A study on a low power optimization algorithm for an edge-AI device

Kaneko, Tatsuya; Orimo, Kentaro; Hida, Itaru; Takamaeda, Shinya; Ikebe, Masayuki; Motomura, Masato; Asai, Tetsuya

A study on a low power optimization algorithm for an edge-AI device

Kaneko, Tatsuya; Orimo, Kentaro; Hida, Itaru et al.

Nonlinear theory and its applications, IEICE, 2019, 10(4), 373-389

Permalink : https://hdl.handle.net/2115/76016

このアイテムのアクセス数:699件（2026-08-11 08:53 集計）

閲覧可能ファイル

ファイル	フォーマット	サイズ	閲覧回数	説明
https://doi.org/10.1587/nolta.10.373			21

論文情報

ファイル出力 EndNote Basic出力 Mendeley出力

アクセス権	open access
URI	https://hdl.handle.net/2115/76016
タイトル	A study on a low power optimization algorithm for an edge-AI device
著者	Kaneko, Tatsuya ; Orimo, Kentaro ; Hida, Itaru ; Takamaeda, Shinya ; Ikebe, Masayuki ; Motomura, Masato ; Asai, Tetsuya
著者名 Kaneko, Tatsuya
著者名 Orimo, Kentaro
著者名 Hida, Itaru
e-Rad_Researcher 60738897 著者名 Takamaeda, Shinya 著者別名 Yamazaki, Shinya
e-Rad_Researcher 20374613 著者名 Ikebe, Masayuki
著者名 Motomura, Masato
e-Rad_Researcher 00312380 著者名 Asai, Tetsuya
言語	英語
キーワード	machine learning
	edge AI
	training algorithm
	backpropagation
	quantization
	low power
発行日	2019-10
出版者	電子情報通信学会(The Institute of Electronics, Information and Communication Engineers / IEICE)
収録物名	Nonlinear theory and its applications, IEICE
巻(号)	10(4)
ページ	373-389
抄録	Although research on the inference phase of edge artificial intelligence (AI) has made considerable improvement, the required training phase remains an unsolved problem. Neural network (NN) processing has two phases: inference and training. In the training phase, a NN incurs high calculation cost. The number of bits (bitwidth) in the training phase is several orders of magnitude larger than that in the inference phase. Training algorithms, optimized to software, are not appropriate for training hardware-oriented NNs. Therefore, we propose a new training algorithm for edge AI: backpropagation (BP) using a ternarized gradient. This ternarized backpropagation (TBP) provides a balance between calculation cost and performance. Empirical results demonstrate that in a two-class classification task, TBP works well in practice and compares favorably with 16-bit BP (Fixed-BP).
資源タイプ	学術雑誌論文
出版タイプ	Version of Record
権利情報	https://search.ieice.org/
権利情報	Copyright ©2019 The Institute of Electronics, Information and Communication Engineers
PISSN	2185-4106
関連情報 (isIdenticalTo)	出版社版DOI https://doi.org/10.1587/nolta.10.373