From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!qt.cs.utexas.edu!yale.edu!jvnc.net!netnews.upenn.edu!libra.wistar.upenn.edu Tue Jan 28 12:17:20 EST 1992
Article 3116 of comp.ai.philosophy:
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!qt.cs.utexas.edu!yale.edu!jvnc.net!netnews.upenn.edu!libra.wistar.upenn.edu
>From: weemba@libra.wistar.upenn.edu (Matthew P Wiener)
Newsgroups: comp.ai.philosophy
Subject: Three problems with concluding neural nets are in our noses
Message-ID: <63541@netnews.upenn.edu>
Date: 24 Jan 92 16:31:32 GMT
References: <1992Jan23.213320.19198@hilbert.cyprs.rain.com>
Sender: news@netnews.upenn.edu
Reply-To: weemba@libra.wistar.upenn.edu (Matthew P Wiener)
Organization: The Wistar Institute of Anatomy and Biology
Lines: 37
Nntp-Posting-Host: libra.wistar.upenn.edu
In-reply-to: max@hilbert.cyprs.rain.com (Max Webb)

In article <1992Jan23.213320.19198@hilbert.cyprs.rain.com>, max@hilbert (Max Webb) writes:
>I consider waveforms, lesion behavior, and functionality to be
>'significant'. So does anyone else actually working in the field(s)

How does the evidence answer the following three questions:

(1) Are neural networks too powerful a method mathematically?

This means that as a technique, they can solve things that obviously
have nothing to do with neural networks as a physical mechanism.  So
are you getting the solution by mathematical power or physical realism?

The same problem arises in simulated annealing: the travelling salesman
problem is obviously not thermodynamic, yet SA provides a powerful way
to solve TSP.  SA is also used in molecular modelling--its role is not
so clear there.  People introduce artificial energy constraints, and it
still works.  Is it power or realism?  Genetic algorithms used in DNA
analysis have the same problem.

(2) How does the evidence distinguish internal vs external models?

By this I refer to the major simplification that occurs in olfaction:
the internal network potentials are linearly related to the external
measured potentials (see Freeman MASS ACTION 4.3).  A Marshall-Froehlich
model that relies on pumped phonon condesates could just as well be the
explanation, and, thanks to the above linearity in the case of olfaction,
is mathematically equivalent to neural networks.  How would you tell
which it is?

(3) Does a non-neural network instead generate the seen results?

Chemical concentration fluctuation and possibly immune systems can mimic
neural networks.  Most likely, these would operate in conjunction with
_neural_ neural networks.  A mathematical model would not notice the
difference.  Lesion experiments would not notice the difference either.
-- 
-Matthew P Wiener (weemba@libra.wistar.upenn.edu)


