< retour aux séminaires

Sungjoo YOO, Seoul National University (Full professor)

Theme: Low precision in neural network training and inference
Date: January 7th, 2019 - 10:30-12:00, Viallet - Amphi Gosse


Sungjoo YOO received Ph.D. from Seoul National University in 2000. From 2000 to 2004, he was researcher at system level synthesis (SLS) group, TIMA laboratory, Grenoble France. From 2004 to 2008, he led, as principal engineer, system-level design team at System LSI, Samsung Electronics. From 2008 to 2015, he was associate professor at POSTECH. In 2015, he joined Seoul National University and is now full professor. His current research interests are software/hardware co-design of deep neural networks and machine learning-based optimization of computer architecture.


Low precision can improve the efficiency of resource usage in neural network training and inference. In this talk, we present outlier-aware quantization which reduces by 9 times memory capacity cost in training ResNet-50 for ImageNet without losing accuracy. We also present a 4-bit hardware accelerator realizing outlier-aware quantization which offers by 30 % better energy efficiency than the state-of-the-art zero-skipping 8-bit accelerator.

Finally, we report our recent results of 3-bit weight/activation quantization of ResNet-50 without accuracy loss and discuss the feasibility of 2 or 1-bit quantization.