Implementation of a Hardware-friendly Quantized CNN using Integer Arithmetic Only

Jaemyung Kim; Yongwoo Kim

doi:10.5573/ieie.2020.57.12.60

Implementation of a Hardware-friendly Quantized CNN using Integer Arithmetic Only

Recommendations

Search

Questions

정수 연산만을 사용하는 하드웨어 친화적인 양자화된 CNN 구현

Type: Academic journal

Author: Jaemyung Kim (인하대학교) Yongwoo Kim (상명대학교)

Journal: The Institute of Electronics and Information Engineers Journal of the Institute of Electronics and Information Engineers Vol.57 No.12(Wn.517) KCI Excellent Accredited Journal

Published: 2020.12

Pages: 60 - 69 (10page)

DOI: 10.5573/ieie.2020.57.12.60

Usage

📌

Topic

📖

Background

🔬

Method

🏆

Result

Implementation of a Hardware-friendly Quantized CNN using Integer Arithmetic Only

Abstract· Keywords

Report Errors

Recently, in the field of computer vision, research using CNN has established itself as a universal research method as it shows excellent performance. However, to obtain high performance, a CNN model with a large number of parameters and high computational complexity is required. This model is not suitable in an environment where hardware resource usage is limited. Therefore, while maintaining the structure of the CNN, quantization research that can optimize computational complexity and memory usage by lowering the precision is being actively conducted. However, most quantization techniques are not hardware-friendly methods were used, such as using a floating-point scale factor to maintain performance. In this paper, quantization aware training(QAT) was performed by applying a hardware-friendly quantization technique and floating-point parameters learned by QAT were converted into integer parameters. In addition, we proposed a method for constructing a CNN that combines basic operation blocks using integer arithmetic only operation and a method for verifying the precision of each layer. As a result of evaluating the performance of the proposed CNN using 8-bit quantized parameters with the CIFAR-10 dataset, it was confirmed that the number of parameters compared to floating-point was reduced to 1/4, but the accuracy drop was less than 0.71%. Besides, it was confirmed that the computational complexity could be reduced by using only an integer operator instead of an additional floating-point operator, and the accuracy drop was less than 0.13% when compared with the conventional quantization aware inference method.

#Deep learning #Image classification #Quantization #Integer arithmetic #Hardware-friendly

요약
Abstract
Ⅰ. 서론
Ⅱ. 양자화 관련 연구
Ⅲ. 연구 방법
Ⅳ. 실험
Ⅴ. 결론 및 향후 연구 방향
REFERENCES

References (18)

Add References

S. Han / 2015 / Deep compression : Compressing deep neural networks with pruning, trained quantization and huffman coding / arXiv preprint arXiv:1510.00149

K. B. Lee / 2019 / Unaligned Pruning for Fast DNN Inference on Mobile Devices / Korea Computer Congress :1412~1414

G. Hinton / 2015 / Distilling the knowledge in a neural network / arXiv preprint arXiv:1503.02531

R. Zhao / 2019 / Improving neural network quantization without retraining using outlier channel splitting / arXiv preprint arXiv:1901.09504

M. Nagel / 2019 / Data-free quantization through weight equalization and bias correction / Proceedings of the IEEE International Conference on Computer Vision :1325~1334

Recommendations

It is an article recommended by DBpia according to the article similarity. Check out the related articles!

Related Authors

김재명

Affiliation 인하대학교

Fields Engineering > Electrical and Electronic Engineering > Electronics Engineering

Articles 1 Usage 226

김용우

Affiliation 상명대학교

Fields Engineering > Electrical and Electronic Engineering > Electronics Engineering TOP 5% Engineering > Computer Science

Articles 62 Usage 3,132

Frequently Viewed Together

Post-training Quantization Technique Based on Scaled Weight Normalization for Lightweight Edge DNN

Jong-Hee Ban , Kyung-Pil Gwon , Joonhyuk Yoo Proceedings of Symposium of the Korean Institute of communications and Information Sciences 2020 .08

Processing Sliding Window Multi-Joins using a Graph-Based Method over Data Streams

Liang Zhang , Byeong Seob You , Jun Wei Ge and others 3 한국공간정보시스템학회 논문지 2007 .09

An Efficient Simulation Platform for the Development of ABS using Hardware-In-the-Loop System (HILS)

Kyung-Sang Cho , Yeun-Hun Choi , Jae-Yon Hwang and others 3 Transaction of the Korean Society of Automotive Engineers 1999 .02

Comments(0)

Write first comments.

UCI(KEPA) : I410-ECN-0101-2021-569-001451103

구분	그룹	데이터 항목
AI 학습용 데이터	원문	원문 PDF 파일
AI 학습용 데이터	원문 + 메타 (기본/상세)	원문 PDF 파일 및 서지정보 CSV
대량 구매용 데이터	B2B 구독 방식	특정 자료 한정으로 원문 접근 권한 부여
대량 구매용 데이터	URL 전달 방식	바로 PDF 뷰어를 열람할 수 있는 URL 제공

구분	그룹	데이터 항목
AI 학습용 데이터	기본 메타	발행기관명, 간행물명, 권호명, 권(vol), 호(issue), 통권, 발행연도, 발행월, 논문명, 저자명, 시작페이지, 종료페이지, 전체페이지, 상세페이지URL
상세 메타 데이터	발행기관 메타	발행기관 이명, 영문명, 창립연도, 홈페이지URL, 발행기관 소개
	간행물 메타	부제목, 간행물 유형, ISSN, ISBN, 최초발행연도, 폐간연도, 간행빈도, 발행주기, 등재사항, 이용수, 피인용수, 권호수, 논문수, 표지이미지
	논문 메타	작성 언어, 부제목, 대등제목, 목차, 키워드, 초록, 이미지, 참고문헌, 이용수, 피인용수, 논문활용도, DBpia통합주제분류, KDC분류, DDC분류, 한국연구재단분류, UCI, DOI
	저자 메타	소속기관, 소속부서, 직급, 연구분야, 연구키워드, 이용수, 피인용수, 저자 논문활용도

구분	그룹	데이터 항목
※ 결합형/맞춤형 메타 데이터는 신청 내용에 따라 다양하게 제공 가능
이용순위 정보	주제분야별 많이 이용된 논문	“인문학”에서 많이 이용된 논문 TOP100
	이용기관별 많이 이용된 논문	“중고등학교”에서 많이 이용된 논문 TOP100
	세부기관별 많이 이용된 논문	“서울대학교”에서 많이 이용된 논문 TOP100
	키워드별 많이 이용된 논문	“Chat GPT”에서 많이 이용된 논문 TOP100
키워드 정보	많이 이용된 키워드	특정기간/분야/저널 내 많이 이용된 키워드
	많이 발행된 키워드	특정기간/분야/저널 내 많이 발행된 키워드
	많이 검색된 키워드	특정기간/분야/저널 내 많이 검색된 키워드
	연구 트렌드 키워드	특정 키워드 연관 연구동향 분석 데이터 키워드

논문 기본 정보

Abstract· Keywords

AI Summary

Topic

Background

Method

Result

주요내용

Contents

References (18)

Recommendations

Related Authors

Frequently Viewed Together

Recently viewed articles

Comments(0)