Weight Exchange in Decentralized Distributed Machine Learning for Resource-Constrained IoT Edges

• Item type: Symposium(1)
• 公開日: 2022-01-28
• タイトル: Weight Exchange in Decentralized Distributed Machine Learning for Resource-Constrained IoT Edges
• タイトル / 言語: en
• タイトル: Weight Exchange in Decentralized Distributed Machine Learning for Resource-Constrained IoT Edges
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: Tokyo Institute of Technology
• 著者所属: Tokyo Institute of Technology
• 著者所属(英): en
• 著者所属(英): Tokyo Institute of Technology
• 著者所属(英): en
• 著者所属(英): Tokyo Institute of Technology
• 著者名: Naoya, Yokota; Yuko, Hara-Azumi
• 著者名(英): Naoya, Yokota; Yuko, Hara-Azumi
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Although a Gossip Stochastic Gradient Descent (SGD) algorithm is known to be suitable for decentralized distributed machine learning, it has a non-convergence problem for heterogeneous datasets between multiple devices. In this paper, we propose a Gossip Swap SGD to address this problem by employing a weight swapping method between devices. Our evaluation demonstrated that our proposed method successfully improves higher accuracy without increasing computation load than the original Gossip SGD.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Although a Gossip Stochastic Gradient Descent (SGD) algorithm is known to be suitable for decentralized distributed machine learning, it has a non-convergence problem for heterogeneous datasets between multiple devices. In this paper, we propose a Gossip Swap SGD to address this problem by employing a weight swapping method between devices. Our evaluation demonstrated that our proposed method successfully improves higher accuracy without increasing computation load than the original Gossip SGD.
• 書誌情報: Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform; 巻 2021, p. 94-95, 発行日 2022-01-28
• 出版者 / 言語: ja
• 出版者: 情報処理学会