On Thu, 18 May 2000, Andras Lorincz wrote:
> >la> How about the first two words learned, say 'dad' or 'mam'?
> >la> BTW, my son is different. In his case the first word was 'auto'
> >la> and he meant it.
>
> sh> Dad, mum and auto are trivial categories. Try chicken-sexing, or
> sh> naming all the people you've seen...
>la> As far as I see you define two types of categories:
>la>
>la> (1) you can tell the difference based on special higher order features
>la>
>la> (2) you can separate individual elements of the categories but can not
>la> "tell" the difference
>la>
>la> Under these conditions you say (??) that (2) requires "toil" and can
>la> not gain anything by "theft".
I am afraid you are conflating two things:
(a) acquiring lower-order/higher-order categories,
(b) acquiring categories by (sensorimotor)theft/(symbolic)toil,
Sensorimotor categories can be both lower-order and higher-order.
(Categorization is a hierarchy of abstraction.)
I might learn the lower-order category "white" and "red-dotted"
directly (= by toil), by trying to point to and identify "white" and
"red-dotted" things, and being fed whenever I get it right and starved
whenever I get it wrong.
I might acquire the higher-order category "edible mushroom" directly (=
by toil) too, by trying to eat different kinds of mushrooms, getting
sick when I eat the wrong kind, and eventually learning which kinds
are edible.
Alternatively, I might acquire the higher-order category "edible
mushroom" indirectly (= by theft) by being told that the edible kind are
the kind that are white and red-dotted.
So it is not the degree of abstraction that makes a category symbolic,
but whether it has been acquired directly, via sensorimotor
learning (toil), or indirectly, via symbol strings (theft).
But for the latter, the symbols in the symbol strings have to be
grounded, either directly (toil) or indirectly (theft).
And it cannot be theft all the way down.
Nor can it be toil all the way up (what is the direct sensorimotor
grounding of "true" or "good"), although it must all be grounded in toil.
>la> In my view the point is that if we work with two categories then there
>la> is no difference between trial-and-error or naming. We need more than
>la> two categories to see the difference.
Dichotomous categories are the simplest kind (and if we consider any
category plus its complement, it is dichotomous -- e.g., all animals
other than giraffes are non-giraffes), but most nontrivial categories
are neither dichotomous nor unidimensional. There are many kinds of
things that a thing can be, and each kind of thing it is is determined
by many features.
There is no opposition between trial-and-error and naming. A "name" is
any differential (usually also arbitrary) response I perform to
differentiate different kinds of things; naming them is one such
response, eating them is another, but either we, I first have to be able
to tell them apart. And learning to tell them apart (i.e., to sort and
identify them with a differential response) is what category learning is
about.
It is uninformative to think of category learning as merely the
association of well-differentiated objects with their names. For, that,
we may as well not talk about categories (= kinds), or category-learning
problems at all, and simply talk about trivial pairwise association of
objects (some of which happen to be called "names," but all of them
fully differentiable and differentiated).
Otherwise put, in nontrivial category learning, it is the trial and error
that eventually makes the category learner able to associate the name
with the (kind of object). Until then, he doesn't know which is which.
I took your point to be about whether mere pairing (or repeated pairing)
of the (kind of) object with its "name" (or correct response) is enough
for you to learn to name them on your own. And I was suggesting that in
a nontrivial problem this would not be so. Mere passive exposure and
pairing is not enough; active sensorimotor trial-and-error, guided by
feedback from the consequences of miscatgorization, are necessary in
order eventually to find the critical invariant features on the basis of
which error-free categorization (differential responding, naming) is
THEN possible.
>la> Assume that we have a space with parameters and according to the
>la> parameters we have four different categories with sharp decision
>la> surfaces.
You last me in this transition from I/O performance to model. So far,
categories are kinds of objects. What are "decision surfaces"? A
mechanism inside my head "decides" for me whether this mushroom is or is
not edible (and if it has learned correctly, it makes the right decision
for me). But it is still quite unspecified HOW it does so; presumably by
having learned the right invariant features. But how did it do that?
Jumping to a "decision boundary" here sounds rather premature; at the
least, the model in question has to be specified.
>ls> One person is given trial-and-error feedback (whether he has
>la> chosen the good category or not), and another person is given the name
>la> of the category (this latter corresponds to your example with the
>la> students and the teacher). Who will be faster in learning the proper
>la> categorization?
If it is a trivial category (like black and white), there will be no
difference after a few trials. If it is nontrivial, the active
corrective-feedback learner will be faster, the learner passively given
the name with the category will either learn much more slowly or (if the
critical features are sufficiently underdetermined) may not learn at
all.
--------------------------------------------------------------------
Stevan Harnad harnad@cogsci.soton.ac.uk
Professor of Cognitive Science harnad@princeton.edu
Department of Electronics and phone: +44 23-80 592-582
Computer Science fax: +44 23-80 592-865
University of Southampton http://www.cogsci.soton.ac.uk/~harnad/
Highfield, Southampton http://www.princeton.edu/~harnad/
SO17 1BJ UNITED KINGDOM
This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:07 GMT