categorization (though not
    necessarily necessary or exhaustive, for, especially with
    underdetermination, there might be other features that would
    suffice too). The putative alternatives to the classical
    necessary/sufficient-features approach to categorization --
    originating with Rosch (Rosch & Lloyd 1978) and attributed to
    Wittgenstein (1953) -- seem to be based on confusions among the
    following additional (and independent) factors:

    (i) Some categorization is not all-or-none; there may be no X's,
    just things that are X to greater or lesser degrees (e.g., the
    category big).
    (ii) Some categorization performance may not be reliable; subjects may
    sometimes miscategorize, or there may be some instances whose
    membership is uncertain or graded or probabilistic (e.g., the category
    (iii) The subject may not be aware of the features he is using; the
    ones he verbalizes may indeed be neither necessary nor sufficient, but
    then they're not the ones he's using.
    (iv) There is an element of arbitrariness in what one does and does not
    choose to call a feature (as opposed to a metafeature); there is no
    logical or practical reason why features cannot be disjunctive,
    negative, conditional, relational, polyadic or probabilistic -- or even
    derivable only by complex computational, constructive, algorithmic,
    propositional or model-driven processes -- as long as they are grounded
    in reliable, detectable invariant properties of the instances being
    categorized and they are sufficient to subserve successful

    Hence, at least insofar as our reliable, overlearned, all-or-none,
    bounded categories are concerned -- and these are the categories
    (e.g., bird and pet) that tend to be used in the experiments
    stimulated by Rosch's work -- both the existence and the use of
    (singly) sufficient (and disjunctively necessary) sets of features
    seems inescapable. The origin of the putative alternatives to this
    -- non-necessary/sufficient prototypes and family resemblances --
    seems to be attributable to a focus on typicality judgments and
    reaction times rather than categorization per se, together with a
    reliance on the subject's (and perhaps the experimenter's)
    introspections as to the basis for the categorization. The real
    basis for categorization can only be found by inference, as tested
    by models that attempt to generate reliable categorization
    performance when confronted with the same instances that subjects
    can categorize successfully. [footnote end]

This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:44 GMT