|
Research Magazine > ARCHIVE > Spring
93 > Article
Artificial Intelligence Authentic Innovation
by David L. Hart
If two heads are better than one, a scientist would probably enjoy
taking an extra brain to work. The spare brain could toil away at a research
project long after the researcher slips home for supper.
Of all the gadgets in the scientific toolbox, computers are the closest thing
to an extra brain that researchers can muster so far. And, frankly, computers
just don't measure up.
Sure, a computer can instantly recall billions of bits of data, process them
and flash them up on a screen.
But the human mind can still do one thing computers can't even dream about:
Make a judgment call.
All a computer can do is what some human being programmed it to do.
So far, anyway.
An eclectic assembly of University of Georgia scientists is working to change
that. They are teaching computers to learn from their own mistakes and answer
questions computers have never faced before.
In laboratories across the campus, scientists are putting "neural networks" --
named after the neurons in the human brain -- to work on problems as disparate
as how to identify a specific carbohydrate molecule from thousands of potential
patterns, or how to detect a ripe tomato.
Those are tasks that traditional computers just cannot do. Unlike humans,
most computers cannot cope with situations they are not explicitly programmed
to
handle. When something doesn't fit the program, it simply "does not compute."
Neural networks are different. They do more than just crunch numbers. They
use experience, almost like people, to deal with new situations.
So far, scientists have found at least two ways in which artificial neural
networks and human brains are similar:
- Both work surprisingly well at diverse
tasks that pose problems for traditional computing techniques;
- And no one knows precisely how either one works.
"Neural networks are still an art," said Dr. Ron McClendon, professor
of biological and agricultural engineering. "You have to experiment
some."
That experimenting has engaged scientists from a variety of academic
disciplines -- and it has put computers to work in ways that have
never been tried before.
For one, researchers at the university's Complex Carbohydrate
Research Center are using computers to identify molecules by something
akin to a fingerprint.
Instead of a detective's fingerprint pad, computers scan a "fingerprint" based
on the magnetic properties of the molecular nuclei, called the nuclear magnetic
resonance (NMR) spectrum.
Before the scientists can find out what a particular carbohydrate does, they
may spend as much as a year figuring out how the molecule is put together.
The NMR fingerprint is their first clue, according to Dr. Bernd Meyer, associate
professor in biochemistry at the CCRC.
The natural first step is to compare an unidentified molecule with others in
a file of other molecular fingerprints. However, matching one fingerprint at
a time from thousands on file takes a long time for computers and scientists
alike. Searching the fingerprint file by name is easy, but to do that, scientists
must first identify the mystery molecule's structure -- which was the original
problem.
Molecular Fingerprints
Conventional computer methods can help, but despite
being slow, they are inflexible. Fingerprints are often "smudged," or
filled with analytical noise from equipment pushed to its sensitivity limits.
If the files contained
a perfect match, computers could find it, but smudged prints give them fits.
A more useful file-matcher would handle these smudged prints. And if it couldn't
find an exact match, it would be nice if it returned the closest matches, which
provide important clues for the molecular detective. Ideally, it would be even
nicer if it could distinguish between one version of a molecule that has nearly
the same structure as another, but has a side group that points in a different
direction.
This subtle property, called "chirality," can turn a useful medication
into a harmful one. Not even skillful scientists can detect this difference
from the NMR fingerprints alone.
After reading about neural networks in the journal Science, Dr. Jan Thomsen,
a project director at the CCRC, decided that they might solve the researchers'
circular dilemma.
"I was maybe gambling a little on this one," Thomsen said. But his
gamble paid off. "It was much more successful than anyone had anticipated."
Meyer and Thomsen stored their NMR spectra in a neural network database system
they call the ANalytical CHemical Object Verification sYstem, or ANCHOVY for
short.
ANCHOVY matches mystery fingerprints in a fraction of a second, and when there
is no exact match, it returns the closest matches. Plus, it can track down
a match from a fingerprint that has more noise than humans can deal with. In
fact, it can be trained to identify molecules from spectra that have up to
25 times the noise usually allowed, Meyer said.
ANCHOVY can even do scientists one better: It can learn to use fingerprints
to differentiate molecules by their chirality.
"It has always worked better than you would expect from a scientist," Meyer
said. "We all have been surprised over and over again at how easy it is
and how powerful it is."
The major stumbling block for the system is training large databases. ANCHOVY
may take four days to memorize 1,000 molecules. Days can seem like forever
in computer time, but ANCHOVY's networks only have to be trained once. Afterwards,
it has split-second recall.
However, to be useful to a pharmaceutical company, Meyer said, the system would
have to store upwards of 100,000 fingerprints, which would make training unmanageable.
Meyer and Thomsen solved their problem by breaking the database into manageable
chunks. For 10,000 fingerprints, the database uses 10 networks. ANCHOVY then
rapidly searches the 10 separate databases to find the closest matches.
Thomsen pointed out that separate databases makes logical sense as well, since
there is no real reason to store, for example, carbohydrate and steroid fingerprints
in the same database.
Meyer sees great potential for ANCHOVY in storing a lab's collection of
fingerprints. "Every
chemical research lab in the world would want such a system," he said.
While the CCRC researchers found that neural networks could solve a very
complex task, others in the College of Agricultural and Environmental Sciences
used them to solve an easy one -- in fact, so easy for people that it seems
boring, repetitive and tailor-made for a computer.
There was just one problem: Conventional computer methods couldn't do it.
The task is "egg candling," the process of scanning eggs for
cracks and other defects. For every egg you buy in the supermarket, a pair
of eyes
at a processing plant has made sure it is crack-free. To find defects, each
egg is placed in front of a light -- it used to be a candle -- which highlights
cracks, dirt stains and blood spots.
Two human candlers might inspect 240,000 eggs a day. To give people a break,
so to speak, Dr. Ron McClendon, Dr. John Goodrum and graduate student Veren
Patel taught a neural network to find cracks in the computerized image of a
candled egg, which the machine learned to do better than a novice human candler.
Based on the success of this network, they are developing networks for other
applications, such as detecting dirt stains and blood spots.
The egg candling network takes about an hour to train, using a set of 90 good
and 90 cracked eggs. The machine examines all 180 eggs each second, so in an
hour it looks at 650,000 eggs, give or take a few thousand.
Of course, once it is trained, the network takes a fraction of a second to
compute its result. But if it only trains with 180 eggs, how could it possibly
learn to distinguish a quarter million different eggs a day?
The answer is, it doesn't.
People don't either, for that matter. A human training for this job wouldn't
memorize 180,000 or even 180 eggs. Instead, by trial and error a person learns
to judge whether an egg is cracked by learning what cracks look like in general.
That's the concept that makes neural networks so novel; they can learn to generalize
from 180 eggs to just about any egg.
It also sets the egg-candling network apart from the ANCHOVY database. Both
projects use the same type of network and learn by trial and error, but the
egg-candling system learns to generalize and ANCHOVY learns to memorize.
DOWN-TO-EARTH APPLICATIONS
To attack a completely different problem with a computer normally requires
writing a new program from scratch. As ANCHOVY and the egg-candling network
show, however, the same basic components can learn different tasks. Only
the training changes.
Using the same variety of network from the egg-candling project, McClendon
has addressed other projects which also require the computer to make a judgment
call.
Being able to generalize allows neural networks to surpass more common statistical
models. For example, a computer simulation can determine the ideal greenhouse
temperature for a given set of conditions.
But the simulation is slow -- too slow to be implemented in a small, quick
computerized thermostat.
So McClendon, with Dr. Ido Seginer, a visiting professor from the Technion
in Israel, taught a neural network to remember what the simulation calculates
for each situation.
The network short cuts the simulation and recalls the right temperature quickly.
In projects with Dr. Gerrit Hoogenboom, assistant professor of biological and
agricultural engineering at the Georgia Agricultural Experiment Station in
Griffin, McClendon has used neural networks to improve the performance of traditional
statistical models for forecasting solar radiation and for predicting crop
development.
In another agricultural application, a neural network suggests when to spray
a crop with pesticide and seems to outperform an expert system for the same
task, McClendon said.
But the average person doesn't often make judgment calls about the amount
of solar radiation or crop development. "Real" people have other
problems to worry about, like finding ripe vegetables at the grocery store.
Dr. Chi Thai, associate professor of biological and agricultural engineering
at the Griffin experiment station, said he thinks neural networks have potential
here, too. His goal is to teach them how shoppers choose fresh produce in grocery
stores. "Sure, everyone wants to buy a tomato. But exactly how red?" he
asked.
It's not just a rhetorical question. If his networks can mimic just how ripe
shoppers like their tomatoes -- and do that in conjunction with mathematical
models that predict how long it will be until they get that ripe -- grocery
stores can arrange to have tomatoes that will have the best color.
One of Thai's networks maps measured values for tomato color to subjective
consumer preference scores. Another network has learned to differentiate between
the `Sunny' and `Sunbeam' varieties of tomato by their color spectra, and yet
another has attempted to predict how fast tomato color will change from green
to red. (This one only succeeded for one variety -- for reasons Thai can't
quite explain.)
Traveling Salesmen and the Human Genome
Simple stuff, you say? Not on your life. Try explaining your mental processes
when you choose ripe tomatoes. Better still, explain how you would find the
shortest route that goes through every city in a given region and get back
to the starting point. Researchers call it the traveling salesman problem.
On the surface, this problem doesn't seem to have much to do with genetics.
But for Dr. Jonathan Arnold, associate professor of genetics and statistics,
finding the salesman's route is akin to constructing a map of the human genome,
and he's using a neural network to tackle the problem.
"The most important medical diagnostic tool of the 20th century will be
a map of the human genome," Arnold said. "About one-third of all human
diseases have a genetic basis."
With that much at stake, how do you go about drawing the map?
To study long DNA strands, researchers break them into fragments they can examine.
They are then left with the immensely complex task of reassembling the pieces
-- in exactly the right order. A wrong turn for the salesman is the same as
a misplaced gene in the genome.
"It would be nice to have an idea of where the fragment came from in the
chromosome," Arnold said. "You're usually only interested in one little
piece."
Dr. Hubert Chen, a professor of statistics, and statistics graduate student
Momiao Xiong are working with Arnold on the DNA mapping problem. They have
devised a new neural network model that addresses these so-called "optimization" problems.
Its success in solving the DNA mapping problem could mean that a whole class
of "unsolvable" problems might also be addressed more quickly and
more precisely with neural networks.
But Chen and Xiong have set their sights even higher. They want to know exactly
how neural networks do what they do. Until now, that's been a mystery.
"Roughly speaking, the neural network is a very complicated (mathematical)
function," Chen said. "It uses some kind of learning process and empirical
knowledge to predict the future or predict an unknown situation."
If a neural network is a mathematical function, it should be possible to figure
out how the machine learns. In fact, Chen and Xiong eventually want to present
a unified learning theory that explains how all the varieties of neural networks
work. This would help researchers understand when and how neural networks would
apply to particular situations.
"Because this neural network is new to most fields, it can be used in genetics,
biology, computer science, statistics, education, engineering, business -- almost
anything you can name," Chen said. "Because the real world is so complicated,
there is no [single] model that can fit."
The mystery of neural networks isn't solved yet. Even after much success building
and training neural networks, McClendon and other researchers know them only
as well as most drivers know their cars. They can fill up the tank, turn the
key and drive, but the details of internal combustion remain hidden under the
hood.
Researchers and neural network "artists" must know when networks
might work and know what their limits are. When they build these little brains,
researchers still must decide how many nodes to use. Too few, and the network
cannot discriminate well enough; too many, and the network starts to memorize
instead of generalize.
They must also consider when to stop training, since that also affects whether
a network generalizes or memorizes. When Thai started using neural networks,
he fed them all sorts of information. But networks, unlike people, aren't smart
enough to filter the junk from the relevant information. A network bases its
results on everything it has available. So rather than being an extra brain
in a researcher's toolbox, networks might be more like a sophisticated hammer.
"I consider them just another tool, like statistical regression," Thai
said.
Neural networks aren't ready to replace people just yet. Although they were
modeled after the brain because of its flexibility and power, a neural network
is actually the ultimate one-track mind.
"Our brain is much more complicated than the computer," Chen said. "It
is our computer. There is a similarity, therefore we can use the computer to
solve some real complicated problems."
The brain still outperforms neural networks, but Chen is optimistic about networks.
"We think we can get very close," he said.
David L. Hart is a graduate student at UGA's Henry W. Grady College
of Journalism and Mass Communication. He earned bachelor's and master's
degrees in computer science at the Georgia Institute of Technology and
Carnegie-Mellon University, respectively.
Research
Communications, Office of the VP for Research, UGA
For comments or for information please e-mail the editor: rcomm@uga.edu
To contact the webmaster please email: ovprweb@uga.edu
|
| |
|
|
|
|