Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.csd.uwm.edu!gem.mps.ohio-state.edu!ginosko!husc6!brutus.cs.uiuc.edu!lll-winken!arisia!kanga!chrisley
From: chrisley@kanga.uucp (Ron Chrisley UNTIL 10/3/88)
Newsgroups: comp.ai.neural-nets
Subject: Re: : Step Function
Keywords: learning,generalization
Message-ID: <2795@arisia.Xerox.COM>
Date: 6 Sep 89 01:05:30 GMT
References: <1060@rex.cs.tulane.edu> <6980@sdcsvax.UCSD.Edu> <11308@boulder.Colorado.EDU> <1750@cbnewsl.ATT.COM>
Sender: news@arisia.Xerox.COM
Reply-To: k.karn@macbeth.stanford.edu (Ron Chrisley)
Organization: Xerox Palo Alto Research Center
Lines: 19

Tony Russo writes:

"*IF* this is true (and I'm not sure it is), then for a classifier,
from an information-theoretic point of view, it then makes the most sense
to give the network examples that are on the borderline
of a rule or class. These "borderline" patterns should contain more
information about biases than good examples of a class.

In ways this makes sense -- it is akin to telling the network only where the
boundaries between classes are."

This is exactly the learning procedure behind Kohonen's LVQ2 (Learning Vector
Quantization) algorithm.  See his paper in ICNN '88.  He derives the optimality
of this method from Bayesian arguments.  I do not see how the fact that
generalization = bias implies the optimality of learning the boundary
condsitions, and would be very interested in having you elaborate on why you
think it might.

Ron Chrisley