Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression

Wan Ting Chang; Chih Hung Kuo; Li Chun Fang

doi:10.1109/VLSI-DAT54769.2022.9768055

Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression

Wan Ting Chang, Chih Hung Kuo, Li Chun Fang

Department of Electrical Engineering

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

2 Citations (Scopus)

Abstract

This paper presents a model compression frame-work for both pruning and quantizing according to the channel distribution information. We apply the variational inference technique to train a Bayesian deep neural network, in which the parameters are modeled by probability distributions. According to the characteristic of the probability distribution, we can prune the redundant channels and determine the bit-width layer by layer. The experiments conducted on the CIFAR10 dataset with the VGG16 show that the number of parameters can be saved by 58.91x. The proposed compression approach can help implement hardware circuits for efficient edge and mobile computing.

Original language	English
Title of host publication	2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings
Publisher	Institute of Electrical and Electronics Engineers Inc.
ISBN (Electronic)	9781665409216
DOIs	https://doi.org/10.1109/VLSI-DAT54769.2022.9768055
Publication status	Published - 2022
Event	2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Hsinchu, Taiwan Duration: 2022 Apr 18 → 2022 Apr 21

Publication series

Name	2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings

Conference

Conference	2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022
Country/Territory	Taiwan
City	Hsinchu
Period	22-04-18 → 22-04-21

All Science Journal Classification (ASJC) codes

Hardware and Architecture
Electrical and Electronic Engineering
Safety, Risk, Reliability and Quality

Access to Document

10.1109/VLSI-DAT54769.2022.9768055

Cite this

Chang, W. T., Kuo, C. H., & Fang, L. C. (2022). Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression. In 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings (2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings). Institute of Electrical and Electronics Engineers Inc.. https://doi.org/10.1109/VLSI-DAT54769.2022.9768055

Chang, Wan Ting ; Kuo, Chih Hung ; Fang, Li Chun. / Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression. 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings. Institute of Electrical and Electronics Engineers Inc., 2022. (2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings).

@inproceedings{cacba869b0c34f189d55871624d3d777,

title = "Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression",

abstract = "This paper presents a model compression frame-work for both pruning and quantizing according to the channel distribution information. We apply the variational inference technique to train a Bayesian deep neural network, in which the parameters are modeled by probability distributions. According to the characteristic of the probability distribution, we can prune the redundant channels and determine the bit-width layer by layer. The experiments conducted on the CIFAR10 dataset with the VGG16 show that the number of parameters can be saved by 58.91x. The proposed compression approach can help implement hardware circuits for efficient edge and mobile computing.",

author = "Chang, {Wan Ting} and Kuo, {Chih Hung} and Fang, {Li Chun}",

note = "Publisher Copyright: {\textcopyright} 2022 IEEE.; 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 ; Conference date: 18-04-2022 Through 21-04-2022",

year = "2022",

doi = "10.1109/VLSI-DAT54769.2022.9768055",

language = "English",

series = "2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings",

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

booktitle = "2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings",

address = "United States",

}

Chang, WT, Kuo, CH & Fang, LC 2022, Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression. in 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings. 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings, Institute of Electrical and Electronics Engineers Inc., 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022, Hsinchu, Taiwan, 22-04-18. https://doi.org/10.1109/VLSI-DAT54769.2022.9768055

Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression. / Chang, Wan Ting; Kuo, Chih Hung; Fang, Li Chun.
2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings. Institute of Electrical and Electronics Engineers Inc., 2022. (2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings).

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

TY - GEN

T1 - Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression

AU - Chang, Wan Ting

AU - Kuo, Chih Hung

AU - Fang, Li Chun

PY - 2022

Y1 - 2022

N2 - This paper presents a model compression frame-work for both pruning and quantizing according to the channel distribution information. We apply the variational inference technique to train a Bayesian deep neural network, in which the parameters are modeled by probability distributions. According to the characteristic of the probability distribution, we can prune the redundant channels and determine the bit-width layer by layer. The experiments conducted on the CIFAR10 dataset with the VGG16 show that the number of parameters can be saved by 58.91x. The proposed compression approach can help implement hardware circuits for efficient edge and mobile computing.

AB - This paper presents a model compression frame-work for both pruning and quantizing according to the channel distribution information. We apply the variational inference technique to train a Bayesian deep neural network, in which the parameters are modeled by probability distributions. According to the characteristic of the probability distribution, we can prune the redundant channels and determine the bit-width layer by layer. The experiments conducted on the CIFAR10 dataset with the VGG16 show that the number of parameters can be saved by 58.91x. The proposed compression approach can help implement hardware circuits for efficient edge and mobile computing.

UR - http://www.scopus.com/inward/record.url?scp=85130468864&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85130468864&partnerID=8YFLogxK

U2 - 10.1109/VLSI-DAT54769.2022.9768055

DO - 10.1109/VLSI-DAT54769.2022.9768055

M3 - Conference contribution

AN - SCOPUS:85130468864

T3 - 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings

BT - 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings

PB - Institute of Electrical and Electronics Engineers Inc.

T2 - 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022

Y2 - 18 April 2022 through 21 April 2022

ER -

Chang WT, Kuo CH, Fang LC. Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression. In 2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings. Institute of Electrical and Electronics Engineers Inc. 2022. (2022 International Symposium on VLSI Design, Automation and Test, VLSI-DAT 2022 - Proceedings). doi: 10.1109/VLSI-DAT54769.2022.9768055

Variational Channel Distribution Pruning and Mixed-Precision Quantization for Neural Network Model Compression

Abstract

Publication series

Conference

All Science Journal Classification (ASJC) codes

Access to Document

Other files and links

Fingerprint

Cite this