Growing a Grid at Skidmore College
Flip Phillips found an "insightful" way to accelerate
his research this summer at Skidmore College. Phillips and his colleagues
are studying human vision, primarily 3D shape perception, or "how
we see what we see."
Phillips's focus is on empirical studies and actual laboratory testing. When
these empirical studies do not match up to existing theoretical models,
the onus is on Phillips to prove or disprove their validity, and he is
using gridMathematica as
one of his primary tools for doing so.
Testing the existing theoretical models requires computations--and
lots of them. Phillips, who has an arguably fast 2 GHz dual-processor
PowerMac G5, was faced with a daunting task when he realized that
his machine would take 1.5 to 2 months for each individual
test of a popular theoretical model against his current empirical
data. So he took advantage of Skidmore's Unlimited site program for
all Wolfram Research products and sought out the resources of the
gridMathematica computing cluster in the Skidmore Computer
Science department.
Skidmore's computer science lab includes a rack of dual-processor
Apple Xserve G4 machines. Even though using this setup significantly
reduced his computing time to only 2 to 2.5 weeks per test, Phillips
still wanted a faster way to get results. This past summer, when
computer usage on campus was low, it occurred to Phillips to extend
his grid by "scavenging" for all the Macintosh machines sitting idle
throughout the campus network.
Phillips wrote a series of programs, run on the main grid, to go out
and find Macs with free time to use. When a free machine is found, the
program loads gridMathematica and any needed data, and starts
to calculate. If the machine is needed before the current process is
completed, Mathematica
finishes its calculations at a slowed pace in the background until
the machine is free again.
Using gridMathematica, Phillips has been able to reduce computation
times for the exact same processes from 1.5 to 2 months down to only 4 to 6
hours. With some adjustments, such as reviewing intermediate results and
"pruning off" undesirable branches, the process could be made even faster.
Phillips was inspired by an experience from when he worked at Pixar in the
1980s. After a friend won the student Academy Award for best animated short
film, part of the prize winnings were used to send the sketches overseas
for colorization. When the thousands of original drawings were lost in
transit, they found themselves trying to transfer the images from film to
computer data in order to recreate them. To help accelerate the
process--which still took about a year--they wrote a program that
performed a similar "scavenging" process on all of the Pixar laser
printers.
Having finally found a way to perform the necessary calculations in
much less time than ever before, Phillips and his colleagues are
preparing a paper for publication. Interestingly enough, what they
have found is that the model they set out to disprove is not entirely
incorrect. The theoretical model predicts the empirical results for a
constrained set of circumstances. However, for a whole range of other
conditions the model failed to accurately predict the empirical
results.
Using gridMathematica has allowed Phillips to take his vision
research to the next level. Rather than rejecting or accepting a theory as
a whole due to computational constraints, he is now able to find out what
parts of the theory may be usable, and under what conditions.
| 
Any questions about topics on this page? Click here to get an individual response.
![[ja]](/common/images2003/lang_bottom_ja.gif)