Evolutionary structure minimization of deep neural networks for motion sensor data

Publikation: Beitrag in Buch/Bericht/Sammelwerk/KonferenzbandAufsatz in KonferenzbandForschungPeer-Review

Forschungs-netzwerk anzeigen

Details

OriginalspracheEnglisch
Titel des SammelwerksKI 2018
UntertitelAdvances in Artificial Intelligence - 41st German Conference on AI, 2018, Proceedings
Herausgeber/-innenAnni-Yasmin Turhan, Frank Trollmann
Herausgeber (Verlag)Springer Verlag
Seiten243-257
Seitenumfang15
ISBN (Print)9783030001100
PublikationsstatusVeröffentlicht - 2018
Veranstaltung41st German Conference on Artificial Intelligence, KI 2018 - Berlin, Deutschland
Dauer: 24 Sept. 201828 Sept. 2018

Publikationsreihe

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Band11117 LNAI
ISSN (Print)0302-9743
ISSN (elektronisch)1611-3349

Abstract

Many Deep Neural Networks (DNNs) are implemented with the single objective to achieve high classification scores. However, there can be additional objectives like the minimization of computational costs. This is especially important in the field of mobile computing where not only the computational power itself is a limiting factor but also each computation consumes energy affecting the battery life. Unfortunately, the determination of minimal structures is not straightforward. In our paper, we present a new approach to determine DNNs employing reduced structures. The networks are determined by an Evolutionary Algorithm (EA). After the DNN is trained, the EA starts to remove neurons from the network. Thereby, the fitness function of the EA is depending on the accuracy of the DNN. Thus, the EA is able to control the influence of each individual neuron. We introduce our new approach in detail. Thereby, we employ motion data recorded by accelerometer and gyroscope sensors of a mobile device. The data are recorded while drawing Japanese characters in the air in a learning context. The experimental results show that our approach is capable to determine reduced networks with similar performance to the original ones. Additionally, we show that the reduction can improve the accuracy of a network. We analyze the reduction in detail. Further, we present arising structures of the reduced networks.

ASJC Scopus Sachgebiete

Zitieren

Evolutionary structure minimization of deep neural networks for motion sensor data. / Lückehe, Daniel; Veith, Sonja; von Voigt, Gabriele.
KI 2018: Advances in Artificial Intelligence - 41st German Conference on AI, 2018, Proceedings. Hrsg. / Anni-Yasmin Turhan; Frank Trollmann. Springer Verlag, 2018. S. 243-257 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Band 11117 LNAI).

Publikation: Beitrag in Buch/Bericht/Sammelwerk/KonferenzbandAufsatz in KonferenzbandForschungPeer-Review

Lückehe, D, Veith, S & von Voigt, G 2018, Evolutionary structure minimization of deep neural networks for motion sensor data. in A-Y Turhan & F Trollmann (Hrsg.), KI 2018: Advances in Artificial Intelligence - 41st German Conference on AI, 2018, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), Bd. 11117 LNAI, Springer Verlag, S. 243-257, 41st German Conference on Artificial Intelligence, KI 2018, Berlin, Deutschland, 24 Sept. 2018. https://doi.org/10.1007/978-3-030-00111-7_21
Lückehe, D., Veith, S., & von Voigt, G. (2018). Evolutionary structure minimization of deep neural networks for motion sensor data. In A.-Y. Turhan, & F. Trollmann (Hrsg.), KI 2018: Advances in Artificial Intelligence - 41st German Conference on AI, 2018, Proceedings (S. 243-257). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Band 11117 LNAI). Springer Verlag. https://doi.org/10.1007/978-3-030-00111-7_21
Lückehe D, Veith S, von Voigt G. Evolutionary structure minimization of deep neural networks for motion sensor data. in Turhan AY, Trollmann F, Hrsg., KI 2018: Advances in Artificial Intelligence - 41st German Conference on AI, 2018, Proceedings. Springer Verlag. 2018. S. 243-257. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-030-00111-7_21
Lückehe, Daniel ; Veith, Sonja ; von Voigt, Gabriele. / Evolutionary structure minimization of deep neural networks for motion sensor data. KI 2018: Advances in Artificial Intelligence - 41st German Conference on AI, 2018, Proceedings. Hrsg. / Anni-Yasmin Turhan ; Frank Trollmann. Springer Verlag, 2018. S. 243-257 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).
Download
@inproceedings{05d02004a15f4e74b485ccb554015cd9,
title = "Evolutionary structure minimization of deep neural networks for motion sensor data",
abstract = "Many Deep Neural Networks (DNNs) are implemented with the single objective to achieve high classification scores. However, there can be additional objectives like the minimization of computational costs. This is especially important in the field of mobile computing where not only the computational power itself is a limiting factor but also each computation consumes energy affecting the battery life. Unfortunately, the determination of minimal structures is not straightforward. In our paper, we present a new approach to determine DNNs employing reduced structures. The networks are determined by an Evolutionary Algorithm (EA). After the DNN is trained, the EA starts to remove neurons from the network. Thereby, the fitness function of the EA is depending on the accuracy of the DNN. Thus, the EA is able to control the influence of each individual neuron. We introduce our new approach in detail. Thereby, we employ motion data recorded by accelerometer and gyroscope sensors of a mobile device. The data are recorded while drawing Japanese characters in the air in a learning context. The experimental results show that our approach is capable to determine reduced networks with similar performance to the original ones. Additionally, we show that the reduction can improve the accuracy of a network. We analyze the reduction in detail. Further, we present arising structures of the reduced networks.",
keywords = "Deep learning, Evolutionary Algorithm, Japanese characters, Motion sensor data, Neuroevolution, Pruning",
author = "Daniel L{\"u}ckehe and Sonja Veith and {von Voigt}, Gabriele",
note = "Publisher Copyright: {\textcopyright} Springer Nature Switzerland AG 2018.; 41st German Conference on Artificial Intelligence, KI 2018 ; Conference date: 24-09-2018 Through 28-09-2018",
year = "2018",
doi = "10.1007/978-3-030-00111-7_21",
language = "English",
isbn = "9783030001100",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer Verlag",
pages = "243--257",
editor = "Anni-Yasmin Turhan and Frank Trollmann",
booktitle = "KI 2018",
address = "Germany",

}

Download

TY - GEN

T1 - Evolutionary structure minimization of deep neural networks for motion sensor data

AU - Lückehe, Daniel

AU - Veith, Sonja

AU - von Voigt, Gabriele

N1 - Publisher Copyright: © Springer Nature Switzerland AG 2018.

PY - 2018

Y1 - 2018

N2 - Many Deep Neural Networks (DNNs) are implemented with the single objective to achieve high classification scores. However, there can be additional objectives like the minimization of computational costs. This is especially important in the field of mobile computing where not only the computational power itself is a limiting factor but also each computation consumes energy affecting the battery life. Unfortunately, the determination of minimal structures is not straightforward. In our paper, we present a new approach to determine DNNs employing reduced structures. The networks are determined by an Evolutionary Algorithm (EA). After the DNN is trained, the EA starts to remove neurons from the network. Thereby, the fitness function of the EA is depending on the accuracy of the DNN. Thus, the EA is able to control the influence of each individual neuron. We introduce our new approach in detail. Thereby, we employ motion data recorded by accelerometer and gyroscope sensors of a mobile device. The data are recorded while drawing Japanese characters in the air in a learning context. The experimental results show that our approach is capable to determine reduced networks with similar performance to the original ones. Additionally, we show that the reduction can improve the accuracy of a network. We analyze the reduction in detail. Further, we present arising structures of the reduced networks.

AB - Many Deep Neural Networks (DNNs) are implemented with the single objective to achieve high classification scores. However, there can be additional objectives like the minimization of computational costs. This is especially important in the field of mobile computing where not only the computational power itself is a limiting factor but also each computation consumes energy affecting the battery life. Unfortunately, the determination of minimal structures is not straightforward. In our paper, we present a new approach to determine DNNs employing reduced structures. The networks are determined by an Evolutionary Algorithm (EA). After the DNN is trained, the EA starts to remove neurons from the network. Thereby, the fitness function of the EA is depending on the accuracy of the DNN. Thus, the EA is able to control the influence of each individual neuron. We introduce our new approach in detail. Thereby, we employ motion data recorded by accelerometer and gyroscope sensors of a mobile device. The data are recorded while drawing Japanese characters in the air in a learning context. The experimental results show that our approach is capable to determine reduced networks with similar performance to the original ones. Additionally, we show that the reduction can improve the accuracy of a network. We analyze the reduction in detail. Further, we present arising structures of the reduced networks.

KW - Deep learning

KW - Evolutionary Algorithm

KW - Japanese characters

KW - Motion sensor data

KW - Neuroevolution

KW - Pruning

UR - http://www.scopus.com/inward/record.url?scp=85054523618&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-00111-7_21

DO - 10.1007/978-3-030-00111-7_21

M3 - Conference contribution

AN - SCOPUS:85054523618

SN - 9783030001100

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 243

EP - 257

BT - KI 2018

A2 - Turhan, Anni-Yasmin

A2 - Trollmann, Frank

PB - Springer Verlag

T2 - 41st German Conference on Artificial Intelligence, KI 2018

Y2 - 24 September 2018 through 28 September 2018

ER -

Von denselben Autoren