ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory

Sifan Sun; Jinming Ge; Jinyu Bai; Wang Kang

doi:10.1109/NVMSA58981.2023.00018

ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory

Sifan Sun
, Jinming Ge
, Jinyu Bai
, Wang Kang

School of Integrated Circuit Science and Engineering

Beihang University

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

1 Scopus citations

Abstract

Network quantization can effectively reduce the complexity without changing the network structures, which is conducive to deploying deep neural networks (DNN) on edge devices. However, most of the existing methods set the quantization precision manually and rarely consider the case that the computing array is limited, such as computing-in-memory (CIM). In this paper, we introduce a novel method named ES-MPQ, which employs evolutionary search to achieve mixed precision quantization with a small calibration dataset. The ES-MPQ can optimize multiple objectives to achieve better hardware efficiency. The experimental results for ResNet-18 on CIFAR-10 show that the proposed ES-MPQ can reduce the parameter size and energy consumption by up to 1.89x and 2.81x, respectively, compared with the fixed bit-width (8 bits) quantization, while losing only 0.59% accuracy.

Original language	English
Title of host publication	Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023
Publisher	Institute of Electrical and Electronics Engineers Inc.
Pages	38-43
Number of pages	6
ISBN (Electronic)	9798350344967
DOIs	https://doi.org/10.1109/NVMSA58981.2023.00018
State	Published - 2023
Event	12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023 - Niigata, Japan Duration: 30 Aug 2023 → 1 Sep 2023

Publication series

Name	Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023

Conference

Conference	12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023
Country/Territory	Japan
City	Niigata
Period	30/08/23 → 1/09/23

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

SDG 7 Affordable and Clean Energy

Keywords

Mixed precision quantization
computing-in-memory
evolutionary search

Access to Document

10.1109/NVMSA58981.2023.00018

Cite this

Sun, S., Ge, J., Bai, J., & Kang, W. (2023). ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory. In Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023 (pp. 38-43). (Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/NVMSA58981.2023.00018

Sun, Sifan ; Ge, Jinming ; Bai, Jinyu et al. / ES-MPQ : Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory. Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023. Institute of Electrical and Electronics Engineers Inc., 2023. pp. 38-43 (Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023).

@inproceedings{8b3ba16246e4455697916d944ee616d7,

title = "ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory",

abstract = "Network quantization can effectively reduce the complexity without changing the network structures, which is conducive to deploying deep neural networks (DNN) on edge devices. However, most of the existing methods set the quantization precision manually and rarely consider the case that the computing array is limited, such as computing-in-memory (CIM). In this paper, we introduce a novel method named ES-MPQ, which employs evolutionary search to achieve mixed precision quantization with a small calibration dataset. The ES-MPQ can optimize multiple objectives to achieve better hardware efficiency. The experimental results for ResNet-18 on CIFAR-10 show that the proposed ES-MPQ can reduce the parameter size and energy consumption by up to 1.89x and 2.81x, respectively, compared with the fixed bit-width (8 bits) quantization, while losing only 0.59\% accuracy.",

keywords = "Mixed precision quantization, computing-in-memory, evolutionary search",

author = "Sifan Sun and Jinming Ge and Jinyu Bai and Wang Kang",

note = "Publisher Copyright: {\textcopyright} 2023 IEEE.; 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023 ; Conference date: 30-08-2023 Through 01-09-2023",

year = "2023",

doi = "10.1109/NVMSA58981.2023.00018",

language = "英语",

series = "Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023",

publisher = "Institute of Electrical and Electronics Engineers Inc.",

pages = "38--43",

booktitle = "Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023",

address = "美国",

}

Sun, S, Ge, J, Bai, J & Kang, W 2023, ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory. in Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023. Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023, Institute of Electrical and Electronics Engineers Inc., pp. 38-43, 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023, Niigata, Japan, 30/08/23. https://doi.org/10.1109/NVMSA58981.2023.00018

ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory. / Sun, Sifan; Ge, Jinming; Bai, Jinyu et al.
Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023. Institute of Electrical and Electronics Engineers Inc., 2023. p. 38-43 (Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - ES-MPQ

T2 - 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023

AU - Sun, Sifan

AU - Ge, Jinming

AU - Bai, Jinyu

AU - Kang, Wang

PY - 2023

Y1 - 2023

N2 - Network quantization can effectively reduce the complexity without changing the network structures, which is conducive to deploying deep neural networks (DNN) on edge devices. However, most of the existing methods set the quantization precision manually and rarely consider the case that the computing array is limited, such as computing-in-memory (CIM). In this paper, we introduce a novel method named ES-MPQ, which employs evolutionary search to achieve mixed precision quantization with a small calibration dataset. The ES-MPQ can optimize multiple objectives to achieve better hardware efficiency. The experimental results for ResNet-18 on CIFAR-10 show that the proposed ES-MPQ can reduce the parameter size and energy consumption by up to 1.89x and 2.81x, respectively, compared with the fixed bit-width (8 bits) quantization, while losing only 0.59% accuracy.

AB - Network quantization can effectively reduce the complexity without changing the network structures, which is conducive to deploying deep neural networks (DNN) on edge devices. However, most of the existing methods set the quantization precision manually and rarely consider the case that the computing array is limited, such as computing-in-memory (CIM). In this paper, we introduce a novel method named ES-MPQ, which employs evolutionary search to achieve mixed precision quantization with a small calibration dataset. The ES-MPQ can optimize multiple objectives to achieve better hardware efficiency. The experimental results for ResNet-18 on CIFAR-10 show that the proposed ES-MPQ can reduce the parameter size and energy consumption by up to 1.89x and 2.81x, respectively, compared with the fixed bit-width (8 bits) quantization, while losing only 0.59% accuracy.

KW - Mixed precision quantization

KW - computing-in-memory

KW - evolutionary search

UR - https://www.scopus.com/pages/publications/85174185865

U2 - 10.1109/NVMSA58981.2023.00018

DO - 10.1109/NVMSA58981.2023.00018

M3 - 会议稿件

AN - SCOPUS:85174185865

T3 - Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023

SP - 38

EP - 43

BT - Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023

PB - Institute of Electrical and Electronics Engineers Inc.

Y2 - 30 August 2023 through 1 September 2023

ER -

Sun S, Ge J, Bai J, Kang W. ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory. In Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023. Institute of Electrical and Electronics Engineers Inc. 2023. p. 38-43. (Proceedings - 2023 12th IEEE Non-Volatile Memory Systems and Applications Symposium, NVMSA 2023). doi: 10.1109/NVMSA58981.2023.00018

ES-MPQ: Evolutionary Search Enabled Mixed Precision Quantization Framework for Computing-in-Memory

Abstract

Publication series

Conference

UN SDGs

Keywords

Access to Document

Other files and links

Fingerprint

Cite this