Swarm Learning for Deep Neural Networks
| Project/Area Number |
17K12734
|
|---|
| Research Category |
Grant-in-Aid for Young Scientists (B)
|
| Allocation Type | Multi-year Fund |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Tokyo Institute of Technology |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2017-04-01 – 2022-03-31
|
| Project Status |
Completed (Fiscal Year 2021)
|
| Budget Amount *help |
¥4,550,000 (Direct Cost: ¥3,500,000、Indirect Cost: ¥1,050,000)
Fiscal Year 2020: ¥650,000 (Direct Cost: ¥500,000、Indirect Cost: ¥150,000)
Fiscal Year 2019: ¥650,000 (Direct Cost: ¥500,000、Indirect Cost: ¥150,000)
Fiscal Year 2018: ¥2,340,000 (Direct Cost: ¥1,800,000、Indirect Cost: ¥540,000)
Fiscal Year 2017: ¥910,000 (Direct Cost: ¥700,000、Indirect Cost: ¥210,000)
|
|---|
| Keywords | 半教師あり学習 / 群知能 / 深層学習 / 機械学習 / 合意形成 / 自己学習 / 分散学習 |
|---|
| Outline of Final Research Achievements |
Deep learning usually requires a huge amount of training data to compute a precise model. Preparing training data is costly, and hence it is a drawback of the deep leaning. This study aims at reducing the amount of training data while keeping the preciseness of the computed model. To achieve this objective, this project proposes a new semi-supervised learning method inspired by the studies in the field of swarm intelligence. The effectiveness of the proposed method is evaluated for various tasks, and the training mechanisms behind the proposed method are experimentally analysed.
On the one hand, the project concluded that the proposed model is competitive to other methods proposed in related work in terms of preciseness, provided that it is effective for some specific tasks and when the computation cost does not matter. On the other hand, the project also figured out the advantage of the proposed method; it can reduce the cost to design an optimal network structure beforehand.
|
| Academic Significance and Societal Importance of the Research Achievements |
深層学習はいまやAI研究の根幹をなす技術であり，自動運転車における周辺環境の知覚タスクや，カメラ映像などを入力とした自動監視システムなど，さまざまな実応用がなされている．今後さまざまな新たなサービスやシステムへの適用も期待される技術である一方，特に新規参入分野では訓練データの収集や最適なネットワーク構造の設計などに多大なコストを要求される．本研究は，こうしたコストを削減することが期待できる．提案手法は，訓練データ数の削減については競合手法と同程度の性能を保ちつつ，複数の有望なネットワーク構造の並列学習によって，利用者に最終的に使用するモデルについて複数の選択肢を与えることが可能である．
|
Report
(6 results)
Research Products
(8 results)
