National Tsing Hua University, Taiwan
Hardware-Friendly Convolutional Neural Network Architecture Design
Recent advance in computer vision is almost empowered by convolutional neural networks (CNN). Numerous CNN architectures have been proposed for different tasks and deployed across different platforms. Most popular design consideration are in accuracy, parameter size and number of MACs. We show that the number of MACs is not a proper measure of inference time nor energy consumption, both are important design merits. Instead, we propose taking off-chip DRAM traffic into consideration when designing a CNN architecture. Experiments demonstrate that a proposed architecture called HarDNet delivers superior performance in such tasks as image classification, object detection, and semantic segmentation.
Youn-Long Lin has been with National Tsing Hua University, Taiwan, since 1987, where he is a Chair professor of computer science. He received his PhD from the University of Illinois, Urbana-Champaign, in 1987. His research interest includes electronic design automation, video coding architecture design, and neural network architecture design. He is a co-founder of Global Unichip Corp. and the founder and chairman of Neuchips Corp.
If you wish to modify any information or update your photo, please contact the Publicity Chair at the following address: