Logo image
LG-LSQ: Learned Gradient Linear Symmetric Quantization for Low-Precision Integer Hardware
Conference paper

LG-LSQ: Learned Gradient Linear Symmetric Quantization for Low-Precision Integer Hardware

Shih-Ting Lin, Zhaofang Li, Yu-Hsiang Cheng, Hao-Wen Kuo, Rui-Hsuan Wang, Nai-Jen Sung, Chih-Cheng Lu and Kea-Tiong Tang
AICAS 2023 - IEEE International Conference on Artificial Intelligence Circuits and Systems, Proceeding
2023

Abstract

Machine Learning Model Compression Quantization Artificial Intelligence Computer Vision and Pattern Recognition Hardware and Architecture Information Systems Electrical and Electronic Engineering
To design a deep learning accelerator hardware in edge devices, lower precision weights and activation have advantages regarding area and power. We propose learned gradient linear symmetric quantization (LG-LSQ) as a method for quantizing weights and activation functions to low bit-widths with high accuracy and an integer-only accelerator as hardware for LG-LSQ. First, we introduce the scaling simulated gradient (SSG) method for determining the appropriate gradient for the scaling factor of the linear quantizer during the training process. Second, we introduce the arctangent soft round (ASR) method, which differs from the straight-through estimator (STE) method in its ability to prevent the gradient from becoming zero, thereby solving the discrete problem caused by the rounding process. Finally, to bridge the gap between full-precision and low-bit quantization networks, we propose the minimize discretization error (MDE) method to determine an accurate gradient in backpropagation. In our evaluation on ImageNet, the proposed quantizer achieved full-precision baseline accuracy in various 3-bit networks, including ResNet18, ResNet34, and ResNet50, and an accuracy drop of less than 1% in the quantization of 4-bit weights and 4-bit activations in lightweight models such as MobileNetV2 and ShuffleNetV2.

Metrics

1 Record Views

Details

Logo image