How about parallel computing with finance
Robert G. Brown
rgb at phy.duke.edu
Wed Dec 6 09:56:00 PST 2000
On Wed, 6 Dec 2000, Terrence E. Brown wrote:
> I am also interested in the business and managerial application as well as other
> industrial apps.
> I would certainly like to talk with another with similar thoughts. I have even
> started an org dedicated to that objective.
> "Horatio_B._Bogbindero_ (Horatio B. Bogbindero )" wrote:
> > i would just like to know about building neural networks in clusters.
> > i am not into neural network but some people here in the university
> > maybe interested. however, we do not know where to start. i would to know
> > where i can get some sample NN code. maybe for something trivial.
Hmmm, I don't know how much such of such a discussion should occur on
this list. The following (up to <shameless marketing> is probably
Neural networks (and the genetic algorithms that underlie a really good
one in a problem with high dimensionality) are certainly fascinating
things. They are even in some sense a fundamentally parallel thing, as
their processing capabilities result from a tiered composition of
relatively simple (but nonlinear) transfer functions. A general
discussion of NN's and how they work is clearly not appropriate for this
list though. There are some particular issues that are.
In practice, most the parallelization issues of NN's are a small part of
the overall problem UNLESS you are interested in constructing custom
hardware or building NN ASIC's or the like. This is because computers
generally run neural network SIMULATORS and use what amounts to
relatively small-scale linear algebra (transmogrified through an e.g.
logistic function) to do a net evaluation. Since this is so small that
it will often fit into even L1 (and almost certainly into L2) there is
no possible way that it can be distributed in parallel except via
(embarrassingly parallel) task division in a profitable way.
Evaluation of networks' values applied to training/trial set data makes
up the bulk of the numerical effort in building a network and is at the
heart of the other tasks (e.g. regression or conjugate gradient
improvement of the weights). For large training/trial sets and "big"
networks, this can be split up (and my experiences splitting it up are
recorded in one of the talks available on the brahma website). For
small ones, the ratio of the time spent doing parallel work to the time
spent doing parallel communication isn't favorable and one's parallel
scaling sucks. As in even two nodes may complete in more time than one
I'm working on an improved algorithm that splits up NN
construction/training in a way that is more functionally coherent. That
way one or two of the distinct tasks can be parallelized very
efficiently and thoroughly and the results fed back into a mostly serial
or entirely serial step further down the pipeline. I expect that this
will permit a very nice master/slave implementation of a neural network
constructor where nodes are slaves that can be working on any of a
number of parallelized tasks according to the directions of the master
(quite possibly with internode IPC's, though), and all the serial work
can be done on the master.
NN's (parallelized or not) are, as one might expect, incredibly useful
and potentially profitable. After all, a successful predictive model
"tells the future", at least probabilistically, by construction, and
does even better than a delphic oracle ever did in that they can often
provide a quantitative (although probabilistic) answer to "what if"
questions as well. In ancient times the words of the oracle were just
fate and nothing you could do would change them. In business, one would
like to predict what is likely to happen if you follow plan A instead of
plan B. Just about any business manager has a list of questions about
the future (what if or otherwise) they would love to have the answers
to. That's one of Market Driven's foci -- providing answers and
expertise in business optimization.
Anyway, let me know if you're interested in more discussion of this (or
how NN's work or how they and predictive modeling in general can be
applied in business and managerial situations) offline.
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
More information about the Beowulf