Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!pipex!sunsite.doc.ic.ac.uk!hgmp.mrc.ac.uk!news.dcs.warwick.ac.uk!not-for-mail
From: B.Barnes@dcs.warwick.ac.uk (Benjohn Barnes)
Subject: An idea (A GOOD one?)
Message-ID: <1994Nov16.124002.11714@dcs.warwick.ac.uk>
Summary: Spliting up tasks
Keywords: function
Sender: news@dcs.warwick.ac.uk (Network News)
Nntp-Posting-Host: granite
Organization: Department of Computer Science, Warwick University, England
Date: Wed, 16 Nov 1994 12:40:02 GMT
Lines: 50

I was sitting in my Psychology lecture this mornig and an idea popped into my
mind. I think it has been forming for quite a while. The chances are that I am
re-inventing the wheel, or talking rubish. Anyway here it is :-

Although our brain is parallel it is also organised in a modular way.
Especially the more "instinctive" parts, those dealing with motivation and
processing sences. When we store information we don't remember it like a bit
map, storring everything we see, we remember things in terms of other
things, more like a structured drawing pakage I supose. If a given thing
crops up lots of times it will just be stored once and used again and
again.

I think that it is the same when we learn. We don't learn how to integrate
multi dimensional functions, we start with adding peas and work our way
up. When learnig to walk, we very early on get the hang of these limb things,
then we learn to crawl about, then to walk, maybe with a push along thing on
wheels. Finally we get to walking freely, then stairs and more complex
situations.

Maybe when we learn to read it is a similar situation. We start bm learning
about lines and curves and shapes. Then we put these together to make letters
and these into words. Actually I'm not so sure about this last example, but I
hope it makes the point.

What this is getting to is maybe some of the things that NN's are being trained
to do are perfectly posible but could be implemented far more quickly and
with less data if they were done in stages.

A given task could be split into rough sections. Like, as a bad example,
reading text could maybe be split up into:-

	turning a bit map into lines and curves,
	turning the lines and curves into letters.

So you would give a net a  load of pictures with lines and curves in and try
to get it to describe the lines and curves in some way. You would then plug
another "lump" of nodes onto this. And train this to recognise letters or
pictures. In the second training the waitings in the first "lump" would be
kept steady, or alowed to change but much less easilly than those in the new
"lump".

I think that by doing this, much more complex tasks could be assembled that
would usually be very hard, or imposible to solve.

Like I said, I don't know if this has been done before. It strikes me as
being a fairly neat idea, and worth looking into, so it probaably has. Hope
it is some help to someone. Sorry about the lousy spelling, need to get a
spell checker going.

	Benjohn
