NetNews Usenet Archive 1992 #18/NN_1992_18.iso/spool/comp/ai/neuraln/3181

Text File | 1992-08-13 | 1.9 KB | 41 lines

Newsgroups: comp.ai.neural-nets
Path: sparky!uunet!destroyer!ubc-cs!alberta!arms
From: [e-mail] (Bill Armstrong)
Subject: Re: Perceptrons
Message-ID: <arms.713731475@spedden>
Sender: [e-mail] (News Administrator)
Nntp-Posting-Host: spedden.cs.ualberta.ca
Organization: University of Alberta, Edmonton, Canada
References: <[e-mail]>
Date: Thu, 13 Aug 1992 18:44:35 GMT
Lines: 28

[e-mail] (Robert B Lambert) writes:

>To the NN community in general.
>To what extent to the objections raised in perceptrons carry over to ANNs
>other than BP? (In particular self-organizing NNs, recurrent NNs, ALNs).
> ... I believe ANNs only have a future if all the principal objections
>raised in Perceptrons are addressed.

Since ALNs are special cases of multilayer perceptrons, of course they
can't have any greater theoretical computing power than the usual
multilayer perceptrons, whether used recurrently or not.  On the other
hand, with appropriate encodings of real numbers, ALNs can approximate
at least any continuous function (like those output by an ordinary
multilayer perceptron, say).  The two thus have equivalent power to
realize mappings in the finite digital domain.

In"Perceptrons", Minsky and Papert don't place emphasis on speed of
learning or execution, an attribute that may be considered an
advantage of one or the other type of net in a given practical
situation.  They were concerned about the limits of what can be
computed.  The backpropagation and ALN training procedures have made
it possible to train multilayer nets since they wrote the book, but
that doesn't give the resulting nets any more computing power.
--
***************************************************
Prof. William W. Armstrong, Computing Science Dept.
University of Alberta; Edmonton, Alberta, Canada T6G 2H1
[e-mail] Tel[phone] FAX [phone]