Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news.harvard.edu!news2.near.net!MathWorks.Com!yeshua.marcam.com!usc!howland.reston.ans.net!news.sprintlink.net!uunet!psinntp!psinntp!pls.com!lef
From: lef@pls.com (Lawrence E Fitzpatrick)
Subject: Re: Is OO Appropriate for NNs? (Was Re: A good book about NN in C++)
Message-ID: <1994Sep16.220722.1622@pls.com>
Organization: Personal Library Software, Inc.
References: <354fvj$7ma@hannibal.psych.ucla.edu>
Date: Fri, 16 Sep 1994 22:07:22 GMT
Lines: 32

In article <354fvj$7ma@hannibal.psych.ucla.edu>,
Jim Kroger <kroger@hannibal.psych.ucla.edu> wrote:
>
>Why do you do this? Would it in fact be slower to do
>it in a c++ fashion? If you are using loops and arrays,
>you must not be coding nodes and connections as objects
>with their own update functions, or you are violating
>the spirit of OOP if your loops are accessing the 
>object's data members without calling function members.


Ahhh, now we're getting onto familiar turf, and Jim seems to be
hitting the crux of the matter.  There is this little matter of
"what's an object."

Seems that folks sometimes forget that if performance REALLY matters,
then the object model takes that into account.  If it doesn't then,
well, a different factoring of the problem space can occur.  I am not
convinced that in the general case the "logical" versus "physical"
factorings are at all incompatible with each other.  I cannot say for
the nn problem domain, as I am still inexperienced.  However, I am
quite familiar with database internals, and in this domain no one ever
(with commercial success) implements the logical models directly, they
are always mapped onto some physical model (e.g., indexes are used to
speed logical operations).

So, would folks mind sharing their decompositions in more detail, so
us beginners can better understand?

Thanks,
lef

