A Graph Theoretic Framework of Recomputation Algorithms for Memory-Efficient Backpropagation

• Item type: SIG Technical Reports(1)
• 公開日: 2020-05-02
• タイトル: A Graph Theoretic Framework of Recomputation Algorithms for Memory-Efficient Backpropagation
• タイトル / 言語: en
• タイトル: A Graph Theoretic Framework of Recomputation Algorithms for Memory-Efficient Backpropagation
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: 招待講演
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者名: Mitsuru, Kusumoto; Takuya, Inoue; Gentaro, Watanabe; Takuya, Akiba; Masanori, Koyama
• 著者名(英): Mitsuru, Kusumoto; Takuya, Inoue; Gentaro, Watanabe; Takuya, Akiba; Masanori, Koyama
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Recomputation algorithms collectively refer to a family of methods that aims to reduce the memory consumption of the backpropagation by selectively discarding the intermediate results of the forward propagation and recomputing the discarded results as needed. In this paper, we will propose a novel and efficient recomputation method that can be applied to a wider range of neural nets than previous methods. We use the language of graph theory to formalize the general recomputation problem of minimizing the computational overhead under a fixed memory budget constraint, and provide a dynamic programming solution to the problem. Our method can reduce the peak memory consumption on various benchmark networks by 36%～81%, which outperforms the reduction achieved by other methods.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Recomputation algorithms collectively refer to a family of methods that aims to reduce the memory consumption of the backpropagation by selectively discarding the intermediate results of the forward propagation and recomputing the discarded results as needed. In this paper, we will propose a novel and efficient recomputation method that can be applied to a wider range of neural nets than previous methods. We use the language of graph theory to formalize the general recomputation problem of minimizing the computational overhead under a fixed memory budget constraint, and provide a dynamic programming solution to the problem. Our method can reduce the peak memory consumption on various benchmark networks by 36%～81%, which outperforms the reduction achieved by other methods.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AN1009593X
• 書誌情報: 研究報告アルゴリズム（AL）; 巻 2020-AL-178, 号 4, p. 1-1, 発行日 2020-05-02
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8566
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会