Re: Categorisation and Prototypes

From: HARNAD Stevan (
Date: Fri Jun 07 1996 - 19:58:11 BST

> From: "Naden, Christopher" <>
> Date: Mon, 27 May 1996 04:39:43 +0100 (BST)
> Categorization is the practice of sorting things: making groups of
> similare objects.

That's how it works with unsupervised learning, but with supervised
learning it's not just grouping similar objects, it's grouping together
the RIGHT objects (the mushrooms you can or cannot eat; the microscope
slides of cancerous vs noncancerous cells): Their "similarities" may
lead you astray. In supervised learning you have to find the right
features; in other words, the relevant was in which they are similar,
which may not be obvious at all from the "raw" similarities.

> The classical theory of categorzation suggests that
> the way we do this is to learn a list of necessary and sufficient
> features with which to classify categories.

A set of features such that having those features is necessary to be in
the category (no members failt to have them) and sufficient to be in the
category (no nonmembers have them them).

> For example, the features
> of the category mammal are [a] Give birth to live young, and [b] suckle
> their young on milk.

Actually, you should say the features "might be," because I'm not sure
these really are the necessary and sufficient conditions for being a
mammal! Remember, we're not too good at introspecting the features our
mind actually uses.

> Each of these features are necessary, in other
> words anything lacking them is not a mammal, but neither are
> sufficient: in other words BOTH are needed to indicate mammal status,
> one alone will not do.
> The Prototype theory of categorization suggests that instead of this
> feature analysis method, we in fact categorize by learning a
> "prototype", an 'ideal' image of the perfect member of that category,
> against which we match propective members. If, for example, we are
> categorizing a swallow, it is closer to the 'prototype' image of the
> category 'bird' than to the 'prototype' image of the category 'camel':
> thus it gets categorized as a bird.

To keep it parallel, I'd have contrasted it with mammal rather than
camel, which seems more parallel with swallow...

> The evidence for this theory was
> propoed in 1978 buy Eleanor Rosch, who noted that not only do people
> not always know the features by which they classify categories, but
> they find that certain members of categories semm to be 'more typical'
> than others (a swallow seems a more typical 'bird' than and ostrich).
> She also noted that they tend to classify these more 'typical' examples
> quicker than less typical ones.

more quickly: watch grammar and use spell-checker (don't follow my poor
example lately, when I am rushing so much to comment on your 120
questions in time that I have no time to re-read, much less spell-check
my comments!)

> Further evidence for the prototype theory was added by Ludwig
> Witgenstein when cited the example of the category 'game'. It is
> clearly a category, yet no-one can actualy point out the necessary and
> sufficient features of the category.

Not exactly evidence, since Ludwig didn't do experiments... More like
arguments in favour...

> The 'protoype' theory only really applies consistenly to categories
> which are a matter of degree: for example the category 'long'. It would
> be difficult to build up a list of features for this category as it is
> not only continuous but relative; however a prototype against which
> each example could be checked might make this easier to categorize.

Good job, but too literally based on what I said and the examples I
used. For an A, relate it to neural nets, categorical perception,
or the imagery debate.

This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:45 GMT