| Item type |
Symposium(1) |
| 公開日 |
2022-01-28 |
| タイトル |
|
|
タイトル |
Edge Domain Adaptation through Stepwise Cross-Domain Distillation |
| タイトル |
|
|
言語 |
en |
|
タイトル |
Edge Domain Adaptation through Stepwise Cross-Domain Distillation |
| 言語 |
|
|
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 |
|
|
Tokyo Institute of Technology |
| 著者所属(英) |
|
|
|
en |
|
|
Tokyo Institute of Technology |
| 著者名 |
Taisei, Yamana
Yuko, Hara-Azumi
|
| 著者名(英) |
Taisei, Yamana
Yuko, Hara-Azumi
|
| 論文抄録 |
|
|
内容記述タイプ |
Other |
|
内容記述 |
Machine learning is now required to be built on embedded systems to realize edge-AI devices, where not only weight reduction but also accuracy degradation that stems from domain shift need to be addressed. This paper proposes Stepwise Cross-Domain Distillation (SCDD) that employs unsupervised domain adaptation for lightweight models. By distilling knowledge from a pre-domain-adapted large model stepwisely through a teaching assistant model, the final lightweight student model can effectively achieve good accuracy in a target domain. We also provide insights obtained through quantitative evaluations to improve stepwise knowledge distillation in various domain shifts.Code is available at https://github.com/TaiseiYamana/SCDD.git |
| 論文抄録(英) |
|
|
内容記述タイプ |
Other |
|
内容記述 |
Machine learning is now required to be built on embedded systems to realize edge-AI devices, where not only weight reduction but also accuracy degradation that stems from domain shift need to be addressed. This paper proposes Stepwise Cross-Domain Distillation (SCDD) that employs unsupervised domain adaptation for lightweight models. By distilling knowledge from a pre-domain-adapted large model stepwisely through a teaching assistant model, the final lightweight student model can effectively achieve good accuracy in a target domain. We also provide insights obtained through quantitative evaluations to improve stepwise knowledge distillation in various domain shifts.Code is available at https://github.com/TaiseiYamana/SCDD.git |
| 書誌情報 |
Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform
巻 2021,
p. 1-7,
発行日 2022-01-28
|
| 出版者 |
|
|
言語 |
ja |
|
出版者 |
情報処理学会 |
IPSJ-APRIS2021001.pdf (2.3 MB)
