Efficient predictive density for transfer learning under high-dimensional settings
| Project/Area Number |
17H06570
|
|---|
| Research Category |
Grant-in-Aid for Research Activity Start-up
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
Statistical science
|
|---|
| Research Institution | The University of Tokyo |
|---|
Principal Investigator |
Yano Keisuke 東京大学, 大学院情報理工学系研究科, 助教 (20806070)
|
|---|
| Research Collaborator |
Andrew Barron
Kato Kengo
Komaki Fumiyasu
Gourab Mukherjee
|
|---|
| Project Period (FY) |
2017-08-25 – 2019-03-31
|
| Project Status |
Completed (Fiscal Year 2018)
|
| Budget Amount *help |
¥1,950,000 (Direct Cost: ¥1,500,000、Indirect Cost: ¥450,000)
Fiscal Year 2018: ¥910,000 (Direct Cost: ¥700,000、Indirect Cost: ¥210,000)
Fiscal Year 2017: ¥1,040,000 (Direct Cost: ¥800,000、Indirect Cost: ¥240,000)
|
|---|
| Keywords | 予測分布 / 高次元統計 / 擬似ベイズ / 機械学習 / 分布予測 / 転移学習 / 統計数学 / 数理工学 |
|---|
| Outline of Final Research Achievements |
This research develops A)the strategies of reducing the computational cost of Bayesian methods by quasi-posteriors; B)the efficient predictive densities for sparse count data.
In Research A, theoretical properties of quasi-posteriors (posteriors based on handy or mis-specified likelihoods) such as information losses or performance in uncertainty quantification are studied under high dimensional settings. This research shows a possibility of constructing predictive densities with both low computational costs and high performance by leveraging quasi-posteriors.
In Research B, efficient predictive densities for sparse count data are constructed. This research shows compatibility of high performance and low computational cost in constructing predictive densities under high dimensional settings by focusing on sparsity or quasi-sparsity of data.
|
| Academic Significance and Societal Importance of the Research Achievements |
予測とは,現在の観測量をもとに予測したい量(予測量)の振る舞いを推測することで
ある.地震予測,交通予測,遺伝子機能予測等,様々な予測が社会で活用されている.統計的な予測手法には,予測量の平均を推定する点予測と予測量の従う分布を推定する分布予測がある.予測量の従う分布が分かれば,検定や予測区間の構成ができるため,分布予測がより重要である.
転移学習とは,ある領域での観測量を利用して別の領域にある予測量を予測することである.転移学習は統計学と機械学習で近年注目されており,例えば、深層学習の精度向上に利用されている.転移学習の理論的性質が分かると,既存の学習手法の精度は飛躍的に向上するため重要である.
|
Report
(3 results)
Research Products
(11 results)
