Re: Cangelosi/Harnad Symbols

From: Stevan Harnad (harnad@coglit.ecs.soton.ac.uk)
Date: Wed May 17 2000 - 21:26:58 BST


On Wed, 17 May 2000, Lorincz, Andras wrote:

> sh> For toilers, the knowledge is
> sh> grounded in having learned the features of the edible mushrooms; so the
> sh> toilers can forage alone; for the thieves, the only "feature" of an
> sh> edible mushroom is the vocalization of the toilers: "edible". So when
> sh> there are no more toilers, the thieves cannot eat.
>
>la> I buy this.
>
> sh> The solution is always to have the bottom-level categories learned by
> sh> toil, by everyone. The categories acquired by "theft" need to be
> sh> higher-order categories, describable by (e.g., boolean) combinations of
> sh> the names of the lower-order categories, as in the Cangelosi/Harnad
> sh> paper.
>
>la> I do not see this if the thieves can "associate" other features to
>la> the vocalization of the toilers: "edible". On the other hand, if
>la> "association" qualifies as "toil" then I see it. However, this seems
>la> to contradict with your final conclusion:

No contradiction. "Toil" is learning the hard way, by active trial and
error, with corrective feedback (from the consequences of making
errors). The toilers learn by trying to eat mushrooms, and getting sick
if they eat inedible ones (or going hungry if they fail to eat edible
ones).

They thereby learn (the hard way) the mushroom-features (sensory) that
distinguish the edible and inedible ones.

As long as there are toilers around who vocalize (correctly), the only
"feature" a mushroom needs to have in order to make it edible is the
auditory feature, which is a toiler nearby, vocalizing "edible." Take
away the toilers and you take away that feature.

Now you are right that in eating the mushrooms that have the (auditory)
feature "edible" and not-eating the mushrooms that lack the (auditory)
feature "edible," it is possible that the real (visual) features are
being passively learned by differential association alone, and with
very easy features (such as black/white), that might be enough -- but
for those trivial feature-learning problems, there is not much
difference between toil and theft in the first place.

For nontrivial features, requiring nontrivial learning (e.g.,
chicken-sexing), active trial-and-error toil (reinforcement learning)
and feedback are needed, and not mere passive exposure and
association.

By way of an intuitive example, if a school-teacher had 30 students in
every class, and taught 7 different classes a day (and never for longer
than a few months), and every child every day wore his name written on
his forehead -- and every day the teacher had to interact with each
child several times by name -- what do you think would be the "hit
rate" of that interaction if, on the last day of class, all the
students came in without their forehead-labels?

That hit-rate would reflect the strength of passive association. (I
don't think it would be very high! Facial features are hard enough to
learn WITH active trial-and-error toil and feedback; passive association
would not stand much of a chance.)

Not zero chance, I agree; but in the context of what it would take to
sustain language, with nontrivial features and and categories to learn,
close enough to zero as to be a nonstarter (in my opinion).

(The chicken-sexing counterpart would be to sit passively by the
assembly-line watching the grand-master examine the chickens and say
whether they were male or female, for several months, and then to try it
on their own -- compared to having done it on their own for all those
months, and having gotten the feedback from the grandmaster each time
only AFTER having tried to say which was male and which was female
themselves.

> sh> The solution is always to have the bottom-level categories learned by
> sh> toil, by everyone. The categories acquired by "theft" need to be
> sh> higher-order categories, describable by (e.g., boolean) combinations of
> sh> the names of the lower-order categories, as in the Cangelosi/Harnad
> sh> paper."
>
>la> unless you allow (as a special case) that "lower-order categories" could
>la> be raw sensory information and not more.

I don't know what "raw sensory information" means; but what I mean here
is certainly sensory features (of the object -- not of its associated
label: it is the label we are actively trying to learn to apply to the
right objects).

Describing something (with words/symbols) only helps if you already
have grounded SOME symbols, at least, in something other than just more
descriptions. Otherwise it's all just meaningless squiggles and
squoggles. (That's the symbol grounding problem.)

But I don't think mere passive exposure and association is enough to
ground symbols either.

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