| 研究課題/領域番号 |
22KJ0878
|
|---|
| 補助金の研究課題番号 |
22J12681 (2022)
|
|---|
| 研究種目 |
特別研究員奨励費
|
| 配分区分 | 基金 (2023)
補助金 (2022) |
|---|
| 応募区分 | 国内 |
|---|
| 審査区分 |
小区分61030:知能情報学関連
|
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
SUN YUWEI 東京大学, 情報理工学系研究科, 特別研究員(DC2)
|
|---|
| 研究期間 (年度) |
2023-03-08 – 2024-03-31
|
| 研究課題ステータス |
完了 (2023年度)
|
| 配分額 *注記 |
1,700千円 (直接経費: 1,700千円)
2023年度: 800千円 (直接経費: 800千円)
2022年度: 900千円 (直接経費: 900千円)
|
|---|
| キーワード | transfer learning / multi-modal / Transformer / modularity / attention / machine learning / neural networks / life-long learning / AI security / data privacy / decentralized ML / edge computing |
|---|
| 研究開始時の研究の概要 |
This project aims to develop a trustworthy decentralized AI framework for the life-long representation learning within multi-modal AI models. A feasible and scalable system hinges on overcoming key challenges of reusable knowledge transfer, communication efficiency, and adversarial robustness.
|
| 研究実績の概要 |
This project delved into localized learning in multi-modal distributed machine learning to address the out-of-distribution problem. The study initially explored knowledge transfer among a group of expert models observing partial environments in a federated learning setup. It demonstrated that generalization could be attained through the coordination of localized models by extracting domain-invariant knowledge with a global model. An additional approach is the Markov chain-based Homogeneous Learning, where a meta-observer learns an efficient communication policy of individual models.
Overall, this project proposed novel approaches to reusable neural modules for distributed machine learning in real-world learning settings. The project contributed to invited talks and multiple publications in both journals and top conferences in the field.
|
