< BK21+ BEST Seminar Series Announcement> 

Time and Date : 11:00 ~ 12:00 Thursday 01/30/2020

Place : D503, Engineering Building #4

Title : Towards Sub-4-bit Quantization of Optimized Mobile Networks
Abstract:
4-bit quantization of optimized mobile models such as MobileNet-v3 is challenging due to recently adopted architectural optimizations. According to our observation, it is mainly limited by activation instability incurred by weight quantization (AIWQ). In order to mitigate the effect of AIWQ, a novel training method called BN last (BLast) is proposed. In addition, a differentiable quantization method called DuQ is proposed that improves upon the state-of-the-art quantization method. Our experiments show that the proposed methods offer 4-bit weight/activation quantization of MobileNet-v3 on ImageNet at a negligible accuracy loss (0.916 %) from the full-precision top-1 accuracy.

Presenter: Eunhyuk Park, Researcher / Seoul National University

Host: Prof.Kim, Hanjun, Yonsei EEE