Deep learning is performing wonders, automating everything from driving to recognizing speech to composing music. Now, it's poised to fire up scientific discovery, thanks to new software created by scientists at the U.S. Department of Energy's Oak Ridge National Laboratory (ORNL).
Working on the GPU-accelerated Titan supercomputer at ORNL, the team developed an algorithm that automatically generates neural networks. Modeled loosely on the connections in the human brain, these do the "learning" in deep learning.
MENNDL, short for Multi-node Evolutionary Neural Networks for Deep Learning, evaluates, tests and recommends neural networks for unique datasets like those that scientists collect. And with GPU acceleration, it's fast, reducing what can be a months-long endeavor to a matter of weeks.
"MENNDL is about saving people time and making scientific discovery happen faster," said Steven Young, a research scientist who's part of ORNL's Nature Inspired Machine Learning team.
AI for Scientists
Although the ORNL team created MENNDL for scientists, it has the potential to transform AI more broadly. By training a neural network, researchers create software to perform certain tasks. ORNL's software creates the network itself, eliminating the trial-and-error process normally required to configure one.
Scaled across Titan's 18,688 Tesla GPUs, the algorithm simultaneously tests and trains thousands of potential networks to predict those best suited to the job.
In many fields, researchers use existing neural networks or datasets as a launching pad for their deep learning efforts. That's not possible for scientists, whose data comes from scientific instruments and looks a lot different from what's used to teach computers to recognize faces or understand speech.
"At the lab we work with data that's pulled from a neutrino detector, electron microscope or some other scientific instrument," Young said. "It's a far cry from pictures of cats and dogs."
MENNDL is already speeding research in neutrino physics. Neutrinos are subatomic particles that scientists believe could shed light on such mysteries as the origins of the universe and the nature of matter.
Because neutrinos are notoriously hard to detect, scientists at DOE's Fermi National Accelerator Laboratory (Fermilab) use high-intensity beams to study how they react with ordinary matter. That produces mountains of data, which researchers must analyze to identify precisely where the interaction occurred.
In the past, the Fermilab team would have spent months testing neural networks to find one that would work for their problem, Young said. MENNDL did it in just 24 hours.
"Instead of having scientists play around with deep learning frameworks for months, MENNDL gives them a network that will work with their data in just a day," Young said.
That lets researchers conduct more experiments in less time -- and advance science more quickly.
For more information, see ORNL's paper, Optimizing Deep Learning Hyper-Parameters Through an Evolutionary Algorithm.
To learn more about the latest research at ORNL on deep learning and supercomputing, attend the GPU Technology Conference, March 26-29, in Silicon Valley. Register now.
* The main image for this story shows the inside of the MiniBooNE neutrino detector at Fermi National Accelerator Laboratory.