A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis

Publikation: Bidrag til bog/antologi/rapport/konference-proceedingKonferencebidrag i proceedingsForskningpeer review

Resumé

Reservoir computing (RC) features with the rich computational dynamics is a kind of powerful machine learning paradigm that is well suited for non-linear time-series prediction and classification problems. However, this impressive performance comes with a cost of complex arithmetic operations and high memory usage that make it significantly challenging to deploy on embedded systems. Solutions based on CPU and/or GPU-based designs, provides flexibility but suffers from a lack of efficiency in terms of power, performance, and area (PPA). Although hardware-accelerated solutions can improve efficiency, it takes longer design cycles and is time-consuming. Furthermore, it may happen that design spec requires run change due to the fact that the network is retrained with the new data set to improve the performance. It leads to extra effort in the redesign of the hardware-accelerated solution. This preliminary work presents the design and implementation of a hardware generator for RC-ESNs (echo state networks) to tackle the problem. The proposed methodology is demonstrated by various offline-trained network parameters and topologies. Compared to existing solutions, the proposed framework provides scalability with the support of DSE in agile hardware design.

OriginalsprogEngelsk
Titel2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings
Antal sider6
ForlagIEEE
Publikationsdato15. jul. 2019
Artikelnummer8760065
ISBN (Trykt)9781728117393
ISBN (Elektronisk)9781728117409
DOI
StatusUdgivet - 15. jul. 2019
Begivenhed8th Mediterranean Conference on Embedded Computing, MECO 2019 - Budva, Montenegro
Varighed: 10. jun. 201914. jun. 2019

Konference

Konference8th Mediterranean Conference on Embedded Computing, MECO 2019
LandMontenegro
ByBudva
Periode10/06/201914/06/2019
NavnMediterranean Conference on Embedded Computing, MECO
Vol/bind2019
ISSN2377-5475

Fingeraftryk

Embedded systems
Computer hardware
Hardware
Program processors
Learning systems
Scalability
Time series
Topology
High level synthesis
Data storage equipment
Costs

Citer dette

Huang, N. S., Braun, J. M., Larsen, J. C., & Manoonpong, P. (2019). A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis. I 2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings [8760065] IEEE. Mediterranean Conference on Embedded Computing, MECO , Bind. 2019 https://doi.org/10.1109/MECO.2019.8760065
Huang, Nan Sheng ; Braun, Jan Matthias ; Larsen, Jorgen Christian ; Manoonpong, Poramate. / A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis. 2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings. IEEE, 2019. (Mediterranean Conference on Embedded Computing, MECO , Bind 2019).
@inproceedings{1ead4ac40d7d4d78bd4c4283a005dc82,
title = "A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis",
abstract = "Reservoir computing (RC) features with the rich computational dynamics is a kind of powerful machine learning paradigm that is well suited for non-linear time-series prediction and classification problems. However, this impressive performance comes with a cost of complex arithmetic operations and high memory usage that make it significantly challenging to deploy on embedded systems. Solutions based on CPU and/or GPU-based designs, provides flexibility but suffers from a lack of efficiency in terms of power, performance, and area (PPA). Although hardware-accelerated solutions can improve efficiency, it takes longer design cycles and is time-consuming. Furthermore, it may happen that design spec requires run change due to the fact that the network is retrained with the new data set to improve the performance. It leads to extra effort in the redesign of the hardware-accelerated solution. This preliminary work presents the design and implementation of a hardware generator for RC-ESNs (echo state networks) to tackle the problem. The proposed methodology is demonstrated by various offline-trained network parameters and topologies. Compared to existing solutions, the proposed framework provides scalability with the support of DSE in agile hardware design.",
keywords = "Echo State Networks, Embedded Systems, Hardware Accelerator, High-Level Synthesis, Neural Networks, Reservoir Computing",
author = "Huang, {Nan Sheng} and Braun, {Jan Matthias} and Larsen, {Jorgen Christian} and Poramate Manoonpong",
year = "2019",
month = "7",
day = "15",
doi = "10.1109/MECO.2019.8760065",
language = "English",
isbn = "9781728117393",
booktitle = "2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings",
publisher = "IEEE",
address = "United States",

}

Huang, NS, Braun, JM, Larsen, JC & Manoonpong, P 2019, A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis. i 2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings., 8760065, IEEE, Mediterranean Conference on Embedded Computing, MECO , bind 2019, 8th Mediterranean Conference on Embedded Computing, MECO 2019, Budva, Montenegro, 10/06/2019. https://doi.org/10.1109/MECO.2019.8760065

A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis. / Huang, Nan Sheng; Braun, Jan Matthias; Larsen, Jorgen Christian; Manoonpong, Poramate.

2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings. IEEE, 2019. 8760065 (Mediterranean Conference on Embedded Computing, MECO , Bind 2019).

Publikation: Bidrag til bog/antologi/rapport/konference-proceedingKonferencebidrag i proceedingsForskningpeer review

TY - GEN

T1 - A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis

AU - Huang, Nan Sheng

AU - Braun, Jan Matthias

AU - Larsen, Jorgen Christian

AU - Manoonpong, Poramate

PY - 2019/7/15

Y1 - 2019/7/15

N2 - Reservoir computing (RC) features with the rich computational dynamics is a kind of powerful machine learning paradigm that is well suited for non-linear time-series prediction and classification problems. However, this impressive performance comes with a cost of complex arithmetic operations and high memory usage that make it significantly challenging to deploy on embedded systems. Solutions based on CPU and/or GPU-based designs, provides flexibility but suffers from a lack of efficiency in terms of power, performance, and area (PPA). Although hardware-accelerated solutions can improve efficiency, it takes longer design cycles and is time-consuming. Furthermore, it may happen that design spec requires run change due to the fact that the network is retrained with the new data set to improve the performance. It leads to extra effort in the redesign of the hardware-accelerated solution. This preliminary work presents the design and implementation of a hardware generator for RC-ESNs (echo state networks) to tackle the problem. The proposed methodology is demonstrated by various offline-trained network parameters and topologies. Compared to existing solutions, the proposed framework provides scalability with the support of DSE in agile hardware design.

AB - Reservoir computing (RC) features with the rich computational dynamics is a kind of powerful machine learning paradigm that is well suited for non-linear time-series prediction and classification problems. However, this impressive performance comes with a cost of complex arithmetic operations and high memory usage that make it significantly challenging to deploy on embedded systems. Solutions based on CPU and/or GPU-based designs, provides flexibility but suffers from a lack of efficiency in terms of power, performance, and area (PPA). Although hardware-accelerated solutions can improve efficiency, it takes longer design cycles and is time-consuming. Furthermore, it may happen that design spec requires run change due to the fact that the network is retrained with the new data set to improve the performance. It leads to extra effort in the redesign of the hardware-accelerated solution. This preliminary work presents the design and implementation of a hardware generator for RC-ESNs (echo state networks) to tackle the problem. The proposed methodology is demonstrated by various offline-trained network parameters and topologies. Compared to existing solutions, the proposed framework provides scalability with the support of DSE in agile hardware design.

KW - Echo State Networks

KW - Embedded Systems

KW - Hardware Accelerator

KW - High-Level Synthesis

KW - Neural Networks

KW - Reservoir Computing

U2 - 10.1109/MECO.2019.8760065

DO - 10.1109/MECO.2019.8760065

M3 - Article in proceedings

SN - 9781728117393

BT - 2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings

PB - IEEE

ER -

Huang NS, Braun JM, Larsen JC, Manoonpong P. A scalable Echo State Networks hardware generator for embedded systems using high-level synthesis. I 2019 8th Mediterranean Conference on Embedded Computing, MECO 2019 - Proceedings. IEEE. 2019. 8760065. (Mediterranean Conference on Embedded Computing, MECO , Bind 2019). https://doi.org/10.1109/MECO.2019.8760065