A Challenge for Deep Neural Network Design - Automatic Architecture Optimization

カテゴリ: 研究会(論文単位)

論文No: CT22088

グループ名: 【C】電子・情報・システム部門 制御研究会

発行日: 2022/11/02

タイトル(英語): A Challenge for Deep Neural Network Design - Automatic Architecture Optimization

著者名: 李 恒毅(立命館大学),孟 林(立命館大学)

著者名(英語): Hengyi LI(Ritsumeikan University ),Lin MENG(Ritsumeikan University )

キーワード: DNNs|automatic architecture optimization|lightweight|high efficiency

要約(日本語): As deep neural networks (DNNs) achieve great success in various domains, the heavy intensity of computation and memory is also a serious burden that blocks the development and applications of the technique. However, it has been proved that the current DNNs are severely over-parameterized with numerous amounts of redundancy. In addition, there is no explicit theory basis for determining the exact architecture of DNNs such as the channel configuration of each layer. Then, numerous studies have been focusing on designing the high-efficient DNNs by cutting off the redundant computation and memory overheads. In this paper, we propose an automatic architecture optimization method for designing lightweight and high-efficient DNNs. Based on the proposal, a lightweight DNN model for a specific AI task is generated automatically with the exact configuration of the architecture.

要約(英語): As deep neural networks (DNNs) achieve great success in various domains, the heavy intensity of computation and memory is also a serious burden that blocks the development and applications of the technique. However, it has been proved that the current DNNs are severely over-parameterized with numerous amounts of redundancy. In addition, there is no explicit theory basis for determining the exact architecture of DNNs such as the channel configuration of each layer. Then, numerous studies have been focusing on designing the high-efficient DNNs by cutting off the redundant computation and memory overheads. In this paper, we propose an automatic architecture optimization method for designing lightweight and high-efficient DNNs. Based on the proposal, a lightweight DNN model for a specific AI task is generated automatically with the exact configuration of the architecture.

本誌: 2022年11月5日制御研究会

本誌掲載ページ: 7-11 p

原稿種別: 英語

PDFファイルサイズ: 338 Kバイト