How to exploit sparsity in RNNs on event-driven architectures

Jarno Brils; Luc Waeijen; Arash Pourtaherian

doi:10.1145/3493229.3493302

How to exploit sparsity in RNNs on event-driven architectures

Jarno Brils, Luc Waeijen, Arash Pourtaherian

Mathematics and Computer Science

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

Abstract

Event-driven architectures have been shown to provide low-power, low-latency artificial neural network (ANN) inference. This is especially beneficial on Edge devices, particularly when combined with sparse execution. Recurrent neural networks (RNNs) are ANNs that emulate memory. Their recurrent connection enables the reuse of previous output for the generation of new output. However, when trying to use RNNs in a sparse context on event-driven architectures, novel challenges in synchronization and the usage of sparse data are encountered. In this work, these challenges are systematically analyzed, and mechanisms to overcome them are proposed. Experimental results of a monocular depth estimation use case on the NeuronFlow architecture show that sparsity in RNNs can be exploited effectively on event-driven architectures.

Original language	English
Title of host publication	Proceedings of the 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021
Editors	Sander Stuijk
Publisher	Association for Computing Machinery, Inc
Pages	17-22
Number of pages	6
ISBN (Electronic)	978-1-4503-9166-5
DOIs	https://doi.org/10.1145/3493229.3493302
Publication status	Published - 1 Nov 2021
Event	24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021 - Virtual, Online, Eindhoven, Netherlands Duration: 1 Nov 2021 → 2 Nov 2021 Conference number: 24 https://scopesconf.org/scopes-21/

Conference

Conference	24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021
Abbreviated title	SCOPES 2021
Country/Territory	Netherlands
City	Eindhoven
Period	1/11/21 → 2/11/21
Internet address	https://scopesconf.org/scopes-21/

Funding

The authors acknowledge Funda??o para a Ci?ncia e a Tecnologia (FCT) for financial support (grant numbers SFRH/BD/44108/2008, SFRH/BD/78639/2011 and SFRH/BPD/91397/2012 and projects PTDC/EQU-ERQ/102771/2008 and project UID/QUI/00100/2013).

Keywords

asynchronous
event-driven
monocular depth estimation
neural networks
recurrent neural networks
sparsity

Access to Document

10.1145/3493229.3493302

Cite this

@inproceedings{94c61885b67343f5b5bfb1bc77a57e02,

title = "How to exploit sparsity in RNNs on event-driven architectures",

abstract = "Event-driven architectures have been shown to provide low-power, low-latency artificial neural network (ANN) inference. This is especially beneficial on Edge devices, particularly when combined with sparse execution. Recurrent neural networks (RNNs) are ANNs that emulate memory. Their recurrent connection enables the reuse of previous output for the generation of new output. However, when trying to use RNNs in a sparse context on event-driven architectures, novel challenges in synchronization and the usage of sparse data are encountered. In this work, these challenges are systematically analyzed, and mechanisms to overcome them are proposed. Experimental results of a monocular depth estimation use case on the NeuronFlow architecture show that sparsity in RNNs can be exploited effectively on event-driven architectures. ",

keywords = "asynchronous, event-driven, monocular depth estimation, neural networks, recurrent neural networks, sparsity",

author = "Jarno Brils and Luc Waeijen and Arash Pourtaherian",

note = "Funding Information: The authors acknowledge Funda??o para a Ci?ncia e a Tecnologia (FCT) for financial support (grant numbers SFRH/BD/44108/2008, SFRH/BD/78639/2011 and SFRH/BPD/91397/2012 and projects PTDC/EQU-ERQ/102771/2008 and project UID/QUI/00100/2013). Publisher Copyright: {\textcopyright} 2021 ACM.; 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021, SCOPES 2021 ; Conference date: 01-11-2021 Through 02-11-2021",

year = "2021",

month = nov,

day = "1",

doi = "10.1145/3493229.3493302",

language = "English",

pages = "17--22",

editor = "Sander Stuijk",

booktitle = "Proceedings of the 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021",

publisher = "Association for Computing Machinery, Inc",

address = "United States",

url = "https://scopesconf.org/scopes-21/",

}

Brils, J, Waeijen, L & Pourtaherian, A 2021, How to exploit sparsity in RNNs on event-driven architectures. in S Stuijk (ed.), Proceedings of the 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021. Association for Computing Machinery, Inc, pp. 17-22, 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021, Eindhoven, Netherlands, 1/11/21. https://doi.org/10.1145/3493229.3493302

How to exploit sparsity in RNNs on event-driven architectures. / Brils, Jarno; Waeijen, Luc; Pourtaherian, Arash.
Proceedings of the 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021. ed. / Sander Stuijk. Association for Computing Machinery, Inc, 2021. p. 17-22.

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

TY - GEN

T1 - How to exploit sparsity in RNNs on event-driven architectures

AU - Brils, Jarno

AU - Waeijen, Luc

AU - Pourtaherian, Arash

N1 - Conference code: 24

PY - 2021/11/1

Y1 - 2021/11/1

N2 - Event-driven architectures have been shown to provide low-power, low-latency artificial neural network (ANN) inference. This is especially beneficial on Edge devices, particularly when combined with sparse execution. Recurrent neural networks (RNNs) are ANNs that emulate memory. Their recurrent connection enables the reuse of previous output for the generation of new output. However, when trying to use RNNs in a sparse context on event-driven architectures, novel challenges in synchronization and the usage of sparse data are encountered. In this work, these challenges are systematically analyzed, and mechanisms to overcome them are proposed. Experimental results of a monocular depth estimation use case on the NeuronFlow architecture show that sparsity in RNNs can be exploited effectively on event-driven architectures.

AB - Event-driven architectures have been shown to provide low-power, low-latency artificial neural network (ANN) inference. This is especially beneficial on Edge devices, particularly when combined with sparse execution. Recurrent neural networks (RNNs) are ANNs that emulate memory. Their recurrent connection enables the reuse of previous output for the generation of new output. However, when trying to use RNNs in a sparse context on event-driven architectures, novel challenges in synchronization and the usage of sparse data are encountered. In this work, these challenges are systematically analyzed, and mechanisms to overcome them are proposed. Experimental results of a monocular depth estimation use case on the NeuronFlow architecture show that sparsity in RNNs can be exploited effectively on event-driven architectures.

KW - asynchronous

KW - event-driven

KW - monocular depth estimation

KW - neural networks

KW - recurrent neural networks

KW - sparsity

UR - http://www.scopus.com/inward/record.url?scp=85119688608&partnerID=8YFLogxK

U2 - 10.1145/3493229.3493302

DO - 10.1145/3493229.3493302

M3 - Conference contribution

AN - SCOPUS:85119688608

SP - 17

EP - 22

BT - Proceedings of the 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021

A2 - Stuijk, Sander

PB - Association for Computing Machinery, Inc

T2 - 24th International Workshop on Software and Compilers for Embedded Systems, SCOPES 2021

Y2 - 1 November 2021 through 2 November 2021

ER -

How to exploit sparsity in RNNs on event-driven architectures

Abstract

Conference

Funding

Keywords

Access to Document

Other files and links

Fingerprint

Cite this