Implementation of a batch normalized deep LSTM recurrent network on a smartphone for human activity recognition

Tahmina Zebin; Ertan Balaban; Krikor Ozanyan; Alex Casson; Niels Peek

doi:10.1109/BHI.2019.8834480

Implementation of a batch normalized deep LSTM recurrent network on a smartphone for human activity recognition

Tahmina Zebin, Ertan Balaban, Krikor Ozanyan, Alex Casson, Niels Peek

Division of Informatics, Imaging & Data Sciences (L5)

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

436 Downloads (Pure)

Abstract

In this paper we present a Long-Short Term Memory (LSTM) deep recurrent neural network (RNN) model for the classification of human daily life activities by using the accelerometer and gyroscope data of a smartphone. The proposed model was trained by using the open-source TensorFlow library, optimised and deployed on an Android smartphone. Hardware resource requirements for the implementation are empirically investigated and the effect of data quantization on the accuracy of the implementation is discussed. In addition, we profile the power budget for running the proposed model on smartphone. Results of this work will be of use for deep learning implemented on edge computing devices, which leverages the user privacy as the raw data never leaves the person.

Original language	English
Title of host publication	IEEE-EMBS BHI 2019
DOIs	https://doi.org/10.1109/BHI.2019.8834480
Publication status	Published - 12 Sept 2019
Event	IEEE-EMBS International Conference on Biomedical and Health Informatics - Chicago, United States Duration: 19 May 2019 → 22 May 2019

Conference

Conference	IEEE-EMBS International Conference on Biomedical and Health Informatics
Abbreviated title	IEEE-EMBS BHI 2019
Country/Territory	United States
City	Chicago
Period	19/05/19 → 22/05/19

Access to Document

10.1109/BHI.2019.8834480

1570526264Accepted author manuscript, 1.66 MB

Cite this

@inproceedings{bb185dc1dc5743e885c17f356dc9c682,

title = "Implementation of a batch normalized deep LSTM recurrent network on a smartphone for human activity recognition",

abstract = "In this paper we present a Long-Short Term Memory (LSTM) deep recurrent neural network (RNN) model for the classification of human daily life activities by using the accelerometer and gyroscope data of a smartphone. The proposed model was trained by using the open-source TensorFlow library, optimised and deployed on an Android smartphone. Hardware resource requirements for the implementation are empirically investigated and the effect of data quantization on the accuracy of the implementation is discussed. In addition, we profile the power budget for running the proposed model on smartphone. Results of this work will be of use for deep learning implemented on edge computing devices, which leverages the user privacy as the raw data never leaves the person.",

author = "Tahmina Zebin and Ertan Balaban and Krikor Ozanyan and Alex Casson and Niels Peek",

year = "2019",

month = sep,

day = "12",

doi = "10.1109/BHI.2019.8834480",

language = "English",

booktitle = "IEEE-EMBS BHI 2019",

note = "IEEE-EMBS International Conference on Biomedical and Health Informatics, IEEE-EMBS BHI 2019 ; Conference date: 19-05-2019 Through 22-05-2019",

}

TY - GEN

T1 - Implementation of a batch normalized deep LSTM recurrent network on a smartphone for human activity recognition

AU - Zebin, Tahmina

AU - Balaban, Ertan

AU - Ozanyan, Krikor

AU - Casson, Alex

AU - Peek, Niels

PY - 2019/9/12

Y1 - 2019/9/12

N2 - In this paper we present a Long-Short Term Memory (LSTM) deep recurrent neural network (RNN) model for the classification of human daily life activities by using the accelerometer and gyroscope data of a smartphone. The proposed model was trained by using the open-source TensorFlow library, optimised and deployed on an Android smartphone. Hardware resource requirements for the implementation are empirically investigated and the effect of data quantization on the accuracy of the implementation is discussed. In addition, we profile the power budget for running the proposed model on smartphone. Results of this work will be of use for deep learning implemented on edge computing devices, which leverages the user privacy as the raw data never leaves the person.

AB - In this paper we present a Long-Short Term Memory (LSTM) deep recurrent neural network (RNN) model for the classification of human daily life activities by using the accelerometer and gyroscope data of a smartphone. The proposed model was trained by using the open-source TensorFlow library, optimised and deployed on an Android smartphone. Hardware resource requirements for the implementation are empirically investigated and the effect of data quantization on the accuracy of the implementation is discussed. In addition, we profile the power budget for running the proposed model on smartphone. Results of this work will be of use for deep learning implemented on edge computing devices, which leverages the user privacy as the raw data never leaves the person.

U2 - 10.1109/BHI.2019.8834480

DO - 10.1109/BHI.2019.8834480

M3 - Conference contribution

BT - IEEE-EMBS BHI 2019

T2 - IEEE-EMBS International Conference on Biomedical and Health Informatics

Y2 - 19 May 2019 through 22 May 2019

ER -

Implementation of a batch normalized deep LSTM recurrent network on a smartphone for human activity recognition

Abstract

Conference

Access to Document

Fingerprint

Cite this