NetNews Usenet Archive 1992 #18/NN_1992_18.iso/spool/comp/ai/neuraln/3209

Text File | 1992-08-17 | 1.8 KB | 36 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!cs.utexas.edu!zaphod.mps.ohio-state.edu!caen!destroyer!ubc-cs!unixg.ubc.ca!kakwa.ucs.ualberta.ca!uofapsy.uucp!mike
From: [e-mail] (Mike Dawson)
Subject: Re: Reducing Training time vs Generalisation
Message-ID: <[e-mail]>
Keywords: back propagation, training, generalisation
Sender: [e-mail]
Organization: Psychology, University of Alberta, Edmonton
References: <[e-mail]> <[e-mail]> <arms.714014919@spedden>
Date: Mon, 17 Aug 1992 16:01:57 GMT
Lines: 23

Bill Armstrong's point about defining generalization is an extremely
good one.  In my view, most ANN researchers view generalization from
an "output performance" perspective, as noted by Bill:  you train a
net on a subset of possible stims, and then measure the network's
responses to the remaining ones.

In a 1989 review of pattern classification by ANNs, Lippmann points
out an equally valid notion of generalization, in terms of "learning
performance".  Specifically, you train a network on some subset of
stims, and then train the network on the remaining patterns.  The
issue is whether learning on the new sets of patterns is helped
in any way from the previous training.  In the memory literature in
cognitive psychology, this effect is called "savings".

Given that it may be extremely difficult to interpolate judgements
about a function because of its complexity, or potential discontinuity,
it may be that generalization of neural networks is more appropriately
viewed from this savings perspective.
--
Michael R.W. Dawson                       email: [e-mail]
Biological Computation Project, Department of Psychology
University of Alberta, Edmonton, AB CANADA T6G 2E9
Tel:  [phone]   Fax: [phone]