Machine learning, the science that makes it possible for devices to “think” on their own, is taking the next step forward thanks to UT College of Engineering assistant professor Jeremy Holleman and associate professor Itamar Arel, both of whom are in the Department of Electrical Engineering and Computer Science.
Holleman’s expertise comes in the area of analog computational circuits, which map mathematical operations onto transistors and often provide lower power consumption than the more commonly used digital circuits.
Much as brains control the bodies of animals, Arel had the concept for developing a machine-based “brain” model to help control a wide array of devices—an idea based on studies of human cognition.
“We recognized that the merger of his technology with the kind of machine learning systems I was studying had great potential in moving the field forward,” said Arel. “We spent over a year successfully developing a chip that would implement some of the basic building blocks of deep learning, which has given us lots of insight into how such research can grow in the future to yield even more impactful intelligent systems.”
By using Holleman’s analog designs along with Arel’s overall concept, the team was able to reduce energy requirements down to about one-third of one percent of what a digital system would have needed, while also making it possible to put the system on a microchip.
Such a breakthrough paves the way for smaller, more efficient devices in everything from implantable medical devices to military-grade equipment.
“One of the exciting things for us with this project is the new areas that it could open up,” said Holleman. “We’re really excited about the possibilities in store for this technology.”
For example, as devices capable of thinking become smaller less dependent on frequent recharging, the opportunity to make self-dosing medicinal implants or devices capable of synthesizing nerve impulses to damaged limbs becomes more real.
“There are a lot of ideas out there where the required energy makes it impossible to put intelligence where it is needed,” said Holleman. “Brain interfaces to restore movement for paralysis patients is one area where this sort of technology could really help people.”
While such breakthroughs hold promise for a number of medical ailments, Holleman was quick to caution that it could still be years before they become commonplace.
In the meantime, the two professors have secured grants from the National Science Foundation, the Intelligence Advanced Research Projects Activity, and the Defense Advanced Research Projects Agency, which is funding their current project.
The DARPA UPSIDE (Unconventional Processing of Signals for Intelligent Data Exploitation) grant, which came through UT’s Center for Intelligent Systems and Machine Learning and includes fellow associate professor Gong Gu as well as researchers at Stanford University and Oak Ridge National Laboratory, came in at $4.8 million over four years and funds the use of the team’s unconventional computing strategies to improve the video processing systems used by drones.
“This has been a great example of collaboration,” said Arel. “We’ve been able to recruit seven graduate students to study this project, an experience that is really valuable for them.”
Interestingly enough, the path for collaboration was paved by a chance meeting between the two.
While interviewing for a job at UT, Holleman was explaining his ideas to a group that just happened to include Arel.
“I was excited to learn that there was someone in the department that I could work together with in this area,” said Holleman. “We realized how we could help one another immediately.”
As the DARPA project is government-sponsored, the content of this release should not necessarily be viewed as having been endorsed by nor the policy of the government. This release is approved for public viewing with unlimited distribution.
C O N T A C T :
David Goddard (865-974-0683, firstname.lastname@example.org)