Researchers from Stanford University and the SLAC National Accelerator Laboratory have just taught the world something brand new about neural networks, a kind of artificial intelligence (AI). Not only can these networks be used to accurately analyze gravitational lenses, they can also produce results 10 million times more quickly than traditional methods. Their work has been published in Nature.
“Analyses that typically take weeks to months to complete, that require the input of experts and that are computationally demanding, can be done by neural nets within a fraction of a second, in a fully automated way and, in principle, on a cell phone’s computer chip,” study co-author Laurence Perreault Levasseur explained in a press release.
Gravitational lensing provides insight into how mass is distributed in space, both now and over time. This helps us understand how the universe is changing and expanding, and it also provides valuable insight into the dark matter that makes up the lion’s share of our universe — about 85 percent.
Until now, though, analysis of gravitational lensing has been a slow, deliberate process. This is because it required researchers to compare actual images of lensing from space — taken using telescopes like Hubble and the in-development Large Synoptic Survey Telescope (LSST) — with computer simulations of mathematical lensing models.
To achieve the same results in seconds using neural networks, the researchers first trained the networks by showing them approximately half a million simulations of gravitational lenses over the course of a single day. That was all the training the neural networks needed to produce analytical results that were as precise as those provided by their human trainers — and they were almost instantaneous.
This breakthrough is more advanced than any other recent astrophysics applications of neural networks.
Previously, they had been limited to solving problems of classification, such as deciding whether an image showed a gravitational lens or not. Now, the networks go far beyond simple identification of the phenomena, providing all of the analytical insight a human scientist would.
This capacity for processing and analyzing massive amounts of data will become even more critical very soon. Astrophysicists will have access to more images from space than ever before once the LSST is in operation. The device is predicted to increase the detected number of gravitational lenses, for example, from the few hundred we’re aware of now to tens of thousands.
“We won’t have enough people to analyze all these data in a timely manner with the traditional methods,” explained Perreault Levasseur. “Neural networks will help us identify interesting objects and analyze them quickly.”
Furthermore, although these initial tests were conducted on the Stanford Research Computing Center’s Sherlock high-performance computing cluster, they could have been completed on a laptop or even on a smartphone. This will leave the more powerful systems free for scientists to use for non-uniform work — something they’ll have time for now that the neural networks are ready to assist with the more tedious tasks.
“This will give us more time to ask the right questions about the universe,” said Perreault Levasseur.