Kartik Hegde Awarded Facebook Fellowship For Plan to Rev Up Deep Learning For Mobile Devices
2/13/2019 9:55:53 AM
Now Hegde will have the chance to advance the impact of mobile computing on a much larger scale, working later this year as a Facebook Fellow. The second-year PhD student was one of 21 Fellows and seven Emerging Scholars announced by Facebook in January from more than 900 applications from around the world.
“I work on building efficient processors for machine learning. Deep learning is (increasingly) in our day-to-day lives,” Hegde said. “When we interact with our phone, when you open your Siri or open any applications, there is some deep learning on the back end that’s helping us.”
The fellowship will begin this fall after Hegde interns at the company over the summer. He plans to use the fellowship to continue research aimed at allowing more of that computing to be done on the phones themselves.
“For Facebook, this is at the core of their business,” said Hegde, whose advisor is Assistant Professor Chris Fletcher. “You have Facebook applications running on everyone’s smartphone, and more and more intelligence is being enabled, thanks to the rapid growth in the field of machine learning.”
For that to happen now, most of the work is done in the cloud – data is sent to the cloud, the computing happens, and then the result comes back to the phone. But there is a communication delay, and the user needs a constant internet connection, Hegde said.
During the fellowship, he’ll work on developing a hardware accelerator that he hopes will contribute to the development of a more efficient processor that would be used in phones. That accelerator, he hopes, will be able to run several deep learning-related algorithms, providing both efficiency and flexibility.
The fellowship will cover Hegde’s tuition and fees for two years, and provide a $37,000 grant and an additional $5,000 for travel to conferences.
Beyond the potential improvements in local computing for phone, Hegde hopes his work can have more widespread impact -- for instance, in the sensors that IoT technology relies on.
“If you can add some small logic that helps do some processing before (a sensor) sends data off to the cloud for higher processing … there is so much possibility there, if there are efficient processors,” he said.
Hegde’s first glimpse of the possibility to make tasks smoother and easier started with an undergraduate project at the National Institute of Technology Karnataka, Surathkal, the one in which he built irrigation device for his father.
The family’s farm in southern India grows a variety of crops -- areca nuts, bananas, and others. Home, Hegde said, is a good distance away from the farm itself. Every time his father wanted to turn irrigation pumps on or off, he’d have to make the trip to the farm.
So Hegde installed an embedded system on the pump with an application that lets his father control it by text.
“It goes anywhere with you – I can literally sit here and operate the pump there,” Hegde said during a recent interview in the Department of Computer Science.
The pump device is far simpler than the sort of deep learning leap Hegde hopes to enable with his research at Facebook.
“If I can do some research that brings deep learning to all the devices around us, then that’s pretty impactful,” he said.