2022 Fiscal Year Annual Research Report
大規模IoTデータ知識獲得のための信頼される分散型人工知能
| Project/Area Number |
22J12681
|
|---|
| Allocation Type | Single-year Grants |
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
SUN YUWEI 東京大学, 情報理工学系研究科, 特別研究員(DC2)
|
|---|
| Project Period (FY) |
2022-04-22 – 2024-03-31
|
| Keywords | machine learning / neural networks / life-long learning / AI security / data privacy / decentralized ML / edge computing / multi-modal |
|---|
| Outline of Annual Research Achievements |
The project aims to address the fundamental challenges of decentralized deep learning (DDL) to make AI feasible and scalable for everyone. The proliferation of edge AI applications has been reshaping the contours of future high-performance edge computing, and DDL is a key enabler that would benefit society through distributed model training and globally shared knowledge. In the first phase of the project, we focused on developing collaborative representation learning techniques for different neural network models at the edge. We identified critical challenges and proposed new solutions to privacy protection, edge heterogeneity, and adversarial attacks and defenses in DDL. The proposed methods enable improved model generality to unseen data and robustness against adversarial attacks.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
Representation learning in collaborative and competing neural network models could improve the systems' generality to unseen samples by reusing previous knowledge and experience. Therefore, we set the goal for the first fiscal year as implementing a decentralized deep learning (DDL) framework that incorporates collaborative representation learning techniques and privacy protection measures. Regarding this goal, we focused on developing collaborative representation learning techniques identifying the critical challenges of privacy protection, edge heterogeneity, and adversarial robustness. Experiments were to conducted to evaluate effectiveness and trustworthiness. Given these reasons, we judged that this fiscal year has progressed as planned, and this project is progressing well.
|
| Strategy for Future Research Activity |
In the next phase of the project, we would focus on building reusable knowledge components in conventional large foundation models such as Transformer models. We aim to understand how a decentralized framework coupled with reusable knowledge representations would facilitate an intelligence systems that is capable of life-long learning and swiftly adapting to unseen situations by reusing learned knowledge from past experiences and tasks. We would also continue to investigate such systems' robustness to adversarial attacks and devise effective defense mechanisms. In addition, we look forward to further testing and validating our solutions in real-world scenarios for the learning with multi-modal edge data.
|
Research Products
(10 results)
