Yiyu Shi, professor of computer science and engineering, has been named the recipient of the 2021 IEEE Transactions on Computer-Aided Design Donald O. Pederson Best Paper Award for the work titled “Hardware/Software Co-Exploration of Neural Architectures.”
The winning paper, which was selected from among more than 800 papers published by the journal in the last two years, proposed hardware/software co-exploration for deep neural network accelerators.
“Before our work, neural networks were typically optimized for a fixed hardware system,” Shi said. “We show that by simultaneously optimizing the hardware design and neural network search architecture — a co-exploration framework — a much better trade-off between energy, speed, and accuracy can be achieved.”
Co-exploration is especially important as the number of applications and devices that combine edge computing and machine learning — cell phones, laptops, smart homes, self-driving vehicles, unmanned air vehicles (UAVs), and security cameras — is exploding. Users want faster, more accurate, and expanded performance options. But these devices have limited hardware resources and battery life. Existing neural architecture search frameworks cannot keep up with usage demands.
The framework developed by Shi, his Notre Dame team, and in collaboration with Associate Professor Jingtong Hu at the University of Pittsburgh, expands hardware design freedom, accelerating applications such as implantable medical devices, UAVs, and more.
“In the same way that hardware-software co-design enabled faster, more accurate computing, our co-exploration framework promotes faster real-time response, higher accuracy, and prolonged battery life to make the most of the limited hardware resources on edge devices,” Shi said.
The award will be presented during the International Conference on Computer Aided Design, taking place virtually Nov. 1-4, 2021.
— Originally published by the College of Engineering on Nov. 1, 2021.