High-performance DNNs at the Edge: Co-optimization of Model Architectures, Compiler and Accelerator

Wednesday, May 22, 2:10 PM - 2:40 PM
Summit Track: 
Enabling Technologies
Exhibit Hall ET 1

Horizon Robotics provides hardware and software solutions for deep learning-based visual and audio perception at the edge. To make these perception systems accurate while also achieving energy efficiency, we jointly design three critical elements in-house: the neural network models, the compiler and the AI accelerator engine. Co-design of these three elements enables us to achieve better performance and efficiency than would be possible if they were optimized independently. In this talk, we present our approach, including manual and automatic techniques for optimization of CNN models. We also briefly present our CNN compiler and CNN accelerator architecture. Finally, we show how this jointly optimized solution meets the needs of several real-world applications.


Yuan Li

Director, Applied AI Lab, Horizon Robotics

Yuan Li is a founding member of Horizon Robotics’ Silicon Valley branch and Director of the company’s Applied AI Lab. Before joining Horizon Robotics, she led teams at Google Research working on computer vision algorithms deployed in various Google products such as Image Search, Photos, Lens and Cloud Vision API. Yuan earned her bachelor’s degree in computer science from Tsinghua University and her master’s degree in computer science from the University of Southern California. She has published work in top tier conferences and journals, and authored the Best Student Paper at CVPR 2008.

See you at the Summit! May 18-21 in Santa Clara, California!