Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!wuarchive!zaphod.mps.ohio-state.edu!uakari.primate.wisc.edu!aplcen!jhunix!ins_atge
From: ins_atge@jhunix.HCF.JHU.EDU (Thomas G Edwards)
Newsgroups: comp.ai
Subject: Re: Against AI
Message-ID: <6526@jhunix.HCF.JHU.EDU>
Date: 27 Sep 90 02:15:57 GMT
References: <94y2P6w163w@bluemoon.UUCP>
Reply-To: ins_atge@jhunix.UUCP (Thomas G Edwards)
Organization: The Johns Hopkins University - HCF
Lines: 53

In article <94y2P6w163w@bluemoon.UUCP> feedback (Bryan Bankhead) writes:
>        1/ Intellingence occurs at the wrong hierarchial level 
>                        We may not be capable of programming what is going 
>on in our minds because intellingence is produced at levels our software 
>is not capable of "backstrapping" (a term I just coined) itself to, that 
>is our software may not be capable of instantiating our 'other order' 
>operations in the form that our 'awareness' is capable of processing.

I think we are on fairly safe ground if we claim cognitive processing
done by "symbolic AI" is not what is done on the ground of neurons.
The big question is whether or not processing done by "symbolic AI"
is similar to what is being done by millions of neurons.

I have experience with artificial neural networks.  ANN's have developed
to the point where we can solve a good number of fairly small
"toy" problems, and even some "slightly-less-than-toy" problems.
The problem with neural networks has always been one of scaling.
Big, complex problems are still not easily tackled by ANN's.
However, the problems they can solve by induction still seem very
neat, especially compared to "symbolic AI."  For example, an ANN
can learn to perform a pole-balancing task in a fairly short time
using reinforcement learning. 

There are some things which "symbolic AI" does really well.  For instance,
MACSYMA can do all kinds of symbolic algebra, partial fraction
expansion, taking derivatives, integrals, Laplacian transforms, etc.
Many "symbolic AI" programs use deductive reasoning to solve their
problems (i.e. build up to a goal using known subgoals).

I think that more impressive cognitive systems will eventually be built
utilizing both of these modalities.  Sub-goals can be inductively 
learned using neural networks.  These sub-goals will be farly simple
problems.  The networks will then be strung together in methods
similar to traditional symbolic structures to solve large difficult
goals.  The heuristics of goal solving from "symbolic AI" will be
helped by neural networks building sub-goal networks using
inductive-style learning.

Of course, I could be totally wrong.

>There are known examples of this in computer sci..  for instance it has 
>been proven that von nueman programming constructs are incapable of 
>determining if they themselves are self terminating.

This is kind of a red herring.  I could easily write a program which
could say whether it is self-terminating.  It would say "yes."
Not a "theoretical" program, but a real one running on a real machine.
It has a very high percentage chance of being right (especially
if I don't pay the electric bill :-)
People determining whether they are self-terminating stand about as
good a chance.

-Thomas Edwards