Re: DM: neural/discrete

[e sommer]

At 09:16 AM 3/25/98 +0000, Omer F Rana wrote:
>Hi, 
>
>I was wondering if anyone has come up with using neural networks
>for discrete data - such as the one generally found in databases?


i prefer to think of the dichotomy as being between
symbolic and sub-symbolic representations. a lit
search of those two terms might already turn up
a few interesting papers.

more specifically, Jude Shavlik and his students
have made a career of this topic, mapping symbolic
input into a nn, and even trying to retranslate the
trained net back into a symbolic representation.
the excercise always struck me as slightly l'art pour l'art,
but they have published a wealth of paper, so there's
much to read for you there.

the one ref i have at hand is from AAAI'90:
Towell, Shavlik, Noordewier: 
Refinement of approximately correct domain theories
by knowledge-based neural networks.


there was, at least one workshop, " Integrating ... Neural .. 
Symbolic .."
at AAAI-92. 


F. Kurfess (then) at the U of Ulm, Germany, has done some
more recent work, but i don't know the current status of that.

my favourite paper on the topic, though more philosophical,
is by Smolensky: Connectionist AI, Symbolic AI and the Brain.
Artificial Intelligence Review, 1:95 - 109, 1987.

fun!
eddi




--
 alive in the superunknown
                        -- soundgarden