Mixture of Deterministic and Stochastic Quantization Schemes for Lightweight CNN

Abstract

There has been a breakthrough in the field of image classification and object detection, owing to the development of GPU and deep learning. However, because of the huge computation of deep learning, it is hard to use the deep learning algorithms in an embedded platform or a mobile device. Therefore, many compression studies have been conducted, and one of the most popular methods is a parameter quantization. In this paper, we propose an adaptive quantization scheme that reduces the loss of accuracy due to the quantization by properly mixing deterministic and stochastic quantization methods, while retaining the characteristics of the hardware-friendly fixed-point quantization method. By applying the proposed method to the weight parameters of image classification and object detection networks, the proposed method shows better mean average precision (mAP) of up to 0.44% in image classification and 0.91 % in object detection.