Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!rutgers!rochester!udel!burdvax!bigburd.PRC.Unisys.COM!pastor
From: pastor@bigburd.PRC.Unisys.COM (Jon Pastor)
Newsgroups: comp.ai.neural-nets
Subject: Re: request for philosophic reactions to connectionism
Keywords: connectionism philosophy materialism representations
Message-ID: <10139@burdvax.PRC.Unisys.COM>
Date: 3 May 89 13:50:54 GMT
References: <370@eurtrx.UUCP> <935@syma.sussex.ac.uk>
Sender: news@PRC.Unisys.COM
Organization: Unisys Corporation, Paoli Research Center; Paoli, PA
Lines: 72

In article <935@syma.sussex.ac.uk> aarons@syma.sussex.ac.uk (Aaron Sloman) writes:
>
>Here's my pennyworth.
>   [...lots of cogent remarks omitted...]
>It seems to me absurd to argue over whether either
>connectionist models or conventionalist AI models provide better
>theories of the nature of mind when it is patently clear both are
>still miles away from accounting for more than highly simplified
>versions of tiny fragments of human ability.
>
>Instead of silly squabbles we need to work both top-down (collecting
>requirements for adequate models and explanations), and bottom up
>(trying to investigate different kinds of mechanisms and finding out
>what can and cannot be achieved by putting them together in
>different ways).
>
>It seems very likely that the final story (if we ever find it) will
>involve many different kinds of mechanisms put together in a complex
>variety of ways. Attempts to do it all using one kind of technique
>(Production systems, Logic, PDP mechanisms) will then just look
>silly.
>
>
I was going to send e-mail directly to Aaron, but I believe that his
message is so critical, and I am so grateful to him for expending the
effort that it quite obviously took to prepare it, that I wanted to thank
him publicly.

The arguments one hears about the relative merits of this or that
computational model (architecture, paradigm, programming language, etc.)
tend to generate considerable friction, and thus much heat -- but not much
light.  As Aaron notes, it is inconceivable that any single AI model will
provide adequate support for understanding human cognition, or even the
relatively simpler -- but still staggeringly difficult -- problem of
building intelligent systems, even in limited domains.  I will add to
Aaron's comments the observation that proponents of various schools of
thought act as though their objectives are the only ones that matter.  For
example, biological plausibility is critical in Neural Networks if and only
if you are using the NNs to model, and thus understand, human cognition;
those of us who are primarily interested in using NNs as a computational
tool are not concerned with biological plausibility except from an abstract
intellectual perspective -- we want to know whether the NN will do the
tasks we want it to do, and how well.  

Similarly, mathematical rigor (e.g., proofs of convergence) is undeniably a
Good Thing, but many of us got into AI because rigorous solution techniques
often require assumptions and restrictions that do not hold in the real world.
For example, Expected Utility Theory presumes rationality on the part of the
decision-maker, but there is ample and incontrovertible evidence that
decision-makers depart from the rationality axioms in numerous ways
(intransitivity of preference, inability to make objective probability
judgements, inability to cope with problems of high dimensionality,
sensitivity to time and other pressures, etc.).  I don't think that it's an
accident that I've seen the same heuristic search techniques applied in
Symbolic AI, Connectionist AI, and Operations Research; we're all trying to
solve problems that do not have closed-form solutions, and for which
analytic solutions are generally intractable.  Nor do I think that it's
coincidental that the most widely-used tool in each of these three disciplines 
(forward-chaining rule-based systems, BP with some form of steepest-descent 
in the error space, and Linear Programming, respectively) has no convergence
proof -- but works quite well in practice.  Only a fool would claim that
mathematical rigor is unimportant, but practitioners will gladly use a tool
that has strong empirical support while the theoreticians continue looking
for formal results vindicating the empirical results.


Thank you, Aaron, for stating the case for eclecticism, and against
parochialism, so eloquently.  I hope that future discussions in this
newsgroup will take heed of his message that all of us need to understand
and use each other's tools, and that squabbles about the relative merits of
one tool over another without regard for the problem at hand are not
constructive, instructive, or good science.