Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!bbn!rochester!PT!cadre!geb
From: geb@cadre.dsl.PITTSBURGH.EDU (Gordon E. Banks)
Newsgroups: comp.ai
Subject: Re: Neural Networks & Unaligned fields
Message-ID: <802@cadre.dsl.PITTSBURGH.EDU>
Date: Wed, 9-Sep-87 09:54:23 EDT
Article-I.D.: cadre.802
Posted: Wed Sep  9 09:54:23 1987
Date-Received: Fri, 11-Sep-87 02:06:34 EDT
References: <277@ndmath.UUCP> <3523@venera.isi.edu>
Reply-To: geb@cadre.dsl.pittsburgh.edu.UUCP (Gordon E. Banks)
Organization: Decision Systems Lab., Univ. of Pittsburgh, PA.
Lines: 21

In article <3523@venera.isi.edu> smoliar@vaxa.isi.edu.UUCP (Stephen Smoliar) writes:
>In article <277@ndmath.UUCP> milo@ndmath.UUCP (Greg Corson) writes:
>>Ok, here's a quick question for anyone who's getting into Neural Networks.
>>If you setup the type of network described in BYTE this month, or the
>>type used in the program recently posted to the net, what happens if you
>>feed it an input image that is not aligned right?

I didn't see the Byte article, but the simple neural networks that
I have seen (such as the one that solves the T-C problem by Hinton
& Rummelhart in the PDP book) do not generalize very well.  You can
train the hidden units with a given input, but then if you shift the
pattern, it won't work.  I asked Rummelhart about this, and he said
that once the hidden units develop the patterns (such as edge detectors
and center-surround, etc.) you do not need to retrain for each translation
of the pattern, but you need to add more units to the network.  These
units have the same weights as the previously trained units, but they
have a different field of view.  You have to have another set of units
for each region which can possibly contain the image.  Alternatively,
you have to have a scheme for making sure the image is "centered" in
the field of view.  Sounds like there is some room for interesting
research here, maybe a thesis.