Path: utzoo!attcan!utgpu!news-server.csri.toronto.edu!rutgers!cs.utexas.edu!sun-barr!lll-winken!tristan!loren
From: loren@tristan.llnl.gov (Loren Petrich)
Newsgroups: comp.ai.philosophy
Subject: Re: Reasoning Paradigms
Message-ID: <69377@lll-winken.LLNL.GOV>
Date: 6 Oct 90 14:44:11 GMT
References: <3586@media-lab.MEDIA.MIT.EDU> <69347@lll-winken.LLNL.GOV> <3593@media-lab.MEDIA.MIT.EDU>
Sender: usenet@lll-winken.LLNL.GOV
Organization: Lawrence Livermore National Laboratory
Lines: 91

In article <3593@media-lab.MEDIA.MIT.EDU> minsky@media-lab.media.mit.edu (Marvin Minsky) writes:
>
>
>I agree with most of what loren@tristan.llnl.gov (Loren Petrich) said
>in article 62.  The only problem I have is with his assertion 
>
>> I feel that there is much more promise in NN's than in traditional
>> AI, which has been dependent on working out decision rules explicitly.

	You're right. I goofed. I concede that there are things that
traditional AI techniques can do better than most NN's. I doubt that
NN's will ever pose much competition in fields like computer algebra,
where most of the inference rules are straightforward and unambiguous,
and have been well understood for a long time.

	There are other difficulties with NN's, at least at the
present time. For instance, NN's are generally constructed around data
structures that are linear and whose lengths are fixed. This is OK for
a wide range of problems, but there are difficulties for representing
data structures whose length may vary, and even which are nonlinear,
an example being a treelike one. There are tricks I have seen for
getting around that, but even there, a NN will probably have to be
"managed" by some outside system.

	But my point was, why attempt to painstakingly work out
hundreds of complicated and imprecise inference rules when the whole
job can be done automatically?

>  1.  Yes: systems with compact rules with very few input terms are not
>good at recognizing patterns which need many inputs.  So AI systems
>restricted to compact rules must be supplemented by NN-like
>structures.  
>  2.  No: the NN-like structures cannot replace the "reasoning
>systems" of "traditional AI", unless we supply architectures that
>embody those goal-oriented processes.  For example, "annealing" does
>not replace all other kinds of intelligent heuristic search.  

	I agree.

> ... In this sense, then, NN solutions, in
>contrast, tend to be dead ends, simply because what you end
>up with, after your 100,000 steps of hill-climbing, is an opaque
>vector of coefficients.  You have solved the prob lem, all right.  You
>have even _learned_ the solution!  But you don't end up with anything
>you can THINK about!

	I understand your point. However, my colleagues and I have
occasionally been able to interpret the weight values produced by
NN's. One project we did was to evaluate spectra produced in etching
chips. By examining them, we hoped to train a NN to determine how much
hydrogen was in the etching chamber. We discovered that the weights
were largest in some small regions of the spectrum. These corresponded
to lines of H and one of CO, a reaction product. It was surprising to
us that the NN might have been using a CO line as a diagnostic for the
amount of hydrogen.

	An improved version might be set up to look only at H and CO
lines, given what the first one ended up focusing on.

	I think that the difficulty of not learning too much about
what one wants to recognize is far from fatal in practice, however
desirable in theory may be.

>Here is a simple, if abstract, example of what I mean.  Consider one
>of the most powerful ideas in traditional AI -- the concept of
>acheiving a goal by detecting differences between the present
>situation ("what you have") and a target situation ("what you want").
>The Newell and Simon 'GPS' system did such things (and worked in many
>cases, but not all) by trying various experiments and comparing the
>results, and then applying strategies designed (or learned) for
>'reducing' those differences.

	I see the point. It seems to me very difficult to imagine how
to get a NN to do something like that.

>In any case, I want to thank Loren for endless thoughtful observations
>about many other topics.  I intend to think more about what he said here.

	Thank you.

	I have been interested in AI and I have become rather
disappointed at its slow rate of progress over the years. However, it
is good to know that now we can make progress somewhere.


$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
Loren Petrich, the Master Blaster: loren@sunlight.llnl.gov

Since this nodename is not widely known, you may have to try:

loren%sunlight.llnl.gov@star.stanford.edu