Lin wins Best Paper for work on energy-efficient, reliable mobile computing techniques

02/27/2017

CSL graduate student Yingyan Lin has won a Best Paper award at the IEEE International Workshop on Signal Processing Systems for her work on energy-efficient convolutional neural networks. Lin’s research will help improve the energy efficiency of a certain class of machine learning algorithms used for mobile computing, while maintaining reliability and performance.

Yingyan Lin
Yingyan Lin
“In order to continue to advance mobile computing applications, we need to increase on-device computing capabilities, instead of offloading work to the cloud,” said Lin, who is pursuing her Ph.D. in electrical and computing engineering at Illinois. “But these applications are very computionally intensive and drain batteries, which is a real challenge for mobile devices.”

To run many of today’s applications, devices send jobs to the cloud, which uses supercomputers in data centers to process the data and then sends it back to the device. This transfer to and from the cloud could require as much as nine times more energy than it would take to process the data on the device.

However, mobile devices have limited battery power, making it difficult to do all computations locally. While engineers are working to improve energy-efficiency by aggressively reducing the supply voltage, a common side effect is a loss in reliability.

Working with advisor Naresh Shanbhag and ECE Illinois graduate Sai Zhang , Lin developed a new statistical error compensation technique that leverages the inherent redundancy that exists within the algorithms for low-cost error detection and compensation. With this architecture, individual underlying devices and circuits can make errors without jeopardizing the performance of the entire system. The approach could provide energy savings up to 10 times while enhancing system robustness by more than 100 times over a conventional approach, according to Lin

“Yingyan’s work shows how Shannon-inspired statistical computing techniques can be used to great effect in reducing the energy requirements of complex machine learning kernels in current day application scenarios," said Shanbhag, the Jack Kilby Professor of Electrical and Computer Engineering. "Her work pushes the limits of what can be achieved in terms of energy efficiency and robustness in scaled semiconductor process technologies.”

The work was funded through the SONIC Center, a multi-university research center led by Illinois that designs robust, energy efficient, and intelligent nanoscale computing platforms. SONIC is a STARnet project, sponsored by the Semiconductor Research Corporation and DARPA.

Lin’s work could pave the way for new applications, such as mobile medical devices that can monitor health and communicate results with doctors.

“In addition to having more reliable and energy-efficient devices, we could offer greater privacy of medical data, because all the computation is being done locally instead being sent to the cloud,” she said. “I really feel like this research can help improve lives.”