|Date||March 24, 2017|
|Title||Cognitive IoT: From RoboBees to Machine Learning on the Edge|
|Abstract||Dennard scaling ended several generations ago. Moore’s Law scaling is slowing down. Many new device technologies have been proposed over the years, but none ready to supplant silicon CMOS. Power dissipation is the main performance limiter. What does this all mean for us chip designers? Doom and gloom? Or are we entering a golden age of chip design? We believe there is much to be done, especially given the upcoming explosion of IoT devices. Beginning with an overview of the highly energy- and weight-constrained RoboBees developed at Harvard, efforts and results from the “BrainSoC” motivate our recent exploration of alternative computing approaches to address the energy-constrained, but often compute-intense demands of miniature edge devices. Deep learning methods have transformed many aspects of computing replacing decades-old canonical approaches in many domains. This revolution has been driven by enormous amounts of data, novel machine learning algorithms, and inexpensive, high-performance computing through the cloud. Cognitive IoT is a natural evolution of this technology, but IoT devices are typically tiny devices at the far edge of the network, suffering from low compute power, poor communication bandwidth, and intermittent connectivity.This talk presents Minerva, a highly automated co-design approach across the algorithm, architecture, and circuit levels to optimize deep neural network (DNN) hardware accelerators. Results from a 28nm DNN acclerator test chip confirm performance and energy efficiency enhancements by leveraging specific optimizations and inherent resilience attributes of machine learning algorithms. A common theme for our projects is the importance of specialized hardware acceleration to improve performance and energy efficiency, where cross-layer innovations in software, computer architecture, and chip design can meet the demands of cognitive IoT devices.|
|Bio||Gu-Yeon Wei is Gordon McKay Professor of Electrical Engineering and Computer Science at the John A Paulson School of Engineering and Applied Sciences at Harvard University. He received his BS, MS, and PhD degrees all in electrical engineering from Stanford University in 1994, 1997, and 2001, respectively. After an eighteen month stint at a high-speed links startup in Oregon, he joined the Harvard faculty in 2002. His research interests span a broad range of topics in mixed-signal and digital circuits and systems, computer architecture, design tools, low- and high-voltage power converters, robotics, and more.|
These seminars supported by the Ming Hsieh Institute.