Coffee and Kudos
In December 2010, in an effort to encourage the development of better algorithms, an international group of astronomers issued a challenge to computer geeks everywhere: What is the best way to measure gravitational lensing, or the distorting effect that dark matter has on the light from distant galaxies? David Kirkby read about the GREAT10 (GRavitational lEnsing Accuracy Testing 2010) Challenge on Wired.com and decided to give it a go.
Kirkby, a physicist at the University of California, Irvine, and his graduate student won the contest using a modified version of a neural network algorithm that he had previously developed for the BABAR experiment, a large physics collaboration investigating the asymmetry of matter and antimatter. The victory earned Kirkby a co-author credit on the recent paper detailing the contest, easing his switch from the field of particle physics to astrophysics. Also, with the prize money, “we bought a top of the line espresso machine for the lab,” he said.
GREAT10 was one of a growing number of “data challenges” designed to find solutions to specific problems faced in creating and analyzing large physics and astronomy databases, such as the best way to reconstruct the shapes of two galaxies that are aligned relative to Earth and so appear blended together.
“One group produces a set of data — it could be blended galaxies — and then anybody can go out and try and estimate the shape of the galaxies using their best algorithm,” explained Connolly, who is involved in generating simulations of future LSST images that are used to test the performance of algorithms. “It’s quite a lot of kudos to the person who comes out on top.”
Many of the data challenges, including the GREAT series, focus on teasing out the effects of dark matter. When light from a distant galaxy travels to Earth, it is bent, or “lensed,” by the gravity of the dark matter it passes through. “It’s a bit like looking at wallpaper through a bathroom window with a rough surface,” Kirkby said. “You determine what the wallpaper would look like if you were looking at it directly, and you use that information to figure out what the shape of the glass is.”
Each new data challenge in a series includes an extra complication — additional distortions caused by atmospheric turbulence or a faulty amplifier in one of the detectors, for example — moving the goal posts of the challenge closer and closer to reality.
Data challenges are “a great way of crowd-sourcing problems in data science, but I think it would be good if software development was just recognized as part of your productivity as an academic,” Kirkby said. “At career reviews, you measure people based on their scientific contributions even though software packages could have a much broader impact.”
The culture is slowly changing, the scientists said, as the ability to analyze data becomes an ever-tightening bottleneck in research. “In the past, it was usually some post-doc or grad student poring over data who would find something interesting or something that doesn’t seem to work and stumble across some new effect,” Tyson said. “But increasingly, the amount of data is so large that you have to have machines with algorithms to do this.”
Dark Side of the Universe