What have neural networks achieved?

[Michael Arbib]

The following extract from the debate on what AI has achieved may be of
interest to connectionists engaged in the present discussion, and in the
one on
symbolic representation.

One might ask:  Does AI need neural networks to be really successful?


Subject: Re: What have neural networks achieved?
>Date: Wed, 2 Sep 1998 20:50:12 -0700 (PDT)
>From: John McCarthy <jmc at Steam.Stanford.EDU>
>Subject: challenges to AI
>
>Paul Rosenbloom asks David McAllester:
>
>   Can you elaborate a bit on what
>   you would find necessary before you would see "any evidence for the
>   feasibility of the grand goals"?
>
>David McAllester replied:
>
>     Some convincing ability to discuss, say, daily events in the
>     life of a child.  A human-level theorem proving machine for
>     UNRESTRICTED conceptual mathematics would also be evidence
>     for me.  I am quite familiar with the current state of the
>     art in theorem proving and I feel like I have climbed a tree
>     while staring at the moon.  While current formal methods do
>     have applications, I do not believe that current
>     applications constitute evidence for the grand goals.  It
>     seems popular among AI researchers to take a "sour grapes"
>     attitude toward grand-goal problems like the Turing test and
>     human-level understanding of general conceptual mathematics
>     --- "oh those can't be done but they're not important
>     anyway".
>
>       David
>
>David McAllester's complaints are similar in some respects to those
>Hubert Dreyfus and Lotfi Zadeh presented at the recent Wonderfest in
>Berkeley.  The difference is that McAllester knows a lot more about
>AI.  At the Berkeley meeting, I tried to get Dreyfus and Zadeh to say
>what was the *easiest* task they considered infeasible to AI.  After
>some discussion, this came down to what task would not be accomplished 
>in the next ten years.  I think I got something out of each of them.
>
>Dreyfus gave the example of "Jane saw the puppy in the window of the
>pet store.  She pressed her nose against it."  The problem is to
>get the referent of "it" to be the window in an honest way, i.e. not
>building in too much.  This requires returning to the task of using
>world knowledge to get the referents of pronouns.  Mere scripts
>wouldn't do it.
>
>Zadeh's example was to get a car out of a parking garage with columns
>and lots of other cars some of which would have to be moved.  That one 
>doesn't seem very hard.  Zadeh thinks AI without fuzzy logic
>won't work.
>
>I have some sympathy with McAllester's point of view.  Much
>present work in AI uses methodologies that are limited in what
>they can ultimately do.  I discuss this in my
>
>FROM HERE TO HUMAN-LEVEL AI
>
>http://www-formal.stanford.edu/jmc/human.html.
>
>That paper lists some of the problems that must be solved to
>reach human-level AI.  I reread it and plan to improve it.
>
>I would like David McAllester to list some of the problems that
>he sees.  If there are any relatively concrete problems that he
>thinks can't be solved in the next ten years, this would serve as 
>a worthwhile challenge to AI research.  He should list the
>*easiest* problem he thinks won't be solved in ten years.
>
>It hasn't helped AI much to be challenged only by the ignorant,
>but McAllester isn't ignorant, so his challenges, if he can make
>them more concrete than in his message, will be helpful.
>
>To give an example, I don't think the methods that have been
>moderately successful at chess will succeed with Go.
>I say "moderately successful", because I think the amount of
>computer power used by Deep Blue is disgracefully large and
>conceals a lack of understanding.  Almost all of those 200
>million positions examined each second would be rightfully
>ignored by a more sophisticated program.
>
>See http://www-formal.stanford.edu/jmc/newborn.html which
>appeared in _Science_.
>
>
>
> 
> 
*************************
Michael Arbib
Director, USC Brain Project
University of Southern California
Los Angeles, CA 90089-2520
(213) 743-6452     FAX (213) 740-5687
arbib at pollux.usc.edu
http://www-hbp.usc.edu/HBP/