Re: Aaron Sloman's "The Irrelevance of Turing Machines to AI" article

From: Sergio Navega (snavega_at_intelliwise.com)
Date: 08/04/04 

Date: Wed, 4 Aug 2004 10:44:05 -0300

"Traveler" <traveler@nospam.com> escreveu na mensagem
news:lm9vg0hktlcupsefrf9dgc3tmnstrlufs3@4ax.com...
> In article <410fa4ce$1_3@news.athenanews.com>, "Sergio Navega"
> <snavega@intelliwise.com> wrote:
>
> >I agree, but there's an important point to be made. The domains where
> >children usually refine their discrimination abilities are linked to
> >conceptual categories (as opposed to perceptual categories).
> >Successive conceptually refined categories (such as "edible stuff",
> >and then "food", later "liquid food" and finally "milk") are categories
> >that are refined mostly because of top/down processes. The perception of
> >stimuli in these cases may be the same for an adult and an infant,
> >but the former has developed these categories while the latter hasn't.
> >Perceptual categories (or bottom-up categories) are those which seek
> >for similarity and clustering based on the raw elements captured by
> >the senses. After some time, our brain becomes unable to discriminate
> >all the sort of things that an infant appears to be discriminating
> >(the example of the japanese children is evidence of such a thing).
>
> Three questions:
>
> 1. What makes you believe that conceptual categorization is a top/down
> process?

That almost seems to be the definition of it. Top-down processes
involve the influence of previous abstractions in the perception of
new stimuli. These previously generated abstractions help (or even
"compel") the "separation" of classes of things.

>
> 2. Whether or not categorization is a top-down process, have you given
> any thought as to the nature of the mechanisms thatmay be responsible
> for categorization, both conceptual and perceptual?

That is, by far, the most important question of all. I could say that
once we understand how this happens, we will have almost half of our
entire abstract model of the mind. There are several theories, but I
tend to stick with the ones which propose to compute the similarity
of two instances based on some operation done with the features that
characterize them. An apple is similar to another because they
"almost" share the same set of features. Theories differ as to how
they select the features and how to "weigh" them (some features are
more relevant than others, etc; to further complicate things, this
is often variable with context).

>
> 3. What is the essential difference, in your view, between conceptual
> categorization and perceptual learning? I realize that you believe one
> is bottom-up and the other is top-down but what is the reasoning
> underlying your stance?

Conceptual categorization involves the effect of previously acquired
abstract categories over what is perceived. This influence can also
(and is often) influenced by linguistic input. Even children are
subject to this effect. Purely perceptual learning, on the other
hand, is much more dependent on the statistical properties of
the stimuli (such as what happens during the learning of phonological
characteristics of one's native language). But this is not as clear-cut
as it seems. For instance, the learning of whether a tomato is ripe
or not one have mixtures of those processes (and things get much more
complicated when one takes the word "ripe" in the way a famous chef does).

Sergio Navega.

Relevant Pages

  • Tverskys formula for similarity
    ... the problem of categorization as done by humans is very ... > be used as the basis for categorization: ... = Similarity of instance I compared to prototype P ... f= function of the features not in common between P and I ...
    (sci.cognitive)
  • Re: Organizing my work
    ... features a TreeView of categories and automatic item categorization! ... Although it's shareware, I am allowing free registration for this first ...
    (microsoft.public.office.misc)
  • ANN: New WordStat 5.0
    ... features that dramatically improve the flexibility of processes for text ... WordStat 5 delivers improved performances and ... - Automatic document classification using Naive Bayes and k-Nearest ... developer's kit for the integration of WordStat categorization and ...
    (sci.lang)
  • Re: Tverskys formula for similarity
    ... > f= function of the features not in common between P and I ... > similarity between an instance and a prototype is directly proportional ... > to the quantity of common features between I and P and is reduced ... > theory nowadays is related to the exemplar account of categorization, ...
    (sci.cognitive)
  • Re: Aaron Slomans "The Irrelevance of Turing Machines to AI" article
    ... >children usually refine their discrimination abilities are linked to ... >Perceptual categories (or bottom-up categories) are those which seek ... What makes you believe that conceptual categorization is a top/down ...
    (sci.cognitive)