Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!rochester!cornell!travelers.mail.cornell.edu!news.kei.com!news.mathworks.com!usenet.eel.ufl.edu!news-feed-1.peachnet.edu!gatech!newsxfer.itd.umich.edu!nntp.cs.ubc.ca!news.UVic.CA!butterfly!nmicklew
From: nmicklew@butterfly.UVic.CA (Nancy  Micklewright)
Subject: Re: What do you mean by "epoch"? (abused term ?)
Message-ID: <1995Mar12.084531.16983@sol.UVic.CA>
Sender: news@sol.UVic.CA
Nntp-Posting-Host: butterfly.uvic.ca
Organization: University of Victoria, Victoria, BC, Canada
Date: Sun, 12 Mar 95 08:45:31 GMT
Lines: 48

Tonight I have noted 3 relevant postings to this subject,
which has persisted for quite some time.  With all due respect
to people who are far more experienced in ANN's than me, the
reason for this persistence is very specific.  The use of
language to describe 'epoch' has been egregiously imprecise.

Let me quote the 3 examples I have just read.
One person goes thru a brief historical scan of the use of
'epoch', and quotes the following:
     "'sweeps' thru the training data [updating after each 
     example]."
The first part of this quote means nothing, since 'training
data' remains wholly undefined. Fortunately, the bracketed
amplification defines it better, so long as we are all
agreed on what 'each example' means.
A second quote from the same source:
     "sweeps thru the set of patterns [ie training examples,
     cases]"
Here again, the amplification is clear, but 'set of patterns'
is about as fuzzy as you can get, at least in my working
vocabulary.

Another person waxes religious, but is not saved (at least
from the evil of imprecision):
     "one pass thru a subset of training samples"
What is 'a subset of training samples?'

Finally, one person (whose comments are almost always very
helpful to me) says he prefers Masters (of "one pass thru a
subset of training samples" fame).  I wish I could share 
his enthusiasm. He then goes on to quote Neuralware's manual:
     "the number of training presentations between weight
     updates."
Groan.

I don't think this has to be so hard, or jargony, or imprecise
or whatever else is wrong here.  Data is not dumped into an
ANN like some child dumping a bag of legos onto the floor.
It is clearly structured.  Rows and columns, cases and variables,
fields and records, are all complementary terms from other
usages.  They provide some parallax.  Until these or other
complementary terms are used as modifiers, the terminology
evolving from ANN's is going to remain a hinderance to the
field. 

Steve Berer
nmicklew@nero.uvic.ca

