| Project/Area Number |
18K19785
|
|---|
| Research Category |
Grant-in-Aid for Challenging Research (Exploratory)
|
| Allocation Type | Multi-year Fund |
|---|
| Review Section |
Medium-sized Section 60:Information science, computer engineering, and related fields
|
|---|
| Research Institution | Osaka Prefecture University |
|---|
Principal Investigator |
Iwamura Masakazu 大阪府立大学, 工学(系)研究科(研究院), 准教授 (80361129)
|
|---|
| Project Period (FY) |
2018-06-29 – 2020-03-31
|
| Project Status |
Completed (Fiscal Year 2019)
|
| Budget Amount *help |
¥6,370,000 (Direct Cost: ¥4,900,000、Indirect Cost: ¥1,470,000)
Fiscal Year 2019: ¥650,000 (Direct Cost: ¥500,000、Indirect Cost: ¥150,000)
Fiscal Year 2018: ¥5,720,000 (Direct Cost: ¥4,400,000、Indirect Cost: ¥1,320,000)
|
|---|
| Keywords | 深層学習 / 物体認識 / 学習の阻害 / 正則化 |
|---|
| Outline of Final Research Achievements |
In this research, on the regularization method we have proposed and named ShakeDrop which improves the accuracy of object recognition (aka, image classification) using "interference of learning process," we have completed (1) improvement of learning ability, (2) development of a method that requires less training data, and (3) unraveling the mechanism.
Through various experiments, we have shown (1) and (2) hold.
Regarding (3), we have got an explanation that "promotion of learning process of deep learning by interference" is achieved by data augmentation in the feature space.
|
| Academic Significance and Societal Importance of the Research Achievements |
ShakeDropは、物体認識のためのデータベースであるCIFAR-100において、一時世界最高精度を達成した正則化手法である。現在は他の手法がより良い精度を達成しているが、現在最高精度を達成している手法も最高精度を達成するためにShakeDropを使用している。
本研究では、ShakeDropのメカニズムを解明し、更にShakeDropが前述のCIFAR-100データベース以外においても高い認識精度を達成できることを実験的に示した。
|
