Re: mind and body

Ricardo Ribeiro Gudwin (gudwin@DCA.FEE.UNICAMP.BR)
Wed, 3 Jun 1998 16:12:40 -0300


John J. Kineman wrote:

> I am actually quite surprised to read from Ricardo's comments
> that AI programmers have not incorporated this need for context or "Umwelt."

This is not actually 100% correct. What happens is that usually, AI tries to
modelthe higher capabilities of human reasoning i.e. logic and inference
capabilities, but
in a level that it is disconnected of real environment. You can not simply show
a chair
and say to a computer: "chair" !. Usually, the computer does not share the same
perceptions we have. The question of context is a little bit different, because
we
are able to set a context, in logic terms, and still stay away from sharing
perceptions.
For example, I can say to a computer:
man(X) -> mortal(X)
Then I set a context:
man(Socrates)
Based on this, the computer is able to know that mortal(Socrates).
But it is not able either to know what is "Socrates", or either "mortal" else
"man", and
relate those concepts to the real world. This is the symbol grounding problem.
The whole area of Computational Intelligence grew with the aim of solving the
symbol
grounding problem. But this is not completelly solved. There is actually a lot
of people
worried with the symbol grounding problem in AI. But all the temptatives to
solve this
problem ignoring the question of "Umwelt" are destinated to fail. One of these
temptatives
tries, e.g. to grow a symbol grounding, relating terms to other terms. For
example, ... ,
in order to understand what is "hot", I associate "hot", e.g. with "sun",
"warm", "fire", "heater",
"friction", etc, etc. Maybe, I can oppose it to "cold", "ice", "refrigerator",
"snow", etc. But this
is not a real symbol grounding. This is only the "dictionary game" (for example,
I get
a word in the dictionary: "swangtem" - swangtem is a set of gimksoilers. Then I
go to
"gimksoiler": gimksoiler is a member of a swangtem). This is like when we try to
make
a blind person understand what is "red". We may associate "red" with hot, with
blood,
with the comunists, but even if we make all possible associations, we are not
able to make
a blind person understand what really is "red", because he is not able to share
this perception
with us. One of the most successfull tries, from the computational viewpoint, in
solving
this problem is due to the fuzzy sets theory. Fuzzy sets theory allows us to
start growing
the symbol grounding. But unfortunately, fuzzy sets are good only for grounding
adjectives.
For example, if we measure the wave length of a light, we are able to detect and
classify,
for example if something is "red". Even for concepts that require a more
contextual
information, like judging something "small" or "large". But we are not still
able to detect if
something is "a chair", or if something is "dropping", or even if "a given chair
is dropping down".
A promise for this comes from the Computational Semiotics theory, where things
like that
are investigated more deeply. Unfortunately, Computational Semiotics is still in
its childhood.

> Is it true that AI hasn't embraced this requirement? If so, what is the
> alternative philosophy?

As I said, Computational Intelligence (which includes fuzzy sets, neural
networks and
evolutionary systems) are a good try, but they didn't solve the problem at all.
A possible
alternative comes from semiotics, and its computational version, Computational
Semiotics,
that tries to analyze such matter in a more formal way. Other good tries are the
outline
for a theory of intelligence, from J. Albus and the situated control and
semiotic analysis
from A. Meystel and D. Pospelov.

--
                                                   //\\\
                                                   (o o)
 +-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-oOO--(_)--OOo-=-=-+
 \                   Prof. Ricardo Ribeiro Gudwin                /
 /             Intelligent Systems Development Group             \
 \    DCA - FEEC - UNICAMP    |           INTERNET               /
 /     Caixa Postal 6101      |     gudwin@dca.fee.unicamp.br    \
 \   13081-970 Campinas, SP   |       gudwin@fee.unicamp.br      /
 /          BRAZIL            |      gudwin@correionet.com.br    \
 +-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-+
 \ URL:        http://www.dca.fee.unicamp.br/~gudwin/            /
 / Telephones: +55 (19) 788-3819 DCA/Unicamp (University)        \
 \             +55 (19) 254-0184 Residencia  (Home)              /
 / FAX:        +55 (19) 289-1395                                 \
 +-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-+