A Low-Cost Neural ODE with Depthwise Separable Convolution for Edge Domain Adaptation on FPGAs

High-performance deep neural network (DNN)-based systems are in high demand in edge environments. Due to its high computational complexity, it is challenging to deploy DNNs on edge devices with strict limitations on computational resources. In this paper, we derive a compact while highly-accurate DN...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEICE Transactions on Information and Systems 2023/07/01, Vol.E106.D(7), pp.1186-1197
Hauptverfasser: KAWAKAMI, Hiroki, WATANABE, Hirohisa, SUGIURA, Keisuke, MATSUTANI, Hiroki
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page 1197
container_issue 7
container_start_page 1186
container_title IEICE Transactions on Information and Systems
container_volume E106.D
creator KAWAKAMI, Hiroki
WATANABE, Hirohisa
SUGIURA, Keisuke
MATSUTANI, Hiroki
description High-performance deep neural network (DNN)-based systems are in high demand in edge environments. Due to its high computational complexity, it is challenging to deploy DNNs on edge devices with strict limitations on computational resources. In this paper, we derive a compact while highly-accurate DNN model, termed dsODENet, by combining recently-proposed parameter reduction techniques: Neural ODE (Ordinary Differential Equation) and DSC (Depthwise Separable Convolution). Neural ODE exploits a similarity between ResNet and ODE, and shares most of weight parameters among multiple layers, which greatly reduces the memory consumption. We apply dsODENet to a domain adaptation as a practical use case with image classification datasets. We also propose a resource-efficient FPGA-based design for dsODENet, where all the parameters and feature maps except for pre- and post-processing layers can be mapped onto on-chip memories. It is implemented on Xilinx ZCU104 board and evaluated in terms of domain adaptation accuracy, inference speed, FPGA resource utilization, and speedup rate compared to a software counterpart. The results demonstrate that dsODENet achieves comparable or slightly better domain adaptation accuracy compared to our baseline Neural ODE implementation, while the total parameter size without pre- and post-processing layers is reduced by 54.2% to 79.8%. Our FPGA implementation accelerates the inference speed by 23.8 times.
doi_str_mv 10.1587/transinf.2022EDP7149
format Article
fullrecord <record><control><sourceid>proquest_cross</sourceid><recordid>TN_cdi_proquest_journals_2850468009</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2850468009</sourcerecordid><originalsourceid>FETCH-LOGICAL-c466t-c71e87e54a66835dd46c8b2008d48ab98ca5e5d1d778cd0fc62a9061984cf54b3</originalsourceid><addsrcrecordid>eNpNkF9LwzAUxYMoOKffwIeAz51Jm6TpY1m7KQw3_PNqSNN06-iammQOv73VuTm4cC_c8zsHDgC3GI0w5fG9t7J1dVuNQhSGebaIMUnOwADHhAY4YvgcDFCCWcBpFF6CK-fWCGEeYjoA7ymcmV0wNs7DJ721soHzLIe72q9gpju_2tVOwxfdSSuLRsOxaT9Ns_W1aWFlLMzLpYaZ2ci6hWkpOy9_X_1MFtPUXYOLSjZO3_ztIXib5K_jh2A2nz6O01mgCGM-UDHWPNaUSMZ4RMuSMMWLECFeEi6LhCtJNS1xGcdclahSLJQJYjjhRFWUFNEQ3O19O2s-ttp5sTZb2_aRIuQUEcYRSnoV2auUNc5ZXYnO1htpvwRG4qdJcWhSnDTZY897bO28XOojJK2vVaP_oRwjJjIRH44Tk6NYraQVuo2-Af4AhUc</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>2850468009</pqid></control><display><type>article</type><title>A Low-Cost Neural ODE with Depthwise Separable Convolution for Edge Domain Adaptation on FPGAs</title><source>J-STAGE Free</source><source>Elektronische Zeitschriftenbibliothek - Frei zugängliche E-Journals</source><creator>KAWAKAMI, Hiroki ; WATANABE, Hirohisa ; SUGIURA, Keisuke ; MATSUTANI, Hiroki</creator><creatorcontrib>KAWAKAMI, Hiroki ; WATANABE, Hirohisa ; SUGIURA, Keisuke ; MATSUTANI, Hiroki</creatorcontrib><description>High-performance deep neural network (DNN)-based systems are in high demand in edge environments. Due to its high computational complexity, it is challenging to deploy DNNs on edge devices with strict limitations on computational resources. In this paper, we derive a compact while highly-accurate DNN model, termed dsODENet, by combining recently-proposed parameter reduction techniques: Neural ODE (Ordinary Differential Equation) and DSC (Depthwise Separable Convolution). Neural ODE exploits a similarity between ResNet and ODE, and shares most of weight parameters among multiple layers, which greatly reduces the memory consumption. We apply dsODENet to a domain adaptation as a practical use case with image classification datasets. We also propose a resource-efficient FPGA-based design for dsODENet, where all the parameters and feature maps except for pre- and post-processing layers can be mapped onto on-chip memories. It is implemented on Xilinx ZCU104 board and evaluated in terms of domain adaptation accuracy, inference speed, FPGA resource utilization, and speedup rate compared to a software counterpart. The results demonstrate that dsODENet achieves comparable or slightly better domain adaptation accuracy compared to our baseline Neural ODE implementation, while the total parameter size without pre- and post-processing layers is reduced by 54.2% to 79.8%. Our FPGA implementation accelerates the inference speed by 23.8 times.</description><identifier>ISSN: 0916-8532</identifier><identifier>EISSN: 1745-1361</identifier><identifier>DOI: 10.1587/transinf.2022EDP7149</identifier><language>eng</language><publisher>Tokyo: The Institute of Electronics, Information and Communication Engineers</publisher><subject>Accuracy ; Adaptation ; Artificial neural networks ; Convolution ; Differential equations ; distillation ; domain adaptation ; edge device ; Feature maps ; Field programmable gate arrays ; FPGA ; Image classification ; Inference ; neural ODE ; Ordinary differential equations ; Parameters ; Resource utilization</subject><ispartof>IEICE Transactions on Information and Systems, 2023/07/01, Vol.E106.D(7), pp.1186-1197</ispartof><rights>2023 The Institute of Electronics, Information and Communication Engineers</rights><rights>Copyright Japan Science and Technology Agency 2023</rights><lds50>peer_reviewed</lds50><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed><citedby>FETCH-LOGICAL-c466t-c71e87e54a66835dd46c8b2008d48ab98ca5e5d1d778cd0fc62a9061984cf54b3</citedby><cites>FETCH-LOGICAL-c466t-c71e87e54a66835dd46c8b2008d48ab98ca5e5d1d778cd0fc62a9061984cf54b3</cites></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>314,776,780,1877,27901,27902</link.rule.ids></links><search><creatorcontrib>KAWAKAMI, Hiroki</creatorcontrib><creatorcontrib>WATANABE, Hirohisa</creatorcontrib><creatorcontrib>SUGIURA, Keisuke</creatorcontrib><creatorcontrib>MATSUTANI, Hiroki</creatorcontrib><title>A Low-Cost Neural ODE with Depthwise Separable Convolution for Edge Domain Adaptation on FPGAs</title><title>IEICE Transactions on Information and Systems</title><addtitle>IEICE Trans. Inf. &amp; Syst.</addtitle><description>High-performance deep neural network (DNN)-based systems are in high demand in edge environments. Due to its high computational complexity, it is challenging to deploy DNNs on edge devices with strict limitations on computational resources. In this paper, we derive a compact while highly-accurate DNN model, termed dsODENet, by combining recently-proposed parameter reduction techniques: Neural ODE (Ordinary Differential Equation) and DSC (Depthwise Separable Convolution). Neural ODE exploits a similarity between ResNet and ODE, and shares most of weight parameters among multiple layers, which greatly reduces the memory consumption. We apply dsODENet to a domain adaptation as a practical use case with image classification datasets. We also propose a resource-efficient FPGA-based design for dsODENet, where all the parameters and feature maps except for pre- and post-processing layers can be mapped onto on-chip memories. It is implemented on Xilinx ZCU104 board and evaluated in terms of domain adaptation accuracy, inference speed, FPGA resource utilization, and speedup rate compared to a software counterpart. The results demonstrate that dsODENet achieves comparable or slightly better domain adaptation accuracy compared to our baseline Neural ODE implementation, while the total parameter size without pre- and post-processing layers is reduced by 54.2% to 79.8%. Our FPGA implementation accelerates the inference speed by 23.8 times.</description><subject>Accuracy</subject><subject>Adaptation</subject><subject>Artificial neural networks</subject><subject>Convolution</subject><subject>Differential equations</subject><subject>distillation</subject><subject>domain adaptation</subject><subject>edge device</subject><subject>Feature maps</subject><subject>Field programmable gate arrays</subject><subject>FPGA</subject><subject>Image classification</subject><subject>Inference</subject><subject>neural ODE</subject><subject>Ordinary differential equations</subject><subject>Parameters</subject><subject>Resource utilization</subject><issn>0916-8532</issn><issn>1745-1361</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2023</creationdate><recordtype>article</recordtype><recordid>eNpNkF9LwzAUxYMoOKffwIeAz51Jm6TpY1m7KQw3_PNqSNN06-iammQOv73VuTm4cC_c8zsHDgC3GI0w5fG9t7J1dVuNQhSGebaIMUnOwADHhAY4YvgcDFCCWcBpFF6CK-fWCGEeYjoA7ymcmV0wNs7DJ721soHzLIe72q9gpju_2tVOwxfdSSuLRsOxaT9Ns_W1aWFlLMzLpYaZ2ci6hWkpOy9_X_1MFtPUXYOLSjZO3_ztIXib5K_jh2A2nz6O01mgCGM-UDHWPNaUSMZ4RMuSMMWLECFeEi6LhCtJNS1xGcdclahSLJQJYjjhRFWUFNEQ3O19O2s-ttp5sTZb2_aRIuQUEcYRSnoV2auUNc5ZXYnO1htpvwRG4qdJcWhSnDTZY897bO28XOojJK2vVaP_oRwjJjIRH44Tk6NYraQVuo2-Af4AhUc</recordid><startdate>20230701</startdate><enddate>20230701</enddate><creator>KAWAKAMI, Hiroki</creator><creator>WATANABE, Hirohisa</creator><creator>SUGIURA, Keisuke</creator><creator>MATSUTANI, Hiroki</creator><general>The Institute of Electronics, Information and Communication Engineers</general><general>Japan Science and Technology Agency</general><scope>AAYXX</scope><scope>CITATION</scope><scope>7SC</scope><scope>8FD</scope><scope>JQ2</scope><scope>L7M</scope><scope>L~C</scope><scope>L~D</scope></search><sort><creationdate>20230701</creationdate><title>A Low-Cost Neural ODE with Depthwise Separable Convolution for Edge Domain Adaptation on FPGAs</title><author>KAWAKAMI, Hiroki ; WATANABE, Hirohisa ; SUGIURA, Keisuke ; MATSUTANI, Hiroki</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-c466t-c71e87e54a66835dd46c8b2008d48ab98ca5e5d1d778cd0fc62a9061984cf54b3</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2023</creationdate><topic>Accuracy</topic><topic>Adaptation</topic><topic>Artificial neural networks</topic><topic>Convolution</topic><topic>Differential equations</topic><topic>distillation</topic><topic>domain adaptation</topic><topic>edge device</topic><topic>Feature maps</topic><topic>Field programmable gate arrays</topic><topic>FPGA</topic><topic>Image classification</topic><topic>Inference</topic><topic>neural ODE</topic><topic>Ordinary differential equations</topic><topic>Parameters</topic><topic>Resource utilization</topic><toplevel>peer_reviewed</toplevel><toplevel>online_resources</toplevel><creatorcontrib>KAWAKAMI, Hiroki</creatorcontrib><creatorcontrib>WATANABE, Hirohisa</creatorcontrib><creatorcontrib>SUGIURA, Keisuke</creatorcontrib><creatorcontrib>MATSUTANI, Hiroki</creatorcontrib><collection>CrossRef</collection><collection>Computer and Information Systems Abstracts</collection><collection>Technology Research Database</collection><collection>ProQuest Computer Science Collection</collection><collection>Advanced Technologies Database with Aerospace</collection><collection>Computer and Information Systems Abstracts – Academic</collection><collection>Computer and Information Systems Abstracts Professional</collection><jtitle>IEICE Transactions on Information and Systems</jtitle></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>KAWAKAMI, Hiroki</au><au>WATANABE, Hirohisa</au><au>SUGIURA, Keisuke</au><au>MATSUTANI, Hiroki</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>A Low-Cost Neural ODE with Depthwise Separable Convolution for Edge Domain Adaptation on FPGAs</atitle><jtitle>IEICE Transactions on Information and Systems</jtitle><addtitle>IEICE Trans. Inf. &amp; Syst.</addtitle><date>2023-07-01</date><risdate>2023</risdate><volume>E106.D</volume><issue>7</issue><spage>1186</spage><epage>1197</epage><pages>1186-1197</pages><artnum>2022EDP7149</artnum><issn>0916-8532</issn><eissn>1745-1361</eissn><abstract>High-performance deep neural network (DNN)-based systems are in high demand in edge environments. Due to its high computational complexity, it is challenging to deploy DNNs on edge devices with strict limitations on computational resources. In this paper, we derive a compact while highly-accurate DNN model, termed dsODENet, by combining recently-proposed parameter reduction techniques: Neural ODE (Ordinary Differential Equation) and DSC (Depthwise Separable Convolution). Neural ODE exploits a similarity between ResNet and ODE, and shares most of weight parameters among multiple layers, which greatly reduces the memory consumption. We apply dsODENet to a domain adaptation as a practical use case with image classification datasets. We also propose a resource-efficient FPGA-based design for dsODENet, where all the parameters and feature maps except for pre- and post-processing layers can be mapped onto on-chip memories. It is implemented on Xilinx ZCU104 board and evaluated in terms of domain adaptation accuracy, inference speed, FPGA resource utilization, and speedup rate compared to a software counterpart. The results demonstrate that dsODENet achieves comparable or slightly better domain adaptation accuracy compared to our baseline Neural ODE implementation, while the total parameter size without pre- and post-processing layers is reduced by 54.2% to 79.8%. Our FPGA implementation accelerates the inference speed by 23.8 times.</abstract><cop>Tokyo</cop><pub>The Institute of Electronics, Information and Communication Engineers</pub><doi>10.1587/transinf.2022EDP7149</doi><tpages>12</tpages><oa>free_for_read</oa></addata></record>
fulltext fulltext
identifier ISSN: 0916-8532
ispartof IEICE Transactions on Information and Systems, 2023/07/01, Vol.E106.D(7), pp.1186-1197
issn 0916-8532
1745-1361
language eng
recordid cdi_proquest_journals_2850468009
source J-STAGE Free; Elektronische Zeitschriftenbibliothek - Frei zugängliche E-Journals
subjects Accuracy
Adaptation
Artificial neural networks
Convolution
Differential equations
distillation
domain adaptation
edge device
Feature maps
Field programmable gate arrays
FPGA
Image classification
Inference
neural ODE
Ordinary differential equations
Parameters
Resource utilization
title A Low-Cost Neural ODE with Depthwise Separable Convolution for Edge Domain Adaptation on FPGAs
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-02-08T23%3A38%3A36IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-proquest_cross&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=A%20Low-Cost%20Neural%20ODE%20with%20Depthwise%20Separable%20Convolution%20for%20Edge%20Domain%20Adaptation%20on%20FPGAs&rft.jtitle=IEICE%20Transactions%20on%20Information%20and%20Systems&rft.au=KAWAKAMI,%20Hiroki&rft.date=2023-07-01&rft.volume=E106.D&rft.issue=7&rft.spage=1186&rft.epage=1197&rft.pages=1186-1197&rft.artnum=2022EDP7149&rft.issn=0916-8532&rft.eissn=1745-1361&rft_id=info:doi/10.1587/transinf.2022EDP7149&rft_dat=%3Cproquest_cross%3E2850468009%3C/proquest_cross%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=2850468009&rft_id=info:pmid/&rfr_iscdi=true