Neural Network applications using Beowulf

Eray Ozkural eozk at
Wed Nov 20 01:00:53 PST 2002

Hi Thomas,

It is very hard to pin down what will pave the way for next generation AI 
research. In machine learning it is well known that there is no specific 
method that will be fitting for all domains. Sometimes an ANN will do the 
job, and most of the times it won't. You have to be very pragmatic. Of 
course, I'm very pleased with the information you've given. I'd be delighted 
to read the paper and evaluate its theoretical content.

As we all know even a small MLFF ANN with back-propagation learning takes a 
very long time (with a scary time complexity) to converge, it is usually  
infeasible for larger networks and really hard problems. However, if 
Hecht-Nielsen's networks are comprised of thousands of artificial neurons 
then we could in principle parallelize the learning algorithm in a smart way. 
(Since we have n >> p, there is a practical way to partition) It then becomes 
a supercomputing problem. My concern is whether those networks are doing 
anything useful because in ANN research it is often the case that you are 
achieving a result that a simple ordinary algorithm could do in a much more 
efficient manner. In supercomputing efficiency becomes even a more important 
aspect because people don't build supercomputers for fun, they build them to 
solve what was not solvable before. (As we all appreciate)

That is, one has to make sure that he is not doing ANNs for the sake of ANN, 
especially on a supercomputer. If you're an AI researcher you should also 
have noticed that most ANN or GA/hybrid systems research focuses on getting 
results for a problem that could be solved better with existing algorithms. 
That is why I'm a little skeptical of the quality of a work that bears the 
"ANN" label. I think this puts me on the harder edge of algorithms world, but 
that's what I think :>

Nevertheless, in some cases, ANNs are better. For instance, when you are 
predicting the stock index in stock market time series. So if we have a 
problem that are better solved by Hecht-Nielsen's networks, and if they 
require a supercomputer it would prove to be a new challenge for 
supercomputing specialists.


On Wednesday 20 November 2002 01:05 am, Thomas Zheng wrote:
> Hi Eray,
> What you said was pretty much true until recently.  In WCCI2002
> ( this May, Dr. Hecht-Nielsen from Univeristy of
> California, San Diego, announced his new thalamocortical information
> processing theory, which, I think, is paving the way for next generation AI
> research.  In his special lecture, he showed a couple slides of
> parallel-computing machines he used in his lab.  Even though he never got
> into the details of these machines, from what i know about associative
> memory networks, which are the building blocks of his new theories, it does
> demonstrate highly parallel-implementable features.  And we are talking
> about thousands of nodes  as minimum requirements for these networks to be
> functional.
> In my opinions, there are tremendous potentials for parallel computing in
> the neural network arena.  The question is what kind of practical/useful
> applications would come out of it.
> Regards
> Thomas Zheng
> At 09:40 AM 11/14/2002 +0200, you wrote:
> >On Tuesday 12 November 2002 06:24, Robert G. Brown wrote:
> > > I actually think that there is room to do a whole lot of interesting
> > > research on this in the realm of Real Computer Science.
> > >
> > > Too bad I'm a physicist...;-)
> >
> >Note that most artificial neural network applications don't fall in the
> > realm of supercomputing since they would be best suited to hardware
> >implementations, or more commonly, serial software.
> >
> >
> >We had discussed this with colleagues back at bilkent cs department and we
> >could not find great research opportunities in this area. It is a little
> >similar to stuff like parallel DFA/NFA systems. You first need an
> > application to prove that there is need for problems of that magnitude
> > (more than what a serial computer could solve!). What good is a
> > supercomputer for an artificial neural network that is comprised of just
> > 20 nodes?
> >
> >If of course somebody showed an application that did demand the power of a
> >supercomputer it would be very different, then we would get all of our
> >combinatorial tools to partition the computational space and parallelize
> >whatever algorithm there is :)
> >
> >Neural networks being Turing-complete, I assume such a network would bear
> > an arrangement radically different from the "multi-layer feed-forward"
> > networks that EE people seem to be obsessed with. I have lost my interest
> > in that area since they don't seem to demand parallel systems and they
> > are not biologically plausible.
> >
> >Regards,
> >
> >--
> >Eray Ozkural (exa) <erayo at>
> >Comp. Sci. Dept., Bilkent University, Ankara
> >www:  Malfunction:
> > GPG public key fingerprint: 360C 852F 88B0 A745 F31B
> >  EA0F 7C07 AE16 874D 539C
> Regards,
> Thomas Zheng
> _______________________________________________
> Beowulf mailing list, Beowulf at
> To change your subscription (digest mode or unsubscribe) visit

Eray Ozkural
GPG public key fingerprint: 360C 852F 88B0 A745 F31B  EA0F 7C07 AE16 874D 539C

More information about the Beowulf mailing list