BIT INTERPRETATION FOR CONVOLUTIONAL NEURAL NETWORK INPUT LAYER

In order to efficiently execute deep convolutional neural networks (CNNs) on edge devices (e.g., wearable device such as Apple Watch or FitBit), it is necessary to reduce a bit width of network parameters down to 1-bit. Typically, binarization at a first layer of the CNN is not performed because the...

Full description

Saved in:
Bibliographic Details
Main Authors DUERICHEN ROBERT, PETERS CHRISTIAN, ROCZNIK THOMAS
Format Patent
LanguageEnglish
Korean
Published 09.06.2020
Subjects
Online AccessGet full text

Cover

Loading…
Abstract In order to efficiently execute deep convolutional neural networks (CNNs) on edge devices (e.g., wearable device such as Apple Watch or FitBit), it is necessary to reduce a bit width of network parameters down to 1-bit. Typically, binarization at a first layer of the CNN is not performed because the binarization leads to an increase in an output validation error of input data. Method and systems provided herein include a binary input layer (BIL) which accepts binary input data by learning bit-specific binary weights. By executing the CNN using the binary input data, the method and the system according to the present invention result in a reduction in a consumed chip area and used energy used in contrast to CNN models executed by using floating point input data. A method for implementing a convolutional neural network includes the steps of: receiving input data for the convolutional neural network; filtering the input data by applying a bitwise weight algorithm that learns bit-specific relevance without a predefined ordinal structure provided to generate direct binary input data; and providing the direct binary input data to a convolutional layer within the convolutional neural network. 에지 디바이스들(예컨대, 애플 워치 또는 FitBit와 같은 착용 가능 디바이스) 상에서 심층 컨볼루션 신경망들(CNN)을 효율적으로 실행하기 위해, 1-비트 아래로 네트워크 파라미터들의 비트폭을 감소시키는 것이 필요할 수 있다. CNN의 제 1 층에서의 이진화는 통상적으로 그것이 입력 데이터의 출력 검증 에러에서 증가를 야기할 수 있기 때문에 수행되지 않는다. 제공된 방법 및 시스템들은 비트 특정 이진 가중치들을 학습함으로써 이진 입력 데이터를 수용하는 이진 입력 층(BIL)을 포함한다. 이진 입력 데이터를 사용하여 CNN을 실행함으로써, 본 방법 및 시스템은 부동 소수점 입력 데이터를 사용하여 실행된 CNN 모델들과 대조적으로 소비된 칩 면적 및 사용된 에너지에서의 감소를 야기할 수 있다.
AbstractList In order to efficiently execute deep convolutional neural networks (CNNs) on edge devices (e.g., wearable device such as Apple Watch or FitBit), it is necessary to reduce a bit width of network parameters down to 1-bit. Typically, binarization at a first layer of the CNN is not performed because the binarization leads to an increase in an output validation error of input data. Method and systems provided herein include a binary input layer (BIL) which accepts binary input data by learning bit-specific binary weights. By executing the CNN using the binary input data, the method and the system according to the present invention result in a reduction in a consumed chip area and used energy used in contrast to CNN models executed by using floating point input data. A method for implementing a convolutional neural network includes the steps of: receiving input data for the convolutional neural network; filtering the input data by applying a bitwise weight algorithm that learns bit-specific relevance without a predefined ordinal structure provided to generate direct binary input data; and providing the direct binary input data to a convolutional layer within the convolutional neural network. 에지 디바이스들(예컨대, 애플 워치 또는 FitBit와 같은 착용 가능 디바이스) 상에서 심층 컨볼루션 신경망들(CNN)을 효율적으로 실행하기 위해, 1-비트 아래로 네트워크 파라미터들의 비트폭을 감소시키는 것이 필요할 수 있다. CNN의 제 1 층에서의 이진화는 통상적으로 그것이 입력 데이터의 출력 검증 에러에서 증가를 야기할 수 있기 때문에 수행되지 않는다. 제공된 방법 및 시스템들은 비트 특정 이진 가중치들을 학습함으로써 이진 입력 데이터를 수용하는 이진 입력 층(BIL)을 포함한다. 이진 입력 데이터를 사용하여 CNN을 실행함으로써, 본 방법 및 시스템은 부동 소수점 입력 데이터를 사용하여 실행된 CNN 모델들과 대조적으로 소비된 칩 면적 및 사용된 에너지에서의 감소를 야기할 수 있다.
Author PETERS CHRISTIAN
DUERICHEN ROBERT
ROCZNIK THOMAS
Author_xml – fullname: DUERICHEN ROBERT
– fullname: PETERS CHRISTIAN
– fullname: ROCZNIK THOMAS
BookMark eNrjYmDJy89L5WSwd_IMUfD0C3ENCghyDXEM8fT3U3DzD1Jw9vcL8_cJBfEdfRT8XEODwFRIuH-QN1B9QGiIgo9jpGsQDwNrWmJOcSovlOZmUHZzDXH20E0tyI9PLS5ITE7NSy2J9w4yMjAyMDAwMzMyNHU0Jk4VAIUaLXk
ContentType Patent
DBID EVB
DatabaseName esp@cenet
DatabaseTitleList
Database_xml – sequence: 1
  dbid: EVB
  name: esp@cenet
  url: http://worldwide.espacenet.com/singleLineSearch?locale=en_EP
  sourceTypes: Open Access Repository
DeliveryMethod fulltext_linktorsrc
Discipline Medicine
Chemistry
Sciences
Physics
DocumentTitleAlternate 컨볼루션 신경망 입력 층을 위한 비트 해석
ExternalDocumentID KR20200066215A
GroupedDBID EVB
ID FETCH-epo_espacenet_KR20200066215A3
IEDL.DBID EVB
IngestDate Fri Jul 19 14:49:43 EDT 2024
IsOpenAccess true
IsPeerReviewed false
IsScholarly false
Language English
Korean
LinkModel DirectLink
MergedId FETCHMERGED-epo_espacenet_KR20200066215A3
Notes Application Number: KR20190155396
OpenAccessLink https://worldwide.espacenet.com/publicationDetails/biblio?FT=D&date=20200609&DB=EPODOC&CC=KR&NR=20200066215A
ParticipantIDs epo_espacenet_KR20200066215A
PublicationCentury 2000
PublicationDate 20200609
PublicationDateYYYYMMDD 2020-06-09
PublicationDate_xml – month: 06
  year: 2020
  text: 20200609
  day: 09
PublicationDecade 2020
PublicationYear 2020
RelatedCompanies ROBERT BOSCH GMBH
RelatedCompanies_xml – name: ROBERT BOSCH GMBH
Score 3.227932
Snippet In order to efficiently execute deep convolutional neural networks (CNNs) on edge devices (e.g., wearable device such as Apple Watch or FitBit), it is...
SourceID epo
SourceType Open Access Repository
SubjectTerms CALCULATING
COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
COMPUTING
COUNTING
PHYSICS
Title BIT INTERPRETATION FOR CONVOLUTIONAL NEURAL NETWORK INPUT LAYER
URI https://worldwide.espacenet.com/publicationDetails/biblio?FT=D&date=20200609&DB=EPODOC&locale=&CC=KR&NR=20200066215A
hasFullText 1
inHoldings 1
isFullTextHit
isPrint
link http://utb.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwfV3dT8IwEL8gfr4pSvxAs0Szt0U6BmYPC4F9BB1spNkQn8i6dYnRAHEz_vteCyhPPDXtNZf2kuvdtXe_AjxkvNkyuc60BJ1bzdDTjsZInms6yQkjKXogbVE7PAo6g9h4mbanFfjc1MJInNAfCY6IGpWivpfyvF7-X2I5MreyeGTvOLToepHlqOvoWMTHTVN1-pY7Dp3QVm3b8qka0BVNoJ2Tdm8P9tGRfhL64E76oi5luW1UvFM4GCO_eXkGlY9FDY7tzd9rNTgarZ-8a3AoczTTAgfXelicQ7f_HCkSzRYlGMlrJgXDOcUOg0k4jFcAt0rgxlQ20WtIfZw_jiNl2Htz6QXce25kDzRc0-xPBDOfbm-gVYfqfDHnl6C0CMk4STqJSZjB0B1jqQC4yQyeJTmSrqCxi9P1bvINnIiuTI0yG1Atv775LRrhkt1J2f0CI8WE3Q
link.rule.ids 230,309,783,888,25576,76876
linkProvider European Patent Office
linkToHtml http://utb.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwfV3dT8IwEL8gfuCbosYP1CWavS1SGJg9EALbyHBfZOkQn8i6j8RoBpEZ_32vZShPPDW5ay7tJde7a-9-BXhM0mZbS1tMiTC4VdRW3FUYyTKlRTLCSIwRSIf3Drte1wrVl1lnVoHPTS-MwAn9EeCIaFEx2nshzuvl_yWWIWorV0_sHUmL_oj2DLnMjnl-3NRkY9gzJ77h67Ku9-xA9oI1j6Odk85gD_YxyH7m9mBOh7wvZbntVEYncDBBeXlxCpWPRR1q-ubvtTocueWTdx0ORY1mvEJiaYerM-gPx1QSaLaoQSqumSRM5yTd96a-E64BbiXPDAMx0Fc_sHH-JKSSM3gzg3N4GJlUtxRc0_xPBXM72N5A-wKq-SJPL0FqE5KkJOpGGmEqw3CMxRzgJlHTJMqQdQWNXZKud7PvoWZR15k7Y8--gWPOEmVSWgOqxdd3eosOuWB3Qo-_iwGH0A
openUrl ctx_ver=Z39.88-2004&ctx_enc=info%3Aofi%2Fenc%3AUTF-8&rfr_id=info%3Asid%2Fsummon.serialssolutions.com&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Apatent&rft.title=BIT+INTERPRETATION+FOR+CONVOLUTIONAL+NEURAL+NETWORK+INPUT+LAYER&rft.inventor=DUERICHEN+ROBERT&rft.inventor=PETERS+CHRISTIAN&rft.inventor=ROCZNIK+THOMAS&rft.date=2020-06-09&rft.externalDBID=A&rft.externalDocID=KR20200066215A