WEKO3
アイテム
Distributed K-FAC Over Unstable Networks (unreferred)
https://ipsj.ixsq.nii.ac.jp/records/237574
https://ipsj.ixsq.nii.ac.jp/records/2375747303a3f6-f1f5-4db0-8a8b-7122d1cefba8
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
IPSJ-HPC24195013.pdf (1.7 MB)
2026年8月1日からダウンロード可能です。
|
Copyright (c) 2024 by the Information Processing Society of Japan
|
|
| 非会員：¥660, IPSJ:学会員：¥330, HPC:会員：¥0, DLIB:会員：¥0 | ||
| Item type | SIG Technical Reports(1) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 公開日 | 2024-08-01 | |||||||||
| タイトル | ||||||||||
| タイトル | Distributed K-FAC Over Unstable Networks (unreferred) | |||||||||
| タイトル | ||||||||||
| 言語 | en | |||||||||
| タイトル | Distributed K-FAC Over Unstable Networks (unreferred) | |||||||||
| 言語 | ||||||||||
| 言語 | eng | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | 深層学習 | |||||||||
| 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 | |||||||||
| 著者所属 | ||||||||||
| Graduate School of Science and Technology, University of Tsukuba | ||||||||||
| 著者所属 | ||||||||||
| Center for Computational Sciences, University of Tsukuba | ||||||||||
| 著者所属(英) | ||||||||||
| en | ||||||||||
| Graduate School of Science and Technology, University of Tsukuba | ||||||||||
| 著者所属(英) | ||||||||||
| en | ||||||||||
| Center for Computational Sciences, University of Tsukuba | ||||||||||
| 著者名 |
Mingzhe, Yu
× Mingzhe, Yu
× Osamu, Tatebe
|
|||||||||
| 著者名(英) |
Mingzhe, Yu
× Mingzhe, Yu
× Osamu, Tatebe
|
|||||||||
| 論文抄録 | ||||||||||
| 内容記述タイプ | Other | |||||||||
| 内容記述 | This study introduces a novel fault-tolerance approach for distributed training of deep learning models using the K-FAC optimizer under unstable network conditions. Inspired by the Federated Averaging algorithm, our method departs from traditional Distributed Data Parallel frameworks by periodically averaging weights among online nodes. We demonstrate that this approach significantly mitigates the impact of network instability on model training, effectively maintaining model accuracy even under high offline probabilities. The experimental results reveal that offline probability greatly affects test accuracy, while the maximum duration of offline iterations and the number of concurrently offline nodes exert lesser impacts. Our findings suggest that the block-diagonal approximation of the Fisher Information Matrix (FIM) used in K-FAC remains effective for guiding gradient descent, even with heterogeneous and outdated information. The study lays a groundwork for applying these insights to asynchronous training and federated learning in similar conditions. | |||||||||
| 論文抄録(英) | ||||||||||
| 内容記述タイプ | Other | |||||||||
| 内容記述 | This study introduces a novel fault-tolerance approach for distributed training of deep learning models using the K-FAC optimizer under unstable network conditions. Inspired by the Federated Averaging algorithm, our method departs from traditional Distributed Data Parallel frameworks by periodically averaging weights among online nodes. We demonstrate that this approach significantly mitigates the impact of network instability on model training, effectively maintaining model accuracy even under high offline probabilities. The experimental results reveal that offline probability greatly affects test accuracy, while the maximum duration of offline iterations and the number of concurrently offline nodes exert lesser impacts. Our findings suggest that the block-diagonal approximation of the Fisher Information Matrix (FIM) used in K-FAC remains effective for guiding gradient descent, even with heterogeneous and outdated information. The study lays a groundwork for applying these insights to asynchronous training and federated learning in similar conditions. | |||||||||
| 書誌レコードID | ||||||||||
| 収録物識別子タイプ | NCID | |||||||||
| 収録物識別子 | AN10463942 | |||||||||
| 書誌情報 |
研究報告ハイパフォーマンスコンピューティング（HPC） 巻 2024-HPC-195, 号 13, p. 1-7, 発行日 2024-08-01 |
|||||||||
| ISSN | ||||||||||
| 収録物識別子タイプ | ISSN | |||||||||
| 収録物識別子 | 2188-8841 | |||||||||
| 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 | |||||||||
