Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!oitnews.harvard.edu!purdue!lerc.nasa.gov!magnus.acs.ohio-state.edu!math.ohio-state.edu!howland.reston.ans.net!vixen.cso.uiuc.edu!uwm.edu!lll-winken.llnl.gov!uop!pacbell.com!amdahl.com!amd!amd.com!txnews.amd.com!news
From: Joseph Bridgewater <bridgwtr@vanzandt.amd.com>
Subject: Re: RFD: My Learning/Thinking Neural Network
Content-Type: text/plain; charset=us-ascii
Message-ID: <DD1syp.BJE@txnews.amd.com>
Sender: news@txnews.amd.com
Nntp-Posting-Host: corgi
Content-Transfer-Encoding: 7bit
Organization: Advanced Micro Devices, Inc.
References: <DCyy7A.K4B@txnews.amd.com> <4085ho$d7f@xdm084.ccc.cranfield.ac.uk>
Mime-Version: 1.0
Date: Wed, 9 Aug 1995 14:44:45 GMT
X-Mailer: Mozilla 1.1 (Windows; U; 16bit)
Lines: 36

"S.Read" <me944p> wrote:
>Fascinating. How about some more details? How long have you been working
>on it? How many neurons? Synapses? Which type: Hopfield/Feedforward/
>recursive/other ??? What is the learning algorithm like?

We worked an aggregate of about 4 years on research and development.  The number of neurons is 
not limited.  The synapses are implied in the software implementation of the paradigm.  As to 
what type (Hopfield, etc), I would need to review our books to see what term best fits, 
however as I remember it similiar to Hopfield, however significiantly more advanced.  

The learning is through an advanced backpropogation technique.  You supply an input data set 
and present and output data set that is the correct answer, and it proceeds to self modify to 
be able to produce that answer with that input, perhaps with a slight error.  Retraining is 
likly to be neccessary after a period of usage.  Even during (re)trainging, ouput can be 
elicited from output nodes where trainging data is not presented

>
>And.. What does it do?

What does it do?  Well one way of describing that is that once taught all the input=output 
relationships you can, it will provide, in essence, a fuzzy quantification of all other input 
sets.  It was not designed to be limited to do one thing; it can be used for anything where 
thinking can be applied.  

Although the design is complete, the implementation is extremelly complex.  It is the most 
comlpex C code I've seen.  I have an advanced bug reporting system in the code to help me find 
bugs, but there are one or two I need to hunt and kill.  There is also a little more coding to 
do for the low level functions.  Basically, the code will be just the paradigm.  We do not 
have a front end or shell of any type.  One reason for this is that we want to be able to run 
this on every kind of computer in the world and figure the shell should be suited to the 
computer and there are lots of shells (ie Windows NT, X, etc.) that could be made to interact 
with the paradigm.

Please read/write other post for more discussion.


