Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!zaphod.mps.ohio-state.edu!think!nrl-cmf!tedwards
From: tedwards@nrl-cmf.UUCP (Thomas Edwards)
Newsgroups: comp.ai.neural-nets
Subject: Re: Transputer?  Was: Re: request: implementation NN on CM / transputers
Summary: Connection Machine NN
Keywords: transputer
Message-ID: <78@nrl-cmf.UUCP>
Date: 20 Mar 90 22:59:38 GMT
References: <527@fwi.uva.nl> <9968@spool.cs.wisc.edu>
Reply-To: tedwards@cmsun.UUCP (Thomas Edwards)
Organization: NRL Connection Machine Facility, Washington, DC
Lines: 22

>(Patrick van der Smagt) writes:
>>Does anyone have any references or results about the implementation of
>>neural networks on the Connection Machine?  What about transputers?

A Technical Report has been produced by Thinking Machines concerning
various implementations of backpropagation on the Connection Machine.
Contact David Singer at Thinking Machines.   

I myself have implemented backprop on the CM.  If you think a bit about
how to get the most out of the parallel structure, you can create very
speedly learning implementations on the CM for neural networks. 
My implementation used TMC written matrix algebra routines which
utitlized a very fast systollic array routine.
I happened to have needed very large nets, with few training 
exemplars.  If you have alot of training data, you could put one network
and training exemplar on each processor, and run all 64K training
exemplars at once, and then add up all of the weight deltas using
the systollic array addition commands.  The throughput can be amazing!

-Tom