<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2//EN">
<HTML>
<HEAD>
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
<META NAME="Generator" CONTENT="MS Exchange Server version 5.5.2650.12">
<TITLE>RE: How about parallel computing with finance</TITLE>
</HEAD>
<BODY>

<P><FONT SIZE=2>I've been trying to get my e-mail client to sent text only (no HTML).  I apologize in advance if this is HTML.</FONT>
</P>

<P><FONT SIZE=2>This is exactly where I've been focusing my energy on with beowulf technology.  And in health care, like I think in many industries, making the technology work for your finance department is key.  Do great science, and people in finance raise an eyebrow.  Save the company six figures in operating costs or process their data faster, and you get applauded, invited to meet with directors  . . . the whole nine yards.</FONT></P>

<P><FONT SIZE=2>Richard Schilling</FONT>
<BR><FONT SIZE=2>Web Integration Programmer</FONT>
<BR><FONT SIZE=2>Affiliated Health Services</FONT>
<BR><FONT SIZE=2>Mount Vernon, WA</FONT>
</P>
<BR>

<P><FONT SIZE=2>> -----Original Message-----</FONT>
<BR><FONT SIZE=2>> From: Terrence E. Brown [<A HREF="mailto:tbrown@lector.kth.se">mailto:tbrown@lector.kth.se</A>]</FONT>
<BR><FONT SIZE=2>> Sent: Wednesday, December 06, 2000 11:14 PM</FONT>
<BR><FONT SIZE=2>> To: Robert_G._Brown_ (Robert G. Brown <rgb@phy.duke.edu>)</FONT>
<BR><FONT SIZE=2>> Cc:</FONT>
<BR><FONT SIZE=2>> Horatio_B._Bogbindero__(Horatio_B._Bogbindero_<wyy@cersa.admu.</FONT>
<BR><FONT SIZE=2>> edu.ph>)_;</FONT>
<BR><FONT SIZE=2>> liuxg__; beowulf@beowulf.org__</FONT>
<BR><FONT SIZE=2>> Subject: Re: How about parallel computing with finance</FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> Probably , not much discussion should take place here. But I </FONT>
<BR><FONT SIZE=2>> will have a</FONT>
<BR><FONT SIZE=2>> another venue for it in a few days.</FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> "Robert_G._Brown_ (Robert G. Brown )" wrote:</FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> > On Wed, 6 Dec 2000, Terrence E. Brown wrote:</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > > I am also interested in the business and managerial </FONT>
<BR><FONT SIZE=2>> application as well</FONT>
<BR><FONT SIZE=2>> > as other</FONT>
<BR><FONT SIZE=2>> > > industrial apps.</FONT>
<BR><FONT SIZE=2>> > ></FONT>
<BR><FONT SIZE=2>> > > I would certainly like to talk with another with similar </FONT>
<BR><FONT SIZE=2>> thoughts. I</FONT>
<BR><FONT SIZE=2>> > have even</FONT>
<BR><FONT SIZE=2>> > > started an org dedicated to that objective.</FONT>
<BR><FONT SIZE=2>> > ></FONT>
<BR><FONT SIZE=2>> > > Terrence</FONT>
<BR><FONT SIZE=2>> > ></FONT>
<BR><FONT SIZE=2>> > > "Horatio_B._Bogbindero_ (Horatio B. Bogbindero )" wrote:</FONT>
<BR><FONT SIZE=2>> > ></FONT>
<BR><FONT SIZE=2>> > > > i would just like to know about building neural </FONT>
<BR><FONT SIZE=2>> networks in clusters.</FONT>
<BR><FONT SIZE=2>> > > > i am not into neural network but some people here in </FONT>
<BR><FONT SIZE=2>> the university</FONT>
<BR><FONT SIZE=2>> > > > maybe interested. however, we do not know where to </FONT>
<BR><FONT SIZE=2>> start. i would to</FONT>
<BR><FONT SIZE=2>> > know</FONT>
<BR><FONT SIZE=2>> > > > where i can get some sample NN code. maybe for </FONT>
<BR><FONT SIZE=2>> something trivial.</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > Hmmm, I don't know how much such of such a discussion </FONT>
<BR><FONT SIZE=2>> should occur on</FONT>
<BR><FONT SIZE=2>> > this list.  The following (up to <shameless marketing> is probably</FONT>
<BR><FONT SIZE=2>> > reasonable.</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > Neural networks (and the genetic algorithms that underlie a </FONT>
<BR><FONT SIZE=2>> really good</FONT>
<BR><FONT SIZE=2>> > one in a problem with high dimensionality) are certainly fascinating</FONT>
<BR><FONT SIZE=2>> > things.  They are even in some sense a fundamentally </FONT>
<BR><FONT SIZE=2>> parallel thing, as</FONT>
<BR><FONT SIZE=2>> > their processing capabilities result from a tiered composition of</FONT>
<BR><FONT SIZE=2>> > relatively simple (but nonlinear) transfer functions.  A general</FONT>
<BR><FONT SIZE=2>> > discussion of NN's and how they work is clearly not </FONT>
<BR><FONT SIZE=2>> appropriate for this</FONT>
<BR><FONT SIZE=2>> > list though.  There are some particular issues that are.</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > In practice, most the parallelization issues of NN's are a </FONT>
<BR><FONT SIZE=2>> small part of</FONT>
<BR><FONT SIZE=2>> > the overall problem UNLESS you are interested in constructing custom</FONT>
<BR><FONT SIZE=2>> > hardware or building NN ASIC's or the like.  This is </FONT>
<BR><FONT SIZE=2>> because computers</FONT>
<BR><FONT SIZE=2>> > generally run neural network SIMULATORS and use what amounts to</FONT>
<BR><FONT SIZE=2>> > relatively small-scale linear algebra (transmogrified </FONT>
<BR><FONT SIZE=2>> through an e.g.</FONT>
<BR><FONT SIZE=2>> > logistic function) to do a net evaluation.  Since this is </FONT>
<BR><FONT SIZE=2>> so small that</FONT>
<BR><FONT SIZE=2>> > it will often fit into even L1 (and almost certainly into </FONT>
<BR><FONT SIZE=2>> L2) there is</FONT>
<BR><FONT SIZE=2>> > no possible way that it can be distributed in parallel except via</FONT>
<BR><FONT SIZE=2>> > (embarrassingly parallel) task division in a profitable way.</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > Evaluation of networks' values applied to training/trial </FONT>
<BR><FONT SIZE=2>> set data makes</FONT>
<BR><FONT SIZE=2>> > up the bulk of the numerical effort in building a network </FONT>
<BR><FONT SIZE=2>> and is at the</FONT>
<BR><FONT SIZE=2>> > heart of the other tasks (e.g. regression or conjugate gradient</FONT>
<BR><FONT SIZE=2>> > improvement of the weights).  For large training/trial sets </FONT>
<BR><FONT SIZE=2>> and "big"</FONT>
<BR><FONT SIZE=2>> > networks, this can be split up (and my experiences </FONT>
<BR><FONT SIZE=2>> splitting it up are</FONT>
<BR><FONT SIZE=2>> > recorded in one of the talks available on the brahma website).  For</FONT>
<BR><FONT SIZE=2>> > small ones, the ratio of the time spent doing parallel work </FONT>
<BR><FONT SIZE=2>> to the time</FONT>
<BR><FONT SIZE=2>> > spent doing parallel communication isn't favorable and </FONT>
<BR><FONT SIZE=2>> one's parallel</FONT>
<BR><FONT SIZE=2>> > scaling sucks.  As in even two nodes may complete in more </FONT>
<BR><FONT SIZE=2>> time than one</FONT>
<BR><FONT SIZE=2>> > working alone.</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > I'm working on an improved algorithm that splits up NN</FONT>
<BR><FONT SIZE=2>> > construction/training in a way that is more functionally </FONT>
<BR><FONT SIZE=2>> coherent.  That</FONT>
<BR><FONT SIZE=2>> > way one or two of the distinct tasks can be parallelized very</FONT>
<BR><FONT SIZE=2>> > efficiently and thoroughly and the results fed back into a </FONT>
<BR><FONT SIZE=2>> mostly serial</FONT>
<BR><FONT SIZE=2>> > or entirely serial step further down the pipeline.  I </FONT>
<BR><FONT SIZE=2>> expect that this</FONT>
<BR><FONT SIZE=2>> > will permit a very nice master/slave implementation of a </FONT>
<BR><FONT SIZE=2>> neural network</FONT>
<BR><FONT SIZE=2>> > constructor where nodes are slaves that can be working on any of a</FONT>
<BR><FONT SIZE=2>> > number of parallelized tasks according to the directions of </FONT>
<BR><FONT SIZE=2>> the master</FONT>
<BR><FONT SIZE=2>> > (quite possibly with internode IPC's, though), and all the </FONT>
<BR><FONT SIZE=2>> serial work</FONT>
<BR><FONT SIZE=2>> > can be done on the master.</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > <shameless marketing></FONT>
<BR><FONT SIZE=2>> > NN's (parallelized or not) are, as one might expect, </FONT>
<BR><FONT SIZE=2>> incredibly useful</FONT>
<BR><FONT SIZE=2>> > and potentially profitable.  After all, a successful </FONT>
<BR><FONT SIZE=2>> predictive model</FONT>
<BR><FONT SIZE=2>> > "tells the future", at least probabilistically, by construction, and</FONT>
<BR><FONT SIZE=2>> > does even better than a delphic oracle ever did in that </FONT>
<BR><FONT SIZE=2>> they can often</FONT>
<BR><FONT SIZE=2>> > provide a quantitative (although probabilistic) answer to "what if"</FONT>
<BR><FONT SIZE=2>> > questions as well.  In ancient times the words of the </FONT>
<BR><FONT SIZE=2>> oracle were just</FONT>
<BR><FONT SIZE=2>> > fate and nothing you could do would change them.  In </FONT>
<BR><FONT SIZE=2>> business, one would</FONT>
<BR><FONT SIZE=2>> > like to predict what is likely to happen if you follow plan </FONT>
<BR><FONT SIZE=2>> A instead of</FONT>
<BR><FONT SIZE=2>> > plan B. Just about any business manager has a list of </FONT>
<BR><FONT SIZE=2>> questions about</FONT>
<BR><FONT SIZE=2>> > the future (what if or otherwise) they would love to have </FONT>
<BR><FONT SIZE=2>> the answers</FONT>
<BR><FONT SIZE=2>> > to.  That's one of Market Driven's foci -- providing answers and</FONT>
<BR><FONT SIZE=2>> > expertise in business optimization.</FONT>
<BR><FONT SIZE=2>> > </shameless marketing></FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > Anyway, let me know if you're interested in more discussion </FONT>
<BR><FONT SIZE=2>> of this (or</FONT>
<BR><FONT SIZE=2>> > how NN's work or how they and predictive modeling in general can be</FONT>
<BR><FONT SIZE=2>> > applied in business and managerial situations) offline.</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> >    rgb</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > --</FONT>
<BR><FONT SIZE=2>> > Robert G. Brown                        <A HREF="http://www.phy.duke.edu/~rgb/" TARGET="_blank">http://www.phy.duke.edu/~rgb/</A></FONT>
<BR><FONT SIZE=2>> > Duke University Dept. of Physics, Box 90305</FONT>
<BR><FONT SIZE=2>> > Durham, N.C. 27708-0305</FONT>
<BR><FONT SIZE=2>> > Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb@phy.duke.edu</FONT>
<BR><FONT SIZE=2>> ></FONT>
<BR><FONT SIZE=2>> > To: "Terrence E. Brown" <tbrown@lector.kth.se></FONT>
<BR><FONT SIZE=2>> > From: "Robert G. Brown" <rgb@phy.duke.edu></FONT>
<BR><FONT SIZE=2>> > Subject: Re: How about parallel computing with finance</FONT>
<BR><FONT SIZE=2>> > Date: Wed, 6 Dec 2000 12:56:00 -0500 (EST)</FONT>
<BR><FONT SIZE=2>> > cc: "Horatio_B._Bogbindero_ (Horatio B. Bogbindero</FONT>
<BR><FONT SIZE=2>> > <wyy@cersa.admu.edu.ph>)" <wyy@cersa.admu.edu.ph>, liuxg_</FONT>
<BR><FONT SIZE=2>> > <beowulf@eyou.com>, "beowulf@beowulf.org_" <beowulf@beowulf.org></FONT>
<BR><FONT SIZE=2>> > Message-ID: </FONT>
<BR><FONT SIZE=2>> <Pine.LNX.4.30.0012061029560.15373-100000@ganesh.phy.duke.edu></FONT>
<BR><FONT SIZE=2>> > Received: from ganesh.phy.duke.edu (ganesh.phy.duke.edu </FONT>
<BR><FONT SIZE=2>> [152.3.183.52]) by</FONT>
<BR><FONT SIZE=2>> > lector_gw.lector.kth.se with SMTP id MSGTTTQE; Wed, 6 Dec </FONT>
<BR><FONT SIZE=2>> 2000 17:56:02 GMT</FONT>
<BR><FONT SIZE=2>> > Received: from localhost (rgb@localhost) by ganesh.phy.duke.edu</FONT>
<BR><FONT SIZE=2>> > (8.9.3/8.9.3) with ESMTP id MAA15684; Wed, 6 Dec 2000 12:56:00 -0500</FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> --</FONT>
<BR><FONT SIZE=2>> -----------------------------------------</FONT>
<BR><FONT SIZE=2>> Terrence E. Brown, Ph.D.</FONT>
<BR><FONT SIZE=2>> Stockholm School of Entrepreneurship</FONT>
<BR><FONT SIZE=2>> Assistant Professor</FONT>
<BR><FONT SIZE=2>> Royal Technical Institute</FONT>
<BR><FONT SIZE=2>> INDEK</FONT>
<BR><FONT SIZE=2>> Drottning Kristina Väg 35D</FONT>
<BR><FONT SIZE=2>> S- 100 44 Stockholm</FONT>
<BR><FONT SIZE=2>> Sweden</FONT>
<BR><FONT SIZE=2>> Tel: +46(0)8-7906174 (work)</FONT>
<BR><FONT SIZE=2>> Fax: +46(0)8-7906741 (work)</FONT>
<BR><FONT SIZE=2>> Email: tbrown@lector.kth.se</FONT>
<BR><FONT SIZE=2>>  or terrence.brown@sses.se</FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> </FONT>
<BR><FONT SIZE=2>> _______________________________________________</FONT>
<BR><FONT SIZE=2>> Beowulf mailing list</FONT>
<BR><FONT SIZE=2>> Beowulf@beowulf.org</FONT>
<BR><FONT SIZE=2>> <A HREF="http://www.beowulf.org/mailman/listinfo/beowulf" TARGET="_blank">http://www.beowulf.org/mailman/listinfo/beowulf</A></FONT>
<BR><FONT SIZE=2>> </FONT>
</P>

</BODY>
</HTML>