2023 Fiscal Year Annual Research Report
高信頼識別のための最適リジェクトの理論および応用研究
| Project/Area Number |
22KJ2398
|
|---|
| Allocation Type | Multi-year Fund |
|---|
| Research Institution | Kyushu University |
|---|
Principal Investigator |
JI XIAOTONG 九州大学, システム情報科学府, 特別研究員(DC1)
|
|---|
| Project Period (FY) |
2023-03-08 – 2024-03-31
|
| Keywords | Reliability / Learning with Rejection / Top-rank learning / machine learning / feature / patter recognition / outlier |
|---|
| Outline of Annual Research Achievements |
My research aimed at enhancing the reliability of machine learning models, particularly in contexts where errors are intolerable, such as medical image recognition and signature verification.
Year 1: the focus was on optimizing feature spaces for rejection. I developed a pioneering method employing ranking algorithms, accepted at the 15th IAPR International Workshop on Document Analysis System. This approach prioritized "absolute" positive samples, significantly improving model reliability.
Year 2: I delved deeper into rejection operations and top-rank learning. I proposed a novel framework for highly reliable signature verification, which is accepted by Pattern Recognition. This framework, integrating rejection methods and top-rank learning, addresses the fundamental challenge of ensuring high reliability in scenarios where errors are unacceptable. By selectively removing samples with ambiguous confidence scores and prioritizing "absolute" positive samples, the model's reliability was significantly enhanced.
Year 3: I focused on refining the top-rank learning approach. While this methodology provides high reliability by focusing solely on top-ranking samples, it is vulnerable to outliers. To address this, I combined rejection methods with top-rank learning, aiming to mitigate the impact of outliers. This approach ensures that the model maintains robustness and reliability by rejecting outliers during the training stage. The resulting framework was submitted to an international conference.
|
