Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!news.kei.com!nntp.et.byu.edu!netline-fddi.jpl.nasa.gov!elroy.jpl.nasa.gov!swrinde!howland.reston.ans.net!news.sprintlink.net!noc.netcom.net!ix.netcom.com!netcom.com!perry
From: perry@netcom.com (Perry West)
Subject: Re: How much hardware needed for image analysis?
Message-ID: <perryD9wq25.93v@netcom.com>
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
X-Newsreader: TIN [version 1.2 PL1]
References: <3r7nd7$grb@netnews.upenn.edu>
Date: Fri, 9 Jun 1995 13:13:17 GMT
Lines: 17
Sender: perry@netcom10.netcom.com

When you are looking at 65000 input nodes and who knows how many hidden 
and output nodes.  The number of connections becomes enormous.  Have you 
considered the size of the training set you will need with this network?

Yes, computational power will be an issue, but disk storage for the 
training set may be the real limitation.

I our experience with image analysis using neural networks, we find it is 
necessary to reduce the size of the input data vector.  Either we select 
a feature set appropriate to the images and objectives of our task, or we 
use a transform that condenses the data.  (Please don't ask me about the 
transforms we use.  I can't give out that information.  And no, if I told 
you I wouldn't have to kill you.  I'd have to kill myself.)

Regards,
Perry

