Memory Efficient Nonuniform Quantization for Deep Convolutional Neural Network

Fangxuan Sun, Jun Lin

Published 2016-07-10Version 1

Convolutional neural network (CNN) is one of the most famous algorithms for deep learning. It has been applied in various applications due to its remarkable performance. The real-time hardware implement of CNN is highly demanded due to its excellent performance in computer vision. However, the cost of memory of a deep CNN is very huge which increases the area of hardware implementation. In this paper, we apply several methods in the quantization of CNN and use about 5 bits for convolutional layers. The accuracy lost is less than $2\%$ without fine tuning. Our experiment is depending on the VGG-16 net and Alex net. In VGG-16 net, the total memory needed after uniform quantization is 16.85 MB per image and the total memory needed after our quantization is only about 8.42 MB. Our quantization method has saved $50.0\%$ of the memory needed in VGG-16 and Alex net compared with the state-of-art quantization method.