Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices

Sebastiaan B.H.C. Hofstee; Duc V. Le

doi:10.1007/978-3-031-18872-5_16

Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices

Sebastiaan B.H.C. Hofstee^*, Duc V. Le

^*Corresponding author for this work

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

22 Downloads (Pure)

Abstract

Pruning of neural networks is a technique often used to reduce the size of a machine learning model, as well as to reduce the computation cost for model inference. This research provides an analysis on four current pruning techniques that theoretically efficiently reduce the machine learning model size, where efficiency is defined by the relation between the compression of the model and the accuracy of the model. Furthermore, this research will assess in what way these four neural network pruning techniques affect the total energy consumption during model inference on a Raspberry Pi 4B board, applied to MobileNetV2, a machine learning model architecture optimized for image classification on embedded devices. Lastly, the research will analyze the trade-offs between energy consumption, model size and model accuracy for each of the assessed pruning algorithms applied to one of the most commonly used neural network architectures, MobileNetV2, on a Raspberry Pi 4B prototyping board. The research is expected to provide engineers a reference providing guidance upon deciding what pruning technique to use for a machine learning model to be deployed on an embedded device.

Original language	English
Title of host publication	Internet of Things. IoT through a Multi-disciplinary Perspective
Subtitle of host publication	5th IFIP International Cross-Domain Conference, IFIPIoT 2022, Amsterdam, The Netherlands, October 27–28, 2022, Proceedings
Editors	Luis M. Camarinha-Matos, Luis Ribeiro, Leon Strous
Place of Publication	Cham
Publisher	Springer
Pages	274-292
Number of pages	19
ISBN (Electronic)	978-3-031-18872-5
ISBN (Print)	978-3-031-18871-8, 978-3-031-18874-9
DOIs	https://doi.org/10.1007/978-3-031-18872-5_16
Publication status	Published - 2022
Event	5th IFIP International Cross-Domain Conference on Internet of Things, IFIPIoT 2022 - Amsterdam, Netherlands Duration: 27 Oct 2022 → 28 Oct 2022 Conference number: 5

Publication series

Name	IFIP Advances in Information and Communication Technology
Publisher	IFIP
Volume	665
ISSN (Print)	1868-4238
ISSN (Electronic)	1868-422X

Conference

Conference	5th IFIP International Cross-Domain Conference on Internet of Things, IFIPIoT 2022
Abbreviated title	IFIPIoT 2022
Country/Territory	Netherlands
City	Amsterdam
Period	27/10/22 → 28/10/22

Keywords

Deep learning
Efficiency
Embedded devices
Energy consumption
Machine Learning (ML)
Neural networks
Pruning
2024 OA procedure

Access to Document

10.1007/978-3-031-18872-5_16

ArticleFinal published version, 1.34 MBLicence: Taverne

Cite this

Hofstee, S. B. H. C., & Le, D. V. (2022). Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices. In L. M. Camarinha-Matos, L. Ribeiro, & L. Strous (Eds.), Internet of Things. IoT through a Multi-disciplinary Perspective: 5th IFIP International Cross-Domain Conference, IFIPIoT 2022, Amsterdam, The Netherlands, October 27–28, 2022, Proceedings (pp. 274-292). (IFIP Advances in Information and Communication Technology; Vol. 665). Springer. https://doi.org/10.1007/978-3-031-18872-5_16

Hofstee, Sebastiaan B.H.C. ; Le, Duc V. / Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices. Internet of Things. IoT through a Multi-disciplinary Perspective: 5th IFIP International Cross-Domain Conference, IFIPIoT 2022, Amsterdam, The Netherlands, October 27–28, 2022, Proceedings. editor / Luis M. Camarinha-Matos ; Luis Ribeiro ; Leon Strous. Cham : Springer, 2022. pp. 274-292 (IFIP Advances in Information and Communication Technology).

@inproceedings{8ff5b03591954d67b919e2e3be1f4fc0,

title = "Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices",

abstract = "Pruning of neural networks is a technique often used to reduce the size of a machine learning model, as well as to reduce the computation cost for model inference. This research provides an analysis on four current pruning techniques that theoretically efficiently reduce the machine learning model size, where efficiency is defined by the relation between the compression of the model and the accuracy of the model. Furthermore, this research will assess in what way these four neural network pruning techniques affect the total energy consumption during model inference on a Raspberry Pi 4B board, applied to MobileNetV2, a machine learning model architecture optimized for image classification on embedded devices. Lastly, the research will analyze the trade-offs between energy consumption, model size and model accuracy for each of the assessed pruning algorithms applied to one of the most commonly used neural network architectures, MobileNetV2, on a Raspberry Pi 4B prototyping board. The research is expected to provide engineers a reference providing guidance upon deciding what pruning technique to use for a machine learning model to be deployed on an embedded device.",

keywords = "Deep learning, Efficiency, Embedded devices, Energy consumption, Machine Learning (ML), Neural networks, Pruning, 2024 OA procedure",

author = "Hofstee, {Sebastiaan B.H.C.} and Le, {Duc V.}",

note = "Publisher Copyright: {\textcopyright} 2022, IFIP International Federation for Information Processing.; 5th IFIP International Cross-Domain Conference on Internet of Things, IFIPIoT 2022, IFIPIoT 2022 ; Conference date: 27-10-2022 Through 28-10-2022",

year = "2022",

doi = "10.1007/978-3-031-18872-5_16",

language = "English",

isbn = "978-3-031-18871-8",

series = "IFIP Advances in Information and Communication Technology",

publisher = "Springer",

pages = "274--292",

editor = "Camarinha-Matos, {Luis M.} and Luis Ribeiro and Leon Strous",

booktitle = "Internet of Things. IoT through a Multi-disciplinary Perspective",

address = "Germany",

}

Hofstee, SBHC & Le, DV 2022, Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices. in LM Camarinha-Matos, L Ribeiro & L Strous (eds), Internet of Things. IoT through a Multi-disciplinary Perspective: 5th IFIP International Cross-Domain Conference, IFIPIoT 2022, Amsterdam, The Netherlands, October 27–28, 2022, Proceedings. IFIP Advances in Information and Communication Technology, vol. 665, Springer, Cham, pp. 274-292, 5th IFIP International Cross-Domain Conference on Internet of Things, IFIPIoT 2022, Amsterdam, Netherlands, 27/10/22. https://doi.org/10.1007/978-3-031-18872-5_16

Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices. / Hofstee, Sebastiaan B.H.C.; Le, Duc V.
Internet of Things. IoT through a Multi-disciplinary Perspective: 5th IFIP International Cross-Domain Conference, IFIPIoT 2022, Amsterdam, The Netherlands, October 27–28, 2022, Proceedings. ed. / Luis M. Camarinha-Matos; Luis Ribeiro; Leon Strous. Cham: Springer, 2022. p. 274-292 (IFIP Advances in Information and Communication Technology; Vol. 665).

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

TY - GEN

T1 - Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices

AU - Hofstee, Sebastiaan B.H.C.

AU - Le, Duc V.

N1 - Conference code: 5

PY - 2022

Y1 - 2022

N2 - Pruning of neural networks is a technique often used to reduce the size of a machine learning model, as well as to reduce the computation cost for model inference. This research provides an analysis on four current pruning techniques that theoretically efficiently reduce the machine learning model size, where efficiency is defined by the relation between the compression of the model and the accuracy of the model. Furthermore, this research will assess in what way these four neural network pruning techniques affect the total energy consumption during model inference on a Raspberry Pi 4B board, applied to MobileNetV2, a machine learning model architecture optimized for image classification on embedded devices. Lastly, the research will analyze the trade-offs between energy consumption, model size and model accuracy for each of the assessed pruning algorithms applied to one of the most commonly used neural network architectures, MobileNetV2, on a Raspberry Pi 4B prototyping board. The research is expected to provide engineers a reference providing guidance upon deciding what pruning technique to use for a machine learning model to be deployed on an embedded device.

AB - Pruning of neural networks is a technique often used to reduce the size of a machine learning model, as well as to reduce the computation cost for model inference. This research provides an analysis on four current pruning techniques that theoretically efficiently reduce the machine learning model size, where efficiency is defined by the relation between the compression of the model and the accuracy of the model. Furthermore, this research will assess in what way these four neural network pruning techniques affect the total energy consumption during model inference on a Raspberry Pi 4B board, applied to MobileNetV2, a machine learning model architecture optimized for image classification on embedded devices. Lastly, the research will analyze the trade-offs between energy consumption, model size and model accuracy for each of the assessed pruning algorithms applied to one of the most commonly used neural network architectures, MobileNetV2, on a Raspberry Pi 4B prototyping board. The research is expected to provide engineers a reference providing guidance upon deciding what pruning technique to use for a machine learning model to be deployed on an embedded device.

KW - Deep learning

KW - Efficiency

KW - Embedded devices

KW - Energy consumption

KW - Machine Learning (ML)

KW - Neural networks

KW - Pruning

KW - 2024 OA procedure

UR - http://www.scopus.com/inward/record.url?scp=85142729595&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-18872-5_16

DO - 10.1007/978-3-031-18872-5_16

M3 - Conference contribution

AN - SCOPUS:85142729595

SN - 978-3-031-18871-8

SN - 978-3-031-18874-9

T3 - IFIP Advances in Information and Communication Technology

SP - 274

EP - 292

BT - Internet of Things. IoT through a Multi-disciplinary Perspective

A2 - Camarinha-Matos, Luis M.

A2 - Ribeiro, Luis

A2 - Strous, Leon

PB - Springer

CY - Cham

T2 - 5th IFIP International Cross-Domain Conference on Internet of Things, IFIPIoT 2022

Y2 - 27 October 2022 through 28 October 2022

ER -

Hofstee SBHC, Le DV. Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices. In Camarinha-Matos LM, Ribeiro L, Strous L, editors, Internet of Things. IoT through a Multi-disciplinary Perspective: 5th IFIP International Cross-Domain Conference, IFIPIoT 2022, Amsterdam, The Netherlands, October 27–28, 2022, Proceedings. Cham: Springer. 2022. p. 274-292. (IFIP Advances in Information and Communication Technology). doi: 10.1007/978-3-031-18872-5_16