2023 Fiscal Year Final Research Report
Designing deep network architectures to solve domain shift
| Project/Area Number |
21K17756
|
|---|
| Research Category |
Grant-in-Aid for Early-Career Scientists
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Basic Section 61010:Perceptual information processing-related
|
|---|
| Research Institution | Tohoku University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2021-04-01 – 2024-03-31
|
| Keywords | 深層学習 / 画像認識 |
|---|
| Outline of Final Research Achievements |
Deep neural networks can solve various problems with high accuracy, but they face the challenge of not performing as expected when dealing with data that slightly differs from the training data (the domain shift problem). To address this issue, various deep neural networks were evaluated and analyzed with the goal of discovering network structures that are robust to domain shifts. As an example of domain shift, this study tackled the problem of image classification for degraded images and was able to gain insights into network structures that demonstrate robustness to domain shifts.
|
| Free Research Field |
コンピュータビジョン
|
| Academic Significance and Societal Importance of the Research Achievements |
深層ニューラルネットワークは，近年発展の著しい人工知能（AI）の中枢技術であり，実応用も数多く行われている．しかしながら，実応用の上で頻出するドメインシフト問題に対して決定打となる方法がないという課題があった．本研究は，ネットワーク構造の観点からこの問題に取り組み，ネットワーク構造のドメインシフトに与える影響について，多くの知見を得た．これらの成果は，今後の研究や実応用において参考になりうる．
|
