[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
No Subject
Making the task more challenging, many things that are not life forms appear
lifelike, while many true life forms appear to be non-life.
Buying or creating a single computer to conduct the search is out of the
question since at least 100 million images will have to be stored digitally
and scanned, and classifying these images will require 10,000 times the
computing power it took to produce the animated feature film Toy Story , one
of the current standards in supercomputing.
Instead, Noever - working together with Dr. Subbiah Baskaran, a visiting
scientist from the University of Vienna Institute for Molecular
Biotechnology, and Helen Matsos of NASA/Marshall - plans to borrow a few
thousand computers to build what might be called the first D'Arcy Machine, a
computer dedicated to classifying images for tell-tale biological shapes.
Before considering extraterrestrial sources of life, however, the technology
must be in place for an extensive classification of the only life forms we
know - life on Earth.
With a little help from my friends
Named after the original morph man, the D'Arcy Machine will borrow
processing power from volunteer computers connected to the Internet around
the world to perform the giant task.
"We hope to get young scientists from elementary school through college to
help us with the search by linking their computers to the D'Arcy Machine,"
said Noever.
Noever and his colleagues plan to develop the "Book of Life"
technology using neural networks and evolvable hardware -
rewriteable computer chips capable of learning multiple patterns or images
as they process information. Testing the system's image recognition ability
and cataloging life forms from Earth will be the first of three project
phases.
One of the Allan Hills meteorites was
cut off for examination. Studying large specimens at high magnifications
will be like scouring a desert by hand in search of fossil fragments.
"In Phase One, we will construct image-based family trees of living forms as
distinct from inorganic shape features," said Noever, who plans to feed the
new machine at least 100,000 images to get it started. The goal for this
phase is peer-reviewed publication and presentation at the 1998 conference
"On Growth and Form" highlighting scientific progress in the 50 years since
D'Arcy Thompson's death.
In the second phase, the D'Arcy Machine will use trained neural networks
from Phase One while being re-trained to simultaneously acquire and classify
new, often ambiguous images. Noever and his colleagues will also throw the
machine some curve balls with artificial data to test its performance.
The goal of the third phase is for the D'Arcy machine to automatically
acquire and classify images with minimal human supervision. At this stage,
the machine will be equipped for future search scenarios, including the
examination of meteorites found on Earth and lunar or interplanetary samples
retrieved from new space missions.
A lab assistant that doesn't get tired
"The most exciting aspect of artificial intelligence is the way it can be
applied to so many different problems," Noever said, such as his work on the
In Virtuo program which Discover magazine has selected as the top computer
software innovation the year. This software grew from earlier work funded by
NASA's biotechnology research program to investigate the structures of
proteins.
Whereas traditional methods of searching for drugs, or searching for life on
Mars for that matter, require scientists to labor through a lengthy process
of trial and error, artificial intelligence software evolves as it searches.
Noever likes to compare it to solving Rubik's Cube. A supercomputer randomly
working all possible solutions would take about a billion years to get the
right answer. In 1983, a Los Angeles high school student set the world's
record at just under 23 seconds. If a random search takes too long, then
teaching a computer to see patterns like a human might interpret them
becomes the challenge to AI researchers: How to empower a software program
with some kind of autonomous learning?
AI software starts with a few mediocre solutions to problems, and
then develops several variations on these solutions based on the
outcome of initial calculations. The process repeats itself again and again
until a workable number of refined solutions are found for human review.
Like evolution, Noever's AI technology finds the fittest candidates.
"Before putting the engineer's precision to the final candidate, we first
let the computer go to work for us" said Noever.
But computers aren't doing all the work. Noever is conducting innovative
research in space flight experiments to make improved forms of Aerogel, a
superinsulation with broad applications, and other areas.
Papers
1. Noever, D. A., S. Baskaran, "Steady-State vs. Generational Genetic
Algorithms: A Comparison of Time Complexity and Convergence
Properties", 1992, Santa Fe Institute, 92-06-032 (online abstract)