Add a few points to your PC's
Neural-network programs simulate human
By Andrew Maykuth
years ago, Connecticut currency trader James O'Sullivan bought some new
software, one of a new class of brainy programs known as neural networks.
He stuffed the software full of historical data about inflation, money
supply, interest rates, and currency prices. Then he entered the current
economic indicators. The system rewarded him with straightforward advice:
Buy yen. Sell marks. Hold dollars.
are superior," says British-born O'Sullivan, who religiously applies
the neural network's recommendations to one portfolio. "I'm not
saying what I've got is perfect, but it does seem to work."
O'Sullivan's got is Brainmaker, from California Scientific Software
in Grass Valley, Calif. It was one of the first of a growing number of
neural-network packages available for PCs.
networks were envisioned as a computer imitation of the human brain's web
of interconnected neurons, using thousands of processors operating
simultaneously in a massively parallel structure. Such hardware has proven
difficult and expensive to design, however, and in recent years, software
developers have devised simulated neural networks using a newly discovered
algorithm called back-propagation. This algorithm compares the relation
between various inputs and outputs through a process of trial and error.
Make your own rules.
Neural networks are
good at the things that have traditionally tripped up computers: making associations and generalizations, drawing
inferences, recognizing patterns, and disregarding errors. Unlike expert
systems and other conventional rule-based programs, neural-net software
does not need to be preprogrammed to consider every conceivable aspect of
a problem. Instead, users "train" neural networks by filling
them with massive amounts of data. The programs make exhaustive
comparisons of the relationships among the data and arrive at their own
The result is
applications that are eerily similar to human behavior. Gerald Tesauro, a
staff member at the IBM Research Center in Yorktown, N.Y., trained a
neural network to play backgammon by showing it thousands of random moves.
Unlike chess, which is rule-based, "backgammon relies more heavily on
judgment than it does on search," says Tesauro, who constructed his
game on a Unix-based workstation. "Chess programs rely on massive,
brute-force search programs that project various options into the future.
But they don't have to be very smart." When the training was over,
Tesauro's computer could respond to moves it had never before encountered.
In fact, it blew away all commercial backgammon programs.
Not everyone has had
as much success. William H. Harder, a Los Angeles sales representative for
Quotron Systems, says he achieved only mixed results predicting stock
market indexes with Braincel, a neural-net add-in for Microsoft Excel 3.0
from Promised Land Technology (New Haven, Conn.). On the other hand,
Harder may simply be keeping a good thing to himself. "If it does
work, I guarantee you won't know about it," he says. "I'll be
using it myself."
Many companies have
had significant success with neural networks. Chase Manhattan Bank, for
example, has begun to train neural nets to make initial assessments of
loan applicants. "Banks and insurance companies are getting 3%, 4%,
and 5% improvements," says Steven J. Weaver, a senior scientist at
Computer Sciences Corp. in Laurel, Md. "That's literally money in the
bank when you're lending out hundreds of millions."
Using Neural Works
from NeuralWare Inc. of Pittsburgh, Weaver recently devised a neural
network for a more astronomical task: diagnosing problems in the
trajectory of the Gamma Ray Observatory, which was launched in April from
the space shuttle Atlantis.
From soup to
The possible applications are virtually limitless. The Defense
Department has poured millions of dollars into classified systems that
identify targets for fighter aircraft or submarines. On a simpler scale,
Sumitomo Heavy Industries in Japan has created a system-based on
ExploreNet 3000 from HNC Inc. in San Diego-that can distinguish subtle
differences in apples streaming down a conveyor belt. At a rate of 40
apples per second, the machine counts and evaluates defects before
deciding how to grade each piece of fruit.
The Ford Motor Co. is
working on a neural-net application that "listens" to a stream
of data from a car's microprocessors the way an expert mechanic might
listen to an engine's purr. "Long term, it looks like a ripe area for
on-board diagnostics," says Lee Feldkamp, a Ford researcher.
"Near term, we're probably talking about diagnostics at the end of
the production line or at the dealer."
Jumping headlong into
this market, a small stampede of software vendors have sold about 40,000
neural networks in the last two years, according to Tom Schwartz, head of
The Schwartz Associates in Mountain View, Calif.
calls itself the industry sales leader, sells several versions of Neural
Works at prices starting at $ 1,895. California Scientific Software sells
Brainmaker for $ 195 and a professional version for $ 795. Prices for
HNC's ExploreNet 3000 range from $ 600 to $ 1,500. Science Applications
International Corp. (San Diego) sells a neural-net program called Ansim
for $ 495. Neurix (Boston) sells a $ 1,495 Macintosh package called
MacBrain. And Murray Ruggiero, Promised Land Technology's vice president,
says his company may adapt Braincel, the $ 250 Excel add-in, to run with
Most of the vendors
say their products can import 1-2-3 files. The cheaper versions can
process only smaller files and lack some of the features of more advanced
versions, such as C compilers, which let users embed the neural network in
another application. Some of the packages come with proprietary
accelerator boards or math coprocessors to speed up the training process.
But even with a math
coprocessor, nearly everyone who has used a neural net cautions against
expecting instant results. The programs take time to set up, and training
neural-network software on a complex data set can tie up even the most
powerful desktop machines for the weekend.
And after that, it
may take time to learn to trust the results. "Neural networks are not
for the dilettante," confirms O'Sullivan, the currency trader.
"It took me a good year to develop enough comfort and security to
commit my capital."